text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'Following the methods used by Derksen-Weyman in [@DW11] and Chindris in [@Chi08], we use quiver theory to represent the generalized Littlewood-Richardson coefficients for the branching rule for the diagonal embedding of ${\text{GL}}(n)$ as the dimension of a weight space of semi-invariants. Using this, we prove their saturation and investigate when they are nonzero. We also show that for certain partitions the associated stretched polynomials satisfy the same conjectures as single Littlewood-Richardson coefficients. We then provide a polytopal description of this multiplicity and show that its positivity may be computed in strongly polynomial time. Finally, we remark that similar results hold for certain other generalized Littlewood-Richardson coefficients.'
address: '<span style="font-variant:small-caps;">Mathematics Department, Fitchburg State University, 160 Pearl St, MA, 01420</span>'
author:
- Brett Collins
title: 'Generalized Littlewood-Richardson coefficients for branching rules of ${\text{GL}}(n)$ and extremal weight crystals'
---
[^1]
Introduction
============
Context and motivation
----------------------
Littlewood-Richardson coefficients appear in many contexts in representation theory, such as the coefficients in the decomposition of a product of symmetric polynomials or the tensor product of irreducible representations of ${\text{GL}}(n)$. One way to define Littlewood-Richardson coefficients is as follows: let $V$ be a complex vector space of dimension $n$ and $\lambda = (\lambda_1,\ldots, \lambda_n)$ a weakly decreasing sequence of $n$ integers. Denote the irreducible rational representation of ${\text{GL}}(n)$ with highest weight $\lambda$ by $S^\lambda(V)$. Given three weakly decreasing sequences of $n$ integers $\lambda(1), \lambda(2), \lambda(3)$, the Littlewood-Richardson coefficient $c^{\lambda(2)}_{\lambda(1),\lambda(3)}$ is defined to be the multiplicity of $S^{\lambda(2)}(V)$ in $S^{\lambda(1)}(V)\otimes S^{\lambda(3)}(V)$, that is, $$c^{\lambda(2)}_{\lambda(1),\lambda(3)} = \operatorname{dim}_{\mathbb{C}}{\text{Hom}}_{{\text{GL}}(V)}(S^{\lambda(2)}(V), S^{\lambda(1)}(V) \otimes S^{\lambda(3)}(V)).$$
Similarly, sums of products of these coefficients, which we call generalized Littlewood-Richardson coefficients throughout this paper, appear naturally in the decompositions of various algebraic objects. In particular, generalized Littlewood-Richardson coefficients describe the multiplicities in the branching rules of restricted representations of ${\text{GL}}(n)$, as described in [@HJ09] and [@HTW05]. While there is no known way to describe the multiplicities in each of these branching rules using quiver theory, we show that we can do exactly such for one of them, that is, we describe the coefficients as the dimension of a weight space of semi-invariants for a certain quiver, dimension vector, and weight. More generally, we do this for the generalized Littlewood-Richardson coefficient $$\label{one}
f(\lambda(1),\ldots, \lambda(m)) :=
\sum c^{\lambda(1)}_{\alpha(1),\alpha(2)} c^{\lambda(2)}_{\alpha(2),\alpha(3)}\cdots c^{\lambda(m-1)}_{\alpha(m-1),\alpha(m)} c^{\lambda(m)}_{\alpha(m),\alpha(1)}$$ for $m \geq 4$ and even, where the summation ranges over all partitions $\alpha(i)$. This multiplicity describes the coefficients arising from the branching rule for the diagonal embedding ${\text{GL}}(n) \subseteq {\text{GL}}(n) \times {\text{GL}}(n)$ in the case $m=6$. We remark in the that similar techniques likewise represent two other generalized Littlewood-Richardson coefficients in this manner.
Recently, Littlewood-Richardson coefficients have been of vital interest in geometric complexity theory which seeks to determine the complexity of computational problems by using tools from algebraic geometry and representation theory to provide lower bounds, and the complexity is quite commonly compared to that of computing certain multiplicities like Littlewood-Richardson coefficients. Understanding the complexity of certain cases of these generalized coefficients or even whether they’re nonzero can then be used in comparison to other computational problems. A common technique in combinatorics is to associate a polytope to a multiplicity in such a way that the number of lattice points of the polytope is precisely this number. Because the polytope is defined by a system of linear inequalities, combinatorial optimization may then be used to determine the complexity of the multiplicities as well as the properties of the polytope.
Knutson and Tao [@KT99] provided a polytopal description of Littlewood-Richardson coefficients, allowing them to give a combinatorial proof of the saturation of the coefficients (Theorem \[LR-saturation\]) and complete the proof of Horn’s conjecture (Theorem \[Horn\]). Derksen and Weyman [@DW00a] then reproved the saturation property in the context of quiver representations by using the saturation of weight spaces of semi-invariants. The motivation for this paper may then be summarized as providing an explicit quiver theoretic interpretation of these generalized Littlewood-Richardson coefficients in order to prove their saturation and use results of quiver theory to study their combinatorial and geometric properties.
Main results
------------
One of the main and most useful results is that of the saturation of this multiplicity.
\[saturation-sun\] Let $\lambda(1),\ldots, \lambda(2k)$ be weakly decreasing sequences of $n$ integers for $k \geq 2$. For every integer $r \geq 1$, $$f(r\lambda(1),\ldots, r\lambda(2k)) \neq 0 \Longleftrightarrow
f(\lambda(1),\ldots, \lambda(2k)) \neq 0.$$
We extend the results of [@KT99] by using their hive models to provide a polytopal description of the generalized Littlewood-Richardson coefficients in Section \[polytope\]. By using results in combinatorial optimization theory and the above saturation property, we prove the following theorem.
\[sun-complexity\] Determining whether the multiplicity *(\[one\])* is positive or not can be decided in polynomial time. Even more, it can be decided in strongly polynomial time in the sense of [@Tar86].
Horn’s conjecture (Theorem \[Horn\]) relates the set of possible eigenvalues arising from a sum of Hermitian matrices to the nonvanishing of Littlewood-Richardson coefficients. To describe a corresponding statement of the conjecture to multiplicity (\[one\]), we need to define some notation. For an $m$-tuple $(I_1,\ldots, I_m)$ of subsets of $\{1,\ldots, n\}$, define the following weakly decreasing sequences of integers (the notation is explained in the at the end of this section): $${\underline{\lambda}}(I_i) = \begin{cases}
\lambda'(I_i) & i \text{ even} \\
\lambda'(I_i) -((|I_i| - |I_{i-1}| - |I_{i+1}|)^{n-|I_i|}) & i \text{ odd},
\end{cases}$$ where we identify $I_0$ and $I_m$. Define the set $K(n,m) \subseteq {\mathbb{R}}^{mn}$, $m \geq 4$ and even, to be all $m$-tuples $(\lambda(1),\ldots, \lambda(m))$ of weakly decreasing sequences of $n$ reals that satisfy $\sum_{i \text{ even}} |\lambda(i)| = \sum_{i \text{ odd}} |\lambda(i)|$ and $$\sum_{j \in I_i} \sum_{i \text{ even}} \lambda(i)_j \leq \sum_{j \in I_i} \sum_{i \text{ odd}} \lambda(i)_j$$ for every tuple $(I_1, \ldots, I_m)$ such that ${\underline{\lambda}}(I_i)$, $1 \leq i \leq m$, are partitions and $$f({\underline{\lambda}}(I_1), \ldots, {\underline{\lambda}}(I_m)) \neq 0,$$ This makes $K(n,m)$ a rational convex polyhedral cone in ${\mathbb{R}}^{mn}$. A corresponding statement of Horn’s conjecture for this multiplicity is then as follows, where we describe the generalized eigenvalue problem for $f$ in Section \[Hermitian\].
\[generalization-Horn\] The following statements are true.
1. The cone $K(n,m) \subseteq {\mathbb{R}}^{nm}$, $m \geq 4$ and even, consists of all $m$-tuples $(\lambda(1),\ldots, \lambda(m))$ of weakly decreasing sequences of $n$ reals satisfying $\sum_{i \text{ even}} |\lambda(i)| = \sum_{i \text{ odd}} |\lambda(i)|$ and $$\sum_{j \in I_i} \sum_{i \text{ even}} \lambda(i)_j \leq \sum_{j \in I_i} \sum_{i \text{ odd}} \lambda(i)_j$$ for every tuple $(I_1, \ldots, I_m)$ such that the ${\underline{\lambda}}(I_i)$, $1 \leq i \leq m$, are partitions and $$f({\underline{\lambda}}(I_1), \ldots, {\underline{\lambda}}(I_m)) = 1.$$
2. If $(\lambda(1),\ldots, \lambda(m)) \in K(n,m)$, then the tuple satisfies the generalized eigenvalue problem for $f$.
3. If $\lambda(1),\ldots, \lambda(m)$ are weakly decreasing sequences of $n$-integers, then $$(\lambda(1),\ldots, \lambda(m)) \in K(n,m) \Longleftrightarrow f(\lambda(1),\ldots,\lambda(m)) \neq 0.$$
4. $\operatorname{dim}K(n,m) = mn-1$.
In particular, this provides a recursive procedure for finding all nonzero generalized Littlewood-Richardson coefficients of this type. We use this description to describe all facets of the cone of effective weights in the case $n=2,\,m=6$ and find the minimal set of inequalities on the $\lambda(i)$ (see Example \[n=2\] and the ).
One consequence of this description of the sequences in the cone $K(n,m)$ is the following factorization formula.
\[factorization\] Let $(\lambda(1), \ldots, \lambda(m)) \in K(n,m) \cap {\mathbb{Z}}^{mn-1}$. For any tuple of subsets $I=(I_1, \ldots, I_m)$ of $S=\{1,\ldots, n\}$ satisfying the conditions defining $K(n,m)$, we have the factorization $$f(\lambda(1),\ldots, \lambda(m)) = f(\lambda(1)^*,\ldots, \lambda(m)^*) \cdot f(\lambda(1)^\#,\ldots, \lambda(m)^\#),$$ where $$\lambda(p)^* = (\lambda(p)_{i_{j_1}}, \ldots, \lambda(p)_{i_{j_r}}), \quad I_j = \{i_{j_1}, \ldots, i_{j_r}\}, \qquad \lambda(p)^\# = (\lambda(p)_{\tilde{i}_{j_1}}, \ldots, \lambda(p)_{\tilde{i}_{j_{n-r}}}), \quad S \backslash I_j = (\tilde{i}_{j_1}, \ldots, \tilde{i}_{j_{n-r}}).$$
In addition, we investigate the stretched Littlewood-Richardson polynomials $f(N\lambda(1),\ldots, N\lambda(m))$ for certain $\lambda(1),\ldots, \lambda(m)$ in Section \[stretched-polynomials\]. The tuples we investigate turn out to have the same behavior as the stretched Littlewood-Richardson polynomials for a single coefficient, namely, they satisfy conjectures of King, Tollu, and Toumazet [@KTT04] and of Fulton, providing evidence that the conjectures for the stretched polynomials for a single Littlewood-Richardson coefficient extend to those of generalized Littlewood-Richardson coefficients. Our examples are based on corresponding examples in [@Fei15] for the quiver associated to a single Littlewood-Richardson coefficient. As opposed to Fei’s examples, ours do not always lie on an extremal ray of the cone of effective weights.
The organization of this paper is as follows. In Section \[preliminaries\] we provide background on quiver invariant theory and state a certain saturation property for effective weights of quivers proven by Derksen and Weyman [@DW00a]. The quiver associated to multiplicity (\[one\]) is defined in Section \[section-saturation\] and its saturation property is proven. After recalling more detailed descriptions of the facets of the cone of effective weights for acyclic quivers in Section \[section-facets\], we describe the facets of our quiver in Section \[section-facets-quivers\], which allows a description of the Horn-type inequalities of the multiplicity in Section \[section-Horn\]. A moment map description of the cone associated to our quiver is provided by the generalized eigenvalue problem in Section \[Hermitian\] while we use the Horn-type inequalities to prove a factorization formula in Section \[section-factorization\]. In Section \[stretched-polynomials\], we explicitly calculate the stretched Littlewood-Richardson polynomials in certain cases and verify that they share some of the same properties as the stretched polynomials for single Littlewood-Richardson coefficients. We provide a polytopal description of the multiplicity in Section \[polytope\] and prove the complexity of computing its positivity. Finally, we discuss in Section \[section-others\] that our methodology can be used to prove similar results for other generalized Littlewood-Richardson coefficients, in particular, for a multiplicity arising from another branching rule of ${\text{GL}}(n)$ and a multiplicity related to extremal weight crystals. We state without proof the corresponding main results for these multiplicities.
Relation to existing literature
-------------------------------
Horn made his famous in 1962 [@Hor62], yet the motivation for it goes back to Weyl in 1912 [@Wey12]. Weyl was interested in necessary and sufficient inequalities on the eigenvalues of Hermitian matrices such that one matrix was the sum of the other two due to questions in solid mechanics. Many advances were made over the next 50 years (see [@Ful97b] for a survey of the history and results pertaining to this problem), and the connection was made between the eigenvalue problem and Schubert calculus, resulting in Horn’s conjecture.
The second part of the conjecture provides a recursive process for determining all triples $(I_1,I_2,I_3)$ of subsets of $\{1,\ldots, n\}$ which are necessary to determine if $(\lambda(1),\lambda(2), \lambda(3))$ is such a solution. The first major step in proving the conjecture was made when Klyachko [@Kly98] found necessary and sufficient homogeneous linear inequalities for the eigenvalues. It remained, however, to find a minimal set of inequalities. Klyachko had claimed that these inequalities were independent, but Woodward [@AW98] showed that many inequalities were redundant, and later Belkale [@Bel01] showed that all the inequalities for which $c^{\lambda(I_2)}_{\lambda(I_1),\lambda(I_3)} > 1$ are redundant, which includes the set found by Woodward. The remaining inequalities would be irredundant by a theorem of Klyachko provided the saturation of Littlewood-Richardson coefficients.
\[LR-saturation\] For weakly decreasing sequences of $n$ integers $\lambda, \mu, \nu$, $c^{N\nu}_{N \lambda, N\mu} \neq 0$ for some positive $N$ if and only if $c^{\nu}_{\lambda, \mu} \neq 0.$
The statement that $c^{\nu}_{\lambda, \mu} \neq 0$ implies $c^{N\nu}_{N\lambda,N\mu} \neq 0$ for all positive integers $N$ follows immediately from the Littlewood-Richardson rule or by these multiplicities forming a semigroup (see [@Zel99]). Knutson, Tao, and Woodward [@KTW04] used combinatorial gadgets called honeycombs and hive models to prove the saturation conjecture, completing the proof of Horn’s conjecture; a similar proof of the conjecture using only hive models is found in [@Buc00]. A geometric proof of the saturation conjecture using Schubert calculus was also given by Belkale [@Bel06]. Derksen, Schofield, and Weyman [@DSW07] showed that the number of subrepresentations of a specific dimension of a given dimension vector for an acylic quiver may be determined using Schubert calculus, which is why Derksen and Weyman defined a specific quiver and dimension vector for which this number was a given Littlewood-Richardson coefficient. In this way, they were able to extend results about Littlewood-Richardson coefficients to the quiver setting and prove the saturation conjecture for Littlewood-Richardson coefficients in [@DW00a] by using results from quiver theory.
The multiplicity (\[one\]) for the branching rule of the diagonal embedding of ${\text{GL}}(n)$ was first proven in [@Kin71]. The proof may also be found in [@HTW05] (see [@Koi89] and [@HJ09] for further discussion).
Notation
--------
A partition $\lambda$ of length $n$ is a weakly decreasing sequence of $n$ positive integers, denoted $\lambda = (\lambda_1, \ldots, \lambda_n)$. We identify two partitions $\lambda$ and $\mu$ if one can be written as the other by adjoining (finitely) many zeros at the end of the sequence. As such, we say $\lambda$ is a partition of at most $n$ (nonzero) parts if $\lambda = (\lambda_1, \ldots, \lambda_n) \in {\mathbb{Z}}^n$ with $\lambda_1 \geq \ldots \geq \lambda_n \geq 0$. If $\lambda = (\lambda_1, \ldots, \lambda_n)$ and $\mu$ are weakly decreasing sequences (not necessarily of integers), we define $r \lambda = (r\lambda_1, \ldots, r\lambda_n)$ for $r \in {\mathbb{R}}^+$ and $\lambda + \mu$ is defined by adding componentwise after extending the sequences by zeros as necessary. Every partition $\lambda$ may be identified with a Young diagram, and we denote the conjugate partition as $\lambda'$, which is the partition associated to the reflection of the Young diagram of $\lambda$ across the main diagonal, i.e., switching the rows and columns. If $I=\{z_1 < \ldots < z_r\}$ is an $r$-tuple of integers, $\lambda(I)$ is defined by $\lambda(I) = (z_r-r, \ldots, z_1-1)$. For an integer $c$ and a non-negative integer $r$, we denote the $r$-tuple $(c,\ldots, c)$ more simply as $(c^r)$. For a sequence of real numbers $\lambda = (\lambda_1, \ldots, \lambda_n)$, we define $|\lambda| = \sum_{i=1}^n \lambda_i$.
For a complex vector space $V$ of dimension $n$ and a weakly decreasing sequence of $n$ integers $\lambda = (\lambda_1, \ldots, \lambda_n)$, we denote the irreducible rational representation of ${\text{GL}}(V)$ with highest weight $\lambda$ as $S^\lambda(V)$. Given any three weakly decreasing sequences of $n$ integers $\lambda(1), \lambda(2), \lambda(3)$, the Littlewood-Richardson coefficient $c_{\lambda(1), \, \lambda(3)}^{\lambda(2)}$ is defined to be $$c_{\lambda(1), \, \lambda(3)}^{\lambda(2)} = \operatorname{dim}_{\mathbb{C}}{\text{Hom}}_{{\text{GL}}(V)}(S^{\lambda(2)}(V), S^{\lambda(1)}(V) \otimes S^{\lambda(3)}(V)),$$ that is, the multiplicity of $S^{\lambda(2)}(V)$ in $S^{\lambda(1)}(V) \otimes S^{\lambda(3)}(V)$.
Preliminaries
=============
Preliminaries {#preliminaries}
-------------
A quiver $Q = (Q_0,Q_1,t,h)$ consists of a finite set of vertices $Q_0$, a finite set of arrows $Q_1$, and functions $t,h : Q_1 \to Q_0$ that assign the tail $ta$ and head $ha$ of each arrow $a$, commonly denoted $ta {\xrightarrow}{a} ha$. Note that we allow multiple arrows between two vertices and loops in the directed graph $Q$.
Throughout this paper we always work over the complex numbers ${\mathbb{C}}$. A representation $V$ of $Q$ is a family of finite-dimensional vector spaces (over ${\mathbb{C}}$) $\{V(x) \mid x \in Q_0\}$ together with a family of linear transformations $\{V(a):V(ta) \to V(ha) \mid a \in Q_1\}$. For a representation $V$, its dimension vector ${\underline{\textbf{dim}} }\, V$ is defined by ${\underline{\textbf{dim}} }\, V(x) = \operatorname{dim}_{\mathbb{C}}V(x)$ for all $x \in Q_0$. The dimension vectors of representations of $Q$ then lie in $\Gamma = {\mathbb{Z}}^{Q_0}$, the set of integer-valued functions on $Q_0.$ For each vertex $x \in Q_0$, there is a simple representation $S_x$ defined by the dimension vector $e_x(y) = \delta_{x,y}$ for all $y \in Q_0$, where $\delta_{x,y}$ is the Kronecker delta.
Given two representations $V$ and $W$ of $Q$, define a morphism $\phi : V \to W$ of representations to be a collection of linear maps $\{\phi(x): V(x) \to W(x) \mid x \in Q_0\}$ such that for every arrow $a \in Q_1$ we have $\phi(ha)V(a) = W(a) \phi(ta)$, meaning the diagram $$\xymatrix{
V(ta) \ar[r]^{\phi(ta)} \ar[d]_{V(a)} & W(ta) \ar[d]^{W(a)} \\
V(ha) \ar[r]^{\phi(ha)} & W(ha)
}$$ commutes. Define ${\text{Hom}}_Q(V,W)$, or simply ${\text{Hom}}(V,W)$, to be the ${\mathbb{C}}$-vector space of all morphisms from $V$ to $W$. We thus obtain the abelian category $\text{Rep}(Q)$ of all quiver representations of $Q$. We call $V'$ a subrepresentation of $V$ if $V'(x)$ is a subspace of $V(x)$ for all vertices $x \in Q_0$ and $V'(a) = V(a)|_{V'(ta)}$ for all arrows $a \in Q_1$.
For any $\alpha, \beta \in \Gamma$, define the Euler form by $$\langle \alpha, \beta \rangle = {\displaystyle}\sum_{x \in Q_0} \alpha(x) \beta(x) - \sum_{a \in Q_1} \alpha(ta) \beta(ha).$$
Semi-invariants for quivers
---------------------------
For a dimension vector $\beta$ of a quiver $Q$, the representation space of $\beta$-dimensional representations of $Q$ is defined as $$\operatorname{Rep(Q,\beta)}= \bigoplus_{a \in Q_1} {\text{Hom}}\left( {\mathbb{C}}^{\beta(ta)}, {\mathbb{C}}^{\beta(ha)} \right).$$ Note that this is simply an affine space. If ${\text{GL}}(\beta) = \prod_{x \in Q_0} {\text{GL}}(\beta(x))$, then there is a natural action of ${\text{GL}}(\beta)$ on $\operatorname{Rep(Q,\beta)}$ given by simultaneous conjugation: for $g = (g(x))_{x \in Q_0} \in {\text{GL}}(\beta)$ and $V = \{V(a)\}_{a \in Q_1} \in \operatorname{Rep(Q,\beta)}$, $g \cdot V$ is defined by $$(g \cdot V)(a) = g(ha)V(a)g(ta)^{-1} \quad \forall a \in Q_1.$$ Hence, $\operatorname{Rep(Q,\beta)}$ is a rational representation of the linearly reductive group ${\text{GL}}(\beta)$ and the ${\text{GL}}(\beta)$-orbits parameterize the isomorphism classes of $\beta$-dimension representations of $Q$ since the action is simply base change (with respect to a specified basis). If $Q$ is without oriented cycles, there is only one closed ${\text{GL}}(\beta)$-orbit in $\operatorname{Rep(Q,\beta)}$ (specifically, the orbit of the unique $\beta$-dimensional semisimple representation $\bigoplus_{x \in Q_0} S_x^{\beta(x)})$, so the invariant ring ${\mathbb{C}}[\operatorname{Rep(Q,\beta)}]^{{\text{GL}}(\beta)}$ is simply ${\mathbb{C}}$. However, while there are only constant ${\text{GL}}(\beta)$-invariant polynomial functions on $\operatorname{Rep(Q,\beta)}$, the action descends to that of the subgroup ${\text{SL}}(\beta)$, and the invariant ring under the action of this group is highly nontrivial.
Let ${\operatorname{SI}(Q,{\beta})} = {\mathbb{C}}[\operatorname{Rep(Q,\beta)}]^{{\text{SL}}(\beta)}$ be the ring of semi-invariants. Since ${\text{GL}}(\beta)$ is linearly reductive and ${\text{SL}}(\beta)$ is the commutator subgroup of ${\text{GL}}(\beta)$, we have the weight space decomposition $${\operatorname{SI}(Q,{\beta})} = \bigoplus_{\sigma \in X^*({\text{GL}}(\beta))} {\operatorname{SI}(Q,{\beta})}_\sigma,$$ where $X^*({\text{GL}}(\beta))$ is the group of rational characters of ${\text{GL}}(\beta)$ and $${\operatorname{SI}(Q,{\beta})}_\sigma = \{f \in {\mathbb{C}}[\operatorname{Rep(Q,\beta)}] \mid g \cdot f = \sigma(g)f \; \; \forall g \in {\text{GL}}(\beta)\}$$ is the space of semi-invariants of weight $\sigma$. A character (or weight) of ${\text{GL}}(\beta)$ is of the form $$\{g(x) \mid x \in Q_0\} \in {\text{GL}}(\beta) \mapsto \prod_{x \in Q_0} (\det g(x))^{\sigma(x)}$$ for $\sigma(x) \in {\mathbb{Z}}$ for all $x \in Q_0$, so we may identify $X^*({\text{GL}}(\beta))$ with ${\mathbb{Z}}^{Q_0}$. For an integer-valued function $\alpha$ on $Q_0$, define $\sigma = \langle \alpha, \cdot \rangle$ by $$\sigma(x) = \langle \alpha, e_x \rangle = \alpha(x) - \sum_{y \to x} \alpha(y), \quad \forall x \in Q_0.$$ One can similarly define $\sigma = \langle \cdot, \alpha \rangle.$
Given a quiver $Q$ and dimension vector $\beta$, define the set $\Sigma(Q,\beta)$ to be the set of (integral) effective weights: $$\Sigma(Q,\beta) = \{\sigma \in {\mathbb{Z}}^{Q_0} \mid {\operatorname{SI}(Q,{\beta})}_\sigma \neq 0\}.$$ Schofield [@Sch91] constructed distinguished semi-invariants for quivers that proved to be quite useful in studying the ring of semi-invariants. Derksen and Weyman [@DW00a] (see also [@SB01]) showed that these semi-invariants in fact span all spaces of semi-invariants. An important consequence of this result is the following saturation property.
\[saturation\] If $Q$ is a quiver without oriented cycles and $\beta$ is a dimension vector, then the set $$\Sigma(Q,\beta) = \{\sigma \in {\mathbb{Z}}^{Q_0} \mid {\operatorname{SI}(Q,{\beta})}_\sigma \neq 0\}$$ is saturated, that is, if $\sigma$ is a weight and $r \geq 1$ an integer, $${\operatorname{SI}(Q,{\beta})}_\sigma \neq 0 \Longleftrightarrow {\operatorname{SI}(Q,{\beta})}_{r \sigma} \neq 0.$$
We will later use this theorem to prove the saturation of the multiplicity (\[one\]) and give an explicit description of the nonzero generalized Littlewood-Richardson coefficients of this form.
Representing multiplicity (\[one\]) as the dimension of a weight space of semi-invariants {#section-saturation}
=========================================================================================
Saturation theorem
------------------
In this section we will show that the multiplicity (\[one\]) described in the branching rule for the diagonal embedding for ${\text{GL}}(n)$ stated in the introduction arises as the dimension of the weight space of semi-invariants for a certain quiver and dimension vector which we construct. A proof of the saturation of the multiplicity will then follow from Theorem \[saturation\].
Sun quiver
----------
Construct a quiver $Q$ in the following way: for $k \geq 2$, start with a regular $2k$-gon with the vertices labeled $(n,i)$, $1 \leq i \leq 2k$, which we call the central vertices, and an arrow connecting $(n,i)$ with $(n,i+1)$ (we will always consider $(n,2k+1) = (n,1)$), where the arrows alternate in direction. At each central vertex $(n,i)$ attach an equioriented $A_n$ quiver, called a flag and denoted ${\mathscr{F}}(i)$, where each $A_n$ is directed the same way as the arrows at the central vertex $(n,i)$. We will later associate each flag with a weakly decreasing sequence with at most $n$ parts and each central arrow with some other partition. For instance, for $k=3$ the quiver looks like $$\xymatrix{&&&& \ar@{~>}[dl]^{\lambda(2)} \\
&& 3 \ar@{~>}[ul]_{\lambda(3)} & 2 \ar[l]_{\alpha(2)} \ar[dr]^{\alpha(1)} \\
\ar@{~>}[r]_{\lambda(4)} & 4 \ar[ur]^{\alpha(3)} \ar[dr]_{\alpha(4)} &&& 1 \ar@{~>}[r]_{\lambda(1)}& \\
&& 5 \ar@{~>}[dl]_{\lambda(5)} & 6 \ar[l]^{\alpha(5)} \ar[ur]_{\alpha(6)} \\
&&&& \ar@{~>}[ul]_{\lambda(6)}}$$ with $n$ vertices along each flag, denoted here by wavy arrows. Label the $j^{th}$ vertex along the $i^{th}$ flag by $(j,i)$, numbered so that $(n,i)$ denotes each center vertex or simply $i$ when it is understood. For consistency we’ll always have ${\mathscr{F}}(i)$ going into the central vertex if $i$ is even and out if $i$ is odd.
Define the dimension vector $\beta$ as $\beta(j,i) = j$ for each $1\leq i \leq 2k$ and $1 \leq j \leq n$. We will show that the generalized Littlewood-Richardson coefficient in (\[one\]) is the dimension of the weight space of semi-invariants for a certain weight for this quiver and the associated dimension vector $\beta$. We have labeled each flag and central arrow by the sequence we will want to eventually associate to it when we calculate the dimension of this particular weight space of semi-invariants. More specifically, we’ll associate the weakly decreasing sequence $\lambda(i)$ to flag ${\mathscr{F}}(i)$ with central arrows $\alpha(i-1), \, \alpha(i)$ both entering vertex $(n,i)$ when $i$ is odd and leaving when $i$ is even, and $\alpha(j)$ will likewise denote the partition associated to this arrow. Throughout the rest of this paper all results except in section \[section-others\] will be for $k \geq 2$ and the quiver $Q$ with $2k$ flags, which we call the [***sun quiver***]{} or the [***$2k$-sun quiver***]{} when we want to emphasize the number of flags, and $\beta$ is the dimension vector defined by $\beta(j,i) = j$.
\[sun-2\] Let $\sigma \in {\mathbb{Z}}^{Q_0}$ be a weight for the $2k$-sun quiver, $k \geq 2$. If $\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_\sigma \neq 0$, then the weight must satisfy $(-1)^i \sigma(j,i) \geq 0$ for all $1 \leq j \leq n, \, 1 \leq i \leq 2k$. Furthermore, $$\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_\sigma = \sum c^{\phi(1)}_{\alpha(1), \alpha(2)} c^{\phi(2)}_{\alpha(2), \alpha(3)} \, \cdots \, c^{\phi(2k)}_{\alpha(2k), \alpha(1)},$$ where the sum ranges over all partitions $\alpha(1), \ldots, \alpha(2k),$ and $\phi(i) = (n^{(-1)^i\sigma(n,i)}, \ldots, 1^{(-1)^i \sigma(1,i)})'$ for $1 \leq i \leq 2k$.
Define $V_j(i) = {\mathbb{C}}^{\beta(j,i)}$ as the vector space assigned to vertex $(j,i)$. A standard calculation using Cauchy’s rule (see [@Ful97a], page 121) shows that the affine coordinate ring ${\mathbb{C}}[\operatorname{Rep(Q,\beta)}]$ decomposes as a sum of tensor products of irreducible representations of the general linear groups ${\text{GL}}(V_j(i))$ with contributions from each of the flags and central arrows. Specifically, if ${\mathscr{F}}(i)$ is a flag going out of a central vertex, meaning when $i$ is odd, then the $n-1$ arrows of the flag contribute $${\displaystyle}\bigoplus_{\phi^1(i), \ldots, \phi^{n-1}(i)} S^{\phi^1(i)} V_1(i)^* \otimes \bigotimes_{j=2}^{n-1} (S^{\phi^{j-1}(i)} V_j(i) \otimes S^{\phi^j(i)} V_j(i)^*) \otimes S^{\phi^{n-1}(i)} V_n(i)$$ for some partitions $\phi^1(i), \ldots, \phi^{n-1}(i)$. We want to determine when these terms give nonzero semi-invariants of weight $\sigma$. Because the $j^{th}$ term of ${\text{SL}}(\beta)$ acts trivially on each $S^{\phi^m(i)} V_k(i)$ whenever $j \neq k$, the terms of ${\text{SL}}(\beta)$ distribute to the corresponding terms of the sum across the tensor products.
The term $(S^{\phi^1(i)} V_1(i)^*)^{{\text{SL}}(V_1(i))} \neq 0$ if and only if $\phi^1(i)$ is of size $w \times \operatorname{dim}V_1(i) = w \times 1$ for some $w \in {\mathbb{Z}}_{\geq 0}$. Hence, in this case, the space is one-dimensional and is spanned by a semi-invariant of weight $-w$. Therefore, $(S^{\phi^1(i)} V_1(i)^*)^{{\text{SL}}(V_1(i))}$ contains a nonzero semi-invariant of weight $\sigma (1,i)$ if and only if $\sigma(1,i) < 0$ and $\phi^1(i)$ is of size $-\sigma(1,i)\times 1 = (1^{-\sigma(1,i)})'$. We know from this that $(S^{\phi^1(i)} V_1(i)^*)^{{\text{SL}}(V_1(i))}$ is nonzero if and only if it is one-dimensional.
Next, $(S^{\phi^1(i)} V_2 (i) \otimes S^{\phi^2(i)} V_2(i)^*)^{{\text{SL}}(V_2(i))}$ is nonzero if and only if $\phi^2(i)_p - \phi^1(i)_p = k \in {\mathbb{Z}}_{\geq 0}$ for all $p$. That is, $\phi^2(i)$ is $\phi^1(i)$ plus some extra columns, which must be of length $\operatorname{dim}V_2(i) =\beta(i,2) = 2$. In this case, the space being nonzero is equivalent to it being spanned by a semi-invariant of weight equal to the negative of the number of extra columns. Hence, $(S^{\phi^1(i)} V_2 (i) \otimes S^{\phi^2(i)} V_2(i)^*)^{{\text{SL}}(V_2(i))}$ contains a semi-invariant of weight $\sigma(2,i)$ if and only if the space is one-dimensional and $\phi^2(i)=(2^{-\sigma(2,i)}, 1^{-\sigma(1,i)})'.$
Reasoning this way and continuing by sorting the spaces of semi-invariants in ${\operatorname{SI}(Q,{\beta})}$ of weight $\sigma$, we have that $\phi^1(i)$ is of size $-\sigma(1,i) \times 1$ and $\phi^j(i)$ is attained from $\phi^{j-1}(i)$ by adjoining a rectangle of size $-\sigma(j,i) \times j$ to the left of it. Thus, $\phi^{n-1}(i) = ((n-1)^{-\sigma(n-1,i)}, \ldots, 1^{-\sigma(1,i)})'$ and the contribution of the flag ${\mathscr{F}}(i)$ for odd $i$ to ${\operatorname{SI}(Q,{\beta})}_{\sigma}$ is precisely $S^{\phi^{n-1}(i)} V_n(i)$.
Similarly, if ${\mathscr{F}}(i)$ is a flag going into a central vertex, meaning $i$ is even, then $\sigma(j,i) \geq 0$ for all $1 \leq j \leq n-1$ and the contribution of the flag ${\mathscr{F}}(i)$ to $\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_{\sigma}$ is $\operatorname{dim}S^{\phi^{n-1}(i)} V_n(i)^*$ with $$\phi^{n-1}(i) = ((n-1)^{\sigma(n-1,i)}, \ldots, 1^{\sigma(1,i)})'.$$
In addition, the $2k$ central arrows give unspecified partitions $\alpha_i$ with at most $n$ parts each. By taking into account the weights at the central vertices and denoting $V_n(i)$ as simply $V(i)$, we may tensor the contributions from the central vertices to the space of semi-invariants with appropriate powers of the determinant to obtain ${\text{GL}}$-representations, the dimensions of which will be the Littlewood-Richardson coefficients we want. To be precise, we get $$\begin{array}{rcl}
\operatorname{dim}\left((S^{\phi^{n-1}(2i-1)}V(2i-1) \otimes S^{\alpha_{2i-2}} V(2i-1)^* \otimes S^{\alpha_{2i-1}}V(2i-1)^* \otimes {\text{det}}_{V(2i-1)}^{-\sigma(n,2i-1)})^{{\text{GL}}(V(2i-1))}\right) & = & c^{\phi(2i-1)}_{\alpha_{2i-2}, \alpha_{2i-1}} \\\\
\operatorname{dim}\left((S^{\phi^{n-1}(2i)}V(2i)^* \otimes S^{\alpha_{2i}} V(2i) \otimes S^{\alpha_{2i-1}}V(2i) \otimes {\text{det}}_{V(2i)}^{-\sigma(n,2i)})^{{\text{GL}}(V(2i))} \right) &= &c^{\phi(2i)}_{\alpha_{2i-1}, \alpha_{2i}}
\end{array}$$ for each $i=1, \ldots, k$ (recall that we consider the central vertex $(n,2k)$ to be the same as $(n,0)$ and so on). Putting these together, the dimension of ${\operatorname{SI}(Q,{\beta})}_{\sigma}$ is as stated.
For weakly decreasing sequences $\lambda(1), \ldots, \lambda(2k) $ of $n$ integers, define the weight $\sigma_1$ as $$\label{weight-sigma_1}
\sigma_1(j,i) = \begin{cases}
(-1)^i(\lambda(i)_j - \lambda(i)_{j+1}) & 1 \leq i \leq 2k, \, 1 \leq j \leq n-1 \\
(-1)^i \lambda(i)_n & 1 \leq i \leq 2k, \, j = n. \end{cases}$$
The following is immediate from calculating what the $\phi(i)$ are with respect to this weight.
\[si-saturation\] Let $\lambda(1),\ldots, \lambda(2k)$, $k\geq 2$, be weakly decreasing sequences of $n$ integers. Then for every integer $r \geq 1$, we have $$f(r \lambda(1), \ldots, r\lambda(2k)) = \sum c^{r\lambda(1)}_{\alpha(1), \alpha(2)} c^{r\lambda(2)}_{\alpha(2), \alpha(3)} \, \cdots \, c^{r\lambda(2k)}_{\alpha(2k), \alpha(1)} = \operatorname{dim}{\operatorname{SI}(Q,{\beta})}_{r\sigma_1}.$$ In particular, when $k=3$ and $r=1$ the dimension of this weight space of semi-invariants is the multiplicity of the branching rule in (\[one\]).
Because the general case is proven in precisely the same way, assume $r=1$. The proof is then the same as that of Lemma \[sun-2\] because $\phi(i) = \lambda(i)$ with this weight.
While it is clear that $\lambda(1),\ldots, \lambda(2k)$ must be partitions if $f(\lambda(1),\ldots, \lambda(2k))$ is to be nonzero, this is verified from the conditions for $\sigma$ in Lemma \[sun-2\] and the description of $\sigma_1$.
By representing the multiplicity as the dimension of a weight space of semi-invariants as in Lemma \[si-saturation\], the saturation of this multiplicity immediately follows from Theorem \[saturation\].
In this way, we have written the generalized Littlewood-Richardson coefficient $f(\lambda(1), \ldots, \lambda(m))$ as the dimension of a certain weight space of semi-invariants of some quiver. However, we can only express the generalized Littlewood-Richardson coefficient in terms of quiver invariant theory when $m$ is even and at least four. When $m \geq 3$ is odd, this process fails because the first and last flags will be oriented the same direction which would require the central arrow connecting the first and last central vertices to be pointed both directions, an impossibility, while if $m=2$ we would have an oriented cycle.
The facets of the cone of effective weights {#section-facets}
===========================================
Recall that for a quiver $Q$ and dimension vector $\beta$, the set of (integral) effective weights is $$\Sigma(Q,\beta)= \{\sigma \in {\mathbb{Z}}^{Q_0} \mid {\operatorname{SI}(Q,{\beta})}_\sigma \neq 0 \}.$$ If $\sigma \in {\mathbb{R}}^{Q_0}$ is a real-valued function on the set of vertices $Q_0$ and $\alpha$ is an integer-valued function on $Q_0$, define $\sigma(\alpha)$ by $$\sigma(\alpha) = \sum_{x \in Q_0} \sigma(x) \alpha(x).$$
The condition $\sigma(\beta)=0$ is clearly necessary for $\sigma$ to be effective. This is because the action of the one-dimensional torus $\{(t \operatorname{Id}_{\beta(i)})_{i \in Q_0} \mid t \in k \backslash \{0\}\}$ on $\operatorname{Rep(Q,\beta)}$ is trivial, so if $f$ is a nonzero semi-invariant of weight $\sigma$ and $g_t = (t\operatorname{Id}_{\beta(i)})_{i \in Q_0} \in {\text{GL}}(\beta)$, then $$g_t \cdot f = t^{\sigma(\beta)} \cdot f,$$ which implies $\sigma(\beta)=0$. Surprisingly, satisfying a certain set of linear homogenous inequalities is sufficient for a weight to be effective (see Theorems \[spanning\] and \[facets\]).
King [@Kin94] gave the following numerical criterion for $\sigma$-(semi-)stability for finite-dimensional algebras based on the Hilbert-Mumford criterion from GIT. (King’s criterion differs in sign from our convention, which is why the inequalities in the following theorem go the opposite direction as the ones in his original paper.)
Let $Q$ be a quiver, $\beta$ a dimension vector, and $V \in \operatorname{Rep(Q,\beta)}$. Suppose $\sigma \in {\mathbb{Z}}^{Q_0}$ is a weight such that $\sigma(V) = 0$. Then
1. $V$ is $\sigma$-semi-stable if and only if $\sigma({\underline{\textbf{dim}} }\,{V'}) \leq 0$ for every subrepresentation $V'$ of $V$;
2. $V$ is $\sigma$-stable if and only if $\sigma({\underline{\textbf{dim}} }\, {V'}) < 0$ for every proper nontrivial subrepresentation $V'$ of $V$.
We call $\beta$ *$\sigma$-(semi)-stable* if there exists a $\sigma$-(semi-)stable representation in $\operatorname{Rep(Q,\beta)}$.
Because of this description of the $\sigma$-(semi)-stable representations, there is a full subcategory of $\text{Rep}(Q)$ consisting $\sigma$-(semi)-stable representations. This is an abelian category with simple objects being the $\sigma$-stable representations, and moreover because every representation has finite length, the subcategory is Artinian and Noetherian, so any $\sigma$-semi-stable representation has a Jordan-Hölder filtration with $\sigma$-stable factors.
The following result is quite useful for calculations.
*([@DW00a], Corollary 1)* \[reciprocity\] For any dimension vectors $\alpha, \beta$ and quiver $Q$ without oriented cycles, we have $$\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_{\langle \alpha, \cdot, \rangle} = \operatorname{dim}{\operatorname{SI}(Q,{\alpha})}_{-\langle \cdot, \beta \rangle}.$$
Denote this common value of the dimensions of the weight spaces by $\alpha \circ \beta$. By the saturation of effective weights (Theorem \[saturation\]) and the reciprocity property, we have $$\alpha \circ \beta \neq 0 \Longleftrightarrow r\alpha \circ s \beta \neq 0, \; \forall r,s \geq 1.$$
Schofield defined useful semi-invariants which span the weight spaces of semi-invariants, and together with his study of general representations the set $\Sigma(Q,\beta)$ may be described in the following way. We use the notation $\alpha \hookrightarrow \beta$ to mean that every $\beta$-dimensional representation has a subrepresentation of dimension $\alpha$.
\[spanning\]
Let $Q$ be a quiver and $\beta$ a sincere dimension vector, [*i.e.* ]{}, $\beta(x) \neq 0$ for all $x \in Q_0$. If $\sigma = \langle \alpha, \cdot \rangle \in {\mathbb{Z}}^{Q_0}$ is a weight with $\alpha \in {\mathbb{Z}}^{Q_0}$, then the following statements are equivalent:
1. $\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_\sigma \neq 0;$
2. $\sigma(\beta) = 0$ and $\sigma(\beta') \leq 0$ for every $\beta' \hookrightarrow \beta;$
3. $\alpha$ is a dimension vector, $\sigma(\beta) = 0$, and $\alpha \hookrightarrow \alpha + \beta$.
Some of the necessary and sufficient linear homogeneous inequalities above turn out to be redundant. In order to describe a minimal list among these, we need the following result.
\[Schofield-Schur\] Let $\beta \in {\mathbb{Z}}_{\geq 0}^{Q_0}$ be a dimension vector. The following are equivalent:
1. there exists a $\beta$-dimensional representation $V$ such that End$_Q(V,V) \cong {\mathbb{C}};$
2. $\sigma_\beta(\beta') < 0$ for all $\beta' \hookrightarrow \beta$, $\beta' \neq 0,\, \beta$, where $\sigma_\beta = \langle \beta, \cdot \rangle - \langle \cdot, \beta \rangle.$
Any such $\beta$ is called a *Schur root* and a representation with these properties is a *Schur representation*.
With this theorem, we can describe the effective weights by semi-stability conditions. In particular, we will be able to associate the dimension vectors of the subrepresentations with subsets of $\{1, \ldots, n\}$ because of the following theorem.
\[facet-decomp\] Let $Q$ be a quiver without oriented cycles and $N$ vertices, and $\beta$ a Schur root. Then
1. $\operatorname{dim}C(Q, \beta) = N-1$, and
2. $\sigma \in C(Q,\beta)$ if and only if $\sigma(\beta) = 0$ and $\sigma(\beta_1) \leq 0$ for every decomposition $\beta = c_1 \beta_1 + c_2 \beta_2$ with $\beta_1, \beta_2$ Schur roots, $\beta_1 \circ \beta_2 = 1$, and $c_i = 1$ whenever $\langle \beta_i, \beta_i \rangle < 0$.
This then allows us to describe the cone of effective weights as $${C(Q,\beta)}= \{\sigma \in {\mathbb{H}}(\beta) \mid \sigma(\beta') \leq 0 \text{ for all } \beta' \hookrightarrow \beta\},$$ where ${\mathbb{H}}(\beta) = \{ \sigma \in {\mathbb{R}}^{Q_0} \mid \sigma(\beta) = 0\}.$
\[minimal-list\] While we could replace $\beta_1 \circ \beta_2 =1$ with $\beta_1 \circ \beta_2 \neq 0$ in Theorem , this would give a longer list of inequalities. The condition $\beta_1 \circ \beta_2 =1$ ensures a complete and irredundant list of necessary and sufficient inequalities.
A more precise description of the facets of the cone $C(Q,\beta)$ is given by the following.
For a dimension vector $\beta$, define $W_2(Q,\beta)$ as the set of all ordered pairs $(\beta_1, \beta_2)$ such that:
1. $\beta = c_1 \beta_1 + c_2 \beta_2$ for some integers $c_1, c_2 \geq 1$;
2. $\beta_1$ and $\beta_2$ are Schur roots;
3. $s_1 \beta_1 \circ s_2 \beta_2 =1$ for all $s_1, s_2 \geq 1$;
4. $c_i=1$ whenever $\langle \beta_i, \beta_i \rangle < 0$.
\[facets\] Let $Q$ be a quiver without oriented cycles and $\beta$ a Schur root. Let $\mathcal{F}(Q,\beta)$ denote the set of all facets of $C(Q,\beta)$. Then the map $$W_2(Q,\beta) \to \mathcal{F}(Q,\beta)$$ defined by $$(\beta_1, \beta_2) \in W_2(Q,\beta) \mapsto C(Q,\beta_1) \bigcap C(Q,\beta_2) = \mathbb{H}(\beta_1) \bigcap C(Q,\beta)$$ is a bijection. Thus, a minimal list of linear homogeneous inequalities defining the cone $C(Q,\beta)$ is obtained by $\sigma(\beta) = 0$ and $\sigma(\beta_1) \leq 0$ for all $(\beta_1, \beta_2) \in W_2(Q,\beta)$.
The facets of the cone of effective weights for the sun quiver {#section-facets-quivers}
==============================================================
In order to use the results in the previous section to describe the facets of $C(Q,\beta)$, we’ll first show that the dimension vector $\beta$ is Schur and determine conditions on the $\beta_1'$s that can appear in the decompositions.
\[Schur-root\] The dimension vector $\beta$ for the sun quiver is Schur.
The dimension vector $\beta$ is indivisible, meaning the greatest common divisor of its coordinates is one. By a result of Kac ([@Kac82], Theorem B(d)), to show $\beta$ is Schur, it suffices to show that $\beta$ is in the fundamental region of the graph, meaning that the support of $\beta$ is a connected graph and $\tau_i(\beta) \leq 0$ for all vertices $i \in Q_0$, where $e_i$ denotes the dimension vector of the simple representation at vertex $i$ and $\tau_i(\cdot) := \langle e_i, \cdot \rangle + \langle \cdot, e_i \rangle$. This is immediately checked to hold for all $n \geq 1$.
\[dim-cones\] For the sun quiver $Q$, $\operatorname{dim}C(Q,\beta) = mn-1$.
This immediately follows from Lemma \[Schur-root\] and Theorem \[facet-decomp\].
Now consider the following dimension vectors $\beta_1$ for the sun quiver, where $e_{(j,i)}$ denotes the dimension vector of the simple representation at vertex $(j,i)$:
1. $\beta_1 = e_{(j,i)}$ for a flag $i$ going out of the central vertex, or $\beta_1 = \beta - e_{(j,i)}$ for a flag $i$ going into a central vertex;
2. $\beta_1 \neq \beta$, $\beta_1 \circ (\beta - \beta_1) = 1$, and $\beta_1$ is weakly increasing with jumps of at most one along each of the $m$ flags.
Denote the set of such $\beta_1$ by $\mathcal{D}$. We show in the next lemma that each $\beta_1 \in \mathcal{D}$ defines a facet of ${C(Q,\beta)}$, which is called a *regular facet* if $\beta_1$ is in the form described in (2), while a facet defined by some $\beta_1$ as described in (1) is called *trivial*. The interpretation of the inequalities arising from the $\beta_1 \in \mathcal{D}$ is given in Remark \[facets-interpretation\].
\[facets-2\] For the sun quiver $Q$, the regular facets of $C(Q,\beta)$ are of the form $$\mathbb{H}(\beta_1) \bigcap C(Q,\beta),$$ where $\beta_1$ is weakly increasing with jumps of at most one along the flags, $\beta_1 \neq \beta$, and $\beta_1 \circ (\beta - \beta_1) = 1$.
The proof is the same as the one given in [@Chi08], Lemma 5.2, which we include here for completeness. By Theorem \[facets\], a facet $\mathcal{F}$ is of the form $$\mathbb{H}(\beta_1) \bigcap C(Q,\beta),$$ where $\beta_1, \beta_2$ are Schur roots, $\beta_1 \circ \beta_2 = 1$, and $\beta = c_1\beta_1 + c_2\beta_2$ for some $c_1, c_2 \geq 1$.
Suppose that $\beta_1$ is not of the form as in (1). We’ll show that $\beta_1$ is weakly increasing with jumps of at most one along the flags. Denote $c_1\beta_1 = \beta_1'$ and $c_2\beta_2 = \beta_2'$. Because it is clear that $s_1 \beta_1 \circ s_2 \beta_2 \geq \beta_1 \circ \beta_2$ for all $s_1,s_2 \geq 1$, $\beta_1' \circ \beta_2' \neq 0$. By Theorem \[spanning\], it follows that any representation of dimension vector $\beta$ has a subrepresentation of dimension vector $\beta_1'$. Choose a $\beta-$dimensional representation which is injective along the flags going into a central vertex and surjective along the flags going out of the central vertex. Then $\beta_1'$ is weakly increasing along the flags going in and has jumps of at most one (from the end of the flag towards the center vertex) along the flags going out, or else the maps couldn’t be surjective.
We’ll show that $\beta_1'$ is weakly increasing along each flag $\mathcal{F}(i)$ going out of a central vertex. Suppose to the contrary that $\beta_1'(l+1) - \beta_1'(l) < 0$ for some $l \in \{1, \ldots, n-1\}$. Then $\beta_1' - e_l \hookrightarrow \beta'_1$. Moreover, $\beta_1' \circ \beta_2' \neq 0$ is equivalent to $\beta_1'$ being $-\langle \cdot, \beta_2' \rangle-$semi-stable by reciprocity (Theorem \[reciprocity\]). Thus, $\langle \beta_1' - e_l, \beta_2' \rangle \geq 0$, so $\beta'_2(l) \leq \beta_2'(l-1)$, implying $\beta_1'(l) \geq 1 + \beta_1'(l-1)$. As we previously showed that $\beta_1'$ has jumps of at most one along such a flag, we must have $\beta_1'(l) = 1 + \beta_1'(l-1)$. Thus, $c_1 = 1$ and $e_l \hookrightarrow \beta_1'$. We then have that $\beta_1' = \beta_1$ is a Schur root by assumption, hence is $\sigma_{\beta_1'}$-semistable by Theorem \[Schofield-Schur\], and $e_l, \, \beta_1' - e_l \hookrightarrow \beta_1'$, with $\beta_1' \neq e_l$. Therefore, by the same theorem, $\sigma_{\beta'_1}(e_l)< 0$ and $\sigma_{\beta'_1}(\beta_1'-e_l)< 0,$ which is a contradiction. Thus, $\beta_1'$ must be weakly increasing along the flags going out of a central vertex. By a similar argument, $\beta_1'$ will have jumps of at most one along each flag going in.
Finally, we’ll show that $c_1=c_2=1$. Because $\beta_1' = c_1\beta_1$ has jumps of at most one along each flag, we have $0 \leq c_1(\beta_1(l+1,i) - \beta_1(l,i)) \leq 1$ for all $l \in \{1, \ldots, n-1\}$ and $i \in \{1, \ldots, m\}.$ If there are no $l,i$ such that $\beta_1(l+1,i) - \beta_1(l,i) \neq 0$, then $c_1 = 1$, while otherwise there must exist an $i$ such that $\beta_1'(l,i) =1,$ so $c_1=1$. Similarly, $c_2=1$.
Thus, $\beta = \beta_1 + \beta_2$ with $\beta_1$ weakly increasing of jumps of at most one along the flags.
\[inequalities\] Let $\sigma \in {\mathbb{H}}(\beta)$ for the sun quiver. Then $\sigma \in {C(Q,\beta)}$ if and only if the following are true:
1. $(-1)^i \sigma(e_{(j,i)}) \geq 0$ for all $1 \leq j \leq n-1, \, 1 \leq i \leq m$;
2. $\sigma(\beta_1) \leq 0$ for every $\beta_1 \neq \beta$ weakly increasing with jumps of at most one along the flags and $\beta_1 \circ (\beta - \beta_1) = 1$.
The description of the regular facets in Lemma \[facets-2\] proves one direction, while if $\sigma \in {C(Q,\beta)}$, then $\sigma(\beta_1) \leq 0$ for every $\beta_1 \in \mathcal{D}$ by Theorem \[spanning\], which is equivalent to (1) and (2).
\[facets-interpretation\] Let $\sigma_1$ be the weight we defined for the sun quiver in equation (\[weight-sigma\_1\]). In particular, $$\sigma_1(e_{(j,i)}) = (-1)^i(\lambda(i)_j - \lambda(i)_{j+1}), \quad \, \, 1 \leq i \leq m, \, 1 \leq j \leq n-1.$$ The inequalities arising from a trivial facet of $C(Q,\beta)$, as described in (1) in the preceding lemma, called the *chamber inequalities*, simply state that the sequences $\lambda(i)$ are weakly decreasing sequences of real numbers. Because we will always assume this, we exclude these $\beta_1$ from our considerations. The inequalities arising from (2) in the lemma are called the *regular inequalities*, and the corresponding facet is regular.
Horn-type inequalities {#section-Horn}
======================
In this section let $\beta_1$ be a dimension vector which is weakly increasing with jumps of at most one along each of the flags towards the central vertices. Define the following jump sets: $$I_i = \{l \mid \beta_1(l,i) > \beta_1(l-1,i), \, 1 \leq l \leq n\}$$ with the convention $\beta_1(0,i) =0$ for all $i$. Because $\beta_1$ defines a tuple $I=(I_1, \ldots, I_m)$, we’ll commonly denote $\beta_1$ by $\beta_I$. Note that $|I_i| = \beta_I(n,i)$ for each $i$.
Conversely, each tuple $I = (I_1, \ldots, I_m)$ of subsets of $\{1, \ldots, n\}$ defines a dimension vector $\beta_I$ because if $$I_i = \{z_1(i) < \cdots < z_r(i)\},$$ then $\beta_I(j,i) = j-1$ for all $z_{k-1}(i) \leq j < z_k(i)$ for all $1 \leq k \leq r+1$, with the convention that $z_0(i) = 0$ and $z_{r+1}(i) = n+1$ for all $i$. This means that going towards the center vertex on the $i^{th}$ flag, the dimension at a vertex is 0 until the vertex $z_1(i)$, at which it becomes 1 and continues to be 1 until the vertex $z_2(i)$, at which point it becomes $2$, and so on.
For the dimension vector $\beta$ associated to the sun quiver, define $T(n,m)$ to be the set of all tuples $I=(I_1, \ldots, I_m)$ such that $\beta_I \neq \beta$ (equivalently, $|I_i|<n$ for some $i$), and $\beta_I \circ (\beta - \beta_I) = 1$.
We give a description of the set $T(n,m)$ without reference to the sun quiver and only in terms of partitions in Lemma \[sun-Horn\].
\[Horn-type-inequalities\] Suppose $\lambda(1),\ldots, \lambda(m)$ are weakly decreasing sequences of $n$ real numbers. For the sun quiver $Q$ and dimension vector $\beta$, $\sigma_1 \in C(Q,\beta)$ if and only if $\sum_{i \text{ even}} |\lambda(i)| = \sum_{i \text{ odd}} |\lambda(i)| $ and $$\sum_{j \in I_i} \sum_{i \text{ even}} \lambda(i)_j \leq \sum_{j \in I_i} \sum_{i \text{ odd}} \lambda(i)_j$$ for every tuple $(I_1, \ldots, I_m) \in T(n,m)$.
We prove the statement by using the definition of the set $T(n,m)$, Theorem \[facets\], and Lemma \[inequalities\]. Calculating directly from the definition of the weight $\sigma_1$ in equation (\[weight-sigma\_1\]) of Section \[section-saturation\], $$\sigma_1(\beta) = \sum_{i \text{ odd}} \sum_{j=1}^{n-1}[(\lambda(i)_{j+1} - \lambda(i)_j) \beta(j,i)]
+ \sum_{i \text{ even}} \sum_{j=1}^{n-1}[(\lambda(i)_j - \lambda(i)_{j+1}) \beta(j,i)]
+ \sum_{i=1}^m (-1)^i \lambda(i)_n \beta(n,i)$$ Substituting $\beta(j,i)=j$, $\sigma_1(\beta) = 0$ precisely when the equality holds. Replacing $\beta$ with $\beta_I$, we get a similar expression, and after noting that $\lambda(i)_j$ contributes to the inequality exactly when $\beta_I(j,i) \neq \beta_I(j-1,i)$, meaning $j \in I_i$, we obtain the inequality. Because the $\beta_I$ are precisely those described in Lemma \[inequalities\], this proves the equivalence.
We now want to better understand the set $T(n,m)$. Specifically, we would like to describe the set without any intrinsic reference to $\beta_I$ but rather in terms of partitions. For any tuple $(I_1, \ldots, I_m)$ of subsets of $\{1,\ldots,n\}$, define the following decreasing sequences of integers, where we identify $I_0$ and $I_m$: $${\underline{\lambda}}(I_i) = \begin{cases}
\lambda'(I_i) & i \text{ even} \\
\lambda'(I_i) -((|I_i| - |I_{i-1}| - |I_{i+1}|)^{n-|I_i|}) & i \text{ odd}.
\end{cases}$$
\[sun-Horn\] The set $T(n,m)$ for the sun quiver consists of all tuples $I = (I_1, \ldots, I_m)$ such that:
1. at least one of the subsets $I_1, \ldots, I_m$ has cardinality $< n$;
2. ${\underline{\lambda}}(I_i)$ is a partition for all $1 \leq i \leq m$;
3. $f({\underline{\lambda}}(I_1), \ldots, {\underline{\lambda}}(I_m)) = 1$.
Denote the weight $\langle \beta_1, \cdot \rangle$ by $\sigma_I$. Describing $\beta_1$ by $\beta_I$ as previously and letting $e(j,i)$ be the dimension vector of the simple representation with support at vertex $j$ on the $i^{th}$ flag, the contribution to $\sigma_I(\alpha)$ at a vertex $l \in \{1, \ldots, n-1\}$ on a flag $i$ is $$\langle \beta_I, e(l,i) \rangle = \beta_I(l) - \beta_I(l-1)$$ if $i$ is even with $\beta_I(0) = 0$, and $$\langle \beta_I, e(l,i) \rangle = \beta_I(l) - \beta_I(l+1)$$ if $i$ is odd. Since $\beta_I$ is weakly increasing with jumps of at most one along the flags, this translates to $$\label{eq:weight-even}
\sigma_I(l,i) = \left\{\begin{array}{lr}
1 & \text{ if }l \in I_i \\
0 & \text{otherwise}
\end{array} \right.$$ for $l \in \{1, \ldots, n-1\}$ and $i$ even, and $$\label{eq:weight-odd}
\sigma_I(l,i) = \left\{ \begin{array}{lr}
-1 & \text{ if } l+1 \in I_i \\
0 & \text{otherwise}
\end{array} \right.$$ for $l \in \{1, \ldots, n-1\}$ and $i$ odd. We then only need to describe the weight at the central vertices. We have $$\sigma_1(n,i) = \begin{cases}
\beta_I(n,i) - \beta_I(n-1,i) & i \text{ even} \\
\beta_I(n,i) - \beta_I(n,i-1) - \beta_I(n,i+1) & i \text{ odd},
\end{cases}$$ where $\beta_I(n,0) = \beta_I(n,m)$. We use the fact that $\beta_I(n,i) = |I_i|$ to finish defining the weight $\sigma_I$ in terms of the subsets $I_i$, where we identify $I_0$ as $I_m$ below:
$$\sigma_I(n,i) = \left\{\begin{array}{lr}
0 & i \text{ even}, \; n \not \in I_i, \\
1 & i \text{ even}, \; n \in I_i,\\
|I_i| - |I_{i-1}| - |I_{i+1}| & i \text{ odd}.
\end{array}
\right.$$
Using this explicit description of this weight $\sigma_I$, we can calculate the corresponding partitions in determining $\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_{\sigma_I}$. The calculation is done precisely the same as before (see Lemma \[sun-2\] for the details). The partitions we want to consider for describing the contribution of the first $n-1$ vertices along each flag are $$\gamma_{n-1}(i) = (\beta_2(n-1,i)^{(-1)^i\sigma_I(n-1,i)}, \ldots, \beta_2(1,i)^{(-1)^i \sigma_I(1,i)})', \qquad 1 \leq i \leq m.$$ Recalling the contributions of the central vertices to the space of semi-invariants, we have that $$\gamma(i) = \gamma_{n-1}(i) + (((-1)^i \sigma_I(n,i))^{\beta_2(n,i)}),$$ are the partitions we want for $1 \leq i \leq m$, and $$\operatorname{dim}\operatorname{SI}(Q,\beta_2)_{\sigma_1} = f(\gamma(1),\ldots,\gamma(m)).$$
Since $\beta_2 = \beta - \beta_1$, if $I_i = \{z_1(i) < \ldots < z_r(i)\}$, then $$\beta_2(z_j(i),i) = z_j(i) - j = z_j(i) - 1 - (j-1) = \beta_2(z_j(i)-1,i),$$ and in particular, $\beta_2(n,i) = n - \beta_1(n,i) = n - |I_i|$. With this, ${\underline{\lambda}}(I_i) = \gamma(i)$ for each $i$.
Thus, if $I = (I_1, \ldots, I_m) \in T(n,m)$, then condition (a) is satisfied by definition, (b) is satisfied since $\gamma(i)$ is a partition for each $i$ and $\gamma(i) = {\underline{\lambda}}(I_i)$, and (c) is true because $$1 = \beta_I \circ (\beta - \beta_I) = f(\gamma(1), \ldots, \gamma(m)) = f({\underline{\lambda}}(I_1), \ldots, {\underline{\lambda}}(I_m)).$$ Conversely, if conditions (a)$ - $(c) are satisfied by a tuple $I= (I_1, \ldots, I_m)$ of subsets of $\{1,\ldots, n\}$, then we can construct a dimension vector $\beta_I$ in the usual way and the associated decreasing sequences of integers ${\underline{\lambda}}(I_i), \, 1 \leq i \leq m$. Necessarily, $\beta_I \neq \beta$ and $$\beta_I \circ (\beta - \beta_I) = f({\underline{\lambda}}(I_1), \ldots, {\underline{\lambda}}(I_m)) = 1$$ Thus, $I \in T(n,m)$.
\[conditions-weight-cone\] Let $\lambda(1), \ldots, \lambda(m)$ be weakly decreasing sequences of $n$ reals, $m \geq 4$ and even. The following are equivalent for the sun quiver $Q$:
1. $\sigma \in {C(Q,\beta)}$;
2. the numbers $\lambda(i)_j$ satisfy $$\sum_{i \text{ even}} |\lambda(i)| = \sum_{i \text{ odd}} |\lambda(i)|$$ and $$\sum_{j \in I_i} \sum_{i \text{ even}} \lambda(i)_j \leq \sum_{j \in I_i} \sum_{i \text{ odd}} \lambda(i)_j$$ for every tuple $(I_1, \ldots, I_m)$ for which $|I_i| < n$ for some $i$, the ${\underline{\lambda}}(I_i)$ are partitions, $1 \leq i \leq m$, and $$f({\underline{\lambda}}(I_1), \ldots, {\underline{\lambda}}(I_m)) \neq 0;$$
3. the numbers $\lambda(i)_j$ satisfy the same conditions as in part (2), and it suffices to check only the tuples $(I_1,\ldots, I_m)$ which satisfy those conditions and $$f({\underline{\lambda}}(I_1), \ldots, {\underline{\lambda}}(I_m)) = 1.$$
This follows from Proposition \[Horn-type-inequalities\], Lemma \[sun-Horn\], and Remark \[minimal-list\].
We can deduce minor conditions on the sizes of the $I_i$.
Let $I = (I_1,\ldots, I_m)$ be a tuple of subsets of $\{1, \ldots, n\}$ and define $s_i$ to be the smallest $k \in \{0, \ldots, |I_i|\}$ such that $n - k \notin I_i$. Then $$\max \{|I_{i-1}|, |I_{i+1}| \} \leq |I_i| \leq |I_{i-1}| + |I_{i+1}| + s_i$$ if $I \in T(n,m)$ and $i$ is odd.
If $\beta_1 \circ \beta_2 \neq 0$, then any representation $V$ of dimension $\beta_1 + \beta_2$ has a subrepresentation of dimension $\beta_1$. Choosing $V$ such that $V(a)$ is invertible for every arrow $a$ between central vertices, we immediately have $$\text{max} \{|I_{i-1}|, |I_{i+1}| \} \leq |I_i|$$ for the stated $i$. Let $\langle \beta_1, \cdot \rangle = \sigma_1$. In order for $\operatorname{dim}\operatorname{SI}(Q,\beta_2)_{\sigma_1}$ to be nonzero, each ${\underline{\lambda}}(I_i)$ must be a partition, meaning, in particular, that it has non-negative parts. Note that $s_i$ is precisely the smallest part of $\lambda'(I_i)$ for each $i$. We have ${\underline{\lambda}}(I_i) = \lambda'(I_i) - (\sigma_1(i)^{n-|I_i|})$ for each of the specified $i$, as well as $\sigma_1(i) = |I_i| - |I_{i-1}| - |I_{i+1}|$. Thus, ${\underline{\lambda}}(I_i)$ is a partition if and only if $$0 \leq |I_{i-1}| + |I_{i+1}| - |I_i| + s_i.$$
We note that the proof above does not extend to determining how the sizes of the other subsets compare or whether some subset contains $n$ or is nonempty. This is because no central vertex in the sun quiver shares two flags, so we can’t cancel the sizes of any two consecutive subsets that appear in $\sigma_1$.
\[n=2\] For the case $n=2$ and $m=6$ for the sun quiver, we can compute all decompositions $\beta = \beta_1 + \beta_2$ such that both $\beta_1, \beta_2$ are Schur roots and $\beta_1 \circ \beta_2 = 1$, or equivalently by Theorem \[facets\], a description of all facets of ${C(Q,\beta)}$ in this case. This is done by first computing all tuples $({\underline{\lambda}}(I_1),\ldots, {\underline{\lambda}}(I_6))$ such that $f({\underline{\lambda}}(I_1),\ldots, {\underline{\lambda}}(I_6))=1$ and such that the other conditions of Proposition \[conditions-weight-cone\] are satisfied. Next, we compute all tuples $(I_1,\ldots, I_6)$ arising from these and omit any such that the corresponding $\beta_1$ and $\beta_2$ are not Schur as these will produce redundant inequalities. The inequalities are immediately determined from these tuples $(I_1,\ldots, I_6)$. We list the corresponding dimension vectors $\beta_1$ in the .
The corresponding inequalities of the partitions arising from these decompositions then provide a complete and minimal list of linear homogeneous inequalities defining when a weight $\sigma$ is in ${C(Q,\beta)}$ for the sun quiver $Q$ in this case. Specifically, $\sigma \in {C(Q,\beta)}$ if and only if the defining partitions satisfy $$|\lambda(1)| + |\lambda(3)| + |\lambda(5)| = |\lambda(2)| + |\lambda(4)| + |\lambda(6)|,$$ and $$\begin{array}{ll}
\lambda(2)_1 \leq \operatorname{\lambda}(1)_1 + \operatorname{\lambda}(3)_1 &
\operatorname{\lambda}(2)_2 \leq \operatorname{\lambda}(1)_1 + \operatorname{\lambda}(3)_2\\\\
\operatorname{\lambda}(2)_1 + \operatorname{\lambda}(4)_2 \leq \operatorname{\lambda}(1)_1 + \operatorname{\lambda}(3)_1 + \operatorname{\lambda}(5)_1 &
\operatorname{\lambda}(2)_2 + \operatorname{\lambda}(4)_2 \leq \operatorname{\lambda}(1)_1 + \operatorname{\lambda}(3)_2 + \operatorname{\lambda}(5)_1\\\\
\operatorname{\lambda}(2)_2 + \operatorname{\lambda}(6)_2 \leq \operatorname{\lambda}(1)_1 + \operatorname{\lambda}(3)_2 + \operatorname{\lambda}(5)_1 &
|\operatorname{\lambda}(2)| \leq |\operatorname{\lambda}(1)| + |\operatorname{\lambda}(3)| \\\\
\operatorname{\lambda}(2)_1 + \operatorname{\lambda}(4)_1 \leq \operatorname{\lambda}(1)_1 + |\operatorname{\lambda}(3)| + \operatorname{\lambda}(5)_1 &
\operatorname{\lambda}(2)_1 + \operatorname{\lambda}(4)_2 \leq \operatorname{\lambda}(1)_1 + |\operatorname{\lambda}(3)| + \operatorname{\lambda}(5)_2 \\\\
\operatorname{\lambda}(2)_2 + \operatorname{\lambda}(4)_2 \leq \operatorname{\lambda}(1)_2 + |\operatorname{\lambda}(3)| + \operatorname{\lambda}(5)_2 &
\operatorname{\lambda}(2)_1 + \operatorname{\lambda}(4)_2 + \operatorname{\lambda}(6)_2 \leq \operatorname{\lambda}(1)_1 + |\operatorname{\lambda}(3)| + \operatorname{\lambda}(5)_1 \\\\
\operatorname{\lambda}(2)_1 + \operatorname{\lambda}(4)_2 + \operatorname{\lambda}(6)_2 \leq \operatorname{\lambda}(1)_1 + \operatorname{\lambda}(3)_1 +|\operatorname{\lambda}(5)| &
\operatorname{\lambda}(2)_2 + \operatorname{\lambda}(4)_2 + \operatorname{\lambda}(6)_2 \leq \operatorname{\lambda}(1)_1 + |\operatorname{\lambda}(3)| + \operatorname{\lambda}(5)_2 \\\\
|\operatorname{\lambda}(2)| + \operatorname{\lambda}(4)_1 + \operatorname{\lambda}(6)_2 \leq |\operatorname{\lambda}(1)| + |\operatorname{\lambda}(3)| + \operatorname{\lambda}(5)_1 &
|\operatorname{\lambda}(2)|+ \operatorname{\lambda}(4)_2 + \operatorname{\lambda}(6)_2 \leq |\operatorname{\lambda}(1)| + |\operatorname{\lambda}(3)| + \operatorname{\lambda}(5)_2 ,
\end{array}$$ along with the inequalities obtained by permutations of the indices that respect the symmetries of the sun quiver. This likewise provides a description of all the $(I_1, \ldots, I_6)$ in $T(2,6)$, as described in Proposition \[Horn-type-inequalities\].
Generalized eigenvalue problem {#Hermitian}
==============================
The original motivation in [@DW00a] for describing Littlewood-Richardson coefficients in terms of quivers was to provide a solution to a famous conjecture of Horn [@Hor62].
\[Horn\] Let $\lambda(i) = (\lambda_1(i),\ldots, \lambda_n(i))$, $i \in \{1,2,3\}$, be weakly decreasing sequences of $n$ real numbers. Then the following are equivalent:
1. there exist $n \times n$ complex Hermitian matrices $H(i)$ with eigenvalues $\lambda(i)$ such that $$H(2) = H(1) + H(3);$$
2. the numbers $\lambda_j(i)$ satisfy $$|\lambda(2)| = |\lambda(1)| + |\lambda(3)|$$ together with $$\sum_{j \in I_2} \lambda_j(2) \leq \sum_{j \in I_1} \lambda_j(1) + \sum_{j \in I_3} \lambda_j(3)$$ for every triple $(I_1,I_2,I_3)$ of subsets of $\{1,\ldots, n\}$ of the same cardinality $r < n$ and $c^{\lambda(I_2)}_{\lambda(I_1),\lambda(I_3)} \neq 0;$
3. if $\lambda_j(i)$ is an integer for each $1 \leq j \leq n, \, i \in \{1,2,3\}$, (1) and (2) are equivalent to $c^{\lambda(2)}_{\lambda(1),\lambda(3)} \neq 0$.
Klein [@Kle68] noted the equivalence of (3) with short exact sequences of finite abelian $p$-groups, while Klyachko [@Kly98] proved the equivalence of (1) and (2) in Horn’s conjecture. In the same paper, Klyachko noted the connection between this problem and Littlewood-Richardson coefficients. A key step in the proof is the saturation of Littlewood-Richardson coefficients (Theorem \[LR-saturation\]), which was first proved combinatorially by Knutson and Tao [@KT99] and later by Derksen and Weyman [@DW00a] in the context of quivers by using the saturation of weight spaces of semi-invariants.
Klyachko found that the set of solutions in part (1) in the statement of Horn’s conjecture forms a rational convex polyhedral cone $K(n,3)$ in ${\mathbb{R}}^{3n}$, now known as *Klyachko’s cone*. In [@KTW04], Knutson, Tao, and Woodward used honeycombs to describe the facets of Klyachko’s cone. They found that $K(n,3)$ consists of triples $(\lambda(1), \lambda(2), \lambda(3))$ satisfying the conditions in part (2), and restricting to the triples such that $c_{\lambda(I_1), \, \lambda(I_3)}^{\lambda(I_2)} =1$ provides a minimal list.
As we’ve previously proven corresponding statements for parts (2) and (3) for the generalized Littlewood-Richardson coefficient $f$, we now want to find the Weyl-type eigenvalue problem for the non-vanishing of this multiplicity.
Necessary lemmas
----------------
Before we state the generalized eigenvalue problem for the sun quiver, we state two results from linear algebra that will be fundamental for us.
\[matrix-equations\] Let $Q$ be a quiver without oriented cycles, $\beta$ a dimension vector, and $\sigma \in {\mathbb{R}}^{Q_0}$. The following are equivalent:
1. $\sigma \in {C(Q,\beta)}$;
2. there exists $W = \{W(a)\}_{a \in Q_1} \in \operatorname{Rep(Q,\beta)}$ satisfying $$\sum_{\substack{a \in Q_1 \\ ta = x}} W(a)^*W(a) - \sum_{\substack{a \in Q_1 \\ ha = x}} W(a)W(a)^* = \sigma(x) \operatorname{Id}_{\beta(x)}$$ for all $x \in Q_0$, where $W(a)^*$ is the adjoint of $W(a)$ with respect to the standard Hermitian inner product on ${\mathbb{C}}^n$.
\[Hermitians\] Let $\sigma(1), \ldots, \sigma(n-1)$ be non-positive real numbers. The following are equivalent:
1. there exist $W_i \in \operatorname{Mat}_{i \times (i+1)}({\mathbb{C}}), \, 1 \leq i \leq n-1,$ such that $$\begin{array}{rcl}
W_i W_i^* - W_{i-1}^* W_{i-1} & =& - \sigma(i) \operatorname{Id}_{i} \quad \text{ for } 2 \leq i \leq n-1, \\
W_1 W_1^* & = & -\sigma(1);
\end{array}$$
2. the matrix $H=W_{n-1}^*W_{n-1}$ is Hermitian and has eigenvalues $$v(k) = - \sum_{j=k}^{n-1} \sigma(j), \quad \forall \, 1 \leq k \leq n-1 \quad \text{and} \quad v(n) = 0.$$
See Section 3.4 of [@CG02].
This result will allow us to build up an $n \times n$ Hermitian matrix along each flag, though the result as stated only applies to flags that are going out from a central vertex. For the flags going into a central vertex, we need to use the dual form of the above lemma. Namely, $\sigma(i)$ is a non-negative number for each $1 \leq i \leq n-1$, so replace each $-\sigma(i)$ above with $\sigma(i)$, and switch the order of multiplication of $W_i$ and $W_i^*$ in each case for $W_i \in \operatorname{Mat}_{(i+1)\times i}({\mathbb{C}})$.
Generalized eigenvalue problem for $f$
--------------------------------------
Recall the construction of the $2k$-sun quiver in Section \[section-saturation\]. The weight for this quiver is
$$\sigma_1(j,i) = \begin{cases}
(-1)^i(\lambda(i)_j - \lambda(i)_{j+1}) & 1 \leq i \leq 2k, \, 1 \leq j \leq n-1 \\
(-1)^i \lambda(i)_n & 1 \leq i \leq 2k, \, j = n. \end{cases}$$
From Proposition \[matrix-equations\], $\sigma_1 \in {C(Q,\beta)}$ if and only if there is a representation $W \in \operatorname{Rep(Q,\beta)}$ satisfying the specified matrix equations. These equations are essentially the same as those in Lemma \[Hermitians\] for a flag ${\mathscr{F}}(i)$ going out of a central vertex, meaning when $i$ is odd, or in the dual statement when ${\mathscr{F}}(i)$ is going into a central vertex, meaning when $i$ is even. In either case, the first $n-1$ vertices provide $n \times n$ Hermitian matrices $H(i)$ with eigenvalues $$(\lambda(i)_1 - \lambda(i)_n, \ldots, \lambda(i)_{n-1} - \lambda(i)_n, 0)$$ for each $1 \leq i \leq m$.
We now consider the equations arising from the central vertices. Denote the $(n-1)^{th}$ arrow along the $i^{th}$ flag as simply $b^i$ and denote the arrows between the central vertices by the usual partition labeling. The equations arising from the central vertices are, for $1 \leq i \leq k$, $$\begin{array}{rcl}
W(b^{2i-1})^*W(b^{2i-1}) - W(\alpha(2i-1))W(\alpha(2i-1))^* - W(\alpha(2i-2))W(\alpha(2i-2))^* &=& -\lambda(2i-1)_n\operatorname{Id}_n, \\\\
W(\alpha(2i-1))^*W(\alpha(2i-1)) + W(\alpha(2i))^*W(\alpha(2i)) - W(b^{2i})W(b^{2i})^* &=& \lambda(2i)_n\operatorname{Id}_n,
\end{array}$$ where $\alpha(0) = \alpha(m)$. We may rewrite these equations by making a few simple observations. Lemma \[Hermitians\] gives the Hermitian matrices as $H(i) = W(b^i)^*W(b^i)$, or $W(b^i)W(b^i)^*$ depending on the direction of the flag. Clearly, since each $H(i)$ is Hermitian with spectrum $(\lambda(i)_1 - \lambda(i)_n, \ldots, \lambda(i)_{n-1} - \lambda(i)_n, 0)$, $H(i) + \lambda(i)_n \operatorname{Id}_n$ is Hermitian with spectrum $\lambda(i)$; denote this new Hermitian matrix again by $H(i)$. We may conjugate the equations by unitary matrices, if necessary. Moreover, for any $n \times n$ matrix $A$, both $AA^*$ and $A^*A$ are positive semi-definite and have the same spectra, and any positive semi-definite Hermitian matrix $B$ can be written as $WW^*$ or $W^*W$, so we may simplify the forms of the equations. We conclude that $\sigma_1 \in {C(Q,\beta)}$ if and only if there are Hermitian matrices $H(i)$ with spectra $\lambda(i), \, 1 \leq i \leq m,$ and positive semi-definite $n \times n$ matrices $B(\alpha(i))$ such that $$\label{Horn-Sun}
H(i) = B(\alpha(i)) + B(\alpha(i-1)), \quad 1 \leq i \leq m, \;\;$$ where $B(\alpha(0)) = B(\alpha(m))$. Solving for any of the $B(\alpha(i))$ gives $\sum_{i \text{ even}} H(i) = \sum_{i \text{ odd}} H(i)$. Furthermore, because each $H(i)$ is a sum of positive semi-definite matrices, each $H(i)$ must have non-negative eigenvalues. In addition, we get several other conditions on the Hermitian matrices, namely, we can express alternating sums of an odd number of consecutive indexed matrices as a sum of positive semi-definite matrices. Specifically, $$H(i) - H(i+1) + \cdots - H(i+j-1) + H(i+j) = B(i-1) + B(i+j), \quad j \in \{0,2,4,\ldots, 2k-2\},$$ where we are taking $H(m+1) = H(1)$, and so on in cyclic fashion. Thus, each such alternating sum is positive semi-definite. (There is, of course, some redundancy in this statement and the previously stated conditions on the $H(i)$.) These are all the conditions on the $H(i)$ which we can conclude from (\[Horn-Sun\]). Thus, we’ve found the necessary conditions, stated above, posing the following problem and proving the subsequent statement.
**Generalized eigenvalue problem for $f$**. For which weakly decreasing sequences $\lambda(1),\ldots, \lambda(2k)$, $k \geq 2$, of $n$ non-negative real numbers do there exist $n \times n$ complex Hermitian matrices $H(1), \ldots, H(2k)$ with eigenvalues $\lambda(1),\ldots, \lambda(2k)$ such that $$\sum_{i \text{ even}} H(i) = \sum_{i \text{ odd}} H(i),$$ and such that $$H(i) - H(i+1) + \cdots - H(i+j-1) + H(i+j), \quad j \in \{0,2, \ldots, 2k-2\},$$ has non-negative eigenvalues, where $H(2k+1) = H(1)$ and so on in cyclic fashion?
\[eigenvalue-sun\] Suppose $\lambda(1),\ldots, \lambda(2k)$, $k \geq 2$, are weakly decreasing sequences of $n$ non-negative real numbers, and let $Q$ be the sun quiver, $\beta$ the standard dimension vector, and $\sigma_1$ the weight defined in equation (\[weight-sigma\_1\]). If $\sigma_1 \in {C(Q,\beta)}$, then there exist $n \times n$ complex Hermitian matrices $H(1), \ldots, H(2k)$ with eigenvalues $\lambda(1),\ldots, \lambda(2k)$ that solve the generalized eigenvalue problem for the multiplicity $f$.
While an effective weight defines Hermitian matrices satisfying these conditions, the conditions on the matrices are not sufficient; counterexamples are easily found. Alone, they do not determine a weight because we cannot recapture the decompositions of each $H(i)$ into a sum of the particular positive semi-definite matrices, no canonical choice being available. Any additional conditions would need to record the “linkage" between the consecutive $H(i)$, that is, the fact that they share a common positive semi-definite matrix in their decompositions.
Define the set $K(n,m) \subseteq {\mathbb{R}}^{mn}$, $m \geq 4$ and even, to be all $m$-tuples $(\lambda(1),\ldots, \lambda(m))$ of weakly decreasing sequences of $n$ reals that satisfy $\sum_{i \text{ even}} |\lambda(i)| = \sum_{i \text{ odd}} |\lambda(i)|$ and $$\sum_{j \in I_i} \sum_{i \text{ even}} \lambda(i)_j \leq \sum_{j \in I_i} \sum_{i \text{ odd}} \lambda(i)_j$$ for every tuple $(I_1, \ldots, I_m)$ such that the ${\underline{\lambda}}(I_i)$, $1 \leq i \leq m$, are partitions and $$f({\underline{\lambda}}(I_1), \ldots, {\underline{\lambda}}(I_m)) \neq 0.$$ This makes $K(n,m)$ a rational convex polyhedral cone in ${\mathbb{R}}^{mn}$, which we call the [***generalized Klyachko’s cone***]{} for this eigenvalue problem.
The first and third statements follow from Proposition \[conditions-weight-cone\], while the second follows from additionally Proposition \[eigenvalue-sun\]. Letting $Q$ denote the sun quiver, there is a map of cones $$K(n,m) \to C(Q, \beta) \qquad \quad
(\lambda(1), \ldots, \lambda(m)) \mapsto \sigma_1.$$ This map is an isomorphism of cones by the chamber inequalities in Lemma and Proposition \[eigenvalue-sun\]. We found the dimension of $C(Q, \beta)$ to be $mn-1$ in Corollary \[dim-cones\], which proves the last statement.
Factorization formula {#section-factorization}
=====================
Derksen and Weyman showed the following result for the star quiver (the quiver they used to represent a single Littlewood-Richardson coefficient as the dimension of a weight space of semi-invariants), where $\beta$ is the corresponding dimension vector.
\[decomp\] For all $\beta_1, \beta_2$ such that $\beta = \beta_1 + \beta_2$, where $\beta_1, \beta_2$ are nondecreasing along the flags and $\beta_1 \circ \beta_2 =1$, the inequality $\sigma(\beta_1) \leq 0$ defines a wall of ${\mathbb{R}}^+ C(Q, \beta)$. Furthermore, all nontrivial walls are obtained in this way.
The same proof for the first part of the theorem applies to the sun quiver because the proof only relies on $Q$ being acyclic, $\beta$ being Schur, and the $\beta_1, \beta_2$ satisfying the given assumptions. The second part of the statement also easily extends to our setup. Thus, the each weight $\sigma_1$ satisfying the previously defined conditions defines a wall of ${\mathbb{R}}^+ C(Q_\beta)$, and all nontrivial walls of the cone are defined in this way. Because of this, we can extend their proof of the factorization formula for Littlewood-Richardson coefficients to generalized ones.
Let $Q$ be a quiver without oriented cycles, $\alpha$ a dimension vector, and $\sigma$ a weight such that $\sigma(\alpha) = 0$. We call $\alpha$ [*$\sigma$-(semi-)stable*]{} if a general representation of dimension $\alpha$ is $\sigma$-(semi-)stable. We write a decomposition of $\alpha$ into smaller dimension vectors as $\alpha = \alpha_1 \dotplus \ldots \dotplus \alpha_s$ and call this the [*$\sigma$-stable decomposition*]{} of $\alpha$ if a general representation $V$ of dimension $\alpha$ has a Jordan-Hölder filtration with composition factors of dimension $\alpha_1, \ldots, \alpha_s$, in some order, including multiplicity.
We may rewrite the $\sigma$-stable decomposition of a dimension vector $\alpha$ by grouping together the common sub-dimension vectors. If $\alpha_i$ occurs $c_i$ times as the dimension vector of a composition factor in the Jordan-Hölder filtration of $\alpha$, we write the $\sigma$-stable decomposition of $\alpha$ as $$\alpha = c_1 \cdot \alpha_1 \dotplus c_2 \cdot \alpha_2 \dotplus \ldots \dotplus c_s \cdot \alpha_s,$$ where $c_i \in {\mathbb{Z}}^+$ for all $i$ and $\alpha_i \neq \alpha_j$ if $i \neq j$.
\[si-factorization\] Suppose $\sigma$ is an indivisible weight. If $$\alpha = c_1 \cdot \alpha_1 \dotplus c_2 \cdot \alpha_2 \cdot \dotplus \ldots \dotplus c_r \cdot \alpha_r$$ is the $\sigma$-stable decomposition of $\alpha,$ then for any $s \in {\mathbb{Z}}$, there is an equality $$\operatorname{dim}\operatorname{SI}(Q,\alpha)_{s \sigma} = \prod_{i=1}^r \operatorname{dim}(S^{c_i} (\operatorname{SI}(Q, \alpha_i)_{s\sigma})),$$ where $S^{c_i}$ is the $c_i^{th}$-symmetric power.
Theorem \[factorization\] rephrases Theorem 7.14 in [@DW11] for generalized Littlewood-Richardson coefficients, and the proof is essentially the same.
The conditions on the subsets $I_j$ are precisely those defining the set $T(n,m)$ as shown in Lemma \[sun-Horn\]. This set describes when the multiplicity $f$ is nonzero, as shown in Proposition \[Horn-type-inequalities\]. Furthermore, because the multiplicity $f$ agrees with the dimension of the respective weight space of semi-invariants, we can use Theorem \[decomp\].
If $\sigma_1$ is in the interior of the wall, then the $\sigma_1$-stable decomposition of $\beta$ is $\beta_1 \dotplus \beta_2$. The weight $\sigma_1$ is indivisible, so we may use Theorem \[si-factorization\] to get $$f(\lambda(1),\ldots, \lambda(m))= \operatorname{dim}\operatorname{SI}(Q,\beta)_{\sigma_1} = \alpha \circ \beta = (\alpha \circ \beta_1)(\alpha \circ \beta_2) = f (\lambda(1)^*,\ldots, \lambda(m)^*) \cdot f (\lambda(1)^\#,\ldots, \lambda(m)^\#).$$ If, on the other hand, $\sigma_1$ is not in the interior of a wall, then the $\sigma_1$-stable decompositions of $\beta_1$ and $\beta_2$ are of the form $$\beta_1 = c_1\cdot \gamma_1 \dotplus \ldots \dotplus c_s \cdot \gamma_s, \qquad \beta_2 = d_1 \cdot \delta_1 \dotplus \ldots \dotplus d_t \cdot \delta_t.$$ Thus, the $\sigma_1$-stable decomposition of $\beta$ is the sum of these. Because the sets $\{\gamma_1, \ldots, \gamma_s\}$ and $\{\delta_1, \ldots, \delta_t\}$ are disjoint and $\gamma_i \circ \delta_j = 1$ for all $i,j$, Theorem \[si-factorization\] again gives $$\begin{array}{rcl}
f(\lambda(1),\ldots, \lambda(m)) & =& \alpha \circ \beta \\
& = & \prod (\alpha \circ (c_i \cdot \gamma_i)) \prod (\alpha \circ (d_i \cdot \delta_i)) \\
& = & (\alpha \circ \beta_1)(\alpha \circ \beta_2) \\
& = & f (\lambda(1)^*,\ldots, \lambda(m)^*) \cdot f (\lambda(1)^\#,\ldots, \lambda(m)^\#).
\end{array}$$
Level-1 weights and stretched polynomials {#stretched-polynomials}
=========================================
The stretched function $f(N) = c^{N\nu}_{N\lambda, N\mu}$ for $N \in {\mathbb{Z}}^+$ for fixed partitions $\lambda, \mu, \nu$ has interesting combinatorial properties and has been studied by many people (see, for instance, [@KTT04], [@KTT06a], and [@KTT06b]). Inspired by Kirillov’s proof that stretched Kostka numbers are polynomial in the stretching factor $N$ for fixed partitions, King, Tollu, and Toumazet made a similar conjecture for stretched Littlewood-Richardson numbers.
For all partitions $\lambda, \mu,$ and $\nu$, there exists a polynomial $P^\nu_{\lambda, \mu}(N)$ in $N$ with nonnegative rational coefficients such that $P^\nu_{\lambda, \mu}(1) =c^\nu_{\lambda,\mu}$ and $P^\nu_{\lambda, \mu}(N) = c^{N\nu}_{N\lambda, N\mu}$ for all positive integers $N$.
Along with this, Fulton conjectured that if $P(1) = 1$, then $P(N) = 1$ for all $N \geq 1$, while King, Tollu, and Toumazet conjectured that if $P(1) =2$, then $P(N) = N+1$. Derksen and Weyman proved the polynomiality conjecture [@DW02] and Rassart [@Ras04] proved it again shortly afterwards, while Fulton’s conjecture was first proven combinatorially by Knutson, Tao, and Woodward [@KTW04], then later geometrically by Belkale [@Bel07], and again using quivers [@DW11]. The conjecture that $P(N) =N+1$ when $P(1)=2$ was first proven combinatorially by Ikenmeyer [@Ike16] and then geometrically and by using quivers by Sherman in [@She15] and [@She17], respectively.
In this section we explicitly compute the stretched function for certain weights for the sun quiver and verify that similar statements hold for the respective generalized Littlewood-Richardson coefficients. We also note that while similarly defined weights for the star quiver all lie on extremal rays of the cone of effective weights, this is not true for our case.
Level-$1$ weights
-----------------
In [@Fei15], Fei defines a weight for the star quiver to be level-$m$ if the weight has value $m$ at the central vertex. In Lemma $2.3$ of the paper, he classifies all level-$1$ effective weights and shows that they lie on an extremal ray. For the star quiver, the Littlewood-Richardson coefficient arising from any effective level-1 weight is of the form $c^{1^{i+j}}_{1^i, 1^j}$. We use this idea to describe similar weights for the sun quiver.
We define a [***level-1 weight***]{} for the sun quiver to be a (for now, not necessarily effective) nonzero weight with at most one nonzero entry along any flag, with the nonzero entry being $1$ for the flags going out and $-1$ for the flags going in. This will correspond to at most one jump along each flag for the defining dimension vector. Because $\sigma(\beta)=0$ is a necessary condition, if $j_1, \ldots, j_m$ are the vertices along the flags for which $\sigma(j_i) \neq 0$, counting the vertices towards the central ones (so $j_i$ is vertex $(j_i,i)$), with $j_i=0$ to mean that the weight is trivial along flag $i$, then $\sum_{i \text{ odd}} j_i = \sum_{i \text{ even}} j_i$ is necessary, though not sufficient.
We’ll find it useful in this section to describe the effective weights directly in terms of the jumping numbers $j_i$ rather than by the conditions found in Proposition \[eigenvalue-sun\].
\[level-1-weight\] Let $\sigma$ be a level-$1$ weight for the sun quiver $Q$ and let $j_1, \ldots, j_m$ be the vertices along flags $1, \ldots, m$ for which $\sigma(j_i) \neq 0$ with $j_i = 0$ if $\sigma$ is trivial on flag $i$. Then the following are equivalent:
1. $\sigma \in {C(Q,\beta)}$;
2. ${\displaystyle}\sum_{i \text{ odd}} j_i = \sum_{i \text{ even}} j_i$ and $j_i - j_{i+1} + j_{i+2} \geq 0$ for $1 \leq i \leq m$, where $j_{m+1} = j_1$ and $j_{m+2} = j_2$.
Because the partition arising from a flag with the only nonzero weight being $1$ or $-1$ at vertex $j_i$ is $(1^{j_i})$ if $j_i \neq 0$ and $(0)$ if $j_i = 0$, the generalized Littlewood-Richardson coefficient arising from this weight is $${\displaystyle}f((1^{j_1}), \ldots, (1^{j_m}))= \sum c^{(1^{j_1})}_{\alpha_1,\alpha_2}\cdot c^{(1^{j_2})}_{\alpha_2,\alpha_3}\cdots c^{(1^{j_{m-1}})}_{\alpha_{m-1}, \alpha_m} \cdot c^{(1^{j_m})}_{\alpha_m, \alpha_1}.$$ As observed above, $\sum_{i \text{ odd}} j_i = \sum_{i \text{ even}} j_i$ is equivalent to $\sigma(\beta) = 0$ and because $$\alpha_i = j_i - \alpha_{i+1} =j_i - j_{i+1} + \alpha_{i+2} = j_i - j_{i+1} + j_{i+2} - \alpha_{i+3},$$ the condition $j_i - j_{i+1} + j_{i+2} \geq 0$ is necessary. It’s easy to check that the inequalities on the $j_i$’s imply that $j_i \leq j_{i+1}$ and $j_{i-1} \leq j_{i-2}$ for some $i$, which will be sufficient to prove that such a weight is effective. After reindexing, suppose $j_1 \leq j_2$ and $j_m \leq j_{m-1}$. Then the choice $\alpha(1)=0$ uniquely determines the other $\alpha(i)$ and it follows from the above conditions that each $\alpha(i) \geq 0$, making the weight effective.
There are fewer conditions on the $j_i$ for the level-$1$ weights to be effective than those stated for general weights in the context of generalized eigenvalue problem for Hermitian matrices in Proposition \[eigenvalue-sun\]. This is because in the general case we can’t say that either $H(i) - H(i+1)$ or $H(i+1)-H(i)$ is positive semi-definite, while we can make such a direct comparison of $j_i$ and $j_{i+1}$. This allowed us to have an $i$ such that $j_i \leq j_{i+1}$ and $j_{i-1} \leq j_{i-2}$, proving that such a weight was effective. However, this at least tells us that if there is an $i$ such that $H(i+1) - H(i)$ and $H(i-2) - H(i-1)$ are positive semi-definite (along with the other conditions on the $H(i)$), then the $H(i)$ solve the generalized eigenvalue problem. These conditions, though, are not necessary.
As opposed to the case for the star quiver, not every effective level-1 weight lies on an extremal ray for the sun quiver. We found several such weights lying on the facets in the case $n=2, m=6$. It can be checked that the first weight in the provides an instance of a level-1 weight on an extremal ray while the second weight in the first row provides one which is not.
We now want to determine the value $\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_\sigma$ for a level-1 weight $\sigma$.
\[value-level-1\] Let $\sigma \in {C(Q,\beta)}$ be a level-1 weight for the sun quiver $Q$. Let $j_1,\ldots, j_m$ be the jumping numbers defining the weight and define $J_i = j_i-j_{i+1}+j_{i+2}$, where $j_{m+1} = j_1$ and $j_{m+2} = j_2$. If $s=\min\{j_i,J_i \mid 1 \leq i \leq m\}$, then $\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_\sigma = s+1$.
We first show that $\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_\sigma \leq s+1$. Throughout we will denote $(1^{j_i})$ as simply $j_i$. Clearly, any choice of some $\alpha_i$ completely determines each of the other $\alpha_j's$. Moreover, because each partition is of the form $(1^{j_k})$, $c^{j_k}_{\alpha_k, \alpha_{k+1}} = 1$ whenever it’s nonzero. If $i$ is such that $J_i$ or $j_i$ is minimal among the set, then consider the factors $$c^{j_i}_{\alpha_i,\alpha_{i+1}} \cdot c^{j_{i+1}}_{\alpha_{i+1},\alpha_{i+2}} \cdot c^{j_{i+2}}_{\alpha_{i+2},\alpha_{i+3}}$$ in the summation. Then $\alpha_i \leq j_i$ and similarly because $\alpha_{i+3} = j_{i+2}-j_{i+1}+j_i - \alpha_i = J_i - \alpha_i$, we must have $\alpha_i \leq J_i$ in order for this factor to be nonzero. Hence, there are at most $s+1$ choices for $\alpha_i$ resulting in this factor being nonzero.
Suppose $s =j_i$ for some $i$. To show equality, we only need to show that $\alpha_k \geq 0$ for each $k$ for each choice of $\alpha_i \in \{0,\ldots, s\}$. This is quickly done since $i$ was chosen to be such that $j_i \leq j_k$ and $j_i \leq J_k$ for each $k$ along with $\alpha_i \leq j_i$. Similarly, if $s = J_i$, each $\alpha_k$ will be nonnegative after noticing that $j_i \leq j_{i+1}$ in this case since $J_i \leq j_{i+2}$. Thus, the only choices for $\alpha_i$ resulting in a nonzero term in the summation are $0,1,\ldots, s$, and each such choice results in adding one to the summation.
Stretched weights
-----------------
For a level-1 weight $\sigma$, we’re interested in the stretched weights $N \sigma$, $N \in {\mathbb{Z}}^+$. If $j_1, \ldots, j_m$ are the jumping numbers of $\sigma$, the corresponding partitions will be $(N^{j_i})$. Because $|(N^{j_i})| = Nj_i$ for all $i$, Lemma \[level-1-weight\] generalizes immediately to stretched level-1 weights. Similarly, Lemma \[value-level-1\] generalizes in this case because of the following lemma which is quickly checked by using the Littlewood-Richardson rule.
\[rectangular-partitions-1\] Let $\lambda, \mu$ be partitions and $\nu = (N^n)$ a rectangular partition. Then $c^\nu_{\lambda, \mu}$ is either 0 or 1. It is equal to 1 if and only if $\lambda_i + \mu_{n+1-i}=N$ for $i=1,\ldots, n$.
In the proof of the next lemma, we use a partial ordering on the set of rectangular partitions $(N^n)$ for a fixed $N$ defined by $\lambda_1 \leq \lambda_2$ to mean that the Young diagram of $\lambda_1$ fits inside that of $\lambda_2$, meaning $\lambda_2 - \lambda_1$ is a partition. With this, $\lambda_1 \dotplus \lambda_2$ means stacking the corresponding Young diagrams on top of each other (or in terms of partitions, $(N^{n_1}) + (N^{n_2}) = (N^{n_1+n_2}))$, so $\lambda_3 \leq \lambda_1 \dotplus \lambda_2$ means the Young diagram of $\lambda_3$ fits inside the stacked diagrams of $\lambda_1$ and $\lambda_2$, or equivalently, $(N^{n_1 + n_2 - n_3})$ is a partition. We will use the notation $\lambda_1 \dotplus (-\lambda_2)$ to mean we are instead subtracting $\lambda_2$ from the bottom of the diagram of $\lambda_1$.
\[effective stretched weight\] For a level-1 weight $\sigma$ for the sun quiver, let $N \in {\mathbb{Z}}^+$, $j_1, \ldots, j_m$ the corresponding jumping numbers, and $J_i = j_i - j_{i+1} + j_{i+2}$ for $i=1, \ldots, m$, with $j_{m+1} = j_1$ and $j_{m+2} = j_2$. The following are equivalent:
1. $N\sigma \in C(Q,\beta)$;
2. ${\displaystyle}\sum_{i \text{ odd}} j_i = \sum_{i \text{ even}} j_i$ and $J_i \geq 0$ for all $i$.
If $N\sigma$ is effective, then $\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_{N\sigma} = {N+s \choose N}$, where $s= \min\{j_i, J_i \mid 1 \leq i \leq m\}$.
As mentioned above, the necessary and sufficient conditions for $N \sigma$ to be effective are proven the same way as in Lemma \[level-1-weight\]. We adapt the proof of Lemma \[value-level-1\] to compute the value of the dimension of the weight space for the stretched case.
First suppose $s=j_i$, and without loss of generality, suppose $i=1$. The number of partitions $\alpha_1$ such that $\alpha_1 \leq (N^{j_1})$ is ${N+j_1 \choose N}$. This is because if we consider the Young diagram corresponding to $(N^{j_1})$, then we choose how many entries of $\alpha_1$ have value $N$, then how many have value $N-1$, and so on, which is the same as choosing where to place $N$ dividers among $j_1$ entries. Because each $\alpha_1$ uniquely determines the other $\alpha_i$ and because each $c^{(N^{j_i})}_{\alpha_i, \alpha_{i+1}}$ is equal to one when nonzero by Lemma \[rectangular-partitions-1\], dim ${\operatorname{SI}(Q,{\beta})}_{N\sigma} \leq {N + j_1 \choose N}$. The other direction is proved in a similar way as Lemma \[value-level-1\].
The value of $f(\lambda(1),\ldots,\lambda(m))$ is independent of the value of $n$, the length of each flag. This number can only enlarge the value of the coefficient, which is instead determined by the smallest $j_i$ or $J_i$. This formula also agrees with the value that we found in Lemma \[value-level-1\] since in that case $N=1$.
For a fixed weight $\sigma$ for the sun quiver, we showed in Lemma \[si-saturation\] that for each $N \geq 1$, $$f(N\lambda(1),\ldots, N\lambda(m)) = \dim {\operatorname{SI}(Q,{\beta})}_{N\sigma},$$ where $\lambda(1),\ldots, \lambda(m)$ are the partitions arising from $\sigma$ as stated in equation (\[weight-sigma\_1\]). Clearly, this is a polynomial as each stretched function of a single Littlewood-Richardson coefficient is a polynomial. The above formula allows us to calculate $\operatorname{dim}{\operatorname{SI}(Q,{\beta})}_{N \sigma}$ for any level-1 weight immediately.
\[LR-sun-polynomial\] Let $\lambda(1),\ldots, \lambda(2k)$, $k \geq 2$, be partitions of at most $n$ parts and of the form $(1^{j_i})$ if $j_i \neq 0$ and zero if $j_i = 0$ for some integers $0\leq j_1, \ldots, j_m \leq n$. Suppose the $j_i$ satisfy the following conditions:
1. ${\displaystyle}\sum_{i \text{ odd}} j_i = \sum_{i \text{ even}} j_i$;\
2. ${\displaystyle}J_i := j_i - j_{i+1} + j_{i+2} \geq 0$, where $j_{2k+1} = j_1, \; j_{2k+2} = j_2$.
Then for any $N \in {\mathbb{Z}}^+$, the stretched Littlewood-Richardson polynomial $f(N\lambda(1),\ldots, N\lambda(2k))$ is equal to ${N + s \choose N}$, where $s = \min \{j_i,J_i \mid 1 \leq i \leq 2k\}$. If either (1) or (2) is not satisfied, then $f(N\lambda(1),\ldots, N\lambda(2k)) = 0$.
This follows immediately from Lemma \[effective stretched weight\] and Theorem \[saturation-sun\].
Proposition \[LR-sun-polynomial\] shows that the conjectures of King, Tollu, and Toumazet, and of Fulton are true for these generalized coefficients in the cases that the partitions are of the stated forms. Namely, as a function of $N \in {\mathbb{Z}}^+$, the stretched Littlewood-Richardson function is a polynomial $P$ and whenever $P(1) = 1$, $P(N) =1$, and when $P(1) = 2$, $P(N) = N+1$. In addition, we showed the saturation property (Theorem \[saturation-sun\]) saying $P(1) = 0$ implies $P(N) = 0$. Furthermore, it has been conjectured ([@KTT04], Conjecture 3.3) that $P(1)=3$ implies $P(N)$ is either $2N+1$ or ${N+2 \choose N}$, which also agrees with our results. It would be interesting to see if similar conjectures for these generalized coefficients hold for all weights.
Polytopal description and complexity {#polytope}
====================================
In this section we examine the complexity of the branching multiplicity by defining a polytope whose number of lattice points is equal to the multiplicity. The main result is Theorem \[sun-complexity\], which states that the positivity of the multiplicity, that is, whether or not it is zero, can be calculated in strongly polynomial time.
Geometric complexity theory
---------------------------
Geometric complexity theory (GCT) was introduced by Mulmuley and Sohoni in a series of papers (see [@MS07], [@MS01a], [@MS01b], [@MS08], [@MNS12], [@MS17], [@Mul10], [@Mul11]) in the early 2000’s with the purpose of approaching fundamental problems in complexity theory, such as P vs. NP, through algebraic geometry and representation theory. Previously, [@KT01] and [@LM06] had independently shown that the positivity of Littlewood-Richardson coefficients could be computed in polynomial time while [@Nar05] had shown that the actual computation of these numbers was a \#P-complete problem, the complexity class for problems for which (unless P=NP) there does not exist a polynomial time algorithm for computing them (rather, it takes an exponential time in the worst case), and such that the computation is at least as difficult as every P problem.
The following is the main theorem of [@MS05], where the *bit length* of a partition $\lambda = (\lambda_1, \ldots, \lambda_k), \, \lambda_k > 0, $ is the bit length of the specifications: $\sum_{i=1}^k\log_2 \lambda_i$.
Deciding whether $c^{\nu}_{\lambda,\mu}$ is positive can be computed in strongly polynomial time in the sense of [@GLS93]. This means that the number of arithmetic steps is polynomial in the number of positive parts of $\nu$ (say $n$), does not depend on the bit lengths of $\lambda_i, \mu_j, \nu_k$, and the bit length of every intermediate operand that arises in the algorithm is polynomial in the total bit length of $\lambda, \mu, \nu$.
In fact, by attaching zeros to the partitions, one can subsume the dependence on $n$ into the bit lengths of $\lambda,\mu,$ and $\nu$. This is especially amazing as the the dimension of the Weyl module $S^\nu(V)$ is exponential in $n$ and the bit lengths of the $\nu_k$’s, yet deciding if an exponential dimensional object $S^\nu(V)$ arises in the decomposition of another exponential dimensional object $S^\lambda(V) \otimes S^\mu(V)$ can be decided in time that is polynomial in only $n$ and the bit lengths of the labels $\lambda,\mu,$ and $\nu$.
Because of results such as these along with the ubiquity of the plethysm problem and related problems in representation theory, GCT allows one to compare the complexity of several problems. The proof of deciding the positivity of a Littlewood-Richardson coefficient relies on two main points: a polyhedral interpretation of these numbers and the saturation theorem. While we define a polytope for the generalized Littlewood-Richardson coefficients to prove a similar result, it would be nice to have a purely combinatorial algorithm, such as those of max-flow or weighted matching problems in combinatorial optimization. Much work has been made towards this for single Littlewood-Richardson coefficients (see [@BI09], [@BI13], and [@Ike16]).
Polytopal description
---------------------
In order to determine the complexity of the positivity of multiplicity (\[one\]), we will define a polytope by determining a system of homogeneous linear inequalities whose number of integer-valued solutions is precisely the multiplicity. The idea is to use the Littlewood-Richardson hives defined by Knutson and Tao in [@KT99].
To define the polytope associated with multiplicity (\[one\]), subdivide a regular $m$-gon into $m$ triangles with $n+1$ vertices along each exterior edge and a common vertex at the center. Subdivide each of these triangles into $n^2$ triangles of the same size, so the hexagon is divided into $mn^2$ total triangles. We label the edges in the $r^{th}$ triangular array in the following way: the first subscript $i$ refers to the row from bottom to top while the second subscript $j$ refers to the diagonal from left to right, and $0 \leq i,j \leq n-1$. The edges along increasing diagonals are labeled $e_{ij},$ the edges along decreasing diagonals are labeled $f_{ij}$, and the horizontal edges are $g_{ij}$. The superscript $r$ refers to which triangular array is used, though this is often neglected. For instance, when $n=3$ the $r^{th}$ triangular array is labeled $$\begin{tikzpicture}
\draw (0,0) -- ++(0:6) -- ++(120:6) -- cycle;
\draw (0,0) -- ++(0:2) -- ++(120:2) -- cycle;
\draw (2,0) -- ++(0:2) -- ++(120:2) -- cycle;
\draw (4,0) -- ++(0:2) -- ++(120:2) -- cycle;
\draw (2,0) -- ++(60:2) -- ++(180:2) -- cycle;
\draw (4,0) -- ++(60:2) -- ++(180:2) -- cycle;
\draw (3,1.7) -- ++(60:2) -- ++(180:2) -- cycle;
\coordinate [label={below left: }] ($a_{00}$) at (0,0);
\coordinate [label={below: }] ($a_{10}$) at (2,0);
\coordinate [label={below:}] ($a_{20}$) at (4,0);
\coordinate [label={below:}] ($a_{30}$) at (6,0);
\coordinate [label={left:}] ($a_{01}$) at (1,1.7);
\coordinate [label={below: }] ($a_{11}$) at (3,1.7);
\coordinate [label={right:}] ($a_{21}$) at (5,1.7);
\coordinate [label={left:}] ($a_{02}$) at (2,3.4);
\coordinate [label={right:}] ($a_{12}$) at (4,3.4);
\coordinate [label={above:}] ($a_{03}$) at (3,5.15);
\foreach \i in {$a_{00}$,$a_{01}$,$a_{02}$,$a_{03}$,$a_{10}$,$a_{11}$,$a_{12}$,$a_{20}$,$a_{21}$,$a_{30}$}
\fill (\i) circle (2pt);
\coordinate [label={below: $g_{00}$ }] () at (1,0);
\coordinate [label={below: $g_{01}$}] () at (3,0);
\coordinate [label={below: $g_{02}$}] () at (5,0);
\coordinate [label={below: $g_{10}$}] () at (2,1.7);
\coordinate [label={below: $g_{11}$}] () at (4,1.7);
\coordinate [label={below: $g_{20}$ }] () at (3,3.5);
\coordinate [label={left: $e_{00}$}] () at (.5,.8);
\coordinate [label={left: $f_{00}$}] () at (1.6,.8);
\coordinate [label={left: $e_{01}$}] () at (2.5,.8);
\coordinate [label={left: $f_{01}$}] () at (3.6,.8);
\coordinate [label={right: $e_{02}$}] () at (4.5,.8);
\coordinate [label={right: $f_{02}$}] () at (5.7,.8);
\coordinate [label={left: $e_{10}$}] () at (1.5,2.6);
\coordinate [label={left: $f_{10}$}] () at (2.6,2.6);
\coordinate [label={right: $e_{11}$}] () at (3.5,2.6);
\coordinate [label={right: $f_{11}$}] () at (4.6,2.6);
\coordinate [label={left: $e_{20}$}] () at (2.5,4.4);
\coordinate [label={right: $f_{20}$}] () at (3.5,4.4);
\end{tikzpicture}$$
Let $E$ be the set of hive edges and ${\mathbb{R}}^E$ the labelings of these edges by real numbers. There are three ways that two adjacent triangles inside a single triangular array can form a rhombus: $$\begin{tikzpicture}
\coordinate [label={above: $g_{i+1j}$}] () at (-2,.5);
\coordinate [label={below: $g_{ij}$}] () at (-3,-.5);
\coordinate [label={left: $e_{ij}$}] () at (-3.5,0);
\coordinate [label={right: $e_{i+1j}$}] () at (-1.5,-.2);
\coordinate [label={below left:}] (a) at (-4,-.5);
\coordinate [label={above left:}] (b) at (-3,.5);
\coordinate [label={below right:}] (c) at (-2,-.5);
\coordinate [label={above:}] (d) at (-1,.5);
\draw (a)--(b)--(d)--(c)--cycle;
\draw (b)--(c);
\foreach \i in {a,b,c,d}
\fill (\i) circle (2pt);
\coordinate [label={left: $e_{i+1j}$}] () at (1.5,.5);
\coordinate [label={right: $f_{i+1j}$}] () at (2.5,.5);
\coordinate [label={left: $f_{ij}$}] () at (1.5,-.6);
\coordinate [label={right: $e_{ij+1}$}] () at (2.5,-.6);
\coordinate [label={above:}] (a) at (2,1);
\coordinate [label={left:}] (b) at (1,0);
\coordinate [label={right:}] (c) at (3,0);
\coordinate [label={below:}] (d) at (2,-1);
\draw (a)--(b)--(d)--(c)--cycle;
\draw (b)--(c);
\foreach \i in {a,b,c,d}
\fill (\i) circle (2pt);
\coordinate [label={left: $f_{ij}$}] () at (5.3,-.1);
\coordinate [label={above: $g_{i+1j}$}] () at (6,.5);
\coordinate [label={below: $g_{ij+1}$}] () at (7,-.5);
\coordinate [label={right: $f_{ij+1}$}] () at (7.5,0.2);
\coordinate [label={above:}] (a) at (5,.5);
\coordinate [label={above right:}] (b) at (7,.5);
\coordinate [label={below: }] (c) at (6,-.5);
\coordinate [label={below right:}] (d) at (8,-.5);
\draw (a)--(b)--(d)--(c)--cycle;
\draw (b)--(c);
\foreach \i in {a,b,c,d}
\fill (\i) circle (2pt);
\end{tikzpicture}$$
We say these rhombi satisfy the [***rhombus inequalities***]{} if for each triangle and rhombus appearing, we have
$$\label{rhombus-edge-inequalities}
e_{ij} \geq e_{ij+1}, \;\;\; g_{ij} \geq g_{i+1 j}; \qquad
f_{i+1 j} \geq f_{ij}, \;\;\; e_{i j+1} \geq e_{i+1 j}; \qquad
f_{ij} \geq f_{i j+1}, \;\;\; g_{i+1 j} \geq g_{i j+1};$$
$$e_{ij} + f_{ij} = g_{ij}, \qquad \quad e_{i \, j+1} + f_{ij} = g_{i+1 \,j}.$$
Define an [***$(m,n)$-LR sun hive***]{} to be a regular $m$-gon subdivided into $m$ triangular arrays with $n+1$ vertices along each edge that satisfies the rhombus inequalities and the border conditions $$\label{boundary-edge-conditions}
\sum_{i=0}^{n-1} e^r_{i0} + \sum_{i=0}^{n-1} f^r_{i n-i} = \sum_{j=0}^{n-1} g_{0j}$$ for each $1 \leq r \leq m$. It is [***integral***]{} if the labeling lies in ${\mathbb{Z}}^E$. These inequalities define a convex polyhedral cone, denoted $C \subseteq {\mathbb{R}}^E$. An [***LR hive***]{} is a single triangular array that satisfies the rhombus inequalities and border conditions for that array, so an $(m,n)$-LR sun hive consists of $m$ LR hives with $n$ edges along each side of the boundary of a regular $m$-gon and with the respective conditions and appropriately shared sides. Let $B$ be the set of border edges $g^k_{0j}$ for $1\leq r \leq m, \; 0 \leq j \leq n-1$, and $\rho: {\mathbb{R}}^E \to {\mathbb{R}}^B$ the restriction map of an LR sun hive to its border. For each $b \in {\mathbb{R}}^B$, the fiber $\rho^{-1}(b) \cap C$ is a compact polytope, called the [***$m$-sun hive polytope***]{} over $b$.
We recall the main result of [@KT99]. For three $n$-tuples $$\lambda = (\lambda_1,\ldots, \lambda_n), \qquad
\mu = (\mu_1,\ldots, \mu_n), \qquad
\nu = (\nu_1, \ldots, \nu_n)$$ that satisfy the boundary condition $|\nu| = |\lambda| + |\mu|$, the triangular array with border determined by $\lambda, \mu, \nu$ is the one with specified border edges $$\begin{tikzpicture}
\coordinate (l0) at (10,0);
\coordinate (n1) at (12,0);
\coordinate (n2) at (14,0);
\coordinate (n3) at (16,0);
\coordinate (n4) at (18,0);
\coordinate (l1) at (11,1.7);
\coordinate (ln1) at (13,1.7);
\coordinate (ln2) at (15,1.7);
\coordinate (l2) at (12,3.4);
\coordinate (ln3) at (14,3.4);
\coordinate (l3) at (13,5.1);
\coordinate (l4) at (14,6.8);
\coordinate (m1) at (15,5.1);
\coordinate (m2) at (16,3.4);
\coordinate (m3) at (17,1.7);
\draw (l0)--(l4)--(n4)--cycle;
\draw (l3)--(m1)--(ln3)--(m2)--(ln2)--(m3)--(n3)--cycle;
\draw (l3)--(ln3)--(l2)--(ln1)--(n1)--(l1)--cycle;
\draw (ln3)--(ln2)--(ln1)--cycle;
\draw (ln1)--(ln2)--(n2)--cycle;
\draw (l1)--(ln1)--(n1)--cycle;
\foreach \i in {l0,l1,l2,l3,l4,n1,n2,n3,n4,m1,m2,m3,ln1,ln2,ln3}
\fill (\i) circle (2pt);
\tkzLabelSegment[left=2pt](l0,l1){$\lambda_1$}
\tkzLabelSegment[left=2pt](l1,l2){$\lambda_2$}
\tkzLabelSegment[left=2pt](l2,l3){$\ldots$}
\tkzLabelSegment[left=2pt](l3,l4){$\lambda_n$}
\tkzLabelSegment[right=2pt](l4,m1){$\mu_1$}
\tkzLabelSegment[right=2pt](m1,m2){$\mu_2$}
\tkzLabelSegment[right=2pt](m2,m3){$\ldots$}
\tkzLabelSegment[right=2pt](m3,n4){$\mu_n$}
\tkzLabelSegment[below=2pt](l0,n1){$\nu_1$}
\tkzLabelSegment[below=2pt](n1,n2){$\nu_2$}
\tkzLabelSegment[below=2pt](n2,n3){$\ldots$}
\tkzLabelSegment[below=2pt](n3,n4){$\nu_n$}
\end{tikzpicture}$$
\[LR-hives\] The Littlewood-Richardson coefficient $c^\nu_{\lambda, \, \mu}$ is the number of integer LR hives with boundary labels determined by $\lambda, \mu$, and $\nu$.
Because we are only interested in integer LR hives it suffices to restrict to when $\lambda,\mu,\nu$ are partitions. Further, if each partition has at most $n$ parts, then the LR hives are LR $n$-hives. Though the border conditions are obvious from the necessary condition that $|\nu| = |\lambda| + |\mu|$ for $c^\nu_{\lambda,\mu}$ to be nonzero, the rhombus inequalities seem mysterious at first. Their inspiration comes from Gelfand-Tsetlin patterns and Cauchy’s Interlacing Theorem for the eigenvalues of Hermitian matrices. The first two pairs of inequalities ensure that the tuples are weakly decreasing while the third pair gives a way of associating a contratableau satisfying the Littlewood-Richardson rule; for more on this, see [@Buc00].
Because one LR hive is used to calculate one Littlewood-Richardson coefficient, it stands to reason that “gluing" multiple LR hives together appropriately should be used to calculate our generalized Littlewood-Richardson coefficients. Before stating and proving this we first make precise how we intend to “glue" the LR hives. Given one LR hive we combine it with another LR hive by requiring the two to share a side other than the base. This results in the second LR hive being flipped. Of course, we need to verify the values assigned to the shared edges coincide and make precise the edge labelings along with the rhombus inequalities for the flipped hive. For the flipped hive the edges on a descending diagonal are now labeled by the $e$’s while the ascending diagonal edges are labeled by the $f$’s.
Under this notation, $e^{k+1}_{j0} = f^{k}_{n-1-j \, j}$. Flipping the triangular arrays causes each type of rhombus to be flipped, but by also switching the labels for the $e$’s and $f'$s the same rhombus inequalities in (\[rhombus-edge-inequalities\]) hold. For these flipped arrays we will want to specify when $n$-tuples $\lambda, \mu, \nu$ such that $|\nu| = |\lambda| + |\mu|$ determine the border and align along shared edges, but we will wait to do this depending on which side of the flipped hive we want to have labeled $\nu$.
When defining the $(m,n)$-LR sun hive the rhombus inequalities did not include those two types arising from rhombi of adjacent triangles from different hives. Because one hive is flipped, there is no natural way of determining which direction the inequality should be and the direction may differ in different examples. Interestingly, though, the direction of the inequalities arising from adjacent hives is the same within each individual example examined.
With this notation we may now prove the polytopal description of the generalized Littlewood-Richardson coefficient (\[one\]).
\[sun hive\] For partitions $\lambda(1), \ldots, \lambda(2k)$, $k \geq 2$, of no more than $n$ parts, the generalized Littlewood-Richardson coefficient $$\sum c^{\lambda(1)}_{\alpha(1),\alpha(2)} c^{\lambda(2)}_{\alpha(2),\alpha(3)} \cdots c^{\lambda(2k-1)}_{\alpha(2k-1),\alpha(2k)} c^{\lambda(2k)}_{\alpha(2k),\alpha(1)}$$ is equal to the number of integer $(2k,n)$-LR sun hives with external boundary labels determined by the $\lambda(i)$ in cyclic orientation (so that the edge labeled $\lambda(r)$ is between the edges labeled $\lambda(r+1)$ and $\lambda(r-1)$). For instance, the boundary labels of a $(6,n)-$LR sun hive are $$\begin{tikzpicture}
\draw (-2,2) -- ++(0:2) -- ++(300:2) -- ++(240:2) -- ++(180:2) -- ++(120:2) --++(60:2) --++(0:2) -- cycle;
\coordinate [label={above:$\lambda(2)$}] () at (-.9,2);
\coordinate [label={right:$\lambda(1)$}] () at (.6,1.3);
\coordinate [label={right: $\lambda(6)$}] () at (.5,-.8);
\coordinate [label={below:$\lambda(5)$}] () at (-.9,-1.5);
\coordinate [label={left: $\lambda(4)$}] () at (-2.4,-.8);
\coordinate [label={left:$\lambda(3)$}] () at (-2.4,1.3);
\end{tikzpicture}$$
By Theorem \[LR-hives\] the Littlewood-Richardson coefficient $c^{\lambda(r)}_{\alpha(r),\alpha(r+1)}$ is equal to the number of integer LR $n$-hives with $\lambda(r)$ as the base which satisfy the boundary conditions and rhombus inequalities, where $\alpha(r), \alpha(r+1)$ are some tuples of no more than $n$ parts forming the other two sides of the $r^{th}$ triangular array. Necessarily the tuple $\alpha(r)$ is also a boundary of the $(r-1)^{th}$ triangular array while $\alpha(r+1)$ is a boundary of the $(r+1)^{th}$. We use the previously defined notation for each hive and adjacent (flipped) hive, so we only need to specify the border labels. If the base labeled $\lambda(r)$ has edges labeled $\lambda(r)_1, \ldots, \lambda(r)_n$ from *left to right*, then the adjacent base labeled $\lambda(r+1)$ has edges labeled $\lambda(r+1)_1, \ldots, \lambda(r+1)_n$ from *right to left*. In this way, edges labeled by $\alpha(r)$ and $\alpha(r+1)$ in the $r^{th}$ LR hive are $\alpha(r)_1, \ldots, \alpha(r)_n, \alpha(r+1)_1, \ldots, \alpha(r+1)_n$ clockwise while the edges in the adjacent hive are labeled $\alpha(r+2)_1,\ldots, \alpha(r+2)_n, \alpha(r+1)_1,\ldots, \alpha(r+1)_n$ counterclockwise. The multiplicity $$\sum c^{\lambda(1)}_{\alpha(1),\alpha(2)} c^{\lambda(2)}_{\alpha(2),\alpha(3)} \cdots c^{\lambda(2k-1)}_{\alpha(2k-1),\alpha(2k)} c^{\lambda(2k)}_{\alpha(2k),\alpha(1)}$$ is then equal to the number of integer $(2k,n)$-LR sun hives with these choices of $\alpha(1),\ldots, \alpha(2k)$. The total number of integer $(2k,n)$-LR sun hives with the boundaries $\lambda(1),\ldots, \lambda(2k)$ is then the sum over all possible integer tuples $\alpha(1), \ldots, \alpha(2k)$ with at most $n$ parts.
There is a characterization of LR hives with vertex labels rather than edge labels. Though the two labelings may be used interchangeably for all results concerning a single Littlewood-Richardson coefficient, the vertex labeling fails in the case of the generalized coefficients because the vertices along a shared boundary would not necessarily agree. For instance, the vertex at the center of the regular $n$-gon could only be zero while this would force the external boundary labels to not be the $\lambda(i)$.
In the previous theorem, it is necessary that the number of partitions be at least four and even. We saw that adjacent LR hives must “flip" in order to line up the boundary edges and in our description the edges on the side determined by $\lambda(i)$ are labeled by $\lambda(i)_1, \ldots, \lambda(i)_n$ from left to right for odd $i$ and in reverse order for even $i$. If the number of partitions, $m$, were odd, then the first and $m+1$ hives are the same, yet these have different parities, so we get two different labelings. The number of hives must then be even and it is easily checked that $m=2$ fails.
As we’ve seen, the rhombus inequalities (\[rhombus-edge-inequalities\]) and boundary conditions (\[boundary-edge-conditions\]) determine a polytope whose number of lattice points corresponds to the multiplicity (\[one\]) when the external boundaries are determined by the $\lambda(r)$. For each $1 \leq r \leq 2k$, these inequalities may be solved into a linear program $A_r \mathbf{x_r} \leq \mathbf{b_r}$, where $A_r$ is a matrix with entries $0,1,-1$, $\mathbf{x_r}$ is the vector of interior edges $e^r_{ij},\, f^r_{ij}, \, g^r_{ij} \; 0 \leq i \leq n-1, \, 0 \leq j \leq n-1$, and the entries of $\mathbf{b_r}$ are homogeneous, linear forms in the entries of $\lambda(r)$ (which are integral when $\lambda(r)$ is a partition). Because this can be done for each $r$, we can express all of these as a single linear program $A \mathbf{x} \leq \mathbf{b}$, where $A$ is the block sum of the matrices $A_r$ and similarly for $\mathbf{x}$ and $\mathbf{b}$. Again, $\mathbf{b}$ will be homogeneous, which is necessary for the proof of the complexity of the positivity of the generalized Littlewood-Richardson coefficient. In this way, the multiplicity is equal to the number of integer-valued vector solutions $\mathbf{x}$ to this inequality. This proves the following.
\[linear-program\] For partitions $\lambda(1), \ldots, \lambda(2k)$, $k \geq 2$, there exists a linear program $A\mathbf{x} \leq \mathbf{b}$, where the matrix $A$ has entries $0,1,-1,$ $\mathbf{b}$ is a vector of homogeneous, integral, linear forms in terms of the parts of $\lambda(1), \ldots, \lambda(2k)$, and the multiplicity $$\sum c^{\lambda(1)}_{\alpha(1),\alpha(2)} c^{\lambda(2)}_{\alpha(2),\alpha(3)} \cdots c^{\lambda(2k-1)}_{\alpha(2k-1),\alpha(2k)} c^{\lambda(2k)}_{\alpha(2k),\alpha(1)}$$ is equal to the number of solution vectors $\mathbf{x}$ with integer entries.
With this, we can prove the complexity of the positivity of this multiplicity.
First, we claim that the $m$-sun hive polytope, whose number of lattice points equals $$f(\lambda(1),\ldots, \lambda(2k)) = \sum c^{\lambda(1)}_{\alpha(1),\alpha(2)} c^{\lambda(2)}_{\alpha(2),\alpha(3)} \cdots c^{\lambda(2k-1)}_{\alpha(2k-1),\alpha(2k)} c^{\lambda(2k)}_{\alpha(2k),\alpha(1)}$$ contains an (integer) $(2k,n)-$LR sun hive if and only if it is nonempty, which is equivalent to the multiplicity being nonzero. Note that because the polytope is defined by a linear system $A \mathbf{x} \leq \mathbf{b}$ where $\mathbf{b}$ is homogeneous (Lemma \[linear-program\]), for any integer $N$, $f(N\lambda(1), \ldots, N \lambda(2k))$ is the number of integer $(2k,n)$-LR sun hives in the polytope with scaled external boundaries.
One direction of the claim is trivial, so suppose the polytope is nonempty. In particular, the polytope has a vertex. One characterization of a vertex of a polytope (see, for instance, [@Sch03]) defined by such a system of inequalities $A \mathbf{x} \leq \mathbf{b}$ is a point $\mathbf{v}$ of the polytope such that $A\mathbf{v} = \mathbf{b}$. Because $A$ is of full rank (because the defined polytope is nonempty) and the entries of $A$ and $\mathbf{b}$ are all integers, Cramer’s rule implies that all the vertices of the polytope have rational coefficients. There is then an integer $N$ for which the scaled polytope contains a $(2k,n)$-LR sun hive. The saturation theorem \[saturation-sun\] ensures that $f(\lambda(1), \ldots, \lambda(m))$ is positive, so the original polytope contains a $(2k,n)$-LR sun hive.
Determining whether the polytope is nonempty or not can be determined in polynomial time using linear programming, such as the ellipsoid or interior point algorithm. Furthermore, because the linear program $A \mathbf{x} \leq \mathbf{b}$ is combinatorial, positivity can be determined in strongly polynomial time by using the algorithm in [@Tar86].
Other generalized Littlewood-Richardson coefficients {#section-others}
=====================================================
The same techniques used in this paper can be used to prove similar results for two other generalized Littlewood-Richardson coefficients, and as such we state the results for these multiplicities without proof. The details for the calculations and proofs may be found in the author’s PhD thesis.
Context and motivation
----------------------
The two other multiplicities are $$\label{two}
f_1(\lambda(1),\ldots, \lambda(m)) :=
\sum c^{\alpha(1)}_{\lambda(1),\lambda(2)} c^{\lambda(3)}_{\alpha(1),\alpha(2)}\cdots c^{\lambda(m-2)}_{\alpha(m-4),\alpha(m-3)} c^{\alpha(m-3)}_{\lambda(m-1), \lambda(m)}$$ for $m\geq 4$, and $$\label{three}
f_2(\lambda(1),\ldots, \lambda(m)):=
\sum c^{\lambda(2)}_{\lambda(1), \alpha(1)} c^{\lambda(3)}_{\alpha(1), \alpha(2)} \cdots c^{\lambda(m-2)}_{\alpha(m-4), \alpha(m-3)} c^{\lambda(m-1)}_{\alpha(m-3), \lambda(m)},$$ for m $\geq 3$, where the case $m=3$ is the Littlewood-Richardson coefficient $c^{\lambda(2)}_{\lambda(1),\lambda(3)}$, and each summation ranges over all partitions $\alpha(i)$. The first multiplicity describes the branching rule for the direct sum embedding ${\text{GL}}(n) \times {\text{GL}}(n') \subseteq {\text{GL}}(n+n')$ when $m=6$. This was first proven in [@Kin70], and is also proven in [@HTW05] (see also [@HJ09] and [@Koi89]). The second multiplicity describes the tensor product multiplicities for extremal weight crystals of type $A_{+\infty}$, using a combinatorial rule found by Kashiwara [@Kas90] similar to the Littlewood-Richardson rule that described the irreducible components of the tensor product of irreducible representations of the quantized universal enveloping algebra of a symmetrizable Kac-Moody algebra $\mathfrak{g}$ as described in [@Kwo09] (see also [@Kwo10]) again when $m=6$. This generalized multiplicity is described in [@Chi08] and [@Chi09], and is found to have connections with long exact sequences of finite, abelian $p$-groups, parabolic affine Kazhdan-Lusztig polynomials, and decomposition numbers for $q$-Schur algebras.
Statement of results
--------------------
The quiver representing multiplicity (\[two\]) is the generalized star quiver used to describe multiplicity (\[three\]) in [@Chi08], except that the first two flags are oriented in the same direction and likewise the last two are going the same direction. The dimension vector $\beta$ is defined like before as $\beta(j,i) = j$ and the weight is defined similar to $\sigma_1$ in equation (\[weight-sigma\_1\]). Similar calculations to the ones in Section \[section-saturation\] prove the following.
Let $\lambda(1),\ldots, \lambda(m)$, $m \geq 4$, be weakly decreasing sequences of $n$ integers. For every integer $r \geq 1$, $$f_1(r\lambda(1),\ldots, r\lambda(m)) \neq 0 \Longleftrightarrow f_1(\lambda(1),\ldots, \lambda(m)) \neq 0.$$
The saturation for multiplicity (\[three\]) is Theorem 1.4 of [@Chi08]. In the same paper, Chindris provides the Horn-type inequalities and a generalization of Horn’s conjecture for the second multiplicity in Theorem 1.6.
The corresponding inequalities and a generalization of Horn’s conjecture for the first multiplicity above are slightly complicated by the fact that they depend on the parity of $m$. We state below only the results for $m$ even and remark that similar results hold true for $m$ odd.
**Generalized eigenvalue problem for $f_1$.** For which weakly decreasing sequences $\lambda(1),\ldots, \lambda(2k),\, k\geq 2$, of $n$ real numbers do there exist $n \times n$ complex Hermitian matrices $H(1),\ldots, H(2k)$ with eigenvalues $\lambda(1),\ldots, \lambda(2k)$ such that $${\displaystyle}H(1) + \sum_{i=1}^{k-1} H(2i) = H(2k) + \sum_{i=1}^{k-1} H(2i+1)$$ and such that $$\begin{array}{rcl}
{\displaystyle}H(1) + H(2), & {\displaystyle}(-1)^j(H(1)+H(2)) + \sum_{i=3}^j (-1)^{j+i} H(i) & 3 \leq j \leq 2k-2
\end{array}$$ have non-negative eigenvalues?
We define the generalized Klyachko’s cone for this multiplicity as the rational convex polyhedral cone of $m$-tuples $(\lambda(1),\ldots, \lambda(m))$ of solutions to this generalized eigenvalue problem. We denote this cone as $K_1(n,m) \subseteq {\mathbb{R}}^{nm}$.
For subsets $I_i \subseteq \{1,\ldots, n\}, \, 1 \leq i \leq 2k$, define the following tuple of weakly decreasing sequences of integers: $${\underline{\lambda}}_1(I_i) = \begin{cases}
\lambda'(I_i) & 1 \leq i \leq 2k-3, \; i \text{ odd} \\
\lambda'(I_2) - ((|I_2| - |I_3|)^{n-|I_2|}) & i=2\\
\lambda'(I_i) - ((|I_i| - |I_{i-1}| - |I_{i+1}|)^{n-|I_i|}) & 4 \leq i \leq 2k-2, \; i \text{ even} \\
\lambda'(I_{2k-1}) - ((|I_{2k-1}| - |I_{2k-1} \backslash \{n\} - |I_{2k} \backslash \{n\}|)^{n-|I_{2k-1}|}) & i=2k-1 \\
\lambda'(I_{2k} \backslash \{n\}) & i=2k.
\end{cases}$$
The generalization of Horn’s conjecture for multiplicity (\[two\]) may then be stated as follows (for $m$ even).
Let $\lambda(1),\ldots, \lambda(m),\, m=2k \geq 4$, be weakly decreasing sequences of $n$ real numbers. Then the following conditions are equivalent:
1. $(\lambda(1),\ldots, \lambda(m)) \in K_1(n,m)$;
2. the numbers $\lambda(i)_j$ satisfy $$\sum_{i=1}^{k-1} |\lambda(2i+1)| + |\lambda(2k)| =
|\lambda(1)| + \sum_{i=1}^{k-1} |\lambda(2i)|$$ and $$\sum_{j \in I_i} \sum_{i=1}^{k-1} \lambda(2i+1)_j+ \sum_{j \in I_{2k}} \lambda(2k)_j \leq
\sum_{j \in I_1} \lambda(1)_j + \sum_{j \in I_i} \sum_{i=1}^{k-1} \lambda(2i)_j$$ for every tuple $(I_1, \ldots, I_{m})$ for which $|I_i|<n$ for some $i$, $|I_1| = |I_2|$ and $|I_{m-1}| = |I_m|$, ${\underline{\lambda}}_1(I_i)$ are partitions, $1 \leq i \leq m$, and $$f_1({\underline{\lambda}}_1(I_1),\ldots, {\underline{\lambda}}_1(I_m)) \neq 0.$$ If the $\lambda(i)$ are sequences of integers, conditions $(1)-(2)$ are equivalent to
3. $f_1(\lambda(1),\ldots, \lambda(m)) \neq 0$.
Note that in this case we get an equivalence of the tuples that satisfy the generalized eigenvalue problem and those that satisfy the Horn-type inequalities, as opposed to the case for the sun quiver. This again provides a recursive method for finding all nonzero generalized Littlewood-Richardson coefficients of this type. Like before, the list may be shortened.
The following statements are true for $m=2k \geq 4$:
1. We have $\operatorname{dim}K_1(n,m) = mn-1.$
2. The cone $K_1(n,m)$ consists of all tuples $(\lambda(1),\ldots, \lambda(m))$ of weakly decreasing sequences of $n$ reals such that $$\sum_{i=1}^{k-1} |\lambda(2i+1)| + |\lambda(2k)| =
|\lambda(1)| + \sum_{i=1}^{k-1} |\lambda(2i)|$$ and $$\sum_{j \in I_i} \sum_{i=1}^{k-1} \lambda(2i+1)_j+ \sum_{j \in I_{2k}} \lambda(2k)_j \leq
\sum_{j \in I_1} \lambda(1)_j + \sum_{j \in I_i} \sum_{i=1}^{k-1} \lambda(2i)_j$$ for every tuple $(I_1, \ldots, I_{m})$ for which $|I_i|<n$ for some $i$, $|I_1| = |I_2|$ and $|I_{m-1}| = |I_m|$, ${\underline{\lambda}}_1(I_i)$ are partitions, $1 \leq i \leq m$, and $$f_1({\underline{\lambda}}_1(I_1),\ldots, {\underline{\lambda}}_1(I_m)) = 1.$$
Factorization formulas can be found for these multiplicities by using these Horn-type inequalities. Likewise, we can define level-1 weights for the corresponding quivers for (\[two\]) and (\[three\]), yet these prove to be less interesting than for (\[one\]): if a level-1 weight is effective in either case, then the Littlewood-Richardson polynomial has value one and stretching it by a factor of $N$ does not change this. This at least shows that Fulton’s conjecture remains true for these multiplicities.
A similar method of gluing LR hives together produces a polytopal description of these multiplicities. In fact, the two have the same generalized LR hives (differing in shape depending on the parity of $m$), except the boundary labels for (\[two\]) use the related partitions $\widetilde{\lambda(i)}$ for $i=1,2,m-1,m$. Because of this polytopal description and the saturation properties, we again can determine the complexity of their positivity.
Determining whether multiplicities *(\[two\])* and *(\[three\])* are positive or not can be decided in polynomial time. Even more, each can be decided in strongly polynomial time in the sense of [@Tar86].
Appendix
========
The following are the dimension vectors $\beta_1$ that correspond to a complete and minimal list of Schur roots $\beta_1$ and $\beta_2$ such that $\beta = \beta_1 + \beta_2$ and $\beta_1 \circ \beta_2 = 1$ in the case $(n,m) =(2,6)$, up to permutations of the flags respecting the symmetries of the sun quiver.
$$\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 1 \arrow[dl] \\
&& 1\arrow[ul] & 1 \arrow[l] \arrow[dr] \\
0\arrow[r] & 0 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 1 \\
&& 0 \arrow[dl] & 0 \arrow[l] \arrow[ur] \\
&0&&& \arrow[ul] 0
\end{tikzcd} \hspace{2in}
\begin{tikzcd}[column sep=small, row sep=small]
&0&&& 0 \arrow[dl] \\
&& 1 \arrow[ul] & 1\arrow[l] \arrow[dr] \\
0\arrow[r] & 0 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 1 \\
&& 0 \arrow[dl] & 0 \arrow[l] \arrow[ur] \\
&0&&& \arrow[ul] 0
\end{tikzcd}$$
$$\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 1 \arrow[dl] \\
&& 1\arrow[ul] & 1 \arrow[l] \arrow[dr] \\
0\arrow[r] & 1 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 1 \\
&& 1 \arrow[dl] & 0 \arrow[l] \arrow[ur] \\
&1&&& \arrow[ul] 0
\end{tikzcd} \hspace{2in}
\begin{tikzcd}[column sep=small, row sep=small]
&0&&& 0 \arrow[dl] \\
&& 1 \arrow[ul] & 1\arrow[l] \arrow[dr] \\
0\arrow[r] & 1 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 1 \\
&& 1 \arrow[dl] & 0 \arrow[l] \arrow[ur] \\
&1&&& \arrow[ul] 0
\end{tikzcd}$$
$$\begin{tikzcd}[column sep=small, row sep=small]
&0&&& 0 \arrow[dl] \\
&& 1\arrow[ul] & 1 \arrow[l] \arrow[dr] \\
0\arrow[r] & 0 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 1 \\
&& 1 \arrow[dl] & 1 \arrow[l] \arrow[ur] \\
&1&&& \arrow[ul] 0
\end{tikzcd} \hspace{2in}
\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 1 \arrow[dl] \\
&& 2 \arrow[ul] & 2\arrow[l] \arrow[dr] \\
0\arrow[r] & 0 \arrow[ur] \arrow[dr]&&& 2 \arrow[r]& 1 \\
&& 0 \arrow[dl] & 0 \arrow[l] \arrow[ur] \\
&0&&& \arrow[ul] 0
\end{tikzcd}$$
$$\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 1 \arrow[dl] \\
&& 2\arrow[ul] & 1 \arrow[l] \arrow[dr] \\
1\arrow[r] & 1 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 1 \\
&& 1 \arrow[dl] & 0 \arrow[l] \arrow[ur] \\
&1&&& \arrow[ul] 0
\end{tikzcd} \hspace{2in}
\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 1 \arrow[dl] \\
&& 2 \arrow[ul] & 1\arrow[l] \arrow[dr] \\
0\arrow[r] & 1 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 1 \\
&& 1 \arrow[dl] & 0 \arrow[l] \arrow[ur] \\
&0&&& \arrow[ul] 0
\end{tikzcd}$$
$$\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 0 \arrow[dl] \\
&& 2 \arrow[ul] & 1 \arrow[l] \arrow[dr] \\
0\arrow[r] & 1 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 0 \\
&& 1 \arrow[dl] & 0 \arrow[l] \arrow[ur] \\
&0&&& \arrow[ul] 0
\end{tikzcd} \hspace{2in}
\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 1 \arrow[dl] \\
&& 2 \arrow[ul] & 1\arrow[l] \arrow[dr] \\
0\arrow[r] & 1 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 1 \\
&& 1 \arrow[dl] & 1 \arrow[l] \arrow[ur] \\
&1&&& \arrow[ul] 0
\end{tikzcd}$$
$$\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 1 \arrow[dl] \\
&& 1\arrow[ul] & 1 \arrow[l] \arrow[dr] \\
0\arrow[r] & 1 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 1 \\
&& 2 \arrow[dl] & 1 \arrow[l] \arrow[ur] \\
&1&&& \arrow[ul] 0
\end{tikzcd} \hspace{2in}
\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 0 \arrow[dl] \\
&& 2 \arrow[ul] & 1\arrow[l] \arrow[dr] \\
0\arrow[r] & 1 \arrow[ur] \arrow[dr]&&& 1 \arrow[r]& 1 \\
&& 1 \arrow[dl] & 1 \arrow[l] \arrow[ur] \\
&0&&& \arrow[ul] 0
\end{tikzcd}$$
$$\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 1 \arrow[dl] \\
&& 2\arrow[ul] & 2 \arrow[l] \arrow[dr] \\
1\arrow[r] & 1 \arrow[ur] \arrow[dr]&&& 2 \arrow[r]& 1 \\
&& 1 \arrow[dl] & 1 \arrow[l] \arrow[ur] \\
&1&&& \arrow[ul] 0
\end{tikzcd} \hspace{2in}
\begin{tikzcd}[column sep=small, row sep=small]
&1&&& 1 \arrow[dl] \\
&& 2 \arrow[ul] & 2\arrow[l] \arrow[dr] \\
0\arrow[r] & 1 \arrow[ur] \arrow[dr]&&& 2 \arrow[r]& 1 \\
&& 1 \arrow[dl] & 1 \arrow[l] \arrow[ur] \\
&0&&& \arrow[ul] 0
\end{tikzcd}$$
Acknowledgement {#acknowledgement .unnumbered}
===============
The author would like to thank Calin Chindris for his support during the duration of this project, especially suggesting the problem and the many helpful discussions.
The author is also extremely grateful to Velleda Baldoni, Michèle Vergne, and Michael Walter for their interest in this paper and for pointing out mistakes in the calculations in Example \[n=2\] in an earlier draft. They corroborated the inequalities as they are now stated using the techniques developed in [@BVW18].
[10]{}
S. Agnihotri and C. Woodward. Eigenvalues of products of unitary matrices and quantum Schubert calculus. , 5(9): 817–836 (1998).
P. Belkale. Local systems on $\mathbb{P}^1 \backslash S$ for $S$ a finite set. 129(1): 67–86 (2001).
P. Belkale. Geometric proofs of Horn and saturation conjectures. 15(1): 133–173 (2006).
P. Belkale. . , 216: 346–357 (2007).
V. Baldoni, M. Vergne, M. Walter. . Preprint, <https://arxiv.org/pdf/1804.00431.pdf> (2018).
A. Buch. The saturation conjecture (after A. Knutson and T. Tao). 46(2), no. 1–2: 46–60 (2000).
P. Bürgisser, C. Ikenmeyer. A max-flow algorithm for positivity of Littlewood-Richardson coefficients. , 267–278 (2009).
P. Bürgisser, C. Ikenmeyer. Deciding positivity of Littlewood-Richardson coefficients. , 27(4): 1639–1681 (2013).
C. Chindris. Quivers, long exact sequences, and Horn type inequalities. , 320(1): 128–157 (2008).
C. Chindris. Quivers, long exact sequences, and Horn type inequalities II. , 51(2): 201–217 (2009).
W. Crawley-Boevey, Ch. Geiss. Horn’s problem and semi-stability for quiver representations. , I, II: 40–-48, 2002.
H. Derksen, J. Weyman. Semi-invariants of quivers and saturation for Littlewood-Richardson coefficients. , 13(3): 467–479 (2000).
H. Derksen, J. Weyman. On the $\sigma$-stable decomposition of quiver representations. Preprint, <http://www.math.lsa.umich.edu/~hderksen/preprints/stab.pdf> (2000).
H. Derksen, J. Weyman. On Littlewood-Richardson polynomials. , 255(2): 247–257 (2002).
H. Derksen, A. Schofield, J. Weyman. On the number of subrepresentations of a general quiver representation. , 76(2): 135–147 (2007).
H. Derksen, J. Weyman. The combinatorics of quiver representations. , 61(3): 1061–1131 (2011).
J. Fei. Cluster algebras and semi-invariant rings: triple flags. Preprint, <https://arxiv.org/pdf/1411.4693.pdf> (2015).
W. Fulton. Young tableaux. With applications to representation theory and geometry. , volume 35 (1997).
W. Fulton. Eigenvalues of sums of Hermitian matrices. , 845: 255–269 (1997–98)
W. Fulton. Eigenvalues of majorized Hermitian matrices and Littlewood-Richardson coefficients. , 319(1–3): 23–36, 2000.
M. Grotschel, L. Lovasz, A. Schrijver. Geometric algorithms and combinatorial optimization. Algorithms and Combinatorics, 2. [*Springer-Verlag*]{} (1993).
R. Howe, S. Jackson, et. al. Toric degeneration for branching algebras. , 220(6): 1809–1841 (2009).
A. Horn. Eigenvalues of sums of Hermitian matrices. , 12: 225–241 (1962).
R. Howe, E. Tan, J. Willenbring. Stable branching rules for classical symmetric pairs. , 357(4): 1601–1626 (2005).
C. Ikenmeyer. Small Littlewood-Richardson coefficients. , 44(1): 1–-29 (2016).
V. Kac. Infinite root systems, representations of graphs and invariant theory II. , 78(1): 141–162 (1982).
M. Kashiwara. Crystalizing the $q$-analogue of universal enveloping algebras. , 133(2): 249-–260 (1990).
R. King. Generalized Young tableaux and the general linear group. 11: 280–293 (1970).
R. King. Modification rules and products of irreducible representations for the unitary, orthogonal, and symplectic groups. 12: 1588–1598 (1971).
A. D. King. Moduli of representations of finite-dimensional algebras. , 45(180): 515–530 (1994).
T. Klein. The multiplication of Schur-functions and extension of $p$-modules. , 43: 280–284 (1968).
A. Klyachko. Stable bundles, representation theory and Hermitian operators. , 4(3): 419–445 (1998).
K. Koike. On the decomposition of tensor products of the representations of the classical groups. 74(1): 57–86 (1989).
A. Knutson, T. Tao. The honeycomb model of $\text{gl}_n({\mathbb{C}})$ tensor products. I. Proof of the saturation conjecture. , 12(4): 1055–1090 (1999).
A. Knutson, T. Tao. Honeycombs and sums of Hermitian matrices. 48(2):175–186, 2001.
R. King, C. Tollu, F. Toumazet. Stretched Littlewood-Richardson and Kostka coefficients. 34: 99–112 (2004).
R.C. King, C. Tollu, F. Toumazet. The hive model and the factorisation of Kostka coefficients. , 54A (2006).
R.C. King, C. Tollu, F. Toumazet. The hive model and the polynomial nature of stretched Littlewood-Richardson coefficients. , 54A (2006).
A. Knutson, T. Tao, Ch. Woodward. The honeycomb model of $\text{gl}_n({\mathbb{C}})$ tensor products. II. Puzzles determine facets of the Littlewood-Richardson cone. , 17(1): 19–48 (2004).
J. Kwon. Differential operators and crystals of extremal weight modules. , 222(4): 1339–1369 (2009).
J. Kwon. Littlewood-Richardson coefficients and weight crystals. Preprint, <http://www.kurims.kyoto-u.ac.jp/~kyodo/kokyuroku/contents/pdf/1689-07.pdf> (2010).
J. De Loera, T. McAllister. On the computation of Clebsch-Gordan coefficients and the dilation effect. 15(1): 7–20 (2006).
K. Mulmuley and M. Sohoni. Geometric complexity theory, P vs. NP and explicit obstructions. , 239–261 (2001).
K. Mulmuley, M. Sohoni. Geometric complexity theory: an approach to the P vs. NP and related problems. , 31(2): 496–-526 (2001).
K. Mulmuley, M. Sohoni. Geometric complexity theory III, on deciding positivity of Littlewood-Richardson coefficients. Preprint, <https://arxiv.org/pdf/cs/0501076.pdf> (2005)
K. Mulmuley, M. Sohoni. Geometric complexity theory: Introduction. Lecture notes, <https://arxiv.org/pdf/0709.0746.pdf> (2007).
K. Mulmuley, M. Sohoni. Geometric complexity theory II: Towards explicit obstructions for embeddings among class varieties. 38(3): 1175–1206 (2008).
K. Mulmuley, M. Sohoni. Geometric complexity theory V: on deciding nonvanishing of a generalized Littlewood-Richardson coefficient. 30(1): 225–309 (2017).
K. Mulmuley, H. Narayanan, M. Sohoni. Geometric complexity theory III: on deciding nonvanishing of a Littlewood-Richardson coefficient. 36(1): 103–110 (2012).
K. Mulmuley. Explicit Proofs and The Flip. Preprint <https://arxiv.org/pdf/arXiv:1009.0246> (2010).
K. Mulmuley. Geometric complexity theory VI: The flip via positivity. Technical report, The Computer Science Department, The University of Chicago (2011).
H. Narayanan. The computation of Kostka numbers and Littlewood-Richardson coefficients is \#P-complete. Preprint, <https://arxiv.org/pdf/math/0501176.pdf> (2005).
H. Narayanan. On the complexity of computing Kostka numbers and Littlewood-Richardson coefficients. , 24(3): 347–354 (2006).
E. Rassart. A polynomiality property for Littlewood-Richardson coefficients. , 107(2): 161–179 (2004).
A. Schofield, M. van den Bergh. Semi-invariants of quivers for arbitrary dimension vectors. , 12(1): 125–138 (2001).
A. Schofield. Semi-invariants of quivers. , 43(3): 385–395 (1991).
A. Schofield. General representations of quivers. , 65(1): 46–64 (1992).
A. Schrijver. Combinatorial Optimization. Polyhedra and efficiency. (2003).
C. Sherman. Geometric proof of a conjecture of King, Tollu, and Toumazet. Preprint, <https://arxiv.org/pdf/1505.06551.pdf> (2015).
C. Sherman. Quiver generalization of a conjecture of King, Tollu, and Toumazet. , 480: 487–504 (2017).
É. Tardos. A strongly polynomial algorithm to solve combinatorial linear programs. , 34(2): 250–-256 (1986).
H. Weyl Das asymtotische Verteilungsgesetz der Eigenwerte lineare partieller Differential gleichungen. , 71: 441–479 (1912).
A. Zelevinsky. Littlewood-Richardson semigroups. 38: 337–345 (1999).
[^1]: The author was partially supported by the NSA under grant H98230-15-1-0022.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Though it is well known that the roots of any affine polynomial over a finite field can be computed by a system of linear equations by using a normal base of the field, such solving approach appears to be difficult to apply when the field is fairly large. Thus, it may be of great interest to find an explicit representation of the solutions independently of the field base. This was previously done only for quadratic equations over binary finite field. This paper gives an explicit representation of solutions for a much wider class of affine polynomials over a binary prime field.
**Keywords:** Linear equation $\cdot$ Binary finite field $\cdot$ Base of field $\cdot$ Zeros of polynomials $\cdot$ Irreducible polynomials.
author:
- Kwang Ho Kim
- Jong Hyok Choe
- Dok Nam Lee
- Dae Song Go
- Sihem Mesnager
title: 'Solutions of $x^{q^k}+\cdots+x^{q}+x=a$ in ${{\mathbb F}_{2^n}}$'
---
Introduction
============
Define $$T_l^k(x):=x+x^{2^l}+\cdots+x^{2^{l(k/l-2)}}+x^{2^{l(k/l-1)}}$$ when $l|k$, and in particular $$T_k(x):=T_1^k(x)=x+x^2+\cdots+x^{2^{k-2}}+x^{2^{k-1}}.$$ The degree of the polynomial $T_l^k(x)$ is $2^{k-l}$.
This paper gives the explicit representations of all $\overline{{{\mathbb F}_{2}}}-$ and ${{\mathbb F}_{2^n}}-$solutions to the affine equation $$T_l^{k}(x)=a, a\in {{\mathbb F}_{2^n}}.$$ Obviously, this equation has no multiple roots since $(T_l^k)'=1\neq 0$. Throughout this paper, we set $d=\gcd(n,k)$.
To the best of our knowledge, following is the only previous result in this direction.
\[quadratic\_old\](Page 26 of [@BSS1999], 11.1.120 of [@MP2013]) The quadratic equation $$x^2+x+a=0, a\in {{\mathbb F}_{2^n}}$$ has solutions in ${{\mathbb F}_{2^n}}$ if and only if $$T_n(a)=0.$$ Let us assume $T_n(a)=0$. Let $\delta$ be an element in ${{\mathbb F}_{2^n}}$ such that $T_n(\delta)=1$ (if $n$ is odd, then one can take $\delta=1$). Then, $$x_0=\sum_{i=0}^{n-2}(\sum_{j=i+1}^{n-1}\delta^{2^j})a^{2^i}$$ is a solution to the equation.
Some useful facts
=================
\[lem\_properties\] For any positive integers $k, k', l, l'$ such that $s|l|k$ and $l'|k'$, followings hold.
1. (Commutativity) $$T_l^k\circ T_{l'}^{k'}=T_{l'}^{k'}\circ T_l^k.$$
2. (Transitivity) $$T_l^k\circ T_s^l=T_s^k.$$
3. $$T_k\circ T_2(x)=T_k^{2k}(x)=x+x^{2^k}.$$
4. $$T_k\circ T_k\circ T_2 =T_{2k}.$$
All statements can be easily checked by direct calculation.
\[lem\_fielddef\] For any positive integers $n$ and $k$, it holds $$T_k(x)\in {{\mathbb F}_{2^n}}\Longleftrightarrow T_n(x)\in {{\mathbb F}_{2^k}}.$$ In particular, letting $k=1$, we have $$x\in {{\mathbb F}_{2^n}} \Longleftrightarrow T_n(x)\in {{\mathbb F}_{2}}
\Longleftrightarrow T_n\circ T_2(x)=0.$$
Since $T_k(x)+T_k(x)^{2^n}=T_n(x)+T_n(x)^{2^k}$ which is checked by direct computation, it follows $T_k(x)\in
{{\mathbb F}_{2^n}}\Longleftrightarrow T_k(x)^{2^n}=T_k(x) \Longleftrightarrow
T_k(x)+T_k(x)^{2^n}=0 \Longleftrightarrow
T_n(x)+T_n(x)^{2^k}=0\Longleftrightarrow T_n(x)\in {{\mathbb F}_{2^k}}$.
Following fact, though already well-known, can be reformulated.
\[T\_image\] Let $l$ be a divisor of $k$. Then, $$T_l^k({{\mathbb F}_{2^k}})={{\mathbb F}_{2^l}}.$$
This follows from the fact that $T_l(
T_l^k({{\mathbb F}_{2^k}}))=T_k({{\mathbb F}_{2^k}})={{\mathbb F}_{2}}$.
\[gcd\_1\] Let us assume $\gcd(n,k)=1$. Then it holds $$T_k(x)\in {{\mathbb F}_{2^n}}\Longleftrightarrow x\in {{\mathbb F}_{2^n}}+{{\mathbb F}_{2^k}},$$ where ${{\mathbb F}_{2^n}}+{{\mathbb F}_{2^k}}=\{a+b\,|\,a\in{{\mathbb F}_{2^n}}, b\in{{\mathbb F}_{2^k}}\}$.
$(\Longleftarrow)$\
Let $x=a+b$ for $a\in{{\mathbb F}_{2^n}}$ and $b\in{{\mathbb F}_{2^k}}$. Then $a^{2^n}=a$, and $T_k(b)\in {{\mathbb F}_{2}}$ by above proposition, thus $$\begin{aligned}
T_k(a+b)^{2^n}&=(T_k(a)+T_k(b))^{2^n}\\
&=T_k(a)^{2^n}+T_k(b)^{2^n}\\
&=T_k(a^{2^n})+T_k(b)\\
&=T_k(a)+T_k(b)=T_k(a+b),\end{aligned}$$ where the linearity of $T_k$ was exploited. That is $T_k(x)=T_k(a+b)\in{{\mathbb F}_{2^n}}$.
$(\Longrightarrow)$\
Since the necessity in the statement has been proved, in order to prove the sufficiency in the statement, it is enough to show $$\#\{x\in \overline{{{\mathbb F}_{2}}}\,|\,T_k(x)\in
{{\mathbb F}_{2^n}}\}=\#\{{{\mathbb F}_{2^n}}+{{\mathbb F}_{2^k}}\},$$ where $\overline{{{\mathbb F}_{2}}}$ is the algebraic closure of ${{\mathbb F}_{2}}$. To begin with, we have $\#\{x\in \overline{{{\mathbb F}_{2}}}\,|\,T_k(x)\in
{{\mathbb F}_{2^n}}\}=2^{n+k-1}$ because for every $a\in {{\mathbb F}_{2^n}}$ the equation $T_k(x)=a$ has $2^{k-1}$ different solutions.
On the other hand, it also holds $\#\{{{\mathbb F}_{2^n}}+{{\mathbb F}_{2^k}}\}=2^{n+k-1}$. In fact, for $a, a'\in
{{\mathbb F}_{2^n}}$ and $b, b'\in {{\mathbb F}_{2^k}}$, it holds $a+b=a'+b'$ $\iff$ $a+a'=b+b'\in
{{\mathbb F}_{2^n}}\cap{{\mathbb F}_{2^k}}={{\mathbb F}_{2}}$, i.e., ($a=a'$ and $b=b'$) or ($a=a'+1$ and $b=b'+1$). Therefore $\#\{{{\mathbb F}_{2^n}}+{{\mathbb F}_{2^k}}\}=\#\{a+b\,|\,a\in{{\mathbb F}_{2^n}}, b\in{{\mathbb F}_{2^k}}\}=(2^n\cdot2^k)/2=2^{n+k-1}$.
Without the condition $\gcd(n,k)=1$, we give:
It holds $$\{x\in \overline{{{\mathbb F}_{2}}}\,|\,T_k(x)\in {{\mathbb F}_{2^n}}\}=\{T_n^{[n,k]}\circ
T_k^{[n,k]}\circ T_2(x) \,|\,x\in {{\mathbb F}_{2^{2[n,k]}}}\}(\subset
{{\mathbb F}_{2^{2[n,k]}}}),$$ where $[n,k]$ is the least common multiple of two integers $n$ and $k$.
Let $L=[n,k]$. Let us set $y=T_n^{L}\circ T_k^{L}\circ T_2(x)$ for $x\in
{{\mathbb F}_{2^{2L}}}$. To begin with, we will show $T_n\circ T_k(y)\in
{{\mathbb F}_{2}}$ which is equivalent to $T_k(y)\in {{\mathbb F}_{2^n}}$ by Lemma \[lem\_fielddef\]. In fact, $$T_n\circ T_k(y)=T_n^{L}\circ T_n\circ T_k^{L}\circ T_k\circ
T_2(x)=T_L\circ T_L \circ T_2(x)=T_{2L}(x)\in {{\mathbb F}_{2}},$$ where the equalities are from Lemma \[lem\_fielddef\], except for the last equality which is from Lemma \[lem\_properties\].
Obviously, the cardinality of the left side set is $2^{n+k-1}$. On the other hand, the cardinality of the right side set is also $2^{n+k-1}$ as it equals $2^{2L-(L-n)-(L-k)-1}=2^{n+k-1}$ and so the two sets coincide.
By Theorem \[gcd\_1\], we know that $\{x\in
\overline{{{\mathbb F}_{2}}}\,|\,T_k(x)\in {{\mathbb F}_{2^n}}\}\subset {{\mathbb F}_{2^{nk}}}=
{{\mathbb F}_{2^{[n,k]}}}$ when $\gcd(n,k)=1$. However, it is not always the case. Let us consider the case $n=k=2$. $$\begin{aligned}
&\{x\in \overline{{{\mathbb F}_{2}}}\,|\,T_2(x)\in {{\mathbb F}_{2^2}}\}=\{x\in
\overline{{{\mathbb F}_{2}}}\,|\,x+x^2+x^4+x^8=0\}\\
&=\{x\in \overline{{{\mathbb F}_{2}}}\,|\,(x+x^2)(1+x+x^2)(1+x+x^4)=0\}.\end{aligned}$$ The least field that contains this set is ${{\mathbb F}_{2^4}}={{\mathbb F}_{2^{2[2,2]}}}$.
That is, generally, ${{\mathbb F}_{2^{2[n,k]}}}$ is the smallest field including $$\{x\in \overline{{{\mathbb F}_{2}}}\,|\,T_k(x)\in {{\mathbb F}_{2^n}}\}.$$
\[gcd\_lcm\] When $a\in {{\mathbb F}_{2^n}}$, $$T^{[n,k]}_k(a)=T_d^n(a)=T^{[n,n-k]}_{n-k}(a).$$
By definition $T^{[n,k]}_k(a)=\sum_{i=0}^{\frac{[n,k]}{k}-1}a^{2^{ik}}$, $T^{n}_d(a)=\sum_{j=0}^{\frac{n}{d}-1}a^{2^{jd}}$ and $T^{[n,n-k]}_{n-k}(a)=\sum_{i=0}^{\frac{[n,n-k]}{n-k}-1}a^{2^{i(n-k)}}$. Note that the upper bounds of indices in three summations are identical: $\frac{[n,k]}{k}=\frac{n}{d}=\frac{[n,n-k]}{n-k}$. It is easy to check $\{ik \mod n\,|\,0\leq i \leq
\frac{[n,k]}{k}-1\}=\{jd\,|\,0\leq j\leq \frac{n}{d}-1\}=\{i(n-k)
\mod n\,|\,0\leq i \leq \frac{[n,n-k]}{n-k}-1\}$. Since $a^{2^n}=a$, all three summations are identical.
Zeros of $T_l^k$
================
\[corT\_k=0\] Followings are facts.
1. $$\{x\in
\overline{{{\mathbb F}_{2}}}\,|\,T_k(x)=0\}=T_2({{\mathbb F}_{2^k}})=\{x+x^2\,|\,x\in
{{\mathbb F}_{2^k}}\}.$$
2. $$\{x\in
{{\mathbb F}_{2^n}}\,|\,T_k(x)=0\}=T_2({{\mathbb F}_{2^k}})\cap{{\mathbb F}_{2^n}}=\{x+x^2\in
{{\mathbb F}_{2^n}}\,|\,x\in {{\mathbb F}_{2^k}}\}.$$
In particular,
- If $\frac{k}{d}$ is odd, then $$\{x\in {{\mathbb F}_{2^n}}\,|\,T_k(x)=0\}=\{x\in{{\mathbb F}_{2^d}}\,|\,T_d(x)=0\}.$$
- If $\frac{k}{d}$ is even, then $$\{x\in {{\mathbb F}_{2^n}}\,|\,T_k(x)=0\}={{\mathbb F}_{2^d}}.$$
For $x\in {{\mathbb F}_{2^k}}$, by Lemma \[lem\_properties\], $T_k(T_2(x))=x+x^{2^k}=0$, and the two sets $\{x\in
\overline{{{\mathbb F}_{2}}}\,|\,T_k(x)=0\}$ and $T_2({{\mathbb F}_{2^k}})$ have the same cardinality $2^{k-1}$ and so they coincide. As a immediate consequence, we have $\{x\in
{{\mathbb F}_{2^n}}\,|\,T_k(x)=0\}=T_2({{\mathbb F}_{2^k}})\cap{{\mathbb F}_{2^n}}$.
Thus, $\{x\in
{{\mathbb F}_{2^n}}\,|\,T_k(x)=0\}=T_2({{\mathbb F}_{2^k}})\cap{{\mathbb F}_{2^n}}\subset {{\mathbb F}_{2^d}}$, and so $\{x\in {{\mathbb F}_{2^n}}\,|\,T_k(x)=0\}=\{x\in
{{\mathbb F}_{2^d}}\,|\,T_k(x)=0\}=\{x\in {{\mathbb F}_{2^d}}\,|\,\frac{k}{d}T_d(x)=0\}$, which completes the proof.
\[ker\_T\_k\_l\] Let $l$ be a divisor of $k$. Followings are facts.
1. $$\{x\in \overline{{{\mathbb F}_{2}}}\,|\,T_l^k(x)=0\}=T_l\circ
T_2({{\mathbb F}_{2^k}})=\{x+x^{2^l}\,|\,x\in {{\mathbb F}_{2^k}}\}.$$
2. $$\{x \in {{\mathbb F}_{2^n}}\,|\,T_l^k(x)=0\}=T_l\circ
T_2({{\mathbb F}_{2^k}})\cap {{\mathbb F}_{2^n}}=\{x+x^{2^l}\in {{\mathbb F}_{2^n}}\,|\,x\in
{{\mathbb F}_{2^k}}\}.$$
In particular,
- If $\frac{k}{[d,l]}$ is odd, then $$\{x\in
{{\mathbb F}_{2^n}}\,|\,T_l^k(x)=0\}=\{x\in{{\mathbb F}_{2^d}}\,|\,T_{(d,l)}^d(x)=0\}=T_{(d,l)}\circ
T_2({{\mathbb F}_{2^d}}).$$
- If $\frac{k}{[d,l]}$ is even, then $$\{x\in {{\mathbb F}_{2^n}}\,|\,T_l^k(x)=0\}={{\mathbb F}_{2^d}}.$$
Since $T_l^k(T_l\circ T_2({{\mathbb F}_{2^k}}))=T_k\circ T_2({{\mathbb F}_{2^k}})=0$, the set $T_l\circ T_2({{\mathbb F}_{2^k}})$ with cardinality $2^{k-l}$ is a subset of $\{x\in \overline{{{\mathbb F}_{2}}}\,|\,T_l^k(x)=0\}$ with the same cardinality $2^{k-l}$, i.e., the two sets coincide. Thus, $\{x \in
{{\mathbb F}_{2^n}}\,|\,T_l^k(x)=0\}=T_l\circ T_2({{\mathbb F}_{2^k}})\cap {{\mathbb F}_{2^n}}\subset
{{\mathbb F}_{2^d}}$, and we have $$\begin{aligned}
&\{x\in {{\mathbb F}_{2^n}}\,|\,T_l^k(x)=0\}=\{x\in
{{\mathbb F}_{2^d}}\,|\,T_l^k(x)=0\}\\&=\{x\in
{{\mathbb F}_{2^d}}\,|\,\frac{k}{[d,l]}T_l^{[d,l]}(x)=0\}\\&=\begin{cases}{{\mathbb F}_{2^d}},
&\text{if $\frac{k}{[d,l]}$ is
even,}\\
\{x\in{{\mathbb F}_{2^d}}\,|\,T_{(d,l)}^d(x)=0\}=T_{(d,l)}\circ T_2({{\mathbb F}_{2^d}}),
&\text{if $\frac{k}{[d,l]}$ is odd,}\end{cases}\end{aligned}$$ where Proposition \[gcd\_lcm\] was used for the last equality.
Expression of solutions in closed field
=======================================
Let $L$ be any positive integer. For any $a\in{{\mathbb F}_{2^L}}^*$ and $\xi \in \mu_{2^L+1}\setminus \{1\}$,
$$\frac{a}{\xi+1}+{{\mathbb F}_{2^L}}=\{\frac{a}{\xi'+1}\,|\,\xi' \in
\mu_{2^L+1}\setminus \{1\}\}.$$
Let $\eta \in {{\mathbb F}_{2^L}}$. Then we will show $$\frac{a}{\xi+1}+\eta=\frac{a}{\xi'+1}$$ for some $\xi' \in
\mu_{2^L+1}\setminus \{1\}$. Since $\frac{a}{\xi+1}+\eta=\frac{a}{\frac{a\xi+\eta \xi +\eta}{a+\eta \xi
+\eta}+1}$, it is enough to show $$\xi'=\frac{a\xi+\eta \xi
+\eta}{a+\eta \xi +\eta}\in \mu_{2^L+1}\setminus \{1\}.$$ In fact, obviously $\xi'\neq 1$ and $$\xi'^{2^L}=\frac{a\xi^{2^L}+\eta
\xi^{2^L} +\eta}{a+\eta \xi^{2^L} +\eta}=\frac{a/\xi+\eta/ \xi
+\eta}{a+\eta/ \xi +\eta}=1/\xi'.$$
\[T\_kneq0\] Let $a\in {{\mathbb F}_{2^n}}^*$. Let $L$ be any multiple of the least common multiple $[n,k]$ of two integers $n$ and $k$. Then, for any $\xi \in
\mu_{2^L+1}\setminus \{1\}$, $$x_0=T_k^{L}\circ T_2(\frac{a}{\xi+1})$$ is a solution to the equation $T_k(x)=a$. In fact, for any $\xi \in
\mu_{2^L+1}\setminus \{1\}$, $$\begin{aligned}
\{x\in \overline{{{\mathbb F}_{2}}}\,|\, T_k(x)=a \}&=\{T_k^{L}\circ
T_2(\frac{a}{\zeta+1})\,|\,\zeta \in \mu_{2^L+1}\setminus
\{1\}\}\\
&=T_k^{L}\circ
T_2(\frac{a}{\xi+1}+{{\mathbb F}_{2^L}}) \\
&=T_k^{L}\circ T_2(\frac{a}{\xi+1})+T_2({{\mathbb F}_{2^k}}).\end{aligned}$$
Let us set $x=T_k^{L}\circ T_2(\frac{a}{\xi+1})$ for $\xi \in
\mu_{2^L+1}$. Then, by Lemma \[lem\_properties\], one has $$\begin{aligned}
&T_k(x)=T_k^{L}\circ T_k \circ T_2(\frac{a}{\xi+1})\\
&=T_L\circ T_2(\frac{a}{\xi+1})\\
&=\frac{a}{\xi+1}+\left(\frac{a}{\xi+1}\right)^{2^L}\\
&=\frac{a}{\xi+1}+\frac{a}{\xi^{2^L}+1}\\
&=\frac{a}{\xi+1}+\frac{a}{1/\xi+1}=a.\end{aligned}$$ On the other hand, $\#\{x\in \overline{{{\mathbb F}_{2}}}\,|\, T_k(x)=a
\}=2^{k-1}$, and $\#T_k^{L}\circ
T_2(\frac{a}{\xi+1}+{{\mathbb F}_{2^L}})=\#T_k^{L}\circ
T_2({{\mathbb F}_{2^L}})=T_2({{\mathbb F}_{2^k}})=2^{k-1}$. This completes the proof.
Let $a\in {{\mathbb F}_{2^n}}$ and $l$ be a divisor of $k$. Let $L$ be any multiple of the least common multiple $[n,k]$ of two integers $n$ and $k$. Then, for any $\xi \in
\mu_{2^L+1}\setminus \{1\}$, $$x_0=T_l\circ T_k^{L}\circ T_2(\frac{a}{\xi+1})$$ is a solution to the equation $T_l^k(x)=a$. In fact, for any $\xi
\in \mu_{2^L+1}\setminus \{1\}$, $$\begin{aligned}
\{x\in \overline{{{\mathbb F}_{2}}}\,|\, T_l^k(x)=a \}&=\{T_l\circ T_k^{L}\circ
T_2(\frac{a}{\zeta+1})\,|\,\zeta \in \mu_{2^L+1}\setminus
\{1\}\}\\
&=T_l\circ T_k^{L}\circ
T_2(\frac{a}{\xi+1}+{{\mathbb F}_{2^L}})\\
&=T_l\circ T_k^{L}\circ T_2(\frac{a}{\xi+1})+T_l\circ T_2({{\mathbb F}_{2^k}}).\end{aligned}$$
Note that it is easy to take $\xi \in \mu_{2^{L}+1}\setminus \{1\}$: Choose any $s\in{{\mathbb F}_{2^{2L}}}\setminus {{\mathbb F}_{2^{L}}}$, then calculate $\xi=s^{2^{L}-1}$.
Solutions in ${{\mathbb F}_{2^n}}$
==================================
For $a\in {{\mathbb F}_{2^n}}$, the linear equation $$T_k^{2k}(x)=a \text{ (i.e. $x^{2^k}+x=a$)}$$ has solution in ${{\mathbb F}_{2^n}}$ if and only if $$T_{d}^n(a)=0.$$ When $T_{d}^n(a)=0$, this equation $T_k^{2k}(x)=a$ has exactly $2^{d}$ solutions in ${{\mathbb F}_{2^n}}$:
- If $\frac{k}{d}$ is odd, then for any $\xi\in \mu_{2^{n}+1}\setminus\{1\}$ $$\label{odd_T_k_2} \{x\in
{{\mathbb F}_{2^n}}\,|\,x^{2^k}+x=a\}=T_k^{[n,k]}(\frac{a}{\xi+1})+{{\mathbb F}_{2^d}}.$$
- If $\frac{k}{d}$ is even, then for any $\xi\in
\mu_{2^{n}+1}\setminus\{1\}$ $$\label{even_T_k_2} \{x\in
{{\mathbb F}_{2^n}}\,|\,x^{2^k}+x=a\}=T_{n-k}^{[n,n-k]}(\frac{a^{2^{n-k}}}{\xi+1})+{{\mathbb F}_{2^d}}.$$
This easily follows from the fact that the linear operator $T_k^{2k}(x)=x^{2^k}+x$ on ${{\mathbb F}_{2^n}}$ has the kernel of dimension $d$ and, thus, the number of elements in the image of $T_k^{2k}$ is $2^{n-d}$. For any $x\in {{\mathbb F}_{2^n}}$, we have $$\begin{aligned}
T_{d}^n(x^{2^k}+x) &= T_{d}^n(x^{2^k})+T_{d}^n(x)\\
&=T_{d}^n(x)^{2^k}+T_{d}^n(x)\\
&=T_{d}^n(x)+T_{d}^n(x) \text{ (since $T_{d}^n(x)\subset {{\mathbb F}_{2^{d}}}\subset {{\mathbb F}_{2^k}}$) }\\
&=0\end{aligned}$$ leading to the conclusion that the image of $T_k^{2k}$ contains such all elements in ${{\mathbb F}_{2^n}}$ since the total number of such elements in ${{\mathbb F}_{2^n}}$ is exactly $2^{n-d}$.
Let us prove the second part of the theorem. First, let us assume $\frac{k}{d}$ is odd. Then, $x_0=T_k^{[n,k]}(\frac{a}{\xi+1})$ is a solution to the equation since $T_k\circ
T_2(T_k^{[n,k]}(\frac{a}{\xi+1}))=T_{[n,k]}\circ
T_2(\frac{a}{\xi+1})=\frac{a}{\xi+1}+(\frac{a}{\xi+1})^{2^{[n,k]}}=a$. On the other hand, under the condition $T^n_d(a)=0$, this solution really belongs to ${{\mathbb F}_{2^n}}$. In fact, $x_0+x_0^{2^n}=T_k^{[n,k]}(\frac{a}{\xi+1}+(\frac{a}{\xi+1})^{2^n})=T_k^{[n,k]}(a)=T^n_d(a)=0$.
If $\frac{k}{d}$ is even, then we will consider a new equation $x^{2^{n-k}}+x=a^{2^{n-k}}$ instead of the original equation $x^{2^k}+x=a$. As obvious, this new equation shares the same ${{\mathbb F}_{2^n}}$-solution set with the original equation. Since $\frac{n-k}{d}$ is odd as $\frac{k}{d}$ is even, we can apply the solution formula for odd case to this new equation.
\[quadratic\_new\] For any $\xi \in \mu_{2^{n}+1}\setminus \{1\}$, $x_0=T_{n}(\frac{a}{\xi+1})$ and $x_0+1$ are solutions of $x^2+x+a=0$. These solutions are in ${{\mathbb F}_{2^n}}$ if and only if $$T_n(a)=0.$$
\[Solution\_T\_k\] Let $a\in {{\mathbb F}_{2^n}}$. Consider the linear equation $$\label{eqT_k}
T_k(x)=a \text{ (i.e. $x+x^2+\cdots+x^{2^{k-1}}=a$)}$$
1. Let $\frac{k}{d}$ be odd. Then, equation has a solution in ${{\mathbb F}_{2^n}}$ if and only if $$T_d^n(a)\in {{\mathbb F}_{2}}.$$ When $T_d^n(a)\in {{\mathbb F}_{2}}$, the equation $T_k(x)=a$ has exactly $2^{d-1}$ solutions in ${{\mathbb F}_{2^n}}$: for any $\xi\in \mu_{2^{n}+1}\setminus\{1\}$ $$\{x\in
{{\mathbb F}_{2^n}}\,|\,T_k(x)=a\}=T_2\circ
T_k^{[n,k]}(\frac{a}{\xi+1})+T_2({{\mathbb F}_{2^d}}).$$
2. Let $\frac{k}{d}$ be even. Then, equation has a solution in ${{\mathbb F}_{2^n}}$ if and only if $$T_d^n(a)=0.$$ When $T_d^n(a)=0$, the equation $T_k(x)=a$ has exactly $2^d$ solutions in ${{\mathbb F}_{2^n}}$: for any $\xi\in
\mu_{2^{n}+1}\setminus\{1\}$ $$\{x\in
{{\mathbb F}_{2^n}}\,|\,T_k(x)=a\}=T_2\circ
T_{n-k}^{[n,n-k]}(\frac{a^{2^{n-k}}}{\xi+1})+{{\mathbb F}_{2^d}}.$$
Since $T_d^n({{\mathbb F}_{2^n}})={{\mathbb F}_{2^d}}$ by Corollary \[T\_image\], one has $$\label{eq}
T_d^n(T_k({{\mathbb F}_{2^n}}))=T_k(T_d^n({{\mathbb F}_{2^n}}))=T_k({{\mathbb F}_{2^d}})=\begin{cases}{{\mathbb F}_{2}},
&\text{if $\frac{k}{d}$ is odd}\\ 0, &\text{if $\frac{k}{d}$ is
even}.
\end{cases}$$ Let us assume $\frac{k}{d}$ is odd. By Corollary \[corT\_k=0\], we have $\#T_k({{\mathbb F}_{2^n}})=2^{n-d+1}$. Since $\#\{x\in
\overline{{{\mathbb F}_{2}}}\,|\,T_d^n(x)\in {{\mathbb F}_{2}}\}=2^{n-d+1}$ as obvious, by we have $\{x\in \overline{{{\mathbb F}_{2}}}\,|\,T_d^n(x)\in
{{\mathbb F}_{2}}\}=T_k({{\mathbb F}_{2^n}})$, i.e. $T_k(x)=a$ has a solution in ${{\mathbb F}_{2^n}}$ if and only if $T_d^n(a)\in {{\mathbb F}_{2}}$.
At this time, let us assume $\frac{k}{d}$ is even. Then, by Corollary \[corT\_k=0\], $\#T_k({{\mathbb F}_{2^n}})=2^{n-d}$. Since $\#\{x\in \overline{{{\mathbb F}_{2}}}\,|\,T_d^n(x)=0\}=2^{n-d}$, from it follows $\{x\in
\overline{{{\mathbb F}_{2}}}\,|\,T_d^n(x)=0\}=T_k({{\mathbb F}_{2^n}})$, i.e., $T_k(x)=a$ has a solution in ${{\mathbb F}_{2^n}}$ if and only if $T_d^n(a)=0$.
The assertions about the solution number are consequences of Corollary \[corT\_k=0\]. The solution formulas were deduced from , and the fact that if $T_k\circ
T_2(x)=a$, then $y=T_2(x)$ is solution to $T_k(y)=0$.
Let $a\in {{\mathbb F}_{2^n}}$. Consider the linear equation $$\label{eqT_k_l}
T_l^k(x)=a \text{ (i.e.
$x+x^{2^l}+\cdots+x^{2^{l(\frac{k}{l}-1)}}=a$)}$$
1. Let $\frac{k}{[d,l]}$ be odd. Then, the equation has a solution in ${{\mathbb F}_{2^n}}$ if and only if $$T_d^n(a)\in {{\mathbb F}_{2^{(d,l)}}}.$$ When $T_d^n(a)\in {{\mathbb F}_{2^{(d,l)}}}$, the equation has exactly $2^{d-(d,l)}$ solutions in ${{\mathbb F}_{2^n}}$:
- If $\frac{k}{d}$ is odd, then for any $\xi\in \mu_{2^{n}+1}\setminus\{1\}$ $$\{x\in
{{\mathbb F}_{2^n}}\,|\,T_l^k(x)=a\}=T_l\circ T_2\circ
T_k^{[n,k]}(\frac{a}{\xi+1})+T_{(d,l)}\circ T_2({{\mathbb F}_{2^d}}).$$
- If $\frac{k}{d}$ is even, then for any $\xi\in \mu_{2^{d}+1}\setminus\{1\}$ $$\{x\in
{{\mathbb F}_{2^n}}\,|\,T_l^k(x)=a\}=T_l\circ T_2\circ
T_{n-k}^{[n,n-k]}(\frac{a^{2^{n-k}}}{\xi+1})+T_{(d,l)}\circ
T_2(\frac{T_d^n(a)}{\xi+1})+T_{(d,l)}\circ T_2({{\mathbb F}_{2^d}}).$$
2. Let $\frac{k}{[d,l]}$ be even. Then, the equation has a solution in ${{\mathbb F}_{2^n}}$ if and only if $$T_d^n(a)=0.$$ When $T_d^n(a)=0$, the equation has exactly $2^d$ solutions in ${{\mathbb F}_{2^n}}$: for any $\xi\in
\mu_{2^{d}+1}\setminus\{1\}$ $$\{x\in
{{\mathbb F}_{2^n}}\,|\,T_l^k(x)=a\}=T_l\circ T_2\circ
T_{n-k}^{[n,n-k]}(\frac{a^{2^{n-k}}}{\xi+1})+ {{\mathbb F}_{2^d}}.$$
It holds $$\begin{aligned}
T_d^n(T_l^k({{\mathbb F}_{2^n}}))&=T_l^k(T_d^n({{\mathbb F}_{2^n}}))\\&=T_l^k({{\mathbb F}_{2^d}})=\frac{k}{[d,l]}T_l^{[d,l]}({{\mathbb F}_{2^d}})\\&=\frac{k}{[d,l]}T_{(d,l)}^{d}({{\mathbb F}_{2^d}}) \text{ by Proposition \ref{gcd_lcm}}\\
&=\begin{cases} 0, &\text{if $\frac{k}{[d,l]}$ is even,}\\
{{\mathbb F}_{2^{(d,l)}}}, &\text{if $\frac{k}{[d,l]}$ is odd,}
\end{cases}\end{aligned}$$ and on the other hand, Corollary \[ker\_T\_k\_l\] let us know $$\begin{aligned}
\#T_l^k({{\mathbb F}_{2^n}})=\begin{cases} 2^{n-d}, &\text{if $\frac{k}{[d,l]}$ is even,}\\
2^{n-d+(d,l)}, &\text{if $\frac{k}{[d,l]}$ is odd.}
\end{cases}\end{aligned}$$ Since $T_d^n(x)=a$ has $2^{n-d}$ solutions in the closed field (indeed in ${{\mathbb F}_{2^n}}$), thus we conclude $$\begin{aligned}
T_l^k({{\mathbb F}_{2^n}})=\begin{cases} \{a\in{{\mathbb F}_{2^n}}\,|\,T_d^n(a)=0\}, &\text{if $\frac{k}{[d,l]}$ is even,}\\
\{a\in{{\mathbb F}_{2^n}}\,|\,T_d^n(a)\in {{\mathbb F}_{2^{(d,l)}}}\}, &\text{if
$\frac{k}{[d,l]}$ is odd,}
\end{cases}\end{aligned}$$ which is just the sufficient and necessary condition for existence of solution in ${{\mathbb F}_{2^n}}$.
When $\frac{k}{[d,l]}$ be odd and $\frac{k}{d}$ is even, the solution formula can be checked as follows: Consider $T_d^n(a)\in
{{\mathbb F}_{2^d}}$. First, it can be checked that $z_0=T_{(d,l)}\circ
T_2(\frac{T_d^n(a)}{\xi+1})$ is a solution in ${{\mathbb F}_{2^d}}$ of $T_l^k(z)=T_d^n(a)$, i.e. $T_{[d,l]}^l(z)=T_d^n(a)$ i.e. $T_{(d,l)}^d(z)=T_d^n(a)$. For $y_0=T_l\circ T_2\circ
T_{n-k}^{[n,n-k]}(\frac{a^{2^{n-k}}}{\xi+1})\in {{\mathbb F}_{2^n}}$, it is an easy exercise to check by direct calculation $T_l^k(y_0)=T_d^n(a)+a$. So $x_0=y_0+z_0$ is a solution in ${{\mathbb F}_{2^n}}$.
In remained cases, the solution formulas are deduced from theorem \[Solution\_T\_k\], regarding the fact that $T_l(x_0)$ is solution in ${{\mathbb F}_{2^n}}$ of $T_l^k(x)=a$ if $x_0\in {{\mathbb F}_{2^n}}$ is solution of $T_k(x)=a$.
As an immediate consequence of these facts, one has (confirms):
Followings are true.
- $T_l^k(x)$ is a 2-to-1 mapping on ${{\mathbb F}_{2^n}}$ if and only if $d=1$ and $\frac{k}{l}$ is even, or, $d=2$ and both $l$ and $\frac{k}{2l}$ are odd.
- $T_l^k(x)$ is a permutation on ${{\mathbb F}_{2^n}}$ if and only if $\frac{k}{l}$ is odd and $d|l$.
Hence, when $\frac{k}{l}$ is odd, $T_l^k(x)$ is an exceptional polynomial over ${{\mathbb F}_{2}}$.
[10]{} I. Blake, G. Seroussi, N. Smart. Elliptic Curves in Cryptography. Number 265 in London Mathematical Society Lecture Note Series. Cambridge University Press, 1999.
G.L. Mullen and D. Panario. Handbook of Finite Fields. Discrete Mathematics and Its Applications, CRC Press, 2013.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A model for the excitation of a non-linear ion-wake mode by a train of plasma electron oscillations in the non-linear time-asymmetric regime is developed using analytical theory and particle-in-cell based computational solutions. The ion-wake is shown to be a driven non-linear ion-acoustic wave in the form of a cylindrical ion-soliton. The near-void and radially-outwards propagating ion-wake channel of a few plasma skin-depth radius, is explored for application to “Crunch-in" regime of positron acceleration. The coupling from the electron wakefield mode to the ion-mode dictates the long-term evolution of the plasma and the time for its relaxation back to an equilibrium, limiting the repetition-rate of a plasma accelerator. Using an analytical model it is shown that it is the time asymmetric phases of the oscillating radial electric fields of the nearly-stationary electron bubble that excite time-averaged inertial ion motion radially. The electron compression in the back of the bubble sucks-in the ions whereas the space-charge within the bubble cavity expels them, driving a cylindrical ion-soliton structure with on-axis and bubble-edge density-spikes. Once formed, the channel-edge density-spike is sustained over the length of the plasma and driven radially outwards by the thermal pressure of the wake energy in electrons. Its channel-like structure is independent of the energy-source, electromagnetic wave or particle beam, driving the bubble electron wake. Particle-In-Cell simulations are used to study the ion-wake soliton structure, its driven propagation and its use for positron acceleration in the “Crunch-in" regime.'
author:
- 'Aakash A. Sahai'
---
Introduction
============
![ . (a) Electron bubble wakefields in cartesian coordinates (fixed-box) with $\frac{\omega_0}{\omega_{pe}}=10$ driven by a matched laser pulse (vector potential $a_0=4$ and frequency $\omega_0$) with $R_B \simeq 4\frac{c}{\omega_{pe}}$. (b) Non-linear ion-wake in the form of a cylindrical ion-soliton of radius $\simeq 4\frac{c}{\omega_{pe}}$ excited behind the bubble electron wake in a proton plasma. (c) Transverse ion-density profile at $z=15~c/\omega_{pe}$. Notice that the ion density perturbation in this excitation phase is still building up and is a fraction of the background ion density, $\frac{\delta n_i}{n_0} < 1$.[]{data-label="fig1:ion-wake-laser"}](Fig-1-laser-driven-ion-wake-soliton.jpg){width="5in"}
Plasma ions are generally assumed to be stationary in the theory of ultra-relativistic non-linear plasma electron waves [@Akhiezer-Polovin]. Such electron waves are regularly excited as wakefields of high-intensity energy sources such as an ultra-short laser or particle beams and have proved to be promising for accelerating and transporting beams with unprecedented field strengths [@wakefield][@cavitation-laser][@cavitation-beam][@Pukhov-laser-bubble][@Lu-bubble-regime]. Important exceptions to the fundamental assumption of stationary ions occur as the intensities of the energy sources become high enough to lead to significant ion motion within a period of the electron wave. Ion motion also invariably becomes important over several periods of the electron wake train further behind the driver as the energy left-over in electron oscillation modes couples to the ion modes.
The motion of ions has significant implications for plasma acceleration as high-intensity conditions occur when the drive beam (or an accelerated witness beam) has fields that lead to ion trajectories that are a considerable fraction of the electron oscillation trajectory [@ion-motion-intense-beam][@ion-motion-beam-emittance]. These conditions are predicted to arise in the final-stage of ultra-low emittance future plasma-based collider designs at the TeV energy scale. The subject of this paper however is the ion motion at longer timescales, understanding the long-term ion behavior is important to determine the state of the plasma for succeeding bunches in a high repetition rate future plasma-based collider [@long-term-wake][@hot-plasma-wake]. The work presented here shows that a long-lived ion-mode is leftover in the plasma, establishing an upper-limit on the repetition-rate of the plasma-based accelerators.
In this paper, using theoretical analysis and computational modeling, the excitation of a nonlinear ion-wake in the trail of a non-linear “bubble" plasma electron, is shown. The electron wake may be driven by either an intense laser or particle beam energy source [@cavitation-laser-expt][@cavitation-beam-expt]. We show that the time asymmetry of the focusing fields of the bubble leads to the excitation of non-linear ion-acoustic modes in the form of a cylindrical ion-soliton.
The application of the non-linear ion-wake for plasma-based accelerators in “crunch-in" wakefield regime is explored. The “crunch-in" regime of plasma wakefields in an ideal hollow-channel was introduced in [@positron-IPAC-2015][@crunch-in-regime] and Ch.8 of [@Sahai-dissertation]. In this wakefield regime, it was shown that hollow-channel is driven by energy-sources such that the channel-wall electrons collapse to the axis, driving strong wakefields of the order of channel-wall cold-plasma wave-breaking fields. Importantly, it was also shown that the focusing fields excited in this regime have a linear radial dependence of magnitude (with direction favorable for positron transport) and are of the order of accelerating fields [@positron-IPAC-2015] (Fig.2 there-in). The excitation of strong focusing fields in this regime is completely opposite to the conventionally established conclusion that relativistic particles have zero focusing fields in hollow-channels [@Katsouleas-PRL-1998]. This regime is enabled by the ion-wake channel because it is shown to have an initial radius close to the electron wake transverse size while uniquely its length is as long as the acceleration length. Here we show that the ion-wake channel-wall electrons collapse towards the energy-propagation axis resulting in a non-linear on-axis electron density compression many times the near-void background density. The optimal compression is shown to be only possible if the driving beam properties are matched to the channel radius [@positron-IPAC-2015], a strong dependence on the excitation which is a signature of non-linearity. The choice of appropriate channel radius is enabled by launching the driver at an appropriate time, resulting in excitation at an appropriate channel radius during the expansion of the ion-wake channel.
The ion-wake model shows two distinct phases of the non-linear ion-wake: inertial and thermalized phase. In the inertial or excitation phase the time-asymmetry between the attractive and repulsive radial fields of the bubble on the ions excites them into a soliton-like structure. We show that in this phase the inertial response of the ion rings is dictated by an equilibrium or separatrix radius. The ion rings located within this radius collapse towards the axis whereas rings outside are driven outwards. The outward propagating rings are only driven up to the bubble radius beyond which the force of the bubble radial fields rapidly falls off, resulting in the accumulation of the ion rings at the bubble radius.
At later times, the non-linear radial electron oscillations undergo phase mixing [@phase-mixing-longitudinal] leading to coherent electron motion becoming thermalized. The thermalized phase is shown to be a driven non-linear ion-acoustic wave in the form of a cylindrical ion-acoustic soliton. Its characteristics are similar to the solutions of the cylindrical Korteweg - de Vries equation (cKdV) [@Maxon-cyl-soliton][@cyl-soliton-observation][@Berezin-Karpman-1964][@KdV-non-linear-ion-waves]. However, the ion-wake shown here differs from a cylindrical-KdV soliton in several aspects: (a) The bubble wake electron oscillations do not thermalize into an isothermal plasma, so the ion-wake soliton is driven (or forced) by the electron temperature gradient; (b) the ion-wake soliton breaks up into N-solitons as it evolves and (c) at early times there is an ion-density spike on the axis which collapses at a later time. The soliton propagates radially outwards leaving behind a flat residue resulting in a near-void ion-wake channel.
Representative PIC simulation results in Fig.\[fig1:ion-wake-laser\] and Fig.\[fig2:ion-wake-beam\] illustrate the salient features of the non-linear ion-wake. Figure \[fig1:ion-wake-laser\] shows the excitation phase at an early time when the bubble wake-train is still executing orderly oscillations and its fields have begun to excite inertial ion motion resulting in a soliton-like ion-wake structure ($\delta n_i/n_0 \simeq 0.2$) as seen in Fig.\[fig1:ion-wake-laser\](b),(c). At later times as shown in Fig.\[fig2:ion-wake-beam\] the radial oscillations sustaining the bubble train have phase-mixed, converting much of the wave energy into electron thermal energy. The resulting electron thermal pressure drives the ion-soliton ($\delta n_i / n_0 > 1$) outwards. The time evolution of the radial dynamics of the ion rings driven by the time-asymmetric nonlinear electron is shown in a movie in supplementary material [@supplementary-movie]. We show below, it is the longitudinal or time asymmetry of the radial electron wakefields that excites the ion soliton which propagates leaving behind a near-void channel shown in the PIC simulations in Fig.\[fig1:ion-wake-laser\] and Fig.\[fig2:ion-wake-beam\].
![ . (a) Beam-driven ion-wake electron density in cylindrical coordinates (fixed-box). The beam parameters are $n_b=5n_0$, $\sigma_r=0.5c/\omega_{pe}$, $\sigma_z = 1.5 c/\omega_{pe}$, $\gamma_b=38,000$, these beam-plasma parameters are quite similar to [@cavitation-beam]. (b) Corresponding ion density in cylindrical coordinates (fixed-box). Note the N-soliton formation in the ion-density, $50c/\omega_{pe} \leq z \leq 100 c/\omega_{pe}$. The later times in the time-evolution of the ion-wake is also inferred from density snapshots farther behind the beam. (c) Radial electron and ion density profile at $z=150~c/\omega_{pe}$. A full movie of radial electron and ion density dynamics is presented in supplementary material [@supplementary-movie]. []{data-label="fig2:ion-wake-beam"}](Fig-2-beam-driven-ion-wake-soliton.jpg){width="5in"}
The paper is organized into the following sections. In section \[radial-ion-wave\] using the linearized fluid equations for ion dynamics we show the two distinct phases of the ion-wake: the excitation phase and the propagation phase. Using the fact that the thermalizing electron wakefield is non-isothermal with radial electron temperature gradients, we model the non-linear ion-acoustic waves as a driven cylindrical ion soliton. We use an analytical model based on fields of a non-linear plasma wave and simulations to demonstrate the inertial phase of the ion-wake in section \[ion-soliton-excitation-phase\]. In section \[ion-soliton-propagation-phase\] the propagation phase of the ion-wake is analytical modeled with simulations verifying the propagation of the cylindrical ion-soliton driven by the radial temperature gradient of the phase-mixed electrons. Finally, in section \[ion-soliton-positron-wake\] we introduce and analyze the properties of “crunch-in" wakefield regime in an ion-wake channel, using analytical model and simulations. In appendix \[ion-wake-model-considerations\] we present considerations and assumptions made to derive the ion-wake model.
Non-linear Ion-wake:\
as a driven Plasma Ion-wave {#radial-ion-wave}
===========================
To develop insight into the ion wake physics, we consider the 1-D simplified dispersion relation of the ion-acoustic plane waves,
$$\begin{aligned}
% plane-wave DISPERSION relation
\omega^2 = \frac{c_s^2k^2}{ 1 ~ + ~ (c_s/\omega_{pi})^2 ~ k^2 }
\label{ion-wave-dispersion-relation}\end{aligned}$$
where, $\omega_{pi}=\omega_{pe}\sqrt{m_e/m_{i}}$ and $c_s = \sqrt{\Upsilon k_BT_{wk}/m_{i}}$ under the collision-less condition, $T^i_{wk}\ll T^e_{wk}$ and $\Upsilon = 1 + 2/f$ is the adiabatic index with $f$ being the degrees of freedom of the ions.
At early times the ion motion is dominated by inertia, thus ions move over the plasma-ion timescales when driven by time varying and asymmetric fields of non-linear electron plasma-wave. As the ion inertia leads to very small spatial displacement scales $k \rightarrow \infty$, the term $k (c_s/\omega_{pi}) \gg 1$ (where $c_s/\omega_{pi} = \lambda_{De} =\sqrt{\frac{k_BT_e}{4\pi e^2 n_0}}$ is the Debye wavelength). Thus, $\omega \simeq \omega_{pi}$ and the ion-soliton density spikes grow over the plasma-ion frequency timescales, $2\pi \omega_{pi}^{-1}$. The radial electron oscillations sustaining the bubble undergo phase-mixing, the electron trajectories lose orderly motion and thermalize. As the electrons thermalize, the ion motion is driven by thermal pressure of electrons.
When the ions gain significant momentum and start oscillating over larger spatial scales in response to the electron dynamics then $k\lambda_{De} \ll 1$. In this thermally driven phase, the acoustic wave propagation becomes dispersion-less with $\omega = k c_s$.
An ion-acoustic wave growing in amplitude undergoes self-steepening, forming a density spike over much smaller spatial scales; $k$ becomes large while dispersion becomes important. The ion-wake modeled here is non-linear, thus the large $k$ dispersion relation retaining the higher-order terms in $k$ in the Taylor series expansion of eq.\[ion-wave-dispersion-relation\], is, $\omega = c_s k ~ - ~ \frac{c_s}{2} ~ \lambda_{De}^2 ~ k^3$.
Also, at much later times, the ions undergo heating; $T_i$ increases and modifies the sound speed to $c_s = \sqrt{ k_B (\Upsilon_e T_e + \Upsilon_i T_i) / m_{i}}$.
Time-scale separation of Ion-dynamics:\
a simplified driven linearized ion-fluid model
----------------------------------------------
![ (a) electron density in 2D cylindrical real-space, (b) corresponding longitudinal electric field profile and (c) corresponding radial-field profile. Here the beam is located between 170 and 180$\frac{c}{\omega_{pe}}$. The bubbles just behind the driver in Fig. 3a undergo phase-mixing over several cycles. The intermediate stages of the extent of phase-mixing can be inferred from the bubbles that are closer to the beam. The beam-plasma parameters are the same as in Fig.\[fig2:ion-wake-beam\] but the electron-wake is shown at an earlier time $t = 150\omega_{pe}$. []{data-label="fig3:bubble-train"}](Fig-3-bubble-train.jpg){width="5in"}
In this section, we derive the wave equation for the ion-wake in the linear fluid approximation driven by two terms: the electron wakefields and the electron thermal pressure. The linear ion-acoustic wave can be obtained by perturbative expansion of ion density, $n_i$ and ion fluid velocity $v_i$ in the zeroth-order ion fluid continuity equation, $n_0\mathbf{\nabla}\cdot\mathbf{v}^{(1)}_i + \frac{\partial n^{(1)}_i}{\partial t} = 0$. Taking a partial derivative with time, $\mathbf{\nabla}\cdot\frac{\partial \mathbf{v}^{(1)}_i}{\partial t} + \frac{\partial^2 }{\partial t^2} \frac{n^{(1)}_i}{n_0} = 0$. The ion-fluid equation of motion where the electron temperature ($T_e$) has a spatial gradient and electron wakefields ($\bf{E}_{wk}$) still persist is, $m_i\frac{\partial \mathbf{v}^{(1)}_i}{\partial t} = eZ_i\mathbf{E}_{wk} - \Upsilon k_BT_e\mathbf{\nabla}\frac{n^{(1)}_i}{n_0} - \Upsilon k_B\frac{n^{(1)}_i}{n_0}\mathbf{\nabla}T_e$. The assumption of spatial gradient of electron temperature has been used because electron plasma wave oscillations phase-mix into non-isothermal plasma (this is substantiated through numerical results in the simulations section in Fig.\[fig5:elec-temperature-profile\]). Upon substituting the equation of motion in the time-derivative of the linearized continuity equation, $\mathbf{\nabla}\cdot\left(\frac{eZ_i}{m_i}\mathbf{E}_{wk} - \frac{\Upsilon k_BT_e}{m_i}\mathbf{\nabla}\frac{n^{(1)}_i}{n_0} - \frac{ \Upsilon k_B}{m_i} \frac{n^{(1)}_i}{n_0}\mathbf{\nabla}T_e \right) + \frac{\partial^2 }{\partial t^2} \frac{n^{(1)}_i}{n_0} = 0$. Thus, a driven ion-acoustic wave linearized to the first-order in density perturbation has the form,
$$\begin{aligned}
% plane-wave DISPERSION relation
\nonumber \left(\frac{\partial^2 }{\partial t^2} - c_s^2\nabla^2\right) & \frac{n^{(1)}_i({\mathbf r},t)}{n_0} \\
& = - \frac{eZ_i}{m_i}\mathbf{\nabla}\cdot\mathbf{E}_{wk}({\mathbf r},t) \biggr\rvert_{\mathrm{wake}} + \frac{ \Upsilon k_B}{m_i}\frac{n^{(1)}_i}{n_0}\nabla^2T_e \big\rvert_{\mathrm{thermal}}
\label{ion-wave-equation-first-order}\end{aligned}$$
In this first-order approximate ion-fluid model the right-hand side of eq.\[ion-wave-equation-first-order\] shows two separate timescales of the ion-wake.
At earlier times, the first term on the right-hand side dominates. This is the [*formation or inertial*]{} phase of the ion-wake where the bubble electron oscillations undergo ordered radial motion and the bubble radial electric field excites the inertial response of the ions. The group velocity of the electron bubble wake ($\beta_g\approx 3 v_{th}^2/c^2$, in the 1-D limit, where, $v_{th} \simeq \sqrt{k_BT_e/m_e}$ is the mean electron thermal velocity [@Vlasov]) is much smaller than the phase velocity so the bubble fields interact with the background plasma over several oscillations. Fig.\[fig3:bubble-train\] shows the non-linear electron-wake train (electron density in real space in \[fig3:bubble-train\](a)) and its time-asymmetric fields (longitudinal \[fig3:bubble-train\](b) and radial \[fig3:bubble-train\](c)) driven by a near speed-of-light energy-source of high-intensity. The fields lead to the formation of the on-axis and $R_B$ ion density spikes. At later times after the phase-mixing between radial oscillators the electrons thermalize and ${\bf E}_{wk}({\bf r},t)\sim 0$. This is the [*propagation or thermal*]{} phase where the electron thermal pressure gradient drives the cylindrical soliton around $R_B$ radially outwards to many times $R_B$.
Eq.\[ion-wave-equation-first-order\] is not directly solved as it can be separated over the two different timescales. In the inertial or excitation phase when the plasma is cold ($T_e\simeq 0$, $c_s\simeq 0$), a better description is provided by an ion-ring model driven by the fields of the electron wave. The ion-ring model is developed and verified using PIC simulations in section \[ion-soliton-excitation-phase\]. The thermally-driven propagation phase is modeled as a driven ion-acoustic soliton and verified by PIC simulations in section \[ion-soliton-propagation-phase\].
Before proceeding to the solutions in the two time-scales, we illustrate the time-scale separation using PIC simulation snapshot over tens of electron oscillations behind the driver in Fig. \[fig3:bubble-train\]. In the Fig. \[fig3:bubble-train\] (a) the coherent motion in the first four or five oscillations behind the driver is evident; whereas further behind the driver, the electrons begin to de-cohere due to phase-mixing. The falling-off of the electron wakefields to nearly zero as the electrons thermalize is later shown using PIC simulations over a much longer time-scale in sec. \[ion-soliton-propagation-phase\] (see Fig.\[fig6:ion-motion-lineouts\](b) where the radial electric field goes to zero around 200$\omega_{pe}^{-1}$ and the ion-soliton is seen moving outwards radially in \[fig6:ion-motion-lineouts\](a)).
Excitation phase:\
Ion Inertial response to the bubble fields {#ion-soliton-excitation-phase}
==========================================
Since the characteristic time of ion-motion is much longer than the electron oscillations, the longitudinal field ${\mathbf E}_{wk}\cdot\hat{z}$ averages out over the full bubble electron oscillation. So, the ions gain relatively small net longitudinal momentum. However, atypical radial ion-dynamics arise because the radial fields, ${\mathbf E}_{wk}\cdot\hat{r}$ are asymmetric in time as shown in Fig.\[fig4:bubble-ion-dynamics\] and do not average to zero, driving an average radial ion-momentum.
Ion-ring analytical model:\
interaction with time-asymmetric bubble radial fields {#excitation-phase-model}
-----------------------------------------------------
![ . (a) electron density of a bubble in 2D cylindrical real-space. (b) longitudinal on-axis profile of the electron density (black), longitudinal field (blue), focusing field (red). (c) radial-field profile close to the back of the bubble. This is the focussing “crunch-in" phase for the ions. (d) the fields at the center of the ion-cavity of the bubble. This is the defocussing “push-out" phase for the ions. []{data-label="fig4:bubble-ion-dynamics"}](Fig-4-ion-wake-dynamics-bubble.jpg){width="5in"}
The first stage of the ion-wake formation is controlled by the different time-asymmetric phases of ion dynamics inertially responding to the bubble radial field impulses shown in Fig.\[fig4:bubble-ion-dynamics\] namely, “suck-in" due to the electron compression in the back of the bubble $F^{back}$ during $\tau_{back}$ shown in Fig.\[fig4:bubble-ion-dynamics\](c), and the “push-out" due to the mutual-ion space-charge Coulomb repulsion force $F^{sc}$ during $\tau_{cav}$ shown in Fig.\[fig4:bubble-ion-dynamics\](d). The crunch-in force is spatially-periodic at non-linear plasma wavelength, $\lambda_{Np} \approx 2R_B$ with a duty-cycle $\mathcal{D}=\frac{\tau_{back}}{\tau_{back}+\tau_{cav}}\ll 1$. In addition to the plasma wake, the propagating energy sources themselves impart impulses such as the laser ponderomotive force $F^{pm}\tau_{las}$ ($\tau_{las}$ is laser pulse duration) where $F^{pm}_{e}(r,z) = -\frac{m_ec^2}{2\gamma_e}{\bf \nabla}_{r,z}|{\bf a}(r)|^2$ ($\gamma_e$ is the plasma electron Lorentz factor) and the radial force of the drive beam $F_{b}\tau_{b}$ where $F_{b}(r)=- 2\pi e^2 n_{b} r$. The short driver impulses are neglected (below threshold intensity for direct non-linear ion excitation [@ion-motion-intense-beam][@ion-motion-beam-emittance])) because they act on the ions over their sub-wavelength short duration. This is unlike the slowly-propagating wake-plasmon bubbles that undergo continual interaction over many plasma periods. The validity of this assumption is evident from the laser ion-wake in Fig.\[fig1:ion-wake-laser\]. Since the ponderomotive force of a laser driver is an outward force for both the electrons and ions, the on-axis density-spike cannot be from this direct force from the laser. Similarly the ion-density-spike at the radial wake-edge in an electron beam driven ion-motion cannot be excited directly by the space-charge force of the beam, and is caused by the electron wake’s radial-edge density compression.
The Lagrangian fluid model of the ions in a bubble consists of ion-rings under cylindrical symmetry with $m_{i} d^2r_{i}/dt^2 = \Sigma F_{wk}$ (where $F_{wk}$ is the force of the electron wake on the ions). The bared-ion region inside the bubble is assumed to be a positively charged cylinder under steady-state approximation ($R_{B} > r_{Be}$, back of the bubble electron compression radius). The force on the ions from the non-linear electron compression $\delta n_e = n_{Be}\gg n_0$ in the back of the bubble and radius $r_{Be}$, pulls the ion rings towards the axis; and within the bubble, the mutual space-charge force of the ion-rings leads to the ion-rings being driven outwards, away from the axis. The “suck-in” force on the ions is $F^{back} = - Z_{i} 2\pi e^2 n_{Be} \frac{r^2_{Be}}{r_{i}}$. The space-charge force on the ions in the cavity is $F^{sc} = Z_{i} 2 \pi e^2 n_0 r_{i}$. The equation of motion is $m_{i}d^2r_{i}/dt^2 - \frac{c\beta_{\phi}}{\lambda_{Np}}(F^{sc}\tau_{cav} - F^{back}\tau_{back}) = 0$ using, $\omega_{pi}^2 = Z_{i} ~ 4\pi e^2 n_0 / m_{i}$, we have,
$$\begin{aligned}
% Ion-ring collapse equation - asymmetric radial force phases
\frac{d^2r_{i}}{dt^2} + \beta_{\phi} \frac{\omega^2_{pi}}{2}\left(\frac{n_{Be}}{n_0} \frac{\tau_{back}}{\tau_{cav}} \frac{r^2_{Be}}{r_{i}^2} - 1 \right) r_{i} = 0
\label{ion-ring-motion}\end{aligned}$$
where we have assumed that $c\tau_{cav}/\lambda_{Np} \simeq 1$. Therefore the ion dynamics is dictated by an equilibrium or a separatrix ion-ring radius, where the inward and the outward impulses balance out, $r_{i}^{eq}=r_{Be}\sqrt{\frac{n_{Be}}{n_0} ~ \mathcal{D} }$. The ion-rings at $r_{i} \le r_{i}^{eq}$ collapse inwards towards the axis resulting in an on-axis density spike. Whereas the ion-rings at $r_{i} \ge r_{i}^{eq}$ move out away from the axis. For $m_{i}/Z_{i}>m_{p}$ the ion-response is slower but similar.
When the radially outward moving ion-rings reach beyond $R_B$, there is excess net negative charge of the wake electrons within the bubble-sheath. As a result the radially propagating ion rings get trapped and start accumulating just inside the bubble-sheath and cannot freely move beyond, forming a density compression at $R_B$. So, the cylindrical ion soliton is formed around $R_B$. This accumulation of the moving ion-rings is shown in Fig.\[fig1:ion-wake-laser\], where it is seen that the ion and electron density start forming a peak at $R_B$.
The radial location of the excitation of the ion-soliton in the non-linear electron-wave regime is much greater than a skin depth, $c/\omega_{pe}$; thus the ion-wake starts off with a spatial-scale which is over several $c/\omega_{pe}$. This is due to the balance of opposing radial forces on the plasma electrons from the driver and the ion cavity, resulting in their radial accumulation at $R_B$ [@Pukhov-laser-bubble]. In the laser-driven case - the outward ponderomotive force is balanced by the evacuated ion-cavity: $F^{pm}_{las} = -\frac{m_ec^2}{2\gamma_e}{\bf \nabla}_r \lvert{\bf a}_0(r)\rvert^2 \simeq F_{cav} = 2\pi e^2 n_0 R_{B}$ gives $R_{B} \sim (c/\omega_{pe})^2 \frac{1}{\gamma_e} {\bf \nabla}_r \lvert{\bf a}_0(r)\rvert^2$ when simplified using ${\bf \nabla}_r|{\bf a}_0(r)|^2 \simeq a_0^2/R_{B}$ and $\gamma_e \simeq a_0$, $R_{B} \simeq \sqrt{a_0} ~ \frac{c}{\omega_{pe}}$ (computationally, $\simeq2\sqrt{a_0} ~ c/\omega_{pe}$ [@Lu-bubble-regime]). In the electron beam-driven bubble the outward force of the beam on the plasma electrons is balanced by the inward pull of the evacuated ion-cavity: $F_{b}(R_B) = 2\pi e^2 n_b r_b^2 / R_B \simeq F_{cav} = 2 \pi e^2 n_0 R_B$. This gives, $R_B \simeq \sqrt{\Lambda_b/(\pi n_0)}$, where $\Lambda_b = n_b \pi r_b^2$ is the line charge density of the beam, where $r_b$ is the beam-radius computed here as $2.3~\sigma_r$ to account for $95\%$ of beam particles for a radially Gaussian beam profile.
Bubble field time-asymmetry driven ion-soliton:\
simulation results {#excitation-phase-simulations}
------------------------------------------------
The above ion-ring model is verified using $2\frac{1}{2}D$ OSIRIS PIC simulations [@osiris-code-2002] of the ion-wake in the bubble regime by simulating various energy-sources - laser-pulses in cartesian coordinates and electron-beams in cylindrical coordinates. The laser pulse is circularly polarized with radially Gaussian and longitudinally polynomial profile (as in [@laser-pulse-profile]) with $a_0=4$ (not shown $a_0=1.0$ to $40.0$), pulse length of $30\frac{1}{\omega_0}$, matched focal spot-size radius of $40\frac{c}{\omega_0}$, and laser frequency $\omega_0=10\omega_{pe}$ (the pulse dimensions are in the FWHM of the field). The electron beam is initialized with $\gamma_b\sim38,000$, $n_b=5n_0$ (not shown $n_b=0.25n_0$ to $50n_0$) and spatial Gaussian-distribution with $\sigma_r=0.5\frac{c}{\omega_{pe}}$ and $\sigma_z=1.5\frac{c}{\omega_{pe}}$ (the beam spatial dimensions are $5\sigma$ in both the dimensions). The smallest spatial scale, $c/\omega_{pe}$ is resolved in the beam case and $c/\omega_0$ in the laser case (laser frequency $\omega_0$), with 20 cells in the longitudinal direction and 50 cells in the transverse direction. Each of the plasma grid cell has 36 particles. The beam is initialized with 64 particles per cell. The plasma is initialized in the Eulerian specification (non-moving window) and pre-ionized with $Z_i=1$. At the longitudinal boundaries we initialize vacuum space of $50 ~ c/\omega_{pe}$ followed by density ramps of 20 $c/\omega_{pe}$ sandwiching the homogeneous plasma. Absorbing boundary conditions are used for fields and particles.
The electron-beam driven ion-wake soliton structure in theory is compared to the simulations in Fig.\[fig4:bubble-ion-dynamics\](a) and \[fig3:bubble-train\](a). The observed $R_B=2.45 ~ c/\omega_{pe}$ (just behind the beam) whereas the estimated bubble radius is $R_B = \sqrt{n_b/n_0 ~ (2.3\sigma_r)^2} = 2.57 ~ c/\omega_{pe}$ ($r_b= 2.3\sigma_r=1.15c/\omega_{pe}$, the assumption $r_b\ll R_B$ is not strictly satisfied). In Fig.\[fig2:ion-wake-beam\] which is in the propagation-phase, the observed ion-wake soliton is located at $r \simeq 3.3 ~ c/\omega_{pe}$ at 460 $\omega_{pe}^{-1}$ which is about $1.7 \frac{2\pi}{\omega_{pi}}$. The ion-soliton is excited at an early time around $R_B$ and in the snapshot in Fig.\[fig2:ion-wake-beam\] it has propagated outwards. The on-axis density spike drops to a minimum at $r_{i}^{eq} \approx 0.45c/\omega_{pe}$ in Fig.\[fig2:ion-wake-beam\] whereas the estimated $r_{i}^{eq} = 0.5 ~ c/\omega_{pe}$ ($n_{Be}/n_0 \simeq 12$, $\mathcal{D} \simeq 0.1$, $r_{Be}\simeq0.5 ~ c/\omega_{pe}$). The radial ion momentum $p_r-r$ phase-space in Fig.\[fig7:radial-phase-spaces\](b) shows the ions accumulate at the axis and the channel edge, at a time corresponding to Fig.\[fig2:ion-wake-beam\](b). The ions at the channel edge are seen to have a drift velocity and a thermal spread. The radial electron momentum $p_r-r$ phase-space in Fig.\[fig7:radial-phase-spaces\](a) shows that a large density of thermalized electrons are trapped within the ion soliton which is confirmed from the density plots in Fig.\[fig2:ion-wake-beam\](a).
In the laser-driven bubble simulations the expected and observed $R_B \simeq 4 ~ c/\omega_{pe}$ as shown in Fig.\[fig1:ion-wake-laser\](a). In Fig.\[fig1:ion-wake-laser\](c) the ion-wake soliton is created at $r=4.2 ~ c/\omega_{pe}$. The expected and observed on-axis density-spike radius is $r_{i}^{eq} = 0.45 ~ c/\omega_{pe}$ ($n_{Be}/n_0 \simeq 8$, $\mathcal{D} \simeq 0.1$, $r_{Be}\simeq0.5 ~ c/\omega_{pe}$). The model for the excitation of this structure of the non-linear wake has been verified for a range of laser and beam parameters from quasi-linear to strongly non-linear electron wake regime.
Propagation phase:\
Soliton driven by electron thermal pressure gradient {#ion-soliton-propagation-phase}
====================================================
As described in section \[ion-soliton-excitation-phase\] the electron bubble-wake train fields excite a cylindrical ion soliton. Eventually, the electron oscillations phase-mix and thermalize as electron thermal energy on the time-scale of about an ion plasma period. In this section we model the propagation of the cylindrical soliton radially outwards driven by the temperature gradient as shown in eq.\[Cylindrical-KdV-equation\]. This soliton propagation is modeled using a modified cKdV equation in a non-equilibrium condition such that an electron temperature gradient sustains and drives the cylindrical ion soliton.
Thermally-driven ion-acoustic soliton:\
Analytical model {#driven-soliton-model}
---------------------------------------
In the linear regime the homogenous ion-acoustic wave equation predicts sinusoidal radial ion oscillations that support the wave. However, the linearized ion-acoustic wave equation is inadequate to describe the propagating solitary density spike at the ion-wake edge, with ion density accumulation many times the background density.
When the density in the ion perturbation begins to rise to the order of the background density, the electrostatic potential due to charge-separation between the ions and the thermal electrons correspondingly rises. This leads to [*wave-steepening*]{} due to the preferential acceleration of ions in the direction of the ion-acoustic wave velocity. When the potential of the wave is large enough the ions get trapped and co-propagate at the ion-acoustic wave phase velocity, this non-linearity is the basis of the soliton. It should be noted that the linearized kinetic theory does not formally incorporate the trapping of particles at the wave phase-velocity. In this limit the density perturbation shape is therefore not sinusoidal as the co-propagating background ions accumulate and their density perturbation takes the form of an ion-soliton. The co-propagating ion velocity in the soliton can therefore exceed the ion-acoustic phase velocity, $v_i > c_s$ and $\mathcal{M}-1 > 0$ where $\mathcal{M} = v_i/c_s$ is the Mach number. Therefore, non-linear acoustic waves are in the form of a soliton and propagate faster than the ion-acoustic velocity.
To second-order, the non-linear ion-density spike $n_{i}(r,t) > n_0$ propagation is governed by the Korteweg-de Vries (KdV) equation [@KdV-non-linear-ion-waves] which has propagating solutions of the form $\mathcal{U}(r-\mathcal{M}c_st)$ [@Berezin-Karpman-1964] where $\mathcal{U}$ is the ion-acoustic waveform, a soliton solution and $\mathcal{M}$ is the Mach number ($= v_i/c_s$) of the propagating solution. Higher-order contributions to the KdV equation have also been considered by earlier works. However, the more important and relevant here is that the standard form of the cKdV equation assumes an isothermal plasma whereas the bubble-wake phase-mixes into a plasma with a radial electron temperature gradient, whereas the ions are initially cold. In a non-isothermal plasma the effect of trapped electrons in the ion-soliton have been considered using the Bernstein-Greene-Kruskal (BGK) model at the ion-acoustic velocity [@Schamel-trapped-particle-modes].
It is also known that a single ion-soliton under the appropriate conditions can break-up into multiple solitons leading to a N-soliton solution. This is also a phenomenon we observe in the simulations shown in the ion density of the beam-driven case at $z = 60\frac{c}{ \omega_{pe} }$ in Fig.\[fig2:ion-wake-beam\].
We consider a description of the non-linear cylindrical ion-acoustic waves with a radial temperature gradient. We assume that the background electron trapping does not significantly modify the distribution function. We assume that the temperature changes slowly in vicinity of the ion soliton. This assumption is validated by the PIC simulations in Fig.\[fig5:elec-temperature-profile\].
![ . The blue curve shows the root-mean-squared radial electron momentum, $p^e_{th}(r) = \sqrt{ \left[ \Sigma_k ~ p^2_r(k,r) ~ 2\pi r ~ n_e(k,r) \right] / \Sigma_{k} ~ 2\pi r ~ n_e(k,r) }$, profile of the wake electrons corresponding to the time in Fig.\[fig2:ion-wake-beam\] at 460 $\omega_{pe}^{-1}$. This represents the square-root of the electron temperature, $p^e_{th} ~ \propto ~ \sqrt{T_e}$. The radial gradient of the temperature, $\frac{ \partial }{ \partial r } T_e$ is thus computed at the peak of the soliton (red) and in its vicinity (green). It is interesting to note that $\frac{\partial }{\partial r}T_e^{(1)}\biggr\rvert_{peak} = 0$. []{data-label="fig5:elec-temperature-profile"}](Fig-5-electron-temperature-profile.jpg){width="5in"}
To obtain the KdV equation [@Sahai-dissertation] in cylindrical coordinates with radial temperature gradient we normalize with respect to the local electron temperature, the radius: $\hat{r} = \frac{r}{\lambda_D}$, time: $\hat{t} = \omega_{pi}t = t\sqrt{\frac{4\pi e^2 n_0}{m_i}}$, electric field: $\hat{E} = \frac{e \lambda_{De} }{k_B T_e} E$, potential: $\phi = \frac{e}{k_B T_e} \Phi$, ion-density perturbation: $\hat{n}_i = n_i/n_0$, electron-density perturbation: $\hat{n}_e = n_e/n_0$, ion-fluid velocity: $\hat{v}=\frac{v_i}{c_s}$. Under this normalization the cylindrical coordinate equations transform as: electron Boltzmann distribution equation $\frac{\partial \hat{n}_e}{\partial \hat{r} } = -\hat{n_e} \hat{E} - \hat{n}_e \phi \frac{ \partial }{ \partial \hat{r} } ~ \mathrm{ln} T_e$, ion-fluid continuity equation $\frac{\partial }{\partial \hat{t} } \hat{n}_i + 2 \frac{ \hat{n}_i \hat{v}}{\hat{r}} + 2\frac{\partial}{\partial \hat{r}} \hat{n}_i \hat{v} = 0$, ion-fluid equation of motion $\frac{\partial }{\partial \hat{t} } \hat{v} + \hat{v} \frac{\partial}{\partial \hat{r}} \hat{v} = \hat{E}$ and the Poisson equation $\nabla^2\Phi = \frac{1}{\hat{r}} \frac{\partial}{\partial \hat{r}} (\hat{r}\hat{E}) = \hat{n}_i - \hat{n}_e$. The electric field $\hat{E}$ is both due to the thermal pressure and the radial fields of the wake, $\hat{E}_{wk} + \hat{E}_{th}$. But, in the following analysis the propagation of a non-linear ion-acoustic wave is considered, so we assume that the electron oscillations are thermalized and thus the effect of the fields of the wake is negligible, $\hat{E}_{wk} \rightarrow 0$.
We look for a propagating disturbance of $\hat{n}_e$, $\hat{n}_i$, $\hat{v}$ and $\hat{E}$ in a stationary background plasma with uniform background density $n_0$. We consider weakly non-linear ion-acoustic wave and expand all the wave quantities in the powers of $\delta = \mathcal{M} - 1$. We perturbatively expand $\hat{n}_i$, $\hat{n}_e$, $\hat{E}$, $\phi$, $T_e$ and $\hat{v}_i$ and retain all terms up to the order of $\delta^2$. Note that we have assumed that before the electron wake excitation the plasma is cold, $T_e^{(0)} \simeq 0$.
We transform to a moving frame of the steepened ion density perturbation using the coordinate transform $\xi=\delta^{1/2}(\hat{r} - \hat{t})$ and $\tau = \delta^{3/2}\hat{t}$. Using this, $\hat{r} = \delta^{-1/2}(\xi + \delta^{-1}\tau)$ and $\frac{\partial \xi}{\partial \tau} = \frac{\partial \xi}{\partial \hat{t} } \frac{\partial \hat{t} }{\partial \tau} = -\frac{1}{\delta}$. We renormalize the electric field as, $\tilde{E} = \delta^{-1/2}\hat{E}$. Note that in the moving frame the potential gradient is, $E = -\frac{\partial }{\partial \hat{r}} \Phi = -\delta^{1/2} \frac{\partial }{\partial \xi} \Phi $, so $\tilde{E}$ is a more appropriate quantity.
Under the assumption that in the moving-frame the quantities of the disturbance change with small $\delta = \mathcal{M} - 1$, the terms in equations are perturbatively expanded and the terms with same powers of $\delta$ are collected. From the $\delta^{1}$ order terms of all the equations above, we infer $\Phi^{(1)} = n^{(1)}_{e} = v^{(1)} = n^{(1)}_{i}\equiv \mathcal{U}$ and $\frac{\partial }{\partial \xi}\mathcal{U} = -\tilde{E}^{(1)}$.
By collecting the $\delta^{2}$ terms from the Boltzmann’s equation we obtain $\tilde{E}^{(2)} = -\frac{\partial }{\partial \xi} n^{(2)}_{e} + \mathcal{U}\frac{\partial }{\partial \xi}\mathcal{U} - \mathcal{U}\frac{\partial }{\partial \xi}T_e^{(1)}$. Similarly, collecting the $\delta^{2}$ terms from the ion-fluid equation of motion we obtain $\frac{\partial }{\partial \xi} \hat{v}^{(2)} - \frac{\partial }{\partial \xi} n^{(2)}_{e} = \frac{\partial }{\partial \tau} \mathcal{U} + \mathcal{U} \frac{\partial }{\partial \xi} T_e^{(1)}$ and from the Poisson equation we obtain $\frac{\partial^3 }{\partial \xi^3} \mathcal{U} = - \frac{\partial }{\partial \xi}(n^{(2)}_{i} - n^{(2)}_{e})$. Taking the $\delta$-order terms of the continuity equation and substituting $\mathcal{U}$ we obtain, $\mathcal{U} + \tau \left( \frac{\partial }{\partial \tau} \mathcal{U} + 2\mathcal{U}\frac{\partial }{\partial \xi}\mathcal{U} + \left[ \frac{\partial }{\partial \xi} v^{(2)} - \frac{\partial }{\partial \xi} n^{(2)}_{e} \right] \right) - \delta \left(U^2 + v^{(2)} \right) = 0$. Neglecting quantities with $\delta$ times the second-order terms and using the $\frac{\partial }{\partial \tau} \mathcal{U}$ result above, $\frac{\mathcal{U}}{\tau} + 2\frac{\partial }{\partial \tau} \mathcal{U} + 2\mathcal{U}\frac{\partial }{\partial \xi}\mathcal{U} + \left[\frac{\partial }{\partial \xi} n^{(2)}_{e} - \frac{\partial }{\partial \xi} n^{(2)}_{i}\right] = - \mathcal{U}\frac{\partial }{\partial \xi}T_e^{(1)}$.
Using the $\delta^2$ terms of the Poisson equation in the above result and using the self-similarity property of the ion-soliton, we obtain the driven Korteweg-de Vries equation in cylindrical coordinates [@Maxon-cyl-soliton] (a more detailed derivation of this modified cKdV model can be found in [@Sahai-dissertation]),
$$\begin{aligned}
%
% Korteweg-de Vries equation in Cylindrical coordinates
\nonumber & \Phi^{(1)} = n^{(1)}_{e} = v^{(1)} = n^{(1)}_{i}\equiv \mathcal{U} \\
& \frac{\mathcal{U}}{\tau} + 2\frac{\partial }{\partial \tau} \mathcal{U} + 2\mathcal{U}\frac{\partial }{\partial \xi}\mathcal{U} + \frac{\partial^3 }{\partial \xi^3} \mathcal{U} = - \mathcal{U}\frac{\partial }{\partial \xi}T_e^{(1)}
\label{Cylindrical-KdV-equation}\end{aligned}$$
It differs from the cartesian-KdV equation by the term $\frac{\mathcal{U}}{\tau}$ and the temperature-gradient based driver term $- \mathcal{U}\frac{\partial }{\partial \xi}T_e^{(1)}$. The cartesian KdV equation can be analytically solved to obtain two classes of solutions: (a) self-similar solutions which are shown in [@Berezin-Karpman-1964] to be Airy functions and (b) soliton solutions. A “soliton” is a single isolated pulse which retains its shape as it propagates at some velocity, $v_{soliton}$. This means that for a soliton-like solution the $\mathcal{U}$ only depends upon the soliton-frame variable, $\zeta = \xi - \mathcal{M}c_s \tau$ and not on space-like $\xi$ and time-like $\tau$ variables separately. The solution of the cartesian KdV equation in this co-moving frame is $\mathcal{U}(\zeta)=3 v_s ~ \mathrm{sech}^2(\sqrt{ \frac{v_s}{2} } \zeta)$ [@Berezin-Karpman-1964].
The cKdV equation and the driven cKdV equation obtained here cannot be solved analytically. However, the numerical analysis and experimental verification [@cyl-soliton-observation] of the cylindrical-KdV (cKdV) equation show that it supports functions of the form $\mathcal{U} \propto \mathrm{sech}^2(r-\mathcal{M}c_st)$ in the form of a cylindrical ion-soliton. But, the amplitude of the cylindrical soliton changes as it propagates. Here, we show that the wake electron temperature drives the ion soliton for much longer distances than possible in an isothermal plasma. The velocity of the soliton in the cylindrical case is higher than in cartesian case [@Maxon-cyl-soliton]. Since the ion-wake is excited in a non isothermal plasma its velocity changes as it is driven. The mean electron temperature reduces as the soliton propagates radially outwards because the electron thermal energy is distributed over a larger volume. The cKdV equation is also known to support an N-soliton solution, and simulations show N-soliton forming during the propagation phase. We computationally seek the dependence of the non-linear ion-density spike on $(r-\mathcal{M}c_st)$-coordinate.
It should be noted that such soliton solutions are supported under certain limiting condition on the Mach number, $\mathcal{M}$. The strict condition on the existence and stability of ion-soliton arises from a threshold limit on the magnitude of soliton potential to continue trapping the background ions.
Here we find that the speed of ion soliton is nearly equal to and only slightly higher that the ion-acoustic speed calculated using the mean temperature. As this is not an isothermal plasma, there is no well-defined ion-acoustic speed. So, the ion-acoustic wave is phase-mixed and its velocity also changes as it propagates.
The local electron temperature of the ion soliton as shown in Fig.\[fig5:elec-temperature-profile\], is used to calculate the Mach number, $\mathcal{M}$ and thus a stability criterion can be derived. This problem is represented using the condition on the Sagdeev psuedo-potential, $\mathcal{V}(\phi) = - \left( \mathrm{exp}(\phi) -1 + \mathcal{M}(\mathcal{M}^2 - 2\phi)^{1/2} - \mathcal{M}^2 \right)$ that it has to be a real number. This condition is satisfied when $\mathcal{M}^2 - 2\phi \geq 0$ therefore, $\phi < \mathcal{M}^2/2$ and $\phi_{max} = \mathcal{M}^2/2$. Using this we find the well-known condition,
$$\begin{aligned}
%
% Korteweg-de Vries equation in Cylindrical coordinates - SOLUTION EXISTENCE CRITERIA
\nonumber & 1 < \mathcal{M} < 1.6, \quad v_i < 1.6 ~ c_s \\
% Normalized ion-fluid equation - moving frame
& \Phi < \frac{\mathcal{M}_{max}^2}{2} = 1.28
\label{KdV-solution-existence}\end{aligned}$$
As will be shown later, we find from simulations that the Mach number calculated using the mean temperature is well within these bounds, and thus the soliton is stable.
Thermally-driven ion-acoustic soliton:\
simulation results {#driven-soliton-simulations}
---------------------------------------
![ . (a) electron (black) and ion (red) spike radial positions (in terms of $c\omega_{pe}^{-1}$) with time and a third-order fit (green) for the position of the ion density-spike of the soliton. (b) radial wakefields of the electron bubble oscillations (in terms of $m_e c \omega_{pe}e^{-1}$) at the electron density spike (magenta) and at the ion density spike (blue). (c) radial velocity of the ion density spike of the soliton calculated from the third-order fit curve (red). An estimate of the sound speed (green) using the mean temperature, between the axis & the soliton location (green) and in the vicinity of the soliton peak (blue), in the expression $c_s = \sqrt{k_B \langle T_e \rangle / m_i}$. Since the plasma is not isothermal the mean temperature is calculated by averaging the temperature of electrons over the indicated spatial region. (d) gradient of the electron temperature at the soliton ion density peak (blue) and in the vicinity of the peak (red). The vicinity of the ion density peak of the soliton is defined as shown in Fig.\[fig5:elec-temperature-profile\]. []{data-label="fig6:ion-motion-lineouts"}](Fig-6-soliton-time-evolution.jpg){width="5in"}
The channel-edge density spike, with a form similar to the cKdV-solution in the $r-\mathcal{M}c_st$ frame as shown in Fig.\[fig1:ion-wake-laser\] and Fig.\[fig2:ion-wake-beam\] is seen to be propagating radially outwards. The propagation phase starts around $t = 200 \omega_{pe}^{-1}$ as the radial electric fields $\mathbf{E}_{wk} \rightarrow 0$ as shown in Fig.\[fig6:ion-motion-lineouts\](b). The propagation phase is evident in Fig.\[fig6:ion-motion-lineouts\](a) where the red curve is the position of the peak of the ion-soliton in time. The radial position of the peak of the ion-soliton from each of the PIC electrons and ion density snapshot is obtained in the post-processing scripts and this is shown in Fig.\[fig6:ion-motion-lineouts\](a). The cylindrical ion-soliton has propagated from $r_{soliton}(460 \omega_{pe}^{-1}) = 3.3 c/\omega_{pe}$ (also seen in Fig.\[fig2:ion-wake-beam\]) to $r_{soliton}(1100 \omega_{pe}^{-1}) = 4.1 c/\omega_{pe}$ which corresponds to an average speed of $\langle v_{soliton} \rangle = 0.0013c$.
We compare the time-averaged soliton speed $\langle v_{soliton} \rangle$ to the average speed of sound, $c_s/c = p^{e}_{th} \sqrt{\frac{\Upsilon}{2}\frac{m_e}{m_i}}$ where the average $p^{e}_{th} \simeq 0.06$ from the electron phase-space (not shown). This gives $c_s \simeq 0.001c$ ($\Upsilon = 2$ for 2D) in agreement with the average soliton velocity. Using this time-averaged analysis we see that $\mathcal{M} \simeq 1.3$ and so the stability criteria in eq.\[KdV-solution-existence\] is satisfied.
However, as the soliton moves out the volume between the axis and the soliton edge increases. Thus the electron thermal energy re-distributes and spreads over a larger volume. This leads to the reduction in the temperature with time. The soliton is not freely propagating but is driven by the radial gradient of the electron temperature as shown in eq.\[Cylindrical-KdV-equation\]. The soliton speed thus changes in time as shown in the red curve of Fig.\[fig6:ion-motion-lineouts\](c). The sound speed also varies with time and it is estimated using the temperature at that instant using $c_s(t) = \sqrt{k_B \langle T_e(t) \rangle / m_i}$. The plasma is not in thermal equilibrium and its temperature varies radially as shown in Fig.\[fig5:elec-temperature-profile\]. The root-mean-square radial momentum is used to estimate the temperature at an instant of time, and is calculated over radial dimension from the $p_r-r$ phase-space. The mean temperature is calculated by taking the average of the rms radial momentum - over the entire channel: channel-$\langle p^e_{th} \rangle = \left[ \Sigma^{r_{sol}}_{r=0} ~ p_{th}(r) ~ 2\pi r ~ n_e(r) \right] / \Sigma^{r_{sol}}_{r=0} ~ 2\pi r ~ n_e(r) ~ \propto ~ \mathrm{channel} - \sqrt{\langle T_e \rangle }$ or in the vicinity of the soliton: soliton-$\langle p^e_{th} \rangle = \left[ \Sigma^{r_{sol}+\epsilon}_{r_{sol}-\epsilon} ~ p_{th}(r) ~ 2\pi r ~ n_e(r) \right] / \Sigma^{r_{sol}+\epsilon}_{r_{sol}-\epsilon} ~ 2\pi r ~ n_e(r) ~ \propto ~ \mathrm{soliton} - \sqrt{\langle T_e \rangle }$. The instantaneous sound speed, $c_s(t)$ computed with channel-$\langle p^e_{th} \rangle$ is in the green curve in Fig.\[fig6:ion-motion-lineouts\](c) and $c_s(t)$ computed with soliton-$\langle p^e_{th} \rangle$ is in the blue curve in Fig.\[fig6:ion-motion-lineouts\](c). The extent of the vicinity ($\epsilon$) around the soliton peak is shown in Fig.\[fig5:elec-temperature-profile\].
We compare the curves in (i) red: $v_{soliton}(t)$ (from the 3rd-order polynomial curve-fit of the radial position of the ion-density peak as a function of time), (ii) green: $c_s(t)$ from channel-$\langle p^e_{th}(t) \rangle$ and (iii) blue: $c_s(t)$ from soliton-$\langle p^e_{th}(t) \rangle$ in Fig.\[fig6:ion-motion-lineouts\](c). From the comparison it is observed that they are in good agreement. It can be seen that the velocity of the soliton estimated using the location of the ion-density peak (red) lies between $c_s(t)$ calculated using the average temperature over the channel (green) which is the upper limit and $c_s(t)$ calculated using the average temperature over the soliton (blue) which is the lower limit. We also present the radial gradient of the electron temperature, $\frac{ \partial }{ \partial r } T_e(r,t)$ in Fig.\[fig6:ion-motion-lineouts\](d). It is interesting to note from the blue curve in Fig.\[fig6:ion-motion-lineouts\](d) that the temperature gradient at the peak of the ion-soliton is zero, $\frac{ \partial }{ \partial r } T_e(r,t) \big\rvert_{peak} = 0$. In the vicinity of the soliton peak we see that the gradient of the temperature follows the variation in the ion soliton velocity, this follows from eq.\[Cylindrical-KdV-equation\]. The vicinity of the soliton peak is shown as the green curve overlaid on the thermal momentum curve in Fig.\[fig5:elec-temperature-profile\].
In Fig.\[fig2:ion-wake-beam\](b) N-soliton formation is observed in the ion-density at around $z \simeq 60 \frac{c}{\omega_{pe}}$. The single-ion soliton is seen splitting into several solitons. The N-soliton solution can explained by the seeding of different initial momentum of the ion-rings because ion-rings driven in the “push-out" phase have a radial position dependent defocussing force acting on them, $F^{sc}(r_{i}) = Z_{i} 2 \pi e^2 n_0 r_{i}$. This is shown in Fig.\[fig4:bubble-ion-dynamics\](d). Thus the ion-rings originating at a larger radii from the axis are pushed outwards with a force of a higher magnitude and the rings originating at a smaller radii just outside $r_i^{eq}$ are pushed outwards by a smaller force. So, over a longer time the set of ion-rings with a higher initial momentum propagate radially outwards at a larger radial velocity. This break-up of a single ion-soliton into N-solitons occurs over a longer time-scale because the difference in momentum is small compared to the average momentum.
The thermal momentum, $p_{e}^{th}$ at this time is less than one-tenth of the peak wake quiver momentum. There are several reasons for the cooling, such as, transfer of the wake energy to the ions and the trapped electrons [@phase-mixing-trapping], escape of the highest energy electrons and un-trapped ions from the channel edge, energy loss to the bow-shock and the re-distribution of the energy over an expanding volume. The peak radial ion-momentum is $\simeq0.005$ which shows that not all the radially propagating ions are trapped. The un-trapped free-streaming ions at $\simeq 7c/\omega_{pe}$ can be distinguished from the ions at the channel-edge in $p_r-r$ phase-space.
![ . (a) electron $p_r-r$ radial momentum phase-space showing the accumulation of thermalized electrons within the ion-soliton. (b) ion $p_r-r$ radial momentum phase-space showing the on-axis and ion-wake edge ion accumulations. []{data-label="fig7:radial-phase-spaces"}](Fig-7-radial-phase-spaces.jpg){width="5in"}
It should be noted that the long-term stability of the on-axis ion-density spike of the non-linear ion-acoustic wave is not fully modeled here. The on-axis ion-density spike will disintegrate due to mutual Coulomb repulsion of the ions over the sub skin-depth spike radius. This effect of the collapse of the on-axis density-spike will be addressed in future work. We expect that the disintegration of the central structure to be further by azimuthal asymmetries not included in the cylindrically symmetric simulations. Earlier disintegration is seen in cartesian simulations not shown.
In summary, the ion-wake is a near-void channel with sub-skin-depth density-spikes on-axis and at the bubble-edge located at the bubble-radius, $R_{B}$ [@Pukhov-laser-bubble] of several $c/\omega_{pe}$. The ion accumulation in both the density-spikes is many times the background density, and the outre spike propagates outwards as a solitary structure at slightly above the speed of sound.
The time-scale of dissipation of ion-wake and relaxation of the plasma distribution to $v_{th}/c\sim 0$ sets an upper limit on the repetition-rate [@hot-plasma-wake] of the future plasma colliders. It is well known that the ion acoustic wave is damped by collisions and ion-wave Landau damping. The ion dynamics opens questions upon the plasma container walls and the distance needed to avoid damaging them by the significant radially outward ion-flux.
It was earlier suggested that the wakefield energy in the plasma wave could be replenished and sustained [@beam-wakefield-SLAC] by a train of energy source in order to achieve high repetition rate. However, as shown in this paper due to ion motion this is not possible.
Positron acceleration:\
“Crunch-in" regime in the Ion-wake channel {#ion-soliton-positron-wake}
==========================================
We explore the use of the ion-wake channel for positron-beam driven positron wakefield acceleration in a novel and relevant “crunch-in” regime where the channel radius is of a few $c/\omega_{pe}$ as is the case for the ion-wake channel. Such channels are also promising [@hollow-channel-chiou][@hollow-acclerator-1998][@hollow-accelerator-2013] for exciting the well-studied purely electromagnetic electron-wakefields. These pure electromagnetic fields driven in a hollow-channel have proven to have zero focusing forces when driven by relativistic particles [@hollow-acclerator-1998]. Here we show that in the Crunch-in regime driven even in an ion-wake channel, strong accelerating and focusing fields of electrostatic nature are excited by the electron rings crunching in from the channel wall.
The ion-wake enables the “Crunch-in” regime because as it slowly propagates radially outwards the channel radii scans over a variety of $c/\omega_{pe}$, while its length is the energy-source plasma interaction length. Meter-scale propagation of electron beams and few centimeter-scale propagation of laser beams in plasmas while exciting nonlinear electron-waves has been well characterized in experiments. The theoretical model presented above thus provides a mechanism to generate long channels of several skin-depth radii. As we show below, the non-linear “crunch-in” regime requires such channels to optimally match with the driving energy-source.
It is well known [@positron-accln-2001] that in a homogeneous plasma positron beam driven wakes have two major problems [@positron-IPAC-2015] - (i) The plasma electrons collapsing to the axis from different radii arrive at different times, preventing optimal compression. This is because the radial force of the positron beam driving the “crunch-in" decreases with the radii. (ii) The plasma ions located in the path of the positron beam result in a de-focussing force on it. The transport of the positron in a positron-beam driven wake is thus not ideal in a homogenous plasma and has to rely on external focusing optics ahead of the plasma. The use of hollow plasma channels with a few $c/\omega_{pe}$ is shown here to provide possible pathways to overcome these fundamental problems.
The formation of much shorter plasma channels excited by significantly different processes have been shown previously. These processes including using a collimated laser with annular profile [@milchberg-axicon-channel][@bessel-beam-2011], using a hollow capillary discharge [@hollow-channel-discharge], among others [@ting-smlwfa-channel][@ponderomotive-pdpwfa-wake-channel] [@ponderomotive-linear-wake-channel].
As the ion-wake channel is a practical realization of the hypothesized ideal hollow-channel plasma [@positron-accln-2001] of a few skin-depth channel radius, we examine its excitation by a positron-beam and possible use for positron acceleration [@positron-IPAC-2015]. In this section we analyze whether the positron-beam driven wake-fields excited in the ion-wake can be used for the acceleration and transport of a positron beam.
Non-linearly driven Ion-wake channel:\
analytical model {#driven-ion-wake-analytical}
--------------------------------------
Positron acceleration using the ion-wake channel is explored in the non-linear “crunch-in" regime of perturbed electron oscillation radii, $\delta r_e \geq r_{ch}$ under the condition that the peak beam density $n_{pb} > n_0$.
We use the analytical model of the radial electron “crunch-in" based excitation of a positron beam wake in the plasma . The equation of motion of the plasma electron rings at $r$ from the axis, under the positron beam crunch-in force but neglecting the space-charge force of the collapsing electron rings is: $\frac{\operatorname d^2}{\operatorname d\xi^2}r \propto - \frac{1}{r} n_{bp}(\xi) r_{bp}^2(\xi)$, where $\xi = c\beta_{pb}t - z$ is the space just behind the positron beam with velocity $c\beta_{pb}$ driving the collapse. This is a non-linear second-order differential equation of the form, $r'' = f(r,r',\xi)$ where $f$ is not linear in $r$. Under the assumption about the positron-beam properties, $n_{bp}(\xi)$ and $r_{bp}(\xi)$ being constant during the entire interaction of the positron-beam with the hollow-channel over its full length. So, upon dropping the dependence on $\xi$ the equation simplifies to its [*special case*]{} which has analytical solutions, $r'' = f(r,r')$. The solution to this equation is [@positron-IPAC-2015], $r_{ch}\sqrt{\pi} ~ \mathrm{erf} \left( \sqrt{\mathrm{ln}( r_{ch} /r)} \right) = -\sqrt{2\mathcal{C}} ~ \xi$, where $\mathcal{C}=\frac{1}{2\pi\beta_b^2} \frac{n_{bp}}{n_0} \pi \left(\frac{r_{bp}}{c/\omega_{pe}}\right)^2$. Therefore, the collapse time-duration is $\xi_{coll} = -r_{ch}\sqrt{\frac{\pi}{2\mathcal{C}}} $. We note that there is an anomaly that exists in our problem formulation and the solution because we have not taken into account the space-charge force of the compressing electrons as they collapse to the axis and this force balances the crunch-in force of the positron beam. Under these approximations the collapse time in a homogeneous plasma is [@positron-accln-2001]:
$$\begin{aligned}
% electron-ring collapse time - solution
\tau_{c}=\sqrt{\pi} \frac{ r_{ch}}{\omega_{pe} \sqrt{ n_{bp}/n_0 } r_{pb} }
\label{positron-beam-electron-collapse}\end{aligned}$$
This expression shows that the collapse-time even in a homogeneous plasma depends strongly on the properties of the beam and the radius from which the rings are collapsing in.
Also, note that we have neglected the initial expansion velocity of the channel, $dr_{ch}/dt$). For optimal compression avoiding phase-mixing, the electron rings should collapse over, $\tau_c \simeq \mathcal{D}\lambda_{Np}/c$ where $\lambda_{Np}$ is the non-linear wavelength of the positron-driven wake and $\mathcal{D}$ is the duty-cycle of compression phase. So, the optimal channel radius is $r_{ch}^{opt} \simeq 2\sqrt{\pi}\mathcal{D}\frac{\lambda_{Np}}{\lambda_{pe}}\frac{\omega_{pb}}{\omega_{pe}} r_{pb}$. The scaling of the $r_{ch}^{opt}$ with positron beam parameters is shown in [@positron-IPAC-2015].
![ . The radial profile of the normalized electron density (black) in an ion-wake channel (normalized to the maximum electron compression) at longitudinal location of the peak accelerating wakefield ($r_{pb} = 2.3c/\omega_{pe}$, $\gamma_{pb}=38000$, $n_{pb}=1.3 n_0$). The radial profile of the accelerating-wakefield and normalized focussing-wakefield potential (radial field integrated from the edge of the box to a radius).[]{data-label="fig8:positron-acc-field-radial-profiles"}](Fig-8-positron-acc-ion-wake-channel.jpg){width="5in"}
Non-linearly driven Ion-wake channel:\
simulation result {#driven-ion-wake-simulation}
--------------------------------------
Using 2-$\frac{1}{2}$D PIC simulations in a moving window we study the positron beam driven wakefields in cylindrical geometry. We compare positron acceleration in an ideal (Heaviside density function, $n_0\mathsf{H}(r-r_{ch})$) and an ion-wake channels (on-axis and channel-edge density-spike, channel minimum density of $0.1n_0$) with $r_{ch}= 2.5 ~ c/\omega_{pe}$. For non-linear wake parameters $r_{pb}=2.3 c/\omega_{pe}$, $n_{pb} = 1.3 n_0$ and $r_{ch}^{opt}\simeq 2.3 c/\omega_{pe}$ ($\mathcal{D}\frac{\lambda_{Np}}{\lambda_{pe}}=0.25$). Fig.\[fig8:positron-acc-field-radial-profiles\] shows that the peak on-axis accelerating field is $0.4~m_ec\omega_{pe}e^{-1}$ for an ideal channel and $0.2~m_ec\omega_{pe}e^{-1}$ for the ion-wake channel. Fig.\[fig8:positron-acc-field-radial-profiles\] also shows that the focussing potential (normalized to 27.6 $m_ec^2e^{-1}$) is similar and overall focussing in both cases. However, in the ion-channel the radial field is defocussing around the on-axis ion-spike.
Thus the non-linearly driven ion-wake channel is useful for accelerating and transporting positrons despite the lower accelerating and focusing fields in comparison with the ideal channel. We also note that the on-axis density spike has detrimental effect on the focussing fields near the axis. However, the on-axis density spike is unstable over longer time-scales and collapses [@Sahai-dissertation]. The cylindrical simulations used in the current work to model the ion-wake ignore any azimuthal asymmetries in the distribution of electrons and ions in the on-axis density spike. Exploring the collapse of the on-axis density spike will be addressed in future-work. Ideal channels of a few $c/\omega_{pe}$ are technologically challenging whereas the ion-wake channel of radius $r_{ch}\gtrsim R_B$ is formed behind every bubble-wake.
Conclusion
==========
In conclusion, using theory and PIC simulations we have shown the dynamics of the formation and evolution of a non-linear ion-wake excited by the well-characterized time-asymmetric electron bubble-wakefields independent of the type of energy-source. We have shown that the non-linear ion-wake has a characteristic cylindrical ion-soliton solution and evolves to an N-soliton solution over longer time as described by a driven cKdV equation. Thus over the period of persistence of the ion-soliton, a second electron bunch cannot be accelerated in the plasma. This establishes an upper limit on the repetition rate of a plasma collider. We have also shown the feasibility of using the ion-wake channel for positron acceleration in the positron-beam driven “crunch-in" regime within an experimentally relevant parameter regime.
Work supported by the US Department of Energy under DE-SC0010012 and the National Science Foundation under NSF-PHY-0936278. I acknowledge the hospitality of the Dept. of Physics at the Imperial College London and the John Adams Institute, while making corrections to the manuscript. I acknowledge the OSIRIS code [@osiris-code-2002] for PIC simulations presented here. I acknowledge support for experiments by FACET group at Stanford Linear Accelerator Laboratory and Prof. M. Downer’s group at University of Texas at Austin. I acknowledge the 256-node [*Chanakya*]{} server at Duke university.
Considerations in the Non-linear ion-wake model {#ion-wake-model-considerations}
===============================================
There are several considerations and assumptions that underlie the non-linear ion-wake model. Here we briefly describe these and seek to differentiate the ion-wake from other phenomena. Primarily, we establish that the ion-wake is a collision-less phenomena and it is significantly different from diffusion. Secondly, as the ion-wake is formed behind the high phase-velocity non-linear electron plasma waves that are excited as the wakefields of near speed of light energy sources, it is significantly different from hole-boring which occurs in a plasma where the energy source has nearly zero group velocity.
We recognize that to study the time evolution of a wake-excited plasma, for establishing the duration over which it relaxes to thermal equilibrium, both collisional and collision-less dynamics have to be considered along with the physics of recombination modes such as electron-ion recombination. However, in this work the dynamics of plasma is modeled under the collision-less approximation. Thus, diffusion is not important during the timescales over which the ion-wake is studied. We do not discuss recombination except mentioning that the “afterglow" is dominated by volume recombination while localized effects cannot be ruled out.
In order to formally establish the difference between the density wave processes that occur over collision-less timescale in contrast to the ones that start dominating under collisions, we show the assumptions made to arrive at the dynamics of diffusion. The process of diffusion is modeled with a parabolic partial differential equation which is deduced from the ion-fluid equations under the assumption that the inertial response of the ions is much faster than the collisional timescales.
The effect of collisions is introduced as a drag force, $m_in ~ \nu_{coll} ~ \langle \vec{v}_i \rangle$. The collisional drag force modifies the ion-fluid equation of motion as, $m_i n_i\frac{\operatorname d\vec{v}_i}{\operatorname dt} = mn\left(\frac{\partial \langle \vec{v}_i \rangle}{\partial t} + \langle \vec{v}_i \rangle \vec{\nabla} \cdot \langle \vec{v}_i \rangle \right) = \pm e ~ n\vec{E} - \vec{\nabla}\mathcal{P}_e - m_i n_i ~ \nu_{coll} ~ \langle \vec{v}_i \rangle $ where $\nu_{coll}$ is the average electron-ion collision frequency and is obtained from the mean free path. Diffusion of plasma is thus driven by the charge-separation field, $\vec{E}$ and the thermal pressure, $\mathcal{P}_e$ while being impeded by the collisional drag. Upon ignoring the inertia of the ions, $\frac{\partial \vec{v}_i}{\partial t} = 0$, the equation for the ion velocity by diffusion in an isothermal plasma is, $\langle \vec{v}_i \rangle = \pm \frac{e}{m ~ \nu_{coll}} ~ \vec{E} - \frac{k_BT_e}{m ~ \nu_{coll}} \frac{\vec{\nabla}n_i}{n_0}$. The characteristic parameter of diffusion is the diffusion coefficient or diffusivity $D = \frac{k_BT_e}{m ~ \nu_{coll}}$ and mobility $\mu = \frac{e}{m ~ \nu_{coll}}$ which depend upon the collision frequency. Using the gradient of the velocity in the continuity equation and ignoring the mobility $\mu$, leads to a Fick’s law diffusion equation, $\nabla^2\frac{n_i}{n_0} \propto \frac{\partial }{\partial t} \frac{n_i}{n_0}$; characteristic of a parabolic equation.
When the mobility is retained, the fluid equation is a moment of the Fokker-Planck equation which is the kinetic model of the collision-driven drift and diffusion. The diffusion equation thus cannot support wave-like solution because such solutions are characteristic of a hyperbolic partial differential equation.
The solutions of linear and non-linear diffusion equations show the evolution of density profile by diffusion and can be obtained using the self-similar formulation. The self-similar solutions show the spatial and temporal evolution of the density to be exponentially decaying. In the non-linear case, the density can have a sharp-front as it decays. However, a soliton-like propagating solution cannot be described with diffusion equation. Hence, the cylindrical ion-soliton presented here is not diffusion but a wave phenomenon.
The electron bubble wake is excited by a sub-wavelength impulse of an ultra-short driver. In contrast, the ion-wake is excited as the ions undergo sustained interactions with the bubble fields within the spatial extent of the wake over several plasma electron oscillations. This happens because the electron wake-plasmon oscillations [@wakefield] have a near speed-of-light phase-velocity ($\beta_{\phi} \simeq 1$) but negligible group-velocity [@Vlasov] $\beta_g\approx 3 v_{th}^2/c^2$ (in the 1-D limit), where, $v_{th} \simeq \sqrt{k_BT_e/m_e}$ is the mean electron thermal velocity of the background plasma. Therefore a slowly-propagating train of coupled electron plasmons is excited in a cold collision-less plasma [@Vlasov]. A large difference between phase-velocity and the group velocity of the electron oscillations allows sustained field-ion interactions. It should be noted that high phase-velocity plasma electron waves are possible only in a cold plasma with appropriate density, $n_0$ that allows near speed-of-light propagation of the energy sources, $\beta_{es} \simeq 1 \approxeq \beta_{\phi}$. Ion-soliton modeled here is assuming a significant difference between the phase-velocity and the group velocity of the plasma-electron waves.
A time symmetric electron wakefield would excite time symmetric ion oscillations where the ion velocities average to zero. However, the bubble wake is asymmetric in time as the back of the bubble electron compression is a small fraction of the length of electron cavitation. The electron oscillations become non-linear at high driver intensities as all the interacting electrons are displaced radially, $\delta n_e/n_0 > 1$, forming a non-linear bubble-shaped electron spatial structure enclosing ions in its cavity. The wakefields excited in the bubble are useful for accelerating electrons [@cavitation-beam][@cavitation-laser-expt][@cavitation-beam-expt]. High intensities also lead to fields that can directly drive the plasma electrons to velocities near the speed-of-light. This occurs when for a laser pulse $a_0 \geq 1$ and an electron beam $\frac{n_b}{n_0}\left(\frac{r_b}{c/\omega_{pe}}\right)^2\geq 1$ where $a_0$ is the peak normalized laser vector potential, $n_b, r_b$ the peak beam density and radius. The radially expelled electrons oscillate radially under the force of the plasma ions. These oscillations are excited over plasma electron oscillation timescales, $2\pi\omega_{pe}^{-1}$ ($\omega_{pe}=\sqrt{4\pi n_0 e^2/\gamma_em_e}$) where $\gamma_e\beta_em_ec$ is the temporally anharmonic relativistic electron quiver momentum. The normalized quiver momentum of the electrons in the bubble-oscillations is relativistic $\gamma_{\perp}\beta_{\perp} \geq 1$ and the quiver frequency is $\omega_{\perp}=\omega_{pe} \left( \frac{\beta_{\phi}^2}{\gamma(1-\beta_{\phi}^2)} \right)^{1/2}$ [@Akhiezer-Polovin].
We show that non-linear ultra-relativistic electron wakefields interacting with the plasma ions lead to the excitation of a non-linear ion-wake. The non-linear ion-wake $\delta n_i/n_0 > 1$ in Fig.\[fig1:ion-wake-laser\] and Fig.\[fig2:ion-wake-beam\] is excited over timescales $\gg 2\pi\omega_{pe}^{-1}$ in the trail of a bubble-wake train. By shaping the energy source it can be matched or guided to excite a long train of nearly identical plasmons, Fig.\[fig3:bubble-train\]. Since it is the electric field ${\bf E}_{wk}$ of a nearly stationary bubble plasmon that excite collective ion-motion we model the ion dynamics in a single bubble. Using the single bubble ion dynamics, Fig.\[fig4:bubble-ion-dynamics\] we model the ion-wake over the whole bubble-train spanning several hundred plasma skin-depths ($c/\omega_{pe}$).
The wake-plasmon energy density ($\mathcal{E}_{wk} = 0.5 (e\lvert {\bf E}_{p}\rvert/(m_ec\omega_{pe}))^2 ~ m_e c^2 ~ n_0$, where ${\mathbf E}_{p}$ is the wakefield amplitude) is continually partitioned between the field energy and the coherent electron quiver kinetic energy. In our model we do not include heavy beam-loading of the bubble electron wake. Under heavy beam-loading the bubble field energy is efficiently coupled to the kinetic energy of the accelerated beam. In this scenario the bubble collapses and the magnitude of the ion-wake is smaller. The decoherence of the ordered electron quiver to random thermal energy, $\mathcal{E}_{wk} \rightarrow k_BT_{wk}$ due to the phase-mixing [@phase-mixing-longitudinal] of individual electron trajectories caused by the non-linearities and inhomogeneities is further stimulated by the ion motion. The details of the thermalization of the wake electrons under ion motion is beyond the scope of this paper. It is over these timescales upon thermalization that the steepened ion-density expands outwards radially as a non-linear ion-acoustic wave driven by the electron thermal pressure. The energy transfer process observed here is a coupling from the non-linear plasma electron-mode to a non-linear ion-acoustic mode [@lte-napac-2013]. We also observe energy coupling to the bow-shock which is formed behind the bubble, Fig.\[fig4:bubble-ion-dynamics\].
[1]{}
Akhiezer, A. I. and Polovin, R. V., *Theory of wave-motion of an electron plasma*, Zh. Eksp. Teor. Fiz, 30, 915 (1956) \[Sov. Phys. JETP 3, 696 (1956))
Tajima, T., Dawson, J. M., *Laser Electron Accelerator*, [Phys. Rev. Lett. [**43**]{}, pp.267-270 (1979)](http://link.aps.org/doi/10.1103/PhysRevLett.43.267), doi: 10.1103/PhysRevLett.43.267;\
Chen, P., Dawson, J. M., Huff, R. W., Katsouleas, T., *Acceleration of electrons by the interaction of a bunched electron beam with a plasma*, [Phys. Rev. Lett. [**54**]{} (7), 693-696 (1985)](http://link.aps.org/doi/10.1103/PhysRevLett.54.693), doi: 10.1103/PhysRevLett.54.693
Sun, G. Z., Ott, E., Lee, Y. C., Guzdar, P., *Self-focusing of short intense pulses in plasmas*, [Phys. of Fluids [**30**]{}, 526 (1987)](http://dx.doi.org/10.1063/1.866349), doi: 10.1063/1.866349.
Rosenzweig, J. B., Breizman, B., Katsouleas, T., Su, J. J., *Acceleration and focusing of electrons in two-dimensional nonlinear plasma wake fields*, [Phys. Rev. A, [**44**]{}, (1991)](http://dx.doi.org/10.1103/PhysRevA.44.R6189), doi:10.1103/PhysRevA.44.R6189
Pukhov, A., Meyer-Ter-Vehn, J., *Laser wake field acceleration: the highly non-linear broken-wave regime*, [Appl. Phys. B [**74**]{}, pp.355-361 (2002)](http://dx.doi.org/10.1007/s003400200795). doi:10.1007/s003400200795
Lu, W., Huang, C., Zhou, M., Tzoufras, M., Tsung, F. S., Mori, W. B., and Katsouleas, T., *A nonlinear theory for multidimensional relativistic plasma wave wakefields*, [Physics of Plasmas [**13**]{}, 056709 (2006)](http://dx.doi.org/10.1063/1.2203364), doi: 10.1063/1.2203364
Rosenzweig, J. B., Cook, A. M., Scott, A., Thompson, M. C., Yoder, R. B., *Effects of Ion Motion in Intense Beam-Driven Plasma Wakefield Accelerators*, [Phys. Rev. Lett. [**95**]{}, 195002 (2005)](http://dx.doi.org/10.1103/PhysRevLett.95.195002). doi:10.1103/PhysRevLett.95.195002
Gholizadeh, R., Katsouleas, T., Muggli, P., Huang, C. K., Mori, W. B., *Preservation of Beam Emittance in the Presence of Ion Motion in Future High-Energy Plasma-Wakefield-Based Colliders*, [Phys. Rev. Lett., [**104**]{}, 155001 (2010)](http://dx.doi.org/10.1103/PhysRevLett.104.155001) doi:10.1103/PhysRevLett.104.155001
Supplementary Online Material - Movie of the radial dynamics of electron and ion density driven in the Ion-wake channel of a electron-beam driven non-linear electron wake. (in MP4 format)
Sahai, A. A., Katsouleas, T. C., Tsung, F. S., and Mori, W. B., *Long term evolution of plasma wakefields*, [Proc of NAPAC, Sep 2013, Pasadena, CA, USA](http://accelconf.web. cern.ch/accelconf/pac2013/papers/mopac10.pdf) (MOPAC10, ISBN 978-3-95450-138-0), 03-Alternative Acceleration Schemes, A23, Laser-driven Plasma Acceleration.
Sahai, A. A., Katsouleas, T. C., Gessner, S., Hogan, M., Joshi, C., Mori, W. B., *Excitation of wakefields in a relativistically hot plasma created by dying non-linear plasma wakefields*, [AIP Conf. Proc. 1507, 618 (2012)](http://dx.doi.org/10.1063/1.4773768). doi:10.1063/1.4773768 (also SLAC-PUB-15367)
Mangles, S. P. D., Murphy, C. D., Najmudin, Z., Thomas, A. G. R., Collier, J. L., Dangor, A. E., Divall, E. J., Foster, P. S., Gallacher, J. G., Hooker, C. J., et. al., [Nature [**431**]{}, 535 (2004)](http://dx.doi.org/10.1038/nature02939); Geddes, C. G. R., Toth, Cs., van Tilborg, J., Esarey, E., Schroeder, C. B., Bruhwiler, D., Nieter, C., Cary, J. & Leemans, W. P., [Nature [**431**]{}, 538 (2004)](http://dx.doi.org/10.1038/nature02900); Faure, J., Glinec, Y. , Pukhov, A., Kiselev, S., Gordienko, S., Lefebvre, E., Rousseau, J.-P., Burgy, F. & Malka, V., [Nature [**431**]{}, 541 (2004)](http://dx.doi.org/10.1038/nature02963).
Sahai, A. A. and Katsouleas, T. C., *Optimal positron-beam excited wakefields in hollow and ion-wake channels*, [Proceedings of International Particle Accelerator Conference 2015, Richmond, Virginia, May (2015)]()
Sahai, A. A., *Crunch-in regime - Non-linearly driven hollow-channel plasma*, [Oral Contribution at the Advanced Accelerator Conference, 2016, MD, USA. arXiv:1610.03289](https://arxiv.org/abs/1610.03289)
Sahai, A. A., *On Certain Non-linear and Relativistic Effects in Plasma-based Particle Acceleration*, [Ph.D. dissertation, Duke university, July 2015](http://search.proquest.com/docview/1753637333)
Chiou, T. C. and Katsouleas, T., *High Beam Quality and Efficiency in Plasma-Based Accelerators*, [Phys. Rev. Lett. 81, 3411 (1998); doi: 10.1103/PhysRevLett.81.3411](https://dx.doi.org/10.1103/PhysRevLett.81.3411)
Litos, M., Adli, E., An, W., Clarke, C. I., Clayton, C. E., Corde, S., Delahaye, J. P., England, R. J., Fisher, A. S., Frederico, J., et. al., *High-efficiency acceleration of an electron beam in a plasma wakefield accelerator*, [Nature, [**92**]{}, 515, (2014)](http://dx.doi.org/10.1038/nature13882). doi:10.1038/nature13882
Sen-Gupta, S., Kaw, P. K., *Phase Mixing of Nonlinear Plasma Oscillations in an Arbitrary Mass Ratio Cold Plasma*, [Phys. Rev. Lett. [**82**]{}, 1867 (1999)](http://dx.doi.org/10.1103/PhysRevLett.82.1867). doi:10.1103/PhysRevLett.82.1867
Stephen Maxon and James Viecelli, Cylindrical solitons, [Physics of Fluids (1958-1988) 17, 1614 (1974)](http://dx.doi.org/10.1063/1.1694941); doi: 10.1063/1.1694941; Stephen Maxon and James Viecelli, Spherical solitons, [Physical Review Letters, [**32**]{}, p.4 (1974)](http://dx.doi.org/10.1103/PhysRevLett.32.4); doi: 10.1103/PhysRevLett.32.4
Noah Hershkowitz and Thomas Romesser, Observations of Ion-Acoustic Cylindrical Solitons, [Physical Review Letters, [**32**]{}, iss. 11, pp.581-583 (1974)](http://dx.doi.org/10.1103/PhysRevLett.32.581); doi: 10.1103/PhysRevLett.32.581
Berezin, Yu. A., Karpman, V. I., Theory of Non-stationary Finite-amplitude Waves in a Low-density Plasma, [JETP, Vol. 19, No. 5, p. 1265 (November 1964)](http://www.jetp.ac.ru/cgi-bin/e/index/e/19/5/p1265?a=list) \[ZhETF, Vol. 46, No. 5, p. 1880, November (1964)\]
Korteweg, D. J. and de Vries, G., *On the Change of Form of Long Waves Advancing in a Rectangular Canal, and on a New Type of Long Stationary Waves*, [Philosophical Magazine [**39**]{}, 240, pp.422-443, (1895)](http://dx.doi.org/10.1080/14786449508620739). doi:10.1080/14786449508620739;
Schamel, H., Stationary solitary, snoidal and sinusoidal ion acoustic waves [Plasma Phys. [**14**]{}, 905 ??(1972??)](http://dx.doi.org/10.1088/0032-1028/14/10/002). doi:10.1088/0032-1028/14/10/002
Lee, S., Katsouleas, T. C., Hemker, R. G., Dodd, E. S., Mori,W. B., *Plasma-wakefield acceleration of a positron beam*, [Phys. Rev. E, [**64**]{}, 045501 (2001)](http://dx.doi.org/10.1103/PhysRevE.64.045501), doi:10.1103/PhysRevE.64.045501
Vlasov, A. A., *The vibrational properties of an electron gas*, [Zh. Eksp. Tor. Fitz., [**8**]{}, pp.291 (1938)](http://ufn.ru/en/articles/1968/6/a/). [Soviet Physics Uspekhi, [**93**]{}, No.3 and 4, May-June 1968](http://dx.doi.org/10.1070/PU1968v010n06ABEH003709).
Tonks, L. and Langmuir, I., *Oscillations in Ionized Gases*, [Physical Review, [**33**]{}, pp.195-210, Feb. (1929)](http://dx.doi.org/10.1103/PhysRev.33.195); Chen, F. F., *Introduction to Plasma Physics and Controlled Fusion, Plenum Press, New York (1984), ISBN-13: 978-0306413322*
Sahai, A. A., Katsouleas, T. C., Tsung, F. S., Mori, W. B., *Long term evolution of plasma wakefields*, [Proc. of NAPAC 2013, [**MOPAC10**]{}, Sep 2013, Pasadena, CA, USA](http://accelconf.web.cern.ch/accelconf/pac2013/papers/mopac10.pdf)
Ruth, R. D., Chao, A. D., Morton, P. L., Wilson, P. B., *A plasma wake field accelerator*, [Particle Accelerator, [**17**]{}, pp.171-189 (1985)](http://www.slac.stanford.edu/cgi-wrap/getdoc/slac-pub-3374.pdf) SLAC-PUB-3374, July 1984
Fonseca, R. A., Silva, L. O., Tsung, F. S., Decyk, V. K., Lu, W., Ren, C., Mori, W. B., Deng, S., Lee, S., Katsouleas, T., Adam, J. C., *OSIRIS, a three-dimensional fully relativistic particle in cell code for modeling plasma based accelerators*, [Lect. Note Comput. Sci. [**2331**]{}, pp.342-351 (2002)](http://dx.doi.org/10.1007/3-540-47789-6_36). doi:10.1007/3-540-47789-6\_36
Gorbunov, L. M., Mora, P., Solodov, A. A., *Plasma Ions Dynamics in the Wake of a Short Laser Pulse*, [Phys. Rev. Lett., [**86**]{}, (2001).](http://dx.doi.org/10.1103/PhysRevLett.86.3332). doi:10.1103/PhysRevLett.86.3332
Chiou, T. C., Katsouleas, T. C., *High Beam Quality and Efficiency in Plasma-Based Accelerators*, [Phys. Rev. Lett. [**81**]{}, 3411, (1998)](http://dx.doi.org/10.1103/PhysRevLett.81.3411). doi:10.1103/PhysRevLett.81.3411
Vieira, J., Fonseca, R. A., Mori, W. B., Silva, L. O., *Ion Motion in Self-Modulated Plasma Wakefield Accelerators*, [Phys. Rev. Lett. [**109**]{}, 145005 (2012)](http://dx.doi.org/10.1103/PhysRevLett.109.145005). doi:10.1103/PhysRevLett.109.145005
Ting, A., Moore, Krushelnick, K., Manka, C., Esarey, E., Sprangle, P., Hubbard, R., Burris, H. R., Fischer, R., & Baine, M., *Plasma wakefield generation and electron acceleration in a self-modulated laser wakefield accelerator experiment*, [Phys. Plasmas [**4**]{}, iss.5, (1997)](http://dx.doi.org/10.1063/1.872332). doi: 10.1063/1.872332; Durfee, C. G., and Milchberg, H. M., *Light Pipe for High Intensity Laser Pulses*, [Phys. Rev. Lett. [**71**]{}, 2409, (1993)](http://dx.doi.org/10.1103/PhysRevLett.71.2409). doi:10.1103/PhysRevLett.71.2409
Kimura, W. D., Milchberg, H. M., Muggli, P., Li, X., & Mori, W. B., *Hollow plasma channel for positron plasma wakefield acceleration*, [Phys. Rev. ST Accel. Beams [**14**]{}, 041301 (2011)](http://dx.doi.org/10.1103/PhysRevSTAB.14.041301). doi:10.1103/PhysRevSTAB.14.041301
Hubbard, R. F., Ehrlich, Y., Kaganovich, D., Cohen, C., Moore, C. I., Sprangle, P., Ting, A., Zigler, A., *Intense laser pulse propagation in capillary discharge plasma channels*, [AIP Conf. Proc. 472, 394 (1999)](http://dx.doi.org/10.1063/1.58909). doi:10.1063/1.58909
Tzoufras, M., Tsung, F. S., Mori, W. B., Sahai, A. A., *Improving the Self-Guiding of an Ultraintense Laser by Tailoring Its Longitudinal Profile*, [Phys. Rev. Lett. [**113**]{}, 245001 (2014)](http://dx.doi.org/10.1103/PhysRevLett.113.245001). doi:10.1103/PhysRevLett.113.245001
Sahai, A. A., Katsouleas, T. C., Muggli, P., *Self-injection by trapping of plasma $e^-$ oscillating in rising density gradient at the vacuum-plasma interface*, [[**TUPME51**]{}, Proc of IPAC2014, Dresden, Germany](http://accelconf.web.cern.ch/AccelConf/IPAC2014/papers/tupme051.pdf)
Schroeder, C. B., Whittum, D. H., Wurtele, J. S., *Multimode Analysis of the Hollow Plasma Channel Wakefield Accelerator*, [Phys. Rev. Lett., [**82**]{}, 1177 (1999)](http://dx.doi.org/10.1103/PhysRevLett.82.1177). doi:10.1103/PhysRevLett.82.1177
Schroeder, C. B., Benedetti, C., Esarey, E., Leemans, W. P., *Beam loading in a laser-plasma accelerator using a near-hollow plasma channel*, [Physics of Plasmas, [**20**]{}, iss.12, 123115, (2013)](http://dx.doi.org/10.1063/1.4849456) doi:10.1063/1.4849456; Schroeder, C. B., Esarey, E., Benedetti, C., *Control of focusing forces and emittances in plasma-based accelerators using near-hollow plasma channels*, [Physics of Plasmas, [**20**]{}, iss.8, 080701, (2013)](http://dx.doi.org/10.1063/1.4817799). doi:10.1063/1.4817799
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Dynamics of individual DNA undergoing constant field gel electrophoresis (CFGE) is studied by a Brownian dynamics (BD) simulation method which we have developed. The method simulates electrophoresis of DNA in a 3 dimensional (3D) space by a chain of electrolyte beads of hard spheres. Under the constraint that the separation of each pair of bonded beads is restricted to be less than a certain fixed value, as well as with the excluded volume effect, the Langevin equation of motion for the beads is solved by means of the Lagrangian multiplier method. The resultant mobilities, $\mu$, as a function of the electric field coincide satisfactorily with the corresponding experimental results, once the time, the length and the field of the simulation are properly scaled. In relatively strong fields quasi-periodic behavior is found in the chain dynamics, and is examined through the time evolution of the radius of the longer principal axis, $R_l(t)$. It is found that the mean width of a peak in $R_l(t)$, or a period of one elongation-contraction process of the chain, is proportional to the number of beads in the chain, $M$, while the mean period between two such adjacent peaks is proportional to $M^0$ for large $M$. These results, combined with the observation that the chain moves to the field direction by the distance proportional to $M$ in each elongation-contraction motion, yield $\mu \propto M^0$. This explains why CFGE cannot separate DNA according to their size $L\ (\propto M)$ for large $L$.'
author:
- Ryuzo Azuma
- Hajime Takayama
title: |
Brownian Dynamics Studies on DNA Gel Electrophoresis. I.\
Numerical Method and Quasi-Periodic Behavior of Elongation-Contraction Motions
---
Introduction
============
Gel electrophoresis is a most widely used technique to separate DNA according to their length $L$ under an electric field. [@viovy00] The constant field method is quite efficient for DNA of small sizes. It is well known, however, that the separation does not work well for DNA of sizes longer than a certain value which depends on the field and the “pore” diameter of the gel. [@nord91] The non-constant field methods such as the pulsed-field gel electrophoresis [@schcan] have been empirically invented and they have remarkably improved the shortcoming. Nevertheless, the understandings on these gel electrophoresis methods, even on the constant field method, still remain unsatisfactory from the view point of statistical physics of polymers.
By constant field gel electrophoresis (CFGE), DNA molecules with the radius of inertia, $R_{\rm I}$, smaller than the pore diameter of gel, $a_{\rm gel}$, are [*sieved*]{} by gel. [@rodb70; @chra71] In this sieving regime DNA with the smaller $R_{\rm I}$ migrate the faster through the gel. In the opposite case with $R_{\rm I}\gtrsim
a_{\rm gel}$, DNA is considered to move through a “primitive chain” [@DoEd86] which is determined by stochastic movement of its front end in pores of the gel under the field. [@lump82; @lump85; @slat86] This picture, denoted as the biased reptation model, successfully describes various features of CFGE, such as the $L$-dependence of the migration mobility, $\mu$, observed experimentally [@nanc85] in sufficiently weak fields. More recently, the model of the biased reptation with fluctuations has been proposed [@sdv95] which well describes the observed field dependence of $\mu$ in moderate fields. [@viovy00] When the field is further increased, the field-dependence of $\mu$ begins to be saturated. In this regime, interesting behavior of an individual DNA has been observed: “hernias”-like configurations by the computer simulations [@deut88; @deut89] and “quasi-periodic” behavior, i.e., “elongation-contraction” motion in a cyclic way by the real-time fluorescent microscopy experiments. [@masu93; @oana94] Indeed, the latter experimental method combined with optical or magnetic tweezers for a molecule reveals that the elasticity of DNA is originated from the entropy associated with its static conformations. [@smith92; @perk95; @stric96; @aust97] However, dynamic properties of DNA under CFGE are not described in a unified way by this picture alone.
In order to get further insights into the problem mentioned above, we have developed a Brownian dynamics (BD) method which simulates electrophoresis of DNA in a 3 dimensional (3D) space. The method has been extensively applied to the CFGE, in particular, the elongation-contraction motion of the chain under CFGE, the results of which we will discuss in the present series of papers. In the present paper I our numerical method is explained and global aspects, such as the mean mobility $\mu$ and time evolutions of the velocity of the center of mass, $v_{\rm G}(t)$, and of the radius of the longer principal axis, $R_l(t)$, are investigated. In the accompanying paper II [@ourII] we analyze dynamics of ‘defects’ introduced by de Gennes [@deGe71] in a chain under CFGE. Finally in paper III [@ourIII] we will examine conformational changes of the coarse-grained chain under CFGE in details, and will argue that the field-parallel component of its elongation-contraction motion is essentially understood as a deterministic dynamics of an elastic string in a 1D space with an obstacle, thereby an origin of the elasticity will be shown to be the conformational entropy of the chain.
In the Brownian dynamics (BD) simulations extensively performed in the present series of studies, DNA is modeled as a chain of spherical electrolyte beads with a constant radius $1/2$ in a 3D space filled with a solvent. Each bead interacts with other beads of the chain and spatially-fixed beads constituting gel by hard-core repulsive interactions. It also interacts with its two bonded beads by a non-linear elastic force which keeps the distance between the two to be less than a threshold value which we put $\sqrt{2}$. To solve the Langevin equation of motion under these constraints as well as under the field and random forces, we employ a BD algorithm with the Lagrangian multiplier method. It is similar to the one by Deutsch and Madden, [@deut88; @deut89] but is essentially different from their method in the following respects. In their method, the time pitch to solve the equation is adjusted so as to keep a fixed distance between bonded beads under the tensile forces which are evaluated by the Lagrangian multiplier method. In our method, on the other hand, the time pitch is fixed at a certain value, which we regard as the mean period of random forces. Within each time step we evaluate paths of all beads affected by the constraints, i.e., by hard-core interactions and extremely nonlinear elastic interactions between a pair of beads. These effects are reread as the supposed constraining forces with the appropriate Lagrangian multipliers in an iterative way. The resultant constraining forces thus specified yield a new set of bead positions without any violation to the constraints. To our knowledge the present BD analysis is the first extensive simulation of DNA gel electrophoresis in a 3D space based on the Langevin equation involving purely mechanical, microscopic forces which are supposed to act on monomers of DNA.
The purposes of the present paper I are to explain our BD method and to demonstrate its results on some global properties of DNA under CFGE. The resultant mobility $\mu$ as a function of the field turns out to agree well with those observed experimentally once a set of three parameters, $l^*$, $\tau^*$ and $E^*$, are properly specified. They relate respectively the scales of length, time and field strength of the simulation and the experiments.
In the BD simulation under relatively large fields quasi-periodic time evolution is observed in $R_l(t)$ (see Fig. \[fig:Fig-8\] below), as well as in $v_{\rm G}(t)$, as has been observed experimentally. [@masu93; @oana94] Each local maximum of $R_l(t)$ at $t_{\rm max}^n$ with $n=1,2,...$ corresponds to a conformation that a chain just gets rid of trapping due to gel. The local minimum at $t_{\rm min:l}^n$ preceding this maximum is then considered to be the instance that, being trapped by gel, the chain starts to elongate from a certain coiled conformation, while at time $t_{\rm min:r}^n$ of the local minimum subsequent to $t_{\rm max}^n$ the chain retrieves another coiled state. We establish a method to properly classify a whole time sequence of $R_l(t)$ into peak parts, from $t_{\rm min:l}^n$ to $t_{\rm min:r}^n$ and the rest. We call the former parts [*deterministic*]{} ranges and the latter [*stochastic*]{} ones. We then investigate statistics of periods of the deterministic ranges, $D_n \equiv t_{\rm min:r}^n - t_{\rm min:l}^n$, and those of the stochastic ranges, $(1/\lambda)_n \equiv
t_{\rm min:l}^{n+1} - t_{\rm min:r}^n$. It is found that {$D_n$} obey a Gaussian distribution with the mean $D \equiv \overline{D_n} \propto
M$ and the width $\Delta \propto D$, while {$(1/\lambda)_n$} do a Poisson distribution with the mean $1/\lambda \equiv
\overline{(1/\lambda)_n} \propto M^0$ for large $M$, where $M$ is the number of beads of the chain and the overline stands for the averages over $n$, i.e., peak structures. These results, combined with the observation that in each elongation-contraction motion the chain moves to the field direction by the distance proportional to $M$, yield $\mu \propto M^0$ and explain why CFGE cannot separate DNA according to their sizes $L\ (\propto M)$ for large $L$.
The organization of the present paper is as follows. In the next section we explain our BD method. The mobility $\mu$ of the chain obtained by the BD method is presented and compared with the experimental results in Sec. \[sec:mocuge\]. The method of analysis of peak structures in $R_l(t)$ is explained, and statistical nature of the [*deterministic*]{} and [*stochastic*]{} ranges is discussed in Sec. \[sec:ecmw\]. The last section is devoted to further discussions of our results.
The Brownian Dynamics Method {#sec:method}
============================
\[sec:bdm\]
A chain of spherical beads in a continuous 3D space, shown schematically in Fig. \[fig:Fig-5\], is investigated by a new type of Brownian dynamics (BD) calculation. The hard core interaction is assumed in order for centers of any pair of beads not to come closer than 1 (the excluded volume effect), where the radius of beads is set $1/2$. Furthermore distances between neighboring beads, $l$, are restricted to $l<\sqrt{2}$. We suppose that this is realized by the extremely nonlinear elastic interaction between the neighboring beads: it is infinite for $l>\sqrt{2}$ and zero otherwise. These interactions assure the self-avoiding walk of the chain.
The BD method we adopt in the present work is to solve the following Langevin equation of motion for the chain, $$\zeta \dot{\bm x}_i = q{\bm E}_{\rm b}
+ \sum_\alpha \sum_j {\bm S}^\alpha_{ij}
+ \sum_\alpha \left[ {\bm F}^\alpha_i-{\bm F}^\alpha_{i-1}\right]
+ {\bm f}_i,
\label{eqn:lange}$$ where ${\bm x}_i$ is the position of the center of the $i$-th bead, and $\zeta$, $q$, and ${\bm E}_{\rm b}$ are the viscosity, the charge of a bead and the (bare) external field, respectively. The second and the third terms in r.h.s. are the constraining forces required for the beads to satisfy the conditions, $$\begin{aligned}
\|{\bm x}_{ij}\|=&\|{\bm x}_{j}-{\bm x}_i\|&>1\label{eqn:cexcl}\\
\|{\bm l}_i\|=&\|{\bm x}_{i+1}-{\bm x}_i\|&<\sqrt{2}.\label{eqn:cbond}\end{aligned}$$ The vectors ${\bm S}^\alpha_{ij}$ and ${\bm F}^\alpha_i$ are respectively written as $\theta^\alpha_{ij}[{\bm x}_i-{\bm x}_j]$ and $\phi^\alpha_i[{\bm x}_i-{\bm x}_{i+1}]$ by introducing the Lagrangian multipliers $\theta^\alpha_{ij}$ and $\phi^\alpha_i$. The last term in eqn. \[eqn:lange\] stands for the random force from a solvent. Its distribution is characterized by the following averages $$\begin{aligned}
\langle {\bm f}_i \rangle &=& {\bm 0}\\
\langle {\bm f}_i(t){\bm f}_j(t')\rangle &=& 2\zeta k_{\rm B}T\Delta t\delta_{ij}\delta(t-t').\label{eqn:chi}\end{aligned}$$ Here, $k_{\rm B}$, $T$, and $\Delta t$ are the Boltzmann constant, the absolute temperature, and the average period of the random force, respectively.
To solve the one step evolution of eqn. \[eqn:lange\], thereby the terms $\{{\bm S}^{\alpha}_{ij},{\bm F}^{\alpha}_i\}$ with the Lagrangian multipliers are determined, we examine the full Langevin equation with the inertia term and the forces arising from the hard core interaction and the nonlinear elastic interaction mentioned above. We solve the equation in a time interval of $\Delta t > t > 0$, and take the overdamped limit, $m/\zeta\to 0$, where $m$ is the supposed mass of the bead. The constraints of eqns. \[eqn:cexcl\] and \[eqn:cbond\] are supposed to be satisfied by positions of the beads at $t=0$, denoted by $\{{\bm x}^0_i\}$. At this instance ${\bm f}_i$ and $q{\bm E}_{\rm b}\Delta t$ are added as impulsive forces with the integrated strength $\tilde{\bm f}_i$ where $\tilde{\bm f}_i={\bm f}_i+q{\bm E}_{\rm b}\Delta t$. When the constraints are discarded, the solution in the overdamped limit is described by the relation between the displacement $\Delta {\bm x}_i$ in this time interval and the force as $$\label{eqn:pred}
\Delta {\bm x}_i=\frac{1}{\zeta}\tilde{\bm f}_i,$$ which is just what is obtained by simply integrating eqn. \[eqn:lange\] without the second and the third terms.
In the case that positions of a pair of beads ${\bm x}^{01}_i\equiv {\bm x}^0_i +\Delta {\bm x}_i$ and ${\bm x}^{01}_j\equiv {\bm x}^0_j+\Delta {\bm x}_j$ violate the constraint of eqn. \[eqn:cexcl\], we solve the full Langevin equation for this pair with the hard core scattering, which can be either elastic or inelastic. We then obtain, again in the overdamped limit, positions ${\bm x}^1_i$ and ${\bm x}^1_j$ at $t=\Delta t$ as shown in Fig. \[fig:scattering\]a in case of the inelastic scattering. The result is reread as the consequence of the constraining force, ${\bm S}^1_{ij}\equiv \theta^1_{ij}[{\bm x}^1_i-{\bm x}^1_j]$, which is determined by the conditions ${\bm x}^1_i-{\bm x}^0_i=(\tilde{\bm f}_i +{\bm S}^1_{ij})/\zeta$ and ${\bm x}^1_j-{\bm x}^0_j=(\tilde{\bm f}_j -{\bm S}^1_{ij})/\zeta$ (note that ${\bm S}^1_{ji}=-{\bm S}^1_{ij}$). This constraining force ${\bm S}^1_{ij}$ just corresponds to the one in the second term of eqn. \[eqn:lange\]. In the case ${\bm x}^{01}_i$ and ${\bm x}^{01}_{i+1}$ violate eqn. \[eqn:cbond\], the nonlinear elastic force is supposed to reduce the distance $\|{\bm l}_i\|$ to $\sqrt{2}$. This effect is represented by the constraining force ${\bm F}^1_i\ (=-{\bm F}^1_{i+1})$, which is added to the third term of eqn. \[eqn:lange\]. Its Lagrangian multiplier $\phi^1_i$ is determined in such a way that $\|{\bm x}^1_i-{\bm x}^1_{i+1}\| =\sqrt{2}$ with ${\bm x}^1_i={\bm x}^0_i+{\bm F}^1_i\Delta t$ and ${\bm x}^1_{i+1}= {\bm x}^0_{i+1}+{\bm F}^1_{i+1}\Delta t$ (see Fig. \[fig:scattering\]b). If ${\bm x}^{01}_i$ involves violation to more than one constraints of eqns. \[eqn:cexcl\] and/or \[eqn:cbond\], the above-mentioned procedures are applied to evaluate the constraining forces for each pair of beads with the common ${\bm x}^{01}_i$. If, on the other hand, ${\bm x}^{01}_i$ does not involve any violation to eqns. \[eqn:cexcl\] and \[eqn:cbond\], we put ${\bm x}^1_i={\bm x}^{01}_i$.
In the configuration $\{{\bm x}^1_i\}$, it is generally expected that new pairs of beads may violate the constraints since we have tried to remedy the violations in $\{{\bm x}^{01}_i\}$ in a pairwise way. Then we repeat the above-mentioned procedure, in which $\tilde{\bm f}_i$ is replaced by $\tilde{\bm f}_i+{\bm C}^1_i$, where ${\bm C}^1_i=\sum_j{\bm S}^1_{ij}
+{\bm F}^1_i-{\bm F}^1_{i-1}$ is the constraining force determined by the first procedure. If a configuration without any violation to the constraints is obtained by the $\alpha_{\rm con}$ times repetition of this procedure, we regard $\{{\bm x}^{\alpha_{\rm con}}_i\}$ as the solution of eqn. \[eqn:lange\] evolved from $\{{\bm x}^0_i\}$ by one unit of time $\Delta t$. In this manner the multi-scattering processes in the interval of $\Delta t$, including those due to the nonlinear elastic interaction, can be taken into account, and the corresponding forces $\{{\bm S}^{\alpha}_{ij},{\bm F}^\alpha_i\}$ in eqn. \[eqn:lange\] are determined.
Although our BD method is based on the overdamped Langevin equation with the Lagrangian multipliers, it distinctly differs from the one due to Deutsch and Madden. [@deut88; @deut89] In the latter the equidistant condition on all the pairs of beads is imposed, and the constraining tensile forces associated with it are evaluated at the instance when random forces are applied. Then the time interval $\Delta t$ is adjusted so that the configuration at $t + \Delta t$ in fact obeys the equidistant condition within accuracy of $0.1\%$. In contrast, our BD method takes into account a similar, but looser, constraint on the bond length as well as the excluded volume effect of the beads by explicitly examining the supposed multiple scattering within the interval $\Delta t$ which is fixed. Therefore there is no finer time unit other than $\Delta t$ in solutions of our BD method. It is the mean period of random forces $\{{\bm f}_i\}$.
In the present work we simulate dynamics of a chain in a continuous 3D space confined into a cubic box with a volume $360^3$. The periodic boundary condition is imposed to each direction of the box. The gel is represented by a network of immobile bars arranged so that they form a simple cubic lattice. Each bar consists of tightly connected beads with a radius $1/2$, and its interaction with a bead of the mobile chain is computed according to the constraint of eqn. \[eqn:cexcl\]. We employed three lattices of gel with different lattice distances $a_{\rm gel}=10$, $17$, and $20$, which are considered to correspond to the ’pore’ diameter of agarose gel. [@nord91] The lengths of chains, $M$, used in the BD calculation are $M=30$, $40$, $60$, $80$, $160$, $240$, and $320$. We have performed typically $64$ runs whose initial configurations are given independently. The direction of the field is fixed at $(1,1,1)$. The time is measured in unit of $\Delta
t$, i.e., 1 BD step.
Here we want to emphasize that the present BD method correctly reproduces the expected static and dynamic properties of a real (not phantom) polymer in a 3D space under the vanishing field: $R_I\propto M^{\nu_{\rm F}}$ with $\nu_{\rm F}\simeq 0.59 \pm 0.01$ and the Einstein relation $D_G\propto \mu/M$ are confirmed (see Fig. \[fig:mobility\] in Sec. \[sec:mocuge\]). Here, $\nu_{\rm F}$, $R_I$, and $D_G$ are the Flory exponent, the radius of inertia, and the diffusion constant of the center of mass, respectively.
The number of the multiple scattering, $\alpha_{\rm con}$ above mentioned, is expected to be at most of the order of $M^2$. It has turned out that $\alpha_{\rm con}$ becomes relatively larger when the chain is in an extended configuration under gel electrophoresis. We quote here typical figures for a chain with $M=80$: $\alpha_{\rm con}\simeq 4\times 10^1$, $5\times 10^2$ for $E=0$, $0.032$, respectively.
Mobility {#sec:mocuge}
========
The most basic quantity that has been commonly measured in CFGE of DNA is the long time average of velocity divided by the electric field, or the mobility $\mu$. In the present work, we have implemented long time BD simulations with a highly efficient algorithm using a massive parallel computer and have eventually obtained $\mu$, hereafter denoted by $\mu_{\rm BD}$, with enough statistical precision for all the set of parameters investigated. Here we only mention typical figures of computation to get $\mu_{\rm BD}$ of a $M=240$ chain for each run: more than 150 hours are consumed on a single processor of SGI2800 system to pass $10^6$ BD steps. In the present work the following values of the parameters are chosen: $\Delta t=1$, $\zeta= 10$ and $k_{\rm B}T=2$. The results are shown in Fig. \[fig:mobility\] with the experimental data due to Heller [*et al*]{}. [@hdv94] Here and hereafter the value of field $E$ stands for $qE_{\rm b}$ of eqn. \[eqn:lange\].
The simulated $\mu_{\rm BD}$ at $E=0$ in Fig. \[fig:mobility\] is evaluated through the Einstein relation $${\mu_{\rm BD}^0 \over M} = D_{\rm G} \equiv
\lim_{t\rightarrow \infty}{1 \over 6t}
\langle({\bm R}_{\rm G}(t)-{\bm R}_{\rm G}(0))^2\rangle|_{E=0},
\label{eqn:E-relation}$$ where ${\bm R}_{\rm G}(t)$ is the position of the center of mass of the chain. The diffusion constant $D_{\rm G}$ in the r.h.s. is evaluated by the BD simulation in a vanishing field in the time range $t=(5\sim10)\times10^6$. The resultant $\mu_{\rm BD}^0$ turns out to be compatible with the value of $\lim_{E\rightarrow 0}
\mu_{\rm BD}$ as seen in the figure. This confirms that the Einstein relation in fact holds in the present BD simulation for the chain.
In order to compare the BD results with the experimentally observed data, we have to specify proper conversion of values of the parameters in the simulation and the experiment. Let us first consider the length scale of a chain. In the accompanying paper II we examine in detail the embodiment of ‘defects’ due to de Gennes [@deGe71] in our BD chain and obtain the mean distance $\langle n\rangle$ of adjacent ’defects’ along the chain under $E=0$ as $\langle n\rangle \simeq 5.7$. This value, in unit of the mean distance between neighboring beads, $\langle l\rangle$, is almost independent of the set of parameters $a_{\rm gel}$ and $M$ we have examined. Regarding $\langle n\rangle$ as an estimate of the ‘persistent length’ in the BD simulation, ${\cal P}_{\rm BD}$, we put ${\cal P}_{\rm BD}=c\langle n\rangle$, where $c$ is an adjustable parameter whose value is nearly equal to unity (note that $\sqrt{2} \ge \langle l\rangle \ge 1$). The conversion factor of the length-scale $l^*$ is then determined by $${\cal P}_{\rm exp} = l^* {\cal P}_{\rm BD}
= cl^* \langle n\rangle,
\label{eqn:L-scale}$$ where ${\cal P}_{\rm exp}$ is the persistence length experimentally measured. With ${\cal P}_{\rm exp} \sim$ 60nm at $T_{\rm exp} \sim$ 300K taken from Ref. , we obtain $cl^* \sim$ 10nm.
Another length scale of importance is the mean pore size of gel, which is rather difficult to be accurately specified even in the experiments. [@viovy00] Here we simply suppose that DNA in $1\%$ agarose corresponds to a BD chain in the gel with $a_{\rm gel}=17$. With ${\cal P}_{\rm BD}$ above fixed, this value of $a_{\rm gel}$ is about three times larger than that of ${\cal P}_{\rm BD}$.
Next let us relate $E$ in the BD simulation to the experimental field, $E_{\rm exp}$, in V/cm. For this purpose we impose that the ratios $\Delta \varepsilon/k_{\rm B}T$ in the BD calculation and the experiment be identical, where $\Delta \varepsilon$ is the energy that a part of chain of a length equal to the persistence length ${\cal P}$ gains when it moves a distance of ${\cal P}$ under the electric field: $$\frac{\sigma E_{\rm exp}{\cal P}_{\rm exp}^2}{k_{\rm B}T_{\rm exp}}
=\frac{E{\cal P}^2_{\rm BD}}{k_{\rm B}T_{\rm BD}},$$ or $$E_{\rm exp}=E^*E;\ \
E^*={k_{\rm B}T_{\rm exp}{\cal P}_{\rm BD}^2
\over 2\sigma {\cal P}_{\rm exp}^2}
\simeq 5\times 10^2c^{-2}\ {\rm V/cm}.
\label{eqn:rel-E}$$ where $\sigma$ is the density of charge of DNA, and $k_{\rm B}T_{\rm BD}=2$ has been used. Quoting further the experimental value of $\sigma \sim$ 0.6e$^-$/[Å]{}, [@volk94] as well as $cl^*\sim$ 10nm at $T_{\rm exp} \sim$ 300K in eqn. \[eqn:L-scale\], we obtain the last expression for $E^*$ in the above equation. As for the ordinates of Fig. \[fig:mobility\] we may relate $\mu_{\rm BD}$ to $\mu_{\rm exp}
[{\rm cm}^2/{\rm V}\cdot{\rm s}]$ by $$\mu_{\rm exp}=\mu^*\mu_{\rm BD};\ \
\mu^*={l^* \over \tau^* E^* }
\simeq 2 \times 10^{-9}{c \over \tau^*}\ [{\rm cm}^2/{\rm V}\cdot{\rm s}],
\label{eq:rel-mu}$$ where $\tau^*$\[s\] is the real time which corresponds to one discretized time step $\Delta t$ in the BD calculation with $\zeta=10$ (or $\Delta t/\zeta=0.1$).
The two sets of data of $\mu$ in Fig. \[fig:mobility\] are drawn as follows. The data $\mu_{\rm exp}$ are plotted directly using the units indicated on the right ordinate and the upper abscissa, whereas the units for $\mu_{\rm BD}$ are indicated on the left ordinate and the lower abscissa with $c^2=1.60$ and $1.26$ respectively for $M=240$ and $320$. The scales of the two ordinates are related by means of $\tau^*=1.4\times10^{-6}$. We see from the figure that the field dependence of $\mu_{\rm BD}$, particularly that of $M=240$, well coincide with that of $\mu_{\rm exp}$ of DNA with the corresponding length we have specified, i.e., $4.3$ and $6.5$ kbp DNA for $M=240$ and $320$, respectively. This is quite satisfactory if we take into account the fact that only three adjustable parameters are used to draw the two sets of data in Fig. \[fig:mobility\]. We therefore consider that the present BD model catches up even quantitatively essences of DNA dynamics under CFGE.
When we compare $\mu_{\rm BD}$ with $\mu_{\rm exp}$ further in detail, however, there are some unsatisfactory features in the BD results. As compared with $\mu_{\rm exp}$ of $6.5$ kbp DNA, an $E$-independent branch of $\mu_{\rm BD}$ has not been ascertained in the weak field limit for the $M=320$ chain. One possible reason for this may be the insufficient time of simulation (up to $t=10^6$ in finite fields) for this length of chain in weak fields. Thus, within the limited CPU time, it is a rather hard task for the present BD simulation to judge the field dependence of $\mu_{\rm BD}$, proportional to $E$ or $E^2$, in weak fields. [@herb87; @holms90; @dvs94] In the strong field regime, on the other hand, $\mu_{\rm BD}$ tend to saturate, while, $\mu_{\rm exp}$ are still significantly increasing with $E$. These different tendencies are seen more clearly if we compare $\mu_{\rm exp}$ measured in $10$V/cm with the corresponding $\mu_{\rm BD}$ (not seen). The reason of this discrepancy is not clear at the moment.
We have also examined $\mu_{\rm BD}$ of the chains in gels with $a_{\rm gel}=20$ and $10$. If only the volume ratio of gel is considered, they correspond roughly to $0.75\%$ and $3\%$ agaroses, respectively. Qualitatively, $\mu_{\rm BD}$ becomes larger and its $E$-dependence becomes weaker for the larger $a_{\rm gel}$ as expected, but these tendencies are quantitatively weaker than those observed experimentally. The origins of the discrepancy may be attributed to our model for gel, i.e., a perfectly rigid, regular jungle gym.
Elongation-Contraction Motion {#sec:ecmw}
=============================
Motion of an individual chain
-----------------------------
In fields larger than a certain value depending on $M$ and $a_{\rm gel}$, chains in the present 3D BD simulation exhibit quasi-periodic behavior, i.e., they exhibit elongated and contracted shapes alternatively, as observed in the experiments [@masu93; @oana94; @lars95] and in the previous 2D BD models. [@deut88; @matsu94; @mas95p] It is here emphasized that such quasi-periodic behavior has been observed within the field range where we have discussed the $E$-dependence of $\mu$ in Fig. \[fig:mobility\]. We show a typical conformational change of a chain observed in one MD run of CFGE with the parameters $M=240$, $E=0.032$, and $a_{\rm gel} = 17$ in Fig. \[fig:varfigs-new\]. Within a time window of the figure the elongation-contraction motion occurs three times at around $t=$2, 6 and 11 $\times 10^5$. At ranges between them centered around $t=4$ and 8 $\times 10^5$ a chain is in a rather compact form. In this range it can happen, though not so frequently, that the front and rear ends exchange their roles as seen around $t=9\times 10^5$ in the figure. It is also pointed out here that the front end of a chain moves in the field direction with almost a constant rate.
Quasi-periodic behavior of $R_l(t)$ and $v_{\rm G}(t)$ {#sec:quasi-p}
------------------------------------------------------
In Fig. \[fig:Fig-8\] we show the time evolution of the radius of longer principal axis, $R_l(t)$, and the velocity of the center of mass, $v_{\rm G}(t)$, in the MD run whose chain conformational change is shown in Fig. \[fig:varfigs-new\]. The fluctuation in $R_l(t)$ is notable, showing successive $\Lambda$-shaped peaks. We also note that just before the maxima of $R_l(t)$, minima are observed in $v_{\rm G}(t)$. These features are qualitatively in good agreement with those observed in the experiment. [@oana94]
The time evolution of $R_l(t)$ and $v_{\rm G}(t)$ in Fig. \[fig:Fig-8\] looks quasi-periodic. Actually it was found by the experiment that the autocorrelation function of $R_l(t)$, $C_{\rm RR}(t)$, exhibits a damped oscillation. [@oana94] This is also the case for our present BD data. In Fig. \[fig:Fig-19\] $C_{\rm RR}(t)$ evaluated for various $M$ are presented. For $M\ge80$, $C_{\rm RR}(t)$ exhibit the first undershoot below the line of $C_{\rm RR}(t)=0$, and even the first overshoot is clearly seen for $M\ge 160$. From these results we may conclude that, in chains with $M\ge 160$ under $E=0.032$, elongation-contraction motions occur quasi-periodically. In the inset of Fig. \[fig:Fig-19\] we show nearly linear dependence of the period of oscillation $\tau=4t_0$ on $M$, where $t_0$ is defined as the time of the first intercept of $C_{\rm RR}(t)$ with the abscissa. [@oana94]
Analyses on peak structures of $R_l(t)$ {#sec:sop}
---------------------------------------
Oana [*et al.*]{} [@oana94] argued that the steady-state time evolution of DNA under CFGE is classified into two types of behavior. One is the elongation-contraction motion which looks apparently deterministic. We call the time branches in which such a motion is undergoing the [*deterministic*]{} ones (ranges centered at around $t=$2, 6 and 11 $\times 10^5$ in Fig. \[fig:varfigs-new\]). The other is what is observed between two deterministic branches. We call them [*stochastic*]{} branches. Oana [*et al.*]{} approximately described the quasi-periodic behavior of $R_l(t)$ in terms of a simplified function: peaks in $R_l(t)$ are approximated by $\Lambda$-shape branches whose widths are assumed to obey a Gaussian distribution, and others by constants whose duration times are assumed to obey a Poisson distribution. They calculated $C_{\rm RR}(t)$ by means of this model function for various cases and remarked that when $\lambda D$ becomes larger than unity a damped oscillation in $C_{\rm RR}(t)$ becomes prominent. Here $1/\lambda$ and $D$ stand for the mean period of the stochastic branches and the mean width of the deterministic ones, respectively.
Following the classification due to Oana [*et al.*]{}, we analyze our BD results by making use of an algorithm which appropriately picks up peaks of $R_l(t)$. For this purpose we first pick up time sets $\{t^i_{\rm min:l}, t^i_{\rm max}, t^i_{\rm min:r}\}$, i.e., the times of preceding local minimum of the $i$-th maximum, the $i$-th maximum itself, and the subsequent minimum. Then we select out peaks whose height relative to their adjacent minima is larger than a certain threshold value. The latter is chosen in such a way that the sum of periods of peak branches selected surmounts more than $50\%$ of the total duration of the observation. This procedure divides a whole time sequence of $R_l(t)$ into the two types of branches whose numbers are equal to each other. It picks up ensemble of various peaks but not a limited numbers of peaks of special shapes. The number of peaks thus selected is more than a hundred for each set of parameters.
Statistics of deterministic and stochastic branches {#sec:statistics}
---------------------------------------------------
For the set of times $\{t^n_{\rm min:l}, t^n_{\rm max},
t^n_{\rm min:r}\}$ of peaks picked up by the procedure described above we examine the distribution of peak widths (or periods of deterministic branch) $
D_n \equiv t_{\rm min:r}^n-t_{\rm min:l}^n$ and that of duration times between neighboring peaks (periods of stochastic branch) $1/\lambda_n\equiv t_{\rm min:r}^n-t_{\rm min:l}^{n-1}$. The results for $M=240, E=0.032$ and $a_{\rm gel}=17$ are shown in Fig. \[fig:Fig-24\]. They numerically confirm the basic assumptions of Oana [*et al.*]{}, i.e., $\{D_n\}$ nearly obey a Gaussian distribution, while $\{1/\lambda_n\}$ do a Poisson distribution. The mean and the variance of the former are given by $D=3.3\times10^5$ and $\Delta \simeq 0.2D$, respectively, while the mean width of the latter is given by $\lambda D
\simeq 1.7$. The sum $D+1/\lambda$ should be equal to $\tau=4t_0$ evaluated from $C_{\rm RR}(t)$ in Sec. \[sec:quasi-p\], which is verified within our numerical accuracy. The ratio $\lambda D\simeq 1.6$ fulfills the condition for a damped oscillation to be observed in $C_{\rm RR}(t)$.
The ratio $\lambda D$ is found to decrease below unity for $M$ which is smaller than a certain value. In Fig. \[fig:Fig-26\], $D$ and $1/\lambda$ are plotted against $M$ for $E=0.032$ and $a_{\rm gel}=17$. The results strongly suggest that $D\propto M^1$ and $1/\lambda\propto M^0$ for large $M$. In Fig. \[fig:Fig-23\], the $M$-dependence of ratios $\lambda D$ and $\Delta/D$ are presented. As expected from the results shown in Fig. \[fig:Fig-26\], $\lambda D$ is an approximately linearly increasing function for large $M$. On the other hand, $\Delta/D$ is almost constant ($\sim 0.3$), indicating that the Gaussian distribution of $\{D_n\}$ is scaled by the mean $D$ alone. It should be noted that, for $M\nge 100$ where $\lambda D\nge 1$ holds, $C_{\rm RR}(t)$ clearly exhibits a damped oscillation as has been observed in Fig. \[fig:Fig-19\]. Also the result $4t_0 \propto M$ for $M \nge 100$ shown in the inset of Fig. \[fig:Fig-19\] is in accordance with the fact that $\lambda D$ is larger than unity for such $M$. For sufficiently large $M$ the deterministic branches dominate a whole sequence of the time evolution, and so $4t_0 \simeq D \propto M$ holds. These results consistently indicate that there exists a crossover between the regimes with and without elongation-contraction motions at around $M\cong 100$ in CFGE with the parameters studied, i.e., $E=0.032$ and $a_{\rm gel}=17$.
We have also examined the field dependence of $D$ and $1/\lambda$ for a fixed $M (=240)$, the results of which are shown in Fig. \[fig:E-dep-D\]. In strong fields $E \nge 0.06$, $DE$ becomes to be saturated, while $E/\lambda$ decreases with increasing $E$. As a consequence, $D\lambda$ is an increasing function of $E$, indicating that the deterministic branches, or in other words, the elongation-contraction motions survive and even dominate a whole evolution of the chain in sufficiently strong fields.
Lastly, in Fig. \[fig:Fig-13\], the averaged time evolution of individual peaks of $R_l(t)$ are plotted for various $M$ and $E$. The abscissa and ordinate are normalized by the period $\tau\equiv 4t_0$ and the averaged peak height, respectively. It is seen that with these normalizations, the data with different $M$ and $E$ almost lie on a universal curve during the period from $-0.6\tau$ to $0.4\tau$. The results obtained above such as $D \propto M, E^{-1}$, $\Delta
\propto D (\propto M)$ can be derived from the scaling behavior of $R_l(t)$ in Fig. \[fig:Fig-13\]. Such scaling behavior of $R_l(t)$ including the fact that the normalized peak exhibits an anti-symmetric shape is just what was observed experimentally. [@oana94] It was mentioned in Ref. that the ratio of slope of the $\Lambda$-shape immediately after the maximum to that before it is approximately constant ($\sim 3$) for all conditions they employed. The corresponding ratio of the BD result is $\sim 2$. The origin of this quantitative difference is not clear at the moment.
Discussions {#sec:discussions}
===========
We have performed extensive simulations on the DNA constant-field gel electrophoresis (CFGE) by means of a new Brownian dynamics (BD) method which we have recently developed. Our BD method is based on a [ *generalized*]{} bead-spring model for a polymer in a 3D space with rigid, regularly arrayed obstacles. By the word [*‘generalized’*]{} we mean that the ‘spring force’ is not linear but extremely nonlinear: it is infinite when the distance $l$ between two beads connected by the ‘spring’ is larger than $\sqrt{2}$, while it is zero when $l < \sqrt{2}$. Also in the model we strictly take into account of the excluded volume effects between any pair of beads whose diameter is put unity. These forces are supposed to be of a microscopic origin, i.e., molecular forces acting on monomers of a polymer. It should be emphasized that there involves no characteristic scale in these interactions except for the maximum distance between the connected beads ($\sqrt{2}$) relative to the bead diameter (unity).
In a vanishing field the viscosity $\zeta$ (or $\Delta t/\zeta$ with $\Delta t$ being the time unit of the BD dynamics) and the temperature $T$ in eqn. \[eqn:lange\] give rise to a characteristic scale of the model. The two parameters specify the viscous force and the random force from a solvent and are related to each other by eqn. \[eqn:chi\]. As is discussed in the accompanying paper II, the characteristic scale is the ‘persistence length’ of the chain. Actually, with $\zeta=10, \ \Delta t=1$ and $k_{\rm B}T_{\rm BD}=2$ the mean distance between two connected beads ($cl^*$ in eqn. \[eqn:L-scale\]) turns out to be about one six-th of the persistence length. Equating the latter to the persistence length of real DNA, the mean bead distance is roughly estimated as 10nm. Thus a bead in our BD model represents a potion of DNA consisting of a few tens of base-pairs. We may call our BD model with these values of parameters a bead-spring model of a semi-microscopic level. We will argue in paper III that the resultant chain dynamics, when it is coarse-grained in time and space, can be interpreted as dynamics of a charged, elastic string, and that the elasticity is due to the conformational entropy of the original, semi-microscopic chain.
A value of the electric field, $E=qE_b$ in eqn. \[eqn:lange\], is reasonably converted to that of the experimental CFGE on DNA as discussed in Sec. \[sec:mocuge\]. By the procedure adopted there the conversion of the field scale is essentially governed by that of the length scale as expressed in eqn. \[eqn:rel-E\]. The parameter $c$ in eqn. \[eqn:rel-E\] has been adjusted for the chains with different sizes, but within the range that the results are compatible with the condition $\sqrt{2} \ge \langle l\rangle \ge 1$ where $\langle l\rangle$ is the mean distance of connected beads. Then the $E$-dependence of mobility $\mu_{\rm BD}$ has been shown to agree semi-quantitatively with that of $\mu_{\rm exp}$ once absolute magnitudes of the simulational and experimental $\mu$’s are adjusted at the vanishing field limit. This procedure fixes the conversion of the time scales of the simulation and the experiment. To our knowledge, such a semi-quantitative coincidence between the simulation and the experiment has not been so far achieved.
We have also demonstrated that our BD simulation on CFGE in fields stronger than a certain crossover value reproduces the quasi-periodic evolution of a chain whose characteristic aspects are quite similar to those experimentally observed. Actually, as proposed by Oana [*et al.*]{} [@oana94] based on their experimental observations, a whole sequence of time evolution of the chain can be divided into two branches according to behavior of the longer radius of the chain $R_l(t)$. One is the [*deterministic*]{} branch where $R_l(t)$ exhibits a distinct peak structure which corresponds to the elongation-contraction motion of the chain, and the other is the [*stochastic*]{} branch between two neighboring deterministic branches where the chain exhibits a rather compact form.
We have studied statistics of time durations of the two branches and have found that, as assumed by Oana [*et al.*]{}, [@oana94] those of the deterministic branch obey a Gaussian distribution with mean $D$ and variance $\Delta$, while those of the stochastic branch do a Poisson distribution with the mean $1/\lambda$. Furthermore it is found that, in a fixed field, $D\propto M^1,\ \Delta \propto D$ and $1/\lambda\propto M^0$ for large $M$. This means that the deterministic branches become more dominant in a whole time sequence, and that the quasi-periodicity in the chain dynamics becomes more distinct (see Fig. \[fig:Fig-19\]) for larger $M$. These results, combined with the observation that the chain moves to the field direction by the distance proportional to $M$ in each deterministic branch, yield $\mu \propto M^0$. This explains why CFGE cannot separate DNA according to their size $L\ (\propto M)$ for large $L$. We have also checked that, for a fixed $M$ which is sufficiently large, the deterministic branches become dominant as $E$ is increased at least within the field range of our BD simulation (see Fig. \[fig:E-dep-D\]).
By the present BD simulation a crossover between the regimes with and without elongation-contraction motions is expected to occur for a chain with $M \simeq 100$ in $E=0.032$ and $a_{\rm gel}=17$. According to the length and field conversions described in Sec. \[sec:mocuge\], this roughly corresponds to DNA of 2kbp in $1\%$ agarose in $E_{\rm exp}
\simeq 4{\rm V/cm}$. Experimently, on the other hand, the direct measurement of the elongation-contraction motion by the fluorescent microscopy has been rather limited; the shortest is for T7 DNA (38kbp), [@lars95] but mostly for T4 DNA (166kbp) [@oana94] and T2 DNA (164kbp). [@lars95] But, there exist another experiment which strongly indicates the occurrence of the elongation-contraction motion, i.e., the antiresonance of mobility in the filed inversion gel electrophoresis. In such a experiment the antiresonance has been clearly observed for DNA with 9.42kbp. [@kdmo90] Therefore, our criterion mentioned above is considered to be a reasonable estimate for the elongation-contraction motion to occur.
Lastly two further comments are in order. In the whole parameter ranges we have examined, including the ones where the deterministic branches sufficiently dominate the stochastic ones, so-called ’hernias’-like configurations of a chain has been scarcely observed in the present BD simulation. This implies that they do not play a significant role in determining the saturation of $\mu$ at least for chains with moderate sizes we have studied. Another interesting observation has been already pointed out in Sec. \[sec:quasi-p\]: the front end of a chain moves in the field direction with almost a constant rate (Fig. \[fig:varfigs-new\]). In relation to this, we note that the average of $v_{\rm G}(t)$ over the deterministic branches and that over the stochastic branches are found to almost coincide with each other in a whole range of field we have examined. We shall interpret these results in paper III where the details of the chain dynamics in the deterministic branches is discussed.
The authors wish to thank M. Doi and Y. Masubuchi for useful discussions on their experimental and simulational results. The computation in this work has been done using the facilities of the Supercomputer Center, Institute for Solid State Physics, University of Tokyo, and those of the Computer Center of University of Tokyo.
References {#references .unnumbered}
==========
[31]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, , , , ****(), ().
, ****, ().
, ****(), ().
, ****, ().
, ** (, , ).
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****(), ().
, , , , , , , , ****, ().
, , , , , , ****(), ().
, , , ****(), ().
, , , , ****(), ().
, , , , , ****(), ().
, , , , , ****(), ().
, ().
, ****(), ().
, ().
, ****, ().
, , , , , ****(), ().
, ****, ().
, ****, ().
, ****, ().
, ****(), ().
, ****, ().
, , , in **, edited by , , , , , (, , ), pp. .
, ****, ().
\
\
\
\
\
\
\
\
\
\
\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We construct a spectral representation of neutrino propagator in moving matter or in external magnetic field. In both cases there exists fixed four-dimensional axis of polarization, such that the corresponding spin projectors commute with propagator. As a result, all eigenvalues of propagator and, consequently, dispersion laws for neutrino in media are classified according to spin projection onto this axis. Use of the found spin projectors simplifies essentially the eigenvalue problem and allows to build spectral representation of propagator in moving matter or external magnetic field in analogy with the vacuum propagator.'
author:
- 'A.E. Kaloshin'
- 'D.M.Voronin'
title: Neutrino propagation in media and axis of complete polarization
---
Introduction {#intro}
============
Neutrino physics, actively developing in last decades, is related with a wide spectrum of physical problems, including the astrophysical ones. Propagation of neutrinos in a dense matter or magnetic field leads to modification of neutrinos oscillation picture and appearance of new effects. The most prominant effect in neutrinos passing through matter is related with resonance amplification of oscillations (MSW effect) [@q16; @m5656], which solves the solar neutrino problem, see reviews [@q8]-[@Bil18].
After it a huge interest was generated to further investigations of different properties of media and its influence on fine aspects of flavor oscillations. Different methods were developed for solution of corresponding equations in media with varied density [@Haxton:1986dm]-[@Balantekin:1997jp]. The movement of matter and its polarization, which can arise in magnetic field were taken into account [@Nunokawa:1997dp]-[@EstebanPretel:2008ni]. Interesting results were obtained in investigations of spin dynamics in matter and transitions between different spin states [@Lobanov:2001ar], [@Dvornikov:2002rs], [@Lobanov:2002ur]. As for applications in astrophysics, there exists a variety of conditions for neutrinos propagation: in Earth, Sun, in vicinity of Supernova – see reveiws [@Duan:2009cd], [@Volpe:2016bkp]. We mentioned here only some aspects of investigations and only small part of relevent publications.
Possibility for neutrino to have an anomalious magnetic moment and its experimental manifestations was discussed for a long time [@Cis71]-[@Ego00]. In Standard Model (SM) the magnetic moment of neutrino is arised due to loop corrections and is proportional to neutrino mass. It leads to extreme smallness of neutrino magnetic moment in SM, so the present-day interest for this subject (both in theory and experiment) is related first of all with search of new physics beyond the SM [@Zhi11; @Giu15].
Most transparent way to describe mixing and oscillations phenomena in neutrinos system is to use the quantum-mechanical equations (Schrödinger or Dirac), but the most justified is the Quantum Field Theory (QFT) approach, where production, propagation and detection of neutrino looks like a macroscopic Feynman diagram[@Gri96]-[@Dvo11]. The necessary element of such description is the neutrino propagator.
In the present paper we build a spectral representation of neutrino propagator in matter moving with constant velocity or in constant homogenious magnetic field [^1]. In this representation based on the eigenvalue problem a propagator looks as a sum of single poles, accompanied by orthogonal matrix projectors. Such form of propagator gives the simplest and most convenient algebraic construction and means in fact a complete diagonalization. A spectral representation was discussed earlier for dressed fermion propagator in theory with parity violation [@q12] and for dressed matrix propagator with mixing of few fermionic fields [@q98]. Note that the problem of the neutrino propagator in the presence of matter, including account for possible effect of matter motion, was also discussed in [@Pivovarov:2005cu] (see also [@Studenikin:2008qk]).
In solving the eigenvalue problem for neutrino propagator in media, we find a new aspect, related with polarization of neutrino: there exist spin projectors with fixed polarization 4-vector, commuting with propagator. The properties of these spin projectors (both in matter and magnetic field) allows to reduce the algebraic problem for media to the vacuum case, properties of media modify only scalar coefficients of matrix equations.
In Section 2 we construct the spectral representation of neutrino propagator in a matter, moving with constant velocity. The key moment is the presence of generalized spin projectors (\[Sigma\]), commuting with propagator, which allows to simplify the eigenstate problem (\[eq:5\]) and to get answer for inverse propagator of the most general view (\[formula2\]). We discuss also the particular case of propagator in framework of Standard Model, in this case it is easy to see that the spin projection on the complete polarization axis $z^\mu$ is not conserved.
In Section 3 the spectral representation is build for neutrino propagator in a constant magnetic field. In this case there exists the fixed axis of complete polarization $z^\mu$ and again the corresponding spin projector $\Sigma(z)$ commutes with propagator. This property allows to use the same trick (reducing of number of $\gamma$-matrix structures “under observation” of spin projectors) to obtain an analytical expressions for eigenvalues and eigenprojectors.
In Appendix the main facts on spectral representation for matrix of general form and some details of this representation for fermion propagator in vacuum and media are collected.
Propagator in moving matter and spin projectors {#sec:nonrel}
===============================================
Axis of complete polarization and basis
---------------------------------------
When considering a neutrino oscillations in frameworks of quantum field theory, the central object is the neutrino propagator. In media there exist two 4-dimensional vectors: momentum of particle $p$ and matter velocity $u$, so altogether there exist eight $\gamma$-matrix structures in decomposition of propagator, if parity is not conserved. Most general expression for inverse propagator can be written as $$\begin{aligned}
\label{formula2}
S(p,u)=G^{-1} =s_{1}I+s_{2}\hat{p}+s_{3}\hat{u}+s_{4}\sigma^{\mu\nu}p_{\mu}u_{\nu}+\nonumber\\+s_{5}i\varepsilon^{\mu\nu\lambda\rho}\sigma^{\mu\nu}u_{\lambda}p_{\rho}+s_{6}\gamma^{5}+s_{7}\hat{p}\gamma^{5}+s_{8}\hat{u}\gamma^{5},\end{aligned}$$ where $s_{i}$ are scalar function dependent on invariants.
Below we will solve the eigenvalue problem for inverse propagator of general form. As a starting point it is convenient to introduce $\gamma$-matrix basis with simple multiplicative properties.
Firts of all, let us introduce the four-vector $z^{\mu}$, which is a linear combination of two vectors $p$, $u$ and has properties of fermion polarization vector [^2]: $$\label{123sa}
z^{\mu}p_{\mu}=0, \ \ \ z^{2}=-1.$$
Orthogonal to momentum combination is $$\label{vec_z}
z^{\mu}=b\ (p^{\mu}(up)-u^{\mu}p^{2}),$$ where $b$ is the normalization factor, $b=[p^{2}((up)^{2}-p^{2})]^{-1/2}$.
To clarify its properties, let us consider firstly the rest matter $({\bf{u}}=0,u^{0}=1)$. In this case $$z^{\mu}=b({\bf{p}}^{2},p^{0}{\bf{p}})$$ and its square is $$z_{\mu}z^{\mu}=-b^{2}{\bf{p}}^{2}(p_{\nu}p^{\nu}).$$
Thus we see, that for time-like momentum $p^{\mu}$ the vector $z^{\mu}$ is space-like one [^3].
Then, having the vector $z$, one can construct the generalized off-shell spin projectors [^4]: $$\label{Sigma}
\Sigma^{\pm}=\frac{1}{2}(1\pm\gamma^{5}\hat{z}\hat{n}),~~~ \Sigma^{\pm}\Sigma^{\pm}=\Sigma^{\pm},~~~\Sigma^{\pm}\Sigma^{\mp}=0,$$ where $n^{\mu}=p^{\mu}/W,\ \ W=\sqrt{p^{2}}$. The appeared matrix $\gamma^{5}\hat{z}\hat{n}$ may be rewritten as $$\label{}
\gamma^{5}\hat{z}\hat{n}=\gamma^{5} \sigma^{\alpha\beta}z_{\alpha}n_{\beta}=-bW{\gamma^{5}}\sigma^{\alpha\beta}u_{\alpha}p_{\beta},\ \ \ \ \ \sigma^{\alpha\beta}=\frac{1}{2}[\gamma^{\alpha}, \gamma^{\beta}].$$
After it one can see that $\Sigma^{\pm}$ commute with all $\gamma$-matrices in decomposition of inverse propagator (\[formula2\]). Multiplying the inverse propagator $S(p,u)$ (\[formula2\]) by unit matrix $$S=(\Sigma^{+}(z)+\Sigma^{-}(z))S \equiv S^+ + S^- ,$$ one obtains two orthogonal contributions $S^+, S^-$.
One more useful property of $\Sigma^{\pm}$ is that “under observation” of the spin projectors (i.e. in $S^+, S^-$ terms) $\gamma$-matrix structures may be simplified. Namely:\
$\gamma$-matrices, which contain the matter velocity $u^{\mu}$ may be transformed to the set of four matrices without velocity: $I, \hat{p}, \gamma^{5}, \hat{p}\gamma^{5}$. For example, one can rewrite the term $\hat{u}$ in (\[formula2\]) as a linear combination $\hat{p}$ and $\hat{z}$ and to use the projector property $(\Sigma^{+}\cdot{\gamma^{5}}\hat{z}\hat{n}=\Sigma^{+})$: $$\Sigma^{+}\hat{u}=\Sigma^{+}(a_1\hat{p}+a_2\hat{z})=\Sigma^{+}(z)(a_1\hat{p}-\frac{a_2}{W}\hat{p}\gamma^{5}).$$
After this simplification we have the vacuum set of Dirac matrices $I, \hat{p}, \gamma^{5}, \hat{p}\gamma^{5}$ and it’s convenient to introduce the off-shell momentum projections: $$\Lambda^{\pm}=\frac{1}{2}(1\pm\hat{n}),~~~n^{\mu}=\frac{p^{\mu}}{W}$$ orthogonal to each other.
Having the momentum $\Lambda^{\pm}$ and spin projectors $\Sigma^{\pm}$, one can build the basis, which will be used below in the eigenvalue problem $$\begin{aligned}
\label{1}
R_{1}=\Sigma^{-}\Lambda^{+},~~~~~~R_{5}=\Sigma^{+}\Lambda^{+},~~~\nonumber\\
R_{2}=\Sigma^{-}\Lambda^{-},~~~~~~R_{6}=\Sigma^{+}\Lambda^{-},~~~\nonumber\\
R_{3}=\Sigma^{-}\Lambda^{+}\gamma^{5},~~~R_{7}=\Sigma^{+}\Lambda^{+}\gamma^{5},\nonumber\\
R_{4}=\Sigma^{-}\Lambda^{-}\gamma^{5},~~~R_{8}=\Sigma^{+}\Lambda^{-}\gamma^{5}.\end{aligned}$$ Multiplicative properties of the basis (\[1\]) are presented in Table 1, where column elements multiply from left the row elements.
The inverse propagator (\[formula2\]) may be written as decomposition in this basis $$\label{decomp}
S(p,u)= \sum_{i=1}^{4} R_i S_i (p^2,pu) + \sum_{i=5}^{8} R_i S_i (p^2,pu),$$ where these two sums are orthogonal to each other.
$R_{1}$ $R_{2}$ $R_{3}$ $R_{4}$ $R_{5}$ $R_{6}$ $R_{7}$ $R_{8}$
--------- --------- --------- --------- --------- --------- --------- --------- --------- --
$R_{1}$ $R_{1}$ 0 $R_{3}$ 0 0 0 0 0
$R_{2}$ 0 $R_{2}$ 0 $R_{4}$ 0 0 0 0
$R_{3}$ 0 $R_{3}$ 0 $R_{1}$ 0 0 0 0
$R_{4}$ $R_{4}$ 0 $R_{2}$ 0 0 0 0 0
$R_{5}$ 0 0 0 0 $R_{5}$ 0 $R_{7}$ 0
$R_{6}$ 0 0 0 0 0 $R_{6}$ 0 $R_{8}$
$R_{7}$ 0 0 0 0 0 $R_{7}$ 0 $R_{5}$
$R_{8}$ 0 0 0 0 $R_{8}$ 0 $R_{6}$ 0
: Multiplicative properties of the matrix basis (\[1\])[]{data-label="lklhjlksd"}
It can be seen from Table 1, that with use of the basis (\[1\]) the eigenvalue problem for inverse propagator (\[decomp\]) is separeted into two different problems: one for $R_{1}..R_{4}$ and another for $R_{5}..R_{8}$. Every problem has two different eigenvalues.
Spectral representation of propagator in matter {#sec:rel}
-----------------------------------------------
Let us recall that in quantum mechanics the term spectral representation of linear hermitian operator $\hat{A}$ means the following representation [@q10] $$\hat{A}=\sum\lambda_{i}|i\rangle\langle i| =\sum\lambda_{i} \Pi_{i},$$ which contains the eigenvalues $\lambda_{i}$ and eigenprojectors $\Pi_{i}=|i\rangle\langle i|$. $$\hat{A}|i\rangle=\lambda_{i}|i\rangle.$$
Orthonornality of the vector system leads to property of orthogonality of projectors $$\Pi_{i}\Pi_{k}=\delta_{ik}\Pi_{k} .$$ If an operator is not hermitian, to build a spectral representation one needs to solve two eigenvalue problems: left and right ones (see details in \[sec:eigen\]).
We want to construct a spectral representation for inverse propagator of general form (\[formula2\]), (\[decomp\]), so we should solve the eigenvalue problem $$\label{eq:5}
S\Pi_i = \lambda_i \Pi_i .$$
Note that we are solving eigenvalue problem in a matrix form, i.e. from the begining we are looking for eigenprojectors $\Pi_i$ instead of eigenvectors. It can be done with use of $\gamma$-matrix basis and it allows to avoid cumbersome intermediate formulae. As for non-hermi-\
tiance of fermion propagator: it is enough to solve the left problem and to require the orthogonality of projectors, see [@q12].
When we solve for this problem, we get the spectral representation of inverse propagator in moving matter: $$\label{RS}
S(p,u)=\sum_{i=1}^{4} \lambda_{i}\Pi_{i}.$$
If the eigenprojectors set is the complete orthogonal system, then propagator is easily obtained by reversing of (\[RS\]) and looks very simple $$\label{spec_G}
G(p,u)=\sum_{i=1}^{4} \frac{1}{\lambda_{i}}\Pi_{i} ,$$ as a sum of single poles, accompanied by corresponding orthogonal projectors.
The use of the matrix basis (\[1\]) simplifies essentially solution of eigenvalue problem. The eigenprojects also may be found in form of decomposition in this basis, orthogonality of spin projectors (see Table 1) leads to more simple problems, where only first or second quartet in (\[1\]) is involved. It was noted in above that “under observation” of the spin projectors all gamma-matrices in (\[formula2\]) turn into the set $I, \hat{p}, \gamma^{5}, \hat{p}\gamma^{5}$. So, for example, the eigenstate problem for first quartet of basis elements $$\label{RS1}
\left( \sum_{k=1}^{4}R_{k} S_{k} \right)\cdot \left( \sum_{i=1}^{4}R_{i} A_{i} \right) = \lambda \left( \sum_{i=1}^{4}R_{i} A_{i} \right)$$ coinsides in fact with the eigenstate problem for dressed vacuum propagator with parity violation [@q12]. The presence of matter leads only to appearence of spin projector in (\[1\]) (do not changing an algebra) and modification of scalar coefficients. Besides, as compared with the vacuum case, there appears twice as much eigenvalues, which arise from two different square equations.
Repeating the algebraic operations from [@q12], one can write an answer for eigenvalue problem in most general case. Eigenvalues and eigenprojectors are looking as: $$\begin{aligned}
\lambda_{1,2}=\frac{S_{1}+S_{2}}{2}\pm\sqrt{\Big(\frac{S_{1}-S_{2}}{2}\Big)^{2} + S_{3}S_{4}}~, \nonumber\\
\label{eig_1}
\Pi_{1}=\frac{1}{\lambda_{2}-\lambda_{1}}\Big((S_{2}-\lambda_{1})R_{1}+(S_{1}-
\lambda_{1})R_{2}-\nonumber\\-S_{3}R_{3}-S_{4}R_{4}\Big), \\
\Pi_{2}=\frac{1}{\lambda_{1}-\lambda_{2}}\Big((S_{2}-\lambda_{2})R_{1}+(S_{1}-
\lambda_{2})R_{2}-
\nonumber\\-S_{3}R_{3}-S_{4}R_{4}\Big),\nonumber \\ \lambda_{3,4}=\frac{S_{5}+S_{6}}{2}\pm\sqrt{\Big(\frac{S_{5}-S_{6}}{2}\Big)^{2} + S_{7}S_{8}}~, \nonumber\\
\label{eig_2}
\Pi_{3}=\frac{1}{\lambda_{4}-\lambda_{3}}\Big((S_{6}-\lambda_{3})R_{5}+(S_{5}-
\lambda_{3})R_{6}-\nonumber\\-S_{7}R_{7}-S_{8}R_{8}\Big), \\
\Pi_{4}=\frac{1}{\lambda_{3}-\lambda_{4}}\Big((S_{6}-\lambda_{4})R_{5}+(S_{5}-
\lambda_{4})R_{6}-\nonumber\\-S_{7}R_{7}-S_{8}R_{8}\Big)\nonumber.\end{aligned}$$
Here $S_{i}$ are the coefficients of decomposition of inverse propagator in the basis (\[decomp\]). Recall that the indexies $1,2$ refers to $S^{-}$ (i.e to first quartet in (\[decomp\]), and $3,4$ to contribution $S^{+}$.
The obtained eigenprojectors have the following properties:
1. $S\Pi_{k}=\lambda_{k}\Pi_{k}$, k=1 …4 ,
2. $\Pi_{i}\Pi_{j}=\delta_{ij}\Pi_{j}$,
3. $\sum\limits_{i=1}^4\Pi_{i}=1$.
The introduced by us four-vector $z^{\mu}$ (\[vec\_z\]) plays role of the complete polarization axis and all eigenvalues are classified by the projection of spin onto this axis. In contrast to vacuum, this axis is not arbitrary – see some details in \[sec:prop\]. As it will be seen from discussion of Standard Model case, the projection on this axis is not conserved in general case.
Standard Model propagator
-------------------------
In the case of SM a fermion propagator in matter looks like: $$\label{dddf}
S(p,u) = \hat{p} - m-\alpha\hat{u}(1-\gamma^{5}),$$ where $\alpha$ is some constant dependent on properties of media and flavour.
Note, that $\alpha$ and mass term are in fact matrices $n \times n$ because of mixing of neutrinos. But here we consider propagation of neutrino of one type, so in this approximation the propagator will contain only a diagonal element of a flavour matrix. For example, in case of electron neutrino [@Kuo89] $$\alpha^{(\nu_{e})}= \frac{G_{F}}{\sqrt{2}}(n_{e}(1+4\sin^{2}\theta_{W})+n_{p}(1-4\sin^{2}\theta_{W}) - n_{n}),$$ where $n_{e}, n_{p}, n_{n}$ are densities of matter particles, $\theta_{W}$ is the Weinberg angle.
Let us write down the coefficients of decomposition in two bases: $\gamma$-matrix (\[formula2\]) and $R$-basis (\[decomp\]): $$\begin{aligned}
\label{vvbnmkl}
s_{1}=-m,~~~S_{1}=-m+W(1+K^{+});\nonumber\\
s_{2}=1,~~~~~~~S_{2}=-m-W(1+K^{+});\nonumber\\
s_{3}=\alpha,~~~~~~~S_{3}=-WK^{+};~~~~~~~~~~~~~~\nonumber\\
s_{4}=0,~~~~~~~~S_{4}=WK^{+};~~~~~~~~~~~~~~~~\nonumber\\
s_{5}=0,~~~~~~~S_{5}=-m-W(1+K^{-});\nonumber\\
s_{6}=0,~~~~~~~S_{6}=-m+W(1+K^{-});\nonumber\\
s_{7}=0,~~~~~~~~S_{7}=WK^{-};~~~~~~~~~~~~~~~~\nonumber\\
s_{8}=-\alpha,~~~~~S_{8}=-WK^{-}.~~~~~~~~~~~~~~\end{aligned}$$ Here the following notations are introduced: $K^{\pm}=-\alpha \Big((pu)\pm\sqrt{(up)^{2}-W^{2}}\Big)/W^{2}$, $W=\sqrt{p^{2}}$.
The solutions of the eigenvalue problem (\[eq:5\]) in this case have the form: $$\lambda_{1,2}=-m \pm W\sqrt{1+2K^{+}}, \nonumber$$ $$\label{lamSM}
\lambda_{3,4}=-m \pm W\sqrt{1+2K^{-}},$$ $$\Pi_{1,2}=\Sigma^{-}\cdot \frac{1}{2} \left[1\pm \hat{n}\ \frac{1+K^{+} - \gamma^{5}K^{+}}{\sqrt{1+2K^{+}}} \right], \nonumber$$ $$\label{proSM}
\Pi_{3,4}=\Sigma^{+}\cdot \frac{1}{2} \left[1\pm \hat{n}\ \frac{1+K^{-} - \gamma^{5}K^{-}}{\sqrt{1+2K^{-}}} \right].$$
Vanishing of eigenvalues (\[lamSM\]) gives the dispersion equation for neurino in moving matter $$\label{dispSM}
p^2-m^2 -2\alpha \left( (up) -s \sqrt{(up)^2 -p^2}\right) = 0 ,$$ where $s=\pm 1$ is the spin projection on the axis $z$ (\[vec\_z\]). If we restrict ourselves by the first order in $G_F$, then we have from (\[dispSM\]) for solution with positive energy $$\label{dispSM+}
p^0 = \varepsilon + \frac{\alpha}{\varepsilon}\left[ (u^0\varepsilon - {\bf u}{\bf p}) -s \sqrt{(u^0\varepsilon - {\bf u}{\bf p})^2 - m^2}\right] ,$$ where $\varepsilon = \sqrt{{\bf p}^2 +m^2}$. Solution for negative energy is obtained by substitution $\varepsilon \to -\varepsilon$. The dispersion law (\[dispSM\]) coincides with equation obtained in [@Pivovarov:2005cu], [@Studenikin:2008qk] at investigation of neutrino propagator in moving media.
In case of Standard Model it is easy to convince yourself that the spin projection on the axis of complete polarization is not a conservative value. The Hamiltonian is defined by Dirac operator (\[dddf\]) $$H=p^{0}-{\gamma^{0}}S.$$ We can use a known zero commutator $$[R,S]=0,~~~~R={\gamma^{5}}\hat{z}\hat{n},$$ for simple calculation of commutator $R$ with Hamiltonian $$[R,H]={\gamma^{0}}[S,R]+[{\gamma^{0}},R]S=[{\gamma^{0}},R]S,$$ which may be reduced to $[{\gamma^{0}},R]$. With use of the standard representation of $\gamma$-matrices we have $$R=
\begin{pmatrix}
{\bm \sigma}\bf{v}&-i{\bm \sigma}{\bm \xi}\\
-i{\bm \sigma}{\bm \xi}&{\bm \sigma}{\bf{v}}
\end{pmatrix}, ~~{\bf{v}}=n^{0}{\bf{z}}-z^{0}{\bf{n}}, ~~{\bm \xi}=[{\bf{z}}\times {\bf{n}}].$$
If to require $[{\gamma^{0}},R]=0$, we come to condition ${\bm \xi}=0$, i.e. $$\label{conse}
{\bm \xi}\equiv[{\bf{z}}\times {\bf{n}}]=bW[{\bf{p}}\times{\bf{u}}]=0.$$ Thus, a spin projection on the axis $z^{\mu}$ is conserved only in the case ${\bf{u}}_{\perp}=0$, when 3-momentum of propagator coinsides in direction (or opposite) with matter velocity. In this case the found polarization vector $z^\mu$ (\[vec\_z\]) takes the form $$\label{heli}
z^{\mu}=\frac{1}{W} \left( | \mbox{\bf{p}} | ,\ p^{0}\frac{\bf{p}}{|\bf{p}|} \right),$$ which corresponds to helicity state of fermion, but the off-shell one since $W\not= m$.
In general case, at arbitrary direction of matter velocity, in spite of $[\Sigma^{\pm},S]=0$, the spin projection on the axis $z^{\mu}$ is not conserved: $[\Sigma^{\pm},H]\neq 0$. Evidently, that for propagator of more general form than the Standard Model one, the spin projection on the axis of complete polarization $z$ also is not a conservative value.
### Rest matter case
Let us consider in detail a particular case of SM propagator (\[dddf\]), when matter is in the rest (${\bf{u}}=0, u_{0}=1$). In this case, according to Eq. (\[conse\]), spin projection is concerved and polarization vector $z^\mu$ also corresponds to helicity state (\[heli\]).
Straight calculation shows that the generalized spin projectors (\[Sigma\]) in this case are projectors onto the spatial momentum direction $$\Sigma^{\pm}=\frac{1}{2}\Big(1\pm{\bm \Sigma}\frac{\bf{p}}{|\bf{p}|}\Big),~{\bm \Sigma}=\gamma^{0}{\bm \gamma}\gamma^{5}.$$
For the rest matter the eigenvalues and eigenprojectors are particular case of expressions (\[lamSM\]), (\[proSM\]) and look as follows: $$\lambda_{1,2}=-m \pm W\sqrt{1-\frac{2\alpha}{W^{2}}(E+|\bf{p}|)},$$ $$\lambda_{3,4}=-m \pm W\sqrt{1-\frac{2\alpha}{W^{2}}(E-|\bf{p}|)},$$ $$\label{vvvbgrtewsdf}
\Pi_{1}=\frac{1}{4}(1-{\bm \Sigma}\frac{\bf{p}}{|\bf{p}|})\Big(1+\frac{\hat{n}}{B^{+}}[1-\frac{\alpha(E + |\bf{p}|)}{W^{2}}(1-\gamma^{5})]\Big),$$ $$\Pi_{2}=\frac{1}{4}(1-{\bm \Sigma}\frac{\bf{p}}{|\bf{p}|})\Big(1-\frac{\hat{n}}{B^{+}}[1-\frac{\alpha(E + |{\bf{p}}|)}{W^{2}}(1-\gamma^{5})]\Big),$$ $$\Pi_{3}=\frac{1}{4}(1+{\bm \Sigma}\frac{\bf{p}}{|\bf{p}|})\Big(1+\frac{\hat{n}}{B^{-}}[1-\frac{\alpha(E - |\bf{p}|)}{W^{2}}(1-\gamma^{5})]\Big),$$ $$\Pi_{4}=\frac{1}{4}(1+{\bm \Sigma}\frac{\bf{p}}{|\bf{p}|})\Big(1-\frac{\hat{n}}{B^{-}}[1-\frac{\alpha(E - |{\bf{p}}|)}{W^{2}}(1-\gamma^{5})]\Big),$$ where $B^{\pm}=\sqrt{1-\frac{2\alpha}{W^{2}}(E \pm |\bf{p}|)}$.
Thus, for the rest matter the well-known fact [@Man87; @Pan92] is reproduced that neutrino with definite helicity has a definite law of dispersion in matter.
If some eigenvalue is vanished, we obtain a dispersion relation – energy and momentum connection. We have for $\lambda_{1,2}$ $$E^{2}-2\alpha E -m^{2}-{\bf{p}}^{2}-2\alpha |{\bf{p}}|=0,$$ $$E_{1,2}=\alpha\pm\sqrt{(|{\bf{p}}|+\alpha)^{2}+m^{2}},$$ and for $\lambda_{3,4}$: $$E^{2}-2\alpha E -m^{2}-{\bf{p}}^{2}+2\alpha |{\bf{p}}|=0,$$ $$E_{3,4}=\alpha\pm\sqrt{(|{\bf{p}}|-\alpha)^{2}+m^{2}},$$
The propagation of neutrino in an external magnetic field
=========================================================
In previous section we found that in moving matter there exists an axis of complete polarization $z^{\mu}$ (\[vec\_z\]), and corresponding spin projectors (\[Sigma\]) commute with the propagator. A similar situation arises when neutrino propagates in a magnetic field.
An inverse propagator of a neutral fermion with an anomalous magnetic moment $ \mu $ in a constant external electromagnetic field is as follows: $$\label{proppp}
S=\hat{p}-m-\frac{i}{2}\mu\sigma^{\alpha\beta}F_{\alpha\beta}, \ \ \ \ \ \ \sigma^{\alpha\beta}=\frac{1}{2}[\gamma^{\alpha},\gamma^{\beta}],$$ where $F_{\alpha\beta}$ is the electromagnetic field tensor. In the case of a magnetic field, it takes more customary form: $$\label{new_propagator}
S=\hat{p}-m+\mu{\bm \Sigma}{\bf{B}},~~~~ {\bm \Sigma}={\gamma^{0}}{\bm \gamma}{\gamma^{5}}.$$
Having electromagnetic field tensor and 4-momentum, we can construct a polarization vector $z^{\mu}$ ($ z^{2}=-1$ and $z_{\mu}p^ {\mu}=0$): $$\label{zB}
z^{\mu}=b\epsilon^{\mu\nu\lambda\rho}F_{\nu\lambda}p_{\rho},~~~~~~~ {b}=(p_{0}^{2}{\bf{B}}^{2}-({\bf{p}}{\bf{B}})^{2})^{-1/2}.$$
Using this vector[^5] we can construct a spin projector with the same properties as in the case of neutrino propagation in a matter:
$$\label{sb}
\Sigma^{\pm}=\frac{1}{2}(1 \pm \gamma^{5}\hat{z}).$$
It is easy to see that the spin projectors commute with the inverse propagator (\[new\_propagator\]). To this end, let us write down the particular case of the vector $z^{\mu}$ in magnetic field $$\label{zB1}
z^{\mu}=b(({\bf{B}}{\bf{p}}),p^{0}{\bf{B}}),~~~~
b=(p_0^2{\bf B}^2 - ({\bf{B}}{\bf{p}})^2)^{-1/2}$$ and matrix ${\gamma^{5}}\hat{z}$ may be rewritten as $$\label{R_equi}
R \equiv {\gamma^{5}}\hat{z}=b({\gamma^{5}}{\gamma^{0}}({\bf{B}}{\bf{p}})+p^{0}{\gamma^{0}}({\bm \Sigma}{\bf{B}})),~~~~R^{2}=1.$$ After this, it is easy to see that $[S,\Sigma^{\pm}]=0$.
Further we can apply the same trick that was used for the propagator in a matter: it was found that “under observation” of the spin projector, the gamma-matrix structures are simplified. So we can act by unit matrix composed of spin projectors onto the inverse propagator and to obtain two orthogonal terms $S^+$ and $S^-$: $$S=(\Sigma^{+}(z)+\Sigma^{-}(z))S \equiv S^+ + S^-.$$
Since two matrices commute $[S,R]=0$, they have a common eigenvector: $$S\Psi = \lambda\Psi ,\ \ \ {\gamma^{5}}\hat{z}\Psi =\sigma\Psi ,\ \ \ \sigma=\pm1.$$
The eigenvector of the operator $R$ is obvious: $\Psi^{\pm}=\Sigma^{\pm} \Psi_0$, therefore the system looks like this: $$S^{\pm}\Psi^{\pm}=\lambda\Psi^{\pm}, \ \ \ \ \ {\gamma^{5}}\hat{z}\Psi^{\pm}=\pm \Psi^{\pm} .$$
Since the eigenvalues of the matrix $R$ are equal to $\pm1$, from (\[R\_equi\]) we can find the useful relation $$({\bm \Sigma}{\bf{B}})\Psi^{\pm}=\frac{1}{p^{0}}({\gamma^{5}}({\bf{p}}{\bf{B}}) \pm {\gamma^{0}}\frac{1}{b}) \Psi^{\pm}.$$ Then, in analogy with the case of matter, in the $S^{\pm}$ contributions the $\gamma$-matrix structure can be transformed. Instead of (\[new\_propagator\]) we get $$\label{s_plus_minus}
S^{\pm}=\Sigma^{\pm}(z)\Big[\hat{p}-m+\frac{\mu}{p^{0}}({\gamma^{5}}({\bf{p}}{\bf{B}}) \pm {\gamma^{0}}\frac{1}{b})\Big].$$
Let us recall that for covariant matrix of the form $$\label{cov_m}
S=aI+b\hat{p}+c{\gamma^{5}}+d\hat{p}{\gamma^{5}}$$ solutions of the matrix eigenvalue problem are known [@q12] and were used for the propagator in matter (\[eig\_1\]).
The inverse propagator in the external field (\[new\_propagator\]), (\[s\_plus\_minus\]) is non-covariant (in particular, it contains $\gamma^0$), but for algebraic problem this is not so important. At solving of eigenvalue problem with the matrix (\[cov\_m\]), the momentum vector $p^{\mu}$ may be changed by any four numbers. Therefore, if we redefine the vector $p^{\mu}$ in $S^{\pm}$, we can get rid of $\gamma^0$ and use the ready answer for eigenvalues and eigenprojectors.
So, we can introduce “4-vector” $$p_{\pm}^\mu= (p^{0} \pm \frac{\mu}{bp_{0}}, \ \bf{p} )$$ and after this, the inverse propagator takes the form: $$\label{obr_prop}
S^{\pm}=\hat{p}_{\pm}-m+\mu{\gamma^{5}}\frac{({\bf{B}}{\bf{p}})}{p_{0}},$$ in which there are only $I$, $\hat{p}_{\pm} $ and ${\gamma^{5}}$ matrix, and which is algebraically similar to the vacuum propagator. Therefore, we can use the formulas (\[eig\_1\]) for eigenvalues and eigenprojectors: $$\begin{aligned}
\label{field}
\lambda_{1}^{\pm}=-m + \sqrt{W_{\pm }^{2}+\frac{\mu^{2}}{p_{0}^{2}}({\bf{B}}{\bf{p}})^{2}},\\
\lambda_{2}^{\pm}=-m - \sqrt{W_{\pm }^{2}+\frac{\mu^{2}}{p_{0}^{2}}({\bf{B}}{\bf{p}})^{2}},\\
\Pi_{1}^{\pm}=\frac{\Sigma^{\pm}}{2}\Big(1-\frac{1}{A^{\pm}}(\hat{p}_{\pm}+\frac{\mu ({\bf{B}}{\bf{p}})}{p_{0}}{\gamma^{5}})\Big),\\
\Pi_{2}^{\pm}=\frac{\Sigma^{\pm}}{2}\Big(1+\frac{1}{A^{\pm}}(\hat{p}_{\pm}+\frac{\mu ({\bf{B}}{\bf{p}})}{p_{0}}{\gamma^{5}})\Big).\end{aligned}$$ We introduced here the notations: $W_{\pm}=\sqrt{p^{2}_{\pm}}$, $A^{\pm}=\sqrt{W_{\pm }^{2}+\mu^{2} ({\bf{B}}{\bf{p}})^{2}/p_{0}^{2}}$. If the eigenvalue is vanishing, we can obtain the well-known dispersion law for movement of anomalous magnetic moment in magnetic field [@TBKh; @Bag90] $$E^2 = m^2+ {\bf{p}}^2 + \mu^2 {\bf{B}}^2 \pm 2\mu \sqrt{m^2 {\bf{B}}^2 + {\bf{p}}^2 {\bf{B}}^2_\perp}.$$ Here $\pm$ corresponds to different signs in , i.e. to terms $S^\pm$ in propagator, which are accompanied by spin projectors $\varSigma^{\pm}$.
The spectral representation of the inverse propagator with found eigenvalues and eigenprojectors can be written as: $$S=\sum_{i=1}^{2}\lambda_{i}^{+}\Pi_{i}^{+}+\sum_{i=1}^{2}\lambda_{i}^{-}\Pi_{i}^{-}.$$
So, in constant magnetic field all eigenvalues are classified by the spin projection on the fixed axis $z$ (\[zB1\]). It turns out that, as in the case of moving medium, the projection on this axis, in general, is not a conserved quantity.
The inverse propagator (\[new\_propagator\]) may be connected with the Dirac Hamiltonian $$S=\gamma^0 (p^0 - H_D),\ \ \ \ \ H_D= {\bm \alpha}{\bf p} + \beta m + \mu \gamma^0 (\bf{\Sigma} \bf{B}) .$$ Using the zero commutator of the matrix $R=\gamma^5\hat{z}$ with the inverse propagator $$0 = [R,S] = \gamma^0 [H_D, R] + [R, \gamma^0] (p^0 - H_D) ,$$ we can reduce the case to the commutator $[R, \gamma^0]$. Calculating it in the standard representation of gamma-matrices, we have $$[\gamma^5 \hat{z}, \gamma^0] =
\begin{pmatrix}
0 & 2z^0\\
2z^0 & 0
\end{pmatrix},\ \ \ \ \ z^0=b\ ({\bf{B}}{\bf{p}}).$$ So we see that the projection of the spin on the axis of complete polarization (\[zB\]) is conserved only in case of a transverse magnetic field.
Conclusions
===========
In the present paper we have built the spectral representation of neutrino propagator both in a moving matter and in a constant external magnetic field. In this form (\[RS\]), (\[spec\_G\]), which is based on the eigenvalue problem for inverse propagator (\[eq:5\]), the propagator looks like a sum of poles accompanied by own $\gamma$-matrix orthogonal projectors. The advantage of this representation is that a single term in this sum is related only with one dispersion law for particle in media. More exactly, relation of energy and momentum appears as a result of vanishing of one of eigenvalues $\lambda_i=0$ in (\[spec\_G\]).
It turned out that both in matter and in magnetic field there exists the fixed 4-axis of complete polarization $z^\mu$, such that all eigenvalues of propagator (and, consequently, dispersion laws) are classified accoding to spin projection on this axis. The found generalized spin projectors (\[Sigma\]), (\[sb\]) on the axis of complete polarization play a special role in the eigenvalue problem, simplifying essentially algebraic calculations.
In the case of moving matter the states with the definite spin projection on the found axis (\[vec\_z\]) have a definite dispersion law. In particular case of rest matter the operators $\Sigma^{\pm}$ are projectors on the helicity states in correspondence with known earlier results [@Man87; @Pan92]. Let us emphasize that for moving matter or magnetic field the vanishing of commutator with inverse propagator $S$ $[S,\Sigma^{\pm}]=0$ does not lead to conservation of spin projection on this axis, since the spin projectors $\Sigma^{\pm}$ do not commute, generally speaking, with Hamiltonian.
Let us note that the propagator in external magnetic field (after use of the $\Sigma^{\pm}$ properties) is not covariant one, it contains also $\gamma^0$ matrix besides the unit matrix, $\hat{p}$ and $\gamma^5$. But for the eigenvalue problem the covariance is not essential, so after transfer to non-covariant “momentum” $p^\mu_\pm$ one can use an algebraic construction for vacuum propagator.
We considered here the cases of moving non-polarized matter or external magnetic field. In our approach one can take into account the matter polarization: it leads to simple substitution of four-velocity of matter $u^{\mu}$ by some combination of velocity and matter polarization (see [@Lobanov:2001ar]), after it the same algebraic construction is repeated.
The most evident development of this approach is related with neurtino oscillation in matter, in particular, in astrophysical problems, for instance, in propagation of neutrinos through supernova envelope, see e.g. [@Duan:2009cd], [@zas1]. The presence of the fixed axis of complete polarization and reducing the number of gamma-matrix structures should make this problem algebraically similar to the mixing problem in vacuum – see corresponding spectral representation in [@q98]. We suppose that dynamics of the neutrino spin in media in the presence of the off-shell axis of complete polarization also may be interesting.
We are grateful to V.A. Naumov, S.E. Korenblit, S.I. Sinegovsky and A.V.Sinitskaya for useful discussions and comments.
L. Wolfenstein, Phys. Rev. D [**17**]{}, 2369 (1978) S. Mikheev, A. Smirnov, Sov. J. Nucl. Phys. [**42**]{}, 913 (1985) S.M. Bilenky, C. Giunti, W. Grimus, Prog. Part. Nucl. Phys. [**43**]{}, 86 (1999) A.V. Derbin, Phys. Usp. [**57**]{}, 512 (2014) S. Bilenky, Lect. Notes Phys. **947**, 1 (2018)
W. C. Haxton, Phys. Rev. Lett. [**57**]{}, 1271 (1986) S. J. Parke, Phys. Rev. Lett. [**57**]{}, 1275 (1986) S. T. Petcov, Phys. Lett. B [**200**]{}, 373 (1988) V. A. Naumov, Phys. Lett. B [**323**]{}, 351 (1994). A. B. Balantekin, Phys. Rev. D [**58**]{}, 013001 (1998)
H. Nunokawa, V. B. Semikoz, A. Y. Smirnov, J. W. F. Valle, Nucl. Phys. B [**501**]{}, 17 (1997) S. Bergmann, Y. Grossman, E. Nardi, Phys. Rev. D [**60**]{}, 093008 (1999) A. E. Lobanov, A. I. Studenikin, Phys. Lett. B [**515**]{}, 94 (2001) A. Grigoriev, A. Lobanov, A. Studenikin, Phys. Lett. B [**535**]{}, 187 (2002) A. Esteban-Pretel, A. Mirizzi, S. Pastor, R. Tomas, G. G. Raffelt, P. D. Serpico, G. Sigl, Phys. Rev. D [**78**]{}, 085012 (2008)
M. Dvornikov, A. Studenikin, JHEP [**0209**]{}, 016 (2002) A. Lobanov, A. Studenikin, Phys. Lett. B [**564**]{}, 27 (2003)
H. Duan, J. P. Kneller, J. Phys. G [**36**]{}, 113201 (2009) C. Volpe, Acta Phys. Polon. Supp. [**9**]{}, 769 (2016)
A. Cisneros, Astrophys. Space Sci. [**10**]{}, 87 (1971) K. Fujikawa, R.Shrock, Phys. Rev. Lett. [**45**]{}, 963 (1980) M.B. Voloshin, M.I. Vysotskii, Sov. J. Nucl. Phys. **44**, 544 (1986) M.B. Voloshin, M.I. Vysotskii, L.B. Okun’, Zh. Eksp. Teor. Fiz. **91**, 754 (1986) E. Kh. Akhmedov, Phys. Lett. B **213**, 64 (1988) A.M. Egorov, A.E. Lobanov, A.I. Studenikin, Phys. Lett. B [**491**]{}, 137 (2000) Zhi-zhong Xing, Shun Zhou, Neutrinos in Particle Physics, Astronomy and Cosmology, 1. Springer-Verlag Publ., Heidelberg (2011) C. Giunti, A. Studenikin, Rev. Mod. Phys. **87**, 531 (2015) W. Grimus, P. Stockinger, Phys. Rev. D [**54**]{}, 3414 (1996) C.Y. Cardall, D.J.H. Chung, Phys. Rev. D [**60**]{}, 073012 (1999) M. Beuthe, Phys. Rept. [**375**]{}, 105 (2003) C. Giunti, C.W. Kim, Fundamentals of Neutrino Physics and Astrophysics, 1. Oxford Univ. Press Publ., New York (2007) E.Kh. Akhmedov, J. Kopp, JHEP [**008**]{}, 1004 (2010) D.V. Naumov, V.A. Naumov, J. Phys. G [**37**]{}, 105014 (2010) M. Dvornikov, Neutrinos: Properties, Sources and Detection, Nova Science Publishers, New York, [**23**]{} (2011) D.M. Voronin, A.E. Kaloshin, JETP Lett. **106**, 209 (2017)
A.E. Kaloshin, V.P. Lomov, Eur. Phys. J. C [**72**]{}, 2094 (2012) A.E. Kaloshin, V.P. Lomov, Int. J. Mod. Phys. A [**31**]{}, 1650031 (2016)
I. Pivovarov, A. Studenikin, PoS HEP [**2005**]{}, 191 (2006) A. I. Studenikin, J. Phys. A [**41**]{}, 164047 (2008)
A.E. Kaloshin, I.V. Potapova, Russ. Phys. J. [**53**]{}, 643 (2010)
A. Messiah, Quantum mechanics, vol. **1**, North-Holland Publishing Company, Amsterdam (1961) T. Kuo, J.T. Pantaleone, Rev. Mod. Phys. [**61**]{}, 937 (1989)
P. D. Mannheim, Phys. Rev. D [**37**]{}, 1935 (1988) J.T. Pantaleone, Phys. Rev. D [**46**]{}, 510 (1992)
A.I. Akhiezer, V.B. Berestetskii, Kvantovaya elektrodinamika (in Russian), 4th ed., Nauka Publ., Moscow (1981)
I.M. Ternov, V.G. Bagrov, A.M. Khapaev, JETP **21**, 613 (1965) V.G. Bagrov, D.M. Gitman, Exact solutions of relativistic wave equations, 1., Kluwer Academic Publ., Dordrecht (1990).
V.N. Zirakashvili, V.S. Ptuskin, Astropart. Phys. [**78**]{}, 28 (2016)
Spectral representation of matrix of general form {#sec:eigen}
=================================================
In order to build a spectral representation of the matrix $S$ of general form, one needs to solve two eigenvalue problems.
Left eigenvalue problem: $$\label{LL}
S \psi =\lambda \psi$$
and right one: $$\label{RR}
\phi^T S = \phi^T \lambda.$$ Here $S$ is matrix of dimension $n$ and $\psi$, $\phi$ are the columns of this dimension.
Let us indicate the main properties of these problems.
- The spectra of the left and right problems coinsides. Indeed, the eigenvalues of the left problem are defined by equation $det(S-\lambda E) = 0$, as for spectrum of the right – it is defined by transpose matrix $det(S-\lambda E)^T = 0$.
- Orthogonality of eigenvectors. Let us write down two equations $$\label{LL1}
S \psi_i =\lambda_i \psi_i.$$ $$\label{RR1}
\phi_k^T S = \phi_k^T \lambda_k.$$ Let us multiply (\[LL1\]) by $\phi_k^T$ from the left, (\[RR1\]) by $\psi_i$ from the right and subtruct one equation from another. We have $$\label{}
0 = (\lambda_i - \lambda_k) \phi_k^T \psi_i ,$$ i.e. eigenvectors of left and right problems $\phi_k$, $\psi_i$ are orthogonal [^6] at $i \not= k$. $$\label{}
\phi_k^T \psi_i = \psi_i^T \phi_k \equiv (\psi_i, \phi_k)=0 \ \ \mbox{at} \ \ i \not= k$$ One can require the orthonormality of these two sets of vectors $$\label{orto}
(\psi_i, \phi_k)=\delta_{ik}.$$
- Having solutions of both left and right problems with the property (\[orto\]), one can build matrices of the form $$\label{}
\Pi_i = \psi_i \phi_i^T ,\ \ \ \ \ i=1\dots n ,$$ which are the set of orthogonal projectors. $$\label{}
\Pi_i \Pi_k = \delta_{ik} \Pi_k$$ Note that the projectors $\Pi_i$ (eigenprojectors) are the matrix solution of both left and right eigenvalue problems.
- In particular case of hermitian matrix $S$, solutions of left and right problems are related as follows $$\label{}
\phi_i = \psi_i^*$$ and eigenprojectors look like: $$\label{}
\Pi_i = \psi_i \psi_i^\dagger ,\ \ \ \ \ i=1\dots n .$$
Having solutions of left and right problems, one can represent matrix in a form $$\label{}
S = \sum_{i=1}^{n} \lambda_i \Pi_i = \sum_{i=1}^{n} \lambda_i \psi_i \phi_i^T .$$ This is a spectral representation of a general form matrix, which includes solutions of both left $\psi_i $ and right $\phi_i $ eigenvalue problem.
Spectral representation of fermion propagator in vacuum and media {#sec:prop}
=================================================================
Let us consider the dressed fermion propagator in vacuum $S(p)=\hat{p}-m - \Sigma(p)$. In case of P-parity violating the self-energy $\Sigma(p)$ contains $\gamma^5$ and may be written as $$\label{}
\Sigma(p) = a(p^2) + b(p^2)\hat{p}+c(p^2)\gamma^5 + d(p^2) \hat{p}\gamma^5 .$$ Note that in the absence of $\gamma^5$ a spectral representation is rather evident, but parity violation generates non-trivial eigenprojectors. Solving the eigenvalue problem for such matrix $S(p)$ of dimension 4, we will find only two eigenvalues $\lambda_i(p^2)$ and eigenprojectors $\Pi_i(p)$. Evidently, this degeneration is related with spin and complete set of eigenprojectors is constructed with use of spin projectors, dependent on arbitrary polarization four-vector $s$. $$\label{}
\Sigma^{\pm} (s) \Pi_i(p), \ \ \ \ \ \ i=1,2 .$$ The appeared here generalized spin progectors [@q98] commute with propagator and in theories with parity violation take the form $$\label{}
\Sigma^{\pm} (s) =\frac{1}{2}(1 \pm \gamma^5 \hat{s} \hat{n}) ,\ \ \ \ \ \ n^\mu=p^\mu/W,\ \ \ \ \ \ W=\sqrt{p^2} .$$ The additional factor $\hat{n}$ in spin projector plays an essentional role in the dressed propagator, for bare propagator (and in theory without $\gamma^5$) it turnes into unit matrix.
Spectral representation for dressed inverse propagator in vacuum may be written in the following elegant form $$\begin{aligned}
\label{sp_vac}
S(p) = \lambda_1 \Sigma^{+} (s) \Pi_1(p) + \lambda_1 \Sigma^{-} (s) \Pi_1(p) +\nonumber\\+ \lambda_2 \Sigma^{+} (s) \Pi_2(p) +\lambda_2 \Sigma^{-} (s) \Pi_2(p) = \nonumber\\
= \lambda_1 \Pi_1(p) + \lambda_2 \Pi_2(p) .\end{aligned}$$ It should be mentioned that for fermion propagator it is convenient to solve an eigenvalue problem in a matrix form (i.e. to look for eigenprojectors insteed of eigenvectors), using the $\gamma$-matrix basis. In doing so we will find only two eigenvalues for dressed propagator. As for eigenprogectors, the left problem has two matrix solutions $\Pi_1, \Pi_2$ and there is no necessity to solve a right eigenvalue problem — the orthogonality requirement fixes ambiguity completely [@q12].
For propagator in a moving matter (see (\[formula2\])) there appear new $\gamma$-matrix structures and it leads to disappearing of degeneration: in this case we have four different eigenvalues and eigenprojectors (\[eig\_1\]), (\[eig\_2\]). Spectral representation of inverse propagator in matter $$\begin{aligned}
\label{}
S(p,u)= \sum_{i=1}^{4} \lambda_i \pi_i(p,u) = \nonumber \\= \sum_{1}^{2} \lambda_i \Sigma^{-}(z) \Pi_i + \sum_{3}^{4} \lambda_i \Sigma^{+}(z) \Pi_i \end{aligned}$$ contains the generalized spin projectors, depending of the fixed (in contrast to (\[sp\_vac\])) four-vector $z$ (\[vec\_z\]).
[^1]: Short version of this paper without discussion of propagation in external magnetic field was published in [@VK].
[^2]: This vector was used earlier [@q11] for some algebraic simplification of propagator in matter.
[^3]: Note that for space-like momentun $p^{2}<0$ the polarization vector $z^{\mu}$ becomes imaginary. But the product $\hat{z}\hat{n}=\hat{z}\hat{p}/W$ in spin projector (\[Sigma\]) remains real.
[^4]: We call them as generalized because of appearence of additional factor $\hat{n}$. But in fact the Eq. (\[Sigma\]) is the most general form of spin projectors at dressing of fermion propagator in theories with parity violation — see details in [@q98]. The same modification (\[Sigma\]) of a naive spin projector arises in matter – may be not accidently.
[^5]: This vector arises in consideration of motion of a charged relativistic fermion in a constant and homogenious magnetic field, see 4-th edition of textbook [@AB], §1.6 . We consider another situation: neutral fermion with an anomalous magnetic moment in a magnetic field, but it turns out that in this case the constructed spin projector also commutes with the propagator.
[^6]: Case of degenerate eigenvalues – see below an example of spectral representation of propagator.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'There are three kinds of solid states of matter that can exist in physical space: quasicrystalline (quasiperiodic), crystalline (periodic) and amorphous (aperiodic). Herein, we consider the degree of orientational order that develops upon the formation of a solid state to be characterized by the application of quaternion numbers. The formation of icosahedral quasicrystalline solids is considered alongside the development of bulk superfluidity, characterized by a complex order parameter, that occurs by spontaneous symmetry breaking in three-dimensions. Crystalline solid states are viewed as higher-dimensional analogues to phase-coherent topologically-ordered superfluid states of matter that develop in restricted dimensions (Hohenberg-Mermin-Wagner theorem). Lastly, amorphous solid states are viewed as dual to crystalline solids, in analogy to Mott-insulating states of matter that are dual to topologically-ordered superfluids.'
author:
- 'Caroline S. Gorham'
- 'David E. Laughlin'
bibliography:
- '\\jobname.bib'
title: 'On the Formation of Solid States Beyond Perfect Crystals: Quasicrystals, Geometrically-Frustrated Crystals and Glasses'
---
Introduction
============
The sphere-packing problem, which describes the arrangement of atomic constituents in close-packed crystalline solid states, has been of great interest for several centuries [@knudson_stacking_2019]. In three-dimensions, this topic came to be known as Kepler’s conjecture [@kepler_strena_1611; @kepler_six-cornered_2010] which states that no arrangement of equally-sized spheres filling space in three-dimensions has a greater average density [@gauss_Untersuchungen_1831] than $\pi/3\sqrt{2}\approx 0.74$. Cubic close-packing (FCC) and hexagonal close-packing (HCP) arrangements are examples of such sphere-packing in three-dimensions. This conjecture was proved computationally by Hales in 1998 [@hales_kepler_1998; @sloane_keplers_1998; @hales_proof_2005]. More recently, Viazovska [@viazovska_sphere_2017; @cohn_conceptual_2017] proved that the closest-packing of equally-sized spheres in eight-dimensions is the $E8$ lattice which has a density of $\pi^4/384\approx 0.25$. These close-packed arrangements, of equally-sized spheres, describe crystalline ground states with perfect orientational order in either three-dimensions or eight-dimensions.
It has been proposed by the authors that the problem of solidification can be understood by the application of a *quaternion* ($\mathbb{Q}$) order parameter, to characterize the degree of orientational order in the solid states that form from undercooled liquids. In this way, it is possible to frame crystalline solids (orientationally-ordered) in three-dimensions or eight-dimensions as generalizations of phase-coherent superfluids [@kosterlitz_ordering_1973] that are characterized by a *complex* ($\mathbb{C}$) order parameter [@london_superfluids_1961]. Following this approach, we have generated a unified topological framework within which to understand the formation of the three types of solid states in three-dimensions:
- quasicrystalline [@gorham_solidification_2019] (quasiperiodic)
- crystalline [@gorham_topological_2019; @gorham_crystallization_2019] (periodic)
- amorphous [@gorham_topological_2019] (aperiodic)
This unified topological framework, that applies to the solidification of the three kinds of solid states of matter (listed above) is constructed via the application of *topology* and the theory of *Universality classes of phase transitions* [@herbut_modern_2007].
Perfectly orientationally-ordered crystalline structures, i.e., those without any topological defects, are idealized cases of solids that can only be realized at 0 K. At temperature, even these perfect crystalline solids have a finite density of topological defects (e.g., dislocations) whose nucleation is necessary in order to account for crystal entropy. Beyond perfectly orientationally-ordered crystals, *geometrically-frustrated* crystalline solids [@frank_complex_1958; @frank_complex_1959; @sadoc_geometrical_2006] (i.e., topologically close-packed, TCP) that possess 5-fold local icosahedral order [@frank_supercooling_1952] contain topological defects in the ground state and warrant considerable attention. These geometrically-frustrated crystalline solids, which have a periodic spatial distribution of orientational disorder in the ground state, can also be understood within our unified topological framework for solidification. Thus, by considering a quaternion orientational order parameter we go beyond the conventional sphere-packing problem that applies to perfect crystalline solids in order to consider the formation of solid states with a spatial distribution of orientational disorder (curvature).
The main section of this article (Section \[sec:solidification\]) focuses on a topological ordering field theory for solidification, using quaternion numbers to characterize orientational order. Section \[sec:solidification\] presents a discussion on the mechanisms of formation of the three types of solid states of matter. Section \[sec:solidification\] is separated into two distinct subsections that describe solidification in “bulk" dimensions (Section \[sec:quasicrystals\]) and in “restricted" dimensions (Section \[sec:crystal-to-glass\]) separately. Section \[sec:quasicrystals\] describes the solidification of icosahedral quasicrystalline states of matter, which derive by projection from the eight-dimensional E8 lattice, in analogue to the formation of bulk superfluids in three-dimensions. Solidification of crystalline and non-crystalline solids, in the vicinity of an anticipated crystalline-to-glass quantum phase transition that may be realized in “restricted" dimensions, are discussed in Section \[sec:crystal-to-glass\].
Quaternion Field Theory of Solidification {#sec:solidification}
=========================================
Solidification in “Bulk” Dimensions: Quasicrystals {#sec:quasicrystals}
--------------------------------------------------
Quaternion numbers (SU(2)) have a 2-to-1 homomorphism [@mermin_homotopy_1978; @mermin_topological_1979] with the group of rotations about an origin in three-dimensional Euclidean space ($SO(3)$), and may therefore be used to characterize orientational order that develops upon *solidification*. Quaternion numbers are four-dimensional, such that the group of all unit quaternions is the hyperspherical surface ($S^3\in\mathbb{R}^4$) of constant positive curvature. The discrete nature of orientational order in crystalline solid states is accounted for by considering the ground state manifold to be $\mathcal{M}=G/H'$, where G=SU(2) and $H'$ is the *binary polyhedral group* representation of the preferred orientational order group $H\in SO(3)$. The relevant ground state manifold $\mathcal{M}$ rests in the quaternion plane. Thus, the main kind of available topological defects belong to the *third homotopy group* (i.e., $\pi_3(\mathcal{M})$).
Third homotopy group topological defects are higher-dimensional analogues to quantum vortices in superfluid ordered states of matter, in which a complex order parameter characterizes the degree of order (‘Mexican hat’ free energy function [@endres_higgs_2012]). Quantum vortices in superfluids are holes with the superfluid circulating around the vortex axis [@annett_superconductivity_2004]. Just as the free energy cost to introduce a vortex line in three-dimensional superfluids is much too high for them to appear in the absence of external fields [@halperin_resistive_1979], the same argument applies for third homotopy group topological defects in dimensions larger than four. On the other hand, third homotopy group defects are points in the four-dimensional quaternion plane and so, the free energy cost to introduce these defects in four-dimensions should be comparable to $\text{k}_\text{B}\text{T}$ in the absence of external fields [@halperin_resistive_1979].
It has long been understood that icosahedral quasicrystalline (quasiperiodic) solids can be derived from the 8-dimensional lattice E8, whose close-packed neighbor shells are embedded in 7-dimensional spheres ($S^7$). This is described mathematically by making use of the *quaternion Hopf fibration* [@sadoc_e8_1993] decomposition of $E8$. Third homotopy group topological defects in eight-dimensions are higher-dimensional than points, and it is therefore anticipated – in analogy to vortex lines in superfluids – that they may only be introduced in the presence of an applied field [@halperin_resistive_1979]. In this way, in eight-dimensions, the presence of third homotopy group defects should not prevent *spontaneous symmetry breaking* at the melting temperature.
Thereby, it is anticipated that, the E8 lattice may be realized by a conventional disorder-order phase transition that occurs in the “bulk" dimension for the quaternion orientational order parameter [@gorham_solidification_2019]. Quasiperiodic structures in physical space are therefore considered to be most similar to superfluid ordered systems that exist in three-dimensions [@gorham_solidification_2019], which do not form by topological-ordering (in the Berezinskii-Kosterlitz-Thouless sense [@berezinskii_destruction_1971; @kosterlitz_ordering_1973]).
Solidification in “Restricted" Dimensions 4D/(3D+1t): Crystal-to-Glass Quantum Phase Transition {#sec:crystal-to-glass}
-----------------------------------------------------------------------------------------------
Phase transitions, and transport properties of ordered states of matter become more interesting when the degrees of freedom of the system of particles that undergo ordering are “restricted" in a dimensional sense (*Hohenberg-Mermin-Wagner* (HMW) *theorem* [@mermin_absence_1966; @hohenberg_existence_1967; @halperin_hohenbergmerminwagner_2018]). The original HMW theorem states that any system with continuous symmetry ($O(N)$, $N\geq 2$) cannot undergo *spontaneous symmetry breaking* (SSB) at any finite temperature in two- or one-dimensions [@mermin_absence_1966]. In this way, 2D/(1D+1t) may be considered to be “restricted” (because SSB is not possible) for any ordered system that breaks a continuous symmetry group. As a consequence of the original HMW theorem, traditional phase-coherent ordered states cannot occur by any conventional disorder-order phase transition in two- or one-dimensions [@lee_symmetry_1986].
However, low-temperature complex ordered superfluids do exist in “restricted" dimensions. In the case of complex ordered systems (i.e., $N=2$) that exist in 2D/(1D+1t), which belong to the ($2, 2$) Universality class, the nature of the novel phase transition towards the phase-coherent superfluid ground state was explained in terms of the topological-ordering of vortex/anti-vortex point defect pairs by Berezinskii [@berezinskii_destruction_1971] and Kosterlitz and Thouless [@kosterlitz_ordering_1973]. Such complex ordered systems, that exist in “restricted" dimensions, are described mathematically by the application of the $O(2)$ quantum rotor model [@vojta_quantum_2006; @sachdev_quantum_2011; @endres_higgs_2012].
Such $O(N)$ quantum rotor models are essential for the study of quantum phase transitions [@sachdev_quantum_2011], between phase-coherent and phase-incoherent low-temperature ordered states. In particular, in the two-dimensional $N=2$ case, the $O(2)$ quantum rotor model predicts the existence of a quantum phase transition between the superfluid and a Mott-insulator ordered state [@endres_higgs_2012] that has been realized experimentally. In charged superfluids, this is the superconductor-to-superinsulator transition [@vinokur_superinsulator_2008; @baturina_superinsulatorsuperconductor_2013; @sankar_disordered_2018; @diamantini_superconductor-superinsulator_2018].
In the following subsections (\[sec:xtal\], \[sec:fk\], \[sec:glass\]), we summarize our recent generalizations of the notion of “restricted" dimensions [@gorham_su2_2018; @gorham_topological_2019; @gorham_crystallization_2019] (in the HMW theorem sense) to systems with continuous symmetry ($O(N)$, $N\geq 4$) that exist in 4D/(3D+1t). In quaternion ordered systems ($N=4$), this is a consequence of the nature of *third homotopy group* topological defects as points in 4D/(3D+1t) which are spontaneously generated at finite temperatures. These topological defects generate an abundance of misorientational fluctuations in the quaternion order parameter, that prevent *spontaneous symmetry breaking* at finite temperatures. This generalization of the original HMW theorem, to quaternion ordered systems, has been applied (by the authors) to understand the mechanisms of solidification of crystalline and non-crystalline solid states in the vicinity of a first-order quantum phase transition at the Kauzmann point [@kauzmann_nature_1948].
### Perfect crystalline solid states {#sec:xtal}
Crystallization in 4D/(3D+1t) has been described as a defect-driven phase transition [@gorham_su2_2018; @gorham_topological_2019; @gorham_crystallization_2019], that occurs at a finite temperature below the melting temperature, in analogue to the formation of phase-coherent superfluidity in 2D/(1D+1t) systems. Therefore, crystallization in 4D/(3D+1t) belongs to the Berezinskii-Kosterlitz-Thouless universality class [@berezinskii_destruction_1971; @kosterlitz_ordering_1973] of phase transitions for quaternion ordered systems. This particular topological phase transition, occurs as third homotopy group point defects and anti-point defects bind into complementary pairs. This is a direct higher-dimensional analogue to vortex/anti-vortex binding at the prototypical BKT-transition in complex ordered thin-films [@kosterlitz_critical_1974].
In addition to third homotopy group topological defects, as a consequence of discrete orientational order of atomic clusters in undercooled fluids [@frank_supercooling_1952], misorientational fluctuations that develop below the melting temperature also consist of disclinations that belong to the fundamental homotopy group of $\mathcal{M}$. Upon crystallization, the topology of $\mathcal{M}$ changes discontinuously as a result of the development of a translational lattice that corresponds to a Brillouin zone in reciprocal space. This change in the topology of $\mathcal{M}$, from a spherical to a toroidal manifold, constitutes a first-order phase transition and forces the formation of bound states of complementary disclinations that can be regarded as a dislocations [@seung_defects_1988; @chaikin_principles_2000; @landau_theory_1986; @gopalakrishnan_disclination_2013; @pretko_fracton-elasticity_2018] that can be drawn around $\mathcal{M}$.
In this way, due to the formation of bound states of complementary curvature-carrying disclinations, crystallization may be viewed as a *flattening* of an undercooled liquid into a solid state. Perfect orientational-order is obtained for crystalline ground states in the absence of geometrical frustration [@sadoc_geometrical_2006]. This occurs because, in the absence of geometrical frustration, the plasma of topological defects that forms just below the melting temperature is perfectly balanced and all of the defects are able to form bound pairs. This ensures that the perfect crystalline ground state is free of permanent topological defects, and is therefore flat everywhere. Such bound states of disclinations, i.e., dislocations, are imperfections in a crystalline lattice and only become excited at finite temperatures. In the same way, bound states of third homotopy group topological defects are only present at temperatures above 0 K.
### Geometrically-frustrated crystalline solid states (TCP) {#sec:fk}
In analogue to complex ordered systems that exist in “restricted" dimensions, i.e., 2D/(1D+1t), *frustration* of ground states becomes possible for quaternion ordered systems that exist in 4D/(3D+1t). Increasing frustration drives the ordered system towards a *quantum phase transition*, that belongs to the relevant $O(N)$ quantum rotor model [@sachdev_quantum_2011]. In the case of crystalline solid states, the relevant frustration parameter is geometrical [@sadoc_geometrical_2006] and is proportional to the curvature that is associated with atomic vertices whose local orientational order is incompatible with creating a space-filling arrangement [@nelson_symmetry_1984].
The most well-known manifestation of geometrical frustration in crystals occurs in Frank-Kasper phases of complex transition metal alloys [@frank_complex_1958; @frank_complex_1959], that form from icosahedrally-coordinated undercooled fluids [@nelson_symmetry_1984]. Positive curvature that is attributed to each icosahedrally-coordinated atomic vertex is inversely proportional to the radius of the $\{3,3,5\}$ polytope, which consists of 120 particles (the binary icosahedral group) inscribed on the surface of a sphere in four-dimensions. The finite positive curvature, associated with icosahedral clustering in undercooled fluids, biases the plasma of topological defects that comprise the gas of misorientational fluctuations below the melting temperature. This ensures that the space remains flat overall.
Upon crystallization, these excess negative-signed topological defects cannot form bound states and must persist to the ground state. Topological defects in the crystalline ground state, induced by geometrical frustration, form a lattice in analogue to the Abrikosov flux lattice in topologically-ordered superconductors in the presence of an applied magnetic field [@nelson_defects_2002]. In three-dimensional Frank-Kasper crystalline solids, this manifestation of geometrical frustration in the ground state is evidenced as the *major skeleton network* [@nelson_liquids_1983] of disclination lines that concentrate negative curvature and whose periodic arrangement satisfies the third law of thermodynamics. With a sufficient amount of geometrical frustration, the orientationally-ordered crystalline solid state becomes entirely broken down as the distance between topological defects becomes reduced below a critical value.
### Non-crystalline solid states {#sec:glass}
Within this unified topological framework for solidification, amorphous solids are considered to be topologically-ordered phases that may arise in 4D/(3D+1t) uncharged quaternion ordered systems [@gorham_topological_2019] in the vicinity of a crystalline-to-glass quantum phase transition that belongs to the $O(4)$ quantum rotor model. The possibility of the existence of non-crystalline solid states (i.e., orientationally-disordered) is due to the duality of the phase-amplitude uncertainty principle, that applies to the quaternion orientational order parameter wave-function. In this way, glassy solids are considered to be a higher-dimensional analogue to the Mott-insulator or superinsulator ground states [@baturina_superinsulatorsuperconductor_2013] that emerge on the insulating side of the superfluid-to-Mott insulator or superconductor-to-superinsulator transition (SIT) in complex ordered thin films [@diamantini_gauge_1996; @vinokur_superinsulator_2008].
A solidification phase diagram, in the vicinity of the first-order crystalline-to-glass quantum phase transition at the Kauzmann point, has been proposed by the authors [@gorham_topological_2019; @gorham_crystallization_2019]. The glass transition is considered to belong within the Berezinskii-Kosterlitz-Thouless (BKT) (4,4) Universality class, that applies to quaternion ordered systems that exist in “restricted" dimensions. In this sense, the glass transition occurs via the condensation of topological defects and localization of atomic particles [@gorham_topological_2019]. The glass transition in three-dimensions is analogous to the formation of superinsulating phases in Josephson junction chains, that occur via instanton condensation [@polyakov_gauge_1987] and localization of Cooper pairs [@diamantini_confinement_2018].
The duality between ordered states of matter that can be realized in “restricted" dimensions, in the vicinity of a quantum phase transition that belongs to the relevant $O(N)$ quantum rotor model [@sachdev_quantum_2011; @endres_higgs_2012], manifests observably in the transport properties. For example, as originally predicted by t’Hooft [@hooft_phase_1978] (1978), the superinsulator that is the dual to a superconductor exhibits infinite electrical resistance [@diamantini_electrostatics_2019]. In the same way, the thermal conductivity of crystalline and glassy solid states have long been known [@eucken_uber_1911; @kittel_interpretation_1949] to exhibit inverse behavior above approximately 50 K. Many theoretical models exist to interpret the thermal conductivity of glasses in this temperature range, which decreases as a function of decreasing temperatures [@einstein_elementary_1911; @birch_thermal_1940; @kittel_interpretation_1949; @slack_thermal_1979; @cahill_thermal_1987; @nakayama_increase_1999] – above a plateau at low-temperatures that is a universal feature of the thermal conductivity in non-crystalline solids [@zeller_thermal_1971; @freeman_thermal_1986; @graebner_phonon_1986; @leggett_amorphous_1991; @ramos_low-temperature_1997]. The topological framework for solidification of crystals and glasses elucidated herein points towards a novel interpretation of this inverse transport behavior [@gorham_topological_2019].
Conclusions {#sec:summary}
===========
In this treatise on solidification, a unified framework has been presented to consider the phase transitions that give rise to the three types of solid states of matter: quasicrystalline (Section \[sec:quasicrystals\]), crystalline and glassy (Section \[sec:crystal-to-glass\]). This approach to solidification relies on the application of a quaternion order parameter, to characterize the degree of orientational order that develops below the melting temperature. Fundamentally, orientationally-ordered crystalline solids are viewed as generalizations of phase-coherent superfluids.
The Kauzmann point, that occurs at a finite temperature, has been associated with a quantum critical point that identifies a first-order quantum phase transition belonging to the $O(4)$ quantum rotor model. Such a finite-temperature quantum critical point, between crystalline and non-crystalline solid states, is classified as first-order. Topologically, this is owing to the fact that the genus of the orientational order parameter manifold changes discontinuously from a torus (in the crystal) to a sphere (in the glass). This quantum phase transition is a higher-dimensional analogue to the superfluid-to-Mott insulator transition. It follows that, solid states in the vicinity of the Kauzmann point exhibit topological-order in the Berezinskii-Kosterlitz-Thouless sense.
Acknowledgements
================
The authors gratefully acknowledge support from the ALCOA Chair in Physical Metallurgy. C.S.G. also thanks the lecturers and participants of Beg Rohu’s Summer School 2019: Glasses, Jamming and Slow Dynamics, for insightful discussions regarding the glass transition.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Despite the recent works on knowledge distillation (KD) have achieved a further improvement through elaborately modeling the decision boundary as the posterior knowledge, their performance is still dependent on the hypothesis that the target network has a powerful capacity (representation ability). In this paper, we propose a knowledge representing (KR) framework mainly focusing on modeling the parameters distribution as prior knowledge. Firstly, we suggest a knowledge aggregation scheme in order to answer how to represent the prior knowledge from teacher network. Through aggregating the parameters distribution from teacher network into more abstract level, the scheme is able to alleviate the phenomenon of residual accumulation in the deeper layers. Secondly, as the critical issue of what the most important prior knowledge is for better distilling, we design a sparse recoding penalty for constraining the student network to learn with the penalized gradients. With the proposed penalty, the student network can effectively avoid the over-regularization during knowledge distilling and converge faster. The quantitative experiments exhibit that the proposed framework achieves the state-of-the-arts performance, even though the target network does not have the expected capacity. Moreover, the framework is flexible enough for combining with other KD methods based on the posterior knowledge.'
author:
- |
Junjie Liu[^1^]{}, Dongchao Wen[^1^]{}, Hongxing Gao[^1^]{}, Wei Tao[^1^]{},\
[^1^]{}[Canon Information Technology (Beijing) Co., LTD]{}\
[{liujunjie, wendongchao, gaohongxing, taowei}@canon-ib.com.cn]{}
- |
Tse-Wei Chen[^2^]{}, Kinya Osa[^2^]{}, Masami Kato[^2^]{}\
[^2^]{}[Device Technology Development Headquarters, Canon Inc.]{}\
[[email protected]]{}
bibliography:
- 'egbib.bib'
title: 'Knowledge Representing: Efficient, Sparse Representation of Prior Knowledge for Knowledge Distillation'
---
Introduction
============
The deep neural network has achieved the significant improvement in different fields with years, but it also requires higher computational and memory costs. For the purpose to apply these networks to the real-time industrial tasks, the neural network compression [@cheng2017survey] is arguably the most crucial strategy. As for the network compression problem, the typical solutions are designed to *slim* [@sandler2018mobilenetv2; @zhang2018shufflenet] the network directly, or quantify their parameters distributions [@hubara2017quantized; @jacob2018quantization; @rastegari2016xnor], and filter the redundant layer dimensions [@he2017channel; @thinet2017].
![The pipeline of knowledge representing algorithm: The prior knowledge in teacher network is represented by the knowledge aggregation scheme into higher abstract level. Then the sparse recoding penalty is further used to regularize the gradients in student network for efficient learning these prior knowledge.[]{data-label="fig:KD"}](fig1.pdf){width=".47\textwidth"}
In contrast to these techniques which aim at directly compressing the network while preserving its performance as much as possible, an alternative solution is to preset a smaller target network as the student, and employ the knowledge from the larger network as teacher to improve student’s performance. Therefore, knowledge distillation [@Hinton2015Distilling] (KD) is proposed. The KD mainly assumes the samples distribution is anisotropy [@balan2015bayesian], but annotations of the samples are not able to represent this intrinsic. Based on the hypothesis, these methods evaluate the samples in the teacher network to produce the decision boundary as a strong posterior distribution, and then use to regularize the gradients optimization of student network. While this helps prevent the student network from being over-fitting, the extra risk of non-convergence is introduced.
A possible solution is to refine the posterior distribution from the teacher network, in order to provide more valuable knowledge for better distilling. The Neuron Selectively Transfer (NST) [@huang2017like] is proposed to align the distribution selectively with the Maximum Mean Discrepancy (MMD) metric, and the generative adversarial network with KD (KDGAN) [@wang2018kdgan] is further used to produce a more robust decision boundary for student classifier. However, considering the student network which contains a very limited capacity - the representation ability, this limitation gradually becomes a major bottleneck in network training to further improve the performance of knowledge distillation. In a word, the fine-grained posterior distribution is usually underemployed.
With the constraint from network capacity, an instinctive approach is to introduce the parameters distribution [@denil2013predicting] from the teacher network as the prior knowledge [@Romero2014FitNets; @yim2017gift]. For the typical one, Romero et al. [@Romero2014FitNets] constructs the Hint layer to estimate a parameters distribution with less filter numbers, through using the intermediate features representation of the teacher, and it uses these knowledge to guide the update of student parameters. However, the Hint layer suffers from the over-regularization if the teacher network is too deep. In this paper, we produce a KD solution mainly focusing on modeling the prior knowledge, while avoiding the negative impacts from over-regularization, and the solution is flexible enough, for combining with other KD methods based on the posterior knowledge. Specially, we propose a knowledge representing (KR) framework, which aims at representing the prior knowledge at more abstract level, and taking full advantage of these knowledge. For answering the question of how to represent the prior knowledge, a knowledge aggregation scheme is firstly suggested. Inspired by the theory of optimal transportation [@Martin2017; @Na2017A], the scheme is designed to alleviate the phenomenon of residual accumulation in the deeper layers. Then, as for the most critical issue of what the dominant prior knowledge is for better distilling, a sparse recoding penalty is proposed. Through employing a learnable threshold in the penalty, it can enhance the gradients of dominant neurons and smooth inactive ones. With these two proposed terms, the proposed framework can prompt the student network to preserve the key features of teacher network, even without a strong representation ability.
Our paper makes the following contributions:
- A new penalty is proposed to constrain the optimization of knowledge distillation. It helps the student network to avoid the over-regularization and converge faster. Moreover, the penalty can be further applied on other network optimization problems.
- A new scheme is suggested for aggregating the prior knowledge. It is able to produce more abstract features and alleviate the phenomenon of residual accumulation.
- According to the proposed framework, the more flexible architecture is allowable for both teacher and student network, without the constraints from model depths or filter scales.
Related Work
============
The latest deep networks are usually accompanied with carefully designed modules [@denseNet2017; @he2016deep] and enormous parameters. Though the performance of targeted tasks is obviously being improved, the computation and memory cost gradually become the challenge to employ these networks in real-life applications [@jacob2018quantization; @Shen2018Towards]. Comparing to the traditional neural network compression methods [@cheng2017survey] which focus on compressing the original network directly, a solution with the knowledge distillation to compress the deep network attracts more attention from research community in recent years, such as in the tasks of image recognition [@yim2017gift], object detection [@GuobinChen_2017_NIPS; @luo2016face], or recommender systems [@Guorui_2018_AAAI], as the flexibility to obtain an arbitrary architecture of target network. In summary, the KD methods can be categorized into two main groups:
1\) *Distilling the posterior distribution from training data:* Considering the possibility to extract the knowledge in an ensemble (teacher) into a single model (student), Hinton et al. [@Hinton2015Distilling] introduces the idea of knowledge distillation as a regularizer. Through employing a penalized version [@BSS2019; @huang2018data; @zheng2017CKD] of final features of the teacher network, a joint learning is processed with the knowledge from posterior distribution. For refining the posterior distribution to provide more valuable knowledge, the Neuron Selectively Transfer (NST) [@huang2017like] is proposed to align the distribution selectively with the Maximum Mean Discrepancy (MMD) metric. Furthermore, considering the sample bias is unavoidable, the generative adversarial networks for knowledge distillation (KDGAN) [@wang2018kdgan] is further used to produce a more robust posterior distribution for student classifier. However, these methods haven’t take the capacity of student network into consideration, so the fine-grained posterior distribution is underemployed.
2\) *Distilling the prior distribution from model parameters:* An alternative approach is to introduce the parameters distribution from teacher network as the prior knowledge [@Romero2014FitNets; @yim2017gift; @Zagoruyko2017Paying]. Romero et al. [@Romero2014FitNets] designs the Hint layer to estimate the parameters distribution by using the intermediate hidden layers from the teacher, and used the Hint layer to guide the distillation. Net2Net [@Chen2015Net2Net] suggests a function-preserving transform for extracting the prior knowledge from teacher network to initialize the parameters of the student network. And Yim et al. [@yim2017gift] suggests a representation operator named FSP matrix. It uses not only the parameters distribution but the intermediate features from the neighbor layers. However, these methods either are constrained by the depth of teacher network, or suffer from the over-regularization.
Method
======
For obtaining a student network that faithfully preserves the key representation ability of the teacher, Sec. \[3dot1\] presents the objective function of the knowledge representing framework. Accordingly, we firstly answer the key problem of what the most important prior knowledge is for distilling in Sec. \[3dot2\], through introducing the mathematical expression of the sparse recoding penalty. Then, we suggest how to represent the prior knowledge from the teacher network, with a knowledge aggregation scheme in Sec. \[3dot3\]. Finally, Sec. \[3dot4\] shows the optimization procedure of the objective function.
**Initialization:** $\tilde{W^t}$, $W^s$, $MaxIter$ \[alg:frame\]
Knowledge Representing {#3dot1}
----------------------
As one of the most typical feature representation technique, the deep model produces the decision boundary through modeling the data distribution with the parameters in layers. Given a trained decision boundary $y^{t}(x, W^t)$, where $y^t$ is generated by teacher network with data distribution $x$ and the parameters $W^t$, the objective of knowledge distillation is to find the parameters $W^s$ for the student network. Specially, with the $W^s$ and $x$, the $y^s$ from student network is jointly optimized with the $y^t$. Through minimizing the dissimilarity of two decision boundaries, the objective function of knowledge distillation is defined as:
$$\underset{W^s}{\arg\min} \sum_{i=1}^{N}\pounds(y_{i}^{t}(x_i, W^t), y_{i}^{s}(x_i, W^s)) + \lambda \Phi(W^s)
\label{eq:interpret-KD}$$
where $\pounds$ represents the metric for evaluating the similarity between the $y^t$ and $y^s$, and the cross entropy, KDGAN [@wang2018kdgan], or NST [@huang2017like] are allowable. Different from the KD methods only evaluating the decision boundary, we further introduce a penalty $\Phi(\cdot)$ in Eq. \[eq:interpret-KD\], in order to measure the representation ability of student network. However, if the representation ability of student network is weak, the fine-grained posterior distribution will be underemployed. Then, we extend the objective function Eq. \[eq:interpret-KD\] through further introducing the prior knowledge $W^t$ from the teacher network, and the objective function is:
$$\begin{split}
\underset{W^s, \tilde{W^t}}{\arg\min} \sum_{i=1}^{N}&\pounds(y_{i}^{t}(x_i, W^t), y_{i}^{s}(x_i, W^s)) + \pounds(\tilde{W^t}, W^s) \\
&+ \gamma \Psi(\tilde{W^t}, W^t) + \lambda \Phi(W^s)
\label{eq:interpret-KD-revised}
\end{split}$$
Instead of directly employing the parameter distributions $W^t$ from the teacher network as prior knowledge, we firstly represent these distributions as more abstract level, and a knowledge aggregation scheme $\Psi(\cdot)$ is suggested to aggregate $W^t$ into the $\tilde{W^t}$. With the prior knowledge $\tilde{W^t}$, the $\pounds(\tilde{W^t}, W^s)$ is used to guide the update of parameters distributions $W^s$ for the student. Moreover, we propose a sparse recoding penalty to specify the $\Phi(\cdot)$. Through enhancing the magnitude of dominant gradients and filtering the inactive ones, the optimizer no longer requires the parameters distribution $W^s$ of student network to strictly close to the teacher one, and prompts the student network to firstly learn with the most valuable knowledge. In summary, the optimization procedure is represented in Algorithm \[alg:frame\], and we leave over the details in following sections.
Sparse Recoding Penalty {#3dot2}
-----------------------
As demonstrated by previous works [@wen2016learning; @wu2017beyond; @zhang2016l1], prompting the neurons connection being sparse is beneficial for obtaining a well generalization ability. However, such penalties are designed to directly clip the parameters distribution, and the extra risk of over-regularization is introduced. After we analyze the distribution of prorogated gradients in the previous KD methods, we found that major reason for the convergence of oscillatory is that the gradients are not discriminative enough, especially in the student network with a weak representation ability. Therefore, we propose a sparse recoding penalty $\Phi(\cdot)$, which can penalize the prorogated gradients during the training of deep network. Given an input parameters tensor $W$, it enhances the high gradients $g_j$ of dominant neurons, and filters the low gradients of inactive neurons. The function is defined as:
$$\begin{split}
\Phi(W) = \sum_{j}\Phi_0(g_{j})
\label{eq:sp-loss}
\end{split}$$
where
$$\Phi_0(g)=
\begin{cases}
\frac{1}{\varepsilon}(|g| + g^2), &if|g| \geq \varepsilon \\
0, &otherwise
\end{cases}
\label{eq:sp-loss-cases}$$
where $\Phi_0(\cdot)$ is a piecewise function that enhances the gradients when $|g| \geq \varepsilon$, and smooths the $|g|$ by zero in others. The $\varepsilon$ is a learnable threshold within the update of gradient optimization, and it is initialized with the mean value of parameters distribution. For fairly comparing with other penalties, the Fig.\[fig:sp-loss\] shows the curves of $\Phi(\cdot)$ by comparing with the $L_1$ and $L_2$ norms. It exhibits that $\Phi(\cdot)$ is a more strict sparse constraint. Moreover, with different parameter setting, properties of the sparse recoding penalty are shown in the figure, and we leave over the further discussion in experiments.
Deep Knowledge Aggregation {#3dot3}
--------------------------
For representing the prior knowledge as more abstract level, we design a deep knowledge aggregation scheme through stacking the neighbor layers in a very deep network. Specially, with the analysis of prior knowledge distilling in previous methods, we notice that the optimization errors between two networks will be accumulated from layers, since the higher layer in teacher network usually contains a strong representation ability. However, the situation is simply regarded as the phenomenon of gradient vanishing, and cause an over-regularization if the teacher network is too deep. So we name this phenomenon as the residual accumulation, and the proposed scheme will mainly considers this phenomenon. Based on the theory of optimal transportation [@Martin2017; @Na2017A], the scheme try to reduce the residual accumulation during gradient optimization, through minimizing the inter-domain transportation cost. Given a $P_1$ and $P_2$ being two distribution space with probability measures $\mu$ and $\nu$ respectively, the transportation $T$ preserving $P_1 \rightarrow P_2$ has equal total measure
$$\mu(T(p_1)) = \nu(p_2)
\label{eq:layer-ag}$$
where $p_1$ and $p_2$ is any measurable subset of $P_1$ and $P_2$. Then the total transportation cost for sending $p_1 \subset P_1$ to $p_2 \subset P_2$ by transportation cost $\tau(p_1,p_2)$ can be defined by
$$\min_{T:P_1 \rightarrow P_2}\int_{P_1}\tau(p_1,T(p_1))d\mu(p_1)
\label{eq:layer-ag-tp}$$
With minimizing the total transportation cost, the distribution $P_2$ progressively approximates $P_1$ on measures $\mu$. Assuming a series of neighbor layers ${\ell_k, ..., \ell_n}$ as set $\textit{\L}_k^n$, for sending parameter distribution $W_{\textit{\L}_k^n}$ to $\tilde{W}$ with measurable subset $w \subset W_{\ell_k, ..., \ell_n}$, the deep knowledge aggregation scheme merges the neighbouring layers to form the higher abstract parameters knowledge. In this case, the function $\Psi(\cdot)$ is formulated as
$$\Psi(\tilde{W}, W_{\textit{\L}_k^n}) = \min_{T:W_{\textit{\L}_k^n} \rightarrow \tilde{W}}\int_{W_{\textit{\L}_k^n}}\tau(w,T(w))d\mu(w)
\label{eq:layer-ag-details}$$
Optimization {#3dot4}
------------
Instead of directly optimizing the proposed objective function, we design an joint optimization method as the alternative solution. In details, our method uses two stages optimization to alternatingly solve the Eq. \[eq:interpret-KD-revised\].
#### Optimizing $\tilde{W^t}$ with $W^t_{\textit{\L}_k^n}$
Given an elaborate teacher network with parameter distribution $W^t_{\textit{\L}_k^n}$, we first aggregate the knowledge $\tilde{W^t}$ with $T({W^t_{\textit{\L}_k^n}})$ in here as:
$$\begin{split}
&\underset{\tilde{W^t}}{\arg\min} \sum_{i=1}^{N}\pounds(y_{i}^{t}(x_i, W^t), y_{i}^{s}(x_i, W^s))\\
&+ \gamma \int_{W_{\textit{\L}_k^n}^t}\tau(w^t,T(w^t))d\mu(w^t)
\label{eq:interpret-KD-revised-solve}
\end{split}$$
As the Eq. \[eq:interpret-KD-revised-solve\] involves a transportation cost and the definition of probability measures, it is difficult to directly integrate with gradient descent optimizer. In this case, we use the feature representation $\textit{F}_W$ as an approximation probability measures, which means the set of features maps $\textit{F}$ generated by parameters set $W$. If the transportation cost $\tau(\cdot)$ is defined as the simple $L_2$ distance, we revise the Eq. \[eq:interpret-KD-revised-solve\] as:
$$\begin{split}
\underset{\tilde{W^t}}{\arg\min} \sum_{i=1}^{N}&\pounds(y_{i}^{t}(x_i, W^t), y_{i}^{s}(x_i, W^s))\\
&+ \gamma \mu(W^t_{\textit{\L}_k^n})\|\textit{F}_{W^t_{\textit{\L}_k^n}} - \textit{F}_{\tilde{W^t}}\|_2
\label{eq:interpret-KD-further-revised-solved}
\end{split}$$
where $\gamma$ is a predefined parameter to control the penalty from optimal transportation. The $\mu(W^t_{\textit{\L}_k^n})$ as a measures function is used to penalize more on the layer with higher accumulation error, and the standard deviation is employed here. Moreover, we remove the part of terms during the derivation for Eq. \[eq:interpret-KD-revised-solve\] for fast computation. Then, the solution of $\tilde{W^t}$ can be obtained by gradient descent optimization.
#### Optimizing $W^s$ with $\tilde{W^t}$
Given an aggregate knowledge $\tilde{W^t}$, our goal here is further to solve the $W^s$ on student network with sparse recoding penalty, as: $$\begin{split}
\underset{W^s}{\arg\min} \pounds(\tilde{W^t}, W^s) + \lambda \sum_{j}\Phi_0(g_{j}^{s})
\label{eq:interpret-KD-revised-split2}
\end{split}$$
where $\Phi_0(g_{j}^{s})$ is designed for prompting the student network to firstly learn with the penalized gradients, and the parameter $\lambda$ is predefined to control the importance of the sparse recoding penalty.
Instead of directly solving the global optimum for objective function Eq. \[eq:interpret-KD-revised\], the two sub-objective functions Eq. \[eq:interpret-KD-further-revised-solved\] and Eq. \[eq:interpret-KD-revised-split2\] are designed to overcome the conflict between optimizing the prior knowledge and posterior knowledge simultaneously. Through alternatively minimizing the distribution dissimlarity $\pounds(\tilde{W^t}, W^s)$ and $\pounds(y^{t}, y^{s})$, the optimization for Eq. \[eq:interpret-KD-revised\] is regarded as an joint optimization procedure. Once the posterior knowledge is dominant during optimization, the optimizer for prior knowledge will penalize the total loss more, and the opposite is also. The gradient is only allowed to descend on the direction that makes both two optimizers are optimal.
Experiments
===========
In this section, we evaluate the proposed knowledge distillation framework with several benchmark datasets. For the base of experiments, we use the deep residual network [@he2016deep] as the network architecture, and the excerpt of the proposed framework in this architecture is shown in Fig. \[fig:residual\_module\]. The $c$ in residual module means the number of aggregated convolution layers. For the problem of optimizing these layers with different spatial scales, the identity mapping (ID) layer [@Yu2018Learning] is employed also. To ensure a fair comparison, the same data augment strategies are used. Moreover, we employ the similar settings of learning rates, optimization iterations and computation precision (32 float points). The implementation details will be shown in corresponding subsections.
![The excerpt of proposed framework on the residual network.[]{data-label="fig:residual_module"}](fig7.pdf){width=".47\textwidth"}
In Sec. \[sec\_penalty\_analysis\], through comparing with the typical penalties, the property of the sparse recoding penalty is analysed. Then, through comparing with the state-of-the-arts, we evaluate the performance of student networks in general image recognition tasks, and further explore their generalization ability in a revised dataset TCIFAR-100, as described in Sec. \[sec\_performance\_analysis\]. Finally, the discussions about the optimization procedure of the proposed framework is shown in Sec. \[sec\_optimization\_analysis\].
Analysis of Proposed Penalty {#sec_penalty_analysis}
----------------------------
As for the sparse recoding penalty, its property through comparing with typical methods is analysed, and we further explore the reason of why the proposed penalty is able to boost the convergence of knowledge distilling. Based on the experiment result, we address that the proposed penalty can be applied on other network optimization problems if the gradients distribution is not discriminative enough.
#### Penalty Property
Given a specific parameters distribution, the traditional penalties [@wen2016learning; @zhang2016l1] form a convex function and obtain the maximal reward in the unique extreme. It penalizes the parameter with higher value to reduce the total loss, for encouraging the value of parameter to close to 0. In contrast to these methods, the sparse recoding penalty is designed to penalize the gradients directly. For the propagated gradients, it filters the gradients with an equal reward within the learnable threshold, in order to slow down the update of inactive neurons. For the gradients out of the threshold, it boosts the update to highlight the dominant neurons. For validating our hypothesis, we visualize the convolutional kernels with the constraint by different penalties in image recognition tasks. The Fig. \[fig:kernel-analysis\] shows that the sparse recoding penalty can prompt the parameters distribution of the network to be more sparse, through directly regularizing the optimized gradients.
![With the penalized gradients, the sparse recoding penalty is also to produce more discriminative parameters distribution (best see in color).[]{data-label="fig:kernel-analysis"}](fig4.pdf){width=".42\textwidth"}
#### Convergence
We have observed fast convergence in our experiment result. In Fig. \[fig:analysis\_convergence\], it illustrates the training loss on MNIST over the beginning 20,000 iteraitons. The student network with sparse recoding penalty is better than the traditional penalties. We think one possible reason is that the proposed penalty is designed to penalize the gradients firstly, so it can produce a bigger step for gradients descent in the beginning of network training. Moreover, we evaluate the different types for initializing the parameters distribution in the experiment, and we also found the similar conclusion.
![Convergence speed; the traditional penalties [@wen2016learning; @zhang2016l1] and the sparse recoding penalty (best see in color).[]{data-label="fig:analysis_convergence"}](fig5.pdf){width=".47\textwidth"}
Performance Analysis {#sec_performance_analysis}
--------------------
In this section, we firstly conduct the experiments in the image recognition task on CIFAR-10, CIFAR-100 [@Krizhevsky2009] and ILSVRC 2012 [@Deng2009], in order to evaluate the performance of the proposed knowledge representing framework with the state-of-the-arts. Then, we design a TCIFAR-100 based on CIFAR-100, for further verifying their generalization ability. As the focus of this experiment is analysing the performance of student network with a small capacity, so we reserve the comparison on different tasks as future works.
### CIFAR-10
The CIFAR-10 is an image recognition dataset [@Krizhevsky2009] which includes 50,000 training images and 10,000 test images, and per training class has 5,000 images while test class has 1000 images. For all images, they store in RGB format with size of $32 \times 32$. We use a trained teacher network with 26 layers, which is structured as 5 residual modules. For student network, it contains 8 layers with 2 residual modules, which has roughly $1\//3$ parameters of the teacher. In details, with the same parameters settings and training strategies, we reduce about $1\//3$ number of the filters on each layer for the student network, in order to evaluate the case if the target network contains a weak representation ability. And we set the $c$ of knowledge aggregation as 3, which aggregates each three layer of teacher network into higher abstract level for one layer in student network.
Accuracy Params
----------------------------------- ----------- --------------
Teacher ResNet-26 91.91 $\sim$ 0.36M
Student ResNet-8 (Original) 87.91 $\sim$ 0.12M
FitNet [@Romero2014FitNets] 88.57 $\sim$ 0.12M
FSP [@yim2017gift] 88.70 $\sim$ 0.12M
Proposed-Dense 89.11 $\sim$ 0.09M
Proposed **90.65** $\sim$ 0.09M
NTS [@huang2017like] 88.98 $\sim$ 0.12M
KDGAN [@wang2018kdgan] 88.62 $\sim$ 0.12M
Proposed + KDGAN [@wang2018kdgan] **91.35** $\sim$ 0.09M
: ResNet-8 in CIFAR-10 Classification rates(%). Proposed: the KR framework. Proposed-Dense: the KR framework but removing the sparse recoding penalty.[]{data-label="tab:cifar10"}
In Tab. \[tab:cifar10\], it summarizes the obtained results. Based on the proposed framework, the student network which contains less parameters wins the methods [@Romero2014FitNets; @yim2017gift] focusing on prior knowledge with a significant improvement. For the state-of-the-arts [@huang2017like; @wang2018kdgan] by modeling the posterior knowledge, the proposed framework also achieves the comparable performance. For the self-comparison, we remove the sparse recoding penalty in KR framework and name it as the KR-Dense. And the experiment proves the importance to sparsely penalize the gradients during the distilling optimization, if the student network only has a small capacity. Besides, through combining with the KDGAN [@wang2018kdgan], a further improvement confirms that our method is flexible for the extension.
### CIFAR-100
The CIFAR-100 is an augmented version of CIFAR-10. It contains the same amount of images and size of CIFAR-10, which includes 50,000 training images and 10,000 test images, so only has 100 samples per class. Similar the setting to CIFAR-10, we use a trained teacher network with 32 layers as 6 residual modules, and student is composed of 14 layers as 3 residual modules. Besides, the reduction of about $1\//3$ filter number is still used, and $c$ is set as 3.
Tab. \[tab:cifar100\] shows results of student network with evaluated methods. Though the proposed method achieves the comparable performance than the state-of-the-arts [@huang2017like; @wang2018kdgan] with less parameters, the improvement for our method is not obvious. We think one possible reason is that the ResNet-14 has a stronger representation ability that the ResNet-8.
Accuracy Params
------------------------------------------ ----------- --------------
Teacher ResNet-32 64.06 $\sim$ 0.46M
Student ResNet-14 (Original) 58.65 $\sim$ 0.19M
FitNet [@Romero2014FitNets] 61.28 $\sim$ 0.19M
FSP [@yim2017gift] 63.33 $\sim$ 0.19M
Proposed Method 63.95 $\sim$ 0.17M
NTS [@huang2017like] 63.78 $\sim$ 0.19M
KDGAN [@wang2018kdgan] **63.96** $\sim$ 0.19M
Proposed Method + KDGAN [@wang2018kdgan] **63.98** $\sim$ 0.17M
: ResNet-14 in CIFAR-100 Classification rates(%). With the similar network architecture, we further reduce the output channels in each layer for saving the total parameters.[]{data-label="tab:cifar100"}
### ILSVRC 2012
The ILSVRC 2012 classification challenge involves the recognition task to classify one image into 1,000 leaf-node categories in the ImageNet hierarchy [@Krizhevsky2012]. It has about 1.2 million images for training, 50,000 for validation and 100,000 testing images. Although training the very deep network on such enormous datasets to achieve satisfied performance has been a solvable issue, how to obtain the comparable performance with a tiny network by the knowledge distillation still confuses the research community, especially for the methods [@luo2016face; @Romero2014FitNets; @yim2017gift] with prior knowledge. We think the major reason is that the depth of teacher network in ILSVRC 2012 is very deep, so the student network in these methods seriously suffers from the over-regularization.
Tab. \[tab:imagenet1\] shows the errors of Top-1 and Top-5. With the $c$ which is set as 4 in knowledge aggregation scheme, we found the situation of over-regularization is alleviated, and it prompts the KR framework to achieve the better performance.
Top-1 Top-5
------------------------------------------ ----------- ----------
Teacher ResNet-101 22.68 6.58
Student Inception-BN [@Bn2015] 25.74 8.07
FitNet [@Romero2014FitNets] 25.30 7.93
NTS [@huang2017like] 24.34 7.11
KDGAN [@wang2018kdgan] 24.11 6.98
Proposed Method **23.47** **6.85**
Proposed Method + KDGAN [@wang2018kdgan] **23.18** **6.79**
: ImageNet Classification errors (Top-1 and Top-5%). []{data-label="tab:imagenet1"}
### Generalization Ability
We further explore the generalization ability of previous methods and the proposed framework. Based on the data resource from CIFAR-100, we reproduce the CIFAR-100 as the TCIFAR-100 with the data distortion strategies. In details, each image in training and test set is distorted by the artifacts, from a gaussian distribution ($\sigma$ = 1) with the random sample. The Fig. \[fig:tcifar-100ep\] shows the examples. In Tab. \[tab:tcifar100\], it shows the proposed framework achieves a significant improvement than state-of-the-arts. We believe the KR framework is able to produce a student network with stronger generalization ability, since the joint optimization prevents the optimizer from being trapped in local extremum.
![*left* CIFAR-100; *right* TCIFAR-100[]{data-label="fig:tcifar-100ep"}](fig6bk){width=".4\textwidth"}
Accuracy Params
------------------------------------------ ----------- --------------
Teacher ResNet-32 61.25 $\sim$ 0.46M
Student ResNet-14 (Original) 54.37 $\sim$ 0.19M
FitNet [@Romero2014FitNets] 56.77 $\sim$ 0.19M
FSP [@yim2017gift] 57.31 $\sim$ 0.19M
Proposed Method **60.03** $\sim$ 0.17M
NTS [@huang2017like] 57.88 $\sim$ 0.19M
KDGAN [@wang2018kdgan] 58.15 $\sim$ 0.19M
Proposed Method + KDGAN [@wang2018kdgan] **60.33** $\sim$ 0.17M
: ResNet-14 in TCIFAR-100 Classification rates(%). The transformed CIFAR-100 dataset is reproduced by the CIFAR-100.[]{data-label="tab:tcifar100"}
Optimization Discussion {#sec_optimization_analysis}
-----------------------
In this section, we further discuss the implementation details of optimizing the proposed framework, and analysis the optimization procedure with different settings.
#### Implementation Details
As for the training on CIFAR-10 and CIFAR-100, the learning rate for Eq. \[eq:interpret-KD-further-revised-solved\] is set as 0.1, and was changed to 0.01, and 0.001 at two steps (30k and 48k) respectively. The optimizer for Eq. \[eq:interpret-KD-revised-split2\] started at a smaller learning rate 0.01, but also is reduced according to similar strategies. For the ILSVRC 2012, the learning rate for Eq. \[eq:interpret-KD-further-revised-solved\] is set as 0.1 with a ploy decreasing in each 6 epoch, and the optimizer for Eq. \[eq:interpret-KD-revised-split2\] started at learning rate 0.005. The weight decay of 0.00001 and momentum of 0.9 are all used. For the works related to quantization strategies [@jacob2018quantization; @rastegari2016xnor], we try to evaluate the performance if combining these works with our framework. Since the quantization techniques transfer the parameters distribution into a discrete space, we found the optimization will be seriously impacted and convergence performance also be influenced. However, this analysis is out of the scope of this paper, so it is left as future work.
#### Joint Optimization
For optimizing the $\tilde{W^t}$ with $W^t$ by Eq. \[eq:interpret-KD-further-revised-solved\] and the $W^s$ with $\tilde{W^t}$ by Eq. \[eq:interpret-KD-revised-split2\], we use two different optimizers to separately training these two sub-objective functions. Moreover, we tried different initialization techniques for parameters, and we found the objective function is harder to converge, if the initialization on $\tilde{W^t}$ is very different from $W^s$. We also consider the types for different optimizers [@adam2014; @Bottou2012; @rmsprop2012]. Through changing the two optimizers as Adam [@adam2014] or RMS [@rmsprop2012], we found it caused a performance oscillation but less than $1\%$.
Conclusion
==========
In this paper, we propose a knowledge representing (KR) framework mainly focusing on modeling the parameters distribution as prior knowledge. We suggest a knowledge aggregation scheme to represent the parameters knowledge from teacher network into more abstract level, for alleviating the phenomenon of residual accumulation in the deeper layers. We also design a sparse recoding penalty for constraining the student network to learn with the penalized gradients. It helps the student network to avoid the over-regularization during knowledge distilling and converge faster. In conclusion, the proposed framework can prompt the student network to preserve the key features of teacher network, even though the student network does not have a strong representation ability.
#### Acknowledgements.
We thanks all reviewers for providing the constructive suggestions.
| {
"pile_set_name": "ArXiv"
} |
---
address: |
School of Physics and Astronomy, The University of Manchester,\
Oxford Road, Manchester M13 9PL, United Kingdom
author:
- |
Krisztian Peters\
(on behalf of the CDF and 0 Collaborations)
title: 'Higgs Searches at the Tevatron [^1] '
---
Introduction
============
The Higgs boson is the last missing particle in the Standard Model (SM). Its mass is not determined by the SM, there are however several experimental constraints which bound the Higgs mass to values which are within the reach of the Tevatron collider. Lower bounds are given from direct searches at LEP2. These results exclude Higgs masses below 114.4 GeV at the 95% Confidence Level (C.L.) [@lep]. An upper bound on the Higgs mass is obtained by global electroweak fits. Especially radiative corrections to the $W$ mass from the Higgs and top quark play an important role. New precision measurements of the $W$ mass [@top] and the top mass [@W] from the Tevatron favor a light SM Higgs boson and yield an upper value of 144 GeV at 95% C.L. (or 182 GeV if the LEP2 limit is included) [@ewfit].
Experimental environment
========================
The Higgs searches are crucially dependent on performance of the Tevatron accelerator and detectors. Both, CDF and 0 detectors are currently performing close to their optimal design values, taking data with an efficiency of about 90%. The present Tevatron performance is matching the design values in terms of the current weekly integrated and peak luminosity. As of today, more than 2.5 fb$^{-1}$ have been delivered, with weekly integrated luminosity routinely reaching 50 pb$^{-1}$. If the accelerator keeps following the designed luminosity evolution, an integrated luminosity of about 8 fb$^{-1}$ will be achieved by the end of 2009, increasing the potential for a Higgs discovery at the Tevatron significantly.
Standard Model Higgs searches
=============================
Production cross sections for the SM Higgs boson at the Tevatron are rather small. They depend on the Higgs mass and are about 0.1 – 1 pb in the mass range of 100 – 200 GeV. The largest production cross section comes from gluon fusion, where the Higgs is produced via a quark loop. The second largest cross section, almost an order of magnitude smaller, is the associated production with vector bosons. At the mass range covered by the Tevatron, below 135 GeV the highest branching ratio is given by the decay to $b\bar b$ pairs and for masses above 135 GeV the Higgs boson decays mainly to $WW$ pairs.
These production and decay properties lead to the following search strategy at the Tevatron:
- For masses below 135 GeV the main search channels are the associated productions with vector bosons where the Higgs decays into $b\bar b$ pairs. In order to isolate the main background processes to these channels, an efficient b-tagging algorithm and a good dijet mass resolution are essential. The same final state produced via the gluon fusion process leads to a higher cross section but is overwhelmed by the huge multijet QCD background at a hadron collider.
- For masses above 135 GeV the search is mainly focused on the gluon fusion production process where the Higgs decays into $WW$ pairs.
$WH\to \ell\nu b\bar b$, $\ell=e,\mu$
-------------------------------------
For SM Higgs searches the most sensitive production channel at the Tevatron for a Higgs mass below 135 GeV is the associated production of a Higgs boson with a $W$ boson. Dominant backgrounds to the $WH$ signal are $W$ + heavy flavor production, $t\bar t$ and single-top quark production. Both, CDF and 0 performed cut based analyses with a rather similar approach. Both, electron and muon channels are studied here. The channels are separated in events having exactly one “tight” b-tagged jet (ST), and those having two “loose” b-tagged jets (DT) (with no overlap). The resulting four channels are analyzed independently to optimize the sensitivity and are later combined. Both experiments select events with isolated electrons or muons with $p_T>20$ GeV, require missing transverse energy above 20 GeV and two jets with $p_T>20$ GeV (0 ) or $p_T>15$ GeV (CDF). Cross section limits are derived from the invariant dijet mass distribution of the four individual analyses of each experiment and later combined. For $m_H=115$ GeV the observed (expected) limit is 1.3 (1.1) pb at 0 [@WHcutD0] and 3.4 (2.2) pb at CDF [@WHcutCDF], to be compared to the Standard Model cross section expectation of 0.13 pb. Thus the best expected measurement is a factor 8.8 higher than the SM expectation.
0 analyzed this channel also with the Matrix Element technique to separate signal from background. Like in the cut based analysis the four channels ($e$,$\mu$,ST,DT) are analyzed separately and later combined. The matrix-element-based technique attempts to make use of all the available kinematic information in the event to separate signal and background. Therefore leading order Matrix Elements are used to compute the event probabilities for signal and background. The present selection criteria is based on the single top search \cite{} and will be optimized in the future. Although this selection is not optimal for $WH$, the sensitivity of this search is similar to the sensitivity of the cut-based analysis and will improve with an optimized selection. For $m_H=115$ GeV the observed (expected) limit is 1.7 (1.2) pb with this present approach [@WHmatrixD0]. Limits for other Higgs masses together with the cut-based results are displayed in Fig.1.
$ZH\to \ell\ell b\bar b$, $\ell=e,\mu$
--------------------------------------
Similarly to $WH$ the Higgs boson can be produced associated with the $Z$ boson. First we focus on the channel where the $Z$ boson decays to a pair of electrons or muons with opposite sign. Here the $Z$ boson is reconstructed and identified from a pair of high $p_T$ leptons with an invariant mass constraint. Events are required to have b-tagged jets. The dominant backgrounds result from the associated production of a $Z$ boson with jets, among which the $Zb\bar b$ production is an irreducible background. Other main backgrounds are $t\bar t$, $WZ$, $ZZ$, and multijet production from QCD processes.
In the search at 0 at least two b-tagged jets are required. Cross section limits are then derived from the dijet invariant mass distribution within a search window. At CDF only 1 b-tagged jet is required. After this, a two dimensional Neural Network discriminates against the two largest backgrounds which are $Z$ + jets and $t\bar
t$. Limits are derived from the Neural Network distribution. For $m_H=115$ GeV the observed (expected) limit is 2.7 (2.8) pb at 0 [@ZHllCDF] and 2.2 (1.9) at CDF [@ZHllD0], to be compared to the Standard Model cross section expectation of 0.08 pb.
$ZH\to \nu\nu b\bar b$, $WH\to (\ell^{\pm})\nu b\bar b$
-------------------------------------------------------
The $ZH\to \nu\nu b\bar b$ channel benefits from the large $Z\to\nu\nu$ branching ratio. However it is challenging at hadron colliders due to the absence of visible leptons and the presence of only two jets in the final state. The two b-jets from the Higgs are boosted along the direction of the Higgs momentum and so tend to be more acoplanar than the dijet background. There are two major sources of background: physics backgrounds such as $Z$+jets, $W$+jets, electroweak diboson production or top quark production with missed leptons and jets and the instrumental background resulting from calorimeter mismeasurements which can lead to high $E_T$ signals with the presence of jets from QCD processes.
A result on this search channel was presented from CDF. Selecting events with a large $E_T > 75$ GeV and high $p_T$ b-tagged jets (leading jet $p_T > 60$ GeV), vetoing events with isolated leptons or where the missing $E_T$ is aligned in $\phi$ with jets eliminate much of the physics background. Two separate analyses are optimized for one or two b-tagged samples and later combined. Since the $WH$ channel with an undetected lepton has the same signature those events are taken into account in this search channel. For $m_H=115$ GeV the expected limit at CDF is a factor 15 higher than the Standard Model expectation [@ZHnnCDF].
{width=".75\textwidth"}
$H\to WW^{(*)}\to \ell^+\ell^-\nu\bar\nu$, $\ell=e,\mu$
-------------------------------------------------------
At Higgs masses above 135 GeV the biggest branching ratio is the decay to $WW$ pairs. With only leptons and missing energy in the final state the main background is $WW$ production without a large overlapping QCD background. Both, CDF and 0 analyzed this channel for the three combinations of electron and muon final states. Later the cross section limits have been combined.
The search strategy is to look for two high $p_T$, isolated, opposite sign leptons, require large missing transverse energy and veto on events with jets. Finally, the spin correlations in the decay of the Higgs boson are used. The leptons of the Higgs decay tend to have a small opening angle, whereas leptons from most of the backgrounds are expected to be back-to-back. Thus a cut on the opening angle between the leptons in the transverse plane $\Delta\phi_{\ell\ell}$ is mainly used to discriminate against the dominant $WW$ background. Since the Higgs mass cannot be directly reconstructed due to the neutrinos in the final state, the cross section limit is derived from the $\Delta\phi_{\ell\ell}$ distribution. For $m_H=160$ GeV, which yields the best sensitivity, the expected limit at CDF [@HWWCDF] is a factor 6 and at 0 [@HWWD0; @HWWmuD0] a factor of 5 higher than the Standard Model expectation.
Combined Standard Model Higgs limits
------------------------------------
The above presented channels can be combined which leads to a much more sensitive cross section limit throughout the whole discussed mass range. Both, 0 and CDF released results on the SM Higgs combination, the obtained results can be found in [@comb]. A further, important increase of the sensitivity can be gained from a combination of the CDF and 0 results. Such a first Tevatron combination limit was released Summer 2006, the result is plotted in Fig.2. The expected combined limits are a factor of 7.5 at $m_H=115$ GeV and a factor of 4 at $m_H=160$ GeV away from the Standard Model expected cross sections. It should be stressed that this result does not include CDF’s new 1 fb$^{-1}$ high mass results and it does not include any of 0 ’s new 1 fb$^{-1}$ low mass results yet. Further significant improvements are expected when all the 1 fb$^{-1}$ results will be included. Such a new Tevatron combination is planned for the Summer 2007.
{width=".75\textwidth"}
MSSM Higgs searches
===================
The Minimal Supersymmetric Standard Model (MSSM) predicts two Higgs doublets leading to five Higgs bosons: a pair of charged Higgs boson ($H^\pm$); two neutral CP-even Higgs bosons ($h$,$H$) and a CP-odd Higgs boson ($A$). At tree level, the Higgs sector of the MSSM is fully described by two parameters, which are chosen to be the mass of the CP-odd Higgs, $m_A$, and $\tan\beta$ , the ratio of the vacuum expectation values of the two Higgs doublets. The Higgs production cross-section is enhanced in the region of low $m_A$ and high $\tan\beta$ due to the enhanced Higgs coupling to down-type fermions. This makes it possible to search in the MSSM for $\tau\tau$ final states, which would be very challenging in the SM due to the large irreducible background of $Z\to\tau\tau$. In the low $m_A$, high $\tan\beta$ region of the parameter space, Tevatron searches can therefore probe several MSSM benchmark scenarios extending the search regions covered by LEP [@lepmssm].
Both, CDF and 0 performed a search for the neutral MSSM Higgs decaying to $\tau$ pairs, where one of the $\tau$-leptons is decaying in the leptonic and the other one in the hadronic mode. 0 ’s result covers so far only the $\mu$-channel, CDF’s result is a combination of the electron and muon channels, including $\tau_e\tau_\mu$.
A set of Neural Networks (NN) is used at 0 to discriminate $\tau$-leptons from jets. An isolated muon is required, separated from the hadronic $\tau$ with opposite sign. A cut on the visible $W$ mass removes most of the remaining $W$ boson background. Further optimized NNs are used for signal discrimination. In the cross section limit calculation the output of the NNs for different tau types is used.
CDF uses a variable cone size algorithm for $\tau$ discrimination. An isolated muon or electron is required, separated from the hadronic $\tau$ with opposite sign. Most of the $W$ background is removed by a requirement on the relative directions of the visible $\tau$ decay products and the missing transverse energy. Cross section limits are derived from the visible mass distribution.
For both experiments the data is consistent with the background only observation. Exclusion regions in the $\tan\beta$ – $m_A$ plane can be derived for different MSSM benchmark scenarios. Both experiments obtained similar results [@tauCDF; @tauD0; @mark]. In the region of $90 < m_A < 200$ GeV, $\tan\beta$ values larger than 40-60 are excluded for the no-mixing and the $m_h^{max}$ benchmark scenarios. Examples of such exclusion regions are shown in Fig.3. In CDF’s result the observed limits are weaker than the expectations due to some excess of events in the data sample with a significance of approximately 2$\sigma$.
![Excluded regions by the CDF and 0 experiments in the $\tan\beta$-$m_A$ plane for $\mu<0$ in the $m_h^{max}$ and the no-mixing scenario. []{data-label="taud0"}](tanb-mupos.eps "fig:"){width=".45\textwidth"}![Excluded regions by the CDF and 0 experiments in the $\tan\beta$-$m_A$ plane for $\mu<0$ in the $m_h^{max}$ and the no-mixing scenario. []{data-label="taud0"}](H29F07.eps "fig:"){width=".55\textwidth"}
Perspectives
============
Today some single channels have cross section limits similar to the combined Tevatron results obtained half a year ago. With Tevatron’s excellent performance matching the designed delivered weekly luminosities, a significant amount of sensitivity will be gained by an increase of the luminosity by about a factor of 8. There is already 2.5 times more data on tape than used for the presented results. In addition, the inclusion of more channels in the Higgs search (for example $\tau$-final states) will gain additional sensitivity. Dijet mass resolution, b-tagging and simulation are important ingredients for Higgs searches and both experiments are continuously improving at these scopes. Still a lot of improvements are expected in analyses techniques. Especially the use of multivariate techniques, like Neural Networks, Decision Trees and Matrix Element analyses shall bring further important improvements. 0 ’s recent evidence for Single Top production and CDF’s $WZ$ observation is an important milestone in the use of these techniques to discriminate very low rate signals in the presence of substantial backgrounds.
Acknowledgments {#acknowledgments .unnumbered}
===============
I would like to thank my colleagues from the CDF and 0 Collaborations working on this exiting topic and for providing material for this talk. I also like to thank the organizers of Rencontres de Moriond for a stimulating conference and the European Union “Marie Curie” Programme for their support.
References {#references .unnumbered}
==========
[99]{}
ALEPH, DELPHI, L3 AND OPAL Collaborations, [*Phys. Lett.*]{} [**B 565**]{} (2003) 61. E. Barberis for the CDF and 0 Collaborations, these proceedings. O. Stelzer-Chilton for the CDF Collaboration, these proceedings. LEP Electroweak Working Group, http://lepewwg.web.cern.ch/LEPEWWG/ 0 Collaboration, 0 note 5357-CONF. CDF Collaboration, CDF/ANAL/EXOTIC/PUBLIC/8390. 0 Collaboration, 0 Note 5365-CONF. CDF Collaboration, CDF/ANAL/EXOTIC/PUBLIC/8422. 0 Collaboration, 0 Note 5275-CONF. CDF Collaboration, CDF/ANAL/EXOTIC/PUBLIC/8442. CDF Collaboration, http://www-cdf.fnal.gov/physics/exotic/r2a/20070131.HWW/\
/HwwCutsBlessed/HwwCutsBlessed.html. 0 Collaboration, 0 Note 5194-CONF. 0 Collaboration, 0 Note 5194-CONF. 0 and CDF Collaborations, CDF Note 8384, 0 Note 5227. The ALEPH, DELPHI, L3 and OPAL Collaborations, LHWG-Note 2005-01. CDF Collaboration, CDF Note 8676. 0 Collaboration, 0 Note 5331-CONF. M. Owen for the 0 Collaboration, these proceedings, .
[^1]: Presented at Recontres de Moriond EW 2007, 10-17 March 2007, La Thuile.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A large majority of cellular networks deployed today make use of Frequency Division Duplexing (FDD) where, in contrast with Time Division Duplexing (TDD), the channel reciprocity does not hold and explicit downlink (DL) probing and uplink (UL) feedback are required in order to achieve spatial multiplexing gain. In order to support massive MIMO, i.e., a very large number of antennas at the base station (BS) side, the overhead incurred by conventional DL probing and UL feedback schemes scales linearly with the number of BS antennas and, therefore, may be very large. In this paper, we present a new approach to achieve a very competitive tradeoff between spatial multiplexing gain and probing-feedback overhead in such systems. Our approach is based on two novel methods: (i)an efficient regularization technique based on Deep Neural Networks (DNN) that learns the Angular Spread Function (ASF) of users channels and permits to estimate the DL covariance matrix from the noisy i.i.d. channel observations obtained freely via UL pilots (UL-DL covariance transformation), (ii)a novel “*sparsifying precoding*” technique that uses the estimated DL covariance matrix from (i) and imposes a *controlled sparsity* on the DL channel such that given any assigned DL pilot dimension, it is able to find an optimal sparsity level and a corresponding sparsifying precoder for which the “*effective*” channel vectors after sparsification can be estimated at the BS with a low mean-square error. We compare our proposed DNN-based method in (i) with other methods in the literature via numerical simulations and show that it yields a very competitive performance. We also compare our sparsifying precoder in (ii) with the state-of-the-art statistical beamforming methods under the assumption that those methods also have access to the covariance knowledge in the DL and show that our method yields higher spectral efficiency since it uses in addition the instantaneous channel information after sparsification.'
author:
- 'Mahdi Barzegar Khalilsarai, Yi Song, Tianyu Yang, Saeid Haghighatshoar, and Giuseppe Caire [^1]'
bibliography:
- 'references2.bib'
title: 'Uplink-Downlink Channel Covariance Transformations and Precoding Design for FDD Massive MIMO'
---
[f\_[ul]{}]{}[f\_]{} [f\_[dl]{}]{}[f\_]{}
Massive MIMO, Sparse Scattering, Angular Spread Function (ASF), Uplink-Downlink Covariance Transformation (UDCT), Deep Neural Networks (DNNs), Sparisfying Precoder.
Introduction {#sec:intro}
============
Massive Multiple-Input Multiple-Output (MIMO) is a variation of conventional multi-user MIMO, where base station (BS) has a much larger number of antennas (or antenna ports) $M \gg 1$, and is considered to be a key technology for the next generation of wireless networks [@larsson2014massive]. Large number of antennas permits to multiplex $K \gg 1$ data streams over the spatial domain to serve $K$ users and guarantees significant advantages such as energy efficiency due to a large beamforming gain, reduced inter-cell interference, and simple user scheduling and rate adaptation due to the well-known channel hardening phenomenon [@larsson2014massive]. To achieve such benefits, especially in the a downlink (DL) scenario we are interested in this paper, the BS needs to learn the channel vectors of $K$ users to $M$ BS antennas in the DL. With [*Time Division Duplexing*]{} (TDD), due to uplink-downlink (UL-DL) channel reciprocity which holds under suitable calibration [@marzetta2006much; @Huh11], this can be done via transmitting mutually orthogonal pilots from the users only in the UL. Unfortunately, channel reciprocity does not hold in [*Frequency Division Duplexing*]{} (FDD) since UL and DL channels lie on disjoint and far-separated frequency bands. Consequently, the only way to learn the DL channel is to devote a fraction of *resource elements* (RE) [@lin20185g] to estimate the DL channel of $K$ users by transmitting DL pilots and then UL feedback. Conventional DL training consists of the transmission of an $M \times {T_{\rm dl}}$ pilot matrix over ${T_{\rm dl}}$ REs, such that ${T_{\rm dl}}\geq M$ to permit each user to estimate its own DL channel vector. Then, the users feed their estimated channels back to the BS via the UL channel. Although this method works quite well for conventional MIMO systems with moderately small number of antennas $M$, it is quite inefficient in massive MIMO since $M\gg 1$ and full training of DL channel wastes at least ${T_{\rm dl}}=M$ REs which may exhaust or be even larger than the whole REs available in the DL. This feedback bottleneck makes implementing massive MIMO in FDD quite challenging.
To overcome this bottleneck, several works have been proposed to reduce DL training and feedback overhead using the sparse structure of the channel in the angle-of-arrival (AoA) domain. This sparsity arises due to the fact that FDD systems in 5G will be mainly used for large cells (while TDD for smaller denser cells) with tower-mounted base stations [@shepard2012argos], where the communication between the users and the BS occurs through a sparse cluster of scatterers with limited angular support (see, e.g., Fig.\[fig:scat\_chan\]).
![Sparse scattering channel between a generic user and the BS. In this example, the channel consists of 3 large scatterers reflecting the power of the user to the BS array.[]{data-label="fig:scat_chan"}](scat_chan2.pdf)
As a result, the effective dimension of the channel $s$ is much less than $M$. Building on this idea, these works proposed using compressed DL pilots and [*Compressed Sensing*]{} (CS)-based channel recovery techniques [@rao2014distributed; @gao2015spatially; @kuo2012compressive]. From standard results in CS, these method require a pilot dimension of order ${T_{\rm dl}}= O(s \log M)$, which may be much less than $M$ for very sparse channels $(s \ll M)$. Although the assumption $s \ll M$ may be met in very sparse Line-of-Sight (LoS) scattering scenarios, it is not fulfilled when the propagation occurs through diffuse clusters of moderately large AoA support size because in those cases the effective dimension $s$ also scales like $s=\alpha M$ where $\alpha$ is proportional to the angular width of the scatterer. In those cases, CS methods incur a pilot dimension overhead of ${T_{\rm dl}}= O(\alpha M \log M)$ which still grows proportionally to $M$. Overall, although the CS methods are able to exploit the sparsity of the channel they are still at the mercy of sparsity induced by the propagation environment since they are unable to “shape the channel sparsity” as desired.
Contribution
------------
In this paper we propose an efficient scheme for realizing FDD massive MIMO. The fundamental assumption we make to achieve this goal is that, the channel is a Gaussian process with an *Angular Spread Function* (ASF) that remains the same for UL and DL. In particular, this implies that although the UL-DL reciprocity may not hold for the instantaneous channel vectors, some sort of statistical reciprocity still holds as the ASF does not vary between UL and DL. We can summarize our contributions in this paper as follows.
[**(i) A Novel UL-DL Covariance Transformation Using Deep Neural Networks.** ]{} We use the reciprocity of the ASF in UL and DL to estimate the DL covariance matrix from the observation of noisy channel vectors freely available through the UL pilots. The ASF reciprocity was also assumed implicitly in several past works [@hochwald2001adapting; @han2010potential; @decurninge2015channel] by considering only ASFs with spike discrete components. Recent works [@haghighatshoar2018multi; @miretti2018fdd; @cavalcante2018error] generalize these results to arbitrary ASFs consisting of both discrete spike and also continuous components, and show that such Uplink-Downlink covariance transformation (UDCT) is still feasible for sufficiently large number of BS antennas $M$ although establishing the stability of UDCT is quite challenging when only a noisy estimate of the UL covariance matrix is available.
One of the aims of this paper is to improve the performance of UDCT by exploiting the natural group sparsity structure of the ASF. Unfortunately, designing good regularization methods that promote the group-sparsity of the ASF (see, e.g., Fig.\[fig:scat\_chan\]) is not trivial. In particular, none of the structured group-sparse estimators widely adopted in sparse signal recovery literature are applicable here [@baraniuk2010model; @eldar2009block]. In this paper, we bridge this gap by using Deep Neural Networks (DNNs). We train DNNs with suitable training data corresponding to group-sparse ASFs and let them learn/capture the notion of group-sparsity. Once the DNN is suitably trained, we use it as a black-box algorithm to estimate the ASF, thus, the DL covariance matrix, from the noisy channel vectors gathered in the UL. We illustrate via numerical simulations that our proposed DNN-based method is quite strong and recovers the DL covariance matrix much better than other methods.
[**(ii)A Novel Sparsifying Precoder.**]{} We use the estimated DL covariance matrix from (i) to design a [*sparsifying precoder*]{}, such that the following two criteria are met: (a)the reduced-dim channel vectors of all the users after sparsification are sparse enough such that they are stably estimated with a very low error and (b)the dimension of the channel after sparsification is kept as large as possible in order to increase the rank of the effective channel matrix of the users after sparsification such that large number of users $K$ can be served. We pose this as a Mixed Integer Linear Program (MILP) and solve it via off-the-shelf MILP solvers. It is also worthwhile to mention that compared with CS methods, which are at the mercy of the sparsity of the channel due to the propagation, our proposed method is able to shape the sparsity of the channel completely flexibly depending on number of REs ${T_{\rm dl}}$ available for channel estimation. We illustrate via numerical simulations that the proposed method has an excellent performance much superior to other methods such as statistical beamforming, which also use the knowledge of the DL covariance matrix.
Notation
--------
We denote vectors/matrices with small/large boldface letters (e.g., $\bfx$/$\bfX$), and sets with calligraphic letters (e.g., $\clX$). We use the $i$-th element of a vector $\bfx$ with $[\bfx]_i$ and the $(i,j)$-th element of a matrix $\bfX$ with $[\bfX]_{i,j}$. For an integer $k$, we use the short-hand notation $[k]$ for $\{1,\dots, k\}$.
Proposed ASF Estimation and Uplink-Downlink Covariance Transformation {#sec:dnn}
=====================================================================
In this section, we explain our proposed method for UDCT using DNNs. For the sake of completeness, we first provide a summary of UDCT problem for the Uniform Linear Array (ULA) we consider in this paper (see, Fig.\[fig:scat\_chan\]). Let us consider a generic use and let us denote the ASF of this user by $\gamma(\xi)$ where $\xi=\sin(\theta) \in [-1,1]$ denotes a parametrization of the AoA $\theta\in[-\frac{\pi}{2}, \frac{\pi}{2}]$ and where $\gamma(\xi)$ is density of the received signal power at the AoA $\xi$. The covariance matrix of the channel vector of this user at UL/DL carrier frequency $f \in \{{f_{\rm ul}}, {f_{\rm dl}}\}$ is given by [@haghighatshoar2018multi] $$\begin{aligned}
\label{cov_cont}
\Sigmam(f)=\int_{-1}^1 \gamma(\xi) \bfa(\xi, f) \bfa(\xi,f)^\herm d\xi,\end{aligned}$$ where $\bfa(\xi,f)$ denotes the array response vector at AoA $\xi$ at frequency $f$ given by $$\begin{aligned}
[\bfa(\xi,f)]_k=e^{j (k-1) \pi d \frac{f}{c_0} }, k \in [M],\end{aligned}$$ where $d$ denotes the antenna spacing and where $c_0$ is the speed of light. Note that in we assumed implicitly that the ASF $\gamma(\xi)$ is the same in both UL and DL frequency range. As explained before, this provides some sort of statistical UL-DL channel reciprocity for FDD (in contrast with the instantaneous channel reciprocity which may not hold). With this notation, can pose the UDCT problem as follows.
[**UDCT Problem:**]{} [*Given the UL covariance matrix or an estimate thereof $\Sigul:=\Sigmam(f_\text{ul})$, find the DL covariance matrix $\Sigdl:=\Sigmam(f_\text{dl})$.*]{} $\lozenge$
For the ULA, we can gain a better understanding of UDCT by looking at the Fourier coefficients of the ASF $\gamma(\xi)$. We first assume that the array has the standard half wavelength spacing in the UL, namely, $d=\frac{\lambda_\text{ul}}{2}$ where $\lambda_\text{ul}=\frac{c_0}{{f_{\rm ul}}}$ denotes the wavelength at UL carrier frequency. Then, it is not difficult to show that $\Sigul$ is a Toeplitz matrix whose first column is given by $\sigul(\gamma) \in \bC^M$ where $$\begin{aligned}
\label{UL_proj}
[\sigul(\gamma)]_k=\int_{-1}^1 \gamma(\xi) e^{j (k-1) \pi \xi} d\xi,\ \ k\in [M],\end{aligned}$$ denotes the $(k-1)$-th Fourier coefficient of $\gamma(\xi)$. Similarly, $\Sigdl$ is also a Toeplitz matrix with first column $\sigdl(\gamma)\in \bC^M$ $$\begin{aligned}
\label{DL_proj}
[\sigdl(\gamma)]_k=\int_{-1}^1 \gamma(\xi) e^{j (k-1) \pi \beta \xi} d\xi,\ \ k\in [M],\end{aligned}$$ where $\beta=\frac{{f_{\rm dl}}}{{f_{\rm ul}}}$ denotes the ratio between the DL and UL carrier frequencies. In current deployments of FDD systems typically ${f_{\rm ul}}>{f_{\rm dl}}$, thus, $\beta>1$. Therefore, in the ideal case where $\Sigul$ is known exactly, one can pose UDCT as recovering the projections of the positive functions $\gamma(\xi)$ on the DL set of harmonic functions $\clH_\text{dl}=\{e^{j (k-1) \pi \beta \xi}: k\in [M]\}$ from the knowledge of its projections on the UL harmonic functions $\clH_\text{ul}=\{e^{j (k-1) \pi \xi}: k\in [M]\}$.
All the UDCT algorithms in the literature implicitly or explicitly do the following: (a)estimate a positive function $\widehat{\gamma}(\xi)$ that has exactly (in the noiseless case) or approximately (in the noisy case) the same projections on the UL harmonic functions $\clH_\text{dl}$ as the original ASF $\gamma(\xi)$, (b)use the resulting estimate $\widehat{\gamma}(\xi)$ to compute the projections onto the DL set $\clH_\text{dl}$ to recover an estimate of the DL covariance matrix. Note that even in the ideal noiseless case, the mapping from the ASF to the DL projections is a linear map from the infinite-dim space of positive functions to the finite-dim space of $M$ projections. Therefore, there are generally a large set of ASFs corresponding to a given UL projection $\sigul(\gamma)$ produced by a generic ASF $\gamma(\xi)$: $$\begin{aligned}
\clA(\gamma):=\big \{\mu(\xi): \int_{-1}^1 \mu(\xi) e^{j (k-1)\pi \xi} d \xi=[\sigul(\gamma)]_k\big \}.\end{aligned}$$ All the UDCT algorithms can be seen in one way or other as simply different strategies for selecting a specific candidate $\alg(\sigul(\gamma)) \in \clA(\gamma)$ according to specific criteria. One such criterion is to assume that the original ASF $\gamma$ belongs to a specific subset $\clA_0$ of structured ASFs. Also, the UDCT algorithm can written more generally as a method that produces $\alg(\sigul(\gamma)) \in \clA(\gamma) \cap \clA_0$ when fed with the UL projections $\sigul(\gamma)$.
[\_[ul]{}]{}[\_]{} A similar argument applies to the more general case where instead of $\Sigul$ one has access to a collection of $N$ i.i.d. noisy UL channel vectors $\bfy(s)={\check{\bfh}_{\rm ul}}(s)+\bfz(s)$, $s\in [N]$, gathered via UL pilot transmission where ${\check{\bfh}_{\rm ul}}(s)$ and $\bfz(s)$ denote the UL channel vector and additive measurement noise over resource block $s \in [N]$. Then, one can compute the UL sample covariance matrix as $$\begin{aligned}
\label{ul_samp_cov}
\widehat{\Sigul} =\frac{1}{N} \sum_{s\in [N]} \bfy(s) \bfy(s)^\herm,\end{aligned}$$ and design an algorithm $\alg$ that produces a structured ASF in $\clA_0$ whose UL covariance matrix matches $\widehat{\Sigul}$ under suitable metric. For example, [@haghighatshoar2018multi; @miretti2018fdd] propose such algorithms using Non-Negative Least Squares (NNLS) and $\ell_2$-norm projections. However, none of these methods are able to capture the group-sparsity of the ASF in the AoA domain as illustrated in Fig.\[fig:scat\_chan\]. For example, as we will illustrate via numerical simulations, the NNLS proposed in [@haghighatshoar2018multi] is able to promote sparsity of ASF in the AoA domain but it does not yield necessarily group-sparse ASF. The $\ell_2$ projection method [@miretti2018fdd], in contrast, is able to produce smooth ASFs but creates out-of-support components in the estimated ASF since $\ell_2$ norm is inherently unable to promote sparsity.
Overall it is generally difficult to design suitable regularization methods that promote notion of group-sparsity of ASF we address here. In particular, none of the structured group-sparse estimators widely adopted in sparse signal recovery literature are applicable here [@baraniuk2010model; @eldar2009block]. In this paper, we develop such a group-sparsity promoting regularization using DNNs as follows.
[**(a) Training Data.**]{} We first consider a class of group-sparse ASFs $\clA_0$ that may potentially arise in practical propagation scenarios. To capture the notion of group sparsity, we assume that each ASF in $\clA_0$ can be written as $\gamma(\xi)=\sum_{i=1}^g \kappa_i p_i(\xi)$ where $g$ denotes the number of groups, where $p_i(\xi)$ is a normalized, i.e., $\int_{-1}^1 p_i(\xi) d\xi=1$, positive function with connected support in $[\xi_i, \xi_i+w_i]$ with $w_i$ denoting the angular width of the ASF of the $i$-th group $p_i(\xi)$, and where $\kappa_i \in [0,1]$ with $\sum_{i=1}^g \kappa_i=1$ are the normalized weights corresponding to ASFs in $g$ groups. For example, Fig.\[fig:scat\_chan\] corresponds to an ASF with $g=3$ groups.
We use the ASFs in $\clA_0$ to train a DNN as follows. Given a training sample size $S$, we select $S$ ASFs from $\clA_0$ completely randomly. For each specific ASF $\gamma$ inside this training set, we compute the first row of the UL covariance matrix as and produce $N$ i.i.d. noisy UL channel vectors and their corresponding sample covariance $\widehat{\Sigul}$ as in . Since the covariance matrices for ULA are Toeplitz, we toeplitizify $\widehat{\Sigul}$ and define the first column of the resulting Toeplitz matrix as $\widehat{\sigul}$ where $$\begin{aligned}
\label{toeplitzif}
[\widehat{\sigul}]_k=\frac{\sum_{i=1}^{M-k+1} [\widehat{\Sigul}]_{i,i+k}}{M-k+1}.
\end{aligned}$$ We define a uniform quantization grid $\clG:=\{\xi_i: i \in [G]\}$ over the set of AoAs $[-1,1]$ of size $G \gg M$ where $\xi_i=-1+\frac{2(i-1)}{G}$ denotes the $i$-th quantization point. We also define the discrete quantization of the ASF $\gamma(\xi)$ over the grid $\clG$ as $\gammam=(\gamma(\xi_1), \dots, \gamma(\xi_G))^\transp\in \bR_+^G$. For simplicity, we always normalize $\gammam$ to make sure that $\sum_{i=1}^G [\gammam]_i=1$. Finally, we use $(\widehat{\sigul}, \gammam)$ as input-output labeled pair for training the DNN illustrated in Fig.\[DNN\]. By repeating this for all the $S$ ASFs selected for training, we obtain a collection of $S$ training samples for DNN.
![The structure of DNN adopted for ASF estimation. DNN consists of $5$ layers with $2M$, $4M$, $8M$, $16M$, and $G$ neurons, where $M$ is the number of antennas and where $G$ is the ASF quantization grid size. The last layer has a *soft-max* activation function and produces positive values for $\gammam\in \bR_+^{G}$. []{data-label="DNN"}](dnn_config.pdf)
[**(b) Supervised Learning.**]{} We use these $S$ training samples to train a *Deep Neural Networks* (DNN). For training, we use the widely-adopted *Stochastic Gradient Descend* (SGD) with $\ell(\gammam, \widehat{\gammam})=\|\gammam - \widehat{\gammam}\|_1$ as the loss function between the true $\gammam$ and the estimate $\widehat{\gammam}$ generated by the network. DNNs have recently been of tremendous importance in Machine Learning and also in many applications in wireless communications, such as signal detection, channel encoding [@farsad2018deep], and decoding [@nachmani2018deep].
[**(c) Structure of DNN.**]{} One of the important factors affecting the performance of the ASF estimation using DNNs is the structure of the DNN consisting of the number of layers, the number of neurons in each layer, and the activation function of each layer. In this paper, we use a fully-connected network illustrated in Fig.\[DNN\], with $5$ layers consisting of $2M$, $4M$, $8M$, $16M$, and $G$ neurons, respectively, where $M$ denotes the number of antennas, where the number of neurons $G$ in the last layer corresponds to the grid size we are adopting for ASF quantization. The activation function of the $4$ initial layers is the RelU function $x \mapsto \max\{x,0\}$. For the last layer we use the *soft-max* activation function, which for an input vectors $(x_1, \dots, x_G) ^\transp$ in the input produces the output as $(x_1, \dots, x_{G})^\transp \mapsto \frac{(e^{x_1}, \dots, e^{x_{G}})^\transp}{\sum_{j=1}^{G} e^{x_j}}$. Note that the summation of the elements produced by soft-max layer is always $1$, which produces a normalized $\gammam$, i.e., $\sum_{i=1}^G\gamma_i=1$, as desired.
[**(d) UDCT using DNN.**]{} Once the DNN was suitably trained we use it as a black-box algorithm for ASF estimation and UDCT. More specifically, given $N$ i.i.d. noisy channel vectors received in UL, we compute the UL sample covariance matrix as in and $\widehat{\sigul}$ as in . Then, we feed $\widehat{\sigul}$ to the trained DNN and obtain an estimate of the quantized ASF $\widehat{\gammam}\in \bR_+^G$ where $G$ is the grid size. Then, we build an estimate of the original continuous ASF $\widehat{\gamma}(\xi)$ by triangular interpolation. Finally, we use $\widehat{\gamma}(\xi)$ to compute the first column of the DL Toeplitz covariance matrix as in . Note that DNNs have the fundamental advantage that one does not need to run any iterative algorithm as in [@haghighatshoar2018multi; @miretti2018fdd], which may require many iteration to converge; instead one immediately computes the estimate by straightforward calculations through the network, which can be done even in parallel to obtain a tremendous speed-up.
Sparsification Precoding {#sec:ch_sparsification}
========================
In this section, we provide an step-by-step method to design the sparsifying precoder for a collection of $K$ users based on their estimated DL covariance matrix denoted by ${\Sigmam_k}$, $k \in [M]$, where for simplicity we drop the label ‘dl’. An essential requirement for the our method is the existence of the *common eigen-vector property* for the array and in particular for the set of $K$ DL covariance matrices ${\Sigmam_k}$, $k \in [K]$. In brief, this property is satisfied when the set of all the DL covariance matrices produced by the array is, at least asymptotically for large $M$, diagonalizable in the same basis. Fortunately, this assumption is met for Uniform Linear/Planar Array (ULA/UPA) widely adopted in wireless applications since the resulting covariance matrices have Toeplitz/Block-Toeplitz structure. And, it is well-known that, asymptotically for large $M$, all such matrices are diagonalizable with appropriate DFT (Discrete Fourier Transform) matrices, thus, common eigen-vector property.
In the following, we provide a detailed derivation of the this property for the ULA. We first calculate the circulant approximation of the estimated DL covariance matrices of all users, so that they all share the same set of eigen-vectors, or equivalently virtual beam-space representation.[^2] The circulant approximation of large Toeplitz matrices imposes an small error as a result of the application of Szegö Theorem [@adhikary2013joint]. Let ${\Sigmam_k}$, $k\in[K]$, be as before and define the diagonal matrices $\mathring{\Lambdam}_k = \text{diag}(\Fm^\herm {\Sigmam_k}\Fm)$ for $k \in [K]$. There are several ways to define a circulant approximation [@zhu2017asymptotic], among which we choose ${\mathring{\Sigmam}_k}= \Fm \mathring{\Lambdam}_k\Fm^\herm$. According to Szegö’s theorem, for large $M$, $\mathring{\Lambdam}_k$ converges to the diagonal eigenvalue matrix $\Lambdam_k$ of ${\Sigmam_k}$, i.e. $\mathring{\Lambdam}_k \rightarrow \Lambdam_k$ as $M\rightarrow \infty$. This shows that, with a small error, we can find a set of common eigenvectors for all the estimated DL covariance matrices ${\Sigmam_k}$, $k \in [K]$. As a consequence, the DL channel covariance of each user $k\in [K]$ is characterized simply by a vector of eigenvalues $\lambdav_{k} \in \bR^M$, with $m$-th element $\lambda^{(k)}_m = [\mathring{\Lambdam}^{(k)}]_{m,m}$. In addition, the DFT matrix whose $(m,n)$-th entry is given by $[\Fm]_{m,n}=\frac{1}{\sqrt{M}} e^{-j2\pi \frac{mn}{M}},~m,n\in [M]$, forms a unitary basis for (approximately) expressing any user channel vector via an (approximated) Karhunen-Loeve expansion. In particular, letting $\fv_n:=[\Fm]_{\cdot,n}$ denote the $n$^th^ column of $\Fm$, we can express the DL channel vector of user $k$ as $$\label{eq:approximate-KL}
\hv^{(k)} \approx \sum_{m=0}^{M-1} g_{m}^{(k)} \sqrt{\lambda_{m}^{(k)}} \fv_n
$$ where $g_{m}^{(k)} \sim \cg (0,1) $. Here, we wish to design the precoder $\Bm$ such that the support of the effective channels ${{\check{\bfh}}_{\rm eff}^{(k)}}= \Bm \hv^{(k)}$ is not larger than ${T_{\rm dl}}$, such that all users have a chance of being served. Let $\Hm = \Lm \odot \bG\in \bC^{M\times K}$ denote the matrix of DL channel coefficients expressed in the DFT basis (\[eq:approximate-KL\]), in which each column of $\Hm$ represents the coefficients vector of a user, where $\Lm$ is a $M \times K$ matrix with elements $[\Lm]_{m,k} = \sqrt{\lambda_m^{(k)}}$, where $\bG \in \bC^{M\times K}$ has i.i.d. elements $[\bG]_{m,k} = g_{m}^{(k)}$, and where $\odot$ denotes the Hadamard (elementwise) product.
To design the sparsifying precoder, we first illustrate the joint user-beam association of all $K$ users by a graphical model. Let $\Am = [\Lm]$ denote a one-bit thresholded version of $\Lm$, such that $[\Am]_{m,k}=1$ if $\lambda_m^{(k)} > \epsilon$, where $\epsilon > 0$ is a suitable small threshold, used to identify the components that are significantly larger than 0 from the “almost zero” ones, and consider the $M \times K$ bipartite graph $\Lc = \left(\Ac,\Kc,\Ec\right)$ with adjacency matrix $\Am$ and weights $w_{m,k} = \lambda_m^{(k)}$ on the edges $(m,k) \in \Ec$. An example of the bipartite graph $\Lc$ and its corresponding weighted adjacency matrix $\Wm$ is illustrated in Figs.\[fig:bi\_graph\] and \[fig:adj\_mat\].
[.24]{} ![(a) An example of a bipartite graph $\Lc$. (b) The corresponding weighted adjacency matrix $\Wm$.[]{data-label="fig:test"}](bi_graph.pdf "fig:"){width="1\linewidth"}
[.24]{} ![(a) An example of a bipartite graph $\Lc$. (b) The corresponding weighted adjacency matrix $\Wm$.[]{data-label="fig:test"}](Table.pdf "fig:"){width="0.6\linewidth"}
Given a pilot dimension ${T_{\rm dl}}$, our goal consists in selecting a subgraph $\Lc'=\left(\Ac',\Kc',\Ec'\right)$ of $\Lc$ in which each node on either side of the graph has a degree at least 1 and such that
1. For all $k \in \Kc'$ we have $\text{deg}_{\Lc'}(k) \le {T_{\rm dl}}$, where $\text{deg}_{\Lc'}$ denotes the degree of a node in the selected subgraph.
2. The sum of weights of the edges adjacent to any node $k \in \Kc'$ in the subgraph $\Lc'$ is greater than a threshold, i.e. $\sum_{m\in \Nc_{\Lc'} (k)} w_{m,k} \ge P_0, ~\forall k\in \Kc'$, where $\Nc_{\Lc'}(k)$ denotes the set of neighbors in $\Lc'$ of node $k$.
3. The channel matrix $\Hm_{\Ac',\Kc'}$ obtained from $\Hm$ by selecting $a \in \Ac'$ (referred to as “selected beam directions") and $k\in \Kc'$ (referred to as “selected users") has large rank.
The first criterion enables the stable estimation of the effective channel of any selected user with only ${T_{\rm dl}}$ common pilot dimensions and ${T_{\rm dl}}$ complex symbols of feedback per selected user. The second criterion makes sure that the effective channel strength of any selected user is greater than a certain desired threshold. The third criterion is motivated by the fact that the DL multiplexing gain is given by ${\rm rank} ( \Hm_{\Ac',\Kc'} ) \times \max \{ 0 , 1 - {T_{\rm dl}}/T\}$, where $T$ denotes the whole number of REs, and it is obtained by serving a number of users equal to the rank of the effective channel matrix.
In [@khalilsarai2018fdd] we show that the rank of $\Hm$ is given, with probability 1, by the size of the largest intersection submatrix whose associated bipartite graph contains a perfect matching.[^3] With this observation, we can formulate the problem mentioned above as follows. Let $\Mc (\Ac',\Kc')$ denote a matching of the subgraph $\Lc' (\Ac',\Kc',\Ec')$ of the bipartite graph $\Lc (\Ac,\Kc,\Ec)$. Find the solution of the following optimization problem:
\[eq:my\_opt\_1\] $$\begin{aligned}
& \underset{\Ac'\subseteq \Ac, \Kc' \subseteq \Kc}{\text{maximize}} && \left\vert \Mc\left(\Ac',\Kc'\right)\right\vert \label{eq:my_opt_1-one} \\
& \text{subject to} && \text{deg}_{\Lc'} (k) \le {T_{\rm dl}}~ \forall k\in \Kc', \label{eq:my_opt_1-two} \\
& ~ && \hspace{-4mm} \sum_{a\in \Nc_{\Lc'} (k) } w_{a,k} \ge P_0, ~ \forall k\in \Kc'. \label{eq:my_opt_1-three}
\end{aligned}$$
It turns out that, this problem can be transformed into the following equivalent mixed integer linear program (MILP), denoted by $\Pc_{\text{MILP}}$ [@khalilsarai2018fdd]:
\[opt:P\_MILP\] [ $$\begin{aligned}
\underset{x_m,y_k,z_{m,k} }{\text{maximize}} & ~~\sum_{m\in \Ac} \sum_{k \in \Kc} z_{m,k} \label{eq:obj_1} + \epsilon \sum_{m \in \Ac} x_m \\
\text{subject to} & ~~~~z_{m,k} \le [\Am]_{m,k} ~~\forall m\in\Ac,k\in \Kc, \label{eq:one}\\
~~ & ~~~~ \sum_{k\in \Kc}z_{m,k} \le x_m ~~ \forall m\in \Ac, \label{eq:two}\\
~~ & ~~~~ \sum_{m\in \Ac}z_{m,k} \le y_k ~~ \forall k\in \Kc, \label{eq:three} \\
~~ & ~~~~ \sum_{m\in \Ac} [\Am]_{m,k} x_m \le {T_{\rm dl}}y_k + M (1-y_k) ~~\forall k\in \Kc, \label{eq:four}\\
~~ & ~~~~ P_0 \, y_k \le \sum_{m \in \Ac} [\Wm]_{m,k} x_m ~~\forall k \in \Kc, \label{eq:five} \\
~~ & ~~~~ x_m \le \sum_{k\in \Kc} [\Am]_{m,k} y_k ~~ \forall m\in \Ac, \label{eq:six} \\
~~ & ~~~~ x_m, y_k \in \{0,1\} ~~\forall a\in \Ac,k\in \Kc,\\
~~ & ~~~~ z_{m,k} \in [0,1] ~~\forall m\in \Ac,k\in \Kc,
\end{aligned}$$ ]{}
The binary variable $x_m$ in determines whether beam $m\in \Ac$ is selected or not, i.e., $x_m=1$ if and only if the $m$-th beam is selected. Similarly, $y_k$ controls the selection of user $k$ for channel estimation (and eventually serving). For a given set of user DL covariance matrices, we denote by $\Bc=\{m:x_m^\ast=1\} = \{m_1,m_2,\ldots, m_{M'}\}$ the set of selected beams directions of cardinality $|\Bc| = M'$ and by $\Kc = \{k: y^\ast_k = 1\}$ the set of selected users of cardinality $|\Kc| = K'$, where $\left\{x_m^\ast \right\}_{m=1}^M$ and $\left\{y_k^\ast \right\}_{k=1}^K$ are solutions to $\Pc_{\text{MILP}}$.
The desired sparsifying precoding matrix $\Bm$ is finally obtained as $\Bm = \Fm_{\Bc}^\herm$, where $\Fm_{\Bc} = [\fv_{m_1}, \ldots, \fv_{m_{M'}}]$ and $\fv_{m}$ denotes the $m$-th column of the $M \times M$ unitary DFT matrix $\Fm$. Using (\[eq:approximate-KL\]), the effective DL channel vectors take on the form $$\label{effch}
{{\check{\bfh}}_{\rm eff}}^{(k)} = \Bm \sum_{m \in \Sc_k} g_m^{(k)} \sqrt{\lambda_m^{(k)}} \fv_m = \sum_{i : m_i \in \Bc \cap \Sc_k} \sqrt{\lambda_{m_i}^{(k)}} g_{m_i}^{(k)} \uv_i,$$ where $\Sc_k$ is the the support of $\hv^{(k)}$ in the DFT domain and where $\uv_i$ denotes a $M' \times 1$ vector with all zero components but a single “1” in the $i$-th position. With this construction, the number of non-identically zero coefficients for each user $k$ are $|\Bc \cap \Sc_k| \leq {T_{\rm dl}}$ and their positions (encoded in the vectors $\uv_i$ in (\[effch\])) are known to the BS. Hence, the effective channel vectors can be estimated from the ${T_{\rm dl}}$-dimensional DL pilot observation with an estimation MSE that vanishes as $1/\SNR$ [@khalilsarai2018fdd] by increasing the Signal-to-Noise Ration (SNR). With the above precoding, we have $\Bm \Bm^\herm = \Id_{M'}$. Furthermore, we can choose the DL pilot matrix $\Psim$ to be proportional to a random unitary matrix of dimension ${T_{\rm dl}}\times M'$, such that $\Psim \Psim^\herm = P_{\rm dl} \Id_{{T_{\rm dl}}}$. After DL pilot transmission, all users $k\in [K]$ send back their noisy observations $\yv^{(k)}$ to the BS via analog unquantized feedback. The BS, having an estimate of the DL channel covariance, performs linear MMSE estimation to obtain an estimate of the DL channel.
We concclude by describing the channel precoding step. Let $\widehat{\bfH}_{{\text{eff}}}= [ \widehat{{\bfh}}_{{\text{eff}}}^{(1)}, \ldots, \widehat{{\bfh}}_{{\text{eff}}}^{(K')}]$ be the matrix of the estimated effective DL channels for the selected users. We consider the ZF beamforming matrix $\Vm$ given by the column-normalized version of the Moore-Penrose pseudoinverse of the estimated channel matrix, i.e., $\Vm =
\left(\widehat{\bfH}_{{\text{eff}}} \right)^\dagger \Gm^{1/2}$, where $\left(\widehat{\bfH}_{{\text{eff}}} \right)^\dagger =
\widehat{\bfH}_{{\text{eff}}} \left ( \widehat{\bfH}_{{\text{eff}}}^\herm \widehat{\bfH}_{{\text{eff}}} \right )^{-1}$ and $\Gm$ is a diagonal matrix that makes the columns of $\Vm$ to have unit norm. A channel use of the DL precoded data transmission phase at the $k$-th user receiver takes on the form $y^{(k)} = \left ( \hv^{(k)} \right )^\herm \Bm^\herm \Vm \Pm^{1/2} \bfd + n^{(k)}$, where $\bfd \in \bC^{K' \times 1}$ is a vector of unit-energy user data symbols and $\Pm$ is a diagonal matrix defining the power allocation to the DL data streams. The transmit power constraint is given by $\trace( \Bm^\herm \Vm \Pm \Vm^\herm \Bm ) = \trace ( \Vm^\herm \Vm \Pm ) = \trace (\Pm) = P_{\rm dl}$, where we used $\Bm \Bm^\herm = \Id_{M'}$ and the fact that $\Vm^\herm \Vm$ has unit diagonal elements by construction. For simulation results, we use the simple uniform power allocation $P_k = P_{\rm dl}/K'$ to each $k$-th user data stream. The received symbol at user $k$ receiver is given by $$\begin{aligned}
y^{(k)} = b_{k,k} d_k + \sum_{k' \neq k} b_{k,k'} d_{k'} + n^{(k)},\end{aligned}$$ where the coefficients $(b_{k,1}, \ldots, b_{k,K'})$ are given by the elements of the $1 \times K'$ row vector $\left ( \hv^{(k)} \right )^\herm \Bm^\herm \Vm \Pm^{1/2}$. Of course, in the presence of an accurate channel estimation we expect that $b_{k,k} \approx \sqrt{G_k P_k}$ and $b_{k,k'} \approx 0$ for $k' \neq k$. To calculate achievable sum-rate, we assume that all coefficients $(b_{k,1}, \ldots, b_{k,K'})$ are known to the corresponding receiver $k$. Including the DL training overhead, this yields the rate expression [ $$\label{eq:rate_ub}
R_{\rm sum} = \left (1 - \frac{{T_{\rm dl}}}{T} \right ) \sum_{k \in \Kc} \bE \left [ \log \left ( 1 + \frac{\left|b_{k,k}\right|^2}{1 + \sum_{k' \neq k} \left|b_{k,k'} \right|^2} \right ) \right ].$$]{}
Simulation Results
==================
In this section, we perform numerical simulations to assess the performance of our proposed method. For simulations, we consider IMT-FDD band as in LTE standard [@LTEbands] with a UL band $[1920,1980]$MHz and a DL band $[2110,2170]$MHz. Thus, we use $f_\text{ul}=1950$ and $f_\text{dl}=2140$, and set the UDCT parameter $\beta$ in to $\beta=\frac{{f_{\rm dl}}}{{f_{\rm ul}}}\approx 1.1>1$. We also set the number of BS antennas to $M=256$.
ASF Estimation and UL-DL Covariance Transformation
--------------------------------------------------
In this section, we perform numerical simulations to compare a DNN trained with group-sparse ASFs with NNLS method in [@haghighatshoar2018multi] and $\ell_2$ projection method in [@miretti2018fdd].
[**ASF Estimation.**]{} We assume that each ASF in the set of feasible group-sparse ASFs consists of two groups $g=2$ where the ASF corresponding to each group is a uniform distribution with a center randomly selected in the range of AoAs $[-1,1]$ and with a random width of size at most $w_{\text{max}}=0.4$. We consider a SNR of 20dB for noisy UL channel vectors used for training and set the sampling ratio to $\frac{N}{M}=2$ (see, e.g., ). We refer to Section \[sec:dnn\] for a more detailed discussion of the data set used for training the DNN. We build a training data set of size $S=10000$ where we use 80% of this data set for training and the remaining 20% for validation.
Fig.\[fig:ASF\] illustrates the simulation results for ASF estimation. It is seen that DNN performs much better than NNLS and $\ell_2$ projection. NNLS estimates the support very well but is quite spiky over the support. In contrast, $\ell_2$ projection method is quite smooth over the support but produces considerable off-support elements.
![Comparison of performance of ASF estimation for DNN, NNLS, and $\ell_2$ Projection method.[]{data-label="fig:ASF"}](ASF_font.pdf){width="47.50000%"}
[**UL-DL Covariance Transformation.**]{} In this part, we compare the UDCT performance of our proposed DNN with that of NNLS and $\ell_2$ projection method for different sampling ratios $\frac{N}{M}\in \{1, \dots, 8\}$. As in the previous part, we assume a SNR of 20dB for the noisy UL channel vectors. Recall that an UDCT algorithm takes $N$ noisy UL channel vectors as the input and produces an estimate of the DL covariance matrix as the output. We use the following distortion metrics:
1. *Normalized Frobenius-norm*: defined as $$E_{\text{NFD}} = \bE \left\{ \frac{\Vert \Sigdl - {\Sigdl}^\star \Vert_\sfF}{\Vert \Sigdl\Vert_\sfF}\right\},$$ where $\|.\|_\sfF$ denotes the matrix Frobenius norm, where $\Sigdl$ and ${\Sigdl}^\star$ denote the true and the estimated DL covariance matrices, and where the expectation is taken over random realizations of the UL channel vectors.
2. *Power-loss*: Suppose that the eigenvalue decomposition of the true and estimated DL covariance matrices are given as $\Sigdl= \Um \Dm \Um^\herm$ and $\Sigdl^\star= \widetilde{\Um} \widetilde{\Dm} \widetilde{\Um}^\herm$, respectively. We define the efficiency metric of order $q\in [M]$ as $$\label{eq:efficiency_def}
\eta_q=\frac{\tr \left( \widetilde{\Um}_{q}^\herm \Sigdl\widetilde{\Um}_{q}\right)}{\tr \left(\Um_{q}^\herm\Sigdl\Um_{q}\right)}\in [0,1],$$ where $\tr(.)$ denotes the trace operator, and where $\bfU_q$ and $\widetilde{\bfU}_q$ denotes $M \times q$ matrices consisting of the first $q$ columns of $\bfU$ and $\widetilde{\bfU}$ respectively. This metric indicates which fraction of the power lying in the dominant $q$-dim subspace of the true DL covariance $\Sigdl$ is captured by the estimated subspace with the same dimension in $\Sigdl^\star$ spanned by $\widetilde{\Um}_{q}$ [@haghighatshoar2016massive; @haghighatshoar2018low]. The closer the efficiency parameter $\eta_q$ is to one, the better the power in the dominant $q$-dim subspace of $\Sigdl$ is captured by the estimated covariance $\Sigdl^\star$. As a worst-case distortion metric that works independent of the dimension of the subspace, we consider $$\begin{aligned}
E_{\text{PLE}} = 1- \bE \{ \min_{q \in [M]} \eta_q \}\end{aligned}$$ where the expected value is taken with respect to the random realization of the noisy UL channel vectors.
Figs.\[fig:UDCT\_fro\] and \[fig:UDCT\_plos\] illustrate the simulation results. It is again seen that, as in the ASF estimation, our proposed method yields better performance in terms of UDCT under both distortion metrics, especially in the practically-relevant regime where the sampling ratio $\frac{N}{M}$ may be very small.
![Comparison of the performance of UDCT for DNN, NNLS, and $\ell_2$ Projection method under Normalized Frobenius-norm distortion.[]{data-label="fig:UDCT_fro"}](fro_err_2.pdf)
![Comparison of the performance of UDCT for DNN, NNLS, and $\ell_2$ Projection method under Power-loss distortion.[]{data-label="fig:UDCT_plos"}](plos_err_2.pdf)
Sparsifying Precoder vs. Statistical Beamforming
------------------------------------------------
In this paper, we designed the sparsification precoding from the estimated DL covariance matrices ${\Sigmam_k}$, $k\in [K]$ (for simplicity, we dropped the label ‘dl’). In the presence of channel sparsification imposed by this precoder, BS still needs to probe the reduced-dim channel in the DL and request the users to feedback their channel measurements during the probing to the BS, from which the BS is able to extract the instantaneous channel state of the users after sparsification.
As an alternative to sparsification precoding, one may use the statistical beamforming, which uses only the DL covariance knowledge to beamform to the users, and especially does not apply any DL probing and UL feedback to extract the instantaneous channel of the users [@wajid2009robust; @zhang2015statistical]. Consequently, the statistical beamforming does not suffer from the fractional rate loss factor $1-\frac{{T_{\rm dl}}}{T}$ in caused because of devoting ${T_{\rm dl}}$ out of $T$ REs to DL channel probing.
We compare the performance of the sparsification precoder with the statistical beamforming method proposed in [@zhang2015statistical]. Let $\{\Sigmam_k\}_{k=1}^K$ denote the DL covariance matrices of the users, either exact or estimated via a UDCT method. In particular, in this section, we provide results for both the case in which the BS has access to exact DL covariances and the case where the DL covariances are estimated using DNN. Given $\{\Sigmam_k\}_{k=1}^K$, the beamforming vector for the user $k$ is given by
$$\uv_k = \uv_{\max} \left \{ \big(N_0\mathbf{I}_M + \sum_{\ell \neq k} \Sigmam_\ell \big)^{-1} \Sigmam_k \right \},$$
where $\uv_{\max}(.)$ denotes the normalized eigenvector corresponding to the maximum eigenvalue of a matrix, and where $N_0$ denotes the normalized noise power. The statistical beamforming matrix is then given by $\Vm= \left[ \uv_1,\ldots,\uv_K \right]$. We also assume uniform power allocation across the users, just as explained in Section \[sec:ch\_sparsification\], and calculate the sum-rate according to .
![Sum-rate comparison of the proposed sparsification precoder with statistical beamforming for $\SNR =20$ dB. We have $M=256$ BS antennas and $K=20$ served users. []{data-label="fig:rate_vs_Tdl"}](SBF_vs_ACS.pdf)
We set the simulation parameters as follows: the BS has $M=256$ antennas and serves $K=20$ users, each having a randomly generated ASF as before. Fig.\[fig:rate\_vs\_Tdl\] illustrates the sum-rate vs. pilot dimension results. The blue curves represent the results by assuming true DL covariances, and the red curves represent the results for estimated DL covariances via DNN. As we can see, the performance of the statistical beamforming (SBF) method is substantially inferior to the performance of the sparsification precoding for almost all range of pilot dimensions. This implies that, it is always worthwhile to spend some of resource blocks to estimate instantaneous channels and design the beamformer upon them. This is true, even for small values of the pilot dimension, which implies that even a coarse estimation of the instantaneous channel using a few pilots yields a better performance than the that of statistical beamforming.
Conclusion
==========
In this paper, we presented a new approach to achieve a very competitive tradeoff between spatial multiplexing gain and probing-feedback overhead in FDD massive MIMO systems. Our approach has two main ingradients: (i)an efficient regularization technique based on Deep Neural Networks (DNN) for structured ASF estimation and UL-DL covariance transformation, (ii)a novel “*sparsifying precoding*” technique that uses the estimated DL covariance matrices from (i) and shapes the channel sparsity depending on the pilot dimension such that the “*effective*” channel vectors after sparsification can be estimated at the BS with a low mean-square error. In particular, the proposed sparsifying precoder is not at the mercy of the channel sparsity induced by the propagation environment, which is a bottleneck for Compressed Sensing based channel estimation methods. We compared our proposed DNN-based method in (i) with other methods in the literature via numerical simulations and showed that it yields a very competitive performance. We also compared our sparsifying precoder in (ii) with the state-of-the-art statistical beamforming methods under the assumption that those methods also have access to the covariance knowledge in the DL and showed that our method yields higher spectral efficiency since it uses in addition the instantaneous channel information after sparsification.
[^1]: The authors are with the Communications and Information Theory Group (CommIT), Technische Universität Berlin ({m.barzegarkhalilsarai, yi.song, tianyu.yang, saeid.haghighatshoar, caire}@tu-berlin.de).
[^2]: Recall that circulant matrices are diagonalizable with the orthogonal DFT matrix.
[^3]: A matching is a set of edges of a graph without common vertices.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Fundamental primitives such as bit commitment and oblivious transfer serve as building blocks for many other two-party protocols. Hence, the secure implementation of such primitives are important in modern cryptography. In this work, we present a bit commitment protocol which is secure as long as the attacker’s quantum memory device is imperfect. The latter assumption is known as the noisy-storage model. We experimentally executed this protocol by performing measurements on polarization-entangled photon pairs. Our work includes a full security analysis, accounting for all experimental error rates and finite size effects. This demonstrates the feasibility of two-party protocols in this model using real-world quantum devices. Finally, we provide a general analysis of our bit commitment protocol for a range of experimental parameters.'
author:
- Nelly Huei Ying Ng
- 'Siddarth K. Joshi'
- Chia Chen Ming
- Christian Kurtsiefer
- Stephanie Wehner
title: 'Experimental implementation of bit commitment in the noisy-storage model'
---
Introduction
============
Traditionally, the main objective of cryptography has been to protect communication from the prying eyes of an eavesdropper. Yet, with the advent of modern communications new cryptographic challenges arose: we would like to enable two parties, Alice and Bob, to solve joint problems even if they do not trust each other. Examples of such tasks include secure auctions or the problem of secure identification such as that of a customer to an ATM. Whereas protocols for general two-party cryptographic problems may be very involved, it is known that they can in principle be built from basic cryptographic building blocks known as oblivious transfer [@kilian] and bit commitment.
The task of bit commitment is thereby particularly simple and has received considerable attention in quantum information. Intuitively, a bit commitment protocol consists of two phases. In the *commit phase*, Alice provides Bob with some form of evidence that she has chosen a particular bit $C \in \01$. Later on in the *open phase*, Alice reveals $C$ to Bob. A bit commitment protocol is secure, if Bob cannot gain any information about $C$ before the open phase, and yet, Alice cannot convince Bob to accept an opening of any bit $\hat{C} \neq C$.
Unfortunately, it has been shown that even using quantum communication none of these tasks can be implemented securely . Note that in quantum key distribution (QKD), Alice and Bob *trust* each other and want to defend themselves against an outsider Eve. This allows Alice and Bob to perform checks on what Eve may have done, ruling out many forms of attacks. This is in sharp contrast to two-party cryptography where there is no Eve and Alice and Bob *do not trust* each other. Intuitively, it is this lack of trust that makes the problem considerably harder. Nevertheless, because two-party protocols form a central part of modern cryptography, one is willing to make *assumptions* on how powerful an attacker can be in order to implement them securely.
Here, we consider *physical* assumptions that enable us to solve such tasks. In particular, can the sole assumption of a limited storage device lead to security? [@Maurer92b] This is indeed the case and it was shown that security can be obtained if the attacker’s *classical* storage is limited [@Maurer92b; @cachin:bounded]. Yet, apart from the fact that classical storage is cheap and plentiful, assuming a limited classical storage has one rather crucial caveat: If the honest players need to store $N$ classical bits to execute the protocol in the first place, *any* classical protocol can be broken if the attacker can store more than roughly $N^2$ bits [@maurer:imposs].
Motivated by this unsatisfactory gap, it was thus suggested to assume that the attacker’s *quantum* storage was bounded [@Bennett84; @serge:new; @serge:bounded; @chris:id1; @chris:id2], or more generally, noisy [@Noisy1; @noisy:robust; @noisy:new]. The central assumption of the noisy-storage model is that during waiting times $\Delta t$ introduced in the protocol, the attacker can only keep quantum information in his quantum storage device ${\mathcal{F}}$. The exact amount of noise can depend on the waiting time. Otherwise, the attacker may be all-powerful. In particular, he can store an unlimited amount of classical information, and perform any computation instantaneously without errors. Note that the latter implies that the attacker could encode his quantum information into an arbitrarily complicated error correcting code, to protect it from noise in his storage device ${\mathcal{F}}$.
The assumption that storing a large amount of quantum information is difficult is indeed realistic today, as constructing large scale quantum memories that can store arbitrary information successfully in the first attempt has proved rather challenging. We emphasize that this model is not in contrast with our ability to build quantum repeaters, where it is sufficient for the latter to store quantum states while making many attempts. A review on quantum memories can be found in [@qmemory], and numerous recent work can also be found in [@Usmani2010; @bonarota2011; @PhysRevLett.108.210501]. While noting that perpetual advances in building quantum memories fundamentally affect the feasibility of all protocols in the noisy storage model, yet we will explain below that given any upper bound on the size and noisiness of a future quantum storage device, security is in fact possible - we merely need to send more qubits during the protocol.
In this work, we have implemented a bit commitment protocol that is secure under the noisy storage assumption. We provide a general security analysis of our protocol for a range of possible experimental parameters. The parameters of our particular experiment are shown to lie within the secure region. The storage assumption in our work is such that a cheating party cannot store more than approximately 900 qubits, which is a reasonable physical constraint given modern day technology of storing quantum information.
Result
======
The Noisy Storage Model {#the-noisy-storage-model .unnumbered}
-----------------------
To state our result, let us first explain what we mean by a quantum storage device, and how does an assumption regarding these devices translate to security conditions in the noisy storage model. A more detailed introduction to the model can be found in e.g. [@noisy:new].
Of particular interest to us are storage devices consisting of $S$ ”memory cells”, each of which may experience some noise ${\mathcal{N}}$ itself. Mathematically, this means that the storage device is a quantum channel \[mathematically, a completely positive trace preserving map (CPTPM)\] of the form ${\mathcal{F}}= {\mathcal{N}}^{\otimes S}$ where ${\mathcal{N}}: {\mathcal{B}}({\mathbb{C}}^d) \rightarrow {\mathcal{B}}({\mathbb{C}}^d)$ is a noisy channel acting on each memory cell mapping input states to some noisy output states. For example, a noise-free storage device consisting of $S$ qubits (i.e.,$d=2$) corresponding to the special case of bounded storage [@serge:bounded] is given by ${\mathcal{F}}= {\mathsf{id}}_2^{\otimes S}$ where ${\mathsf{id}}_2$ is the identity channel with one qubit input and one qubit output. Another example is a memory consisting of $S$ qubits, each of which experiences depolarizing noise according to the channel ${\mathcal{N}}_r(\rho) = r \rho + (1-r) \frac{{\mathsf{id}}}{2}$. The larger $r$ is, the less noise is present. Yet another example is the erasure channel, which models losses in the storage device.
It is indeed intuitive that security should be related to ”how much” information the attacker can squeeze through his storage device. That is, one expects a relation between security and the capacity of ${\mathcal{F}}$ to carry quantum information. Indeed, it was shown that security can be linked to the classical capacity [@noisy:new], the entanglement cost [@entCost], and finally the quantum capacity [@qcextract] of the adversary’s storage device ${\mathcal{F}}$.
When evaluating security, we start with a basic assumption on the maximum size and the minimum amount of noise in an adversary’s storage device. Such an assumption can for example be derived by a cautious estimate based on quantum memories that are available today. Note that these assumptions are for memories that can store arbitrary states on first attempt. Such memories presently exist for a handful of qubits. Given such an estimate, we then determine the number of qubits we need to transmit during the protocol to effectively overflow the adversary’s memory device and achieve security.
Protocol and its security {#protocol-and-its-security .unnumbered}
-------------------------
We consider the bit commitment protocol from [@noisy:new] with several modifications to make it suitable for an experimental implementation with time-correlated photon pairs. Figure \[fig:flowchart\] provides a simplified version of this modified protocol without explicit parameters - the explicit version can be found in the [Supplementary Methods]{}. In the [Supplementary Methods]{}, we also provide a general analysis that can be used for any experimental setup (details on our particular experiment are also provided in the same section).
To understand the security constraints, we first need to establish some basic terminology. In our experiment, Alice holds the source, and both Alice and Bob have four detectors, each one corresponding to one of the four BB84 states [@Bennett84]. If Alice or Bob observes a click of exactly one of their detectors (*symmetrized* with the procedure outlined in [Supplementary Methods]{}), we refer to it as a *valid click*. Cases where more than one detector clicks at the same instant on the same side are ignored. A *round* is defined by a valid click of *Alice’s* detectors. A *valid round* is where both parties Alice and Bob registered a valid click in a corresponding time window, i.e., where a photon pair has been identified.
First, to deal with losses in the channel we introduce a new step in which Bob reports a loss if he did not observe a valid click. Second, to deal with bit flip errors on the channel, we employ a different class of error-correcting codes, namely a random code. Usage of random codes is sufficient for this protocol since decoding is not required for honest parties. The main challenge is then to link the properties of random codes to the protocol security.
Before we can argue about the correctness and security of the proposed protocol, let us introduce four crucial figures of interest that need to be determined in any experimental setup. The first two are the probabilities $p^0_{\rm sent}$ and $p^1_{\rm sent}$, that none or just a single photon was sent to Bob respectively, conditioned on the event that Alice observed a round. The third is the probability $p_{\rm B, no click}^{\rm h}$ that honest Bob registers a round as missing, i.e. Bob does not observe a valid click when Alice does. Again, this probability is conditioned on the event that Alice observed a round. Note that by no-signalling, Alice’s choice of better (or worse) detectors should not influence the probability of Bob observing a round. Finally, we will need the probability $p_{\rm err}$ of a bit flip error, i.e. the probability that Bob outputs the wrong bit even though he measured in the correct basis.
Naturally, since Alice and Bob do not trust each other, they cannot rely on each other to perform said estimation process. Note, however, that the scenario of interest in two-party cryptography is that the honest parties essentially purchase off the shelf devices with standard properties, for which either of them could perform said estimate. It is only the dishonest parties who may be using alternate equipment. Another way to look at this is to say that there exists some set of parameters (i.e., maximum losses, maxmium amount of noise on the channel, etc) such that an honest party has to conform to these requirements when executing the protocol.
Let us now sketch why the proposed protocol remains correct and secure even in the presence of experimental errors. A detailed analysis is provided in the [Supplementary Methods]{}. In our analysis, we take the storage device ${\mathcal{F}}$, as well as a fixed overall security error ${\epsilon}$ as given. Let $M$ be the number of rounds *Alice* registers during the execution of the protocol. Let $n$ be the number of valid rounds. In the description of theoretical parameters found in the [Supplementary Methods]{}, it is shown that $M$ and $n$ are directly related to each other, given some fixed experimental parameters. In particular, $n$ is a function of $M$ and $p_{\rm B, no click}^{\rm h}$ $$n \approx (1 - p_{\rm B, no click}^{\rm h}) M\ .$$ We can now ask, how large does $M$ (or equivalently $n$) need to be in order to achieve security. If $n$ is very small, for example if $n\approx 100$, it is relatively easy to break the protocol since a cheating party might be able to store enough qubits. Also many terms from our finite $n$ analysis reach convergence only for sufficiently large $n$. As these terms depend on experimental parameters, security can be achieved for a larger range of experimental parameters if $n$ is large. By fixing the assumption on quantum storage size, experiment parameters and security error values, our analysis allows us to determine a value of $n$ where security is achievable.
![\[fig:flowchart\] ](figures/flowSidv3){width="\columnwidth"}
[**Correctness:**]{} First of all, we must show that if Alice and Bob are both honest, then Bob will accept Alice’s honest opening of the bit $C$. Note that the only way that honest Bob will reject Alice’s opening is when too many errors occur on the channel, and hence part 2 of Bob’s final check (see Figure \[fig:flowchart\]) will fail. A standard Chernoff style bound using the Hoeffding inequality [@hoeffding] shows the probability of this event is small, i.e., that the deviation from the expected number of $p_{\rm err} n$ errors is not too large.
[**Security against Alice:**]{} Second, we must show that if Bob is honest, then Alice cannot get him to accept an opening of a bit $\hat{C} \neq C$. In our protocol, Alice is allowed to be all powerful, and is not restricted by any storage assumptions. If she is dishonest, we furthermore assume that she can even have perfect devices and can eliminate all errors and losses on the channel. The first part of our analysis, i.e., the analysis of the steps before the syndrome is sent is thereby identical to [@Curty10] (see Figure \[fig:flowchart\]). More precisely, it is shown that up to this step in the protocol, a string $X^n \in \01^n$ is generated such that Bob knows the bits $X_{\mathcal{I}}$ for a randomly chosen subset $\mathcal{I} \subseteq \{1,\ldots,n\}$, where $X_{\mathcal{I}}$ corresponds to the entries of the string $X^n$ indexed by the positions in $\mathcal{I}$. If Alice is dishonest, we want to be sure at this stage that she cannot learn $\mathcal{I}$, that is, she cannot learn which bits of $X^n$ are known to Bob. In the original protocol without experimental imperfections [@noisy:new] this was trivially guaranteed because Bob never sent any information to Alice. In this practical protocol, however, Bob does send some information to Bob, namely which rounds are valid for him, i.e., when he saw a click. In [@Curty10] it was simply assumed that the probability of Bob observing a loss is the same for all detectors, and hence in particular also independent of Bob’s basis choice. This is generally never the case in practise. However, by symmetrizing the losses as outlined in the [Supplementary Methods]{}, one can ensure that the losses become the same for all detectors. In essence, this procedure probabilistically adds additional losses to the better detectors such that in the end all detectors are as lossy as the worst one. As Bob’s losses are then independent of his basis choice, i.e., the detector, this is means that Alice cannot gain any information about $\mathcal{I}$ when Bob reports some rounds as being lost.
The second part of the protocol and its analysis uses the string $X^n$ and Bob’s partial knowledge $X_{\mathcal{I}}$ to bind Alice to her commitment. First, we have that properties of the error-correcting code ensure that if the syndrome of the string (${\mathrm{Syn}}(X^n)$ in Figure \[fig:flowchart\]) matches and Alice passes the first test, then she must flip many bits in the string to change her mind. In the original protocol of [@noisy:new] sending Bob the syndrome of $X^n$ ensured that she must change at least $\frac{d}{2}$ bits of $X^n$ where $d$ is the distance of the error-correcting code, such that Bob will accept the syndrome to be consistent. However, since Alice does not know which bits $X_{\mathcal{I}}$ are known to Bob she will get caught with high probability. This due to the fact that with probability $1-(1/2)^{d/2}$ Alice changed at least a bit known to Bob, and in the perfect case Bob aborts whenever a single bit is wrong. As we have to deal with experimental imperfections we cannot have that Bob aborts whenever a single bit is wrong, as bit flip errors on the channel likely lead errors even when Alice is honest. As such the difference to the analysis of [@noisy:new] is that Bob must accept some incorrect bits in part two of his final check (see Figure \[fig:flowchart\]). Our argument is nevertheless quite similar, but does require a careful tradeoff involving all experimental parameters between the distance of the code and the syndrome length (see below). We hence use a different error-correcting code as compared to [@noisy:new]. In particular, we use a random code which has the property that with overwhelming probability its distance is large (i.e. it is hard for Alice to cheat), while nevertheless having a reasonably small syndrome length (see [Supplementary Discussion]{}). The latter will be important in the security analysis below when Alice herself is honest.
[**Security against Bob:**]{} Finally, we must show that if Alice is honest, then Bob cannot learn any information about her bit $C$ before the open phase. Again, dishonest Bob may have perfect devices and eliminate all errors and losses on the channel. His only restriction is that during the waiting time $\Delta t$ he can store quantum information only in the device ${\mathcal{F}}$.
We first show that Bob’s information about the entire string $X^n$ is limited. We know from [@noisy:new] that Bob’s min-entropy about the string $X^n$ before Alice sends the syndrome, given all his information including his quantum memory can be bounded by $$\begin{aligned}
\label{minEntropy}
{{\rm H_{\rm min}}}(X^n|\mathrm{Bob}) \gtrsim - \log P_{\rm succ}^{\mathcal{F}}(Rn)\ ,\end{aligned}$$ where $P_{\rm succ}^{\mathcal{F}}(Rn)$ is the maximum probability of transmitting $Rn$ randomly chosen bits through the channel ${\mathcal{F}}$ where $R$ is called the rate. This rate is determined using a novel uncertainty relation that we prove for BB84 measurements, and all experimental parameters. The min-entropy itself can thereby be expressed as ${{\rm H_{\rm min}}}(X^n|\mathrm{Bob}) = - \log P_{\rm guess}(X^n|\mathrm{Bob})$, where $P_{\rm guess}(X^n|\mathrm{Bob})$ is the probability that Bob guesses the string $X^n$, maximized over all measurements that he can perform on his system [@krs:entropy].
As Alice sends the syndrome to Bob, Bob gains some additional information which reduces his min-entropy. More precisely, it could shrink at most by the length of the syndrome, i.e., $$\begin{aligned}
{{\rm H_{\rm min}}}(X^n|\mathrm{Bob},{\mathrm{Syn}}(X^n)) \geq {{\rm H_{\rm min}}}(X^n|\mathrm{Bob}) - \log |{\mathrm{Syn}}(X^n)|\ .\end{aligned}$$ Note that this is the reason why we asked for the error-correcting code to have a short syndrome length above.
Finally, we show that knowing little about all of $X^n$ implies that Bob cannot learn anything about $C$ itself. More precisely, when Alice chooses a random two universal hash function ${\mbox{Ext}}(\cdot, R)$ and performs privacy amplification [@renato:diss], Bob knows essentially nothing about the output ${\mbox{Ext}}(X^n,R)=D$ whenever his min-entropy about $X^n$ is sufficiently large. The bit $D$ then acts as a key to encrypt the bit $C$ using a one-time pad. Since Bob cannot know $D$, he also cannot know $C$. Our analysis is thereby very similar to [@noisy:new], requiring only a very careful balance between the distance of the error-correcting code above, and the syndrome length.
We provide a detailed analysis in the [Supplementary Methods]{}, where a general statement for arbitrary storage devices is included. Especially for the case of bounded storage ${\mathcal{F}}= {\mathsf{id}}_2^{\otimes S}$, we can easily evaluate how large $M$ needs to be in order to achieve security against both Alice and Bob, when an error parameter ${\epsilon}$ is fixed. The total execution error of the protocol is obtained by adding up all sources of errors throughout the protocol analysis.
The case where Alice and Bob are both dishonest is not of interest, because the aim of this protocol is to perform correctly while both players are honest, and protect the honest players from dishonest players.
Experiment {#experiment .unnumbered}
----------
We have implemented a quantum protocol for bit commitment that is secure in the noisy-storage model. For this, $n = 250\,000$ valid rounds (see below) were used at a bit error rate of $p_{\rm err} = 4.1\%$ (after symmetrization) to commit one bit with a security error of less than ${\epsilon}= 2 \times 10^{-5}$. Note that $\epsilon$ is the final correctness and security error for the execution of bit commitment in our experiment. This protocol is secure under the assumption that Bob’s storage size is no larger than 972 qubits, where each qubit undergoes a low depolarizing noise with a noise parameter $r=0.9$ (see [Supplementary Methods]{} Section D). We stress that our analysis is done for finite $n$, and all finite size effects and errors are accounted for. The ${\epsilon}$ includes the error in the choice of random code in the protocol, finite size effects that need to be bounded, smoothing parameters from an uncertainty relation, etc. Our experimental implementation demonstrates for the first time that two-party protocols proposed in the bounded and noisy-storage models are well within today’s capabilities.
Discussion
==========
We demonstrated, for the first time, that two-party protocols proposed in the bounded and noisy-storage models can be implemented today. We emphasize that whereas - like so many experiments in quantum information - our experiment is extremely similar to QKD the experimental parameter requirements and analysis is entirely different to QKD. Where there are many experiments carrying out QKD, there are only a handful of implementation results for two party protocols [@nguyen08; @berlin11]. Bit commitment is one of the most fundamental protocols in cryptography. For example, it is known that with bit commitment, coin tossing can be built. Also using additional quantum communication we can build oblivious transfer [@yao:otFromBc], which in turn enables us to solve any two-party cryptographic problem [@kilian]. In the [Supplementary Methods]{}, we provided a detailed analysis of our modified bit commitment protocol including a range of parameters for which security can be shown. Our analysis could be used to implement the same protocol using a different, technologically simpler setup, with potentially lower error rates or losses. Our analysis can also address the case of committing several bits at once.
It would be interesting to see implementations of other protocols in the noisy-storage model.
Finally, note that our analysis rests on a fundamental assumption made in in the analysis of *all* cryptographic protocols, namely that Alice does not have access to Bob’s lab and vice versa. In particular, this means that Alice cannot tamper with the random choices made by Bob, potentially forcing him to measure e.g. only in one basis, or by maniplating apparent detector losses [@makarov:05; @gerhardt:11a].
Methods
=======
Parameter ranges {#parameter-ranges .unnumbered}
----------------
Our theoretical analysis shows security for a general range of parameters as illustrated in Figures \[fig:securityregion2a\], \[fig:securityregion1a\] and \[fig:securityregion3a\]. A fully general theoretical statement can be found in the [Supplementary Methods]{}. These plots demonstrate that security is possible for a wide range of parameters, of which our particular implementation forms a special case. The plots are done for fixed values of $n=250000$ and a total execution error of $\epsilon=3\cdot 10^{-4}$, unless otherwise indicated. Finally, Bob’s storage size is quantified by $S$, the number of qubits that Bob is able to store. The plots assume a memory of $S$ qubits, where each qubit undergoes depolarizing noise with parameter $r=0.9$.
![\[fig:securityregion2a\]**Security region for $p^1_{\rm sent}$ versus $p^{\rm h}_{\rm B, no click}$.** Plots were done for distinct values of $p_{\rm err}$, while storage size is fixed $S=2500$, and $p_{\rm B, no click}^{\rm d} = 0$. For small values of $p_{\rm B, no click}^{\rm h}$ (large amounts of losses), there exists a threshold on $p_{\rm sent}^1$ for the protocol to be secure. This threshold increases with $p_{\rm err}$, and for extremely small storage rates, it gives a maximal tolerable $p_{\rm err}\approx 0.046$. ](figures/g2_v2){width="0.9\columnwidth"}
![\[fig:securityregion1a\] **Security region for some typical parameter ranges.** $p_{\rm B, no click}^{\rm h}$ and $p_{\rm err}$ quantify the amount of erasures and errors in the protocol. For higher summation values of $p_{\rm B, no click}^{\rm d}+p_{\rm sent}^1$, the less multi-photons Bob gets, and erasures have less impact on the protocol security. This implies if the source is ideal, the protocol remains secure for large values of erasures. Dependences in the security region between erasures and errors also become more obvious when $p_{\rm B, no click}^{\rm d}+p_{\rm sent}^1$ is low. Furthermore, large assumptions on $S$ directly decrease the amount of min-entropy, causing tolerable $p_{\rm err}$ to drop consistently for all amounts of erasures.](figures/g1_v2){width="0.9\columnwidth"}
![\[fig:securityregion3a\]**Security region for different storage size $S$ and error rate $p_{\rm err}$, with $p_{\rm sent}^1=0.765$, and $p_{\rm B, no click}^{\rm d}=0.234$ fixed.** This plot shows a monotonic decreasing trend for tolerable $p_{\rm err}$ w.r.t storage size $S$. The sharp cut-off for $S$ varies with $p_{\rm B, no click}^{\rm h}$, since with lower detection efficiency, dishonest Bob can report more missing rounds, hence the lower his storage size has to be for security to hold. Also, the plot shows security for mostly low values of storage rate. The result is non-optimal, since it has been shown [@entCost] that security can be achieved with arbitrarily large storage sizes, if the depolarizing noise parameter $r \lesssim 0.7$. This is because we bound the smooth min-entropy of an adversarial Bob by the *classical capacity* of a quantum memory, while [@entCost] does so in terms of *entanglement cost*. Since the latter is generally smaller than the former, this poses a better advantage for security which is not shown in our analysis. ](figures/g3_v2){width="\columnwidth"}
Experimental Implementation {#experimental-implementation .unnumbered}
---------------------------
We implement this protocol with a series of entangled photons, with the polarization degree of freedom forming our qubits. This allows for reliable measurements in two complementary bases. Basis 1 corresponds to horizontal/vertical (HV) polarization, and basis 2 to $\pm45^\circ$ (+-) linear polarization. The polarization-entangled photon pairs are prepared via spontaneous parametric down conversion (SPDC), collected into single mode optical fibers, and guided to polarization analyzer (PA) located with Alice and Bob (see figure \[fig:setup\]). Each PA consists of a non-polarizing beam splitter (BS) providing a random basis choice, followed by two polarizing beam splitters (PBS) and a pair of silicon avalanche photodiodes (APD) as single photon detectors in each of the BS outputs. A half wave plate before one of the PBS rotates the polarization by 45$^\circ$ degrees. This detection setup was used in a number of QKD demonstrations [@kurtsiefer:02b; @marcikic:06; @ling:08].
![\[fig:setup\] Experimental setup. Polarization-entangled photon pairs are generated via non-collinear type-II spontaneous parametric down conversion of blue light from a laser diode (LD) in a barium-betaborate crystal (BBO), and distributed to polarization analyzers (PA) at Alice and Bob via single mode optical fibers (SF). The PA are based on a nonpolarizing beam splitter (BS) for a random measurement base choice, a half wave plate ($\lambda/2$) at one of the of the outputs, and polarizing beam splitters (PBS) in front of single-photon counting silicon avalanche photodiodes. Detection events on both sides are timestamped (TU) and recorded for further processing. A polarization controller (FPC) ensures that polarization anti-correlations are observed in all measurement bases.](figures/setup){width="\columnwidth"}
The SPDC source is similar to [@ling:08], with a continuous wave free running laser diode (398nm, 10mW) pumping a 2mm thick Barium-betaborate crystal cut for type-II non-collinear parametric down conversion and the usual walk-off compensation to obtain polarization-entangled photon pairs [@kwiat:95]. We collect photon pairs into single mode optical fibers such that we observe an average pair rate $r_\mathrm{p}=2997\pm
82$s$^{-1}$.
Such a source generates photon pairs in a stochastic manner, but with a strong correlation in time. Therefore, valid clicks are timestamped on both sides first. In a classical communication step, detection times $t_{\rm A},t_{\rm B}$ are compared, and valid rounds are identified if valid clicks fall into a coincidence time window of $\tau_\mathrm{c}=3$ns, i.e., $|t_\mathrm{A}-t_\mathrm{B}|\le\tau_\mathrm{c}/2$, similar to [@marcikic:06] with the code in [@kurtsiefer:08]. The visibility of the polarization correlations in the Singlet state are $97.7\pm0.6$% and $94.7\pm0.9$% in the HV and $\rm45^\circ$ linear basis. Individual detection rates on both sides are $r_\mathrm{A}=23758\pm221$s$^{-1}$ and $r_\mathrm{B}=22227\pm247$s$^{-1}$ on Alice and Bob’s side, respectively. In an initial alignment step, the fiber polarization controller was adjusted such that we see polarization correlations corresponding to a singlet state with a quantum bit error ratio (QBER) of about $p_{\rm err}=4.1$%. The QBER is not to be confused with the failure probability of bit commitment protocol. Calculations of the latter are explicitly stated in the [Supplementary Methods]{}. As reported in the summarizing paragraph of our introduction, this quantity is much smaller than the former.
For carrying out a successful bit commitment, we need to determine the parameters $p^1_{\rm sent}$, $p^0_{\rm sent}$, and $p_{\rm B, no click}^{\rm h}$. Depending on these probabilities and the desired error parameter $\epsilon$, we choose a particular error correcting code and number of rounds $M$ needed for a successful bit commitment. To estimate these probabilities out of the experimental parameters of our source/detector combination, we model our setup by a lossless SPDC source emitting only photon pairs at a rate $r_{\rm s}$, and assign all imperfections (losses, limited detection efficiency, and background events) to the detectors at Alice and Bob. Since the coherence time of the photons in our case is much shorter than the coincidence detection time window $\tau_{\rm c}$, the distribution of photon pairs in time can be well described by a Poisson process, which allows an assessment of multiphoton events. A detailed derivation of bounds for the probabilities is given in the [Supplementary Methods]{}, we just summarize the results necessary for evaluating the security of the protocol: $$\begin{aligned}
p_{\rm sent}^0&\le&(r_{\rm A}-r_{\rm p})/r_{\rm A}=0.875\pm0.009\,, \\
p_{\rm sent}^{n>1}&<&{r_{\rm A} r_{\rm B}\over r_{\rm p}}\tau_{\rm c}=5.32\pm0.17\times10^{-4}\,,\\
p_{\rm sent}^1&=&1-p_{\rm sent}^0-p_{\rm sent}^{n>1}>0.125\pm0.009\,,\\
p_{\rm sent}^0+p_{\rm sent}^1&=&1-p_{\rm sent}^{n>1}>0.99947\pm0.000017\,,\\
p_{\rm B, no click}^{\rm h}&=&1-r_{\rm p}/r_{\rm A}=0.875\pm0.009\,.\end{aligned}$$
Due to small differences in the detection efficiency of the APD and imperfections in polarization components in the actual experiment, there is an asymmetry in the probability of detecting each bit in each basis. Furthermore, the beam splitter for the random measurement basis choice are not completely balanced. A summary of these imperfections over a number of bit commitment runs is shown in figure \[fig:asymmetry\]. This can be corrected for by discarding rounds until the probabilities for both bits are equal. Discarded bits can be modeled as losses without affecting the security of the protocol. A detailed analysis of this can be found in the [Supplementary Methods]{}.
![\[fig:asymmetry\]Bias in measurements. Solid lines indicate the probabilities $P(HV)$ of a HV basis choice for both Alice and Bob for data sets of $250000$ events each. Dashed lines indicate the probability $P(H)$ of a H in the HV measurement basis, the dotted lines the probability $P(+)$ of a $+45^\circ$ detection in a $\pm45^\circ$ measurement basis. These asymmetries arise form optical component imperfections and are corrected in a symmetrization step.](figures/asymmetry3a){width="\columnwidth"}
NN, CK and SW designed the research. SJ, CM and CK carried out the experiment. NN wrote the software. NN and SW performed the theoretical analysis. NN, SJ, CK and SW wrote the paper. There are no competing financial interests.
-35cm
Supplementary Figures
=====================
![\[fig:model\]**Model of experimental setup.** An ideal source generates time-correlated photon pairs with a rate $r_{\rm s}$ and sends them to detectors at Alice and Bob. The losses (due to all causes including source imperfections and detection efficiencies) are modeled with attenuators with a transmission $\eta_{\rm A}$ and $\eta_{\rm B}$, respectively. To cater for dark counts in detectors, fluorescence background and external disturbances, we introduce background rates $r_{\rm bA},r_{\rm bB}$ on both sides. Valid rounds are identified by a coincidence detection mechanism that recognizes photons corresponding to a given entangled pair. Event rates $r_{\rm A}$ and $r_{\rm B}$ reflect measurable detection rates at Alice and Bob, while $r_{\rm p}$ indicates the rate of identified coincidences.](figures/model){width="0.9\columnwidth"}
![\[fig:channel\]**The encoding and decoding of a message.** A total of *k* bits were encoded into *n* bits and sent through the noisy channel, then recovered completely after undergoing the transmission process.](figures/tworeloaded){width="0.9\columnwidth"}
![\[fig:reldist\]**Relative distance of random code versus code rate.** A randomly generated code reaches the bound given in Theorem \[randomcode\] with overwhelming probability. Meanwhile, the MRRW2 bound [@errorcorrcodes] is the smallest upper bound derived up to the present, and it is not known if this bound is tight. It is also not known if there exists any linear binary codes at all between the two regions.](figures/gv_reworked){width="0.9\columnwidth"}
Supplementary Tables
====================
Probabilities Description
------------------------------ ---------------------------------------------------------- --
$p^1_{\rm sent}$ Probability that a single photon was sent to Bob.
$p_{\rm B,no click}^{\rm h}$ Probability that honest Bob observes no click.
$p_{\rm B,no click}^{\rm d}$ Probability that dishonest Bob observes no click.
Note: this value is equal to $p_{\rm sent}^0$, i.e., the
probability that no photons were sent to Bob.
$p_{\rm err}$ Probability that the measurement outcome for
honest Alice and honest Bob is different,
when the same basis is used for both parties.
: \[probabilities\]**Parameters required for security proof of bit commitment.** All the above quantities are conditioned on the event that Alice registered a valid click.
Supplementary Discussion
========================
Properties of Error-correcting codes {#section:codes .unnumbered}
====================================
A linear error-correcting code can be defined by specifying its parity check matrix H, which has dimensions $n \times (1-R)n$. Given a vector **x** of length $n$, the parity check syndrome is simply Syn(**x**)=**x$\cdot$H**. [@errorcorrcodes]\
Recall that in our BC protocol, both parties agree on a code beforehand, and during the commit phase, Alice sends the syndrome to Bob. The syndrome mainly serves as a back-checking procedure for Bob during the open phase to confirm that Alice is honest. The longer the length of the syndrome, the more information about $X^n$ is given to Bob, and the harder it becomes for Alice to cheat. Both Alice and Bob agree on the code used, and for our purposes the complications in decoding is unnecessary, as an honest Bob never needs to decode.\
In the theory of error-correcting codes, a question of much significance is depicted in [Supplementary Figures]{} \[fig:channel\] : given a scenario where information is sent through an unavoidable noisy channel, under what conditions does an encoding scheme exist such that the message can be recovered completely after undergoing the communication process? In other words, given a message Y comprising of $k$ bits and a noisy channel for communication, what is the theoretical minimum length of encoded message $n$, such that the decoding can detect errors and recover Y accurately?\
It has been shown by Shannon that for the recovery of information to be possible, the fraction $\frac{k}{n}$ has a theoretical upper bound C, known as the capacity of the channel. For any value $R=\frac{k}{n} > C$, decoding is never possible. For a binary symmetric channel, the capacity is proven to be $$C_{BSC}(p_{\rm err})=1-h(p_{\rm err}),$$ where $${\rm h}(p_{\rm err})= - p_{\rm err}\log_{2}p_{\rm err} - (1-p_{\rm
err})\log_2(1-p_{\rm err})$$ is the binary entropy of the BSC channel.\
For values of code rate R strictly above the channel capacity, the success probability of delivering the message is exponentially decreasing with code length regardless of the encoding/decoding scheme used.\
Besides the code rate, another important quantity of error-correcting codes is the minimum distance $d$. Given an error-correcting code, this quantity shows the minimum hamming distance between two strings that have the same parity check syndrome. The larger the minimum distance, the more effective a code is at correcting errors. In the subsequent section, we investigate the relation of parameters $R$ and $d$ for randomly generated codes, and show that random codes satisfy our requirements on these parameters for the protocol to be secure.
Random Codes {#random-codes .unnumbered}
------------
Given a parity check matrix constructed randomly, we are interested in what is the minimum distance of this code. This problem is a computationally NP-hard one, but we do know some probabilistic facts about the minimum distance, which is stated in the theorem below:
\[randomcode\](Random codes, [@gallager]) Given a randomly generated binary linear code with rate $R$, the probability that minimum distance $d$ is smaller than some $\delta n$ is bounded by the following: $$\Pr[d \leq \delta n] \leq 2^{(R-C_\delta)n},\qquad \mbox{for }0\leq\delta\leq 1.$$
For large block lengths, we can see that this bound approaches a step function where for rates $R<C_\delta$, minimum distance is expected to be larger than $\delta n$ except with extremely small probability. For our choices of block lengths, the randomly generated code will satisfy the bound on minimum distance whenever $R<C_{\delta}$, except for some minimal probability that is later added into the $\epsilon$-error of the protocol. We plot this bound in [Supplementary Figures]{} \[fig:reldist\] with respect to the parameter $\delta = \frac{d}{n}$, which we refer to as the relative distance.\
Given values of $p_{\rm err}$ and reasonably small error parameter $\epsilon$, by referring to the conditions for minimum distance derived in the security analysis, we obtain the upper bound on the achievable rate, namely $C_{\delta}$. This guarantees that for small enough error rates $p_{\rm err}$, it is sufficient to use a randomly generated code for the use of our protocol, which will provide us both a good enough distance and code rate, except with an extremely small probability. By using Theorem \[randomcode\] we account for the probability of error and add it as a source of error for the execution of the protocol.\
### LDPC codes {#ldpc-codes .unnumbered}
Random binary codes are generated by assigning values 0 and 1 randomly to each element of the parity check matrix. They have a high density (large fraction of non-zero elements), which in large block length limit is time-consuming to deal with. For efficiency purposes, it is of interest whether we can construct codes with lower density (less non-zero values).\
In [@gallager], Gallager has shown that a specific ensemble of low density codes do attain the same limit given for the random codes as above, when considering large enough block lengths. These codes involve using random permutations of a submatrix, and the construction is straightforward. This type of codes can be of future interest, because their usage will shorten the calculational time used in the protocol. However, this is achieved at the expense of introducing an additional error probability of constructing a bad code (one with unsatisfactory minimum distance), which is not straightforward to evaluate.\
### Concatenated codes {#concatenated-codes .unnumbered}
In the case of low bit-flip error, classes of explicit concatenated codes might be generated such that the minimum distance is guaranteed without introducing any probabilistic errors from a randomized construction. These codes are constructed by using a Reed-Solomon code as an outer code, while using a smaller binary linear code as an inner code. We state the properties of such concatenated codes in the following theorem:
Given a $[n_1,R_1n_1,d_1]$ outer code, and a linear binary code with parameters $[n_2,R_2n_2,d_2]$. Then the resulting concatenated code has parameters $[n_1n_2,k_1 k_2, d]$, where the code rate $R=R_1R_2$ and $d \geq d_1 d_2$.
For example, by exploiting this construction, a linear binary concatenated code with rate R=0.53 and relative minimum distance $\delta\geq0.052$ can be constructed, where the code length $n=311296$. This value of $\delta$ has a large discrepancy compared to the probabilistic argument for a random code. From here it is clearly shown that, if a definite statement regarding the minimum distance of such large error-correcting codes (without any probabilistic errors) is desired, one can still obtain security for smaller ranges of experimental parameters. For the given example of concatenated RS code, this corresponds to security for bit flip error rates $p_{\rm err}\leq 0.02$, which exceeds the value obtained in our experiment.
Supplementary Methods
=====================
Experimental parameters\[section:expparams\]
--------------------------------------------
To analyze our bit commitment protocol in any practical experiment, several probabilities have to be determined. [Supplementary Tables]{} \[probabilities\] summarizes all the probabilities we will need to estimate. We emphasize that all such probabilities are conditioned on the event that Alice registers a round, i.e. sees a valid click.
A difficulty in estimating the probabilities of success in a “round” arises from the fact that generation of photon pairs in a parametric down conversion source is a stochastic process. Furthermore, losses in the system may occur in the source or in detectors, and we do not have an easy way of assessing the losses reliably. We thus try to estimate bounds of the required probabilities for the bit commitment protocol out of observable quantities both Alice and Bob can agree upon. For this purpose, we model losses and background events in our system in a way shown in [Supplementary Figures]{} \[fig:model\].
The rates (i.e., events per unit of time) observed at Alice are then given by $$r_{\rm A}=\eta_{\rm A}(r_{\rm s}+r_{\rm bA})\,,\label{eq:singlesAlice}$$ where $\eta_{\rm A}$ indicates the detection efficiency and $r_{\rm bA}$ a background event rate; a similar expression holds for Bob. The observed coincidence rate in this model is given by $$r_{\rm p}=\eta_{\rm A}\eta_{\rm B}r_{\rm s}+r_{\rm acc}\,.$$ where $r_{\rm acc}$ reflects the so-called accidental coincidence rate, caused by detection events on both sides happening within the coincidence time window $\tau_{\rm c}$ that are not due to valid clicks form the same photon pair. This rate can be bounded from observed rates $r_{\rm A}$ and $r_{\rm B}$ to $$r_{\rm acc}<r_{\rm acc}^{\rm max}=r_{\rm A}r_{\rm B}\tau_{\rm c}\,,$$ assuming that all detection events on both sides are caused by uncorrelated events. In our experiment, this quantity would result in a value of $r_{\rm
acc}^{\rm max}=14.9\pm 0.18$s$^{-1}$, and is negligible compared to the observed coincidence rate $r_{\rm s}$. This quantity was independently assessed by recording the rate of detection time pairings $t_{\rm A},t_{\rm B}$ in a time window that was displaced by $\tau_{\rm d}=20$ns from the “true” coincidences, i.e., $|t_{\rm A}-t_{\rm B}-\tau_{\rm d}|\le\tau_{\rm c}$ [@marcikic:06]. We found a rate of $r_{\rm acc}=5.3\pm3.3$s$^{-1}$ over the course of several bit commitment runs. Since $r_{\rm acc}\ll
r_{\rm p}$, we from now on neglect these events in the rate estimations, and interpret their occurence just as events that increase the error ratio.
To evaluate the probability $p^1_{\rm sent}$ that exactly one photon was sent to Bob in the interval $\tau_{\rm c}$ around a time when Alice has seen an event, we first consider the probability $p^0_{\rm sent}$ that no photon was sent to Bob, given Alice has seen an event. This can only be caused by a background event with Alice. Thus, $p^0_{\rm sent}$ equals the probability that a detection event on Alice’s side is caused by background, which is given by $$\begin{aligned}
p^0_{\rm sent}&=&{r_{\rm bA}\over r_{\rm bA}+r_{\rm s}}=1-{r_{\rm s}\over r_{\rm bA}+r_{\rm s}}\nonumber \\
&=&1-{\eta_{\rm A} r_{\rm s}\over\eta_{\rm A}(r_{\rm bA}+r_{\rm s})}=1-{\eta_{\rm A}r_{\rm s}\over r_{\rm A}}\nonumber \\
&=&1-{r_{\rm p}\over\eta_{\rm B} r_{\rm A}}\label{eq:p0sent}\end{aligned}$$ Since the efficiency $\eta_{\rm B}$ is not known exactly, we set it to 1 and thereby obtain an upper bound for $p^0_{\rm sent}$: $$p^0_{\rm sent}<1-{r_{\rm p}\over r_{\rm A}}=0.875\pm0.009$$
Next, we consider the probability $p^{n>1}_{\rm sent}$ that more than one photon has been sent to Bob, given that Alice has seen an event. This probability is the product of the probability that Alice’s event was caused by a photon pair, and the probability that at least one other photon pair than the one causing the event on Alice’s side was generated in the coincidence time window $\tau_{\rm c}$. From equation \[eq:p0sent\], the first probability is given by $r_{\rm p}/(\eta_{\rm B}r_{\rm A})$. For the latter, we consider the statistics of photon pairs emerging from a continuously pumped SPDC source. While light emerging from a downconversion process is known to follow thermal photon counting statistics, the coherence time of the photons in our case (0.73ps for an optical bandwidth of 3nm) is much shorter than $\tau_{\rm c}$. In this case, the statistics of several photon pairs in time window $\tau_{\rm c}$ follows a Poisson distribution. Since the creation of an additional photon pair is then independent of the first photon pair, and the probability that no photon pair is created in $\tau_{\rm c}$ is given by $e^{-r_{\rm s}\tau_{\rm c}}$, the probability of creating at least one more photon pair is given by $1-e^{-r_{\rm s}\tau_{\rm c}}$. This brings us to $$\begin{aligned}
p^{n>1}_{\rm sent}&=&{r_{\rm p}\over \eta_{\rm B}r_{\rm A}}(1-e^{-r_{\rm s}\tau_{\rm c}})\nonumber \\
&<&{r_{\rm p}\over \eta_{\rm B}r_{\rm A}}r_{\rm s}\tau_{\rm c} =
{r_{\rm p}\over \eta_{\rm B}r_{\rm A}}{r_{\rm p}\over \eta_{\rm A}\eta_{\rm B}}\tau_{\rm c}\nonumber \\
&=&{r_{\rm p}^2\over r_{\rm A}\eta_{\rm A}\eta_{\rm B}^2}\label{eq:pg1interim}\end{aligned}$$ The efficiencies $\eta_{\rm A}$, $\eta_{\rm B}$ are not accessible directly from the experiment, but can be bounded by $\eta_{\rm A}>r_{\rm p}/r_{\rm B}$ and $\eta_{\rm B}>r_{\rm p}/r_{\rm A}$ via \[eq:singlesAlice\]. With this, we can further bound expression \[eq:pg1interim\] and arrive at $$p^{n>1}_{\rm sent}<{r_{\rm A}r_{\rm B}\over r_{\rm p}}\tau_{\rm c} = 5.32\pm0.17\times10^{-4}\,,$$ which is much smaller than the uncertainty on $p^0_{\rm sent}$. With this, we arrive at $$p^1_{\rm sent}=1-p^0_{\rm sent}-p^{n>1}_{\rm sent}>0.125\pm0.009$$ and $$\label{a9}
p^1_{\rm sent}+p^0_{\rm sent}=1-p^{n>1}_{\rm sent}>0.99947\pm0.000017$$
Finally, the probability for an honest Bob not seeing an event in a coincidence time window if Alice has detected something is the complement to the probability that Bob sees something if Alice has seen something. The latter, by definition, is given by the ratio $r_{\rm p}/r_{\rm A}$. Thus, we have $$p_{\rm B, no click}^{\rm h} = 1-r_{\rm p}/r_{\rm A} = 0.875\pm0.009\,.$$
Symmetrizing losses {#sec:symLoss}
-------------------
In practice, not all detectors have the same efficiency. Losses will be higher for some detectors than for others. This will lead to imbalances in the choice of basis and the choice of BB84 encoded qubit. In our protocol, such imbalances affect the security in two places. First, if Alice is honest, but Bob is trying to cheat, such imbalances give him additional information about which bit or basis was used. His advantage is similar to the advantage that an eavesdropper in QKD would gain from knowing such extra information. Second, if Bob is honest, but Alice is trying to cheat, having higher losses in one basis does reveal information to Alice in which basis Bob measured - if Bob does not report a loss it is more likely that he used the basis for which losses occur less often.
We describe a method to deal with such imbalances securely - the same method can be used to address imbalances on Alice’s and Bob’s side. For simplicity, we outline the procedure in detail for Alice; exactly the same method can be used to symmetrize Bob’s detectors. The essential idea is to make all detectors equally inefficient, by throwing away (i.e., declaring as lost) rounds where detectors with higher efficiencies registered a click. Note that in our protocol, Alice can discard additional rounds without consequences for security parameters. Meanwhile, discarding additional rounds on Bob’s side increases $p_{\rm B, no click}^{\rm h}$. Detection events combining with such post-processing procedures, define the occurrence of a valid round. In other words, if a single click occurred on both sides and was not manually discarded for symmetrizing purposes, this event is considered a valid round.
In our setup, Alice has four detectors, one for each bit in each basis. Let $x,\theta$ label the detector corresponding to a bit $x \in \01$ in basis $\theta \in \01$. Let $p_{\theta}$ denote the probability that basis $\theta$ is chosen, and let $p_{x|\theta}$ denote the probability that bit $x$ occurs given basis $\theta$. Finally, let $t_{x,\theta}$ denote the probability that Alice keeps bit $x$ in basis $\theta$ when the detector $x,\theta$ clicks. That is, Alice discards bit $x$ in basis $\theta$ with probability $1-t_{x,\theta}$ even though a click occurred. Our goal will be to determine the $t_{x,\theta}$ that renders $\Pr[x,\theta|{\rm keep}]$, the probability that $x,\theta$ occurs conditioned on the event that Alice keeps a particular detection event the same for all $x$ and $\theta$.
First of all, note that the probability that a particular detection event is *not* discarded, i.e. Alice accepts it as a round, can be written as $$\begin{aligned}
\Pr[{\rm keep}] = \sum_{x,\theta \in \01} p_{\theta} p_{x|\theta} t_{x,\theta}\ .\end{aligned}$$ By Bayes’ rule $$\begin{aligned}
\Pr[x,\theta|{\rm keep}] &= \frac{\Pr[{\rm keep}|x,\theta] \Pr[x,\theta]}{\Pr[{\rm keep}]}\\
&= \frac{t_{x,\theta} p_{x|\theta} p_{\theta}}{\Pr[{\rm keep}]}\ .\end{aligned}$$ Ideally, all probabilities are the same, i.e., for all $x$ and $\theta$ $$\begin{aligned}
\Pr[x,\theta|{\rm keep}] = \frac{1}{4}\ .\end{aligned}$$ This yields $4$ equations, in $3$ free parameters since $\sum_{x,\theta} t_{x,\theta} = 1$. These can easily be solved for $t_{x,\theta}$.
For our setup, the parameters for symmetrization on Alice’ side are as follows: $$\begin{aligned}
&& t_{0,0} = 1 \nonumber\\
&& t_{0,1} = 0.963077 \nonumber\\
&& t_{1,0} = 0.882305 \nonumber\\
&& t_{1,1} = 0.871353.\end{aligned}$$ Symmetrization on Bob’s side is dealt with in the same manner. This however increases the value of $p_{\rm B, no click}^{\rm h}$, since now an honest Bob deliberately throws away more clicks. This leads to a new value of $$\tilde{p}_{\rm B, no click}^{\rm h} = 1 - (1-p_{\rm B, no click}^{\rm h})\cdot \Pr\left[{\rm keep}\right].$$ For our setup, the parameters for symmetrization on Bob’s side are: $$\begin{aligned}
&& t_{0,0} = 0.679745 \nonumber\\
&& t_{0,1} = 1 \nonumber\\
&& t_{1,0} = 0.665591 \nonumber\\
&& t_{1,1} = 0.662890.\end{aligned}$$
The probability of Bob keeping a click during symmetrization is $\Pr\left[{\rm keep}\right]=0.729646$. This combining with the initial estimate of $p_{\rm B, no click}^{\rm h}$ gives $\tilde{p}_{\rm B, no click}^{\rm h}=0.909$, implying a high amount of losses. Even so, the protocol remains secure due to the fact that the source provides multi-photons to Bob with an extremely small probability, whenever Alice observes only a single detection event. In other words, $p_{\rm sent}^1 + p_{\rm B, no click}^{\rm d}$ is extremely high, as stated in . In such cases, even a high amount of losses do not compromise security of the protocol.
Also, it should be stressed that $p_{\rm err}$ should be evaluated for the set of data after all symmetrization procedures, since there can be bias in the error rates for each bit and basis. For the set of symmetrized data, $p_{\rm err}=0.0412$, in comparison with before symmetrization, $p_{\rm err}=0.0428$.
Theoretical security analysis {#sec:analysis}
-----------------------------
In the security proof, we divide the protocol into two parts: the first part is Weak String Erasure with Errors (WSEE), and the remaining procedure is Bit Commitment (BC).
Theoretical parameters {#theoreticalparam .unnumbered}
----------------------
Next to the experimental parameters defined in the [Supplementary Tables]{} \[probabilities\], our analysis will make frequent use of the following parameter definitions. There are two more basic parameters in this analysis: $M$ and $\epsilon$.
The parameter $\epsilon$ represents a fixed error parameter, i.e., we want to achieve security up to an error of $O({\epsilon})$. This parameter is used to bound the occurrence probability of bad events, and we need to frequently refer to it throughout the analysis. Such bounds are achieved by making use of the Hoeffding inequality. It says that given a random variable $X_{j} \in \lbrace0,1\rbrace$, where $\Pr(X_{i}=0)=1-p$, $\Pr(X_{i}=1)=p$, and $Y=\Sigma_{i=1}^{N} X_{i}$ we have $$\Pr [ Y \leq (p-\alpha)N] = \Pr [Y \geq (p+\alpha)N] = e^{-2\alpha^{2}N}.$$ The way we will use the Hoeffding inequality is that we demand that $e^{-2 \alpha^2 N} \leq {\epsilon}$, and then solve for $\alpha$ such that our demand is satisfied.
Meanwhile, $M$ denotes the number of signals that Alice counted as valid, i.e., she registers a round (but not necessarily Bob as well).
Based on $\epsilon$ and $M$, we will need the following definitions: $$\begin{aligned}
\label{param}
\zeta_{\rm B, no click}^{\rm h} &:= \sqrt{\frac{\ln \frac{2}{\epsilon}}{2M}}\nonumber\\
n &:= (1-p_{\rm B, no click}^{\rm h}-\zeta_{\rm B, no click}^{\rm h}) M \nonumber\\
\alpha_1 &:=\sqrt{\frac{\ln\frac{1}{\epsilon}}{2n}}\nonumber\\
m &:= \left({\frac{1}{2}}- \alpha_1\right)n\nonumber\\
\alpha_2 &:= \sqrt{\frac{\ln\frac{2}{\epsilon}}{2m}}\nonumber\\
\alpha_3 &:= \sqrt{\frac{\ln\frac{1}{\epsilon}}{d}}\end{aligned}$$ where $d$ is the minimum distance of the error-correcting code used in the protocol, and $n$ is the number of valid rounds that remain.
Weak string erasure {#weak-string-erasure .unnumbered}
-------------------
### Definition {#definition .unnumbered}
We first provide an informal definition of weak string erasure with errors (WSEE). A formal definition can be found in [@noisy:new]. When both Alice and Bob are honest, an $(n,\lambda,\epsilon,p_{\rm err})$-WSEE scheme provides Alice with a string $X^n$ and Bob with a randomly chosen subset ${\mathcal{I}}\in [n]$, as well as a substring $\tilde{X}_{\mathcal{I}}$. This substring is thereby given by the substring $X_{\mathcal{I}}$ (the bits of $X^n$ corresponding to the indices in $\mathcal{I}$) passed through a binary symmetric channel that flips each bit of $X_{\mathcal{I}}$ with probability $p_{\rm err}$.
To specify the security condition against dishonest Bob, we first need to quantify the uncertainty of Bob about $X^n$, given access to the entire system of a dishonest Bob denoted as ${\rm B'}$. This is done by lower bounding the min-entropy of $X^n$ conditioned on Bob’s information, $$\begin{aligned}
{{\rm H_{\rm min}}}(X^n|{\rm B'})_{\rho_{X^n{\rm B'}}}&:=&-\log P_{\rm guess} (X^n|{\rm B'}) \nonumber\\[0pt]
P_{\rm guess}(X^n|{\rm B'}) &:=& \max_{ \lbrace D_x \rbrace_x} \displaystyle\sum_x P_X(x) {\mathop{\mathrm{tr}}\nolimits}(D_x\rho_x),\end{aligned}$$ where $P_{\rm guess}$ is referred to as the guessing probability, namely the probability that Bob correctly guesses $X^n$, maximized over all measurement strategies upon his system ${\rm B'}$ [@krs:operational]. The *$\epsilon$-smooth min-entropy* is defined as $${{\rm H}^{\epsilon}_{\rm min}}(X^n|{\rm B'})_{\rho_{X^n{\rm B'}}} := \sup_{\rho'}~{{\rm H_{\rm min}}}(X^n|{\rm B'})_{\rho'}$$ maximized over all states $\rho'$ such that the purified distance $C(\rho',\rho_{X^n{\rm B'}})= \sqrt{1-F^2 (\rho',\rho_{X^n{\rm B'}})} \leq \epsilon$, where $F(\rho,\tau)$ denotes the fidelity of states $\rho$ and $\tau$. Intuitively, this quantity behaves like the min-entropy, except with a probabilistic error $\epsilon$.\
We can now state the security conditions for an $(n,\lambda,\epsilon,p_{\rm err})$-WSEE:\
[**1. Security for Alice:**]{} If Alice is honest, then the amount of information a dishonest Bob holds about $X^n$ is limited, i.e. the $\epsilon$-smooth min entropy of $X^n$ conditioned on a dishonest Bob’s information is lower bounded $$\frac{1}{n}~{{\rm H}^{\epsilon}_{\rm min}}(X^n|{\rm B'}) \geq \lambda,$$ where $\lambda$ is referred to as the smooth min-entropy rate.\
[**2. Security for Bob:**]{} If Bob is honest, then Alice does not have any information ${\mathcal{I}}$. That is, Alice does not learn which bits of $X^n$ are known to Bob.
### Protocol {#protocol .unnumbered}
In principle, WSEE can be achieved experimentally by using any QKD device. However, we emphasize that the experimental requirements and analysis differs entirely. In particular, security of QKD for a particular setup does not imply security of bit commitment.
Recall from the informal statement of the protocol in the main part of our paper that if Alice herself concludes that no photon or a multi-photon has been emitted in a particular time slot, she simply discards this event and tells Bob to discard it as well. Since this action represents no security problem for us, we will for simplicity omit these events all-together when stating the more detailed protocol below. This means that $M$ in the protocol below, actually refers to the set of post-selected pulses that Alice did register as a round. In practice, Alice reports the missing events to Bob after the waiting time has passed. In principle, this could be used to obtain better security bounds as Bob does not yet know which bits are indeed relevant when he uses his storage device. However, we leave such a refined analysis for future work.
In addition, introducing time slots enables Bob to report a particular bit as missing, if he obtained no click in a particular time slot. Alice and Bob will subsequently discard all lost rounds. In the protocol below, we assume the detectors have already been symmetrized appropriately as outlined in the [Supplementary Methods]{} \[sec:symLoss\]. The purpose of symmetrizing is to ensure that losses are independent of basis choice, hence Alice cannot obtain any information about ${\mathcal{I}}$ by observing the rounds reported lost by Bob.
[1]{}[Weak String Erasure with Errors (WSEE)]{} [Outputs: $x^n \in {\{0,1\}}^n$ to Alice, $({\mathcal{I}},z^{|{\mathcal{I}}|}) \in 2^{[n]} \times {\{0,1\}}^{|{\mathcal{I}}|}$ to Bob.]{}\[proto:wse\]
[**Alice:**]{} Chooses a string $x^M \in_R \01^M$ and basis-specifying string $\theta^M \in_R \01^M$ uniformly at random.
[**Bob:** ]{} Chooses a basis string $\tilde{\theta}^M \in_R \01^M$ uniformly at random.
In time slot $i=1,\ldots,M$:
1. [**Alice:**]{} Encodes bit $x_i$ in the basis $\theta_i$ (i.e., as $H^{\theta_i}{|x_i\rangle}$), and sends the resulting state to Bob.
2. [**Bob:**]{} Measures in the basis given by $\tilde{\theta}_i$ to obtain outcome $\tilde{x}_i$. If Bob obtains no click in this time slot, he records round $i$ as lost.
[**Bob:** ]{} Reports missing rounds to Alice.
[**Alice:** ]{} If the number of rounds that Bob reported missing does not lie in the interval $[(p_{\rm B, no click}^{\rm h} - \zeta_{\rm B, no click}^{\rm h})M,(p_{\rm B, no click}^{\rm h} + \zeta_{\rm B, no click}^{\rm h})M]$, then Alice aborts the protocol. Otherwise, she deletes all bits from $x^M$ that Bob reported missing. Let $x^n \in \01^n$ denote the remaining bit string, and let $\theta^n$ be the basis-specifying string for the remaining rounds. Let $\tilde{\theta}^n$, and $\tilde{x}^n$ be the corresponding strings for Bob.
[**Alice:** ]{} Sends the basis information $\theta^n$ to Bob, and outputs $x^n$.
[**Bob:** ]{} Computes ${\mathcal{I}}:= \{i \in [m] \mid \theta_i = \tilde{\theta}_i\}$, and outputs $({\mathcal{I}},z^{|{\mathcal{I}}|}):=({\mathcal{I}},\tilde{x}
_{{\mathcal{I}}})$.
How large is $n$ going to be? Since Alice aborts if Bob reports too many rounds as missing, we have that $n \geq (1 - p_{\rm B, no click}^{\rm h} -
\zeta_{\rm B, no click}^{\rm h}) M$. If a fixed $n$ is desired, we can take $n = (1 - p_{\rm B, no click}^{\rm h} -
\zeta_{\rm B, no click}^{\rm h}) M$ as in , where Alice randomly truncates the resulting string, and informs Bob about the truncation. This is the approach we take here. In our protocol, there is also a possibility that Alice aborts. An abort here means that Alice simply generates a random $x^n$ as output. This means that our protocol does at all times generate a string of length $n
= (1 - p_{\rm B, no click}^{\rm h} - \zeta_{\rm B, no click}^{\rm h}) M$.
### Analysis {#analysis .unnumbered}
The analysis of weak sting erasure with errors has already been performed in [@Curty10]. Essentially, losses allow a dishonest Bob to discard a fraction of single-photon detection events, and keep more multi-photon events so that his chance of guessing $X^n$ correctly is increased. The resulting min-entropy rate $\lambda$ can thereby be calculated as a function of experimental parameters listed in [Supplementary Tables]{} \[probabilities\]. That is, the min-entropy rate is a function of $p_{\rm sent}^1$, $p_{\rm B, no click}^{\rm h}$, $p_{\rm B, no click}^{\rm d}$, and $M$ $$\begin{aligned}
\lambda = \lambda(p_{\rm sent}^1,p_{\rm B, no click}^{\rm h}, p_{\rm B, no click}^{\rm d},M)\ .\end{aligned}$$ In our further analysis, we will hence assume that WSEE has been shown secure, and state security as a function of a fixed parameter $\lambda$ and fixed $n$. We then combine the two to give explicit security paramters as a function of the experimental parameters.
Bit commitment from weak string erasure {#bit-commitment-from-weak-string-erasure .unnumbered}
---------------------------------------
What remains is to analyse security of the bit commitment protocol(BC) based on WSEE. An informal definition of BC was stated in the introduction, and a formal one can be found in [@noisy:new]. The protocol below is very similar to the one proposed in [@noisy:new], which gave a BC protocol for weak string erasure *without* errors (i.e., $p_{\rm err} = 0$). To address the case of $p_{\rm err} > 0$, we introduce modifications to the BC protocol, allowing the modified protocol to stay secure up to a certain amount of bit flip error from the experimental setup.
### Protocol {#protocol-1 .unnumbered}
We present the fully modified BC protocol as Protocol 2, by including WSEE as a sub-protocol. Our protocol allows Alice to commit a string $D^l \in \01^l$ to Bob. However, for our experiment we chose to commit only a single bit $l=1$ which is the scenario typically considered in bit commitment. Our protocol makes use of the parameters defined in .
[2]{}[Non-Randomized Bit Commitment(BC)]{} [By using an binary linear error-correcting code $\mathcal{C}$, let Syn:$\lbrace0,1\rbrace^n \rightarrow \lbrace0,1\rbrace^{n-k}$ be the function that outputs the parity check syndrome for $\mathcal{C}$. Also, select Ext: $\lbrace0,1\rbrace^n \times \mathit{R} \rightarrow \lbrace0,1\rbrace ^l$ from a set of 2-universal hash functions. ]{}\[proto:bc\]
**Commit Phase**
[**Alice and Bob :**]{} execute $(n,\lambda,\epsilon,p_{\rm err})$WSEE. Alice obtains $X^n$ while Bob obtains $\tilde{X}_\mathcal{I}$ and $\mathcal{I}$.
[**Bob :**]{} Checks if $|{\mathcal{I}}| \geq m$. If so, he randomly truncates ${\mathcal{I}}$ until $|{\mathcal{I}}|=m$. Otherwise, he aborts the protocol.
[**Alice:**]{}\
a) computes w=Syn($X^n$) and sends it to Bob.\
b) picks a 2-universal hash function $r \in_R \mathcal{R}$ and sends it to Bob.
[**Alice:**]{} Commits $C^l \in \lbrace0,1\rbrace^l$ by computing $D^\mathit{l}={\mbox{Ext}}(X^n,r)$ and sends $E^l=C^l \oplus D^l$ to Bob.
**Open Phase**
[**Alice:**]{} reveals the complete string $X^n$ to Bob.
[**Bob:**]{} Perform checks:\
a) computing the syndrome and check that it agrees with $w$ sent by Alice.\
b) checking $X^n$ against $\tilde{X}_\mathcal{I}$, ensuring that the number of bits that disagree at positions in ${\mathcal{I}}$ lie in the interval \[$(p_{\rm err}-\alpha_2)m,(p_{\rm err}+\alpha_2)m$\].
[**Bob:**]{} If conditions are satisfied, he accepts commitment and calculates $D^\mathit{l} = {\mbox{Ext}}(X^n,r)$. Both of them output $C^l$.
The are two modifications in this protocol compared to its previous version in [@noisy:new]. The first is the use of a different error-correcting code $\mathcal{C}$. We will first provide a general analysis to prove the security of bit commitment, given that several conditions on both the rate and relative minimum distance of the error-correcting code used are met. We then show that by generating a binary linear code at random, the conditions on minimum distance and rate can be satisfied, except for a small probabilistic error that can be later added upon the total security error of the protocol.
Note that random codes do not pose a problem when executing the protocol since honest parties never need to decode. Details of the properties of error-correcting codes can be found in the [Supplementary Discussion]{}. Secondly, the checks performed by Bob in the open phase have been modified to account for the existence of bit flip errors, such that Bob tolerates a certain limited amount of errors.
### Analysis {#analysis-1 .unnumbered}
Intuitively, bit flip errors in WSEE give Alice more freedom to cheat, as a malicious Alice can avoid the errors and choose to corrupt the string $X^n$ herself. In other words, the actual bit flip error $p_{\rm err}$ in such a scenario equals zero. This makes it harder for Bob to identify a cheating Alice. This is because whenever he finds a discrepancy between $X_\mathcal{I}$ and $X^n$, he cannot be sure if it was due to a bit flip error or a malicious Alice.\
We now proceed to prove that the protocol is secure, except for a minimal probability $\epsilon$ when $p_{\rm err}$ is sufficiently small. The proof is done in three steps as shown:
[: ]{}
If both parties are honest, Bob always accepts the commitment except with minimal probability.
For any dishonest Alice, Bob detects her attempt to cheat and rejects the commitment except with minimal probability.
For any dishonest Bob, he does not obtain information about the committed bit except with minimal probability.
#### Correctness
We shall first prove the correctness of the protocol, namely under the situation that Alice and Bob are both honest, Bob always accepts the protocol except with some minimal probability $\epsilon$.
\[correctness\] If Alice and Bob are both honest, then the protocol is $2\epsilon$-correct.
There are two steps in this proof. Firstly, we show that Bob receives at least $m$ bits from WSEE except with probability $\epsilon$. Secondly, we prove that the number of erroneous bits Bob picks up is close to the expected value $p_{\rm err} m$, except for probability $\epsilon$. The total probability of error for either events occurring is then $2\epsilon$ since failure occurs in either case.\
Since each bit $X_i$ from Alice is obtained by Bob with probability $\frac{1}{2}$, by applying the Hoeffding’s inequality to the random variable $Y=|\mathcal{I}|$, i.e. the length of substring Bob obtains from WSEE, we see that $$\Pr [ Y \leq (\frac{1}{2}-\alpha_1)n] \leq e^{-2\alpha_{1}^{2}n} \leq \epsilon,$$ where n is the length of string $X^n$ Alice has. Using $\alpha_1$ defined in allows Bob to get at least $m$ bits except for probability $\epsilon$. Note that the probability $1/2$ is determined by Bob’s random choice of basis, and is independent of $p_{\rm err}$.\
We then proceed to show that the number of erroneous bits Bob expects to obtain lie within the interval $[(p_{\rm err}-\alpha_2)m,(p_{\rm err}+\alpha_2)m]$ except for some probability $\epsilon$. Note that since previously we have shown that Bob obtains at least $m$ bits, we can safely fix the length of Bob’s substring to be $m$.
Again by applying the Hoeffding’s inequality, with the random variable of interest Z to be the number of erroneous bits Bob obtains, $$\Pr [ |Z-p_{\rm err}m| \geq \alpha_2m] \leq 2e^{-2\alpha_{2}^{2}m} \leq \epsilon$$ again by using $\alpha_2$ as defined in . Hence the correctness of the protocol is guaranteed except for probability $2\epsilon$.
#### Security against dishonest Alice
We now proceed to prove security against Alice. Recall that a malicious Alice can avoid bit flip errors and tamper with the bit string directly. We need to show that no matter how Alice tampers with the string, Bob will detect her cheating with probability close to unity.\
Previously [@noisy:new] for $p_{\rm err} = 0$ , whenever Bob checks $X_{\mathcal{I}}$ against $X^n$ and finds one faulty bit, he aborts the protocol directly. However, in a realistic setup Bob accepts a number of roughly $p_{\rm err} m$ bits. By properties of the error-correcting code, we know from [@noisy:new] that for any attack of Alice she has to change at least $\frac{d}{2}$ such that the Bob will accept the syndrome to be consistent [@noisy:new]. With this, we set a constraint on the code distance used such that whenever Alice attempts to cheat, Bob picks up enough faulty bits to detect the cheating except for some minimal probability.
\[secA\] If Bob is honest, given that Alice and Bob use an error-correcting code with minimum distance $d > \frac{2(p_{\rm err}+\alpha_2)(\frac{1}{2}-\alpha_1)n}{\frac{1}{2}-\alpha_3}$, the pair of protocols(Commit, Open) is $2\epsilon$-binding.
In our proof we assume that a malicious Alice can avoid all bit flip errors in WSEE, so that the scenario reduces to WSE where $p_{\rm err}=0$. From the proof of correctness, we know that Bob obtains enough bits from WSEE except with probability $\epsilon$.\
For any general attack Alice can attempt, to satisfy Bob’s check on the syndrome she has to change at least $\frac{d}{2}$ bits in the original string $X_{n}$, where $d$ is the code distance [@noisy:new]. Since Bob picks up each faulty bit with probability $\frac{1}{2}$, by defining W to be the number of faulty bits where Bob obtains in his substring, and applying Hoeffding’s inequality, $$\Pr \left[ W \leq (\frac{1}{2}-\alpha_3)\frac{d}{2} \right] \leq e^{-\alpha_{3}^{2}d} = \epsilon.$$ We see that Bob picks up at least $(\frac{1}{2}-\alpha_3)\frac{d}{2}$ flipped bits except with probability $\epsilon$, for $\alpha_3$ as defined in . Combining with the fact that Bob accepts at most $(p_{\rm err}+\alpha_2)m$ bits, we require $$(\frac{1}{2}-\alpha_3)\frac{d}{2} > (p_{\rm err}+\alpha_2)m.$$ The requirement for code distance is then given by $$d > \frac{2(p_{\rm err}+\alpha_2)(\frac{1}{2}-\alpha_1)n}{\frac{1}{2}-\alpha_3}.$$ Hence generally when Alice and Bob use a code with minimum distance that satisfies the above requirement, whenever Alice attempts to cheat, the pair of protocols is proven to be $2\epsilon$-binding.
#### Security against dishonest Bob
Subsequently, we prove security against Bob. Recall that a cheating Bob can first make arbitrary measurements and store some classical information, and then keep some quantum information in this noisy-storage device. The overall state of Bob’s system can then be described as a ccq-state $\rho_{X^n K \Theta {\mathcal{F}}({\mathcal{Q}})}$, where $K$ being Bob’s classical information obtained from measurements, $\Theta$ being Alice’s basis information, and ${\mathcal{F}}({\mathcal{Q}})$ being Bob’s quantum information stored in an imperfect quantum memory.\
To quantify the $\epsilon$-smooth min-entropy of Bob’s information about Alice’s string $X^n$, we proceed as in [@noisy:new]: We first bound Bob’s ignorance about $X^n$ based on his classical information $K$ alone. Second, we then relate this bound to his ignorance about $X^n$ given $K$ *and* ${\mathcal{F}}({\mathcal{Q}})$. This yields security statements in terms of the classical capacity of ${\mathcal{F}}$. Note that it is known that for very many channels better security bounds are possible in terms of the entanglement cost [@entCost] and the quantum capacity [@qcextract]. However, the classical capacity is still much better understood and offers explicit parameters for many interesting channels. In contrast, e.g. the quantum capacity of the depolarizing channel is not known.
To bound Bob’s ignorance given $K$ alone, we invoke our results of [@finiten] as stated in the following theorem:
\[smooth\] If Alice is honest, the $\epsilon$-smooth min-entropy of Bob’s information about $X^n$ is $${{\rm H}^{\epsilon}_{\rm min}}(X^n|\Theta^n K) \geq g(s)n +\frac{2\log\epsilon-1}{s},\\$$ where $$g(s) = \frac{-1}{s} \left[ \log(1+2^s)-(1+s)\right]$$ for any $0<s\leq 1$.
Theorem \[smooth\] gives us a bound on the min-entropy rate of Bob’s classical information of $X^n$, whenever Alice is honest. We then use Lemma 2.2 from [@noisy:new] to bound the min-entropy rate of $X^n$ when he has the ccq-state $\rho_{X^n {\mathcal{K}}\Theta{\mathcal{F}}({\mathcal{Q}})}$.
\[qsideinfo\] Consider an arbitrary ccq-state $\rho_{XTQ}$, and let $\epsilon,\epsilon' \geq 0$ be arbitrary. Let ${\mathcal{F}}: {\mathcal{B}}({\mathcal{H}}_{\mathcal{Q}})\rightarrow B({\mathcal{H}}_{\mathcal{Q}})$ be an arbitrary CPTPM representing a quantum channel. Then $$\mathrm{H}_{\mathrm{min}}^{\epsilon+\epsilon'}(X^n|T{\mathcal{F}}({\mathcal{Q}})) \geq -\log P_{\rm succ}^{\mathcal{F}}\left(\lfloor {{\rm H}^{\epsilon}_{\rm min}}(X^n|T) - \log \frac{1}{\epsilon'} \rfloor\right),$$ with $$\begin{aligned}
P_{\rm succ}^{\mathcal{F}}(Rn) := \max_{\{\rho_y\}_y,\{D_y\}_y} \frac{1}{2^{nR}} \sum_{y\in \01^{nR}} {\mathop{\mathrm{tr}}\nolimits}(D_y {\mathcal{F}}(\rho_y))\ ,\end{aligned}$$ where the maximum is taken over all encodings $\{\rho_y\}$ and decoding POVMs $\{D_y\}_y$ of classical symbols $y$.
Note that this bound is dependent on the properties of the quantum storage device ${\mathcal{F}}$. For example, if the memory were to be of arbitrarily large size and noiseless, the success probability $P_{\rm succ}^{\mathcal{F}}(Rn)$ is always 1, and the min-entropy of Bob’s information about $X^n$ would be simply zero. However, the assumption of noisy and bounded storage comes in here to give a sufficiently high min-entropy which is crucial for the security proof. For simplicity in further proofs, we also introduce a simpler version, considering only bounded storage, which is a simple consequence of the chain rule and monotonicity of the min-entropy [@renato:diss; @serge:bounded].
\[boundedqmstor\] Assuming the min-entropy of Bob’s information of $X^n$ be $H_\infty^{\epsilon} (X^n|T)$, and Bob has a perfect quantum memory ${\mathcal{Q}}$ that can store $S$ qubits. $${{\rm H}^{\epsilon}_{\rm min}}(X^n|T{\mathcal{Q}}) \geq {{\rm H}^{\epsilon}_{\rm min}}(X^n|T) - S .$$
With the above Theorem \[smooth\] and Lemma \[qsideinfo\], we can prove security against Bob in two steps. First, we show that if Alice is honest, the min-entropy rate of Bob’s information about the string $X^n$ is lower bounded by $\lambda n$ for some $\lambda$. Then, by using privacy amplification, we show that the cq-state of $C^\mathit{l}$ and Bob’s information is $2\epsilon$-close to a product state, with $C^\mathit{l}$ having uniform distribution over $\lbrace0,1\rbrace^\mathit{l}$. As we are only interested in commiting a single bit in this experiment, we will restrict our statement to the case of $l=1$. A more general statement can be derived analogously.
\[secB\] For a fixed parameter $\epsilon$, define $\lambda$ to be the $\frac{\epsilon}{2}$-smooth min-entropy rate of Bob’s information about $X^n$. If Alice is honest, and if the code rate satisfies $$\label{condonR}
R\geq 1-\lambda+\frac{2\log\frac{1}{\epsilon}}{n},$$ then the pair of protocols (Commit, Open) are $2\epsilon$-hiding for $l=1$, i.e. the commitment of a single bit.
After executing WSEE, by using Theorem \[smooth\] and Theorem \[boundedqmstor\], Bob’s $\frac{\epsilon}{2}$-smooth min-entropy about $X^n$ can be evaluated. Note that by sending the syndrome he obtains additional information about $X^n$, and this is accounted for by the chain rule and monotonicity property of min-entropy [@renato:diss], $$H_{\mathrm{min}}^{\epsilon_o} (X^n | {\rm B'}, {\mathrm{Syn}}(X^n)) \geq H_{\mathrm{min}}^{\epsilon_o} (X^n| {\rm B'}) -\mathit{L}.$$ where $\mathit{L}$ is the length of the syndrome. Recall that the length of syndrome is $L=n-k=(1-R)n$, where $R=\frac{k}{n}$ is the code rate. Hence, we have $$H_{\mathrm{min}}^{\epsilon/2} (X^n | {\rm B'}, {\mathrm{Syn}}(X^n)) \geq (\lambda - 1 + R)n$$ which denotes the min-entropy rate of Bob’s total information about $X^n$ at the end of the commit phase.\
Next, we show that by privacy amplification Bob does not gain knowledge about the committed information $C^\mathit{l}$. Denoting the committed string as $ C^\mathit{l}={\mbox{Ext}}(X^n,R)\in{\{0,1\}}^\mathit{l}$ having length $\mathit{l}$ we have from [@renato:diss] that $$\label{priamp}
\rho_{C^\mathit{l},{\rm B'}{\mathrm{Syn}}(X^n)} \approx_{\epsilon '} \tau_{{\{0,1\}}^\mathit{l}} \otimes \rho_{{\rm B'},{\mathrm{Syn}}(X^n)}$$ where $$\label{priamperror}
\epsilon ' = 2\epsilon_o + 2^{-\frac{1}{2}[{{\rm H}}_{\mathrm{min}}^{\epsilon_o} (X^n|{\rm B'},{\mathrm{Syn}}(X^n))-\mathit{l}]-1}$$ and $\tau_{\mathcal{A}}$ is the uniform distribution over the entire set ${\mathcal{A}}$. Setting $\epsilon_o=\frac{\epsilon}{2}$, $$\label{priamp2}
\epsilon ' = \epsilon + \frac{1}{2}\cdot 2^{-\frac{1}{2}[{{\rm H}}_{\mathrm{min}}^{\epsilon_o} (X^n|{\rm B'},{\mathrm{Syn}}(X^n))-\mathit{l}]}$$
Setting the second term in to be $\epsilon$, and setting $\mathit{l}=1$, this implies $$\lambda -1+R -\frac{1}{n}> -\frac{2\log\epsilon+2}{n},$$
Rearranging gives . In the large n limit, we require $R>1-\lambda$.
With this, we end the security proof against Bob.\
In summary, we have derived conditions on the relative minimum distance $\delta=\frac{d}{n}$ and the code rate R for where the protocol is secure. By combining Lemma \[correctness\], Lemma \[secA\] and Lemma \[secB\], we summarize these results into the following theorem:
\[summary\] Let $n \in \mathbb{N}$, $\epsilon > 0$ and $\lambda > 0$. If the error correcting code used satisfies the following requirements:
[\*]{}
$\delta > \frac{2(p_{\rm err}+\alpha_2)(\frac{1}{2}-\alpha_1)}{\frac{1}{2}-\alpha_3}$.
$R > 1-\lambda + \frac{2\log\frac{1}{\epsilon}}{n}$,
then Protocol 2 is $2\epsilon$-correct, $2\epsilon$-binding and $2\epsilon$-hiding.
The final part of the theoretical analysis is to discuss the feasibility of finding an error-correcting code that satisfies the requirements stated in Theorem \[summary\]. Clearly, there exists a trade-off between parameters of code rate R and relative minimum distance $\delta$. We make use of Theorem \[randomcode\] to argue that once the parity check matrix of a code with rate R is randomly generated, its distance is lower bounded except for an extremely small probability. Subsequently, by using Theorem \[summary\] and Theorem \[randomcode\], we evaluate an optimal parameter $n=2.5\times 10^5$ in the [Supplementary Methods]{} \[ourexp\], which is used in our experiment, and show that for such a block length, the bit commitment protocol is secure except for an error $3\cdot 10^{-4}$.
Also, by combining Theorem \[summary\] with Theorem \[randomcode\], we provide a cleaner expression for the bound on minimum distance accompanied by a lower bound on the block length such that security can be achieved:
\[security\] Let $\epsilon > 0$, $\lambda \geq 0.3$, $\beta\in\left(0,0.01\right]$, $\delta \in [0.05,0.11]$, and $$\label{boundm3}
n\geq\frac{1}{\beta^2}\log\frac{2}{\epsilon}\ .$$ Also, denote $h(x)=-x\log x-(1-x)\log(1-x)$ as the binary entropy function.
If the bit flip error rate satisfies $$\label{condperr}
p_{\rm err} < (1-4\sqrt{5}\beta)\cdot\frac{\delta}{2} -\frac{\beta}{\sqrt{1-2\beta}} ,$$ and the smooth min-entropy rate satisfies $$\label{condlambda}
\lambda > h(\delta)+3\beta^2,$$ then the bit commitment protocol is $3\epsilon$-secure by using a randomly generated error-correcting code.
Fix $\epsilon$ and assume that $$\sqrt{\frac{\ln\frac{2}{\epsilon}}{n}}\leq \sqrt{\frac{\log\frac{2}{\epsilon}}{n}}\leq \beta.$$ This is achieved for any $\beta$, provided is true. Hence from , we have $$\label{alpha1}
0\leq \alpha_1 \leq \beta.$$ Plugging this into $m=({\frac{1}{2}}-\alpha_1)n$ gives $$\label{alpha2}
\alpha_2 = \sqrt{\frac{\ln\frac{2}{\epsilon}}{2m}} \leq \frac{\beta}{\sqrt{1-2\beta}}.$$ Meanwhile, assume that $\delta\geq 0.05$. This leads to the condition that $\lambda\geq 0.3$ which is generally achieved. $$\label{alpha3}
\alpha_3 \leq \sqrt{20}\beta.$$ Plugging , , and into the condition on minimum distance given in Theorem \[summary\], we obtain $$\label{condondelta}
\delta > 2\cdot\frac{p_{\rm err}+\frac{\beta}{\sqrt{1-2\beta}}}{1-4\sqrt{5}\beta} \geq \frac{2(p_{\rm err}+\alpha_2)(\frac{1}{2}-\alpha_1)}{\frac{1}{2}-\alpha_3}.$$ Rearranging, we obtain .
To derive , we first use Theorem \[randomcode\]. By setting the additional error from the generation of random codes to be smaller than $\epsilon$, we obtain $$R<1-h(\delta)-\frac{\log\frac{1}{\epsilon}}{n}.$$ Combining this with the condition on code rate R given in Theorem \[summary\], we have $$\begin{aligned}
1-h(\delta)&>& 1-\lambda+2\frac{\log\frac{1}{\epsilon}}{n}\nonumber\\
\lambda &>& h(\delta) + 3\frac{\log\frac{1}{\epsilon}}{n},\end{aligned}$$ which is satisfied if is true. The total execution error of Protocol 2 becomes $3\epsilon$, where $2\epsilon$ comes from the execution error in Theorem \[summary\], and additional $\epsilon$ accounts for the probability that a randomly generated error-correcting code does not fulfill the requirement on relative minimum distance $\delta$.
provides us with a lower bound on $n$ for a secure implementation. Note that this lower bound is non-tight, due to the approximations made while deriving bounds for , , and . Also, it is stressed that Theorem \[randomcode\] gives a general proof for a randomly generated error-correcting code. This approach is taken because systematic ways of constructing such binary linear codes are not known, and the task of evaluating the minimum distance of a given code is NP-hard. However, as outlined in the [Supplementary Discussion]{} \[section:codes\] it is well known that the probability of generating a code with undesirable properties is minimal and added as a source of error in the protocol. It is also worth noting that a random code allows for easy execution of the protocol, since the only computation involved for honest parties is the calculation of the syndrome. That is, Alice and Bob never need to decode.
Range of experimental parameters for implementation {#section:exp}
---------------------------------------------------
In this section, we provide full statements about the security of commitments, by combining the analysis accounting for both erasures (WSEE) and errors (BC). We work towards a simplified expression for the rate of commitment, i.e. to commit one bit securely, what is the required number of signals to send. We present region plots showing where security holds for the protocol.
For security, we first note that the analysis of bit flip errors requires a minimum guaranteed amount of min-entropy $\hat{\lambda}$ for the commitment to be secure. This is seen in Lemma \[secB\] where the lower bound for code rate R can be translated into a lower bound for the min-entropy rate. The main condition for feasibility of bit commitment is then given by $$\lambda \geq \hat{\lambda}.$$
The following theorem shows $\hat{\lambda}$ by considering a randomly generated binary linear code.
By fixing an error parameter $\epsilon$ which indicates the error for the generation of a random code. If and given the parameter $\delta$, which is the relative minimum distance of the code required, as determined by $p_{\rm err}$ by using , $$\hat\lambda = h(\delta)+\frac{3\log\frac{1}{\epsilon}}{n}.$$
On the other hand, we need to evaluate a bound on the min-entropy rate $\lambda$ created according to the parameters $p_{\rm sent}^1$, $p_{\rm B, no
click}^{\rm h}$ and $p_{\rm B, no click}^{\rm d}$. We begin by defining two relevant fractional quantities: $$\begin{aligned}
m_{\rm left}^1&=p^1_{\rm sent} - p_{\rm B, no click}^{\rm h} + p_{\rm B, no click}^{\rm d} - 3\zeta\\
m_{\rm frac}&=1- p_{\rm B, no click}^{\rm h}-\zeta\end{aligned}$$ where $\zeta = \sqrt{\frac{\ln\frac{2}{\epsilon}}{2M}}$. The smooth min-entropy rate can then be evaluated, by invoking the bounded, or in general noisy storage assumption. We refer to below for some examples.
Example: bounded storage {#example-bounded-storage .unnumbered}
------------------------
We state a theorem describing the number of signals M needed to send for a secure commitment, given the relevant experimental parameters and assuming the case of bounded storage. It is worth stressing again that $n$ denotes the **block length used in the commitment**, while $M$ denotes the **number of signals sent from Alice to Bob**. These quantities are related by the expressions given in .
\[d2\] Let a dishonest Bob’s storage size be bounded by S. For fixed parameters S and $\epsilon$, and given the experimental probabilities listed in Table I, and for some $\beta,\gamma \in (0,0.01]$, let $$\begin{aligned}
m_2 &=& p_{\rm sent}^1-p_{\rm B, no click}^{\rm h}+p_{\rm B, no click}^{\rm d} -3\gamma\nonumber\\
m_3 &=& 1-p_{\rm B, no click}^{\rm h}\nonumber\\
L'&=&\max_{ s \in (0,1] } ~ \frac{-1}{s} \left[\log(1+2^s)-1-s\right] -\frac{3\epsilon}{s}\nonumber\\
\hat{\lambda} &=& h(\delta) + 3\beta^2\nonumber\\
\delta &=& 2\cdot\frac{p_{\rm err}+\frac{\beta}{\sqrt{1-2\beta}}}{1-4\sqrt{5}\beta}\nonumber\\
M_1 &=& \frac{1}{2\gamma^2}\log\frac{2}{\epsilon}\nonumber\\
M_2 &=& \frac{\log\frac{1}{\epsilon}}{\epsilon\cdot m_2}\nonumber\\
M_3 &=& \frac{\log\frac{2}{\epsilon}}{(m_3-\gamma)\beta^2}\nonumber\\
M_4 &=& \frac{S}{m_2 L'-m_3\hat{\lambda}}\end{aligned}$$ For security to hold at all, the following is required: $$\label{condition}
m_2 L' -m_3 \hat{\lambda} > 0.$$ If is true, then bit commitment can be implemented $3\epsilon$-securely by using a randomly constructed error-correcting code, whenever $$\label{boundingm}
M > \max ~ \{M_1,M_2,M_3,M_4\}.$$
By the analysis of [@Curty10], the min-entropy rate has the form $$\lambda = \frac{m_{\rm left}^1\cdot L - \frac{S}{M}}{m_{\rm frac}},$$ where $$L = \max_{ s \in (0,1] } ~ \frac{-1}{s} \left[\log(1+2^s)-1-s\right] -\frac{3\log\frac{1}{\epsilon}}{m_{\rm left}^1\cdot M}.$$ Note that $\zeta$ is dependent on $M$, and its value decreases while $M$ increases. Setting $\zeta \leq \gamma$ which is a constant, provides an lower bound for $M$, which gives the value for $M_1$ depending on the chosen $\gamma$.\
Similarly, setting $\frac{\log\frac{2}{\epsilon}}{m_{\rm left}^1\cdot M}\leq\epsilon$ provides $M_2$ and $L\geq L'$.\
$M_3$ comes from the condition given for $n$ at , while $M=\frac{n}{m_{\rm frac}} \geq \frac{n}{m_3-\gamma}$. Lastly, $$\lambda\geq \frac{m_2\cdot L' - \frac{S}{M}}{m_3} \geq \hat{\lambda}$$ provides the main condition for security to hold at all (S=0), while rearranging gives the value for $M_4$.
Example: noisy storage {#example-noisy-storage .unnumbered}
----------------------
Lemma \[d2\] gives the case for bounded storage model, with no quantum noise assumed for cheating Bob’s storage device. For a more general noisy storage assumption, where the quantum channel satisfies a strong converse relation as in [@noisy:new], the $\frac{\epsilon}{2}$-smooth min-entropy can be evaluated by $$\begin{aligned}
\label{depolarizingstorage}
& H_{\mathrm{min}}^{\epsilon'+\epsilon''}(X^n|\Theta^n K {\mathcal{F}}({\mathcal{Q}})) \nonumber\\
&\geq -\log P_{\rm succ}^{{\mathcal{N}}^{\otimes S}} \left[H_{\mathrm{min}}^{\epsilon'}(X^n|\Theta^n K)-\log\frac{1}{\epsilon''}\right]\nonumber\\
&= S\cdot \gamma^{{\mathcal{N}}} \left(\frac{H_{\mathrm{min}}^{\epsilon'}(X^n|\Theta^n K)-\log\frac{1}{\epsilon''}}{S}\right),\end{aligned}$$ where $\gamma^{{\mathcal{N}}}$ is the strong converse parameter of the quantum channel. For a fixed error parameter $\epsilon$, $\epsilon'$ and $\epsilon''$ should be chosen such that $\epsilon'+\epsilon'' = \frac{\epsilon}{2}$. Compared to the analysis in [@Curty10], where the storage size is determined by introducing a *storage rate* quantity $\nu$, in this analysis we work with the quantity $S$, which is the maximum number of qubits Bob is able to store. To use this quantity in the analysis, we invoke the bounded storage assumption that Bob cannot store more than $S$ qubits, then calculate the conditions for security. For example, in the setting of depolarizing noise for a two-dimensional quantum channel, the strong converse parameter is given by the following expression [@noisy:new]: $$\begin{aligned}
\gamma^{{\mathcal{N}}} (\hat{R}) &=& \max_{\alpha\geq 1} \frac{\alpha-1}{\alpha} (\hat{R}- C) \nonumber\\
C &=& 1 - \frac{1}{1-\alpha}\log\left[ p^\alpha+(1-p)^\alpha\right] \nonumber\\
p &=& \frac{1+r}{2}.\end{aligned}$$
Our experiment {#ourexp .unnumbered}
--------------
We state again the bounds of experimental parameters as derived: $$\begin{aligned}
p_{\rm sent}^1&=&1-p_{\rm sent}^0-p_{\rm sent}^{n>1}>0.125\nonumber\\
p_{\rm sent}^0+p_{\rm sent}^1&=&1-p_{\rm sent}^{n>1}>0.99947\nonumber\\
p_{\rm B, no click}^{\rm h}&=&0.909,\nonumber\\
p_{\rm err} &=& 0.0412.\end{aligned}$$ Note that $p_{\rm B, no click}^{\rm h}$ and $p_{\rm err}$ are values obtained after the symmetrization procedures on both Alice’s and Bob’s side.
Bounds on $n$ (and $M$) derived based on Theorem \[security\] and Lemma \[d2\] are non-tight. Here, we use an optimal block length such that the classical information post-processing is minimal. We summarize the calculations in the following steps:
[: ]{}
[**Fix $\epsilon$ and $n$:**]{} Firstly, we set $\epsilon=0.99\times 10^{-5}$, and $n=2.5\times 10^5$. By doing so, all relevant parameters in can be evaluated, except for $\alpha_3$ which will depend on the error-correcting code. $M$ is known by its relation with $n$ as stated in , where this is justified by a detailed explanation offered right after Protocol 1.
[**Evaluate relative minimum distance $\delta$:**]{} By performing a numerical optimization that satisfies the condition on relative distance, as stated in Theorem \[summary\], we obtain $\delta > 0.998201$.
[**Set $\epsilon_{code}$:**]{} To obtain a code that with $\delta$ that satisfies the condition as evaluated in step 2, we use Theorem \[randomcode\]. First, we need to set an $\epsilon_{code}=2\times 10^{-7}$, which bounds the probabilistic error for generating a bad random code. By doing so, we pose an upper bound upon the code rate $R$. Using $R=0.531$ satisfies this condition. By using Theorem \[summary\], the protocol is secure provided that the $\frac{\epsilon}{2}$-smooth min entropy rate $\lambda > 0.469133$.
[**Evaluate storage assumption:**]{} By using $p_{\rm B, no click}^{\rm h}$, $p_{\rm B, no click}^{\rm h}$, $p_{\rm sent}^1$, $M$ and $\epsilon$, evaluate and optimize $\lambda$ for different storage noise and storage sizes $S$, such that $\lambda$ satisfies the condition in Step 3.
[**Evaluate total execution error:**]{} The total execution error is then evaluated as $\epsilon_{total}=2\epsilon+\epsilon_{code} = 2\times 10^{-5}$.
For the bounded storage assumption, the commitment is secure whenever dishonest Bob’s storage size is bounded by $S_{bounded}=928$ qubits. For the noisy storage assumption, we can use and maximize over all choices of $\epsilon'$ and $\epsilon''$. For a depolarizing noise of noise parameter $r=0.9$, the commitment is secure whenever Bob’s storage size is bounded by $S_{noisy} = 972$ qubits.
[10]{}
J. Kilian, “Founding cryptography on oblivious transfer,” in [*Proceedings of 20th ACM STOC*]{}, pp. 20–31, 1988.
D. Mayers, “Unconditionally secure quantum bit commitment is impossible,” [*Phys. Rev. Letters*]{}, vol. 78, pp. 3414–3417, 1997.
H. Chau and H.-K. Lo, “Making an empty promise with a quantum computer,” [ *Fortschritte der Physik*]{}, vol. 46, pp. 507–520, 1998.
H.-K. Lo, “Insecurity of quantum secure computations,” [*Phys. Rev. A*]{}, vol. 56, p. 1154, 1997.
H.-K. Lo and H. F. Chau, “Is quantum bit commitment really possible?,” [ *Phys. Rev. Letters*]{}, vol. 78, p. 3410, 1997.
G. D’Ariano, D. Kretschmann, D. Schlingemann, and R. Werner, “Quantum bit commitment revisited: the possible and the impossible,” [*Physical Review A*]{}, vol. 76, p. 032328, 2007.
U. Maurer, “Conditionally-perfect secrecy and a provably-secure randomized cipher,” [*Journal of Cryptology*]{}, vol. 5, pp. 53–66, 1992.
C. Cachin and U. M. Maurer, “Unconditional security against memory-bounded adversaries,” in [*Proceedings of CRYPTO 1997*]{}, LNCS, pp. 292–306, 1997.
S. Dziembowski and U. Maurer, “On generating the initial key in the bounded-storage model,” in [*Proceedings of EUROCRYPT*]{}, LNCS, pp. 126–137, 2004.
C. H. Bennett and G. Brassard, “Quantum cryptography: Public key distribution and coin tossing,” [*Proceedings of IEEE International Conference on Computers, Systems and Signal Processing*]{}, pp. 175–179, 1984.
I. B. Damg[å]{}rd, S. Fehr, R. Renner, L. Salvail, and C. Schaffner, “A tight high-order entropic quantum uncertainty relation with applications,” in [ *Proceedings of CRYPTO 2007*]{}, LNCS, pp. 360–378, 2007.
I. B. Damg[å]{}rd, S. Fehr, L. Salvail, and C. Schaffner, “Cryptography in the [B]{}ounded-[Q]{}uantum-[S]{}torage [M]{}odel,” in [*Proc. IEEE FOCS*]{}, pp. 449–458, 2005.
I. B. Damg[å]{}rd, S. Fehr, L. Salvail, and C. Schaffner, “Secure identification and [QKD]{} in the bounded-quantum-storage model,” in [ *Proceedings of CRYPTO 2007*]{}, LNCS, pp. 342–359, 2007.
C. G.-G. N. J. Bouman, S. Fehr and C. Schaffner, “An all-but-one entropic uncertainty relations, and application to password-based identification,” 2011. arXiv:1105.6212v1.
S. Wehner, C. Schaffner, and B. Terhal, “Cryptography from noisy storage,” [*Physical Review Letters*]{}, vol. 100, p. 220502, 2008.
C. Schaffner, B. Terhal, and S. Wehner, “Robust cryptography in the noisy-quantum-storage model,” [*Quantum Information & Computation*]{}, vol. 9, p. 11, 2008. arXiv:0807.1333v3.
R. König, S. Wehner, and J. Wullschleger, “Unconditional security from noisy quantum storage,” 2009. arXiv:0906.1030v4.
A. I. Lvovsky, B. C. Sanders, and W. Tittel, “Optical quantum memory,” [ *Nature Photonics*]{}, vol. 3, pp. 706–714, 2009.
I. Usmani, M. Afzelius, H. de Riedmatten, and N. Gisin, “Mapping multiple photonic qubits into and out of one solid-state atomic ensemble,” [ *Nature Communications*]{}, vol. 1, p. 12, 2010.
M. Bonarota, J.-L. L. Gouet, and T. Chaneliere, “Highly multimode storage in a crystal,” [*New Journal of Physics*]{},2
H.-N. Dai, H. Zhang, S.-J. Yang, T.-M. Zhao, J. Rui, Y.-J. Deng, L. Li, N.-L. Liu, S. Chen, X.-H. Bao, X.-M. Jin, B. Zhao, and J.-W. Pan, “Holographic storage of biphoton entanglement,” [*Phys. Rev. Lett.*]{}, vol. 108, p. 210501, May 2012.
M. Berta, F. Brandao, M. Christandl, and S. Wehner, “Entanglement cost of quantum channels.” 2012.
M. Berta, O. Fawzi, and S. Wehner, “Quantum to classical randomness extractors.” CRYPTO ’12, 2012.
W. Uhlmann, “Probability inequalities for sums of bounded random variables,” [*Journal of the American Statistical Association*]{}, vol. 58(301), 1963.
S. Wehner, M. Curty, C. Schaffner, and H.-K. Lo, “Implementation of two-party protocols in the noisy-storage model,” [*Physical Review A*]{}, vol. 81, p. 052336, 2010.
R. König, R. Renner, and C. Schaffner, “The operational meaning of min- and max-entropy,” [*IEEE Transactions on Information Theory*]{}, vol. 55, pp. 4674–4681, 2009.
R. Renner, “Security of quantum key distribution,” [*Int. J. of Quantum Information*]{}, vol. 6, p. 1, 2008.
C. Kurtsiefer, P. Zarda, M. Halder, P. M. Gorman, P. R. Tapster, J. G. Rarity, and H. Weinfurter, “Long distance free space quantum cryptography,” [ *Proc. SPIE*]{}, vol. 4917, pp. 25–31, 2002.
I. Marcikic, A. Lamas-Linares, and C. Kurtsiefer, “Free-space quantum key distribution with entangled photons,” [*Appl. Phys. Lett.*]{}, vol. 89, p. 101122, 2006.
A. Ling, M. P. Peloso, I. Marcikic, V. Scarani, A. Lamas-Linares, and C. Kurtsiefer, “Experimental quantum key distribution based on a bell test,” [*Phys. Rev. A*]{}, vol. 78, p. 020301, 2008.
P. G. Kwiat, K. Mattle, H. Weinfurter, A. Zeilinger, A. V. Sergienko, and Y. Shih, “New high-intensity source of polarization–entangled photon pairs,” [*Phys. Rev. Lett.*]{}, vol. 75, pp. 4337–4341, 1995.
C. Kurtsiefer, “Qcrypto: an open source code for experimental quantum cryptography.” http://code.google.com/p/qcrypto/, 2008.
A. Nguyen, J. Frison, K. P. Huy, and S. Massar, “Experimental quantum tossing of a single coin,” [*New Journal of Physics*]{}
G. Berlín, G. Brassard, F. Bussières, N. Godbout, J. A. Slater, and W. Tittel, “Experimental loss-tolerant quantum coin flipping,” [*Nature Communications*]{}, vol. 2, p. 561, 2011.
A. C.-C. Yao, “Security of quantum protocols against coherent measurements,” in [*Proceedings of 20th ACM STOC*]{},
V. Makarov and D. R. Hjelme, “Faked states attack on quantum cryptosystems,” [*J. Mod. Opt.*]{}, vol. 52, p. 691, 2005.
I. Gerhardt, Q.Liu, A. Lamas-Linares, J. Skaar, V. Scarani, V. Makarov, and C. Kurtsiefer, ”Experimentally faking the violation of Bell.”
W. C. Huffman and V. Pless, [*Fundamentals of error-correcting codes*]{}. Cambridge University Press, 2003.
R. G. Gallager, [*Low-Density Parity-Check Codes*]{}. PhD thesis, 1963.
R. König, R. Renner, and C. Schaffner, “The operational meaning of min- and max-entropy,” [*IEEE Transactions on Information Theory*]{}, vol. 55, pp. 4674–4681, 2009.
N. Ng, M. Berta, and S. Wehner, “A min-entropy uncertainty relation for finite size cryptography.” arXiv:1205.0842v1, 2012.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We calculate the single-particle Green’s function of electrons that are coupled to acoustic phonons by means of higher dimensional bosonization. This non-perturbative method is [*[not]{}*]{} based on the assumption that the electronic system is a Fermi liquid. For isotropic three-dimensional phonons we find that the long-range part of the Coulomb interaction cannot destabilize the Fermi liquid state, although for strong electron-phonon coupling the quasi-particle residue is small. We also show that Luttinger liquid behavior in three dimensions can be due to quasi-one-dimensional anisotropy in the electronic band structure [*[or in the phonon frequencies]{}*]{}.'
address: |
Institut fűr Theoretische Physik der Universita̋t Gőttingen,\
Bunsenstr.9, D-37073 Gőttingen, Germany
author:
- Peter Kopietz
date: 'July 21, 1995'
title: ' Bosonization of coupled electron-phonon systems'
---
= 10000
A famous theorem due to Migdal[@Migdal58] states that in coupled electron-phonon systems the electron-phonon vertex is for $\sqrt{ \frac{m}{M}} \ll 1$ not renormalized by phonon corrections. Here $m$ is the effective mass of the electrons and $M$ is the mass of the ions. An implicit assumption in the proof of this theorem is that the electronic system is a Fermi liquid[@Fetter71]. However, there exists experimental evidence that the normal-state properties of the high temperature superconductors do not show typical Fermi liquid behavior[@Anderson90]. Therefore it is desirable to study electron-phonon interactions by means of non-perturbative techniques, which do not assume [*[a priori]{}*]{} that the system is a Fermi liquid. Moreover, in heavy fermion systems the parameter $\sqrt{ \frac{m}{M}} $ is not necessarily small, so that Migdal’s theorem may not be valid. In this case the self-consistent renormalization of the phonon energies due to the coupling to the electrons cannot be neglected[@Bardeen55; @Pines89]. In diagrammatic approaches it is often tacitly assumed that the phonons remain well defined collective modes[@Fetter71; @Engelsberg63].
In this work we shall study electron-phonon interactions by means of our functional integral formulation of higher dimensional bosonization[@Kopietz94; @Kopietz95], and show that with this non-perturbative approach one can circumvent the shortcomings mentioned above. Bosonization in arbitrary dimensions has recently been discussed by a number of authors[@Kopietz94; @Kopietz95; @Luther79; @Haldane92; @Houghton93; @Castro94; @Frohlich95]. As compared with more conventional operator methods[@Houghton93; @Castro94], the functional bosonization approach[@Kopietz94; @Kopietz95; @Frohlich95] has the advantage that the retarded interaction between the electrons which is mediated via the phonons can be obtained trivially by integrating over the phonon field at the very beginning of the calculation. Bosonization is most powerful for long-range interactions that are dominated by momentum transfers $| {\bf{q}} | \ll k_{F}$, where $k_{F}$ is the Fermi momentum. In this case the bosonization method developed in Ref.[@Kopietz94] leads to cutoff-independent results. The physically most relevant Coulomb interaction belongs to this category at high densities, where the Thomas-Fermi screening wave-vector $\kappa = ( 4 \pi e^2 \nu )^{\frac{1}{2}}$ is small compared with $k_{F}$. Here $\nu$ is the density of states at the Fermi energy. In this work we shall therefore start from a model of electrons interacting with Coulomb forces. For $ | {\bf{q}} | \ll k_{F}$ the Fourier transform of the interaction can be approximated by its continuum limit $f_{\bf{q}}^{cb} = \frac{ 4 \pi e^{2}}{ {\bf{q}} ^2 }$. The screening problem will be solved explicitly by means of our bosonization approach. In this respect our work is complementary to studies based on Hubbard models[@Kim89], which [*[assume]{}*]{} that the summation of all diagrams describing screening processes effectively leads to the replacement of the long-range Coulomb interaction by a short-range Hubbard-$U$.
Following the classic textbook by Fetter and Walecka[@Fetter71], we use the Debye model to describe the interaction between electrons and longitudinal acoustic (LA) phonons, and approximate the ionic background charge by a homogeneous elastic medium. Although the ions in real solids form a lattice, the discrete lattice structure is unimportant for LA phonons with wave-vectors $| {\bf{q}}| \ll k_{F}$. For a detailed description of this model and its physical justification see Fetter and Walecka[@Fetter71]. However, some subtleties concerning screening and phonon energy renormalization have been ignored in Ref.[@Fetter71]. To clarify these points, we first give a careful derivation of the effective electron-electron interaction in this model via functional integration.
The dynamics of the isolated phonon system is described via the action $$S_{ph} \{ b \} = \beta \sum_{q} [ - i \omega_{m} + \omega_{\bf{q}} ]
b^{\dagger}_{q} b_{q}
\; \; \; ,$$ where $b_{q}$ is a complex field representing the phonons. Here $\beta$ is the inverse temperature, $\omega_{m} = 2 \pi m /
\beta$, and $q = [ {\bf{q}} , i \omega_{m} ]$ is a collective label. For simplicity we first assume LA phonons with dispersion $\omega_{\bf{q}} = c_{0} | {\bf{q}} |$. Here $c_{0}$ is the [*[bare]{}*]{} phonon velocity, which is determined by the [*[short-range part]{}*]{} of the Coulomb potential and all other non-universal forces between the ions. The strong renormalizations due to [*[long-range]{}*]{} Coulomb forces will be treated explicitly in this work. As usual, the dimensionless measure for the strength of the interaction relative to the kinetic energy is $\nu f_{\bf{q}}^{cb} = \frac{ \kappa^2}{ {\bf{q}}^2}$. Because $\nu f_{\bf{q}}^{cb}
\raisebox{-0.5ex}{$\; \stackrel{>}{\sim} \;$} 1$ for $| {\bf{q}} |
\raisebox{-0.5ex}{$\; \stackrel{<}{\sim} \;$} \kappa $, the Thomas-Fermi wave-vector $\kappa$ defines the boundary between the long– and short-range regimes. Note that we are implicitly assuming that the [*[short-wavelength]{}*]{} part of the problem has already been solved, and that the results (for example the phonon velocity $c_0$) are used as input parameters for the calculation of the effect of the long-wavelength modes. Such a strategy is also adopted in renormalization group approaches to condensed matter systems[@Shankar94], and is different from the classic self-consistent treatment of the electron-phonon system by Bardeen and Pines[@Bardeen55]. Of course, the results for physical quantities should be identical. We shall come back to this point below in our discussion of the phonon energy shift.
The electronic degrees of freedom are represented by a Grassmann-field $\psi$, so that the total action of the interacting electron-phonon system is $$S \{ \psi , b \} = S_{0} \{ \psi \} + S_{ph} \{ b \} + S_{int} \{ \psi , b \}
\label{eq:Stotph}
\; \; \; .$$ Here $$S_{0} \{ \psi \} = \beta \sum_{k} [ - i \tilde{\omega}_{n} + \xi_{\bf{k}} ]
\psi^{\dagger}_{k} \psi_{k}$$ represents non-interacting spinless electrons with energy dispersion $\xi_{\bf{k}} = \epsilon_{\bf{k}} - \mu$, where $\mu$ is the chemical potential. $\tilde{\omega}_{n} = 2 \pi ( n + \frac{1}{2} ) / \beta$ is a fermionic Matsubara frequency, and $k = [ {\bf{k}} , i \tilde{\omega}_{n}
]$. The action $S_{int} \{ \psi , b \}$ represents the Coulomb energy associated with all charge fluctuations in the system[@Fetter71], $$S_{int} \{ \psi , b \} =
S_{int}^{el} \{ \psi \} +
S_{int}^{el-ph} \{ \psi , b \} +
S_{int}^{ph} \{ b \}
\; \; \; ,$$ $$\begin{aligned}
S_{int}^{el} \{ \psi \} & = &
\frac{\beta }{2 V} \sum_{q} f_{\bf{q}}^{cb} \rho_{-q} \rho_{q}
\label{eq:Sintel}
\; \; \; ,
\\
S_{int}^{el-ph} \{ \psi , b \} & = &
- \frac{\beta }{2 V} \sum_{q} f_{\bf{q}}^{cb} \left[ \rho_{-q} \rho_{q}^{ion}
+ \rho_{- q}^{ion}
\rho_{q} \right]
\label{eq:Sintelion}
\; \; \; ,
\\
S_{int}^{ph} \{ b \} & = &
\frac{\beta }{2 V} \sum_{q} f_{\bf{q}}^{cb} \rho^{ion}_{-q} \rho^{ion}_{q}
\label{eq:Sintion}
\; \; \; ,
\end{aligned}$$ where $V$ is the volume of the system, and the Fourier coefficients of the densities are $$\begin{aligned}
\rho_{q} & = & \sum_{k} \psi_{k}^{\dagger} \psi_{k+q}
\; \; \; ,
\\
\rho_{q}^{ion} & = & - z \sqrt{N} \frac{ | {\bf{q}} | }{\sqrt{2 M
\omega_{\bf{q}} }}
[ b_{q} + b^{\dagger}_{-q} ]
\; \; \; .
\end{aligned}$$ Here $z$ is the valence of the ions, $z N$ is the total number of conduction electrons in the system, and it is understood that the ${\bf{q}} = 0$-term in the sums should be omitted due to overall charge neutrality. At this point the following two approximations are made in Ref.[@Fetter71]: (a) the bare Coulomb interaction $f_{\bf{q}}^{cb}$ in $S_{int}^{el-ph} \{ \psi , b \}$ is replaced [*[by hand]{}*]{} by the static screened interaction, $ \frac{ 4 \pi e^2}{ \bf{q}^2 } \rightarrow \frac{ 4 \pi e^2}{ \kappa^2}$, and (b) the contribution $S_{int}^{ph} \{ b \}$ is simply dropped. We shall see shortly that the approximation (b) amounts to ignoring the self-consistent renormalization of the phonon frequencies[@Bardeen55; @Pines89]. Although Fetter and Walecka[@Fetter71] argue that these approximations correctly describe the physics of screening, it is not quite satisfactory that one has to rely here on words and not on calculations. Because in our bosonization method screening can be derived from first principles, we do not follow the “screening by hand” procedure of Ref.[@Fetter71], and retain at this point all terms in Eqs.(\[eq:Sintel\])-(\[eq:Sintion\]) with the bare Coulomb interaction.
We are interested in the electronic Green’s function of the interacting many-body system. The Matsubara Green’s function can be written as a functional integral, $$G ( k )
= - \beta \frac{
\int {\cal{D}} \left\{ \psi \right\}
{\cal{D}} \left\{ b \right\}
e^{- {S} \{ \psi , b \} }
\psi_{k} \psi^{\dagger}_{k}
}
{ \int {\cal{D}} \left\{ \psi \right\}
{\cal{D}} \left\{ b \right\}
e^{- S \{ \psi , b \} } }
\; \; \; .
\label{eq:Gphdef}$$ Evidently the $b$-integration in Eq.(\[eq:Gphdef\]) is Gaussian, so that it can be carried out [*[exactly]{}*]{}. We obtain the following exact expression for the interacting Green’s function $$G ( k )
= - \beta \frac{
\int {\cal{D}} \left\{ \psi \right\}
e^{- {S}_{eff} \{ \psi \} }
\psi_{k} \psi^{\dagger}_{k}
}
{ \int {\cal{D}} \left\{ \psi \right\}
e^{- S_{eff} \{ \psi \} } }
\; \; \; ,
\label{eq:Gph2}$$ with $$\begin{aligned}
S_{eff} \{ \psi \} & = & S_{0} \{ \psi \} +
S_{int}^{el} \{ \psi \}
\nonumber
\\
& - & \beta \sum_{q}
\left[
\frac{ \omega_{\bf{q}} G_{\bf{q}}^2 }{ \omega_{m}^2 + \omega_{\bf{q}}^2 +
\omega_{\bf{q}} W_{\bf{q}} } \right]
\rho_{-q} \rho_{q}
\; \; \; ,
\label{eq:Seffdefph}
\end{aligned}$$ with $$\begin{aligned}
W_{\bf{q}} & = &
\left[ \frac{z^2 N}{V} \frac{ {\bf{q}}^2 }{M } \right]
\frac{ f_{\bf{q}}^{cb}}{ \omega_{\bf{q}} }
\; \; \; ,
\\
G_{\bf{q}} & = &
\left[ \frac{ z^2 N}{V} \frac{ {\bf{q}}^2 }{ M } \right]^{\frac{1}{2}}
\frac{ f_{\bf{q}}^{cb}}{ \sqrt{ 2 V \omega_{\bf{q}} } }
\; \; \; .
\end{aligned}$$ The last term in Eq.(\[eq:Seffdefph\]) is the effective interaction between the electrons mediated by the phonons. Combining this term with $S_{int}^{el} \{ \psi \} $, we arrive at $$S_{eff} \{ \psi \}
= S_{0} \{ \psi \} + \frac{\beta}{2 V}
\sum_{q} f_{q}
\rho_{-q} \rho_{q}
\; \; \; ,
\label{eq:Seffph2}$$ with the total effective interaction given by $$f_{q}
=
\frac{f_{\bf{q}}^{cb}}
{ 1 + \nu f_{\bf{q}}^{cb} \frac{ \lambda \omega_{\bf{q}}^2 }
{ \omega_{m}^2 + \omega_{\bf{q}}^2 } }
\; \; \; ,
\label{eq:fefftotdef}$$ where the dimensionless measure for the strength of the electron-phonon coupling is $\lambda = \nu \gamma^2$, with $\gamma^2 =
\frac{ z^2 N }{VM \nu^2 c_{0}^2}$. It is instructive to compare Eq.(\[eq:fefftotdef\]) with the expression that would result from the “screening by hand” procedure[@Fetter71] described above. The approximation (a) amounts to setting $ G_{\bf{q}}^2 \rightarrow \frac{ \gamma^2}{2V}
\omega_{\bf{q}} $ in Eq.(\[eq:Seffdefph\]), while (b) corresponds to the replacement $W_{\bf{q}} \rightarrow 0$. Then one obtains from Eq.(\[eq:Seffdefph\]) the usual result $f_{q} \rightarrow f_{\bf{q}}^{cb} - \gamma^2 \frac{ \omega_{\bf{q}}^2 }
{ \omega_{m}^2 + \omega_{\bf{q}}^2 }$. Evidently the second term can be obtained by expanding Eq.(\[eq:fefftotdef\]) to [*[first order]{}*]{} in $\lambda $ and replacing $f_{\bf{q}}^{cb} \rightarrow \frac{1}{\nu}$ in the phonon term. By performing these replacements, one implicitly ignores the renormalization of the phonon energies due to the coupling to the electrons. On the other hand, the effective interaction in Eq.(\[eq:fefftotdef\]) is an [*[exact]{}*]{} consequence of the microscopic model defined above. We shall see shortly that phonon energy shift and damping can be [*[derived]{}*]{} from this expression.
Because the phonons simply modify the effective density-density interaction, we can obtain a non-perturbative expression for the Green’s function by substituting the interaction given in Eq.(\[eq:fefftotdef\]) into our general bosonization formula for the single-particle Green’s function[@Kopietz94]. The result for the Matsubara Green’s function can be written as $$G (k) =
\sum_{\alpha} \Theta^{\alpha} ( {\bf{k}} )
\int d {\bf{r}} \int_{0}^{\beta} d \tau
e^{ - i [ ( {\bf{k}} - {\bf{k}}^{\alpha}) \cdot {\bf{r}}
- \tilde{\omega}_{n} \tau ] }
G^{\alpha} ( {\bf{r}} , \tau )
\label{eq:Galphaqtildedef}
\; ,
\label{eq:Gkres2}$$ where $\alpha$ labels “boxes” that partition the degrees of freedom close Fermi surface, ${\bf{k}}^{\alpha}$ points to the center of box $\alpha$ on the Fermi surface, and the cutoff function $\Theta^{\alpha} ( {\bf{k}} )$ is unity if ${\bf{k}}$ lies inside box $\alpha$, and vanishes otherwise. For a more detailed description of this geometric construction see Refs.[@Kopietz94; @Kopietz95; @Haldane92; @Houghton93]. The interacting “patch” Green’s function $G^{\alpha} ( {\bf{r}} , \tau )$ is of the form $$G^{\alpha} ( {\bf{r}} , \tau )
=
G^{\alpha}_{0} ( {\bf{r}} , \tau )
e^{Q^{\alpha} ( {\bf{r}} , \tau ) }$$ where $G^{\alpha}_{0} ( {\bf{r}} , \tau )$ is the non-interacting Green’s function, and the [*[Debye-Waller factor]{}*]{} is $$Q^{\alpha} ( {\bf{r}} , \tau ) =
\frac{1}{\beta {{V}}} \sum_{ q } f^{RPA}_{q}
\frac{ 1 -
\cos ( {\bf{q}} \cdot {\bf{r}}
- {\omega}_{m} \tau )
}
{
( i \omega_{m} - {\bf{v}}^{\alpha} \cdot {\bf{q}} )^{2 }}
\label{eq:DWph}
\; \; \; .$$ Here $ {\bf{v}}^{\alpha} = \nabla_{\bf{k}} \xi_{\bf{k}} |_{ {\bf{k}} =
{\bf{k}}^{\alpha} }$, and the effective interaction is $$f^{RPA}_{q} =
\frac{ f^{cb}_{\bf{q}} }{ 1 + f^{cb}_{\bf{q}} {\Pi}_{ph} ( q ) }
\; \; , \; \;
{\Pi}_{ph} ( q ) = \Pi_{0} ( q ) +
\frac{ \nu \lambda \omega_{\bf{q}}^2}{ \omega_{m}^2 + \omega_{\bf{q}}^2 }
\label{eq:Piphdef}
\; \; \; ,$$ where $\Pi_0 ( q )$ is the usual non-interacting polarization due to the electronic degrees of freedom. Diagrammatically Eqs.(\[eq:Gkres2\])-(\[eq:Piphdef\]) are the result of a controlled resummation of the entire perturbation series, which is possible due to an underlying Ward-identity[@Kopietz95; @Castellani94]. Thus, Eqs.(\[eq:Gkres2\])-(\[eq:Piphdef\]) contain vertex corrections to all orders in perturbation theory, [*[including the renormalizations of the electron-phonon vertex which are only negligible if the Migdal theorem is valid]{}*]{}.
Note that ${\Pi}_{ph}^{-1} ( q )$ can be identified with the dressed phonon propagator[@Engelsberg63], so that it is clear that Eq.(\[eq:DWph\]) takes into account that the phonon dispersion changes because of the coupling to the electrons. The renormalized phonon mode appears as a peak in the dynamic structure factor[@Pines89], $$S ( {\bf{q}} , \omega ) = \frac{1 }{\pi} Im \left\{
\frac{ {\Pi}_{ph} ( {\bf{q}} , \omega + i 0^{+} ) }{ 1 + f_{\bf{q}}^{cb}
{\Pi}_{ph} ( {\bf{q}} , \omega + i 0^{+} ) } \right\}
\label{eq:Piphdyn}
\; \; \; .$$ The qualitative behavior of $S ( {\bf{q}} , \omega )$ can be determined from physical considerations[@Pines89]. In the absence of phonons, $S ( {\bf{q}} , \omega )$ consists of a sum of two terms. The first term $S_{col} ( {\bf{q}} , \omega )$ is due to the collective plasmon mode. In $d=3$ this mode approaches at long wave-lengths a finite value, the plasma frequency $\omega_{pl} = \frac{ v_{F} \kappa }{
\sqrt{3}}$, where $v_{F} $ is the Fermi velocity. Within random-phase approximation the plasmon is not damped, so that $S_{col} ( {\bf{q}} , \omega ) = Z^{pl}_{ {\bf{q}}}
\delta ( \omega - \omega_{pl} )$ with[@Pines89] $Z_{\bf{q}}^{pl} = \frac{\nu}{2} \omega_{pl} (
\frac{{\bf{q}} }{ \kappa} )^2$. For $\omega \leq v_{F} | {\bf{q}} |$ the dynamic structure factor has another contribution $S_{sp} ( {\bf{q}} , \omega )$ due to the decay of density fluctuations into single-pair excitations, i.e. Landau damping. In $d=3$ this is a rather featureless function. As long as the [*[renormalized]{}*]{} phonon velocity is small compared with $v_{F}$ and phonon damping is small, we expect that phonons give rise to an additional narrow peak that sticks out of the smooth background due to $S_{sp} ( {\bf{q}} , \omega )$. This is the dressed phonon mode.
Eqs.(\[eq:Piphdef\]) and (\[eq:Piphdyn\]) confirm the above picture. It turns out that phonons are only well defined for $ \sqrt{\lambda} \ll \frac{v_{F}}{c_{0}} $. In the opposite limit the phonons are strongly damped and mix with the plasmon mode. For $ \sqrt{\lambda} \ll \frac{v_{F}}{c_{0}} $ a simple calculation shows that the phonon contribution to the dynamic structure factor can be approximated by $$S_{ph} ( {\bf{q}} , \omega ) \approx
\frac{ Z_{{ \bf{q}}}^{ph} }{\pi}
\frac{ \Gamma_{\bf{q}} }{ ( \omega - {\Omega}_{\bf{q}} )^2 + \Gamma_{\bf{q}}^2
}
\label{eq:Srpacloseph}
\; \; \; ,$$ where the renormalized phonon energy is $${\Omega}_{\bf{q}} =
\omega_{\bf{q}}
\left[ 1 +
\frac{ \lambda }{ 1 + {\bf{q}}^2 / \kappa^2 } \right]^{\frac{1}{2}}
\; \; \; ,
\label{eq:phoenshift}$$ the phonon damping is $${\Gamma}_{\bf{q}} = \frac{\pi}{4} \frac{ \omega_{\bf{q}}^2 }{v_{F} |
{\bf{q}} | }
\frac{\lambda}{ [ 1 + {\bf{q}}^2 / \kappa^2 ]^2 }
\; \; \; ,$$ and the phonon residue can be written as $$Z_{ {\bf{q}}}^{ph} =
\frac{ \nu}{2} \Omega_{\bf{q}} \left( \frac{ \bf{q} }{\kappa} \right)^4
\frac{ \lambda }{ [ 1 + {\bf{q}}^2 / \kappa^2 ]
[ 1 + \lambda + {\bf{q}}^2 / \kappa^2 ] }
\label{eq:Zqph}
\; \; \; .$$ Note that at length scales small compared with $ \kappa^{-1}$ the phonon dispersion is not renormalized, while at distances $ | {\bf{q}} |^{-1} \gg \kappa^{-1}$ density fluctuations are surrounded by a screening cloud which modifies the phonon velocity and leads to phonon damping. In the long-wavelength limit Eq.(\[eq:phoenshift\]) implies that the phonon velocity is renormalized according to $$\tilde{c} \approx c_0 \sqrt{1 + {\lambda}}
\; \; \; .
\label{eq:BohmStarv0}$$ For a spherical three-dimensional Fermi surface this implies in the limit of large $\lambda$ $${c} \approx c_{0} \sqrt{ \lambda} = \sqrt{ \frac{z}{3} \frac{m}{M}}
v_F
\; \; \; .
\label{eq:BohmStarv}$$ This result has first been derived by Bohm and Staver[@Ashcroft76; @Bohm50] by means of a different self-consistent treatment of the coupled electron-phonon system. Note that the renormalized phonon velocity in Eq.(\[eq:BohmStarv\]) is independent of the bare velocity $c_0$.
In order to investigate whether the system is a Fermi liquid, it is sufficient to calculate the quasi-particle residue $Z^{\alpha}$. The evaluation of the full momentum- and frequency dependent Green’s function from Eqs.(\[eq:Gkres2\])-(\[eq:Piphdef\]) is a formidable mathematical problem that can perhaps only be solved numerically. The quasi-particle residue associated with patch $\alpha$ on the Fermi surface is[@Kopietz94] $Z^{\alpha} = e^{R^{\alpha}}$, where $R^{\alpha}$ is the constant part of the Debye-Waller in Eq.(\[eq:DWph\]). Expressing $f^{RPA}_{q}$ in terms of the dynamic structure factor and performing the Matsubara sum in Eq.(\[eq:DWph\]), we obtain in the limit $V , \beta \rightarrow \infty$ $$R^{\alpha} = - \int \frac{ d{\bf{q}}}{ ( 2 \pi )^3} ( f^{cb}_{\bf{q}})^2
\int_{0}^{\infty} d \omega \frac{S ( {\bf{q}} , \omega )}{ ( \omega + |
{\bf{v}}^{\alpha}
\cdot {\bf{q}} | )^2}
\label{eq:R2ph}
\; \; \; .$$ We would like to emphasize that Eqs.(\[eq:Gkres2\])-(\[eq:Piphdyn\]) and (\[eq:R2ph\]) remain valid [*[even if phonons are overdamped and mix with the plasmon mode.]{}*]{} It is not difficult to see that the integral in Eq.(\[eq:R2ph\]) exists for arbitrary $\lambda$, so that the Fermi liquid state is stable. To make progress analytically, we restrict ourselves in this work to the regime $\sqrt{\lambda} \ll \frac{v_{F}}{c_{0}}$, where the phonon damping is negligible. Then we obtain $R^{\alpha} = R^{\alpha}_{el} + R^{\alpha}_{ph}$, where $R^{\alpha}_{el}$ is due to the above mentioned electronic terms in dynamic structure factor. For a spherical Fermi surface we find $R^{\alpha}_{el} = - \frac{r_{3}}{2} ( \frac{ \kappa }{k_{F} } )^2$, where $r_{3} = O(1)$ is a numerical constant[@Kopietz94]. The phonon contribution $R^{\alpha}_{ph}$ is obtained by substituting Eq.(\[eq:Srpacloseph\]) into Eq.(\[eq:R2ph\]). To leading order in $\sqrt{\lambda} \frac{c_{0}}{v_{F}} $ the Lorentzian can be treated like a $\delta$-function, so that $$R^{\alpha}_{ph} = - \int \frac{ d{\bf{q}}}{ ( 2 \pi )^3} ( f^{cb}_{\bf{q}})^2
\frac{Z_{\bf{q}}^{ph}}{ ( \Omega_{\bf{q}} + | {\bf{v}}^{\alpha}
\cdot {\bf{q}} | )^2}
\label{eq:R3ph}
\; \; \; .$$ Substituting the above expressions for $\Omega_{\bf{q}}$ and $Z_{\bf{q}}^{ph}$ into Eq.(\[eq:R3ph\]), the integration can be performed analytically, with the result $$R^{\alpha}_{ph} = - \frac{1}{4} \left( \frac{ \kappa }{ k_{F} } \right)^2
\ln ( 1 + \lambda )
\; \; \; .$$ Using $ ( \frac{ \kappa }{k_{F} } )^2
= \frac{ 2 e^2}{\pi v_{F}}$, we finally obtain for the quasi-particle residue $$Z^{\alpha} = \left[ \frac{ e^{ - {r}_{3} } }{ \sqrt{ 1 + \lambda } }
\right]^{ \frac{e^2}{\pi v_{F}} }
\; \; \; .$$ Note that in the limit $v_{F} \rightarrow \infty$ (corresponding to the infinite density limit) the quasi-particle residue approaches unity. Although our bosonization approach is only controlled for $\kappa \ll k_{F}$, we expect that the result remains qualitatively correct at realistic metallic densities, where $\frac{e^2 }{ \pi v_{F}} $ is of the order of unity. We conclude that at long wavelengths three-dimensional LA phonons always lead to a stable Fermi liquid, although for strong electron-phonon coupling the quasi-particle residue can become small. It should be kept in mind, however, that we have only retained the long-range part of the Coulomb interaction, so that possible instabilities due to processes with large momentum transfers or superconducting pairing are not included in our formalism.
It is straightforward to generalize our results for anisotropic systems. For example, for strictly one-dimensional electron dispersion the polarization in Eq.(\[eq:Piphdef\]) is given by $\Pi_{0} (q ) = \nu \frac{ ( v_{F} q_{x} )^2}{\omega_{m}^2 + ( v_{F} q_{x}
)^2}$. In this case we find that Eqs.(\[eq:Gkres2\])-(\[eq:Piphdef\]) give rise to Luttinger liquid behavior [*[even if the phonon dispersion is three-dimensional]{}*]{}. In the absence of electron-phonon interactions such a model has been studied in Ref.[@Kopietz94b]. Alternatively, we may couple one-dimensional phonons to three-dimensional electrons. Then we should substitute in Eq.(\[eq:Piphdef\]) $\omega_{\bf{q}} = c_{0} | q_{x} | $, while choosing for $\Pi_{0} ( q )$ the usual three-dimensional polarization. It is easy to see that in this case the quasi-particle residue $Z^{\alpha}$ vanishes at the two points ${\bf{k}}^{\alpha} = \pm k_{F} \hat{\bf{x}}$ on the Fermi surface, where $\hat{\bf{x}}$ is a unit vector in the $x$-direction. At these points $| {\bf{v}}^{\alpha} \cdot {\bf{q}} |
= v_{F} q_{x}$, so that [*[the $q_{x}$-integration in Eq.(\[eq:R3ph\]) is decoupled from the remaining phase space.]{}*]{} As a consequence $R^{x}$ is logarithmically divergent. However, the total Debye-Waller factor $Q^{x} ( r_{x} \hat{\bf{x}} , \tau )$ remains finite. (We use the label $\alpha = x$ for the patch with ${\bf{k}}^{\alpha} = k_{F} \hat{\bf{x}}$. Because $G_{0}^{\alpha} ( {\bf{r}} , \tau )$ contains a $\delta$-function of the components of ${\bf{r}}$ orthogonal to ${\bf{v}}^{\alpha}$[@Kopietz94], it is sufficient to set ${\bf{r}} = r_{x} \hat{\bf{x}}$ and calculate $Q^{x} ( r_{x} \hat{\bf{x}} , \tau )$.) At $\tau = 0$ we obtain for $| r_{x} | \gg \kappa^{-1} $ $$Q^{x} ( r_{x} \hat{\bf{x}} , 0 ) \sim - \gamma_{ph} \ln ( \kappa | r_{x} | )
\; \; \; ,$$ with $$\gamma_{ph} = \frac{e^2}{\pi v_{F}} \frac{ c_{0}}{v_{F}}
\left[ \sqrt{ 1 + \lambda } - 1 \right]
\; \; \; .$$ The logarithmic divergence implies anomalous scaling characteristic for Luttinger liquids, with anomalous dimension $\gamma_{ph}$. It is also easy to calculate the quasi-particle residue in the vicinity of the Luttinger liquid points $ \pm k_{F} \hat{\bf{x}}$. A quantitative measure for the deviation from these points is the parameter $\delta = 1 -
\frac{ | {\bf{v}}^{\alpha} \cdot \hat{\bf{x}} |}{ | {\bf{v}}^{\alpha} | }$. For $\delta \ll \frac{c_{0}}{v_{F}}$ we find that the quasi-particle residue vanishes as $$Z^{\alpha} \propto \left[ \frac{ v_{F} \delta }{c_{0} } \right]^{\gamma_{ph}}
\; \; \; .$$
In summary, we have presented a non-perturbative approach to the coupled electron-phonon system, which is [*[not]{}*]{} based on the assumption that the electronic system is a Fermi liquid. We have shown that the long-range Coulomb forces mediated by three-dimensional LA phonons cannot destabilize the Fermi liquid. However, anisotropic electron- or phonon dispersions can lead to small quasi-particle residues $Z^{\alpha}$ even if the electron dispersion is three-dimensional. Although in realistic materials the phonon energies cannot be exactly one-dimensional on general grounds, we know[@Kopietz94b] that the spectral function of Fermi liquids with small quasi-particle residue can exhibit characteristic Luttinger liquid features in experimentally accessible regimes. More generally, it is tempting to speculate that the coupling between electrons and any well defined quasi-one-dimensional collective mode can lead to Luttinger liquid behavior in three-dimensional Fermi systems. In view of the chain-like structure of some of the high temperature superconductors, this result might be important for the explanation the unusual normal-state properties of these materials.
I would like to thank Roland Zeyher for helping me to understand phonons, and Kurt Schőnhammer for discussions and collaborations. I am also grateful to Guillermo Castilla for his advise.
[99]{} A. B. Migdal, Sov. Phys. JETP [**[34]{}**]{}, 996 (1958). A. L. Fetter and J. D. Walecka, [*[Quantum Theory of Many-Particle Systems]{}*]{}, (Mc Graw-Hill, New York, 1971), chapter 12. See, for example, the articles by P. W. Anderson and Y. Ren, B. Batlogg, and P. A. Lee in [*[High Temperature Superconductivity]{}*]{}, edited by K. S. Bedell, D. E. Meltzer, D. Pines, and J. R. Schrieffer (Addison-Wesley, Reading, MA, 1990). J. Bardeen and D. Pines, Phys. Rev. [**[99]{}**]{}, 1140 (1955). D. Pines and P. Nozières, [*[The Theory of Quantum Liquids]{}*]{}, (Addison-Wesley, Redwood City, Ca, 1989). S. Engelsberg and J. R. Schrieffer, Phys. Rev. [**[131]{}**]{}, 993 (1963). P. Kopietz and K. Schőnhammer, Z. Phys. B (in press); P. Kopietz, J. Hermisson and K. Schőnhammer, Phys. Rev. [**[B 52]{}**]{}, October 15 (1995). P. Kopietz, cond-mat/9506132, to appear in the Proceedings of the Raymond L. Orbach Symposium, edited by D. Hone (World Scientific, Singapore); Habilitationsschrift, Universita̋t Gőttingen, 1995. A. Luther, Phys. Rev. [**[B 19]{}**]{}, 320 (1979). F. D. M. Haldane, Helv. Phys. Acta. [**[65]{}**]{}, 152 (1992); in [*[Perspectives in Many-Particle Physics]{}*]{}, Proceedings of the International School of Physics “Enrico Fermi”, Course 121, (North Holland, Amsterdam, 1994). A. Houghton and J. B. Marston, Phys. Rev. [**[B 48]{}**]{}, 7790 (1993); A. Houghton, H.-J. Kwon, and J. B. Marston, Phys. Rev. [**[B 50]{}**]{}, 1351 (1994); A. Houghton, H.-J. Kwon, J. B. Marston, and R. Shankar, J. Phys. [**[C 6]{}**]{}, 4909 (1994); H.-J. Kwon, A. Houghton, and J. B. Marston, Phys. Rev. [**[B 52]{}**]{}, 8002 (1995). A. H. Castro Neto and E. Fradkin, Phys. Rev. Lett. [**[72]{}**]{}, 1393 (1994); Phys. Rev. [**[B 49]{}**]{}, 10877 (1994); [*[ibid.]{}*]{} [**[51]{}**]{}, 4084 (1995). J. Frőhlich, R. Gőtschmann, and P. A. Marchetti, J. Phys. [**[A]{}**]{}: Math. Gen. [**[28]{}**]{}, 1169 (1995). J. H. Kim, K. Levin, R. Wentzcovitch and A. Auerbach, Phys. Rev. [**[B 40]{}**]{}, 11378 (1989); [*[ibid.]{}*]{} [**[44]{}**]{}, 5148 (1991); M. L. Kulić and R. Zeyher, Phys. Rev. [**[B 49]{}**]{}, 4395 (1994); M. Grilli and C. Castellani, Phys. Rev. [**[B 50]{}**]{}, 16880 (1994). R. Shankar, Rev. Mod. Phys. [**[66]{}**]{}, 129 (1994). C. Castellani, C. Di Castro, and W. Metzner, Phys. Rev. Lett. [**[72]{}**]{}, 316 (1994). N. W. Ashcroft and N. D. Mermin, [*[Solid State Physics]{}*]{}, (Holt-Saunders, Philadelphia, 1976). D. Bohm and T. Staver, Phys. Rev. [**[84]{}**]{}, 836 (1950). P. Kopietz, V. Meden and K. Schőnhammer, Phys. Rev. Lett. [**[74]{}**]{}, 2997 (1995).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We perform the path-integral bosonization of the recently proposed noncommutative massive Thirring model (NCMT$_{1}$) \[JHEP0503(2005)037\]. This model presents two types of current-current interaction terms related to the bi-fundamental representation of the group $U(1)$. Firstly, we address the bosonization of a bi-fundamental free Dirac fermion defined on a noncommutative (NC) Euclidean plane $\IR_{\theta}^{2}$. In this case we show that the fermion system is dual to two copies of the NC Wess-Zumino-Novikov-Witten model. Next, we apply the bosonization prescription to the NCMT$_{1}$ model living on $\IR_{\theta}^{2}$ and show that this model is equivalent to two-copies of the WZNW model and a two-field potential defined for scalar fields corresponding to the global $U(1)\times U(1)$ symmetry plus additional bosonized terms for the four fermion interactions. The bosonic sector resembles to the one proposed by Lechtenfeld et al. \[Nucl. Phys. B705(2005)477\] as the noncommutative sine-Gordon for a [*pair*]{} of scalar fields. The bosonic and fermionic couplings are related by a strong-weak duality. We show that the couplings of the both sectors for some representations satisfy similar relationships up to relevant re-scalings, thus the NC bi-fundamental couplings are two times the corresponding ones of the NC fundamental (anti-fundamental) and eight times the couplings of the ordinary massive Thirring and sine-Gordon models.'
---
‘=11 \#1 =by60 =
\#1[[bsphack@filesw [ gtempa[auxout[ ]{}]{}]{}gtempa @nobreak esphack]{} eqnlabel[\#1]{}]{} eqnlabel vacuum \#1
\#1[@underline\#1 $\@@underline{\hbox{#1}}$]{}
‘@=12
by -by -
8.9in
0.5in
6.5in
-.6in
\#1
\#1
===
ß[s/]{} \#1[(\[\#1\])]{} \#1[\[\#1\]]{} \#1 ${\left(}
\def$[)]{}
v
\#1[\^[(R,\#1)]{}]{} ł Ł ø Ø ¶ § 0[\_[+]{}\^[(0)]{}]{} 0[\_[-]{}\^[(0)]{}]{}
\#1\#2\#3 [f\^[\#1\#2]{}\_[\#3]{}]{} 1[[w\_[1+]{}]{}]{} 1[[W\_[1+]{}]{}]{} \#1\#2[r(,)]{} \#1[[O]{}(R\_[\#1]{})]{} \#1[Ad\_[R\_[\#1]{}]{}]{} \#1[Ad\_[R\_[\#1]{}\^]{}]{} \#1[ad\_[R\_[\#1]{}\^]{}]{} \#1["7017[\#1]{}B 23\#1]{} \#1\#2\#3[[*Eur. Phys. J.*]{} [**C\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*J. Geom. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*JHEP*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Phys. Rev. Lett.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Nucl. Phys.*]{} [**B\#1**]{} (\#2) \#3]{} \#1\#2\#3\#4[[*Nucl. Phys.*]{} [**B\#2**]{} \[FS\#1\] (\#3) \#4]{} \#1\#2\#3[[*Commun. Math. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Phys. Rev.*]{} [**D\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Phys. Rev.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Phys. Lett.*]{} [**\#1A**]{} (\#2) \#3]{} \#1\#2\#3[[*Phys. Lett.*]{} [**\#1B**]{} (\#2) \#3]{} \#1\#2\#3[[*J. Math. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*J. Nonl. Math. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Prog. Theor. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Suppl. Prog. Theor. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Annals Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Proc. Natl. Acad. Sci. USA*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Rev. Mod. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Phys. Reports*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Ann. of Math.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Usp. Mat. Nauk*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Funkt. Anal. Prilozheniya*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3
[*Functional Analysis and Its Application*]{} [**\#1**]{}
(\#2) \#3
\#1\#2\#3[[*Bull. Am. Math. Soc.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Trans. Am. Math. Soc.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Invent. Math.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Letters in Math. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Int. J. Mod. Phys.*]{} [**A\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Int. J. Mod. Phys.*]{} [**B\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Advances in Math.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Reports on Math. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Ill. J. Math.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Acta Phys. Polon.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Theor. Mat. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*J. Physics*]{} [**A\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*J. Soviet Math.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Mod. Phys. Lett.*]{} [**A\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Sov. Phys. JETP*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Commun. Anal&Geometry*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Sov. Phys. JETP Lett.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Physica*]{} [**A\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Physica*]{} [**D\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Proc. Japan. Acad.*]{} [**\#1A**]{} (\#2) \#3]{} \#1\#2\#3[[*J. Phys. Soc. Japan*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Sov. J. Part. Nucl.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*J. Korean Phys. Soc.* ]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*J. Phys. IV*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*ibid*]{} [**\#1**]{} (\#2) \#3]{}
c ł Ł ø Ø u ${\Big(}
\def$[)]{} $${\Big[}
\def$$[\]]{}
\#1["7017[\#1]{} B 23\#1]{} \#1\#2[[\#1\#2]{}]{}
[**Bosonized noncommutative bi-fundamental fermion and S-duality**]{}
Harold Blas\
Departamento de Matemática - ICET\
Universidade Federal de Mato Grosso\
Av. Fernando Correa, s/n, Coxipó\
78060-900, Cuiabá - MT - Brazil\
Introduction {#intro}
============
Quantum field theories on non-commutative (NC) space-times are receiving considerable attention in recent years in connection to string and M theories (see, Refs. [@seiberg]). In the case of open string with $N=2$ world-sheet super-symmetries and for target space filled by $n$ coincident $D3-$branes the low-energy NC model corresponds to NC self-dual $U(n)$ Yang Mills theory (NC SDYM) [@Lechtenfeld1]. This theory is integrable classically [@Takasaki] and possesses a factorized S-matrix in the quantum version [@Lechtenfeld1]. It is a remarkable fact that almost all NC integrable models in less than four dimensions can be obtained by dimensional reduction of four dimensional NC SDYM (see e.g. [@lechtenfeld] and references therein).
Some non-commutative versions of the sine-Gordon (NCSG$_{1\,,2}$) [@lechtenfeld]-[@jhep2] and corresponding massive Thirring models (NCMT$_{1,\,2}$) [@jhep2] have been proposed in the literature. Their relevant NC equations of motion have the general property of reproducing the ordinary models in the commutative limit. The Grisaru-Penati version NCSG$_{2}$ [@grisaru1; @grisaru2] introduces a constraint which is non-trivial only in the non-commutative case. However, at the quantum level this model gives rise to particle production as was discovered by evaluating tree-level scattering amplitudes [@grisaru2]. Moreover, introducing an extra field, Lechtenfeld et al. [@lechtenfeld] proposed a novel NCSG$_{1}$ model which presents a factorizable and causal S-matrix at tree level computations. It seems to be that requiring an infinite number of classical conserved quantities does not uniquely reproduce the Moyal deformation of two-dimensional integrable field theories. The models NCSG$_{1,
2}$ were recently derived from the NC self-dual $U(2)$ Yang Mills theory through dimensional and algebraic reduction processes [@grisaru2; @lechtenfeld]. Moreover, these scalar field models (NCSG$_{1, 2}$) were shown to be related to spinorial models on the classical level, possessing soliton solutions in both sectors and a strong-weak mapping of their relevant coupling constants. These models, respectively named as NCMT$_{1,\,2}$, are Moyal extensions of the usual massive Thirring model (MT) defined for a bi-fundamental($bf$) fermion with two types of current-current interaction terms related to the group $U(1)\times U(1)$ [@jhep2]. The NCMT$_{2}$ model requires two copies of the NCMT$_{1}$ theory, this last model being just the NC extension of the massive Thirring for a bi-fundamental fermion.
On the other hand, the equivalence of the ordinary sine-Gordon and massive Thirring models is a two-dimensional example of having different field representations of the same theory. The massive Thirring model (MT) \_[MT]{}&=& i| \^ \_ +m | - j\^ j\_ \[mt0\], is equivalent to the sine-Gordon model (SG) [@coleman] \_[SG]{}&=& \_\^+ $\mbox{cos}
\,\b_{SG}\, \phi -1$ \[sg0\].
This duality is expressed in terms of the well-known abelian bosonization rules. The two field representations (\[mt0\]) and (\[sg0\]) describing the same physics are useful in different coupling regimes which are related by \[dual0\] = 1+ .
The particle/soliton correspondence in the context of this duality has been uncovered through the master Lagrangian and symplectic quantization approaches [@nucl2] of the higher grading $sl(2)$ affine Toda model coupled to matter (Dirac) field [@matter]. A generalization for any (untwisted) affine Lie algebra has been provided in [@jhep1].
Here we consider the problem of extending this duality to the NC spacetime using the path-integral formalism. This program, initiated in [@nunez], considers the abelian bosonization of a free fermion in two noncommuting dimensions [@moreno1; @moreno2] using path-integral techniques developed in refs. [@Naon]-[@Guillou]. We point out that in these NC treatments the Dirac fermion was assumed to be in the anti-fundamental($\bar{f}$) representation of the $U(1)$ group. Next, we summarize the main features that emerged from this bosonization procedure. The free fermion action is bosonized to a NC $U(1)$ WZW-action in which the WZ term in the action gives a non-trivial contribution due to the non-commutativity of spacetime. The procedure resembles closely the conventional non-abelian bosonization [@Witten1] with the bosonization rules being similar both in the non-abelian and NC cases. The massive Thirring model with anti-fundamental fermion is dual to a WZW model plus a NC cosine potential and an additional quartic term emerging from the NC bosonization of the current-current interaction. In this way, except for the quartic term, a copy of the Grisaru-Penati model emerges in the bosonic sector [@nunez]. Recall that the Grisaru-Penati model [@grisaru1; @grisaru2; @jhep2] is defined by the sum of a WZNW action for $e^{i\b_{\bar{f}}\,\phi}_{\star}\in U(1)_{C}$ and a cosine potential $(e^{i\b_{\bar{f}}
\phi}_{\star}+e^{-i\b_{\bar{f}}\phi}_{\star})$, plus some copies of these terms written for $e^{i\b_{\bar{f}}\,\phi^{\dagger}}_{\star}$, $\b_{\bar{f}}$ being the bosonic coupling constant. The same type of duality relationship, eq. (\[dual0\]), emerges up to couplings re-scalings [@nunez] \[dual01\] = 1+ , thus realizing an example of S-duality in NC spacetime. The S-duality in the context of NC geometry has also been discussed in Refs. [@Ganor].
The Grisaru-Penati model, in spite of integrability in the sense that it possesses an infinite number of conserved quantities, features an acausal S-matrix and particle production takes place. It was recently shown, on the classical level, that this model corresponds to the Moyal product extension of two copies of the ordinary massive Thirring model written for a bi-fundamental fermion [@jhep2]. The main features observed in this classical correspondence, apart from the doubling in the number of fermion fields, is that, even though the strong-weak duality in the soliton sector is preserved, only the unitary subgroup of the $U(1)_{C} \times U(1)_{C}$ symmetry of the bosonic model is realized in the spinorial model.
In two-dimensional Minkowski spacetime, a Moyal deformation necessarily implies the noncommutativity of the time coordinate and the causality and unitarity properties of the theory can be spoiled [@Gomis]. However, in an exactly solvable quantum field theory this feature can be improved as has been observed in the NC Lee model, where it was shown that the model is free from the $IR/UV$ mixing in both the space-space and space-time noncommutative cases. Due to the absence of this type of mixing one can expect that the theory is unitary [@Chu1].
In this context, it would be interesting to understand what actually determines the systems to be integrable (in the sense of possessing a factorizable and causal S-matrix) and dual to each other. From the considerations summarized above on NC extensions of SG model, the correspondence NCSG$_{1}$ $\rightarrow$ NCMT$_{1}$ is promising and deserves a further investigation. The main features of the NCSG$_{1}$ model are the presence of multi-soliton solutions, the factorizable and causal S-matrix (verified at tree level) despite of the space-time noncommutativity [@lechtenfeld]. Moreover, these models originate directly through reduction processes starting from the NC WZNW type action for the $GL(2)$ affine Toda field coupled to the higher grading matter (Dirac) fields (NCATM$_{1}$), such that the $U(1) \times U(1)$ symmetry is relevant, both in the construction of the NCSG$_{1}$ model and in the star-localized Noether procedure to construct the $U(1)\times U(1)$ currents in the NCMT$_{1}$ sector [@jhep2]. Then, it is interesting to perform the bosonization of the massive Thirring theory (NCMT$_{1}$) as defined in [@jhep2] and study the corresponding bosonic theory properties; e.g. to answer the question of what features it shares with the classical NCSG$_{1}$ theory.
The paper is organized as follows. In the next section we perform the NC bosonization of the bi-fundamental free fermion. In section \[mth\] we bosonize the NCMT$_{1}$ theory making use of the bosonization rules for the NC free fermion. In subsection \[mthint\] we bosonize the current-current interaction terms and identify the canonical bosonic physical field, related to the global $U(1)$ charge, coming from the kinetic terms and from the boson quartic interactions. In subsection \[mthmass\] we bosonize the mass term by axial symmetry consideration alone and get a two-field potential. Next, considering the physical field and the relevant coupling constants we establish the S-duality for the NC field theories under consideration. Moreover, we record the bosonized model in matrix form which is of the type proposed by Lechtenfeld et al. (NCSG$_{1}$) except for the quartic contributions coming from the current-current interactions. Finally, the conclusions and discussions are presented in section \[conclusion\].
The bi-fundamental fermion and NC bosonization {#freenc}
==============================================
A conventional system of $N$ free massless Dirac fermions is equivalent to a bosonic theory governed by the WZW model for a bosonic field $g \in SU(N)$ and a real bosonic field $\phi$ with a free scalar action [@Witten1]. On the other hand, the massless free fermion in the fundamental representation or anti-fundamental representation [@moreno1; @moreno2; @nunez] of the abelian $U(1)$ group on $\IR_{\theta}^{2}$ is equivalent to a noncommutative version of the WZW model [@moreno1; @Furuta]. In this section we derive the bosonization rules for the bi-fundamental free fermion action defined on $\IR_{\theta}^{2}$. We follow the path integral approach developed in [@Burgess2; @Guillou; @nunez].
We consider the bi-fundamental $U(1)\times U(1)$ free fermion theory \[free0\] S\_[bf]{}=d\^2 xi |\^[j\_[2]{}]{}\_[ i\_[1]{}]{} \_[i\_[2]{}]{}\^[i\_[1]{}]{} \_[j\_[2]{}]{}\^[j\_[1]{}]{}\^[i\_[2]{}]{}\_[j\_[1]{}]{}, where the group indices $i$ corresponds to the first $U(1)$ and the indices $j$ to the second $U(1)$. We will see that this theory is equivalent to a bosonic model by showing that the correlation functions constructed from the both theories are equal. So, let us write the generating functional for the correlations as \[partifree1\] Z\[s\_[1]{}, s\_[2]{}\]= | [D]{} {-d\^2 x |\^[j\_[1]{}]{}\_[i\_[1]{}]{}(i )\^[i\_[1]{}]{}\_[j\_[1]{}]{}} where {D\_\[s\^[(1)]{}, s\^[(2)]{}\] (x)}\^[i\_[1]{}]{}\_[j\_[1]{}]{} $$\pa_{\mu}\d_{\,\,i_{2}}^{i_{1}}
\d_{\,\,j_{1}}^{j_{2}} + i \star \d^{i_{1}}_{\,\,i_{2}}\,
s_{\mu\,\,\,j_{1}}^{(1)j_{2}}- i \d^{j_{2}}_{\,\,j_{1}}\,
s^{(2)i_{1}}_{\mu\,\,\,i_{2}} \star$$ (x)\^[i\_[2]{}]{}\_[j\_[2]{}]{} , with the notation $[\star s^{(1)}_{\mu}(x)]\psi(x)
\equiv \psi(x) \star s^{(1)}_{\mu}(x)$ and $[s^{(2)}_{\mu}(x)\star] \psi(x) \equiv s^{(2)}_{\mu}(x) \star
\psi(x)$.
Notice that in the above generating functional we have considered two types of external sources, i.e. $s_{\,\mu}^{(1)}$ and $s_{\,\mu}^{(2)}$, such that differentiations with respect to each source provide correlation functions between the relevant currents \[current1\] j\^[(1)]{} && |\^ ;\
\[current2\] j\^[(2)]{}&& \_ ( \^)\_ |\_,respectively.
The currents $j^{(1)\, \mu}$ and $j^{(2)\, \mu}$ as defined above differ only by a sign in the commutative limit, whereas on NC Euclidean space they are different [@gracia-bondia]. Next, we briefly describe the steps followed to obtain the currents (\[current1\])-(\[current2\]) associated to a bi-fundamental fermion system. In order to obtain the currents by the Noether procedure we make the global transformation localized, as discussed in [@liao] the star-localized procedure is not unique in the NC case. In fact, the most general $U(1)\mbox{x}
U(1)$ symmetry in NC space given for a charged bi-fundamental field is provided by the transformation rules [@liao; @terashima] \[u1u1nc\] u\_[2]{}(x) u\_[1]{}\^[-1]{}(x);| u\_[1]{}(x) | u\_[2]{}\^[-1]{}(x), u\_[1]{}(x)=e\_\^[-i \_[1]{}(x)]{}, u\_[2]{}(x)=e\_\^[i \_[2]{}(x)]{},where $u_{1,2}(x)$ are independent starred exponentials with $\a_{1,2}(x)=\mbox{real functions}$. The two currents above (\[current1\])-(\[current2\]) correspond to the symmetries $u_{1}$ and $u_{2}$, respectively.
Since the charge is associated to global transformation of the charged field for which there is no difference between the ordinary and non-commutative product, one can conclude that the currents $j^{(1)\,\mu}$ and $j^{(2)\,\mu}$ share the same charge. In fact, for global $u_{1, 2}$ only the product \[global\]u\_[[*global*]{}]{}= u\_[2]{} u\^[-1]{}\_[1]{}is relevant.
In the following we use the approach developed in [@moreno2; @nunez] conveniently adapted to our problem. The measure in (\[partifree1\]) is invariant under the local transformations of the fermion fields (\[u1u1nc\]) and consequently the generating functional (\[partifree1\]) is gauge invariant under separate gauge transformations acting on $s^{(1)}$ and $s^{(2)}$, namely \[identity\] Z\[s\^[(1)]{}, s\^[(2)]{}\]= Z\[s\^[(1)U\_[1]{}]{}, s\^[(2)U\_[2]{}]{}\] provided that the sources transform according to \[gauget\] s\^[(n)]{}\_ s\_\^[(n)U\_[n]{}]{} = U\_[n]{}s\_\^[(n)]{}U\_[n]{}\^[-1]{} + U\_[n]{} \_ U\_[n]{}\^[-1]{},n=1,2.
Then from the identity (\[identity\]) one gets Z\[s\^[(1)]{}, s\^[(2)]{}\] &=& | [D]{} [D]{}U\_[1]{}[D]{} U\_[2]{} {-d\^2 x |( i-ß\^[(1)U\_[1]{}]{}+ß\^[(2)U\_[2]{}]{})}\
&=& U\_[1]{}[D]{} U\_[2]{} ( i -ß\^[(1)U\_[1]{}]{}+ ß\^[(2)U\_[2]{}]{} ),\[integrated\]where the fermions have been integrated out. Notice that the integration over the $U_{n}$’s just amounts to a change of the overall normalization.
It is useful to introduce an explicit representation of the group $g \in U(1)\times U(1)$ in the form \[u1u1\] g = (
[cr]{} e\^[iŁ\_[2]{}]{}\_ & 0\
0 & e\^[-iŁ\_[1]{}]{}\_
) g\_[2]{}g\_[1]{}= g\_[1]{}g\_[2]{}, g\_[2]{} = (
[cr]{} e\^[iŁ\_[2]{}]{}\_& 0\
0 & 1
), g\_[1]{} = (
[cr]{} 1 & 0\
0 & e\^[-iŁ\_[1]{}]{}\_
), where $\L_{1,2}$ are real fields.
In this $2\times 2$ representation the matrix Lie algebra valued fields $\ss^{(n)}\,(n=1,2)$ corresponding to each subalgebra respectively has just one entry different from zero, respectively. Then the matrix of interest becomes ( i-ß\^[(1)U\_[1]{}]{}+ß\^[(2)U\_[2]{}]{})\_[22]{} &=& (
[cc]{} i+ß\^[(2)U\_[2]{}]{}& 0\
0 & i-ß\^[(1)U\_[1]{}]{}
)\
&=&(
[cc]{} i+ ß\^[(2)U\_[2]{}]{} & 0\
0 & 1
)(
[cc]{} 1 & 0\
0 & i -ß\^[(1)U\_[1]{}]{}
),\[matprod\] where in the right hand sides of (\[matprod\]) the fields $s^{(n)\,U_{n}}\,(n=1,2)$ denote just functions instead of Lie algebra valued fields.
Therefore the generator functional factorizes \[factorized1\] Z\[s\^[(1)]{}, s\^[(2)]{}\] &=& U\_[1]{}[D]{} U\_[2]{}( i-ß\^[(1)U\_[1]{}]{}) (i+ß\^[(2)U\_[2]{}]{}),where the Lie algebra valued character has been restored for the fields $\ss^{(n)\,U_{n}}$.
This factorization is expected since $Z[s^{(1)},\,
s^{(2)}]$ is gauge invariant under the separate gauge transformations (\[gauget\]) acting on $s^{(1)}$ and $s^{(2)}$, respectively
Let us introduce the connections through b\_\^[(n)]{} = s\_\^[(n) U\_[n]{}]{},such that the field strengths of $b_{\mu}^{(n)}$ and $s_{\mu}^{(n)}$ satisfy F\_\^[(n)]{}\[b\^[(n)]{}\] =U\_[n]{} F\_\^[(n)]{}\[s\^[(n)]{}\]U\_[n]{}\^[-1]{},n=1,2.
We proceed by changing the $U_{n}$ integrations to integrations over the new connections $b^{(n)}$. In order to achieve that we shall use the identity [@Guillou] \[stren\] b\_\^[(n)]{} [P]{}\[b\_\^[(n)]{}\] =U\_[n]{}[P]{}\[s\_\^[(n) U\_[n]{}]{}\], n=1,2. where ${\cal P}$ is a gauge invariant function.
Making use of the identities (\[stren\]) the generating functional (\[factorized1\]) can be written as Z\[s\^[(1)]{}, s\^[(2)]{}\] = b\_\^[(n)]{} \_[n]{} (b\^[(n)]{}\_[+]{}-s\^[(n)]{}\_[+]{})\
( i-\^[(1)]{})( i+\^[(2)]{}),\[gfix1\]where we have chosen the gauge conditions $b_{+}^{(n)}=s_{+}^{(n)},\,\,n=1,2$; with the $\D_{n}$’s being the relevant Faddeev-Popov determinants.
We define the Lie algebra valued fields $\hat{a}_{n}$ and use them in order to reinforce the delta function conditions, i.e. Z\[s\^[(n)]{}\] &=& \_[n]{} [D]{}b\_\^[(n)]{} \_[n]{} (b\^[(n)]{}\_[+]{}-s\^[(n)]{}\_[+]{}) $i\sp-\star\,\sb^{(1)} $ $i\sp+\sb^{(2)} \star $\
&& \_[n]{} ${\cal
A}_{n}$\[gfix11\]where ${\cal A}_{n} \equiv -
\frac{C_{n}}{8\pi}\mbox{tr}\, \int d^2x \,\hat{a}_{n}\,
\d[\epsilon^{\mu \nu} (F_{\mu\nu}^{(n)}[b^{(n)}]-
F_{\mu\nu}^{(n)}[s^{(n)}])]$ , with the $C_{n}$’s being some constants to be conveniently chosen below. In two dimensions one can parameterize a gauge boson as a pure gauge field, so the source fields $s_{\mu}^{(n)}$ and $b_{\mu}^{(n)}$ can be written in terms of group-valued variables \[par1\]s\_[+]{}\^[(n)]{}&=&i \_[n]{}\^[-1]{} \_[+]{} \_[n]{}\
\[par2\] s\_[-]{}\^[(n)]{}&=&i s\_[n]{} \_[-]{}s\_[n]{}\^[-1]{}\
\[par3\] b\_[+]{}\^[(n)]{}&=&i (\_[n]{}\_[n]{} )\^[-1]{} \_[+]{} ( \_[n]{}\_[n]{})\
\[par4\] b\_[-]{}\^[(n)]{}&=&i (s\_[n]{}b\_[n]{}) \_[-]{}(s\_[n]{}b\_[n]{})\^[-1]{}
The $U(1)\times U(1)$ group valued fields $s_{n},
\widetilde{s}_{n}, b_{n}, \widetilde{b}_{n}$ are in the same representation as the $g_{n}$’s $(n=1,2)$ in (\[u1u1\]).
Notice that the determinant depending on $\sb_{2}$ is associated to the fundamental fermion representation and the one with $\sb_{1}$ resembles to that of the anti-fundamental. The fundamental and anti-fundamental determinants coincide, as discussed in [@moreno2]. The exact effective action in each case is obtained by integrating the chiral anomaly [@moreno1; @moreno2]. On the other hand, in two dimensions the bi-fundamental Dirac fermion carries no mixed anomaly [@nakajima; @Bonora] (for chiral matter see [@Martin]). Therefore, one can make the affective action computation separately for each $U(1)$ sector. Taking into account (\[factorized1\]) the fermion determinant for the parameterizations (\[par1\])-(\[par4\]) can be written as [@moreno1; @moreno2] ( i-\^[(1)]{}+\^[(2)]{} ) = (W\[\_[1]{}\_[1]{} s\_[1]{}b\_[1]{}\]) (W\[\_[2]{}\_[2]{} s\_[2]{} b\_[2]{}\]), where \[wznw\] W\[g\]= \_[§]{} d\^2x (\_g \_g\^[-1]{}) - \_[B]{} d\^3 x \^[ijk]{} (g\^[-1]{}\_[i]{}g g\^[-1]{}\_[j]{}g g\^[-1]{}\_[i]{}g),is the NC extension of the WZNW model [@moreno1; @Furuta]. The manifold $\S$, parameterized by $(x^0,
x^1)$, is the boundary of the three-dimensional manifold $B$. The extra dimension $x^2$ is taken to be commutative.
The Jacobians for the change of variables from $b_{\mu}^{(n)}$ to $b_{n},\,\widetilde{b}_{n}$ become [@nunez] b\_[+]{}\^[(n)]{}[D]{}b\_[-]{}\^[(n)]{}&=& \_ D\_[+]{}\[\_[n]{} \_[n]{}\] \_ D\_[-]{}\[sb\] [D]{}\_[n]{}[D]{}b\_[n]{}\
&=& [D]{}\_[n]{}[D]{}b\_[n]{} (\_[n]{} W\[\_[n]{}\_[n]{}s\_[n]{}b\_[n]{}\]), n=1,2; where $\kappa_{n}$ is the parameter depending on the representation of the covariant derivative involved in the change of variables, and since we shall not use it here we leave it as an unspecified parameter.
The delta functional in (\[gfix11\]), for these change of variables, can be written as (b\^[(n)]{}\_[+]{}-s\^[(n)]{}\_[+]{})= (\^[(n)]{}-I).
Then, from the results above, (\[gfix11\]) can be written as Z\[s\^[(n)]{}\] &=& \_[n]{} [D]{}b\_[n]{} \_[n]{} ([B]{}\_[n]{}) $(1+\kappa_{n})
W[\widetilde{s}_{n}\star s_{n}\star b_{n}]$,\[group1\] where \_[n]{} =i d\^2x (D\_[+]{}\[\_[n]{}\]\_[n]{}) s\_[n]{}b\_[n]{}(\_[-]{}b\_[n]{}\^[-1]{}) s\_[n]{}\^[-1]{}, and the covariant derivative is in the adjoint representation of the corresponding $U(1)$ group.
In the present approach to bosonization one introduces two group-valued fields, each one related to the corresponding $U(1)$ group, say $g_{1}$ and $g_{2}$, which will play the role of the boson fields equivalent to the original bi-fundamental fermionic model and will be invariant under gauge transformations. So, let us introduce a change of variables from the algebra-valued fields $\hat{a}_{n}$ to the group-valued fields $a_{n}$ as D\_[+]{}\[\_[n]{}\]\_[n]{}= \_[n]{}\^[ -1]{}(a\^[-1]{}\_[n]{} \_[+]{} a\_[n]{}).
The Jacobian of these transformations are [@Burgess2; @nunez] \_[n]{} =J\_[L]{}\^[(n)]{} [D]{}a\_[n]{},J\_[L]{}\^[(n)]{}= (\_[n]{} W\[a\_[n]{}\_[n]{}s\_[n]{}\]-\_[n]{} W\[\_[n]{}s\_[n]{}\]),n=1,2.
Therefore, the generating functional becomes Z\[s\^[(n)]{}\] &=& a\_[n]{} [D]{}b\_[n]{} \_[n]{} $(1+\kappa_{n})
W[\widetilde{s}_{n}\star s_{n}\star b_{n}]+ \kappa_{n} W[a_{n}\star \widetilde{s}_{n}\star s_{n}]- \kappa_{n}
W[\widetilde{s}_{n}\star s_{n}]-\nonumber \\&&
\frac{C_{n}}{4\pi} \mbox{tr} \int d^2 x \,
\widetilde{s}_{n}^{\,-1}\star (a^{-1}_{n}\star \pa_{+} a_{n})\star
\widetilde{s}_{n}\star s_{n}\star (b_{n}\star \pa_{-} b_{n}^{-1})
\star s_{n}^{-1} $ \[group2\].
Choosing the up to now arbitrary constants $C_{n}$ to be $C_{n}\equiv (1 + \kappa_{n})$ and making use of the Polyakov-Wiegmnann identity \[PW\] W\[g\_[2]{}g\_[1]{}\]= W\[g\_[2]{}\]+ W\[g\_[1]{}\]- \_[§]{} d\^2x (g\_[2]{}\^[-1]{}\_[+]{} g\_[2]{} g\_[1]{} \_[-]{} g\_[1]{}\^[-1]{} ),
one can write Z\[s\^[(n)]{}\] = a\_[n]{} [D]{}b\_[n]{} \_[n]{} $(1+\kappa_{n})
W[a_{n}\widetilde{s}_{n}s_{n}b_{n}]
+ W[\widetilde{s}_{n}s_{n}]-W[a_{n}\widetilde{s}_{n}s_{n}] $ \[group3\]
Since the integrations over the $b_{n}$’s can be trivially factorized by making the changes $b_{n} \rightarrow
\hat{b}_{n}=a_{n}\widetilde{s}_{n}s_{n}b_{n}$, each one with trivial Jacobian, one has Z\[s\^[(n)]{}\] = a\_[n]{} \_[n]{} $
W[\widetilde{s}_{n}s_{n}]-W[a_{n}\widetilde{s}_{n}s_{n}]$ \[group31\].
The final transformations $a_{n}\widetilde{s}_{n}s_{n} \rightarrow
\widetilde{s}_{n}a_{n}s_{n}$ with trivial Jacobians; together with the Polyakov-Wiegmann identity provide us the boson counterpart of the bi-fundamental fermion model in (\[partifree1\]) (as usual rename $a_{n}$ as $g_{n}$ )Z\[s\^[(n)]{}\_\] &=& \_[n]{} [D]{}g\_[n]{} $$-W[g_{n}] + \frac{i}{4\pi} \mbox{tr}
\int d^2 x \, (s_{+}^{(n)}\star g_{n} \star \pa_{-} g_{n}^{-1} +
s_{-}^{(n)} \star g_{n}^{-1}\star \pa_{+} g_{n})+
\nonumber\\
&& \frac{1}{4\pi} \mbox{tr} \int d^2 x \, (g_{n}^{-1}\star
s_{+}^{(n)}\star g_{n}\star s_{-}^{(n)}-s_{+}^{(n)}\star
s_{-}^{(n)})$$, \[group4\]where the matrices $g_{i}$ are in the representation (\[u1u1\]).
Differentiating (\[group4\]) with respect to any of the components of the sources gives the correlation functions in the relevant chirality sector \[bosonimatrix\] j\_[+]{}\^[(n)]{} g\_[n]{}\^[-1]{}\_[+]{} g\_[n]{}, j\_[-]{}\^[(n)]{} g\_[n]{}\_[-]{} g\_[n]{}\^[-1]{}, n=1,2;or in terms of the field components \[bosoni1\] |\_[+]{} & & e\^[iŁ\_[1]{}]{}\_ \_[+]{} e\^[-iŁ\_[1]{}]{}\_,| \_[-]{} e\^[-iŁ\_[1]{}]{}\_ \_[-]{} e\^[iŁ\_[1]{}]{}\_;\
\[bosoni2\] \_ (\_[+]{})\_ |\_ & & e\^[-iŁ\_[2]{}]{}\_ \_[+]{} e\^[iŁ\_[2]{}]{}\_,\_ (\_[-]{})\_[ ]{} |\_ e\^[iŁ\_[2]{}]{}\_ \_[-]{} e\^[-iŁ\_[2]{}]{}\_.
Taking into account the representation (\[u1u1\]) one can see that the global free fermion symmetries $U(1)\times U(1)$ of type (\[u1u1nc\]) translate to the bosonized model (\[group4\]) as g\_[2]{} U\_[2]{} g\_[2]{},g\_[1]{} U\_[1]{}\^[-1]{} g\_[1]{},U\_[2]{}=(u\_[2]{}, 1),U\_[1]{}=(1, u\_[1]{}). \[globalbosmat\]
Then, considering the product $u_{2} u_{1}^{-1}$ ($\a_{i}=$const.) as in (\[global\]) for the global $U(1)$, the transformations (\[globalbosmat\]) written for the product of field components becomes e\^[iŁ\_[2]{}]{}\_e\^[iŁ\_[1]{}]{}\_ u\_[2]{} u\^[-1]{}\_[1]{} e\^[iŁ\_[2]{}]{}\_e\^[iŁ\_[1]{}]{}\_ \[pm0\].
We consider the bosonic theory (\[group4\]) in terms of the boson fields that appear naturally in the commutative limit corresponding to the field responsible for the global $U(1)$ charge and a decoupled free field, respectively (similar arguments have been used in [@liao] to study a gauge theory of a doubly $U(1)$ gauged matter in NC spacetime). So, in the $\theta \rightarrow 0$ the transformation (\[pm0\]) becomes e\^[iŁ\_[+]{}]{} u\_[2]{} u\^[-1]{}\_[1]{} e\^[iŁ\_[+]{}]{},Ł\_ (Ł\_[2]{}Ł\_[1]{}) \[pm1\].
Then, the charge corresponding to the global $U(1)$ symmetry of the NC free fermion should be related to the field $\L_{+}$ of the bosonic sector. Once the physical fields are identified, one can re-write the action of (\[group4\]) in terms of them on NC spacetime. Thus, writing the action of (\[group4\]) in terms of the fields $\L_{\pm}$ one can see that these fields do not decouple from each other, whereas in commutative space they do. Let us discuss this point in more detail using the Polyakov-Wiegmnann identity (\[PW\]). Since the last term in (\[PW\]) vanishes for the representation (\[u1u1\]), then the sum $W[g_{1}]+ W[g_{2}]$ is equivalent to $W[g_{2}\star g_{1}]$ which couples the fields $\L_{+}$ and $\L_{-}$ in a non-trivial way. However, it can be seen that the fields $\L_{-}$ and $\L_{+}$ decouple in the commutative limit of the action (\[group4\]). In this limit one has $W[g_{2}g_{1}]=
\int d^2x
\frac{1}{16\pi}[(\pa_{\mu}\L_{+})^2+(\pa_{\mu}\L_{-})^2]$. Since the bosonized action of an ordinary free Dirac fermion in Minkowski spacetime corresponds just to one real scalar field, then one can set $\L_{1}=\L_{2}$ ($\L_{-}=0$) in the commutative limit (however, see in the last paragraph of the next section a discussion regarding the complex character of the fields $\L_{i}$ in the bosonization process). Moreover, in this limit the two currents differ only by a minus sign; in fact, by comparing (\[bosoni1\]) and (\[bosoni2\]) one has $j^{(1)}_{\mu}=-j^{(2)}_{\mu}$, as pointed out before. Regarding the discussions above, an important point is that the symmetry properties are simplest in terms of the fields $\L_{1}$ and $\L_{2}$ and the physical interpretation becomes direct when expressed in terms of the fields $\L_{\pm}$.
The NC bi-fundamental massive Thirring model {#mth}
============================================
The (Euclidean) Lagrangian of the non-commutative massive Thirring model (NCMT$_{1}$) [^1] is
\_[NCMT]{}&=& i| \^ \_ +m | - j\^[(1)]{} j\^[(1)]{}\_ - j\^[(2)]{}j\^[(2)]{}\_ \[ncmt0\], defined for Dirac fields and with the currents given in (\[current1\])-(\[current2\]). Here, $\l_{bf}$ is the relevant coupling constant and the group index contractions are being assumed.
The action related to (\[ncmt0\]) in terms of the spinor field components becomes \_[NCMT]{}&=&d\^2x$$\nonumber
i\widetilde{\psi}_{L}\pa_{+}\psi_{L} +
i\widetilde{\psi}_{R}\pa_{-}\psi_{R}+ m \(\widetilde{\psi}_{R}\psi_{L}
+\widetilde{\psi}_{L}\psi_{R}\)-\\
&&\l_{bf}\(\widetilde{\psi}_{R}\star \psi_{R}\star
\widetilde{\psi}_{L}\star \psi_{L}+ \psi_{R}\star
\widetilde{\psi}_{R}\star \psi_{L}\star \widetilde{\psi}_{L}
\)$$\[ncmt\], where again the group index notation has been suppressed for convenience.
Notice that the NCMT model (\[ncmt0\]) contains the non-standard interaction term $j^{(2)}_{\mu}\star j^{(2)\,\mu} \sim
\psi_{R}\star\widetilde{\psi}_{R}\star\psi_{L}\star\widetilde{\psi}_{L}$, which, to our knowledge, has not been considered previously in the literature. Since the currents $j^{(1)\, \mu}$ and $j^{(2)\, \mu}$ differ only by a sign in the commutative limit, the Lagrangian (\[ncmt0\]) reduces to the usual MT model in this limit. The bosonization process of the NC extension of the usual Thirring interaction performed in [@moreno2; @nunez] considers only $j^{(1)}_{\mu}\star j^{(1)\,\mu}$. The inclusion of a second $U(1)$ current will change various aspects of the bosonized model, such as the potential, the duality relationship and the field content, as we will show below.
The $U(1)\times U(1)$ symmetries give the associated conservation equations for the NCMT$_{1}$ model \[curr11\] \_ j\^[(1)]{}&=&0,\
\[curr22\] \_ j\^[(2)]{}&=&0,where the currents $j^{(n)}_{\mu} \,(n=1,2)$ are of the type given in (\[current1\])-(\[current2\]). The conservation laws (\[curr11\])-(\[curr22\]) can be verified as the result of the field equations of motion [@jhep2]. As in the free bi-fundamental fermion case the global $U(1)$ charge is associated to the transformation $u_{2}u_{1}^{-1}$ and the currents $j^{(1)}_{\mu}$ and $j^{(2)}_{\mu}$ share the same charge. In fact, the two currents are related by $\psi_{\a} \leftrightarrow
\psi_{\a}^{\dagger}$. Therefore the Noether $U(1)$ charge is given by \[charge\] Q e\_[n]{}d\^2 x j\^[(n) 0]{},n=1 2;where the $e_{n}$ are some normalization factors, which can be set $e_{1}=1,\,e_{2}=-1$.
The Thirring type interactions and NC bosonization {#mthint}
---------------------------------------------------
Let us bosonize the four-fermion interaction terms of the model (\[ncmt0\]), i.e. \_[ł\_[bf]{}]{}&=& - j\^[(1)]{} j\^[(1)]{}\_ - j\^[(2)]{}j\^[(2)]{}\_ \[ncth1\].
Here we shall use directly the bosonization dictionary for the currents established in (\[bosoni1\])-(\[bosoni2\]) [^2]. Then, the four fermion interactions of (\[ncth1\]) correspond to the following terms in the boson theory \[currboso1\] - d\^2x 2(|\_[+]{})(|\_[-]{}) d\^2x g\_[1]{}\^[-1]{}\_[+]{}g\_[1]{}g\_[1]{} \_[-]{}g\_[1]{}\^[-1]{}\
\[currboso2\] - d\^2x 2(\_(\_[+]{})\_|\_)(\_(\_[-]{})\_|\_) d\^2x g\_[2]{}\^[-1]{}\_[+]{}g\_[2]{}g\_[2]{} \_[-]{}g\_[2]{}\^[-1]{}.
It is known that the bosonized four fermion interaction term in the commutative case adds to the kinetic term an extra contribution, whereas the right hand side terms in (\[currboso1\])-(\[currboso2\]) contribute an infinite series of theta dependent higher derivative terms to the first quadratic term. In this way the NC extension provides additional terms making the bosonized theory in fact very different from the commutative version.
The bosonized Lagrangian for the kinetic and four-fermion interaction terms becomes W\[g\_[1]{}\] + W\[g\_[2]{}\] + d\^2x g\_[2]{}\^[-1]{}\_[+]{}g\_[2]{}g\_[2]{} \_[-]{}g\_[2]{}\^[-1]{}+\
d\^2x g\_[1]{}\^[-1]{}\_[+]{}g\_[1]{}g\_[1]{} \_[-]{}g\_[1]{}\^[-1]{}\[boso1\]
In order to understand better the bosonic sector we need to uncover a canonically normalized scalar field related to the global $U(1)$ symmetry which corresponds to the charge of the fermion field. So, let us consider the scalar fields $\L_{i}$ defined in (\[u1u1\]). Then, the $\theta$ expansion of (\[boso1\]) provides \_[i=1]{}\^[2]{}d\^2x \_[+]{}Ł\_[i]{}\_[-]{}Ł\_[i]{} + \_[i=1]{}\^[2]{}d\^2x \_[+]{}Ł\_[i]{}\_[-]{}Ł\_[i]{}+...
Taking into account these contributions and the definitions (\[pm1\]) one gets $1+ \frac{\l_{bf}}{8\pi}$ d\^2x $$\pa_{+}\L_{+}\pa_{-}\L_{+}+\pa_{+}\L_{-}\pa_{-}\L_{-}$$ + ..., \[kinetic12\]therefore, the canonically normalized scalar related to the fermion charge corresponds to \_[+]{}$$\frac{1}{8\pi} \(1+ \frac{\l_{bf}}{8\pi}\)$$\^[1/2]{} Ł\_[+]{}. \[canonic\]
This result will be used below to relate the two-field NC sine-Gordon and NCMT$_{1}$ coupling constants. Observe that for stability of the bosonic theory we must have $\l_{bf} > -8\pi$. The allowed range of the coupling gets enlarged as compared to the anti-fundamental massless Thirring which is $\l_{f} > -4\pi$ [@nunez]. Recall that for the ordinary massless Thirring model this bound is $\l > -\pi$ [@coleman]. Thus, there is a whole range of coupling constants, depending on the fermion representation, for which the NC massless Thirring is a sensible model.
At this stage we have reproduced two copies of the WZNW action associated, respectively, to the fields $\L_{1}$ and $\L_{2}$, and additional terms in (\[boso1\]) due to the current-current couplings. A two-field potential will emerge from the bi-fundamental fermion mass term, as we will see in the next subsection.
The mass term and NC bosonization {#mthmass}
---------------------------------
Next we consider the bosonization of the bi-fundamental fermion mass term. The relevant term is \[mass\] S\_[M]{}= md\^2x $\psi^{\dagger\, j_{1}}_{R\,\,i_{1}}\star
\psi_{L\,\,j_{1}}^{i_{1}} +\psi^{\dagger\,
j_{1}}_{L\,\,\,i_{1}}\star \psi_{R\,\,j_{1}}^{i_{1}}$
Here, we follow the procedure implemented in Refs. [@Burgess1; @nunez] in which the chiral symmetry plays a central role in the bosonization process. In our case this procedure still applies since the free bi-fundamental fermion possesses a global axial symmetry $U_{5}(1)\times U_{5}(1)$ for which the starred products collapse to the ordinary ones. In fact, the [*free*]{} fermion theory (\[partifree1\]) is symmetric under the global chiral symmetry \[u1u1qui\] ()\^[T]{} $e^{i\a_{2}\g_{5}} \psi $\^[T]{} e\^[i\_[1]{}\_[5]{}]{} ;(|)\^[T]{} $e^{i\a_{1}\g_{5}}
\bar{\psi}$\^[T]{} e\^[i\_[2]{}\_[5]{}]{},where $T$ stands for matrix transpose operation and the suggestive products have been written in spite of their global character. It is assumed that the bosonic theory preserves this symmetry [@Witten1]. Besides, the duality approach to bosonization relies on gauging the vector $U(1)\times U(1)$ symmetries [@Burgess1; @Burgess2]. On the other hand, it is a known fact that the axial symmetry does not survive quantization. The relevant axial anomalies in NC plane are [@moreno1; @moreno2; @nakajima; @ardalan] \[axial1\] \_ j\_[5]{}\^[(2)]{} &=& \^ F\_\^[(2)]{}, j\_[5]{}\^[(2)]{}= \_ (\^\_[5]{})\_|\_ ,\
\_ j\_[5]{}\^[(1)]{} &=& - \^ F\_\^[(1)]{},j\_[5]{}\^[(1)]{}=| \^\_[5]{} ,\[axial2\],where the $F_{\mu\nu}^{(i)}$ are the NC gauge field strengths.
In order to make the axial symmetry manifest during the bosonization process the authors of [@Burgess1] provided the lagrange multiplier with a transformation property to cancel the fermion axial anomaly. In fact, the gauged action would have the terms \~$ \int \-d^2x \[\, \frac{1}{2\pi} \L_{1}\star
\epsilon^{\mu\nu} F_{\mu\nu}^{(1)} - \frac{1}{2\pi} \L_{2}\star
\epsilon^{\mu\nu} F_{\mu\nu}^{(2)}\] $ \[multipliers\]
The key point is that if the fields $\L_{i}$ transform as $\L_{i}
\rightarrow \L_{i}-\a_{i}$ under axial transformations, then the Lagrange multiplier terms (\[multipliers\]) will cancel the axial anomaly terms (\[axial1\])-(\[axial2\]) which emerge in the path-integral, thus making the whole generating functional manifestly chiral symmetric. These considerations tell us how the $\L_{i}$ fields transform under chiral rotations [@Burgess1].
Even though the mass term is no longer axial symmetric, these transformation properties allow us to bosonize the mass term. The observation is that these transformation rules remain the same for the bilinear terms in (\[mass\]). The problem resides in finding the bosonic functionals associated to these bilinear terms, ${\cal F}(\L_{1},\, \L_{2}) \equiv \psi^{\dagger\,
j_{1}}_{R\,\,i_{1}}\star \psi_{L\,j_{1}}^{i_{1}}$ and its corresponding ${\cal F}^{\dagger}(\L_{1},\, \L_{2})$. Here the functionals depend on $\L_{i}$ through $\star$-products. Under the [*global*]{} symmetry transformations (\[u1u1qui\]) one has \^[ j\_[1]{}]{}\_[Ri\_[1]{}]{}\_[Lj\_[1]{}]{}\^[i\_[1]{}]{} e\^[-2i(\_[1]{}+ \_[2]{})]{} \^[ i\_[1]{}]{}\_[Rj\_[1]{}]{}\_[Lj\_[1]{}]{}\^[i\_[1]{}]{},and taking into account the $\L_{i}$ field transformations discussed above we have the functional transformation property \[funceq\] [F]{}(Ł\_[1]{}-\_[1]{}, Ł\_[2]{}-\_[2]{}) = e\^[-2i(\_[1]{}+ \_[2]{})]{} [F]{}(Ł\_[1]{}, Ł\_[2]{}).
Therefore, one has that \[solution\] [F]{}(Ł\_[1]{}, Ł\_[2]{})\~e\^[2iŁ\_[1]{}]{}\_ e\^[2iŁ\_[2]{}]{}\_,uniquely solves (\[funceq\]) for globally defined rotations. With this result the mass term (\[mass\]) becomes \[potential\] S\_[M]{}= d\^2x m \_[0]{}$e^{2i\L_{1}}_{\star} \star
e^{2i\L_{2}}_{\star}+e^{-2i\L_{2}}_{\star} \star
e^{-2i\L_{1}}_{\star} $,where $\a_{0}$ is a parameter to be associated to the zero point energy.
This is precisely the potential proposed by Lechtenfel et al. for the NC sine-Gordon [@lechtenfeld]. In the previous section we associated the field $\L_{+}=(\L_{1}+\L_{2})$ to the charged sector of the bi-fundamental NC free fermion model. It is expected that the same field will be associated to the global $U(1)$ charge of the NCMT$_{1}$ theory as discussed in the paragraph after the Eqs. (\[curr11\])-(\[curr22\]). This field has been isolated from the quadratic and quartic terms in (\[kinetic12\]). So, we would like to uncover this field in the bosonized mass term (\[potential\]) as well. Actually, this is not possible to do exactly due to the $\star-$product involved and only one can pursue it in the $\theta$ perturbation expansion. Then the expansion of the potential provides d\^2x m \_[0]{} $$2\,\mbox{cos}(2\L_{+})+\,\,...$$.
The canonically normalized field $\Phi_{+}$ was defined in (\[canonic\]), then the first potential term becomes d\^2x m \_[0]{} $$2\,\mbox{cos}(\b_{bf}\, \Phi_{+})+\,\,...$$,where \[duality\] = 1+ .
Thus, we have a S-duality relationship. The relation (\[duality\]) constitutes the quantum version of the strong-weak duality obtained in the classical $1-$soliton sector of the NCMT$_{1}$ $\rightarrow$ NCSG$_{1}$ correspondence [@jhep2].
This type of relationship has also been derived in [@nunez] for the NC bosonization of the massive Thirring model with anti-fundamental fermion (see eq. (\[dual01\])). With these results in hand one can establish that the couplings of the various models are related by \[relationships\] ł\_[bf]{}= 2 ł\_[|[f]{}]{}= 8 ł\_[MT]{},\^2\_[bf]{}= 2 \^2\_[|[f]{}]{}= 8 \^2\_[SG]{}, where the $\l_{\bar{f}}$, $\b^2_{\bar{f}}$ couplings correspond to the fermionic and bosonic sector considered in [@nunez], respectively [^3]. The couplings $\l_{MT},\, \b^2_{SG}$ were defined in (\[mt0\])-(\[sg0\]) for the models in ordinary spacetime.
Notice that the duality relation is maintained in each term of the $\theta$ expansions of the potential terms (\[potential\]), as well as the quadratic and quartic terms of (\[boso1\]). To see this fact notice that in order to define the canonically normalized field $\Phi_{+}$ the fields $\L_{i}$ have been re-scaled as $\L_{i} \rightarrow \frac{\beta}{2} \L_{i}$, then the expansions will only produce terms proportional to positive powers of the bosonic coupling constant $\b$.
In the bosonic sector we must have a topological charge corresponding to the global $U(1)$ Noether charge of the fermionic theory. The physical scalar field associated to this topological charge may be identified in the commutative limit $\theta
\rightarrow 0$ of the model. The field $\L_{-}$ decouples completely in this limit since the quartic terms contributions are only of the bilinear kinetic type and the combination $\L_{+}$ alone appears in the potential term. So, it is clear that the field $\Phi_{+}$ defined in (\[canonic\]) carries this charge in the bosonic sector. Then, we can define [@grisaru1; @jhep2][^4]
j\^\_[top]{} \^\_ \_[+]{},Q\_[top]{}= \_[-]{}\^ dx\^[1]{} \_[n=0]{} \^[n]{} Q\_[n]{},\[topo\]where $Q_{top}$ is the topological charge expressed as a power series in $\theta$. This expression is the same as in the classical NCSG$_{1}$ model. In fact, for the classical one (anti-)soliton sector it has been obtained an exact topological charge equal to $\pm 1$, respectively; i.e. one has in (\[topo\]), $Q_{n}=0 \, (n>0)$, since the $1-$soliton solution of the ordinary sine-Gordon model solves also the NC theory [@lechtenfeld; @jhep2; @dimakis]. The Noether and topological currents equivalence is also present in the ordinary MT/SG relationship, on the quantum [@coleman; @Naon] and classical level establishing a soliton/particle correspondence [@nucl2; @orfanidis].
Finally, the NCMT$_{1}$ model (\[ncmt0\]) is equivalent to a bosonic model which comprises the eq. (\[boso1\]) plus the eq. (\[potential\]), which we record below in matrix notation \_[bf]{}\[g\_[1]{}, g\_[2]{}\] &=& W\[g\_[1]{}\] + W\[g\_[2]{}\] + d\^2x $$m \a_{0}\, \(\s_{1} g_{2}^{-2}\star
g_{1}^{-2}\s_{1} \star g_{1}^{2} \star g_{2}^{2}\) +\\&&
\frac{\l_{bf}}{32\pi^2} \, \(
g_{2}^{-1}\star\pa_{+}g_{2}\star g_{2} \star \pa_{-}g_{2}^{-1}+
g_{1}^{-1}\star\pa_{+}g_{1}\star g_{1} \star \pa_{-}g_{1}^{-1} \)$$ \[ncsg12\], where the matrices $g_{1}, g_{2}$ are given in (\[u1u1\]) and $\s_{1} = [\begin{array}{cc}
0 & 1 \\
1 & 0
\end{array}].$ Thus, the bosonic sector comprises two WZW actions plus a potential depending on two scalar fields (thus, the first three terms of (\[ncsg12\]) correspond to the Lechtenfeld et al. proposal [@lechtenfeld] for NC sine-Gordon) and additional quartic terms coming from the current-current interactions. These quartic terms were also absent in the NCSG$_{1}$ model in the purely classical treatments of [@jhep2].
The upper and lower indices of a model defined for bi-fundamental matter can be interchanged since the metric is $\d_{AB}$ in the normalization $\mbox{tr}(T_{A}T_{B})=\frac{1}{2} \d_{AB}$ of the Lie algebra generators. This symmetry can be written manifestly using the group index notation as in (\[free0\]) or (\[mass\]) for the NCMT$_{1}$ model (\[ncmt0\]). Notice that the symmetry $i \leftrightarrow j$ which interchanges the two fundamental representations corresponds to the symmetry $g_{1} \leftrightarrow
g_{2}$ of the bosonic sector (\[ncsg12\]) (recall that $g_{1}\star g_{2}=g_{2}\star g_{1}$ ). The action (\[ncsg12\]) can be written in a more compact form using the Polyakov-Wiegmann identity and the representation (\[u1u1\]) for $g\in U(1)\times U(1)$ \_[bf]{}\[g\] &=& W\[g\] +d\^2x $$m \a_{0}\, \(\s_{1} g^{-2}
\s_{1}\star g^{2}\) + \frac{\l_{bf}}{32\pi^2} \, \(
g^{-1}\star\pa_{+}g\star g \star \pa_{-}g^{-1}\)$$ \[ncsg121\].
This representation is more convenient in order to compare it to its analog model obtained from reduction of the parent action, the $GL(2, C)$ NC affine Toda model coupled to matter (Dirac) fields [@jhep2]. Except for the last quartic term, this is the NCSG$_{1}$ model defined in [@jhep2; @lechtenfeld]. Recall that the theories NCSG$_{1, 2}$ defined in [@jhep2] correspond to the Lechtenfeld et al. and Grisaru-Penati models, respectively. The Grisaru-Penati model has the potential $(e^{i\phi}_{\star}+
e^{-i\phi}_{\star})$ corresponding to one complex scalar field, instead of the potential (\[potential\]) defined for two fields of the Lechtenfeld et al. model. Strictly speaking, in our constructions above the fields $g_{i} \in U(1)_{C}$ (complexified $U(1)$), because the determinants were computed in Euclidean space. This is the case in ordinary commutative space, in which the complexification of $U(1)$, $U(1)_{C}$ is considered [@naculich]. In the construction of the model NCSG$_{1}$ we may consider the abelian subgroup of $GL(2,C)$, $U(1)_{C}\times
U(1)_{C}$ to which the fields $g_{1}$ and $g_{1}$ belong, respectively [@jhep2]. Nevertheless, it is a fact that the NCSG$_{1}$ model possesses real soliton type solutions for the field $\L_{+}(= \L_{1}+\L_{2})$, whereas $\L_{-}(= \L_{1}-\L_{2})$ remains as a decoupled free field [@jhep2; @lechtenfeld].
Conclusions and discussions {#conclusion}
===========================
The NC bosonization of the NCMT$_{1}$ model and some properties emerging from this process have been considered. In order to bosonize the current-current interactions we have used the bosonization dictionary (\[bosonimatrix\]) established for the relevant currents of a bi-fundamental free fermion. In order to bosonize the mass term we have used the axial transformation rules that the fields $\L_{i}$ must undergo in order to maintain these symmetries manifest during the bosonization process of a free fermion as developed in NC space [@nunez], following the ordinary space approach of [@Burgess1]. Even though the mass term breaks the both axial symmetries the axial transformation rules remain the same, thus allowing to construct the boson sector of the fermion mass term. The bosonized mass term reproduces the Lechtenfeld et al. proposal for the NC sine-Gordon potential [@lechtenfeld; @jhep2]. It is uncovered the duality relationship between the coupling constants of the both sectors (\[duality\]), as well as certain relationships between the couplings for the models with bi-fundamental, (anti-)fundamental fermion representations and their bosonic sectors and the ordinary MT/SG couplings (\[relationships\]). The global $U(1)$ charge of the fermionic sector (\[charge\]) defined for either, the current $j^{(1)\,\mu}$ or $j^{(2)\,\mu}$, corresponds to the topological charge (\[topo\]) defined for the field $\Phi_{+}(\sim \L_{+})$. It is shown that the auxiliary field $\L_{-}$ decouples in the commutative limit. However the deformed situation requires the both scalar fields $\L_{\pm}$.
Disregarding for the moment the quantum nature of the terms, such as certain normal ordering prescriptions, the bosonic (\[ncsg12\])\[or (\[ncsg121\])\] and fermionic (\[ncmt0\]) models resemble to their classical counterparts NCSG$_{1}$ and NCMT$_{1}$, respectively, as defined in [@jhep2](except for the last quartic term in (\[ncsg121\]). A remarkable feature of the NCSG$_{1}$ model, as well as some NC integrable systems [@dimakis], is that for $1-$(anti)soliton type solution the $\star-$products in its equations of motion collapse to ordinary ones since one has in general $f(x-v t) \star g(x-v t)= f(x-v t)
g(x-v t)$. Then the $\star-$products for these type of fields in the equations of motion of the model (\[ncsg121\]) reduce to the ordinary ones. In particular, the contributions from the last quartic term in (\[ncsg121\]) are only of the type emerging from the kinetic terms of the action. So, we may say that regarding the classical $1-$(anti)soliton type solution, the bosonized model (\[ncsg121\]) and the classical NCSG$_{1}$ theories are very similar. This is in contrast to the usual MT/SG duality in which the corresponding classical and quantum Lagrangians look very similar, apart from the field renormalizations and the relevant quantum corrections to the coupling constants.
Various aspects of the bosonized model derived above deserve attention in future research, e.g. the properties of the S-matrix, the NC zero-curvature formulation and integrability properties, the NC multi-soliton spectrum and their scattering properties. The ordinary equivalence SG/MT has been used to understand some confinement mechanism in QCD$_{2}$ with one flavor and N colors [@nucl1], on this regard it would be interesting to consider the NC situation. Moreover, the multi-fermion extensions of the NCMT$_{1}$ type model also deserve attention, e.g. in connection to multi-flavor noncommutative QCD$_{2}$. In ordinary space the correspondence between the multi-fermion massive Thirring and multi-boson sine-Gordon models related to any (untwisted) affine Lie algebra has been considered in [@jhep1] on the classical level. The classical correspondence between three-fermion massive Thirring and three-boson sine-Gordon has been addressed in [@jmp] and the quantum field theory aspects studied in [@epjc] through bosonization techniques. It would be interesting to address their NC extensions.
0.4in
[*Acknowledgments*]{}
0.2in The author thanks H.L. Carrion and M. Rojas for collaboration in a previous work and to ICET and Prof. M. C. Araújo at the Mathematics Department-UFMT for hospitality. This work has been supported by CNPq-FAPEMAT.
[\*\*]{} N. Seiberg and E. Witten, ;\
A. Connes, M. R. Douglas and A. Schwarz, 003;\
M. R. Douglas and C. M. Hull, 008. O. Lechtenfeld, A. D. Popov and B. Spendig, . K. Takasaki, . O. Lechtenfeld, L. Mazzanti, S. Penati, A. D. Popov, L. Tamassia, . M. T. Grisaru and S. Penati, . M.T. Grisaru, L. Mazzanti, S. Penati and L.Tamassia, . I. Cabrera-Carnero and M. Moriconi, . H. Blas, H.L. Carrion and M. Rojas, ; hep-th/0502051. S. Coleman, ;\
S. Mandelstam, . H. Blas, ;\
H. Blas and B.M. Pimentel, . L.A. Ferreira, J.L. Gervais, J. Sanchez Guillen and M.V. Savelev, . H. Blas, . C. Nunez, K. Olsen and R. Schiappa, . E. F. Moreno and F. A. Schaposnik, . E. F. Moreno and F. A. Schaposnik, . C. M. Naon, . C. P. Burgess and F. Quevedo, . C. P. Burgess and F. Quevedo, . J. C. Le Guillou, E. Moreno, C. Nunez and F. A. Schaposnik, . E. Witten, . O. J. Ganor, G. Rajesh and S. Sethi, ;\
R. Gopakumar, J. M. Maldacena, S. Minwalla and A. Strominger, ;\
J. X. Lu, S. Roy and H. Singh, ;\
R. G. Cai and N. Ohta, ; ;\
R.G. Cai, J.-X. Lu and N. Ohta, . N. Seiberg, L. Susskind and N. Toumbas, .\
J. Gomis, T. Mehen, . C-S. Chu, J. Lukierski, W.J. Zakrzewski, . K. Furuta and T. Inami, . J. M. Gracia-Bondia and C. P. Martin, . Y. Liao and K. Sibold, . S. Terashima, . T. Nakajima, 065014. L. Bonora and A. Sorin, . C. P. Martin, . F. Ardalan and N. Sadooghi, ; . A. Dimakis and F. Mueller-Hoissen, [*A noncommutative version of the nonlinear schroedinger equation*]{}, [*hep-th*]{}/0007015. S. J. Orfanidis, ;\
S. J. Orfanidis and R. Wang, . S. G. Naculich and H. J. Schnitzer, ;\
R. E. Gamboa Saravi, F. A. Schaposnik and J. E. Solomin, . H. Blas and L.A. Ferreira, ;\
H. Blas, . J. Acosta, H. Blas, , see also [*hep-th*]{}/0407020; H. Blas, , see also [*hep-th*]{}/0409269.
[^1]: Here we consider the fermionic analog of the (bosonic) model NCMT$_{1}$ of Ref. [@jhep2].
[^2]: There exists another prescription to bosonize the current-current coupling using the “completing the square" type Hubbard-Stratonovich identity which also holds in the NC case [@moreno2].
[^3]: It is expected that by formally setting $\l_{f}\equiv \l_{\bar{f}},\,\b^2_{f}\equiv \b^2_{\bar{f}}$ in (\[dual01\]) one can obtain the same duality relationship for fundamental fermion.
[^4]: The bosonization rules (\[bosoni1\])-(\[bosoni2\]) and the Noether charge definition (\[charge\]) might suggest for $j_{top}^{\mu}$ to be some combinations of the type $e^{\pm i\L_{j}}_{\star} \star \pa_{\pm}
e^{\mp i\L_{j}}_{\star}, \,(j=1,2)$. However, it is a fact that only the field $\Phi_{+}(\sim \L_{+})$ appears in the commutative limit of the potential and since the $Q_{top}$ is related to a discrete symmetry of the potential it is plausible to define it as in (\[topo\]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Tree-grass coexistence in savanna ecosystems depends strongly on environmental disturbances out of which crucial is fire. Most modeling attempts in the literature lack stochastic approach to fire occurrences which is essential to reflect their unpredictability. Existing models that actually include stochasticity of fire are usually analyzed only numerically. We introduce new minimalistic model of tree-grass coexistence where fires occur according to stochastic process. We use the tools of linear semigroup theory to provide more careful mathematical analysis of the model. Essentially we show that there exists a unique stationary distribution of tree and grass biomasses.'
address:
- 'Department of Biomathematics, Institute of Mathematics, University of Silesia in Katowice, Bankowa 14, 40-007 Katowice'
- 'Institute of Mathematics, Polish Academy of Sciences, Bankowa 14, 40-007 Katowice'
author:
- Paweł Klimasara
- 'Marta Tyran-Kamińska'
title: 'A model for random fire induced tree-grass coexistence in savannas'
---
[^1]
Introduction
============
Savanna covers around 20% of the Earth’s land surface. It is a mixed woodland-grassland ecosystem with canopy open enough to support the existence of continuous herbaceous layer dominated by grass. In order to find the explanation of such tree-grass codominance many theoretical models were introduced. Beside interspecies competition (e.g. [@eagleson1985water]), this coexistence is believed to have been driven by various environmental disturbances, primarily rainfall (e.g. [@ursino2014eco], [@synodinos2018impact]), grazing and browsing (e.g. [@bodini2016vegetation]), and fire [@sankaran2008woody]. Some models consider additional factors like competition of tree seedlings with grass [@baudena2010idealized] or varying flammability of trees [@beckage2009vegetation]. From the mathematical point of view, models containing many different factors lack stochasticity and differ in methodology (see e.g. loop analysis for graphs in [@bodini2016vegetation] or models based on impulsive differential equations [@tamen2016tree], [@yatat2017impulsive]).
Realistically, the appearance of fire is stochastic and its frequency can vary significantly [@archibald2009limits]. Usually studies with stochastic fire focus on numerical analysis (see e.g. [@d2006probabilistic], [@baudena2010idealized], [@beckage2011grass], [@de2014tree], and [@synodinos2018impact]). We introduce a simple model where fire occurrences are stochastic and study it in terms of linear semigroup theory. We find that biomasses of grass and trees have a unique stationary distribution and hence this simple model can describe stable savannas driven by stochasticity of fires.
Model description
=================
Our model is based on a simplified version of the system of differential equations given in [@beckage2011grass], but instead of putting fire disturbances inside these equations we introduce appropriate stochastic process separately. Similarly to cited authors we consider only amounts of tree and grass biomasses, fires are events discrete in time, and the strength of grass-fire feedback depends on biomass of grass.
In the absence of fires we represent the dynamics of tree biomass $W$ and grass biomass $G$ (both in $\frac{\text{mass}}{\text{area}}$ units) according to the competition model $$\label{no-fire}
\left\{\begin{array}{lll}
W'(t)=r_w W(t)\left(1-\frac{W(t)}{K_w}\right),\\ G'(t)=r_q G(t)\left(1-\frac{G(t)}{K_g}-\frac{W(t)}{K_w}\right),
\end{array} \right.$$ where $r_w, r_g$ are the growth rates and $K_w,K_g$ are the carrying capacities for tree and grass biomasses. It is easily seen that has three stationary states: $(0,0)$, $(K_w,0)$ and $(0,K_g)$. Moreover, the point $(K_w,0)$ is locally stable, while the points $(0,0)$ and $(0,K_g)$ are unstable. So the system of equations (\[no-fire\]) provides a deterministic description of the change of wood and grass biomasses in time where in the long time, due to species competition, the system will end up as a woodland. The solution curves for the system have the qualitative behavior as shown in Figure \[fig-phase\].
![Phase portraits for with parameter values $r_w=0.08$, $r_g=1.5$ (left-hand panel) and $r_w=0.25$, $r_g=0.5$ (right-hand panel), $K_w=K_g=1$[]{data-label="fig-phase"}](mphase-1 "fig:"){height="5cm"} ![Phase portraits for with parameter values $r_w=0.08$, $r_g=1.5$ (left-hand panel) and $r_w=0.25$, $r_g=0.5$ (right-hand panel), $K_w=K_g=1$[]{data-label="fig-phase"}](mphase-2 "fig:"){height="5cm"}
Instead of using actual amount of biomasses we will relate in our model to ratios of these amounts to maximal capacities of wood and grass, respectively: $$w(t)=\frac{W(t)}{K_W},\quad g(t)=\frac{G(t)}{K_G}.$$ Thus $w(t)$ and $g(t)$ take values in the unit interval, i.e. $0\leqslant w(t), g(t)\leqslant1$ for any time $t$. We now allow disturbances of the growth of biomasses due to fires occurring at random times $(t_n)_{n\geqslant 1}$. Let $t_0=0$ and $w(t_0)=w_0$, $g(t_0)=g_0$, where $w_0,g_0\in [0,1]$ are arbitrary. In periods between fire occurrences the growth of normalized tree and grass biomasses is modeled with $$\label{basic.eq}
\left\{\begin{array}{lll}
w'(t)=r_w w(t)\big(1-w(t)\big),\\
g'(t)=r_g g(t)\big(1-g(t)-w(t)\big),
\end{array} \right.$$ for $t\in(t_{n}, t_{n+1})$, $n{\geqslant}0$, and a sequence of random variables $(\tau_n)$ such that $$\label{tau}
\left\{\begin{array}{lll}
t_{n+1}=t_{n}+\tau_{n+1},\\
\Pr\big(\tau_{n+1}>t|w(t_{n})=w_{n}, g(t_{n})=g_{n}\big)=e^{-\int_0^t \lambda(\pi_s(w_n,g_n))ds},
\end{array} \right.$$ where $\pi_t(w_n,g_n)=(w(t),g(t))$ is the solution of (\[basic.eq\]) with initial condition $(w_n,g_n)$ and $\lambda$ is a nonnegative bounded continuous function. At each time $t_{n+1}$ the loss of biomasses is given by $$\label{loss.eq}
\left\{\begin{array}{lll}
w(t_{n+1})=w(t_{n+1}^-)-M_w \, w(t_{n+1}^-),\\
g(t_{n+1})=g(t_{n+1}^-)-M_g \, g(t_{n+1}^-),\quad n{\geqslant}0,
\end{array} \right.$$ where $ M_w,M_g\in (0,1)$ are constants, $v(t^-)=\lim_{s\rightarrow t^-}v(s)$ for $v\in\{w, g\}$. We assume that the function $\lambda\colon [0,1]^2\to \mathbb{R}_+$ satisfies $$\label{lambda}
\lambda(w,0)=0, \quad w{\geqslant}0,\quad \lambda(w,g)>0\quad \text{for } w\geqslant 0,\ g>0.$$ In Figure \[fire-no-fire\] we display graphs of wood and grass biomasses in time, without and with fires. A sample behavior of the overall system in the long run including losses due to random fires is shown in Figure \[wg.fire\].
![Graphs of system (left-hand panel) and of system – (right-hand panel) with parameter values $r_w=0.25$, $r_g=0.5$, $M_w=0.4$, $M_g=0.1$, $\lambda(w,g)=g$ and initial condition $w_0=0.1,\ g_0=0.2$[]{data-label="fire-no-fire"}](wood-grass "fig:"){height="5cm"} ![Graphs of system (left-hand panel) and of system – (right-hand panel) with parameter values $r_w=0.25$, $r_g=0.5$, $M_w=0.4$, $M_g=0.1$, $\lambda(w,g)=g$ and initial condition $w_0=0.1,\ g_0=0.2$[]{data-label="fire-no-fire"}](wood-grass-fire1 "fig:"){height="5cm"}
![Sample trajectories of the stochastic process in – with parameter values $r_w=0.25$, $r_g=0.5$, $M_w=0.4$, $M_g=0.1$, $\lambda(w,g)=g$ and initial condition $w_0=0.01$, $g_0=0.2$ []{data-label="wg.fire"}](savanna1 "fig:"){height="5cm"} ![Sample trajectories of the stochastic process in – with parameter values $r_w=0.25$, $r_g=0.5$, $M_w=0.4$, $M_g=0.1$, $\lambda(w,g)=g$ and initial condition $w_0=0.01$, $g_0=0.2$ []{data-label="wg.fire"}](savanna2 "fig:"){height="5cm"}
The process $\xi(t)=(w(t),g(t))$, $t{\geqslant}0$, with $w,g$ as in –, is a piecewise deterministic Markov process ([@davis84]) with state space $[0,1]^2$. It is an example of a flow with jumps as presented in [@rudnickityran17 Section 4.2.4]. We describe the jumps of the stochastic process by a linear transformation $S$ mapping $(w,g)\longmapsto S(w,g)$, where $$\label{d:S}
S(w,g)=((1-M_w)w,(1-M_g)g),\quad (w,g)\in [0,1]^2.$$ Let $p(t,w,g)$ be the probability density of $(w(t),g(t))$, i.e. $p$ is nonnegative, Borel measurable, and satisfies $$\Pr((w(t),g(t))\in B)=\int_B p(t,w,g)dwdg$$ for any Borel subset of $[0,1]^2$ with the integral being equal to one for $B=[0,1]^2$. Then $p$ is a solution of the following Fokker-Planck type equation $$\begin{gathered}
\label{fokk.pla.eq}
\frac{\partial p(t,w,g)}{\partial t}+\frac{\partial\big(r_ww(1-w)p(t,w,g)\big)}{\partial w}+\frac{\partial\big(r_gg(1-g-w)p(t,w,g)\big)}{\partial g}\\=-\lambda(w,g)p(t,w,g)
+\frac{\lambda\big(S^{-1}(w,g))\big)p\big(t, S^{-1}(w,g)\big)}{(1-M_w)(1-M_g)},
$$ where $S^{-1}$ is the inverse of the transformation $S$ defined in . Equation is supplemented with initial condition $$\label{fokk.pla.eq.ic}
p(0,w,g)=f(w,g),\quad \text{where } \int_{0}^1\int_0^1 f(w,g)\,dwdg=1$$ and $f$ is a nonnegative Borel measurable function, so that $f$ is the probability density of $(w(0),g(0))$. We have the following result - its proof will be given in the next section.
\[main\] There exists a unique density $p_*(w,g)$ which is a stationary solution of . Moreover, every solution of – converges to $p_*$, i.e. $$\lim_{t\rightarrow\infty}\int_0^1\int_0^1 \big|p(t,w,g)-p_*(w,g)\big|dwdg=0.$$
\[r:boun\] Let $w_0,g_0\in [0,1]$ and $w(t),g(t)$ be as in –. If $w_0>0$ and $g_0=0$ then $g(t)=0$ for all $t>0$. In this case, assumption implies that fire can not occur when there is no grass biomass. Hence, $w$ is defined for all $t$ as the solution of the differential equation $w'(t)=r_w w(t)(1-w(t))$ with initial condition $w(0)=w_0$. Thus $w(t)>0$ for all $t>0$ and $w(t)$ converges to $1$ as $t\to \infty$. Consequently, the point measure $\delta_{\{(1,0)\}}$ is an invariant measure for the process $\xi$. Similarly, if $w_0=0$ and $g_0=0$ then $w(t)=0$ and $g(t)=0$ for all $t> 0$. Thus also the point measure $\delta_{\{(0,0)\}}$ is an invariant measure for the process $\xi$. Finally, if $w_0=0$ and $g_0>0$ then $w(t)=0$ and $g(t)>0$ for all $t{\geqslant}0$. In this case, the process $\xi$ has an invariant distribution which is a product of $\delta_{\{0\}}$ and an absolutely continuous measure, see Remark \[r:inv\].
If instead of we have $t_{n+1}=t_n+\tau$, $n{\geqslant}0$ where $\tau$ is a constant then such a model is an example of an impulsive system [@tamen2016tree; @yatat2017impulsive].
Existence and uniqueness of tree and grass biomasses distribution
=================================================================
Methods in this section are mostly taken from the book [@rudnickityran17]. To prove Theorem \[main\] we use the method from [@rudnickityran17 Section 6.3.2]. We begin by recalling some notions for stochastic semigroups. Let the triple $(X,\Sigma,m)$ be a $\sigma$-finite measure space. Denote by $D$ the subset of the space $L^1=L^1(X,\Sigma,m)$ which contains all densities $$D=\{f\in L^1: f{\geqslant}0,\; \|f\|=1\}.$$ A linear mapping $P\colon L^1\to L^1$ is called a *Markov* or *stochastic operator* if $P(D)\subset D$. A family $\{P(t)\}_{t{\geqslant}0}$ of stochastic operators which satisfies conditions:
1. $P(0)=\operatorname{id}$, $P(t+s)=P(t) P(s)$ for $s,\,t{\geqslant}0$,
2. for each $f\in L^1$ the function $t\mapsto P(t)f$ is continuous,
is called a *stochastic semigroup*.
Consider a stochastic semigroup $\{P(t)\}_{t{\geqslant}0}$. A density $f_*$ is called [*invariant*]{} if $P(t)f_*=f_*$ for each $t>0$. The stochastic semigroup $\{P(t)\}_{t{\geqslant}0}$ is called [*asymptotically stable*]{} if there is an invariant density $f_*$ such that $$\lim _{t\to\infty}\|P(t)f-f_*\|=0 \quad \text{for}\quad
f\in D.$$
We will use a result of Pichór and Rudnicki [@pichorrudnicki18] (see also [@rudnickityran17 Theorem 5.6]) which requires the following conditions:
1. For every $y_0\in X$ there exist $\varepsilon >0$, $t>0$, and a measurable function $\eta{\geqslant}0$ such that $\int \eta(x)\, m(dx)>0$ and $$\label{w-eta2}
P(t)f(x){\geqslant}\eta(x)\int_{B(y_0,\varepsilon)} f(y)\,m (dy),$$ where $B(y_0,\varepsilon)=\{y\in X:\,\,\rho(y,y_0)<\varepsilon\}$.
2. There exists a point $x_0\in X$ such that for each $\varepsilon >0$ and for each density $f$ we have $$\label{k:WI}
\int\limits_{B(x_0,\varepsilon)} P(t)f(x)\,m(dx)>0\quad\textrm{for some $t=t(\varepsilon,f)>0$}.$$
3. There exists $\kappa>0$ such that $$\label{k:T}
\sup\limits_{F\in \mathcal F}\limsup_{t\to\infty} \int_F P(t)f(x)\,m(dx){\geqslant}\kappa$$ for $f\in D_0$, where $D_0$ is a dense subset of $D$ and $\mathcal F$ is the family of all compact subsets of $X$.
\[cor:st\] Let $\{P(t)\}_{t{\geqslant}0}$ be a stochastic semigroup on $L^1(X,\Sigma,m)$, where $X$ is a separable metric space, $\Sigma$ is the $\sigma$-algebra of Borel subsets of $X$, and $m$ is a $\sigma$-finite measure. Assume that $\{P(t)\}_{t{\geqslant}0}$ satisfies conditions [(K)]{}, [(WI)]{}, and [(WT)]{}. Then the semigroup $\{P(t)\}_{t{\geqslant}0}$ is asymptotically stable.
Now we are ready to prove the main theorem.
Let $X=(0,1]^2$ and $m$ be the two-dimensional Lebesgue measure on $X$. It follows from [@rudnickityran17 Section 4.2.4] that the process $\xi(t)$, $t{\geqslant}0$, induces a stochastic semigroup $\{P(t)\}_{t{\geqslant}0}$ on $L^1=L^1(X,\Sigma,m)$ and that the solution of – is given by $p(t,w,g)=P(t)f(w,g)$, $t{\geqslant}0$, $(w,g)\in X$. To apply Theorem \[cor:st\] we need to check conditions (K), (WI), and (WT).
We first show that condition (WT) holds. The extended generator $\widetilde{L}$ of the process $\xi$ is of the form $$\widetilde{L}V(x)=\langle b(x),\mathrm{grad}V(x)\rangle +\varphi(x)(V(S(x))-V(x))\quad \text{for }x=(w,g),$$ where $\mathrm{grad}V(x)$ is the gradient of $V(x)$ and $b(x)$ is the vector with coordinates $$b_1(x)=r_w w(1-w),\quad b_2(x)=r_gg(1-w-g), \quad x=(w,g).$$ The domain $\mathcal{D}(\widetilde{L})$ of the extended generator $\widetilde{L}$ (see [@davis84] or [@rudnickityran17 Section 2.3.6]) contains the set of functions $V\colon X\to \mathbb{R}$ such that for each $x\in X$ the function $t\mapsto V(\pi_t(x))$ is absolutely continuous and for each $t{\geqslant}0$, $x\in X$, we have $$\mathbb{E}\Big(\sum_{t_n{\leqslant}t}\big|V(\xi(t_n))-V(\xi(t_n^{-}))\big|\Big|\xi(0)=x\Big)<\infty.$$ Let $V(w,g)=-\log w -\log g$. Since we have $V(\xi(t_n))-V(\xi(t_n^{-}))=-\log(1-M_w)-\log(1-M_g)$ for any $n$, we see that $V$ belongs to $\mathcal{D}(\widetilde{L})$ and that $$\widetilde{L}V(w,g)=-r_w (1-w)-r_g(1-w-g)-\lambda(w,g)(\log(1-M_w)+\log(1-M_g)).$$ The function $\widetilde{L}V$ is bounded on $(0,1]^2$ and $\widetilde{L}V(w,g)\to -r_w-r_g$ as $\|(w,g)\|\to 0$, where $\|\cdot\|$ denotes a norm in $\mathbb{R}^2$. Thus we can find a $\delta\in (0,1)$ such that $\widetilde{L}V(w,g){\leqslant}-(r_w+r_g)/2$ for $\|(w,g)\|< \delta$. Moreover, we have $$\int_X \widetilde{L}V(x)f(x)m(dx)=\int_X V(x)Af(x)m(dx),\quad f\in \mathcal{D}(A)\cap \mathcal{D}_V,$$ where $f\in D_V$ iff $\int_{X}V(x)|f(x)|m(dx)<\infty$ and $(A,\mathcal{D}(A))$ is the generator of the semigroup $\{P(t)\}_{t{\geqslant}0}$. We conclude that $V$ is a Hasminskiĭ function for the semigroup $\{P(t)\}_{t{\geqslant}0}$ and the compact set $F=\{(w,g)\in X: \|(w,g)\|{\geqslant}\delta\}$ implying that condition (WT) holds, by [@rudnickityran17 Corollary 5.8].
To check condition (K) take $x_0\in X$ and define $x_1=S(x_{0})$, $x_2=S(x_1)$, $$\label{a-cs-1}
v_1=S'(x_{1})S'(x_{0})b(x_{0})
-b(x_{2}),\quad v_2=S'(x_{1})b(x_{1})
-b(x_{2}).$$ Since $S$ is a linear transformation, we have $$v_1=S^2(b(x_{0}))
-b(S^2(x_{0})),\quad v_2=S(b(S(x_{0}))
-b(S^2(x_{0})),$$ where $S^2(x)=S(S(x))$. It is easily seen that vectors $v_1$ and $v_2$ are linearly independent for each $x_0\in X$. Since the function $\lambda$ is strictly positive on $(0,1]^2$, we conclude that condition (K) holds (see e.g. [@rudnickityran17 Section 6.3.2] or [@biedrzyckatyran Section 4]).
Observe that condition (WI) holds once we show that there exists $x_0$ such that for each $\varepsilon>0$ and $x\in X$ we can find $n$ and times $s_1,\ldots,s_n,s_{n+1}>0$ such that $\pi_{s_{n+1}}(x_n)\in B(x_0,\varepsilon)$, where $$\label{d:xn}
x_{n}=S(\pi_{s_{n}}(\ldots S(\pi_{s_1}x))).$$ The point $(0,1)$ is a saddle point for the two-dimensional system considered on $\mathbb{R}^2$. Its stable manifold is the set $\{(0,g): g>0\}$ and its unstable manifold contains a curve joining the point $(0,1)$ with the stable point $(1,0)$, see Figure \[fig-phase\]. Let us take $x_0\in X$ from this curve lying close to the point $(1,0)$. For any point $x\in X$ we can find $n$ and $s_1,\ldots,s_n>0$ such that $x_{n}$ defined as in is as close to $(0,0)$ as is needed. Since $\pi_s y\to (1,0)$ for $y\in (0,1)^2$, we can find $s_{n+1}$ such that $\pi_{s_{n+1}} x_n\in B(x_0,\varepsilon)$, which completes the proof.
\[r:inv\] The process $\xi$ restricted to the set $\{(0,g):g\in(0,1]\}$, considered with measure being the product of $\delta_{\{0\}}$ and the Lebesgue, induces a stochastic semigroup on $L^1(\{0\}\times(0,1])$. Using the same type of argument as in the proof of Theorem \[main\] it can be shown that this semigroup satisfies conditions (K), (WI) and (WT), thus this semigroup is asymptotically stable, implying the existence of the invariant measure mentioned in Remark \[r:boun\].
Discussion
==========
We showed that there exists unique, absolutely continuous with respect to the two-dimensional Lebesgue measure, stationary distribution for positive amount of grass and wood biomasses. The stationary density is strictly positive in the region bounded by the axes and the unstable manifold of the point $(0,1)$, in particular in a neighbourhood of the line $\{(w,1-w):w\in (0,1]\}$, showing that the coexistence of trees and grass is possible. Finding the actual shape of this distribution, numerical analysis, and further improvements of the model by adding more coefficients reflecting real-world factors regulating savanna biomasses we leave for future work.
Moreover such analysis can be implemented in models describing different phenomena involving random fires, such as impact of forest fires on population of pines and bark beetles. Modeling attempts usually are deterministic (see e.g. [@chen2014model]) and hence could benefit from involving stochastic nature of fire.
[10]{} S. Archibald, D. P. Roy, V. Wilgen, W. Brian, R. J. Scholes, What limits fire? An examination of drivers of burnt area in Southern Africa, Global Change Biol. 15 (3) (2009) 613–630.
M. Baudena, F. D’Andrea, A. Provenzale, An idealized model for tree–grass coexistence in savannas: the role of life stage structure and fire disturbances, J. Ecol. 98 (1) (2010) 74–80.
B. Beckage, L. J. Gross, W. J. Platt, Grass feedbacks on fire stabilize savannas, Ecol. Model. 222 (14) (2011) 2227–2233.
B. Beckage, W. J. Platt, L. J. Gross, Vegetation, fire, and feedbacks: a disturbance-mediated model of savannas, Am. Nat. 174 (6) (2009) 805–818.
W. Biedrzycka, M. Tyran-Kamińska, Existence of invariant densities for semiflows with jumps, J. Math. Anal. Appl. 435 (1) (2016) 61–84.
A. Bodini, N. Clerici, Vegetation, herbivores and fires in savanna ecosystems: A network perspective, Ecol. Complex. 28 (2016) 36–46.
B. Chen-Charpentier, M. Leite, A model for coupling fire and insect outbreak in forests, Ecol. Model. 286 (2014) 26–36.
M. H. A. Davis, Piecewise-deterministic [M]{}arkov processes: a general class of nondiffusion stochastic models, J. Roy. Statist. Soc. Ser. B 46 (3) (1984) 353–388.
C. De Michele, F. Accatino, Tree cover bimodality in savannas and forests emerging from the switching between two fire dynamics, PloS one 9 (3) (2014) e91195.
P. D’Odorico, F. Laio, L. Ridolfi, A probabilistic analysis of fire-induced tree-grass coexistence in savannas, Am. Nat. 167 (3) (2006) E79–E87.
P. S. Eagleson, R. I. Segarra, Water-limited equilibrium of savanna vegetation systems, Water Resour. Res. 21 (10) (1985) 1483–1493.
K. Pichór, R. Rudnicki, Stability of stochastic semigroups and applications to [S]{}tein’s neuronal model, Discrete Contin. Dyn. Syst. Ser. B 23 (1) (2018) 377–385.
R. Rudnicki, M. Tyran-Kamińska, Piecewise deterministic processes in biological models, Springer Briefs in Applied Sciences and Technology, [S]{}pringer Briefs in Mathematical Methods, Springer, Cham, 2017.
M. Sankaran, J. Ratnam, N. Hanan, Woody cover in african savannas: the role of resources, fire and herbivory, Global Ecol. Biogeogr. 17 (2) (2008) 236–245.
A. D. Synodinos, B. Tietjen, D. Lohmann, F. Jeltsch, The impact of inter-annual rainfall variability on african savannas changes with mean rainfall, J. Theor. Biol. 437 (2018) 92–100.
A. T. Tamen, Y. Dumont, J. J. Tewa, S. Bowong, P. Couteron, Tree–grass interaction dynamics and pulsed fires: Mathematical and numerical studies, Appl. Math. Model. 40 (11-12) (2016) 6165–6197.
N. Ursino, Eco-hydrology driven fire regime in savanna, J. Theor. Biol. 355 (2014) 68–76.
V. Yatat, P. Couteron, J. J. Tewa, S. Bowong, Y. Dumont, An impulsive modelling framework of fire occurrence in a size-structured model of tree–grass interactions for savanna ecosystems, J. Math. Biol. 74 (6) (2017) 1425–1482.
[^1]: This work was partially supported by the Polish NCN grant 2017/27/B/ST1/00100 and by the grant 346300 for IMPAN from the Simons Foundation and the matching 2015-2019 Polish MNiSW fund.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Let $\g = \n\oplus\h\oplus\n_-$ be a simple Lie algebra over ${\mathbb{C}}$ of type $A, D, E$, and let $U_q(L\g)$ be the associated quantum loop algebra. Following Nakajima [@N], Varagnolo-Vasserot [@VV], and the first author [@H1], we study a $t$-deformation $\K_t$ of the Grothendieck ring of a tensor category $\CC_{\mathbb{Z}}$ of finite-dimensional $U_q(L\g)$-modules. We obtain a presentation of $\K_t$ by generators and relations.
Let $Q$ be a Dynkin quiver of the same type as $\g$. Let $DH(Q)$ be the derived Hall algebra of the bounded derived category $D^b({\mathrm{mod}}(FQ))$ over a finite field $F$, introduced by Toën [@T]. Our presentation shows that the specialization of $\K_t$ at $t=\sqrt{|F|}$ is isomorphic to $DH(Q)$. Under this isomorphism, the classes of fundamental $U_q(L\g)$-modules are mapped to scalar multiples of the classes of indecomposable objects in $DH(Q)$.
Our presentation of $\K_t$ is deduced from the preliminary study of a tensor subcategory $\CC_{Q}$ of $\CC_{\mathbb{Z}}$ analogous to the heart $\mod(FQ)$ of the triangulated category $D^b({\mathrm{mod}}(FQ))$. We show that the $t$-deformed Grothendieck ring $\K_{t,\,Q}$ of $\CC_{Q}$ is isomorphic to the positive part of the quantum enveloping algebra of $\g$, and that the basis of classes of simple objects of $\K_{t,\,Q}$ corresponds to the dual of Lusztig’s canonical basis. The proof relies on the algebraic characterizations of these bases, but we also give a geometric approach in the last section.
It follows that for every orientation $Q$ of the Dynkin diagram, the category $\CC_{Q}$ gives a new categorification of the coordinate ring ${\mathbb{C}}[N]$ of a unipotent group $N$ with Lie algebra $\n$, together with its dual canonical basis.
author:
- 'D. Hernandez, B. Leclerc'
title: '**Quantum Grothendieck rings and derived Hall algebras**'
---
Introduction {#sect1}
============
[**1.1**]{}Let $\g$ be a simple Lie algebra of type $A, D, E$ over ${\mathbb{C}}$. We denote by $\g = \n\oplus \h\oplus\n_-$ a triangular decomposition of $\g$. Let $v$ be an indeterminate, and let $$U_v(\g) = U_v(\n)\otimes U_v(\h)\otimes U_v(\n_-)$$ be the corresponding Drinfeld-Jimbo quantum enveloping algebra over ${\mathbb{C}}(v)$, defined via a $v$-analogue of the Chevalley-Serre presentation of $U(\g)$. Using a geometric realization of $U_v(\n)$ in terms of perverse sheaves on varieties of representations of a quiver $Q$ of the same Dynkin type as $\g$, Lusztig [@Lu] has defined a canonical basis $\bB$ of $U_v(\n)$ with favorable positivity properties. This was inspired by a seminal work of Ringel [@Ri], showing that the twisted Hall algebra of the category $\mod(FQ)$ of representations of $Q$ over a finite field $F$, is isomorphic to the specialization of $U_v(\n)$ at $v=\sqrt{|F|}$.
[**1.2**]{}One can associate with $\g$ another quantum algebra. Let $L\g = {\mathbb{C}}[t,t^{-1}]\otimes \g$ be the loop algebra of $\g$. Let $q$ be a nonzero complex number, which is not a root of unity. Via a $q$-analogue of the loop presentation of $U(L\g)$, Drinfeld [@D] has defined the quantum loop algebra $U_q(L\g)$, an algebra over ${\mathbb{C}}$. The finite-dimensional representations of $U_q(L\g)$ have attracted a lot of attention, because of their connection with the trigonometric solutions of the quantum Yang-Baxter equation with spectral parameter. In this paper we focus on a tensor subcategory $\CC_{\mathbb{Z}}$ of the category of finite-dimensional $U_q(L\g)$-modules, whose simple objects are parametrized by a discrete set (for the precise definition of $\CC_{\mathbb{Z}}$ see [@HL §3.7] or §\[subsectCZ\] below). Denote by ${\mathcal{R}}$ the complexified Grothendieck ring of $\CC_{\mathbb{Z}}$. Let $t$ be another indeterminate. By works of Nakajima [@N] and Varagnolo-Vasserot [@VV], the ${\mathbb{C}}$-algebra ${\mathcal{R}}$ has an interesting $t$-deformation ${\mathcal{R}}_t$ over ${\mathbb{C}}(t)$. The first author [@H1] has introduced a slightly different deformation $\K_t$. These $t$-deformations are important because they contain for every simple object $L$ of $\CC_{\mathbb{Z}}$ a “class” $[L]_t$ which can be characterized by axioms similar to those of Lusztig for the canonical basis $\bB$. As a consequence, Nakajima [@N] has shown that one can calculate algorithmically the character of $L$.
[**1.3**]{}Surprisingly, these deformed Grothendieck rings have not been much studied from the ring theoretic point of view, and for instance, to the best of our knowledge, there is no available presentation by generators and relations in the literature. One of the main results of this paper (Theorem \[presentation\]) is a presentation of $\K_t$, with a similar flavor as the familiar Drinfeld-Jimbo presentation of $U_v(\n)$. More precisely, this presentation shows that $\K_t$ is obtained by taking an infinite number of copies of $U_t(\n)$ labelled by $m\in{\mathbb{Z}}$, and then imposing $t$-boson relations between generators of copies sitting at adjacent integers, and $t$-commutation relations between generators of non-adjacent copies.
[**1.4**]{}Let $D^b({\mathrm{mod}}(FQ))$ be the bounded derived category of ${\mathrm{mod}}(FQ)$. Toën [@T] has attached to this triangulated category an associative algebra called the derived Hall algebra of $D^b({\mathrm{mod}}(FQ))$ (see also [@XX]). Let $DH(Q)$ denote the twisted derived Hall algebra obtained by twisting Toën’s multiplication by means of the Ringel form, as in [@S]. It follows from our presentation of $\K_t$ that:
\[Th\_DHall\]
- The specialization of $\K_t$ at $t=\sqrt{|F|}$ is isomorphic to $DH(Q)$.
- Under this isomorphism, the classes of fundamental $U_q(L\g)$-modules are mapped to scalar multiples of the classes of indecomposable stalk complexes in $DH(Q)$, and the basis of classes of standard $U_q(L\g)$-modules is mapped to a rescaling of the natural basis of $DH(Q)$ indexed by isoclasses of objects of $D^b({\mathrm{mod}}(FQ))$.
There is a similar result for the $t$-deformed Grothendieck ring ${\mathcal{R}}_t$ of [@N; @VV], but the twisted derived Hall algebra should be replaced by a non-twisted one (Remark \[Remark\_untwistedDH\]).
[**1.5**]{}To obtain our presentation of $\K_t$ we first consider a tensor subcategory $\CC_{Q}$ of $\CC_{\mathbb{Z}}$ which “looks like ${\mathrm{mod}}(FQ)$ inside $D^b({\mathrm{mod}}(FQ))$”. Recall that in [@HL] we have introduced an increasing sequence $(\CC_\ell)_{\ell>0}$ of subcategories of $\CC_{\mathbb{Z}}$. When $Q$ is a bipartite orientation of the Dynkin diagram and the Coxeter number $h$ is even, $\CC_Q$ is just the subcategory $\CC_\ell$ with $\ell=h/2-1$. The general definition of $\CC_Q$ for an arbitrary orientation $Q$ will be given in §\[sectCh’\] below. Let $\K_{t,\,Q}$ be the subalgebra of $\K_t$ spanned by the elements $[L]_t$ associated with the simple objects $L$ of $\CC_Q$. Note that $\K_t$ and $\K_{t,\,Q}$ are algebras over ${\mathbb{C}}(t^{1/2})$, where $t^{1/2}$ is a square root of $t$.
The quantum algebra $U_v(\n)$ is endowed with a distinguished scalar product. Let $\bB^*$ be the basis of $U_v(\n)$ adjoint to the canonical basis $\bB$ with respect to this scalar product. Let $v^{1/2}$ be a square root of $v$, and set $\U_{v}(\n):={\mathbb{C}}(v^{1/2})\otimes U_v(\n)$. The main step for obtaining the presentation of $\K_t$ is:
\[mainTh\]
- There is a ${\mathbb{C}}$-algebra isomorphism $\Phi \colon \K_{t,\,Q} \overset{\sim}{\to} \U_{v}(\n)$ with $\Phi(t^{1/2}) = v^{1/2}$.
- For every simple object $L$ of $\CC_{Q}$, the image $\Phi([L]_t)$ belongs to $\bB^*$ (up to some half-integral power of $v$).
Nakajima obtained in [@Ncl] similar results for the first subcategory $\CC_1$ of [@HL]. Namely, he showed that the classes $[L]_t$ of simple objects of $\CC_1$ can be identified with a subset of the basis $\widetilde{\bB}^*$ of $U_v(\widetilde{\n})$. Here $\widetilde{\n}$ denotes the positive part of the Kac-Moody algebra of rank $2{\operatorname{rk}}(\g)$ attached to the decorated Dynkin diagram of $\g$, and $\widetilde{\bB}$ is Lusztig’s canonical basis of $U_v(\widetilde{\n})$. For example, if $\g$ has type $A_3$, $\widetilde{\g}$ has type $E_6$. Note that in Theorem \[mainTh\], we do not use $\widetilde{\n}$, but only $\n$.
[**1.6**]{}Let $A_v(\n)$ be the graded dual of the vector space $U_v(\n)$. It can be endowed with a multiplication coming from the comultiplication of $U_v(\g)$, and regarded as the quantum coordinate ring of the unipotent group $N$ with Lie algebra $\n$ (see [@GLS]). The basis $\bB^*$ can be identified with a basis of $A_v(\n)$ called the dual canonical basis. It specializes when $v\mapsto 1$ to a basis $\B$ of the coordinate ring ${\mathbb{C}}[N]$.
By specializing $v^{1/2}$ and $t^{1/2}$ to $1$ in Theorem \[mainTh\], we see that the complexified Grothendieck ring ${\mathcal{R}}_{Q}$ of $\CC_{Q}$ can be identified with the coordinate ring ${\mathbb{C}}[N]$ in such a way that the basis of ${\mathcal{R}}_{Q}$ consisting of the classes of simple objects becomes Lusztig’s dual canonical basis $\B$ of ${\mathbb{C}}[N]$. We can therefore state:
The tensor category $\CC_{Q}$ is a categorification of the ring ${\mathbb{C}}[N]$ and its dual canonical basis $\B$.
Note that, by work of Khovanov-Lauda [@KL], Rouquier [@R], and Varagnolo-Vasserot [@VV2], $({\mathbb{C}}[N], \B)$ has another categorification in terms of KLR-algebras. In type $A_n$, KLR-algebras are isomorphic to blocks of affine Hecke algebras, and the category $\CC_Q$ for an equi-oriented quiver $Q$ is related to a category of representations of affine Hecke algebras through the quantum affine Schur-Weyl duality. It would be interesting to find for other Dynkin quivers $Q$ similar functors between $\CC_{Q}$ and the module categories of the corresponding KLR-algebras.
[**1.7**]{}The first author [@H3] has shown that tensor products of simple objects of $\CC_{\mathbb{Z}}$ have the following remarkable property: a tensor product $L_1\otimes \cdots \otimes L_k$ of simple objects $L_i$ is simple if and only if for every pair $1\le i < j \le k$ the tensor product $L_i\otimes L_j$ is simple. Using Theorem \[mainTh\] this yields the following:
\[corollary\_binary\] A product $b_1\cdots b_k$ of elements $b_i$ of the dual canonical basis $\bB^*$ of $U_v(\n)$ belongs to $\bB^*$ up to a power of $v$ if and only if for every pair $1\le i < j \le k$ the product $b_i b_j$ belongs to $\bB^*$ up to a power of $v$.
Corollary \[corollary\_binary\] was expected in relation with the program of Berenstein and Zelevinsky [@BZ; @BZ2] of describing $\bB^*$ in terms of quantum cluster algebras. But it was only known in a few low rank cases.
[**1.8**]{}Theorem \[mainTh\] also gives new supporting evidence for some conjectures formulated in [@GLS] and [@HL]. It was conjectured in [@HL §13] that for every $\ell \in {\mathbb{N}}$, the Grothendieck ring ${\mathcal{R}}_\ell$ of $\CC_{\ell}$ has a particular cluster algebra structure for which all cluster monomials are classes of simple objects. In [@GLS], it is shown that $U_v(\n)$ has a quantum cluster algebra structure, and it is conjectured that all quantum cluster monomials belong to $\bB^*$. Suppose that $Q$ is bipartite and $h$ is even. Set $h'=h/2-1$. By comparing initial seeds, one sees that the quantum cluster structure of $\K_{t,\,Q}$ obtained by transporting via $\Phi^{-1}$ the quantum cluster structure of $U_v(\n)$ is a $t$-analogue of the cluster structure of ${\mathcal{R}}_{h'}$ conjectured in [@HL]. Thus, by Theorem \[mainTh\], the two conjectures of [@GLS] for $U_v(\n)$ and of [@HL] for ${\mathcal{R}}_{h'}$ are essentially equivalent.
In [@HL] and [@Ncl], the conjecture for ${\mathcal{R}}_\ell$ was proved in the first non trivial case $\ell = 1$. (In [@HL] some combinatorial steps of the proof were only verified for $\g$ of type $A_n$ and $D_4$; the proof of [@Ncl] is general and uses geometric representation theory.) Since $\CC_1$ is a tensor subcategory of $\CC_Q$ (for every $\g$ except $\Sl_2$ and $\Sl_3$), $\K_{t,\,Q}$ contains a subring $\K_{t,1}$ corresponding to $\CC_1$. It is easy to see that $\Phi(\K_{t,1})$ is equal to the subalgebra $\U_{v}(\n(w))$ of [@GLS] where $w = c^2$ is the square of the Coxeter element of the Weyl group of $\g$ corresponding to the bipartite quiver $Q$. This is a quantum cluster algebra of finite cluster type, equal to the Dynkin type of $\g$ in the classification of Fomin and Zelevinsky. Thus, using [@HL; @Ncl; @Qin], Theorem \[mainTh\] readily implies:
\[Cor\_cluster\] Let $w = c^2$ be as above. Then $\bB^* \cap \,U_v(\n(w))$ is equal to the set of quantum cluster monomials of $U_v(\n(w))$.
For $\g$ of type $A_n$, Lampe [@La] has given a direct proof of the fact that the quantum cluster *variables* of $U_v(\n(w))$ belong to $\bB^*$.
[**1.9**]{}Since the bases $\bB^*$ and $\{[L]_t\}$ have geometric origin, it is natural to ask for a geometric explanation of Theorem \[mainTh\] (b). In the final part of the paper, we show (Theorem \[thmgeo\]) that the quiver representation spaces $E_\bd$ used by Lusztig to define the canonical basis of $U_v(\n)$ are isomorphic to some particular graded quiver varieties ${{\mathcal M}}^\bullet_0(W^\bd)$ used by Nakajima for describing the classes $[L]_t$ of the simple objects $L$ of $\CC_Q$. Moreover the intersection cohomology sheaves of closures of $G_\bd$-orbits in $E_\bd$ can be identified with the intersection cohomology sheaves of closures of strata in ${{\mathcal M}}^\bullet_0(W^\bd)$. This is inspired by a similar result of Nakajima [@Ncl] for the category $\CC_1$.
[**1.10**]{}We now give an overview of the structure of the paper. In Section \[sectCartanAuslander\], we set up our notation and introduce an important bijection $\varphi$ between the set of fundamental modules of $\CC_{\mathbb{Z}}$ and the vertices of the Auslander-Reiten quiver of $D^b(KQ)$. We use this bijection to express the entries of the inverse of the quantum Cartan matrix of $\g$ in terms of the Ringel form of $Q$, or in terms of the scalar product of the weight lattice of $\g$ (Proposition \[formula\_inverse\]). By construction, the quantum Grothendieck ring $\K_{t}$ is a subring of a quantum torus $\Y_{t}$ over ${\mathbb{C}}(t^{1/2})$. The $t$-commutation relations between generators of $\Y_{t}$ are expressed in terms of entries of the inverse of the quantum Cartan matrix of $\g$ [@H1], hence by Proposition \[formula\_inverse\], in terms of scalar products of weights of $\g$. The quantum Grothendieck ring $\K_{t,\,Q}$ is a subring of a subtorus $\Y_{t,\,Q}$ of $\Y_{t}$, of rank $r$ equal to the number of positive roots of $\g$. On the other hand, by [@GLS], $\U_{v}(\n)$ has an explicit embedding into a quantum torus $\T_{v,\,Q}$ of rank $r$ over ${\mathbb{C}}(v^{1/2})$, whose generators are certain unipotent quantum flag minors. The explicit $v$-commutation relations between these minors involve scalar products of roots and weights of $\g$. Comparing these two presentations, we show that there is an isomorphism $\Phi\colon \Y_{t,\,Q} \to \T_{v,\,Q}$ mapping $t^{1/2}$ to $v^{1/2}$ (Proposition \[main\_prop\]).
The proof that $\Phi$ restricts to an isomorphism from $\K_{t,\,Q}$ to $\U_{v}(\n)$ is based on some explicit systems of algebraic identities satisfied by the generators of both algebras. In Section \[sect4\], we recall from [@GLS] a system of quantum determinantal identities occuring in $U_{v}(\n)$, and in Section \[sect5\] we derive a quantum $T$-system for the $(q,t)$-characters of the Kirillov-Reshetikhin modules. (In [@N-KR §4], a quantum $T$-system was already obtained for the $t$-deformed product used in [@N; @VV]. A quantum cluster algebra related to the quantum $T$-system of type $A_1$ is also studied in [@DFK].) Comparing these two systems we obtain that $\Phi$ maps the classes of the Kirillov-Reshetikhin modules of $\CC_{Q}$ to certain quantum minors of $\U_{v}(\n)$ (multiplied by explicit half-integral powers of $v$). In particular, $\Phi$ maps the classes of the fundamental modules of $\CC_{Q}$ in $\K_{t,\,Q}$ to the generators of the dual PBW-basis of $\U_{v}(\n)$ associated with $Q$ (up to powers of $v^{1/2}$). This proves the first part of Theorem \[mainTh\]. The second part is deduced from the algebraic characterizations of $\bB^*$ and of the classes $[L]_t$ (Section \[sect6\]). After some examples, we give the proof of Corollary \[corollary\_binary\].
The above-mentioned presentation of $\K_t$ (Theorem \[presentation\]) is deduced from Theorem \[mainTh\] in Section \[sect\_pres\], and in Section \[sect\_Hall\] we prove the isomorphism with the derived Hall algebra $DH(Q)$ stated in Theorem \[Th\_DHall\]. Finally, in Section \[sect7\], we explain our geometric approach to Theorem \[mainTh\] (b) (Theorem \[thmgeo\]).
Cartan matrices and Auslander-Reiten quivers {#sectCartanAuslander}
============================================
Cartan matrix {#notation}
-------------
Let $\g$ be a simple Lie algebra of type $A, D, E$. We denote by $I$ the set of vertices of its Dynkin diagram, and we put $n = |I|$. The *Cartan matrix* of $\g$ is the $I\times I$ matrix $C$ with entries $$C_{ij} =
\left\{
\begin{array}{cl}
2 & \mbox{if } i=j,\cr
-1 & \mbox{if $i$ and $j$ are adjacent vertices of the Dynkin diagram},\cr
0 &\mbox{otherwise.}
\end{array}
\right.$$ We shall often use the shorthand notation $i\sim j$ to say that $C_{ij}=-1$.
We denote by $P$ the weight lattice of $\g$, and by $\vpi_i\ (i\in I)$ its basis of fundamental weights. The simple roots are defined by $$\a_i = \sum_{j\in I} C_{ij}\vpi_j,\qquad (i\in I).$$ The set of simple roots is denoted by $\Pi := \{\a_i\mid i\in I\}$. We denote by $(\cdot, \cdot)$ the scalar product of $P$ defined by $(\a_i,\vpi_j) = \de_{ij}$. Equivalently $(\a_i,\a_j) = C_{ij}$. The Weyl group $W$ is generated by the reflexions $s_i$ acting on $P$ by $$s_i(\la) = \la -(\la,\a_i)\a_i,\qquad (\la\in P,\ i\in I).$$ The root system of $\g$ is $\De:= W\Pi$. It decomposes as $\De = \De_+ \sqcup \De_-$, where $\De_+ = \De \cap (\oplus_{i\in I} {\mathbb{N}}\a_i)$ and $\De_- = -\De_+$. We write $r:=|\De_+|$.
A *Coxeter element* of $W$ is a product of the form $c = s_{i_1}\cdots s_{i_n}$ where $(i_1,\ldots,i_n)$ is an arbitrary ordering of $I$. All Coxeter elements are conjugate in $W$. Their common order is called the *Coxeter number* and denoted by $h$. We have $hn = 2r$.
Quivers {#subsect_quivers}
-------
Let $Q$ be an orientation of the Dynkin diagram of $\g$. In other words, $Q$ is a Dynkin quiver of the same Dynkin type as $\g$.
For $i\in I$, we denote by $s_i(Q)$ the quiver obtained from $Q$ by changing the orientation of every arrow with source $i$ or target $i$. Let $w = s_{i_1}\cdots s_{i_k}\in W$ be a reduced decomposition. We say that $\bi=(i_1,\ldots,i_k)$ is *adapted* to $Q$ if $i_1$ is a source of $Q$, $i_{2}$ is a source of $s_{i_1}(Q)$, …, $i_k$ is a source of $s_{i_{k-1}}\cdots s_{i_1}(Q)$. There is a unique Coxeter element having reduced expressions adapted to $Q$. We shall denote it by $\tau$.
We denote by $Q_1$ the set of arrows of $Q$. A *height function* $\xi\colon I \to {\mathbb{Z}}$ on $Q$ is a function satisfying $$\xi_j = \xi_i-1\quad \mbox{if}\quad i\to j \in Q_1.$$ Since $Q$ is connected, two height functions differ by a constant. We fix such a function $\xi$. Define $$\hI := \{(i,p)\in I\times{\mathbb{Z}}\mid p-\xi_i \in 2{\mathbb{Z}}\}.$$ We attach to $Q$ the infinite *repetition quiver* $\hQ$, defined as the oriented graph with vertex set $\hI$ and two types of arrows:
- if there is an arrow $i\to j$ in $Q$ we have arrows $(i,p)\to (j,p+1)$ in $\hQ$ for all $(i,p)\in\hI$;
- if there is an arrow $i\to j$ in $Q$ we have arrows $(j,q)\to (i,q+1)$ in $\hQ$ for all $(j,q)\in\hI$.
Note that $\hQ$ depends only on the Dynkin diagram, and not on the choice of orientation $Q$. In fact, it is well known that $\hQ$ is the quiver of a ${\mathbb{Z}}$-covering of the preprojective algebra associated with $Q$. In the literature, this quiver is often denoted by ${\mathbb{Z}}Q$. An example is shown in Figure \[FigD4\_1\], where the height function is $\xi_1=\xi_2=0,\ \xi_3=1,\ \xi_4=2$.
$$\[email protected]{
&&&&4\ar[ld]
\\
&&& \ar[lld]\ar[ld]3 &
\\
&1 &2&&
}
\qquad\qquad
\[email protected]{
&{}\save[]+<0cm,2ex>*{\vdots}\restore&{}\save[]+<0cm,2ex>*{\vdots}\restore& (3,3) &{}\save[]+<0cm,2ex>*{\vdots}\restore
\\
&(1,2) \ar[rru]&(2,2)\ar[ru]&&\ar[lu](4,2)
\\
&&&\ar[llu]\ar[lu](3,1)\ar[ru]&
\\
&(1,0)\ar[rru] &(2,0)\ar[ru]&&\ar[lu](4,0)
\\
&&& \ar[llu]\ar[lu](3,-1)\ar[ru]&
\\
&(1,-2)\ar[rru] &(2,-2)\ar[ru]& & \ar[lu](4,-2)
\\
& && \ar[llu]\ar[lu](3,-3)\ar[ru]& \\
&{}\save[]+<0cm,2ex>*{\vdots}\restore&{}\save[]+<0cm,2ex>*{\vdots}\restore&&{}\save[]+<0cm,2ex>*{\vdots}\restore
}$$
Let $\hDe := \De_+\times{\mathbb{Z}}$. We now describe a natural labelling of the vertices of $\hQ$ by $\hDe$. For $i\in I$, let $B(i)$ be the subset of $I$ consisting of all $j$’s such that there is a path from $j$ to $i$ in $Q$. Define $$\ga_i := \sum_{j\in B(i)} \a_j,\qquad (i\in I).$$ We have $\ga_i\in \De_+$. There is a unique bijection $\varphi\colon \hI \to \hDe$ defined inductively as follows:
- $\varphi(i,\xi_i) = (\ga_i,0)$ for $i\in I$;
- suppose that $\varphi(i,p) = (\b,m)$; then
- $\varphi(i,p-2) = (\tau(\b),m)$ if $\tau(\b)\in \De_+$;
- $\varphi(i,p-2) = (-\tau(\b),m-1)$ if $\tau(\b)\in \De_-$;
- $\varphi(i,p+2) = (\tau^{-1}(\b),m)$ if $\tau^{-1}(\b)\in \De_+$;
- $\varphi(i,p+2) = (-\tau^{-1}(\b),m+1)$ if $\tau^{-1}(\b)\in \De_-$.
Note that this second labelling of $\hQ$ depends on $Q$. This is illustrated in Figure \[FigD4\_2\].
$$\[email protected]{
&&&&4\ar[ld]
\\
&&& \ar[lld]\ar[ld]3 &
\\
&1 &2&&
}
\quad
\def\objectstyle{\scriptstyle}
\[email protected]{
&{}\save[]+<0cm,2ex>*{\vdots}\restore&{}\save[]+<0cm,2ex>*{\vdots}\restore& (\a_1+\a_2+\a_3,1) &{}\save[]+<0cm,2ex>*{\vdots}\restore
\\
&(\a_1,1) \ar[rru]&(\a_2,1)\ar[ru]&&\ar[lu](\a_4,0)
\\
&&&\ar[llu]\ar[lu](\a_3+\a_4,0)\ar[ru]&
\\
&(\a_1+\a_3+\a_4,0)\ar[rru] &(\a_2+\a_3+\a_4,0)\ar[ru]&&\ar[lu](\a_3,0)
\\
&&& \ar[llu]\ar[lu](\a_1+\a_2+2\a_3+\a_4,0)\ar[ru]&
\\
&(\a_2+\a_3,0)\ar[rru] &(\a_1+\a_3,0)\ar[ru]& & \ar[lu](\a_1+\a_2+\a_3+\a_4,0)
\\
& && \ar[llu]\ar[lu](\a_1+\a_2+\a_3,0)\ar[ru]&
\\
&(\a_1,0)\ar[rru] &(\a_2,0)\ar[ru]& & \ar[lu](\a_4,-1)
\\
& && \ar[llu]\ar[lu](\a_3+\a_4,-1)\ar[ru]& \\
&{}\save[]+<0cm,2ex>*{\vdots}\restore&{}\save[]+<0cm,2ex>*{\vdots}\restore&&{}\save[]+<0cm,2ex>*{\vdots}\restore
}$$
Auslander-Reiten theory {#subsect_AR}
-----------------------
The quiver $\hQ$ with its labelling by $\hDe$ arises in the representation theory of the path algebra $KQ$ of $Q$ over a field $K$, as we shall now recall. We refer the reader to [@ARS; @ASS; @GaR; @Ri0] for background on quiver representations and Auslander-Reiten theory.
Let ${\mathrm{mod}}(KQ)$ be the abelian category of representations of $Q$ over $K$. For an object $X$ of ${\mathrm{mod}}(KQ)$ we write ${\underline{\dim}}(X)$ for its dimension vector. We define the Ringel bilinear form $$\< X, Y\> := \dim({\operatorname{Hom}}(X,Y)) - \dim({\operatorname{Ext}}^1(X,Y)), \qquad (X,Y\in{\mathrm{mod}}(KQ)),$$ and the symmetric form $(X,Y):= \<X,Y\> + \<Y,X\>$. It is known that these forms depend only on the dimension vectors ${\underline{\dim}}(X)$ and ${\underline{\dim}}(Y)$. Moreover, if we identify in the standard way ${\underline{\dim}}(X)$ and ${\underline{\dim}}(Y)$ with elements of the root lattice of $\g$, then $(X,Y)$ coincides with the natural scalar product $({\underline{\dim}}(X),{\underline{\dim}}(Y))$. In this picture, $\a_i$ is the dimension vector of the simple $KQ$-module $S_i$ supported on vertex $i$, and $\ga_i$ is the dimension vector of its injective envelope $I_i$. Recall that, by Gabriel’s theorem, the isoclasses of indecomposable $KQ$-modules are in natural bijection with $\De_+$. They form the vertices of the Auslander-Reiten quiver $\Ga_Q$ of ${\mathrm{mod}}(KQ)$. The map $\b \mapsto (\b,0)$ identifies $\Ga_Q$ with the full subgraph of $\hQ$ with set of vertices $\De_+\times \{0\}$. The map $\tau$ restricted to the dimension vectors in $\De_+$ of non projective $KQ$-modules is the Auslander-Reiten translation of ${\mathrm{mod}}(KQ)$ [@ARS].
Let $D^b({\mathrm{mod}}(KQ))$ be the bounded derived category of $KQ$. Its indecomposable objects are the stalk complexes $X[i]$, consisting of an indecomposable object $X$ of ${\mathrm{mod}}(KQ)$ sitting in degree $i\in{\mathbb{Z}}$, and zero objects in all other degrees. Thus, the isoclasses of indecomposable objects of $D^b({\mathrm{mod}}(KQ))$ are naturally labelled by $\hDe$. Using this labelling, the quiver $\hQ$ is identified with the Auslander-Reiten quiver of the triangulated category $D^b({\mathrm{mod}}(KQ))$ [@Ha].
Quantum Cartan matrix {#cartan}
---------------------
Let $z$ be an indeterminate, and let $C(z)$ be the matrix with entries $$C_{ij}(z) =
\left\{
\begin{array}{cl}
z+z^{-1} & \mbox{if } i=j,\cr
-1 & \mbox{if $i\sim j$},\cr
0 &\mbox{otherwise.}
\end{array}
\right.$$ Thus $C(1)$ is just the Cartan matrix $C$ of $\g$. Since $\det(C) \not = 0$, $\det(C(z))\not = 0$. We denote by $\widetilde{C}(z)$ the inverse of the matrix $C(z)$. This is a matrix with entries $\widetilde{C}_{ij}(z)\in{\mathbb{Q}}(z)$. Denoting by $A$ the adjacency matrix of the Dynkin diagram we have $$C(z) = (z+z^{-1})I - A,$$ therefore $$\widetilde{C}(z) = \sum_{k\ge 0} (z+z^{-1})^{-k-1}A^k.$$ Hence the entries of $\widetilde{C}(z)$ have power series expansions in $z$ of the form $$\widetilde{C}_{ij}(z) = \sum_{m \ge 1} \widetilde{C}_{ij}(m)\, z^{m},$$ where $\widetilde{C}_{ij}(m)\in{\mathbb{Z}}$. Note that since $C(z)$ is a symmetric matrix, we have $\widetilde{C}_{ij}(m) = \widetilde{C}_{ji}(m)$.
Formula for $\widetilde{C}_{ij}(m)$ {#formula_inverse}
-----------------------------------
We will now give several equivalent expressions for the coefficients $\widetilde{C}_{ij}(m)$. For other expressions of $\widetilde{C}_{ij}(z)$ in type $A_n$ and $D_n$, see [@FR1 Appendix C].
Fix an orientation $Q$ of the Dynkin diagram, and recall from §\[subsect\_quivers\] and §\[subsect\_AR\] the associated notation $\xi_i$, $\ga_i$, the Coxeter transformation $\tau$, and the Ringel form $\<\cdot,\cdot\>$.
\[prop1.1\] Let $m\ge 1$. If $m+\xi_i-\xi_j-1$ is odd then $\widetilde{C}_{ij}(m) = 0$. Otherwise $$\label{eq1}
\widetilde{C}_{ij}(m) = \left(\tau^{(m+\xi_i-\xi_j-1)/2}(\ga_i), \vpi_j\right).$$ Equivalently, $$\label{eq2}
\widetilde{C}_{ij}(m) = \left\< \tau^{(m+\xi_i-\xi_j-1)/2}(I_i), I_j\right\>.$$
Let us denote temporarily by $D_{ij}(m)$ the value of $\widetilde{C}_{ij}(m)$ predicted by the proposition. We want to show that $$\sum_{k\in I,\ m\ge 1} C_{ik}(z) D_{kj}(m) z^m = \de_{ij},\qquad (i,j\in I).$$ Using the definition of $C_{ik}(z)$, this is equivalent to show that $$\label{eqtoprov}
\sum_{m\ge 1}\left(\left(z+z^{-1}\right)D_{ij}(m) - \sum_{k \sim i} D_{kj}(m)\right)z^m = \de_{ij},
\qquad (i,j\in I).$$
The coefficient of $z^0$ in the left-hand side is equal to $D_{ij}(1)$. If $\xi_i-\xi_j$ is odd then by definition $D_{ij}(1) = 0$. Otherwise, if $\xi_i-\xi_j = 2l$, then $D_{ij}(1) = (\tau^l(\ga_i), \vpi_j)$ is the coefficient of $\a_j$ in $\tau^l(\ga_i)$. Let $(\b,m)$ be the vertex of $\hQ$ in the column of $(\ga_i,0)$ and at the same height as $(\ga_j,0)$. Such a vertex exists because $\xi_i-\xi_j$ is even, and clearly $\b = \pm \tau^l(\ga_i)$. Now it is a well-known fact from the combinatorics of Auslander-Reiten quivers that for all vertices $(\ga, s)$ of $\hQ$ at the same height as $(\ga_j, 0)$ the coefficient of $\a_j$ in $\ga$ is 0, except if $(\ga, s) = (\ga_j,0)$ in which case it is equal to 1. Hence we have $D_{ij}(1) = \de_{ij}$.
Consider now the coefficient of $z^m\ (m\ge 1)$ in (\[eqtoprov\]). We need to show that $$\label{eqtoprov2}
D_{ij}(m+1)+D_{ij}(m-1) - \sum_{k:\,k \sim i} D_{kj}(m) = 0,
\qquad (i,j\in I, \ m\ge 1).$$ Note that for $k\sim i$ we have $\xi_k = \xi_i \pm 1$, hence if $m+\xi_i-\xi_j$ is odd, all summands of the left-hand side are zero. Otherwise, writing $m+\xi_i-\xi_j = 2l$, the left-hand side of (\[eqtoprov2\]) is $$\left(\tau^l(\ga_i) + \tau^{l-1}(\ga_i) - \sum_{k\sim i} \tau^{l+(\xi_k-\xi_i+1)/2}(\ga_k)\ ,\ \vpi_j\right).$$ Now it is again a familiar fact from the combinatorics of Auslander-Reiten quivers that $$\tau^l(\ga_i) + \tau^{l-1}(\ga_i) = \sum_{k\sim i} \tau^{l+(\xi_k-\xi_i+1)/2}(\ga_k),$$ since the roots $\tau^{l}(\ga_i)$, $\tau^{l-1}(\ga_i)$, and $\tau^{l+(\xi_k-\xi_i+1)/2}(\ga_k)$ with $k\sim i$, form a mesh. This proves (\[eq1\]).
Finally, if $\b = {\underline{\dim}}X$ then $(\b,\vpi_j)$ is equal to the coefficient of $\a_j$ in $\b$, hence $$(\b,\vpi_j) = \dim({\operatorname{Hom}}(X, I_j)) = \<X,\, I_j\>,$$ because $I_j$ is injective. This proves (\[eq2\]).
[Take $\g$ of type $A_4$. One has for instance $$\begin{array}{ccl}
\widetilde{C}_{11}(z)&=&z^{1}-z^{9}+z^{11}-z^{19}+\cdots \\[2mm]
\widetilde{C}_{12}(z)&=&z^{2}-z^{8}+z^{12}-z^{18}+\cdots \\[2mm]
\widetilde{C}_{13}(z)&=&z^{3}-z^{7}+z^{13}-z^{17}+\cdots \\[2mm]
\widetilde{C}_{14}(z)&=&z^{4}-z^{6}+z^{14}-z^{16}+\cdots \\[2mm]
\widetilde{C}_{21}(z)&=&z^{2}-z^{8}+z^{12}-z^{18}+\cdots \\[2mm]
\widetilde{C}_{22}(z)&=&z^{1}+z^{3}-z^{7}-z^{9}+z^{11}+z^{13}-z^{17}-z^{19}+\cdots \\[2mm]
\widetilde{C}_{23}(z)&=&z^{2}+z^{4}-z^{6}-z^{8}+z^{12}+z^{14}-z^{16}-z^{18}\cdots \\[2mm]
\widetilde{C}_{24}(z)&=&z^{3}-z^{7}+z^{13}-z^{17}+\cdots \\[2mm]
\end{array}$$ Let us choose the sink-source orientation $Q$ with height function $\xi_1=0, \xi_2=1, \xi_3=0, \xi_4=1$. Then $\tau = s_2s_4s_1s_3$, and since $\tau^5 = 1$, the roots $\tau^l(\ga_i)$ are all determined by: $$\begin{array}{llll}
\ga_1 = \a_1+\a_2, & \ga_2 = \a_2, &\ga_3 = \a_2+\a_3+\a_4, &\ga_4 = \a_4,\\[2mm]
\tau(\ga_1) = \a_3+\a_4, & \tau(\ga_2) = \a_1+\a_2+\a_3+\a_4, &\tau(\ga_3) = \a_1+\a_2+\a_3,
&\tau(\ga_4) = \a_2+\a_3,\\[2mm]
\tau^2(\ga_1) = -\a_4, & \tau^2(\ga_2) = \a_3, &\tau^2(\ga_3) = -\a_2, &\tau^2(\ga_4) = \a_1,\\[2mm]
\tau^3(\ga_1) = -\a_2-\a_3, & \tau^3(\ga_2) = -\a_2-\a_3-\a_4, &\tau^3(\ga_3) = -\a_1-\a_2-\a_3-\a_4,
&\tau^3(\ga_4) = -\a_1-\a_2,\\[2mm]
\tau^4(\ga_1) = -\a_1, & \tau^4(\ga_2) = -\a_1-\a_2-\a_3, &\tau^4(\ga_3) = -\a_3, &\tau^4(\ga_4) = -\a_3-\a_4.
\end{array}$$ For instance by Proposition \[prop1.1\], $\widetilde{C}_{23}(6)$ is equal to the coefficient of $\a_3$ in $\tau^3(\ga_2) = -\a_2-\a_3-\a_4$, namely to $-1$. ]{}
For $i,j\in I$ and $m \ge 1$ we have $$\widetilde{C}_{ij}(m+2h) = \widetilde{C}_{ij}(m).$$
Since $\tau^h = 1$, this follows immediately from Proposition \[prop1.1\].
Quantum tori {#sect3}
============
The quantum torus $\YY_t$ {#secY}
-------------------------
Recall from §\[subsect\_quivers\] the labelling set $\hI$ of $\hQ$. Define $$\Y := {\mathbb{C}}\left[Y_{i,p}^{\pm 1}\mid (i,p) \in \hI\right]$$ to be the Laurent polynomial ring generated by a collection of commutative variables $Y_{i,p}$ labelled by $\hI$. This ring is related to a tensor subcategory $\CC_{\mathbb{Z}}$ of the category of finite-dimensional $U_q(L\g)$-modules considered in [@HL] (see below §\[subsectCZ\]).
Let $t$ be an indeterminate. Following [@H1] we introduce a $t$-deformed version $(\YY_t,*)$ of $\Y$, with noncommutative multiplication denoted by $*$. This is the ${\mathbb{C}}(t)$-algebra generated by variables still denoted by $Y_{i,p}$, subject to the $t$-commutation relations $$\label{tcom}
Y_{i,p}*Y_{j,s} := t^{\cN(i,p;j,s)} Y_{j,s}*Y_{i,p},\qquad ((i,p),\,(j,s)\in \hI),$$ where $$\label{defcN}
\cN(i,p;j,s) := \widetilde{C}_{ij}(p-s-1) - \widetilde{C}_{ij}(p-s+1)
-\widetilde{C}_{ij}(s-p-1) + \widetilde{C}_{ij}(s-p+1).$$ Here we have extended the definition of $\widetilde{C}_{ij}(m)$ to every $m\in{\mathbb{Z}}$ by setting $\widetilde{C}_{ij}(m) = 0$ if $m\le 0$. Note that, since $\widetilde{C}(z)$ is symmetric, we have $$\cN(i,p;j,s) = - \cN(j,s;i,p),\qquad (i,j\in I,\ p,s\in{\mathbb{Z}}).$$ If $p=s$ then $\cN(i,p;j,s) = 0$. Otherwise, without loss of generality we can assume that $p<s$. Then, (\[defcN\]) simplifies as $$\label{defcN2}
\cN(i,p;j,s) =
\widetilde{C}_{ij}(s-p+1) - \widetilde{C}_{ij}(s-p-1),\qquad (p<s).$$ We regard the noncommutative ring $(\YY_t,*)$ as a quantum torus of infinite rank.
\[Remark\_differ\] [In [@VV] and [@N], the construction of a $t$-deformed Grothendieck ring is based on a slightly different quantum torus. Namely, in these papers the product is defined by: $$\label{tcom'}
Y_{i,p}\,\cdot\,Y_{j,s} := t^{\cN'(i,p;j,s)}
Y_{j,s}\,\cdot\,Y_{i,p},\qquad ((i,p),\,(j,s)\in \hI),$$ where instead of (\[defcN\]) the following exponent is used: $$\label{defcN'}
\cN'(i,p;j,s) := -2\left(\widetilde{C}_{ij}(p-s-1) -\widetilde{C}_{ij}(s-p-1)\right). $$ For instance, in type $A_3$, we have $$Y_{1,0} * Y_{2,1} = t Y_{2,1} * Y_{1,0},
\qquad
\mbox{whereas}
\qquad
Y_{1,0} \,\cdot\, Y_{2,1} = Y_{2,1} \,\cdot\, Y_{1,0}.$$ In [@N; @VV], the definition of the product comes from a convolution operation for certain perverse sheaves on quiver varieties, and the deformation parameter $t$ encodes the natural grading of complexes of sheaves. Our product $*$ comes from [@H1] and the original construction of $q$-characters. Indeed in [@FR], the variables $Y_{i,p}\in\Y$ are defined as formal power series in elements of $U_q(L\g)$, and they pairwise commute. In [@H1], these formal power series are replaced by certain infinite sums $\tilde{Y}_{i,p}$ in elements of the quantum affine algebra $U_q(\widehat{\g})$ (with non trivial central charge $c$), which can be seen as vertex operators. The original variable $Y_{i,p}$ is just one factor of the complete variable $\tilde{Y}_{i,p}$. The relations of the quantum affine algebra then give rise to $t$-commutation relations between the $\tilde{Y}_{i,p}$, where the parameter $t$ appears as a formal power series with coefficients in ${\mathbb{C}}[c^{\pm 1}]$ [@H1 Theorem 3.11]. The defining relations (\[tcom\]) (\[defcN\]) of $*$ are obtained by replacing $t$ by $t^{-1}$ in those $t$-commutation relations. ]{}
Recall from §\[subsect\_quivers\] the bijection $\varphi\colon \hI \to \hDe$.
\[expression\_N\] Let $(i,p)$ and $(j,s)$ be elements of $\hI$ with $p<s$. There holds $$\cN(i,p;j,s) = \left(\tau^{(s-p+\xi_i-\xi_j)/2}(\ga_i),\ \ga_j\right).$$ Moreover, if $\varphi(i,p) = (\b,m)$ and $\varphi(j,s) = (\de,l)$, then $$\cN(i,p;j,s) = (-1)^{l-m}\,(\b,\de).$$
First note that the definition of $\hI$ implies that $s-p+\xi_i-\xi_j\in 2{\mathbb{Z}}$. By Proposition \[prop1.1\], we have $$\begin{aligned}
\cN(i,p;j,s) &=& \widetilde{C}_{ij}(s-p+1) - \widetilde{C}_{ij}(s-p-1)\\
&=& \left\<\tau^{(s-p+\xi_i-\xi_j)/2}(I_i),\ I_j\right\>
- \left\<\tau^{(s-p+\xi_i-\xi_j)/2-1}(I_i),\ I_j\right\>.\end{aligned}$$ Now recall the classical formula $$\<\tau^{-1}(X),Y\> = - \<Y,X\>,\qquad (X,Y\in{\mathrm{mod}}(FQ)).$$ It follows that $$\begin{aligned}
\cN(i,p;j,s) &=& \left\<\tau^{(s-p+\xi_i-\xi_j)/2}(I_i),\ I_j\right\>
+ \left\<I_j, \tau^{(s-p+\xi_i-\xi_j)/2}(I_i)\right\> \\
&=& \left(\tau^{(s-p+\xi_i-\xi_j)/2}(I_i),\ I_j\right) \\
&=& \left(\tau^{(s-p+\xi_i-\xi_j)/2}(\ga_i),\ \ga_j\right).\end{aligned}$$ This proves the first equality. The second equality is immediately deduced from the first one if we note that, by definition of the bijection $\varphi$, $$\tau^{(\xi_i - p)/2}(\ga_i) = (-1)^m\b,
\qquad
\tau^{(\xi_j - s)/2}(\ga_i) = (-1)^l\de.$$
\[Remark\_differ2\] [For the product of [@VV] and [@N], we have, for $p<s$, a similar expression $$\cN'(i,p;j,s) = 2\left\<\tau^{(s-p+\xi_i-\xi_j)/2-1}(\ga_i),\ \ga_j\right\>
=-2\left<\ga_j,\tau^{(s-p+\xi_i-\xi_j)/2}(\ga_i)\right\>
=(-1)^{l-m+1}2\<\de,\b\>,$$ in which the symmetric scalar product $(\cdot,\cdot)$ is replaced by the non-symmetric Ringel form $\<\cdot,\cdot\>$. ]{}
Commutative monomials {#commut_prod}
---------------------
Let us adjoin a square root $t^{1/2}$ of $t$ and extend the quantum torus $(\YY_t,*)$ to $$(\Y_t,*):= {\mathbb{C}}(t^{1/2})\otimes_{{\mathbb{C}}(t)} (\YY_t,*).$$ We notice that the expression $$t^{\frac{1}{2}\cN(j,s;\,i,p)}\, Y_{i,p}*Y_{j,s} =
t^{\frac{1}{2}\cN(i,p;\,j,s)}\, Y_{j,s}*Y_{i,p} \in \Y_{t}$$ is invariant under permutation of $(i,p)$ and $(j,s)$. We can then denote it as a *commutative* monomial $Y_{i,p}Y_{j,s}=Y_{j,s}Y_{i,p}$, and write $$Y_{i,p}*Y_{j,s} = t^{\frac{1}{2}\cN(i,p;\,j,s)}\, Y_{i,p}Y_{j,s}.$$ More generally, for a family $(u_{i,p}\mid (i,p)\in \hI)$ of integers with finitely many nonzero components, the expression $$t^{\frac{1}{2}\sum_{(i,p)<(j,s)} u_{i,p} u_{j,s} \cN(j,s;\,i,p)}
\hskip-0.4cm
\stackrel{\longrightarrow}{\mbox{\Large\quad *}_{\hskip-0.4cm (i,p)\in\hI}}
Y_{i,p}^{u_{i,p}}$$ does not depend on the chosen ordering of $\hI$ used to define it. We will denote it as a commutative monomial $\prod_{(i,p)\in\hI} Y_{i,p}^{u_{i,p}}$, and write $$\underset{(i,p)\in\hI}{\overset{\longrightarrow}{\mbox{\Large *}}}
\,Y_{i,p}^{u_{i,p}} =
t^{\frac{1}{2}\sum_{(i,p)<(j,s)} u_{i,p} u_{j,s} \cN(i,p;\,j,s)}
\prod_{(i,p)\in\hI} Y_{i,p}^{u_{i,p}}.$$ The commutative monomials form a basis of the ${\mathbb{C}}(t^{1/2})$-vector space $\Y_{t}$. It will be convenient to denote commutative monomials by $$m = \prod_{(i,p)\in \hI} Y_{i,p}^{u_{i,p}(m)}.$$ A commutative monomial $m$ is said to be *dominant* if $u_{i,p}(m) \ge 0$ for every $(i,p)\in \hI$.
The noncommutative product of two commutative monomials $m_1$ and $m_2$ is given by $$\label{mult_mono}
m_1 * m_2 =
t^{\frac{1}{2}D(m_1,\,m_2)} m_1m_2
= t^{D(m_1,\,m_2)} m_2*m_1,$$ where $$D(m_1,m_2) = \sum_{(i,p),(j,s)\in \hI} u_{i,p}(m_1)u_{j,s}(m_2)
\cN(i,p;\,j,s),$$ and $$m_1m_2 = \prod_{(i,p)\in \hI} Y_{i,p}^{u_{i,p}(m_1)+u_{i,p}(m_2)},$$ denotes the commutative product.
The quantum torus $\YY_{t,\,Q}$ {#torusT}
-------------------------------
Recall the bijection $\varphi \colon \hI \to \hDe$ of §\[subsect\_quivers\]. Define $$\hI_Q := \varphi^{-1}\left(\De_+\times \{0\}\right) \subset \hI,$$ and let $\YY_{t,\,Q}$ be the the ${\mathbb{C}}(t)$-subalgebra of $(\YY_t,*)$ generated by the variables $Y_{i,p}\ ((i,p)\in \hI_Q)$. This is a quantum torus of rank $r = |\De_+|$. We will also use the extended torus $$(\Y_{t,\,Q},*):= {\mathbb{C}}(t^{1/2})\otimes_{{\mathbb{C}}(t)} (\YY_{t,\,Q},*).$$
\[exampleYQ\] [ We take $\g$ of type $D_4$ and choose $Q$ as in Figure \[FigD4\_1\] and Figure \[FigD4\_2\]. Comparing the two figures we see that $\YY_{t,\,Q}$ is generated by $$Y_{1,0}^{\pm1},\ Y_{1,-2}^{\pm1},\ Y_{1,-4}^{\pm1},\
Y_{2,0}^{\pm1},\ Y_{2,-2}^{\pm1},\ Y_{2,-4}^{\pm1},\
Y_{3,1}^{\pm1},\ Y_{3,-1}^{\pm1},\ Y_{3,-3}^{\pm1},\
Y_{4,2}^{\pm1},\ Y_{4,0}^{\pm1},\ Y_{4,-2}^{\pm1}.$$ ]{}
The quantum torus $\TT_{v,\,Q}$ {#qtorusA}
-------------------------------
Let $w_0$ be the longest element of $W$. Let $\bi = (i_1,\ldots,i_r)$ be a reduced expression of $w_0$ adapted to $Q$ (see §\[subsect\_quivers\]). Following [@GLS §11], we introduce a quantum torus $\TT_{v,\,Q}$ of rank $r$ over ${\mathbb{C}}(v)$. (The indeterminate $v$ is denoted by $q$ in [@GLS]). Its generators are certain unipotent quantum minors $$D_{\vpi_{i_k},\,\la_k},\qquad (1\le k\le r)$$ in the quantum coordinate ring $A_v(\n)$. Here $\la_k$ is the weight given by $$\la_k = s_{i_1}\cdots s_{i_k}(\vpi_{i_k}),\qquad (1\le k\le r).$$ The definition of $A_v(\n)$ will be recalled in §\[defAqn\] below. At this stage we only need to know the explicit $v$-commutation relations satisfied by these minors. It is shown in [@GLS Lemma 11.2] that for $k<l$ there holds $$\label{commutD}
D_{\vpi_{i_k},\,\la_k} D_{\vpi_{i_l},\,\la_l} =
v^{(\vpi_{i_k}-\la_k,\ \vpi_{i_l}+\la_l)}
D_{\vpi_{i_l},\,\la_l} D_{\vpi_{i_k},\,\la_k}.$$
For $1\le k\le r$, set $k^-:= \max(\{s<k \mid i_s = i_k\} \cup \{0\}).$ Define $$\label{defEk}
Z_k :=
D_{\vpi_{i_{k}},\,\la_{k}} \left(D_{\vpi_{i_{k^-}},\,\la_{k^-}}\right)^{-1},$$ where if $k^-=0$ we understand $D_{\vpi_{i_{k^-}},\,\la_{k^-}} = 1$. Clearly, $Z_k\ (1\le k\le r)$ is another set of generators of $\TT_{v,\,Q}$. Let $$\label{eq_beta}
\b_k = s_{i_1}\cdots s_{i_{k-1}}(\a_{i_k}),\qquad (1\le k\le r).$$ Note that we have $$\label{relation-n}
\la_{k} = \la_{k^-} - \b_{k}, \qquad (1\le k \le r),$$ where if $k^-=0$ we use the convention $\la_{k^-}=\varpi_{i_k}$.
\[lemcruc\] For $1\le k<l \le r$, we have: $$\label{toprove}
Z_k Z_l = v^{-(\b_k,\b_l)} Z_l Z_k.$$
Let us introduce the integers $\mu_{kl}$ and $\nu_{kl}$ such that $$D_{\vpi_{i_k},\la_k}D_{\vpi_{i_l},\la_l} = v^{\mu_{kl}} D_{\vpi_{i_l},\la_l} D_{\vpi_{i_k},\la_k},
\qquad
Z_k Z_l = v^{\nu_{kl}} Z_l Z_k,
\qquad (1\le k,\ l\le r).$$ By definition of $Z_k$ we have $$\nu_{kl} = (\mu_{kl} - \mu_{k^-l}) - (\mu_{kl^-} - \mu_{k^-l^-}),$$ where we use the convention that $\mu_{k^-l} = 0$ if $k^-=0$, $\mu_{kl^-} = 0$ if $l^-=0$, and $\mu_{k^-l^-} = 0$ if $k^-=0$ or $l^-=0$. Since $k^-<k<l$, we have $$\mu_{kl} - \mu_{k^-l} = (\vpi_{i_k}-\la_{k},\ \vpi_{i_l}+\la_{l})
- (\vpi_{i_k}-\la_{k^-},\ \vpi_{i_l}+\la_{l})
= (\b_k, \vpi_{i_l}+\la_{l}).$$
(a)If $k<l^-$ we have similarly $$\mu_{kl^-} - \mu_{k^-l^-} = ((\b_k, \vpi_{i_l}+\la_{l^-})$$ and so $$\nu_{kl} = (\b_k, \vpi_{i_l}+\la_{l}) - (\b_k, \vpi_{i_l}+\la_{l^-}) = -(\b_k,\b_l),$$ as required.
(b)If $k=l^-$ then $\mu_{kl^-}=0$ and $\mu_{k^-l}= (\vpi_{i_k}-\la_{k^-},\ \vpi_{i_k}+\la_{k})$. Hence $$\begin{aligned}
\nu_{kl}& = &(\b_k, \vpi_{i_l}+\la_{l}) + (\vpi_{i_k}-\la_{k^-},\ \vpi_{i_k}+\la_{k})\\
& = &(\b_k, \vpi_{i_k}+\la_{k}) - (\b_k,\b_l) + (\vpi_{i_k}-\la_{k^-},\ \vpi_{i_k}+\la_{k})\\
& = & - (\b_k,\b_l) + (\vpi_{i_k}-\la_{k^-}+\b_k,\ \vpi_{i_k}+\la_{k})\\
& = & - (\b_k,\b_l) + (\vpi_{i_k}-\la_{k},\ \vpi_{i_k}+\la_{k})\\
& = & - (\b_k,\b_l)\end{aligned}$$ as required, because $(\vpi_{i_k}-\la_{k},\ \vpi_{i_k}+\la_{k})=(\vpi_{i_k},\vpi_{i_k})-(\la_k,\la_k) = 0$.
(c)If $k>l^-$ then $\mu_{kl^-} = - \mu_{l^-k} = - (\vpi_{i_l}-\la_{l^-},\ \vpi_{i_k}+\la_{k})$. Hence $$\begin{aligned}
-\mu_{kl^-}+\mu_{k^-l^-}& = &(\vpi_{i_l}-\la_{l^-},\ \vpi_{i_k}+\la_{k}) +
(\vpi_{i_k}-\la_{k^-},\ \vpi_{i_l}+\la_{l^-})\\
&=& -(\b_k, \ \vpi_{i_l}+\la_{l^-}) + (\vpi_{i_l}-\la_{l^-},\ \vpi_{i_k}+\la_{k})
+(\vpi_{i_k}-\la_{k},\ \vpi_{i_l}+\la_{l^-})\\
&=& -(\b_k, \ \vpi_{i_l}+\la_{l^-}) + 2(\vpi_{i_k},\vpi_{i_l}) -2(\la_k,\la_{l^-})\\
&=& -(\b_k, \ \vpi_{i_l}+\la_{l^-}).\end{aligned}$$ Indeed, since $l^-<k<l$, we have $$(\la_k,\la_{l^-}) = (s_{i_{l^-+1}}\cdots s_{i_k}(\vpi_{i_k}),\ \vpi_{i_l})
= (\vpi_{i_k}, s_{i_{k}}\cdots s_{i_{l^-+1}}(\vpi_{i_l})) = (\vpi_{i_k},\vpi_{i_l}).$$ It follows that again, $\nu_{kl} = (\b_k, \vpi_{i_l}+\la_{l}) - (\b_k, \vpi_{i_l}+\la_{l^-}) = -(\b_k,\b_l)$, as required.
An isomorphism {#def_tilde_k}
--------------
It is well known that the roots $\b_k\ (1\le k\le r)$ give an enumeration of $\De_+$. Therefore, for every $(i,p)\in\hI_Q$ there is a unique $k$ such that $\varphi(i,p) = (\b_k,0)$.
\[prop2.2\] The assignment $$\begin{array}{ll}
t \mapsto v,& \\[3mm]
Y_{i,p} \mapsto Z_k, &\mbox{where } (i,p)\in \hI_Q,\ \mbox{and } \varphi(i,p) = (\b_k,0),
\end{array}$$ extends to an isomorphism of quantum tori from $\YY_{t,\,Q}$ to $\TT_{v,\,Q}$.
This follows immediately from Proposition \[expression\_N\] and Proposition \[lemcruc\] if we note that when $\varphi(i,p) = (\b_k,0)$ and $\varphi(j,s) = (\b_l,0)$, $p<s$ implies that $k>l$.
The involution $\si$ and the rescaled generators $X_k$ {#sigma}
------------------------------------------------------
Let $\T_{v,\,Q}:= {\mathbb{C}}(v^{1/2})\otimes_{{\mathbb{C}}(v)} \TT_{v,\,Q}$. For $\gamma = \sum_{i}c_i\a_i$ in the root lattice of $\g$, we set $$\label{defNgamma}
\deg\ga := \sum_i c_i,\qquad N(\ga) := \frac{(\ga,\ga)}{2} - \deg\ga.$$ Following [@GLS], we introduce an involution $\si$ of $\T_{v,\,Q}$, defined as the ${\mathbb{C}}$-algebra anti-automorphism satisfying $$\si(v^{1/2}) = v^{-1/2},\qquad
\si\left(D_{\vpi_{i_k},\la_k}\right) = v^{N(\vpi_{i_k}-\la_k)} D_{\vpi_{i_k},\la_k}.$$
We rescale the generators $Z_k$ of $\TT_{v,\,Q}$ by defining $$X_k
:=
\left\{
\begin{array}{cl}
v^{N(\b_k)/2} Z_k & \mbox{if $1\le k\le n$,}\\[2mm]
v^{N(\b_k)/2+(\vpi_{i_k}-\la_{k-n},\ \b_k)} Z_k & \mbox{if $n+1\le k\le r$}.
\end{array}
\right.$$ Note that these elements live in $\T_{v,\,Q}$.
For $1\le k\le r$ we have: $$\si(X_k) = X_k.$$
For convenience, we set $\la_{k-n} = \vpi_{i_k}$ if $k-n\le 0$. Using (\[defEk\]), (\[commutD\]), and the definition of $\si$, we have $$\si(Z_k) = v^{N(\vpi_{i_k}-\la_k) - N(\vpi_{i_k}-\la_{k-n}) - (\vpi_{i_k}-\la_{k-n},\ \vpi_{i_k}+\la_{k})}Z_k.$$ A simple calculation using (\[relation-n\]) shows that $$N(\vpi_{i_k}-\la_k) - N(\vpi_{i_k}-\la_{k-n}) - (\vpi_{i_k}-\la_{k-n},\ \vpi_{i_k}+\la_{k})
=
N(\b_k)+2(\vpi_{i_k}-\la_{k-n},\ \b_k),$$ and the lemma follows.
Clearly, the rescaled generators $X_k$ satisfy the same commutation relations as the $Z_k$. Therefore, if we define for $\aa:=(a_1,\ldots,a_r)\in{\mathbb{Z}}^r$, $$X^{\aa}:= v^{\frac{1}{2}\sum_{i<j}a_ia_j(\b_i,\,\b_j)}\,X_1^{a_1}\cdots X_r^{a_r},$$ we have by Proposition \[lemcruc\], $$\si(X^{\aa}) = v^{-\frac{1}{2}\sum_{i<j}a_ia_j(\b_i,\,\b_j)}\,X_r^{a_r}\cdots X_1^{a_1}
= X^{\aa}.$$ Thus, $X^{\aa}$ is $\si$-invariant, and more generally an element of $\T_{v,\,Q}$ is $\si$-invariant if and only if all the coefficients of its expansion with respect to the basis $\{X^{\aa}\mid \aa\in{\mathbb{Z}}^r\}$ are invariant under the map $v^{1/2}\mapsto v^{-1/2}$. Moreover, one checks easily that $$\label{multXa}
X^{\aa}X^{\bb} = v^{\frac{1}{2}\sum_{i<j}(a_jb_i-a_ib_j)(\b_i,\,\b_j)}\,X^{\aa+\bb}
= v^{\sum_{i<j}(a_jb_i-a_ib_j)(\b_i,\,\b_j)}\,X^{\bb}X^{\aa}.$$
The isomorphism $\Phi$
----------------------
We can now state the main result of this section, which follows immediately from Proposition \[prop2.2\] and Equations (\[mult\_mono\]), (\[multXa\]).
\[main\_prop\] There is a ${\mathbb{C}}$-algebra isomorphism $\Phi\colon \Y_{t,\,Q} \to \T_{v,\,Q}$ given by $$\Phi(t^{1/2}) = v^{1/2},\qquad
\Phi(Y_{i,p}) = X_k \quad\mbox{for}\quad (i,p)\in \hI_Q\quad \mbox{and}\quad \varphi(i,p) = (\b_k,0).$$ More generally, let $$m = \prod_{(i,p)\in \hI_Q} Y_{i,p}^{u_{i,p}(m)}$$ be a *commutative* monomial in $\Y_{t,\,Q}$, as in §\[commut\_prod\], and let $\aa = (a_1,\ldots,a_r)$ where $a_k = u_{(i,p)}(m)$ if $\varphi(i,p)=(\b_k,0)$. Then we have $$\Phi(m)
= X^{\aa}.$$
Quantum groups {#sect4}
==============
Background {#defAqn}
----------
Let $\n$ denote a maximal nilpotent subalgebra of $\g$. Let $U_v(\n)$ be the Drinfeld-Jimbo quantum enveloping algebra of $\n$ over ${\mathbb{C}}(v)$, with Chevalley generators $e_i\ (i\in I)$ subject to the quantum Serre relations: $$\begin{array}{lll}
&e_i\,e_j - e_j\,e_i = 0 &\mbox{if $C_{ij} = 0$,}\\[2mm]
&e_i^2\,e_j - (v+v^{-1})e_i\,e_j\,e_i + e_j\,e_i^2= 0\qquad &\mbox{if $C_{ij} = -1$.}
\end{array}$$ It is endowed with a natural scalar product $(\cdot,\cdot)$ which we normalize by $(e_i,e_i) = 1$ (see [@GLS §4.3]). We denote by $A_v(\n)$ the graded dual vector space of $U_v(\n)$. The map $x \mapsto (x,\cdot)$ is a vector space isomorphism from $U_v(\n)$ to $A_v(\n)$, which allows to define a multiplication on $A_v(\n)$ by transporting the multiplication of $U_v(\n)$.
Thus, $U_v(\n)$ and $A_v(\n)$ are isomorphic algebras, but they have dual integral forms and therefore they specialize differently at $v=1$. One should regard $A_v(\n)$ as a quantum coordinate ring of the unipotent group $N$ with Lie algebra $\n$. For example, the elements $D_{\vpi_{i_k},\,\la_k}$ of §\[qtorusA\] are quantum analogues of certain generalized minors on $N$. We set $$\U_v(\n) := {\mathbb{C}}(v^{1/2})\otimes_{{\mathbb{C}}(v)}U_v(\n),
\qquad
{\mathcal{A}}_v(\n) := {\mathbb{C}}(v^{1/2})\otimes_{{\mathbb{C}}(v)}A_v(\n).$$ Since the basis involved in Theorem \[mainTh\] (b) is the dual canonical basis $\bB^*$, it is more natural to think of the quantum algebra of Theorem \[mainTh\] (a) as being ${\mathcal{A}}_{v}(\n)$ rather than $\U_{v}(\n)$.
The algebra $U_v(\n)$ has a natural grading by the root lattice of $\g$, given by $\deg(e_i)=\a_i$. The above isomorphism allows to transfer this grading to $A_v(\n)$.
Determinantal identities
------------------------
In [@GLS], it is shown that $A_v(\n)$ has a quantum cluster algebra structure. In particular, an explicit realization of $A_v(\n)$ as a subalgebra of the quantum torus $\TT_{v,\,Q}$ is given. This goes as follows.
For $u, w \in W$ and $\la\in P_+$, one has unipotent quantum minors $D_{u(\la),w(\la)}\in A_v(\n)$ (see [@GLS §5.2]). They satisfy $$D_{u(\la),w(\la)} =
\left\{
\begin{array}{cl}
1 & \mbox{if $u(\la)=w(\la)$,}\\[2mm]
0 & \mbox{if $u(\la)\not \le w(\la)$}.
\end{array}
\right.$$ Let $\ii = (i_1,\ldots,i_r)$ be as in §\[qtorusA\]. In [@GLS §5.4], a system of identities relating the quantum minors $$D(b,d) := D_{s_{i_1}\cdots s_{i_b}(\vpi_{i_b}),\ s_{i_1}\cdots s_{i_d}(\vpi_{i_b})},
\qquad
(0\le b \le d \le r,\ \ i_b=i_d\in I),$$ is derived, which we now recall. By convention, we write $D(0,b) = D_{\vpi_{i_b},\ s_{i_1}\cdots s_{i_b}(\vpi_{i_b})}$. Note that the minors $D(0,b)\ (1\le b\le r)$ form by definition a system of generators of $\TT_{v,\,Q}$. We will also use the following shorthand notation: $$\begin{aligned}
b^-(j) &:=& \max\left(\{s < b \mid i_s = j\}\cup\{0\}\right),\\
b^- &:=& \max\left(\{s < b \mid i_s = i_b\}\cup\{0\}\right),\label{defb-}\\
\mu(b,j) &:=& s_{i_1}\cdots s_{i_b}(\vpi_j).\label{equamu}\end{aligned}$$ In (\[equamu\]) we understand that $\mu(0,j) = \vpi_j$.
\[Tsystem\] Let $1\le b<d\le r$ be such that $i_b = i_d = i$. There holds $$\label{eqTsystem}
v^A\, D(b,d) D(b^-,d^-) =
v^{-1+B}\, D(b,d^-)D(b^-,d)
\ +\ v^C
\prod_{j\sim i}^{\longrightarrow}D(b^-(j),d^-(j))$$ where $$A = (\mu(d,i),\,\mu(b^-,i)-\mu(d^-,i)),\qquad
B = (\mu(d^-,i),\,\mu(b^-,i)-\mu(d,i)),$$ and $$C
=
\sum\limits_{\substack{j<k\\ j\sim i,\ k\sim i}}
\left(\mu(d,j),\,\mu(b,k)-\mu(d,k)\right).$$
This system of identities allows to express inductively every minor $D(b,d)$ as a rational function of the flag minors $D(0,c)$. Moreover, it follows from [@GLS Theorem 12.3] that all these rational functions belong in fact to $\TT_{v,\,Q}$, and that $A_v(\n)$ is the subalgebra of $\TT_{v,\,Q}$ generated by the minors $D(b^-,b)\ (1\le b \le r)$.
The dual canonical basis $\bB^*$ {#bases}
--------------------------------
Let us write $$E^*(\b_k) := D(k^-,k),\qquad (1\le k \le r),$$ and for $\aa = (a_1,\ldots,a_r)\in{\mathbb{Z}}^r$, $$E^*(\aa) = v^{-\sum_{k=1}^r a_k(a_k-1)/2}E^*(\b_1)^{a_1}\cdots E^*(\b_r)^{a_r}.$$ Then $\bE^*=\{E^*(\aa) \mid \aa \in{\mathbb{Z}}^r\}$ is a ${\mathbb{C}}(v)$-basis of $A_v(\n)$, dual to a basis of $U_v(\n)$ of PBW-type, as defined by Lusztig. The basis $\bE^*$ is called the *dual PBW-basis* of $A_v(\n)$.
The involution $\si$ of $\T_{v,\,Q}$ (see §\[sigma\]) can be restricted to $A_v(\n)$. Lusztig [@Lu] has constructed a canonical basis $\bB$ of $U_v(\n)$. The dual basis $\bB^* = \{B^*(\aa) \mid \aa\in{\mathbb{N}}^r\}$ of $A_v(\n)$ can be characterized as follows (see [@GLS]).
\[caracter\] For $\aa=(a_1,\ldots,a_r)\in{\mathbb{N}}^r$, the vector $B^*(\aa)$ is uniquely determined by the following conditions:
- $B^*(\aa) \in E^*(\aa) + \sum_{\cc\not = \aa} v^{-1}{\mathbb{Z}}[v^{-1}] E^*(\cc)$;
- let $\b(\aa) := \sum_{1\le k\le r}a_k \b_k$. Then $\si(B^*(\aa)) = v^{N(\b(\aa))}B^*(\aa)$.
The integer $N(\ga)$ of (b) is defined in (\[defNgamma\]). Note that $\b(\aa)$ is just the weight of $B^*(\aa)$ or $E^*(\aa)$ in the natural grading of $A_v(\n)$ by the root lattice of $\g$. The basis $\bB^*$ is called the *dual canonical basis* of $A_v(\n)$.
Quantum Grothendieck rings {#sect5}
==========================
Background {#background}
----------
For recent surveys on the representation theory of quantum loop algebras, we invite the reader to consult [@CH] or [@L].
Let $L\g$ be the loop algebra attached to $\g$, and let $U_q(L\g)$ be the associated quantum enveloping algebra. We assume that the deformation parameter $q\in{\mathbb{C}}^*$ is not a root of unity.
By [@FR], every finite-dimensional $U_q(L\g)$-module $M$ (of type 1) has a $q$-character $\chi_q(M)$. These $q$-characters generate a commutative ${\mathbb{C}}$-algebra isomorphic to the complexified Grothendieck ring of the category of finite-dimensional irreducible $U_q(L\g)$-modules. Nakajima [@N], Varagnolo and Vasserot [@VV], and Hernandez [@H1], have studied $t$-deformations of the $q$-characters of the standard modules and of the simple modules, as well as corresponding $t$-deformations of the Grothendieck ring. Although slightly different, these $t$-deformed Grothendieck rings are essentially equivalent, and in particular they give rise to the same $(q,t)$-characters for the simple modules. In what follows, we will use the version of [@H1]. Its definition will be recalled in the next sections.
The subcategory $\CC_{\mathbb{Z}}$ {#subsectCZ}
----------------------------------
The simple finite-dimensional irreducible $U_q(L\g)$-modules (of type 1) are usually labelled by Drinfeld polynomials. Here we shall use an alternative labelling by dominant monomials (see [@FR]). Moreover, as in [@HL], we shall restrict our attention to a certain tensor subcategory $\CC_{\mathbb{Z}}$ of the category of finite-dimensional $U_q(L\g)$-modules. The simple modules in $\CC_{\mathbb{Z}}$ are labelled by the dominant monomials in $\Y$, or equivalently, by the dominant commutative monomials in $(\Y_{t^{1/2}},*)$ (see §\[commut\_prod\]), and their $q$-characters belong to $\Y$. We shall denote by $L(m)$ the simple module labelled by the dominant monomial $m$. When $m=Y_{i,p}$ is reduced to a single variable, $L(m)$ is called a *fundamental module*. When $m$ is the only dominant monomial occuring in $\chi_q(L(m))$, $L(m)$ is said to be *minuscule*. Fundamental modules are examples of minuscule modules [@FM].
Standard modules
----------------
To a dominant commutative monomial $m$ is also attached a tensor product of fundamental modules called a *standard module* $M(m)$ defined by $$\label{standard}
M(m) := \ \stackrel{\longrightarrow}{\bigotimes_{(i,p)\in\hI}} L( Y_{i,p})^{\otimes\, u_{i,p}(m)},$$ where the product is ordered according to the following partial order on $\hI$: $$(i,p) < (j,s) \quad \Longleftrightarrow \quad p < s.$$ Note that for any fixed $p\in {\mathbb{Z}}$ and any total order on $I$, the tensor product $$\stackrel{\longrightarrow}{\bigotimes_{i\in I}} L( Y_{i,p})^{\otimes\, u_{i,p}(m)}$$ is irreducible, and its isomorphism class $L\left(\prod_{i\in I}Y_{i,p}^k\right)$ does not depend on the order of the factors, hence (\[standard\]) is well defined up to isomorphism (see [@FM Proposition 6.15]). The classes $[M(m)]$ of the standard modules $M(m)$ form a second basis of the Grothendieck group of $\CC_{\mathbb{Z}}$.
The ring $\K_t$
---------------
We introduce the commutative monomials [@FR] $$\label{defA}
A_{i,p} = Y_{i,p+1}Y_{i,p-1}\prod_{j \sim i} Y_{j,p}^{-1},\qquad ((i,p-1)\in \hI).$$ Recall from §\[commut\_prod\] that commutative monomials in $\Y$ can be regarded as elements of $(\Y_{t},*)$. More generally, the commutative polynomials $$Y_{i,p}\left(1+A_{i,\,p+1}^{-1}\right)
= Y_{i,p}\ + \ Y_{i,p+2}^{-1}\prod_{j \sim i} Y_{j,p+1},\qquad ((i,p)\in \hI)$$ can be regarded as elements of $(\Y_{t},*)$. For $i\in I$, let $\mathcal{K}_{i,t}$ be the ${\mathbb{C}}(t^{1/2})$-subalgebra of $\Y_{t}$ (for the noncommutative product $*$) generated by $$Y_{i,p}\left(1+A_{i,\,p+1}^{-1}\right),\quad
Y_{j,s}^{\pm 1},
\qquad \left((i,p), (j,s) \in \hI,\ j\not = i\right).$$ (In [@H0], $\mathcal{K}_{i,t}$ is identified with the kernel of a $t$-deformed screening operator.) Define $$\mathcal{K}_t := \bigcap_{i\in I}\mathcal{K}_{i,t}.$$ It is shown in [@H1] that an element of $\mathcal{K}_t$ is uniquely determined by the coefficients of its dominant monomials. Moreover, for any dominant monomial $m$, there is a unique $F(m)\in\mathcal{K}_t$ such that $m$ occurs in $F(m)$ with multiplicity $1$ and no other dominant monomial occurs in $F(m)$. These $F(m)$ form a ${\mathbb{C}}(t^{1/2})$-basis of $\mathcal{K}_t$.
Comparison with other $t$-deformations {#differ}
--------------------------------------
The product $*$ used in this paper is the same as that of [@H1], except that we have replaced $t$ by $t^{-1}$. The product of [@H1] is slightly different from the products of [@N] and [@VV] (see Remark \[Remark\_differ\]). However, as shown in [@H1 Proposition 3.16], for every $(i,p), (j,s) \in \hI$ the pairs $(Y_{i,p}, A_{j,s})$ and $(A_{i,p}, A_{j,s})$ are $t$-commutative with the *same exponents of $t$* for the three products of [@N; @VV; @H1]. This implies that the $t$-deformations of the Grothendieck ring ${\mathcal{R}}$ of $\CC_{\mathbb{Z}}$ associated with the three products are essentially equivalent, as will be explained below.
$(q,t)$-characters of standard modules {#sectqtstandard}
--------------------------------------
For a dominant commutative monomial $m\in \Y_{t^{1/2}}$, define $$\label{defMt}
[M(m)]_t := t^{\alpha(m)}\,
\underset{p\in{\mathbb{Z}}}{\overset{\longleftarrow}{\mbox{\Large *}}}
\,F\left(\prod_{i\in I} Y_{i,p}^{u_{i,p}(m)}\right)
\in \mathcal{K}_t.$$ Here, $\alpha(m)\in\frac{1}{2}{\mathbb{Z}}$ is chosen so that $m$ occurs with coefficient $1$ in the expansion of $[M(m)]_t$ on the basis of commutative monomials of $\Y_{t^{1/2}}$. The coefficients of $[M(m)]_t$ on this basis belong to ${\mathbb{Z}}[t^{\pm1}]$ and may therefore be specialized at $t=1$. The obtained specialization of $[M(m)]_t$ at $t=1$ is equal to $\chi_q(M(m))$, the $q$-character of the standard module $M(m)$. Therefore we may use the alternative notation $$\chi_{q,t}(M(m)) := [M(m)]_t,$$ and call this element of $\K_t$ the *$(q,t)$-character of $M(m)$*.
The bar involution {#susbsectinvol}
------------------
One shows that there is a unique ${\mathbb{C}}$-algebra anti-automorphism of $(\Y_{t},*)$ such that $$\overline{t^{1/2}} = t^{-1/2}, \quad
\overline{Y_{i,p}}=Y_{i,p},\qquad ((i,p)\in \hI).$$ Clearly, the *commutative* monomials are bar-invariant, as in [@N; @VV]. The subring $\mathcal{K}_t$ is stable under the bar involution, since each $\K_{i,t}$ is obviously stable. It follows that the elements $F(m)$ are bar-invariant (since $m$ is the unique dominant monomial of $\overline{F(m)}$). Hence the coefficients of the expansion of $F(m)$ on the basis of commutative monomials are unchanged under the replacement of $t$ by $t^{-1}$. Therefore, $F(m)$ is the same as in [@H1]. Since we have used in (\[defMt\]) the reverse product $\overset{\leftarrow}{*}$, the elements $\chi_{q,t}(M(m))$ also coincide with the corresponding elements of [@N; @VV; @H1], the coefficients of their expansion on the basis of commutative monomials are the same.
$(q,t)$-characters of simple modules {#sectqtsimple}
------------------------------------
\[defLt\] For every dominant monomial $m$, there is a unique element $[L(m)]_t$ of $\K_t$ satisfying
- $\overline{[L(m)]_t} = [L(m)]_t$,
- $\ds [L(m)]_t \in [M(m)]_t + \sum_{m' < m} t^{-1}{\mathbb{Z}}[t^{-1}]\, [M(m')]_t.$
Here $m'\leq m$ means that $m(m')^{-1}$ is a product of elements $A_{i,p}$ in $\Y$.
By §\[susbsectinvol\], the elements $[L(m)]_t$ coincide with the corresponding elements of [@N; @VV; @H1]. Using the geometry of quiver varieties, Nakajima has shown:
\[NakaTh\] The specialization of $[L(m)]_t$ at $t=1$ is equal to $\chi_q(L(m))$, and the coefficients of the expansion of $[L(m)]_t$ as a linear combination of monomials in the $Y_{i,p}$’s belong to ${\mathbb{N}}[t^{\pm 1}]$.
Therefore we may use the alternative notation $$\chi_{q,t}(L(m)) := [L(m)]_t,$$ and call this element of $\K_t$ the *$(q,t)$-character of $L(m)$*.
- If $L(m)$ is minuscule, $\chi_{q,t}(L(m)) = F(m)$.
- If $\chi_q(L(m))$ is multiplicity-free, then $\chi_{q,t}(L(m)) = \chi_q(L(m))$ does not depend on $t$ when expressed on the basis of commutative monomials.
By the positivity statement of Theorem \[NakaTh\], every monomial occuring in $\chi_{q,t}(L(m))$ already occurs in $\chi_{q}(L(m))$. Thus, if $L(m)$ is minuscule then $\chi_{q,t}(L(m))$ is an element of $\K_t$ containing the unique dominant monomial $m$, which proves (a). If $\chi_q(L(m))$ is multiplicity-free, then the coefficient of every commutative monomial in $\chi_{q,t}(L(m))$ is of the form $t^k$ for some $k\in{\mathbb{Z}}$. But since $\chi_{q,t}(L(m))$ is bar-invariant, we must have $k=0$, which proves (b).
Multiplicative structure
------------------------
We shall regard the noncommutative ring $(\K_t,*)$ as a *$t$-deformed version of the Grothendieck ring ${\mathcal{R}}$*. But one should be aware that only the simple modules $L(m)$ and the standard modules $M(m)$ have well-defined “classes” $\chi_{q,t}(L(m))$ and $\chi_{q,t}(M(m))$ in $\K_t$.
For any dominant monomials $m_1$ and $m_2$, write $$\chi_{q,t}(L(m_1))*\chi_{q,t}(L(m_2)) = \sum_m c_{m_1,\,m_2}^{m}(t^{1/2})\, \chi_{q,t}(L(m)).$$ Note that every irreducible $(q,t)$-character is of the form $\chi_{q,t}(L(m)) = m(1 + \sum_{k} M_k)$, where the $M_k$ are monomials in the $A_{i,p}^{-1}$ with coefficients in ${\mathbb{N}}[t,t^{-1}]$ (see [@H1]). So, by §\[differ\], the above coefficients $c_{m_1,\,m_2}^{m}(t^{1/2})$ are obtained from the corresponding ones in [@N; @VV] by multiplying by some $t^k$ with $k\in{\mathbb{Z}}/2$. Varagnolo and Vasserot have shown the following positivity result:
\[posmult\] The structure constants $c_{m_1,\,m_2}^{m}(t^{1/2})$ belong to ${\mathbb{N}}[t^{1/2},t^{-1/2}]$.
\[tcomm\] $L(m_1)\otimes L(m_2) \simeq L(m)$ is a simple module if and only if $$\chi_{q,t}(L(m_1))*\chi_{q,t}(L(m_2))=t^{2k}\chi_{q,t}(L(m_2))*\chi_{q,t}(L(m_1))
= t^k\chi_{q,t}(L(m))$$ for some $k\in{\mathbb{Z}}/2$.
If $L(m_1)\otimes L(m_2) \simeq L(m)$ then $\chi_{q}(L(m_1))*\chi_{q}(L(m_2)) = \chi_{q}(L(m))$. Hence $c_{m_1,\,m_2}^{m}(1) = 1$, and it follows from Theorem \[posmult\] that $\chi_{q,t}(L(m_1))*\chi_{q,t}(L(m_2)) = t^k \chi_{q,t}(L(m))$ for some $k\in{\mathbb{Z}}/2$. Applying the bar involution, we get $\chi_{q,t}(L(m_2))*\chi_{q,t}(L(m_1)) = t^{-k} \chi_{q,t}(L(m))$. If conversely $\chi_{q,t}(L(m_1))*\chi_{q,t}(L(m_2))= t^k\chi_{q,t}(L(m))$, then specializing $t$ to $1$ we get $$\chi_{q}(L(m_1))\chi_{q}(L(m_2))= \chi_{q}(L(m_1)\otimes L(m_2)) =\chi_{q}(L(m)),$$ hence $L(m_1)\otimes L(m_2) \simeq L(m)$.
Quantum $T$-system
------------------
For $(i,p)\in \hI$ and $k\in{\mathbb{N}}$, let $m^{(i)}_{k,p}:= Y_{i,p}Y_{i,p+2}\cdots Y_{i,p+2k-2}$. The simple $U_q(L\g)$-module $$W^{(i)}_{k,p} := L\left(m^{(i)}_{k,p}\right)$$ is called a *Kirillov-Reshetikhin module*. (By convention, if $k=0$ then $W^{(i)}_{k,p}$ is the trivial one-dimensional module.) The $q$-characters of the Kirillov-Reshetikhin modules satisfy the following system of algebraic identities called *$T$-system* [@KNS; @N-KR; @H2]. For every $(i,p)\in\hI$ and $k>0$, there holds $$\chi_q\left(W^{(i)}_{k,p}\right)\chi_q\left(W^{(i)}_{k,p+2}\right) =
\chi_q\left(W^{(i)}_{k-1,p+2}\right)\chi_q\left(W^{(i)}_{k+1,p}\right)
+
\prod_{j\not= i} \chi_q\left(W^{(j)}_{k,p+1}\right)^{-C_{ij}}.$$ This can be lifted to a $t$-deformed $T$-system in $\K_t$, as shown by the next proposition (see also [@N-KR §4], where a different $t$-deformed product is used, as explained in Remark \[Remark\_differ\] and §\[differ\]). Before stating it, we note that $\bigotimes_{j\sim i} W^{(j)}_{k,p+1}$ is a simple module, hence by Corollary \[tcomm\] the $(q,t)$-characters $\chi_{q,t}(W^{(j)}_{k,p+1})$ pairwise $t$-commute in $\K_t$. Moreover, it is easy to check that, since $\widetilde{C}(z)$ is symmetric, for any $j\sim i$ and $j'\sim i$, one has $m^{(j)}_{k,p+1}*m^{(j')}_{k,p+1}=m^{(j')}_{k,p+1}*m^{(j)}_{k,p+1}$, hence the $(q,t)$-characters $\chi_{q,t}(W^{(j)}_{k,p+1})$ do in fact pairwise commute in $\K_t$. So we may write $\underset{j\sim i}{*} \chi_{q,t}\left(W^{(j)}_{k,p+1}\right)$ without specifying an ordering of the factors.
\[Tsystem2\] In $\K_t$ there holds: $$\chi_{q,t}\left(W^{(i)}_{k,p}\right)*\chi_{q,t}\left(W^{(i)}_{k,p+2}\right) =
t^{\a(i,k)}\chi_{q,t}\left(W^{(i)}_{k-1,p+2}\right)*\chi_{q,t}\left(W^{(i)}_{k+1,p}\right)
+
t^{\ga(i,k)}
\underset{j\sim i}{*}
\chi_{q,t}\left(W^{(j)}_{k,p+1}\right),$$ where $$\label{qTsyst}
\a(i,k) = -1 + \frac{1}{2}\left(\widetilde{C}_{ii}(2k-1)+\widetilde{C}_{ii}(2k+1)\right),
\qquad
\ga(i,k) = \a(i,k)+1.$$
Using Theorem \[posmult\], we see that the claimed identity holds for some integers $\a(i,k)$ and $\ga(i,k)$, and we only have to check (\[qTsyst\]). To do so it is enough to compare the coefficients of some particular monomials on both sides. We have $m_{k,p}^{(i)}*m_{k,p+2}^{(i)}= t^{\a}m_{k-1,p+2}^{(i)}*m_{k+1,p}^{(i)}$, where $$\begin{aligned}
\a &=& \sum_{a=1}^{k-1}\cN(i,p;\,i,p+2a) + \frac{1}{2}\cN(i,p;\,i,p+2k) \\
&=& \sum_{a=1}^{k-1} \left(\widetilde{C}_{ii}(2a+1) - \widetilde{C}_{ii}(2a-1)\right)
+ \frac{1}{2} \left(\widetilde{C}_{ii}(2k+1) - \widetilde{C}_{ii}(2k-1)\right)\\
&=& -\widetilde{C}_{ii}(1)+\frac{1}{2}\left(\widetilde{C}_{ii}(2k-1) + \widetilde{C}_{ii}(2k+1)\right).\end{aligned}$$ Thus $\a(i,k) = \a = -1 + (\widetilde{C}_{ii}(2k-1)+\widetilde{C}_{ii}(2k+1))/2$, as claimed.
Similarly, $\chi_q(W^{(i)}_{k,p})$ contains the monomial $m:=m_{k,p}^{(i)}A_{i,p+2k-1}^{-1}\cdots A_{i,p+3}^{-1} A_{i,p+1}^{-1}$ with coefficient 1, and we have $m\,m_{k,p+2}^{(i)} = \prod_{j\sim i} m^{(j)}_{k,p+1}$. Now $$m * m_{k,p+2}^{(i)} =
\left(\left(m_{k,p+2}^{(i)}\right)^{-1}\prod_{j\sim i} m^{(j)}_{k,p+1}\right) * m_{k,p+2}^{(i)}=t^{\ga}
\prod_{j\sim i}m^{(j)}_{k,p+1},$$ where $$\begin{aligned}
\ga &=&
\frac{1}{2}\sum_{j\sim i}\sum_{a=1}^k\sum_{b=1}^{k}\cN(j,p+2a-1;\, i,p+2b)\\
&=&
\frac{1}{2}\sum_{j\sim i}\sum_{a=1}^k\sum_{b=1}^{k}
\left(
\widetilde{C}_{ji}(2(a-b)-2) - \widetilde{C}_{ji}(2(a-b)) - \widetilde{C}_{ji}(2(b-a)) + \widetilde{C}_{ji}(2(b-a)+2)
\right)\\
&=&
\frac{1}{2}\sum_{j\sim i}\sum_{a=1}^{k}
\left(
\widetilde{C}_{ji}(2(a-k)-2) - \widetilde{C}_{ji}(2(a-1)) - \widetilde{C}_{ji}(2(1-a)) + \widetilde{C}_{ji}(2(k-a)+2)
\right)\\
&=&
\frac{1}{2}\sum_{j\sim i}\sum_{a=1}^{k}
\left(
- \widetilde{C}_{ji}(2(a-1)) + \widetilde{C}_{ji}(2(k-a)+2)
\right)\\
&=&
\frac{1}{2}\sum_{j\sim i}
\widetilde{C}_{ji}(2k).\end{aligned}$$ Thus $\ga(i,k) = \ga =
\left(\sum_{j\sim i}\widetilde{C}_{ji}(2k)\right)/2 =
\left(\sum_{j\sim i}\widetilde{C}_{ij}(2k)\right)/2 =
(\widetilde{C}_{ii}(2k-1)+\widetilde{C}_{ii}(2k+1))/2$, as claimed. Here, the last equality comes from the definition of $\widetilde{C}(z)$ (see the proof of Proposition \[formula\_inverse\]).
[ (a) Take $\g$ of type $A_1$. We have $$\widetilde{C}(z) = z-z^3+z^5-z^{7}+z^9 -\cdots,\qquad$$ hence $\a(k) = -1$ for all $k>0$. Thus we get $$\chi_{q,t}\left(W_{k,p}\right)*\chi_{q,t}\left(W_{k,p+2}\right) =
t^{-1}\chi_{q,t}\left(W_{k-1,p+2}\right)*\chi_{q,t}\left(W_{k+1,p}\right)
+ 1.$$ (b) Take $\g$ of type $A_3$. Choose $i=1$, $k=1$, and $p=0$. Using for example Proposition \[formula\_inverse\], we can calculate $$\widetilde{C}_{11}(z) = z-z^7+z^9-z^{15}+\cdots,\qquad
\widetilde{C}_{12}(z) = z^2-z^6+z^{10}-z^{14}+\cdots,$$ hence $$\a(1,1) = -1 + \frac{1}{2}\left(\widetilde{C}_{11}(1)+\widetilde{C}_{11}(3)\right) = -\frac{1}{2},
\qquad
\ga(1,1) = \frac{1}{2}\widetilde{C}_{12}(2) = \frac{1}{2}.$$ Thus Proposition \[Tsystem2\] gives $$\chi_{q,t}\left(W^{(1)}_{1,0}\right)*\chi_{q,t}\left(W^{(1)}_{1,2}\right) =
t^{-1/2} \chi_{q,t}\left(W^{(1)}_{2,0}\right) +
t^{1/2} \chi_{q,t}\left(W^{(2)}_{1,1}\right).$$ ]{}
The subcategory $\CC_Q$ {#sectCh'}
-----------------------
Recall the quantum torus $\Y_{t,\,Q}$ of §\[torusT\]. The dominant commutative monomials in $\Y_{t,\,Q}$ parametrize the simple objects of an abelian subcategory $\CC_Q$ of $\CC_{\mathbb{Z}}$. More precisely, we define $\CC_Q$ as the full subcategory of $\CC_{\mathbb{Z}}$ whose objects have all their composition factors of the form $L(m)$ where $m$ is a dominant commutative monomial in $\Y_{t,\,Q}$. When $Q$ is a sink-source orientation of the Dynkin diagram and the Coxeter number $h$ is even, $\CC_Q$ is one of the subcategories $\CC_\ell$ introduced in [@HL]; namely, $\CC_Q = \CC_{h'}$ where $h' = h/2 - 1$.
\[tensorCQ\] $\CC_Q$ is closed under tensor products, hence is a tensor subcategory of $\CC_{\mathbb{Z}}$.
This is a slight modification of the proof of [@HL Proposition 3.2]. Let $L(m)$ and $L(m')$ be in $\mathcal{C}_Q$. This means that $m$ and $m'$ are monomials in the variables $Y_{i,p}$, $(i,p)\in\hat{I}_Q$. If $L(m'')$ is a composition factor of $L(m)\otimes L(m')$ then $m''$ is a product of monomials of $\chi_q(L(m))$ and $\chi_q(L(m'))$. So we have $m''=m m' M$ where $M$ is a monomial in the $A_{j,r}^{-1}$. Then it is checked as in [@HL Section 5.2.4] that, for $m''$ to be dominant, these $(j,r)$ have to satisfy $(j,r-1)\in\hat{I}_Q$ and $(j,r+1)\in\hat{I}_Q$. It follows that $m''$ depends only on the variables $Y_{i,p}$, $(i,p)\in\hat{I}_Q$, because $\hI_Q$ is a “convex slice” of $\hI$, that is, it satisfies:
- if $(i,p),\, (i,p+2k) \in \hI_Q$ for $i\in I,\, p\in{\mathbb{Z}},\, k>0$, then $(i,p+2j)\in \hI_Q$ for $1\le j\le k-1$;
- if $(i,p),\, (i,p+2) \in \hI_Q$ for $i\in I,\, p\in{\mathbb{Z}}$, then for every $j\sim i$ we have $(j,p+1)\in \hI_Q$.
Hence the result.
[We continue Example \[exampleYQ\]. We take $\g$ of type $D_4$ and choose $Q$ as in Figure \[FigD4\_1\]. The simple objects of $\CC_Q$ are of the form $L(m)$, where $$m =Y_{1,0}^{u_{1,0}} Y_{1,-2}^{u_{1,-2}} Y_{1,-4}^{u_{1,-4}}
Y_{2,0}^{u_{2,0}} Y_{2,-2}^{u_{2,-2}} Y_{2,-4}^{u_{2,-4}}
Y_{3,1}^{u_{3,1}} Y_{3,-1}^{u_{3,-1}} Y_{3,-3}^{u_{3,-3}}
Y_{4,2}^{u_{4,2}} Y_{4,0}^{u_{4,0}} Y_{4,-2}^{u_{4,-2}}.$$ and $u_{i,p}\in{\mathbb{N}}$. ]{}
The ring $\K_{t,\,Q}$ and the truncated $(q,t)$-characters {#sectKtCh'}
----------------------------------------------------------
We denote by $\K_{t,\,Q}$ the ${\mathbb{C}}(t^{1/2})$-subalgebra of $\K_t$ spanned by the $(q,t)$-characters $\chi_{q,t}(L(m))$ of the simple objects $L(m)$ in $\CC_{Q}$. We call $\K_{t,\,Q}$ the *$t$-deformed Grothendieck ring of $\CC_{Q}$*.
The $(q,t)$-character of a simple object $L(m)$ of $\CC_{Q}$ contains in general many monomials $m'$ which do not belong to $\Y_{t,\,Q}$. By discarding these monomials we obtain a *truncated $(q,t)$-character*. We shall denote by $\bchi_{q,t}(L(m))$ the truncated $(q,t)$-character of $L(m)$. One checks that for a simple object $L(m)$ of $\CC_{Q}$, all the dominant monomials occuring in $\chi_{q,t}(L(m))$ belong to the truncated $(q,t)$-character $\bchi_{q,t}(L(m))$ (the proof is similar to that of [@HL] for the category $\CC_1$, as for the proof of Lemma \[tensorCQ\] above). Therefore the truncation map $$\chi_{q,t}(L(m))\mapsto \bchi_{q,t}(L(m))$$ extends to an injective algebra homomorphism from $\K_{t,\,Q}$ to $\Y_{t,\,Q}$. In the sequel we shall identify $\K_{t,\,Q}$ with the subalgebra of $\Y_{t,\,Q}$ given by the image of this homomorphism.
An isomorphism between quantum Grothendieck rings and quantum groups {#sect6}
====================================================================
The isomorphism between $\K_{t,\,Q}$ and ${\mathcal{A}}_{v}(\n)$
----------------------------------------------------------------
Recall the isomorphism $\Phi \colon \Y_{t,\,Q} \to \T_{v,\,Q}$ of Proposition \[main\_prop\], and the notation $${\mathcal{A}}_{v}(\n):={\mathbb{C}}(v^{1/2})\otimes A_v(\n).$$ Define the *rescaled dual canonical basis* of ${\mathcal{A}}_{v}(\n)$: $$\widetilde{\bB}^*:=\left\{\widetilde{B}^*(\aa):=v^{N(\b(\aa))/2}B^*(\aa)\mid B^*(\aa) \in \bB^*\right\}.$$ Clearly, the elements of $\widetilde{\bB}^*$ are invariant under the involution $\si$. The next theorem is Theorem \[mainTh\] in a slightly more precise formulation.
\[mainth\]
- $\Phi$ restricts to an isomorphism $$\K_{t,\,Q}\overset{\sim}{\longrightarrow} {\mathcal{A}}_{v}(\n).$$
- The basis of $\K_{t,\,Q}$ consisting of the irreducible truncated $(q,t)$-characters $\bchi_{q,\,t}(L(m))$ is mapped by $\Phi$ onto $\widetilde{\bB}^*$.
We introduce some necessary notation. For $1\le k \le r$, let $k_{\rm min} := \min\{1\le s \le r\mid i_s = i_k\}$. Set $k^{(0)}:= k$ and, for a negative integer $j$, define $k^{(j)}= (k^{(j+1)})^-$, where the notation $b^-$ is as in Equation (\[defb-\]). We also note that, by definition of $\la_k$ and $\b_k$, if $k^-\not = 0$ then $\tau^{-1}(\la_k)=\la_{k^-}$ and $\tau^{-1}(\b_k)=\b_{k^-}$.
Let us fix some $(i,p)\in \hI_Q$. By definition of $\Phi$, we have: $$\Phi(Y_{i,p}) = X_k \quad\mbox{for}\quad (i,p)\in \hI_Q\quad \mbox{and}\quad \varphi(i,p) = (\b_k,0).$$ Note that this relation between $(i,p)$ and $\b_k$ implies in particular that $i_k=i$. Since if $k^-\not = 0$, $$\varphi(i,p+2) = (\tau^{-1}(\b_k),0) = (\b_{k^-},0),$$ we deduce that $\Phi\left(Y_{i,p}Y_{i,p+2}Y_{i,p+4}\cdots Y_{i,\xi_i}\right)$ is equal up to a power of $v$ to $X_kX_{k^-}X_{k^{(-2)}}\cdots X_{k_{\rm min}}$, that is, up to a power of $v$, to $D(0,k)$. Since the commutative monomial $Y_{i,p}Y_{i,p+2}\cdots Y_{i,\xi_i}$ is bar-invariant, its image is $\si$-invariant, so it has to be equal to $v^{N(\vpi_{i_k}-\la_k)/2}D(0,k)$. Now $Y_{i,p}Y_{i,p+2}\cdots Y_{i,\xi_i}$ is equal to the truncated $(q,t)$-character of the Kirillov-Reshetikhin module $W^{(i)}_{1+(\xi_i-p)/2,\,p}$. Hence we have shown that $$\Phi\left(\bchi_{(q,t)}\left(W^{(i)}_{1+(\xi_i-p)/2,\,p}\right)\right)
=
v^{N(\vpi_{i_k}-\la_k)/2}D(0,k).$$
We now want to show that, more generally, for $1\le s\le (\xi_i-p)/2+1$ we have $$\label{toshow}
\Phi\left(\bchi_{(q,t)}\left(W^{(i)}_{s,\,p}\right)\right)
=
v^{N(\la_{k^{(-s)}}-\la_k)/2}D(k^{(-s)},k).$$ This will be proved by comparing Proposition \[Tsystem\] and Proposition \[Tsystem2\]. Let us denote by $$\bD(b,d) := v^{N(\la_b-\la_d)/2}D(b,d)$$ the rescaled quantum minors. Note that $$N(\la_b-\la_d) = \frac{1}{2}(\la_b -\la_d,\, \la_b - \la_d) - \deg(\la_b-\la_d) = (\la_b,\,\la_b-\la_d) - \deg(\la_b-\la_d).$$ We can rewrite Proposition \[Tsystem\] as $$\label{TD}
\bD(b,d) \bD(b^-,d^-) =
v^{X}\, \bD(b,d^-)\bD(b^-,d)
\ +\ v^Y
\prod_{j\sim i}^{\longrightarrow}\bD(b^-(j),d^-(j))$$ where $$X:= -1+B-A+\frac{1}{2}
\left((\la_b,\,\la_b-\la_d) + (\la_{b^-},\,\la_{b^-}-\la_{d^-})
-(\la_{b^-},\,\la_{b^-}-\la_d) - (\la_{b},\,\la_{b}-\la_{d^-})
\right),$$ and $$Y:=C-A+\frac{1}{2}\left((\la_b,\,\la_b-\la_d) + (\la_{b^-},\,\la_{b^-}-\la_{d^-})
-\sum_{j\sim i} \left(\la_{b^-(j)},\,\la_{b^-(j)}-\la_{d^-(j)}\right)
\right).$$ Replacing $A$ and $B$ by their values from Proposition \[Tsystem\], and simplifying the resulting expression, we easily get $$X=-1+\frac{1}{2}(\la_b+\la_{b^-},\,\la_{d^-}-\la_d).$$ Now, writing $i_b=i_d=i$ and $b = d^{(-s)}$, $$(\la_b+\la_{b^-},\,\la_{d^-}-\la_d)
=(\la_b,\la_{d^-})-(\la_{b^-},\la_d)
=(\vpi_i,\tau^{s-1}(\vpi_i))-(\vpi_i,\tau^{s+1}(\vpi_i)).
$$ Hence, using that $\tau^{s-1}(\vpi_i)-\tau^{s+1}(\vpi_i)= \tau^s(\ga_i)+\tau^{s-1}(\ga_i)$, by Proposition \[prop1.1\] we get $$X = -1 + \frac{1}{2}\left(\widetilde{C}_{ii}(2s-1)+\widetilde{C}_{ii}(2s+1) \right).$$ Similarly, replacing $A$ and $C$ by their values from Proposition \[Tsystem\], and simplifying the resulting expression, we get $$Y=(\vpi_i,\,\tau(\vpi_i)-\tau^{s+1}(\vpi_i))
+(\vpi_i,\,\vpi_i-\tau^s(\vpi_i))-\frac{1}{2}\sum_{j\sim i}\sum_{k\sim i} (\tau^{(\xi_j-\xi_k)/2}\vpi_j,\,\vpi_k-\tau^s(\vpi_k)).$$ Using the identities $$\vpi_i - \tau^s(\vpi_i) = \sum_{l=0}^{s-1}\tau^l(\ga_i),
\qquad
\tau^l(\ga_i)+\tau^{l+1}(\ga_i)=\sum_{k\sim i}\tau^{l+(1+\xi_k-\xi_i)/2}(\ga_k),$$ we get $$(\vpi_i,\,\tau(\vpi_i)-\tau^{s+1}(\vpi_i))
+(\vpi_i,\,\vpi_i-\tau^s(\vpi_i))
=
\sum_{k\sim i} \left(\tau^{(\xi_i-\xi_k-1)/2}(\vpi_i),\, \vpi_k-\tau^s(\vpi_k)\right),$$ hence, $$Y=
\frac{1}{2}\,\sum_{k\sim i}\left(
\tau^{(\xi_i-\xi_k-1)/2}\left(2\vpi_i-\sum_{j\sim i}\tau^{(\xi_j-\xi_i+1)/2}(\vpi_j)\right),\,\vpi_k-\tau^s(\vpi_k)
\right).$$ Now, $$2\vpi_i-\sum_{j\sim i}\tau^{(\xi_j-\xi_i+1)/2}(\vpi_j)
=
2\vpi_i-\sum_{j\sim i;\ \xi_j-\xi_i=1} \tau(\varpi_j) -\sum_{j\sim i;\ \xi_j-\xi_i=-1} \varpi_j
=
\a_i+ \sum_{j\sim i;\ \xi_j-\xi_i=1} \ga_j
= \ga_i.$$ Hence $$\begin{aligned}
Y&=&
\frac{1}{2}\,\sum_{k\sim i}\left(
\tau^{(\xi_i-\xi_k-1)/2}(\ga_i),\,\vpi_k-\tau^s(\vpi_k)
\right)\\
&=& - \frac{1}{2}\,\sum_{k\sim i}\left(\tau^{-s+(\xi_i-\xi_k-1)/2}(\ga_i),\,\vpi_k\right)\\
&=& \frac{1}{2}\,\sum_{k\sim i}\left(\tau^{s+(\xi_i-\xi_k-1)/2}(\ga_i),\,\vpi_k\right)\\
&=&
\frac{1}{2}\,\sum_{k\sim i}
\widetilde{C}_{ik}(2s). \end{aligned}$$ Therefore $X=\a(i,s)$, $Y=\ga(i,s)$, and by Proposition \[Tsystem2\] we see that, for any $(i,p')\in \hI_Q$, there holds in $\K_t$: $$\label{TW}
\chi_{q,t}\left(W^{(i)}_{s,\,p'}\right)*\chi_{q,t}\left(W^{(i)}_{s,\,p'+2}\right) =
t^{X}\chi_{q,t}\left(W^{(i)}_{s-1,\,p'+2}\right)*\chi_{q,t}\left(W^{(i)}_{s+1,\,p'}\right)
+
t^{Y}
\underset{j\sim i}{*}
\chi_{q,t}\left(W^{(j)}_{s,\,p'+1}\right).$$ It was shown in [@GLS] that one can express every quantum minor $D(b,d)$ as a (noncommutative) Laurent polynomial in the quantum flag minors $D(0,k)$, by means of an explicit sequence of applications of Proposition \[Tsystem\]. Equivalently, every rescaled quantum minor $\bD(b,d)$ can be expressed as a Laurent polynomial in the $\bD(0,k)$’s with coefficients in ${\mathbb{Z}}[v^{\pm1/2}]$, by means of an explicit sequence of applications of (\[TD\]). By comparing (\[TD\]) and (\[TW\]), we see that the $(q,t)$-character of $W_{s,\ p}^{(i)}$ (where $\varphi(i,p) = (\b_d,0)$ and $b = d^{(-s)}$) can be expressed by the *same* Laurent polynomial in the $(q,t)$-characters $$\chi_{q,t}\left(
W_{1+(\xi_j-p')/2,\ p'}^{(j)}
\right),
\qquad ((j,p')\in\hI_Q),$$ where $v^{1/2}$ is replaced by $t^{1/2}$. This proves (\[toshow\]). In particular, we have $$\Phi\left(\bchi_{q,t}\left(
L(Y_{i,p})
\right)\right)
=
\Phi\left(\bchi_{q,t}\left(
W_{1,p}^{(i)}
\right)\right)
=
\bD(d^-,d)
=v^{N(\b_d)/2}E^*(\b_d)
,
\qquad
((i,p)\in\hI_Q).$$ Thus, $\Phi$ maps the truncated $(q,t)$-characters of the fundamental modules of $\CC_Q$, that is, a set of algebra generators of $\K_{t,\,Q}$, to the rescaled dual PBW generators of ${\mathcal{A}}_{v}(\n)$. This proves (a).
It follows that $\Phi$ maps the truncated $(q,t)$-characters of the standard modules of $\CC_Q$ to the elements of the dual PBW-basis of ${\mathcal{A}}_v(\n)$ up to some power of $v$. Let us calculate this power of $v$. By Proposition \[defLt\], we have that $[M(m)]_t - [L(m)]_t$ is a linear combination of $[L(m')]_t$ with coefficients in $t^{-1}{\mathbb{Z}}[t^{-1}]$, where $[L(m)]_t$ and the $[L(m')]_t$ are bar-invariant. On the other hand, note that the rescaling factor $v^{N(\b(\aa))/2}$ of the dual canonical basis depends only on the weight of the vector $B^*(\aa)$. Hence if we write $\widetilde{E}^*(\aa) = v^{N(\b(\aa))/2} E^*(\aa)$, the transition matrix between the rescaled dual PBW-basis $\{\widetilde{E}^*(\aa)\}$ and the rescaled dual canonical basis $\{\widetilde{B}^*(\aa)\}$ is identical to the transition matrix between the original bases. Thus, by Proposition \[caracter\], $\widetilde{E}^*(\aa) - \widetilde{B}^*(\aa)$ is a linear combination of $\widetilde{B}^*(\aa')$ with coefficients in $v^{-1}{\mathbb{Z}}[v^{-1}]$, where $\widetilde{B}^*(\aa)$ and the $\widetilde{B}^*(\aa')$ are $\si$-invariant. By Proposition \[main\_prop\], $\Phi$ maps the set of bar-invariant elements of $\K_{t,\,Q}$ to the set of $\si$-invariant elements of ${\mathcal{A}}_{v}(\n)$. This implies that $\Phi$ maps the basis of $\K_{t,\,Q}$ given by the truncated $(q,t)$-characters of the standard modules of $\CC_{Q}$, to the rescaled dual PBW-basis $\{\widetilde{E}^*(\aa)\}$ of ${\mathcal{A}}_{v}(\n)$. Finally, using again Proposition \[caracter\] and Proposition \[defLt\], this yields (b).
Let $\g$ be of type $A_2$. Let $Q$ be the quiver of type $A_2$ with height function $\xi_1=2$ and $\xi_2=1$. We have $\bi = (1,2,1)$, and $$D(0,1) = D_{\vpi_1,\,s_1(\vpi_1)},
\quad
D(0,2) = D_{\vpi_2,\,s_1s_2(\vpi_2)},
\quad
D(1,3) = D_{s_1(\vpi_1),\,s_1s_2s_1(\vpi_1)} = D_{\vpi_2,\,s_2(\vpi_2)}.$$ Let $e_1$ and $e_2$ be the Chevalley generators of $U_v(\n)$. In the identification $A_v(\n) \equiv U_v(\n)$ we have $D(0,1)\equiv e_1$ and $D(1,3)\equiv e_2$.
In this case the quantum torus $\Y_{t,\,Q}$ is generated by $Y_{1,0}, Y_{1,2}, Y_{2,1}$, so $\K_{t,\,Q}$ is generated by $\bchi_{q,t}(L(Y_{1,0})), \bchi_{q,t}(L(Y_{1,2})), \bchi_{q,t}(L(Y_{2,1}))$. The isomorphism $\Phi$ of Theorem \[mainth\] satisfies $$\Phi\left(\bchi_{q,t}(L(Y_{1,2}))\right) = D(0,1),
\quad
\Phi\left(\bchi_{q,t}(L(Y_{1,0}))\right)= D(1,3).$$ Thus Theorem \[mainth\] implies that $\bchi_{q,t}(L(Y_{1,2}))$ and $\bchi_{q,t}(L(Y_{1,0}))$ generate $\K_{t,\,Q}$ and satisfy the quantum Serre relations.
This can easily be checked by means of the quantum $T$-system. Indeed we have by Proposition \[Tsystem2\]: $$\chi_{q,t}(L(Y_{1,0}))*\chi_{q,t}(L(Y_{1,2}))
= t^{-1/2} \chi_{q,t}(L(Y_{1,0}Y_{1,2})) + t^{1/2} \chi_{q,t}(L(Y_{2,1})),$$ and by applying the bar-involution $$\chi_{q,t}(L(Y_{1,2}))*\chi_{q,t}(L(Y_{1,0}))
= t^{1/2} \chi_{q,t}(L(Y_{1,0}Y_{1,2})) + t^{-1/2} \chi_{q,t}(L(Y_{2,1})).$$ Eliminating $\chi_{q,t}(L(Y_{1,0}Y_{1,2}))$ we get $$(t^{-1/2}-t^{3/2})\chi_{q,t}(L(Y_{2,1})) =
\chi_{q,t}(L(Y_{1,2}))*\chi_{q,t}(L(Y_{1,0}))
- t \chi_{q,t}(L(Y_{1,0}))*\chi_{q,t}(L(Y_{1,2})),$$ which shows that $\K_{t,\,Q}$ is generated by $\bchi_{q,t}(L(Y_{1,2}))$ and $\bchi_{q,t}(L(Y_{1,0}))$. Finally, using that $$\chi_{q,t}(L(Y_{2,1}))*\chi_{q,t}(L(Y_{1,0})) = t^{-1} \chi_{q,t}(L(Y_{1,0}))*\chi_{q,t}(L(Y_{2,1}))$$ we obtain that $$\begin{aligned}
&&\chi_{q,t}(L(Y_{1,2}))*\chi_{q,t}(L(Y_{1,0}))^2
- t \chi_{q,t}(L(Y_{1,0}))*\chi_{q,t}(L(Y_{1,2}))*\chi_{q,t}(L(Y_{1,0}))\\[2mm]
&&\qquad=\
t^{-1}\chi_{q,t}(L(Y_{1,0}))*\chi_{q,t}(L(Y_{1,2}))*\chi_{q,t}(L(Y_{1,0}))
-\chi_{q,t}(L(Y_{1,0}))^2*\chi_{q,t}(L(Y_{1,2})), \end{aligned}$$ which is the first quantum Serre relation. The second one is obtained similarly.
[In this example, we illustrate the calculations behind the proof of Theorem \[mainth\]. Let $\g$ be of type $A_3$. Let $Q$ be the quiver of type $A_3$ with height function $\xi_1=\xi_3=2$ and $\xi_2=3$. Thus $Q$ has source $2$ and sinks $1$, $3$. We take $\bi = (2,1,3,2,1,3)$, hence $$\b_1=\a_2,\quad
\b_2=\a_1+\a_2,\quad
\b_3=\a_2+\a_3,\quad
\b_4=\a_1+\a_2+\a_3,\quad
\b_5=\a_3,\quad
\b_6=\a_1,$$ and $$\begin{array}{lll}
\la_1=\vpi_2-\a_2,&
\la_2=\vpi_1-\a_1-\a_2,&
\la_3=\vpi_3-\a_2-\a_3,\\[2mm]
\la_4=\vpi_2-\a_1-2\a_2-\a_3,&
\la_5=\vpi_1-\a_1-\a_2-\a_3,&
\la_6=\vpi_3-\a_1-\a_2-\a_3.
\end{array}$$ Note that in this case $w_0 = c^2$ where $c=s_2s_1s_3$ is a Coxeter element. Thus, this example illustrates also Corollary \[Cor\_cluster\]. The quantum unipotent minors generating $\T_{v,\,Q}$ are $$\begin{array}{lll}
D(0,1) = D_{\vpi_2,\,s_2(\vpi_2)},&
D(0,2) = D_{\vpi_1,\,s_2s_1(\vpi_1)},&
D(0,3) = D_{\vpi_3,\,s_2s_1s_3(\vpi_3)},\\[2mm]
D(0,4) = D_{\vpi_2,\,s_2s_1s_3s_2(\vpi_2)},&
D(0,5) = D_{\vpi_1,\,s_2s_1s_3s_2s_1(\vpi_1)},&
D(0,6) = D_{\vpi_3,\,w_0(\vpi_3)}.
\end{array}$$ The generators of the dual PBW-basis are $$\begin{array}{lll}
E^*(\b_1) = D(0,1),&
E^*(\b_2) = D(0,2),&
E^*(\b_3) = D(0,3),\\[2mm]
E^*(\b_4) = D_{s_2(\vpi_2),\,s_2s_1s_3s_2(\vpi_2)},&
E^*(\b_5) = D_{s_2s_1(\vpi_1),\,s_2s_1s_3s_2s_1(\vpi_1)},&
E^*(\b_6) = D_{s_2s_1s_3(\vpi_3),\,w_0(\vpi_3)}.
\end{array}$$ The new generators $X_i$ of $\T_{v,\,Q}$ are $$\begin{array}{lll}
X_1 = D(0,1),&
X_2 = v^{-1/2}D(0,2),&
X_3 = v^{-1/2}D(0,3),\\[2mm]
X_4 = v^{-1}D(0,4)D(0,1)^{-1},&
X_5 = v^{-1}D(0,5)D(0,2)^{-1},&
X_6 = v^{-1}D(0,6)D(0,3)^{-1}.
\end{array}$$ Let us define integers $\la_{ij}$ and $\mu_{ij}$ by $$D(0,i)D(0,j) = v^{\la_{ij}} D(0,j)D(0,i),
\qquad
X_i X_j = v^{\mu_{ij}} X_jX_i,
\qquad
(1\le i,j\le 6).$$ The matrices $L=[\la_{ij}]$ and $M=[\mu_{ij}]$ are given by $$L = \left[
\begin{matrix}
0&-1&-1&0&0&0\cr
1&0&0&0&1&-1\cr
1&0&0&0&-1&1\cr
0&0&0&0&0&0\cr
0&-1&1&0&0&0\cr
0&1&-1&0&0&0
\end{matrix}
\right],
\qquad
M = \left[
\begin{matrix}
0&-1&-1&0&1&1\cr
1&0&0&-1&1&-1\cr
1&0&0&-1&-1&1\cr
0&1&1&0&-1&-1\cr
-1&-1&1&1&0&0\cr
-1&1&-1&1&0&0
\end{matrix}
\right].$$ The generators of $\Y_{t,\,Q}$ are $Y_{1,0}, Y_{3,0}, Y_{2,1}, Y_{1,2}, Y_{3,2}, Y_{2,3}$. The isomorphism $\Phi$ is defined by $$\Phi(Y_{1,0}) = X_5, \quad
\Phi(Y_{3,0}) = X_6, \quad
\Phi(Y_{2,1}) = X_4, \quad
\Phi(Y_{1,2}) = X_2, \quad
\Phi(Y_{3,2}) = X_3, \quad
\Phi(Y_{2,3}) = X_1.$$ The truncated $(q,t)$-characters of the fundamental modules of $\CC_{Q}$ are expressed in terms of commutative monomials by $$\begin{array}{rclrcl}
\bchi_{q,t}(Y_{1,2})&=&Y_{1,2},
&\bchi_{q,t}(Y_{1,0})&=&Y_{1,0}+Y_{1,2}^{-1}Y_{2,1}+Y_{2,3}^{-1}Y_{3,2},
\\[2mm]
\bchi_{q,t}(Y_{2,1})&=&Y_{2,1}+Y_{1,2}Y_{2,3}^{-1}Y_{3,2},\ \
&\bchi_{q,t}(Y_{2,3})&=&Y_{2,3},
\\[2mm]
\bchi_{q,t}(Y_{3,2})&=&Y_{3,2},
&\bchi_{q,t}(Y_{3,0})&=&Y_{3,0}+Y_{3,2}^{-1}Y_{2,1}+Y_{2,3}^{-1}Y_{1,2},
\end{array}$$ Here, we have used the shorthand notation $\bchi_{q,t}(m)$ instead of $\bchi_{q,t}(L(m))$. We also have $$\bchi_{q,t}(Y_{1,0}Y_{1,2})=Y_{1,0}Y_{1,2},\qquad
\bchi_{q,t}(Y_{2,1}Y_{2,3})=Y_{2,1}Y_{2,3},\qquad
\bchi_{q,t}(Y_{3,0}Y_{3,2})=Y_{3,0}Y_{3,2}.$$ Using the expression of $D(0,k)$ in terms of $X_j$’s and the definition of $\Phi$, one checks that $$\begin{array}{lll}
\Phi(Y_{2,3}) = D(0,1),&
\Phi(Y_{1,2}) = v^{-1/2}D(0,2),&
\Phi(Y_{3,2}) = v^{-1/2}D(0,3),\\[2mm]
\Phi(Y_{2,1}Y_{2,3})=v^{-1}D(0,4),&
\Phi(Y_{1,0}Y_{1,2})=v^{-1}D(0,5),&
\Phi(Y_{3,0}Y_{3,2}) = v^{-1}D(0,6),
\end{array}$$ in agreement with Theorem \[mainth\]. By Proposition \[Tsystem\], we have $$v^{-1}D(1,4)D(0,1) = v^{-1}D(1,1)D(0,4) + D(0,2)D(0,3),$$ hence $$v^{-1}D(1,4) = (v^{-1}D(0,4) + D(0,2)D(0,3))D(0,1)^{-1}.$$ Therefore $$\Phi^{-1}\left(v^{-1}D(1,4)\right) = (Y_{2,1}Y_{2,3} + tY_{1,2}*Y_{3,2})*Y_{2,3}^{-1}
= Y_{2,1} + t Y_{1,2}*Y_{3,2}*Y_{2,3}^{-1}
= Y_{2,1} + Y_{1,2}Y_{3,2}Y_{2,3}^{-1},$$ where the last equality follows from (\[mult\_mono\]). Thus we have $$\Phi^{-1}\left(v^{-1}D(1,4)\right) =\Phi^{-1}\left(v^{-1}E^*(\b_4)\right) = \bchi_{q,t}(Y_{2,1}),$$ in agreement with Theorem \[mainth\]. Next, we have again by Proposition \[Tsystem\], $$D(2,5)D(0,2) = v^{-1}D(0,5) + D(1,4).$$ Hence $$\Phi^{-1}(D(2,5))=\left(Y_{1,0}Y_{1,2} + t(Y_{2,1}+ Y_{1,2}Y_{2,3}^{-1}Y_{3,2})\right)*(t^{1/2}Y_{1,2})^{-1}.$$ Now, $$(Y_{1,0}Y_{1,2})*Y_{1,2}^{-1}=t^{1/2}Y_{1,0}*Y_{1,2}*Y_{1,2}^{-1} = t^{1/2}Y_{1,0},$$ and similarly $$Y_{2,1}*Y_{1,2}^{-1} = t^{-1/2}Y_{2,1}Y_{1,2}^{-1},\quad
(Y_{1,2}Y_{2,3}^{-1}Y_{3,2})*Y_{1,2})^{-1} = t^{-1/2} Y_{2,3}^{-1}Y_{3,2}.$$ Therefore, $$\Phi^{-1}(D(2,5))= \Phi^{-1}\left(E^*(\b_5)\right) =
Y_{1,0} + Y_{1,2}^{-1}Y_{2,1} + Y_{2,3}^{-1}Y_{3,2} = \bchi_{q,t}(Y_{1,0}).$$ Similarly, starting from the minor identity $$D(3,6)D(3,0) = v^{-1}D(0,6) + D(1,4).$$ we deduce that $$\Phi^{-1}(D(3,6))=\Phi^{-1}\left(E^*(\b_6)\right)= Y_{3,0} + Y_{3,2}^{-1}Y_{2,1} + Y_{2,3}^{-1}Y_{1,2} = \bchi_{q,t}(Y_{3,0}).$$ Thus we have checked that $\Phi$ maps the fundamental characters $$\bchi_{q,t}(Y_{1,0}),\quad
\bchi_{q,t}(Y_{3,0}),\quad
\bchi_{q,t}(Y_{2,1}),\quad
\bchi_{q,t}(Y_{1,2}),\quad
\bchi_{q,t}(Y_{3,2}),\quad
\bchi_{q,t}(Y_{2,3}),$$ to the rescaled dual PBW generators $v^{N(\b_k)/2}E^*(\b_k)$, in agreement with Theorem \[mainth\]. ]{}
Proof of Corollary \[corollary\_binary\]
----------------------------------------
Let $b_1,\ldots,b_k\in\bB^*$, and let $L_1,\ldots, L_k$ be the simple objects of $\CC_Q$ such that $$\Phi(\bchi_{q,t}(L_i)) \in v^{{\mathbb{Z}}/2} b_i,\qquad (1\le i\le k).$$ We have $\Phi(\bchi_{q,t}(L_1)*\cdots *\bchi_{q,t}(L_k)) \in v^{{\mathbb{Z}}/2} b_1\cdots b_k$, thus, by Theorem \[mainth\], $b_1\cdots b_k \in v^{\mathbb{Z}}\bB^*$ if and only if $\bchi_{q,t}(L_1)*\cdots *\bchi_{q,t}(L_k)$ is the $(q,t)$-character of a simple module up to a power of $v$, that is by Corollary \[tcomm\], if and only if $L_1\otimes\cdots\otimes L_k$ is simple. Hence Corollary \[corollary\_binary\] follows from [@H3].
A presentation of quantum Grothendieck rings {#sect_pres}
============================================
In the remaining sections we drop the symbol $*$ for the $t$-deformed product of $\K_t$, and simply write $xy$ instead of $x*y$.
The generators {#generators}
--------------
Fix an orientation $Q$ of the Dynkin diagram of $\g$. Define an involution $\nu$ of $I$ by $w_0(\a_i) = -\a_{\nu(i)}$. For $i\in I$ write $\varphi^{-1}(\a_i,0) = (k_i,p_i)\in\hI_Q$. Define the following elements of $\K_t$: $$x_{i,m}^Q := \chi_{q,t}\left(L\left(Y_{\nu^m(k_i),p_i+mh}\right)\right),
\qquad
(i\in I,\ m\in{\mathbb{Z}}).$$ The elements $x_{i,0}^Q$ belong to $\K_{t,\,Q}$ and map to the Chevalley generators $D_{\vpi_i,\,s_i(\vpi_i)}\equiv e_i$ of ${\mathcal{A}}_{v}(\n)\equiv \U_{v}(\n)$ under the isomorphism $\Phi$ of Theorem \[mainth\]. Hence $\K_{t,\,Q}$ has a presentation given by the generators $x_{i,0}^Q\ (i\in I)$ subject to the relations (see §\[defAqn\]) $$\begin{array}{lll}
&x_{i,0}^Q\,x_{j,0}^Q - x_{j,0}^Q\,x_{i,0}^Q = 0 &\mbox{if $C_{ij} = 0$,}\\[2mm]
&(x_{i,0}^Q)^2\,x_{j,0}^Q - (t+t^{-1})x_{i,0}^Q\,x_{j,0}^Q\,x_{i,0}^Q + x_{j,0}^Q(\,x_{i,0}^Q)^2= 0\qquad &\mbox{if $C_{ij} = -1$.}
\end{array}$$ In particular, every $\chi_{q,t}\left(L(Y_{i,p})\right)$ with $(i,p)\in \hI_Q$ can be written as a noncommutative polynomial in the $x_{i,0}^Q$’s.
For $m\in{\mathbb{Z}}$, let $\K^{(m)}$ be the subalgebra of $\K_t$ generated by the $x_{i,m}^Q\ (i\in I)$. Thus $\K^{(0)} = \K_{t,\,Q}$, and $\K^{(m)}$ is isomorphic to $\K^{(0)}$ for every $m\in{\mathbb{Z}}$. This comes from the fact that $\K_t$ is generated by the fundamental $(q,t)$-characters $\chi_{q,t}\left(L(Y_{i,p})\right)\ ((i,p)\in\hI)$, and that the assignment $$\chi_{q,t}\left(L(Y_{i,p})\right) \mapsto \chi_{q,t}\left(L(Y_{\nu(i),\,p+h})\right)$$ extends to an algebra automorphism $\Sigma$ of $\K_t$. (In fact, $L(Y_{i,p})$ is the $U_q(L\g)$-module dual to $L(Y_{\nu(i),\,p+h})$ [@CP §5], see also [@FM Cor. 6.10].) Let $\hI_{Q,m} := \varphi^{-1}(\De_+\times \{m\})$. Thus $\chi_{q,t}\left(L(Y_{i,p})\right)\in \K^{(m)}$ for $(i,p)\in \hI_{Q,m}$. Therefore, we have proved:
The elements $x_{i,m}^Q\ (i\in I, m\in{\mathbb{Z}})$ generate $\K_t$.
The presentation
----------------
We start with the following:
\[lemme\_commut\] Let $(i,p)\in\hI$ and $(j,p+h)\in\hI$. Write $V := L(Y_{i,p})\otimes L(Y_{j,\,p+h})$.
- If $j\not = \nu(i)$ then $V$ is simple.
- If $j=\nu(i)$ then $\chi_q(V) = \chi_q\left(L(Y_{i,p}Y_{\nu(i),\,p+h})\right) + 1$.
- In general we have $$\chi_{q,t}\left(L(Y_{i,p})\right) \chi_{q,t}\left(L(Y_{j,p+h})\right)
= t^{-(\a_i,\a_{\nu(j)})} \chi_{q,t}\left(L(Y_{j,p+h})\right) \chi_{q,t}\left(L(Y_{i,p})\right) + \de_{i\,\nu(j)}(1-t^{-2}),$$ where $\de_{ik}$ is the Kronecker symbol $\de$.
Consider the product $\pi:=\chi_q\left(L(Y_{i,p})\right)\chi_q\left(L(Y_{\nu(i),\,p+h})\right)$. By [@FM §6], $\chi_q\left(L(Y_{i,p})\right)$ contains only one dominant monomial, namely $Y_{i,p}$, one anti-dominant monomial, namely $Y_{\nu(i),p+h}^{-1}$, and all its other monomials involve only variables of the form $Y_{j,m}^{\pm 1}$ with $p<m< p+h$. It follows that, if $j\not = \nu(i)$, then $\pi$ contains no other dominant monomial than $Y_{i,p}Y_{j,\,p+h}$, hence $V$ is irreducible and isomorphic to $L(Y_{i,p}Y_{\nu(i),\,p+h})$. This proves (a).
If $j=\nu(i)$ then $\pi$ contains only two dominant monomials, that is, $Y_{i,p}Y_{\nu(i),\,p+h}$ and $1$. Therefore $V$ has at most two composition factors, $L(Y_{i,p}Y_{\nu(i),\,p+h})$ and the trivial one-dimensional representation. Since $L(Y_{i,p}) = L(Y_{\nu(i),\,p+h})^*$, the trivial representation is indeed a composition factor of $V$ because $U_q(L\g)$ is a Hopf algebra. This proves (b).
It follows that $$\chi_{q}\left(L(Y_{i,p})\right) \chi_{q}\left(L(Y_{j,p+h})\right)
= \chi_{q}\left(L(Y_{i,p}Y_{j,p+h})\right) + \de_{i\,\nu(j)}.$$ In $\K_t$, this identity gets $t$-deformed as $$\chi_{q,t}\left(L(Y_{i,p})\right) \chi_{q,t}\left(L(Y_{j,p+h})\right)
= t^{\frac{1}{2}\cN(i,p;j,p+h)}\chi_{q,t}\left(L(Y_{i,p}Y_{j,p+h})\right) + \de_{i\,\nu(j)}.$$ Now using Proposition \[expression\_N\] and a sink-source orientation $Q$ where $i$ is a source, we see that $\cN(i,p;j,p+h) = \cN(i,0;j,h) =-(\a_i,\a_{\nu(j)})$. Using the bar involution, we also have $$\chi_{q,t}\left(L(Y_{j,p+h})\right) \chi_{q,t}\left(L(Y_{i,p})\right)
= t^{-\frac{1}{2}\cN(i,p;j,p+h)}\chi_{q,t}\left(L(Y_{i,p}Y_{j,p+h})\right) + \de_{i\,\nu(j)}.$$ Then (c) follows by eliminating $\chi_{q,t}\left(L(Y_{i,p}Y_{j,p+h})\right)$ between these two equations.
We can now give a presentation of $\K_t$.
\[presentation\] The algebra $\K_t$ is isomorphic to the ${\mathbb{C}}(t^{1/2})$-algebra $\AA$ presented by generators $y_{i,m}\ (i\in I, m\in{\mathbb{Z}})$ subject only to the following relations:
- for every $m\in{\mathbb{Z}}$, $$\begin{array}{lll}
&y_{i,m}\,y_{j,m} - y_{j,m}\,y_{i,m} = 0 &\mbox{if $(\a_i,\a_j) = 0$,}\\[2mm]
&y_{i,m}^2\,y_{j,m} - (t+t^{-1})y_{i,m}\,y_{j,m}\,y_{i,m} + y_{j,m}\,y_{i,m}^2= 0\qquad &\mbox{if $(\a_i,\a_j) = -1$;}
\end{array}$$
- for every $m\in{\mathbb{Z}}$ and every $i,j\in I$, $$y_{i,m}\, y_{j,\,m+1} = t^{-(\a_i,\a_j)} y_{j,\,m+1}\, y_{i,m} + \de_{ij}(1-t^{-2});$$
- for every $p>m+1$ and every $i,j\in I$, $$y_{i,m}\,y_{j,p} = t^{(-1)^{p-m}(\a_i,\a_j)} \,y_{j,p}\,y_{i,m}.$$
We fix a sink-source orientation $Q$. We first check that the $x_{i,m}^Q$ satisfy the above relations. The relations (R1) are the Drinfeld-Jimbo relations for the subalgebra $\K^{(m)}$, as explained in §\[generators\]. The relations (R2) follow from Lemma \[lemme\_commut\] (c) when $\xi_i=\xi_j$. If $\xi_i\not = \xi_j$, then $x_{i,m}^Q\, x_{j,\,m+1}^Q$ corresponds to a tensor product of the form $L(Y_{i,p})\otimes L(Y_{j,p+1})$ or $L(Y_{i,p})\otimes L(Y_{j,p+2h-1})$. These two types of tensor products are always irreducible [@FM Proposition 6.15]. Using Corollary \[tcomm\], it follows that $x_{i,m}$ and $x_{j,m+1}$ $t$-commute, and the exponent of $t$ is calculated by means of Proposition \[expression\_N\]. For the relations (R3) we note that $L(Y_{i,p})\otimes L(Y_{j,s})$ is irreducible if $s-p>h$ [@FM Proposition 6.15], and we conclude similarly.
It follows that we have a surjective homomorphism $F$ from $\AA$ to $\K_t$ given by $y_{i,m}\mapsto x_{i,m}^Q$, and we have to show that this is an isomorphism. Define $\AA^{(m)}$ as we have defined $\K^{(m)}$ before. Then $\AA^{(m)}$ is presented by the relations (1) (with $x_{i,m}^Q$ replaced by $y_{i,m}$), so $F$ restricts to an isomorphism from $\AA^{(m)}$ to $\K^{(m)}$. It follows from the relations (R2) and (R3) that every monomial $M$ in the $y_{i,m}$’s can be rewritten as a linear combination of monomials of the form $M_{k_1}M_{k_2}\cdots M_{k_s}$ with $M_{k_j}\in \AA^{(k_j)}$ and $k_1>k_2>\cdots>k_s$. So we have $\AA = \ds\prod_{m\in{\mathbb{Z}}}^{\leftarrow} \AA^{(m)}$. Now each $\K^{(m)}$ has a basis $\B^{(m)}$ consisting of the $(q,t)$-characters of standard modules that it contains. Taking $$\B':= \{ b_{k_1}b_{k_2}\cdots b_{k_s}\mid b_{k_j}\in\AA^{(k_j)},\ F(b_{k_j})\in \B^{(k_j)},\
k_1>\cdots >k_s,\ s\in {\mathbb{N}}\},$$ we get a spanning set of $\AA$ such that $F(\B')$ is a basis of $\K_t$, consisting of the $(q,t)$-characters of all the standard modules of $\CC_{\mathbb{Z}}$. Hence $\B'$ is a basis of $\AA$ and $F$ is an isomorphism.
[Let $\g = \Sl_2$. By Theorem \[presentation\], $\K_t$ is presented by generators $y_m := \chi_{q,t}(L(Y_{2m}))$ indexed by $m\in{\mathbb{Z}}$, subject to $$\begin{array}{llll}
y_my_{m+1} &=& t^{-2}y_{m+1}y_m + 1-t^{-2},\\[2mm]
y_my_{p} &=& t^{2(-1)^{p-m}}\,y_{p}y_m, &\mbox{if $p>m+1$.}
\end{array}$$ ]{}
(a) It was shown by Frenkel and Reshetikhin [@FR Corollary 2] that the (classical) Grothendieck ring ${\mathcal{R}}$ of $\CC_{\mathbb{Z}}$ is the polynomial ring in the classes of all fundamental modules $L(Y_{i,p})\ ((i,p)\in \hI)$. More recently, a presentation of ${\mathcal{R}}$ in terms of Kirillov-Reshetikhin modules and $T$-systems was given in [@IIKNS Corollary 2.9].
Note that our presentation of $\K_t$ does *not* yield a new presentation of ${\mathcal{R}}$. Indeed, in order to obtain ${\mathcal{R}}$ from $\K_t$ by specializing $t$ at $1$, one needs to use the integral form $K_t$ defined in §\[main\_iso\] below, and the $x_{i,m}^Q$ are not generators of $K_t$ if $\g \not = \Sl_2$.
(b) For $m\in{\mathbb{Z}}$, let $\K^{(m,m+1)}$ denote the subalgebra of $\K_t$ generated by $y_{i,m}, y_{i,m+1}\ (i\in I)$. It follows from Theorem \[presentation\] that $\K^{(m,m+1)}$ is isomorphic to the $t$-deformed boson algebra $\B_t(\g)$ introduced by Kashiwara [@K §3.3].
Derived Hall algebras {#sect_Hall}
=====================
The Hall algebra $H(Q)$ {#Hall}
-----------------------
Let $F$ be a finite field, and let $u := |F|^{1/2} \in {\mathbb{R}}_{>0}$. Let ${\mathrm{mod}}(FQ)$ be the abelian category of representations of $Q$ over $F$. The twisted Hall algebra $H(Q)$, introduced by Ringel, is the ${\mathbb{C}}$-algebra with basis $\{z_X\}$ labelled by the isoclasses of objects in ${\mathrm{mod}}(FQ)$, with multiplication $$z_X z_Y = u^{\<Y,\, X\>}\sum_W g_{X,Y}^W z_W,$$ where $g_{X,Y}^W$ is the number of submodules $T$ of $W$ such that $T\simeq X$ and $W/T\simeq Y$. Ringel [@Ri; @Ri2; @Ri4] has shown that $H(Q)$ is isomorphic to the ${\mathbb{C}}$-algebra $U_u(\n)$ obtained from $U_v(\n)$ by specializing $v$ at $u$. In this isomorphism, the basis $\{z_X\}$ is mapped to a PBW-basis of $U_u(\n)$. In particular, if $S_i$ denotes the 1-dimensional simple supported on $i\in I$, $z_{S_i}$ is mapped to the Chevalley generator $e_i$.
The derived Hall algebra $DH(Q)$
--------------------------------
Let $D^b({\mathrm{mod}}(FQ))$ be the bounded derived category of ${\mathrm{mod}}(FQ)$. Toën [@T §7] has associated with this triangulated category an associative algebra $DH(Q)$ with the following presentation. The generators $z_X^{[m]}$ are labelled by all pairs $(X,m)$ where $X$ is an isoclass of ${\mathrm{mod}}(FQ)$ and $m\in{\mathbb{Z}}$. (The pair $(X,m)$ corresponds to the stalk complex with $X$ in degree $m$.) The relations are:
- for every $m\in{\mathbb{Z}}$, $$z_X^{[m]} z_Y^{[m]} = u^{\<Y,\, X\>}\sum_W g_{X,Y}^W z_W^{[m]};$$
- for every $m\in{\mathbb{Z}}$, $$z_X^{[m]} z_Y^{[m+1]} = u^{-\<Y,\, X\>}\sum_{W,T} u^{-\<W,T\>}\ga_{X,Y}^{T,W} z_T^{[m+1]}z_W^{[m]};$$
- for $p>m+1$, $$z_X^{[m]}z_Y^{[p]} = u^{(-1)^{p-m}(X,Y)} z_Y^{[p]} z_X^{[m]}.$$
Here, the Hall number $\ga_{X,Y}^{T,W}$ is defined by Toën as $$\ga_{X,Y}^{T,W} := \frac{|{\operatorname{Ex}}(W,Y,X,T)|}{|{\operatorname{Aut}}(X)||{\operatorname{Aut}}(Y)|},$$ where ${\operatorname{Ex}}(W,Y,X,T)$ is the finite subset of ${\operatorname{Hom}}(W,Y)\times {\operatorname{Hom}}(Y,X)\times{\operatorname{Hom}}(X,T)$ consisting of exact sequences $0\to W\to Y \to X \to T \to 0$. Note that, as in §\[Hall\], we have twisted the multiplication by inserting in the original Hall product $z_X^{[m]}z_Y^{[p]}$ of [@T] a factor $u^{(-1)^{p-m}\<Y,X\>}$, see [@S].
Consider the elements $z_{i,m} := z_{S_i}^{[m]}$ for $i\in I$ and $m\in{\mathbb{Z}}$. As in §\[generators\], we see that the $z_{i,m}$ generate $DH(Q)$. More precisely, we have:
\[presentationDH\] The algebra $DH(Q)$ is generated by the $z_{i,m}\ (i\in I, m\in{\mathbb{Z}})$ subject only to the following relations:
- for every $m\in{\mathbb{Z}}$, $$\begin{array}{lll}
&z_{i,m}\,z_{j,m} - z_{j,m}\,z_{i,m} = 0 &\mbox{if $(\a_i,\a_j) = 0$,}\\[2mm]
&z_{i,m}^2\,z_{j,m} - (u+u^{-1})z_{i,m}\,z_{j,m}\,z_{i,m} + z_{j,m}\,z_{i,m}^2= 0\qquad &\mbox{if $(\a_i,\a_j) = -1$;}
\end{array}$$
- for every $m\in{\mathbb{Z}}$ and every $i,j\in I$, $$z_{i,m}\, z_{j,\,m+1} = u^{-(\a_i,\a_j)} z_{j,\,m+1}\, z_{i,m} + \de_{ij}\frac{u^{-1}}{u^2-1};$$
- For every $p>m+1$ and every $i,j\in I$, $$z_{i,m}\,z_{j,p} = u^{(-1)^{p-m}(\a_i,\a_j)} \,z_{j,p}\,z_{i,m}.$$
The relations (H1) follow immediately from (D1) and Ringel’s theorem. The relations (H3) follow immediately from (D3). Let us deduce the relations (H2) from (D2).
If $i\not = j$, the only exact sequences $0\to W\to S_j \to S_i \to T \to 0$ are of the form $$0\to S_j\overset{\f}{\to} S_j \overset{0}{\to} S_i \overset{g}{\to} S_i \to 0$$ where $0$ means the zero map, and $f$ and $g$ are isomorphisms. Clearly there are $(|F|-1)^2$ such sequences, and since $|{\operatorname{Aut}}(S_i)|=|{\operatorname{Aut}}(S_j)|=|F|-1$, we get that $\ga_{S_i,S_j}^{S_j,S_i} = 1$. Hence $$z_{i,m}\, z_{j,\,m+1} = u^{-\<S_j,S_i\>} u^{-\<S_i,S_j\>} z_{j,\,m+1}\, z_{i,m}
= u^{-(\a_i,\a_j)} z_{j,\,m+1}\, z_{i,m}.$$
If $i=j$, we have two types of exact sequences $0\to W\to S_i \to S_i \to T \to 0$, namely $$0\to S_i\overset{\f}{\to} S_i \overset{0}{\to} S_i \overset{g}{\to} S_i \to 0,
\quad \mbox{and}\quad
0\to 0\overset{0}{\to} S_i \overset{h}{\to} S_i \overset{0}{\to} 0 \to 0,$$ where $f,g,h$ are isomorphisms. It follows that $$\ga_{S_i,S_i}^{S_i,S_i} = 1,
\qquad\mbox{and}\qquad
\ga_{S_i,S_i}^{0,0} = \frac{1}{|F|-1} = \frac{1}{u^2-1},$$ hence $$z_{i,m}\, z_{i,\,m+1} = u^{-(S_i,S_i)} z_{i,\,m+1}\, z_{i,m} + u^{-\<S_i,S_i\>}\frac{1}{u^2-1}.$$ This proves (H2). Finally, the proof that relations (H1), (H2), (H3) give a presentation of $DH(Q)$ is entirely similar to the proof of the analogous statement in Theorem \[presentation\] (the basis $\B^{(m)}$ is replaced by $\{z_X^{[m]} \mid X \mbox{ isoclass of }{\mathrm{mod}}(FQ) \}$), and we omit it.
The isomorphism between $\K_u$ and $DH(Q)$
------------------------------------------
Define the integral form $$K_t:= \bigoplus_{L} {\mathbb{C}}[t^{1/2},\,t^{-1/2}]\ \chi_{q,t}(L) \subset \K_t,$$ where the sum runs over all isoclasses $L$ of simple objects in $\CC_{\mathbb{Z}}$. By Theorem \[posmult\], this is a subring of $\K_t$. Set $$\K_u := {\mathbb{C}}\otimes_{{\mathbb{C}}[t^{1/2},\,t^{-1/2}]} K_t,$$ where ${\mathbb{C}}$ is regarded as a ${\mathbb{C}}[t^{1/2},\,t^{-1/2}]$-module via the specialization $t^{1/2}\mapsto u^{1/2}$.
For $\b\in \hDe = \De_+\times{\mathbb{Z}}$, we denote by $z_\b^{[m]}$ the basis element $z_X^{[m]}$ of $DH(Q)$ with $X\in{\mathrm{mod}}(FQ)$ indecomposable of dimension vector $\b$.
The following is a slightly more precise formulation of Theorem \[Th\_DHall\].
\[main\_iso\]
There is a ${\mathbb{C}}$-algebra isomorphism $\iota \colon \K_u \overset{\sim}{\to} DH(Q)$ such that:
- the class of the fundamental $U_q(L\g)$-module $L(Y_{i,p})$ of $\CC_{\mathbb{Z}}$ is mapped by $\iota$ to a scalar multiple of $z_\b^{[m]}$, where $(\b,m) = \varphi(i,p)$.
- the basis of classes of standard $U_q(L\g)$-modules of $\CC_{\mathbb{Z}}$ is mapped by $\iota$ to a rescaling of the natural basis of $DH(Q)$ labelled by all isoclasses of objects of $D^b({\mathrm{mod}}(FQ))$.
We first assume, as in the proof of Theorem \[presentation\], that $Q$ is a sink-source orientation of the Dynkin diagram. We can rescale the generators $x_{i,m}^Q$ of $\K_t$ by setting $$\tx_{i,m}^Q := \frac{1}{u^{1/2}(u-u^{-1})}\ x_{i,m}^Q,\qquad (i\in I,\ m\in{\mathbb{Z}}).$$ Clearly the new generators $\tx_{i,m}^Q$ still satisfy the homogeneous relations (R1) and (R3) of Theorem \[presentation\], and the relations (R2) become $$\tx_{i,m}^Q\, \tx_{j,\,m+1}^Q = t^{-(\a_i,\a_j)} \tx_{j,\,m+1}^Q\, \tx_{i,m}^Q + \de_{ij}\frac{1-t^{-2}}{u(u-u^{-1})^2}.$$ Let $\bx_{i,m}^Q = 1\otimes \tx_{i,m}^Q \in \K_u$. By Theorem \[presentation\], Proposition \[presentationDH\], the assignment $\tx_{i,m}^Q\mapsto z_{i,m}$ extends to an algebra isomorphism $\iota$. Indeed, in the relations (R2) we have $$\frac{1-u^{-2}}{u(u-u^{-1})^2} = \frac{u^{-1}}{u^2-1}$$ so the generators $\bx_{i,m}^Q$ of $\K_u$ and $z_{i,m}$ of $DH(Q)$ give rise to identical presentations.
Since the PBW-basis of $U_v(n)$ is orthogonal with respect to the bilinear form of §\[defAqn\], it only differs from the dual PBW-basis $\bE^*$ by scalar multiples. Hence by Ringel’s theorem, it follows from Theorem \[mainth\] that the classes of fundamental modules in $\CC_Q$, which correspond under $\Phi$ to the elements $E^*(\b)\ (\b\in\De_+)$ of $U_v(\n)$, are mapped by $\iota$ to scalar multiples of the $z_{\b}^{[0]}$. So, if $\varphi(i,p)=(\b,0)$ we have $\iota([L(Y_{i,p})]_u) = \la_{i,p} z_\b^{[0]}$ for some $\la_{i,p}\in{\mathbb{C}}$. Therefore, using on one side the automorphism of $\K_u$ given by $[L(Y_{i,p})]_u \mapsto [L(Y_{i,p-2})]_u$, and on the other side the corresponding automorphism of $DH(Q)$ induced by the Auslander-Reiten translation $\tau$ of $D^b(\mod(FQ))$, we get (a).
Since the classes of standard modules are the ordered products of the $[L(Y_{i,p})]_u$ (up to powers of $u$), and the basis elements of $DH(Q)$ are the ordered products of the $z_\b^{[m]}$ (up to powers of $u$), we get (b).
Therefore we have proved Theorem \[main\_iso\] in the case of a sink-source orientation. But the ${\mathbb{C}}$-algebras $\K_u$ and $DH(Q)$ are both independent of $Q$. For $\K_u$ this is clear. On the other hand if $Q'$ is another orientation of the Dynkin diagram, then $D^b({\mathrm{mod}}(FQ))$ and $D^b({\mathrm{mod}}(FQ'))$ are equivalent triangulated categories, so $DH(Q)$ and $DH(Q')$ are isomorphic. Thus $\K_u$ is isomorphic to $DH(Q)$ for an arbitrary orientation. More precisely, recall that the map $\varphi=\varphi_Q\colon \hI \to \hDe$ depends on the choice of $Q$. There is a triangle equivalence $F_{QQ'}\colon D^b({\mathrm{mod}}(FQ)) \to D^b({\mathrm{mod}}(FQ'))$ such that the induced isomorphism $f_{QQ'}\colon DH(Q) \to DH(Q')$ satisfies $$f_{QQ'}\left(z_{\b}^{[m]}\right) = z_{\b'}^{[m']} \quad \mbox{where}
\quad (\b',m') = \varphi_{Q'}\varphi_Q^{-1}(\b,m).$$ Therefore (a) and (b) hold for an arbitrary orientation.
In the proof of Theorem \[presentation\], we have shown that if $Q$ is a sink-source orientation, the generators $x_{i,m}^Q$ of $\K_t$ satisfy the relations (R1), (R2), (R3). We can see now that this holds for any orientation $Q$.
The generators $x_{i,m}^Q$ of $\K_t$ satisfy the same relations for every orientation $Q$ of the Dynkin diagram, namely the relations (R1), (R2), (R3) of Theorem \[presentation\].
Let $Q$ be any orientation, by Theorem \[main\_iso\], the elements $\bx_{i,m}^{Q}$ of $\K_u$ are mapped by $\iota$ to scalar multiples of the generators $z_{i,m}$ of $DH(Q)$. Now the relations (H1), (H2), (H3) satisfied by the $z_{i,m}$ are independent of $Q$. Moreover, they are all homogeneous except for (H2) with $i=j$. Since scalar multiplication does not affect homogeneous relations, the elements $1\otimes x_{i,m}^Q$ of $\K_u$ satisfy the relations (R1), (R2) $(i \not = j)$, (R3) with $t$ replaced by $u=|F|^{1/2}$. Since this is true for every finite field $F$, it follows that the elements $x_{i,m}^Q$ of $\K_t$ satisfy the relations (R1), (R2) $(i\not = j$), (R3) where $t$ is an indeterminate.
Finally, the relations (R2) $(i=j)$ follow from Lemma \[lemme\_commut\] (c) with $i=\nu(j)$.
\[Remark\_untwistedDH\] [Using Remark \[Remark\_differ2\], one can modify the presentation of $\K_t$ to obtain a presentation of the deformed Grothendieck ring ${\mathcal{R}}_t$ of [@N; @VV]. This presentation shows that the specialization of ${\mathcal{R}}_{t}$ at $t=u^{-1}$ is isomorphic to the *non-twisted* derived Hall algebra of $D^b({\mathrm{mod}}(FQ))$ with the opposite product. ]{}
Quiver varieties {#sect7}
================
In this section we show that the variety $E_\bd$ of representations of $Q$ with dimension vector $\bd$ can be regarded as a Nakajima graded quiver variety ${{\mathcal M}}^\bullet_0(W^\bd)$ for an appropriate $\hI$-graded vector space $W^\bd$. Moreover the stratification of $E_\bd$ by $G_\bd$-orbits coincides with Nakajima’s stratification of ${{\mathcal M}}^\bullet_0(W^\bd)$. It follows that the set of perverse sheaves used by Lusztig to define the (dual) canonical basis of $U_v(\n)$ can be identified with a subset of the set of perverse sheaves used by Nakajima for describing the classes $[L]_t$ of simple $U_q(L\g)$-modules. This gives a geometric way of understanding Theorem \[mainth\] (b).
The quiver representation space $E_{\bd}$ {#quiverLusztig}
-----------------------------------------
Let $
\bd = (d_i)_{i\in I} \in {\mathbb{N}}^I
$ denote a dimension vector for $Q$. We will identify $\bd$ with the element $\sum_{i\in I} d_i\a_i$ of the root lattice of $\g$. The variety $E_\bd$ of representations of $Q$ of dimension $\bd$ is by definition $$E_\bd := \bigoplus_{i\to j} {\operatorname{Hom}}_{\mathbb{C}}({\mathbb{C}}^{d_i},\,{\mathbb{C}}^{d_j}),$$ the sum being over all arrows $i\to j$ of $Q$. This is just a ${\mathbb{C}}$-vector space of dimension $\sum_{i\to j} d_id_j$, but the interesting geometry comes from the following stratification. Consider the algebraic group $$G_{\bd} := \prod_{i\in I} GL(d_i,{\mathbb{C}}).$$ It acts on $E_\bd$ by base change. There are finitely many orbits in one-to-one correspondence with the isomorphism classes of representations of $Q$ of dimension $\bd$. Thus, using Gabriel’s theorem, these orbits have a natural labelling by the set $$I_\bd := \left\{\aa = (a_k) \in {\mathbb{N}}^r \mid \sum_{k=1}^r a_k \b_k = \bd \right\},$$ where the positive roots $\b_k$ are enumerated as in (\[eq\_beta\]). Let $\O_\aa$ denote the orbit labelled by the element $\aa$ of $I_\bd$. Let $IC(\overline{\O_\aa})$ be the intersection cohomology complex of $\overline{\O_\aa}$, extended by zero on the complement of $\overline{\O_\aa}$. Let $\H^i(IC(\overline{\O_\aa}))$ be its ith cohomology sheaf, and $\H^i(IC(\overline{\O_\aa}))_{\cc}$ the stalk of this sheaf at a point of $\O_\cc$.
Recall from §\[bases\] the dual PBW basis $\bE^*$ and the dual canonical basis $\bB^*$ of $A_v(\n)$. Write $$E^*(\cc) = \sum_{\aa\in I_\bd} \kappa_{\aa,\cc}(v)\, B^*(\aa).$$ Lusztig has shown:
[[@Lu §9, §10]]{}\[LusztigTh\] The coefficients $\kappa_{\aa,\cc}(v)$ are given by $$\label{Lusztig_form}
\kappa_{\aa,\cc}(v) = v^{\dim\O_\cc - \dim\O_\aa} \sum_{i\ge 0} v^{i} \dim\H^i(IC(\overline{\O_\aa}))_{\cc}.$$
Nakajima’s variety ${{\mathcal M}}^\bullet_0(W)$ {#quiverNakajima}
------------------------------------------------
Let $$W = \bigoplus_{(i,p)\in \hI} W_i(p)$$ be a finite-dimensional $\hI$-graded ${\mathbb{C}}$-vector space. In his geometric construction of representations of $U_q(L\g)$, Nakajima [@N] has associated with $W$ an affine variety ${{\mathcal M}}^\bullet_0(W)$ whose definition we shall now recall.
Let $\hJ:=\{(i,p)\in I\times {\mathbb{Z}}\mid (i,p-1)\in \hI\}$, and let $$V = \bigoplus_{(i,s)\in \hJ} V_i(s)$$ be a finite-dimensional $\hJ$-graded ${\mathbb{C}}$-vector space. Define $$\begin{aligned}
&L^\bullet(V,W)= \bigoplus_{(i,s)\in \hJ} {\operatorname{Hom}}(V_i(s),W_i(s-1)),
\\
&L^\bullet(W,V)= \bigoplus_{(i,p)\in \hI} {\operatorname{Hom}}(W_i(p),V_i(p-1)),
\\
&E^\bullet(V)= \bigoplus_{(i,s)\in \hJ;\ j\sim i} {\operatorname{Hom}}(V_i(s),V_j(s-1)). \end{aligned}$$ Put $M^\bullet(V,W) = E^\bullet(V) \oplus L^\bullet(W,V) \oplus L^\bullet(V,W)$. An element of $M^\bullet(V,W)$ is written $(B,\a,\b)$, and its components are denoted by: $$\begin{aligned}
&B_{ij}(s)\in {\operatorname{Hom}}(V_i(s),V_j(s-1)),
\\
&\a_i(p)\ \in {\operatorname{Hom}}(W_i(p),V_i(p-1)),
\\
&\b_i(s)\ \in {\operatorname{Hom}}(V_i(s),W_i(s-1)).\end{aligned}$$ We denote by $\L^\bullet(V,W)$ the subvariety of the affine space $M^\bullet(V,W)$ defined by the equations $$\label{ADHM}
\a_i(s-1)\b_i(s) + \sum_{j\sim i} \epsilon(i,j)\, B_{ji}(s-1)B_{ij}(s) = 0,
\qquad ((i,s)\in\hJ),$$ where $\epsilon(i,j)=1$ ($\epsilon(i,j)=-1$) if $i\to j$ is an arrow of $Q$ ($i\to j$ is not an arrow of $Q$). The algebraic group $$G_V := \prod_{(i,s)\in\hJ}GL(V_i(s))$$ acts on $M^\bullet(V,W)$ by base change in $V$: $$g\cdot (B,\a,\b) = \left((g_j(s-1)B_{ij}(s) g_i(s)^{-1}),\ (g_i(p-1)\a_i(p)),\ (\b_i(s)g_i(s)^{-1})\right).$$ Note that there is no action on $W$. This action of $G_V$ preserves the subvariety $\L^\bullet(V,W)$. One defines the affine quotient $${{\mathcal M}}^\bullet_0(V,W) := \L^{\bullet}(V,W) \sslash G_V.$$ By definition, the coordinate ring of ${{\mathcal M}}^\bullet_0(V,W)$ is the ring of $G_V$-invariant functions on $\L^{\bullet}(V,W)$, and ${{\mathcal M}}^\bullet_0(V,W)$ parametrizes the closed $G_V$-orbits. If $V_i(s) \subseteq V'_i(s)$ for every $(i,s)\in \hJ$, then we have a natural closed embedding ${{\mathcal M}}^\bullet_0(V,W) \subset {{\mathcal M}}^\bullet_0(V',W)$. Finally, one defines $${{\mathcal M}}^\bullet_0(W) := \bigcup_V {{\mathcal M}}^\bullet_0(V,W).$$ This is an affine variety, acted upon by the algebraic group $$G_W := \prod_{(i,p)\in\hI}GL(W_i(p)).$$ Let ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W)$ be the open subset of ${{\mathcal M}}^\bullet_0(V,W)$ parametrizing the closed [*free*]{} $G_V$-orbits. For a given $W$, we have ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W)\not = \emptyset$ only for a finite number of $V$’s. Nakajima has shown that this gives a stratification of ${{\mathcal M}}^\bullet_0(W)$: $${{\mathcal M}}^\bullet_0(W) = \bigsqcup_V {{\mathcal M}}^{\bullet\,{\rm reg}}_0(V,W).$$ A necessary condition for ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W)$ to be nonempty is that $$\dim W_i(p) - \dim V_i(p+1) - \dim V_i(p-1) + \sum_{j\sim i} \dim V_j(p) \ge 0$$ for every $(i,p)\in\hI$. In this case we say that $(V,W)$ is a *dominant pair*. Equivalently, by (\[defA\]), the pair $(V,W)$ is dominant if and only if the monomial $Y^W A^V \in \Y$ is dominant, where we use the shorthand notation $$Y^W := \prod_{(i,p)\in\hI} Y_{i,p}^{\dim W_i(p)},
\qquad
A^V := \prod_{(i,s)\in\hJ} A_{i,s}^{-\dim V_i(s)}.$$ Note that this stratification of ${{\mathcal M}}^\bullet_0(W)$ is $G_W$-invariant. Hence each stratum is a union of $G_W$-orbits.
Given a dominant pair $(V,W)$ such that ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W) \not = \emptyset$, we denote by $IC_W(V)$ the intersection cohomology complex of the closure of the stratum ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W)$. Let $\H^i(IC_W(V))$ be its $i$th cohomology sheaf, and $\H^i(IC_W(V))_{V'}$ be the stalk of this sheaf at a point of ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V',W)$.
For a dominant monomial $m\in\Y$, recall from §\[sectqtstandard\] and §\[sectqtsimple\] the $(q,t)$-characters $\chi_{q,t}(M(m))$ and $\chi_{q,t}(L(m))$ of the standard and of the simple $U_q(L\g)$-modules labelled by $m$. Write $$\chi_{q,t}(M(m')) = \sum_{m'} \zeta_{m,m'}(t)\, \chi_{q,t}(L(m)).$$ Nakajima has shown:
[[@N §8]]{}\[NakajimaTh\] The coefficients $\zeta_{m,m'}(t)$ are given by $$\label{Nakajima_form}
\zeta_{m,m'}(t) = t^{\dim{{\mathcal M}}^{\bullet\, {\rm reg}}_0(V',W) - \dim{{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W)}
\sum_{i\ge 0} t^{i} \dim\H^i(IC_W(V))_{V'},$$ for any pair of strata ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W)$ and ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V',W)$ such that $m = Y^WA^V$ and $m'=Y^WA^{V'}$.
(a) In order to make the comparison between Theorem \[LusztigTh\] and Theorem \[NakajimaTh\] easier, we stated Nakajima’s formula (\[Nakajima\_form\]) in a different way from the original one. In [@N], Nakajima writes $$t^{\dim{{\mathcal M}}^{\bullet\, {\rm reg}}_0(V',W)}
\sum_{i\ge 0} t^{-i} \dim\H^i(i_{x_{V'}}^{!}IC_W(V))$$ for the right-hand side of (\[Nakajima\_form\]), but in his degree convention the trivial local system on the open stratum $S = {{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W)$ appears in the intersection cohomology complex of $\overline{S}$ in degree $\dim S$, while in Lusztig’s convention it appears in degree 0. Here we follow Lusztig’s convention. Moreover Nakajima uses the costalk $i_x^{!}$ at a point $x$ instead of the stalk $i_x^*$, which explains the change of $t^i$ into $t^{-i}$.
(b) A dominant monomial $m$ can be written in several ways as $m = Y^WA^V$. The fact that the right-hand side of (\[Nakajima\_form\]) depends only on the monomials $m$ and $m'$, and not on the particular choices of spaces $W, V, V'$, follows from a transversal slice argument [@NJAMS §3].
An isomorphism {#isom}
--------------
Let $\bd=(d_i)$ be a dimension vector, as in §\[quiverLusztig\]. Recall the bijection $\varphi\colon \hI \to \hDe$ of §\[subsect\_quivers\]. We define an $\hI$-graded space $W^{\bd}$ by taking $$W^{\bd}_j(p) := {\mathbb{C}}^{d_i} \quad \mbox{if} \quad\varphi(j,p) = (\a_i,0),$$ and $W^{\bd}_j(p):= 0$ for all others $(j,p)\in\hI$. Clearly, the group $G_{W^{\bd}}$ is isomorphic to $G_\bd$ and we may identify $G_{W^{\bd}} \equiv G_\bd$.
\[propgeo1\] There is a $G_\bd$-equivariant closed immersion of affine varieties $$\Psi\colon {{\mathcal M}}^\bullet_0(W^{\bd}){\longrightarrow} E_\bd.$$
The proof of Proposition \[propgeo1\] will follow from the next two lemmas.
\[Lemdim1\] Let $i\not = i'\in I$ and set $(j,p)=\varphi^{-1}(\a_i,0)$, $(j',p')=\varphi^{-1}(\a_{i'},0)$. Assume that $p'\le p$, and write $\varphi(j',\xi_j+p'-p+2) = (\b,0)$. Then, the coefficient of $\a_j$ in the expansion of the root $\b$ on the basis of simple roots is equal to $1$ if there is an arrow $i\to i'$ in $Q$, and to $0$ otherwise.
By definition of $(j,p)$, $(j',p')$, and of $\varphi$ (see §\[subsect\_quivers\]), we have $$\a_i = \tau^{(\xi_j-p)/2}(\ga_j),\qquad \a_{i'} = \tau^{(\xi_{j'}-p')/2}(\ga_{j'}).$$ It follows that $$\b = \tau^{-1+(\xi_{j'}-\xi_j-p'+p)/2}(\ga_{j'}) = \tau^{-1+(p-\xi_{j})/2}(\a_{i'}).$$ Recall the Ringel bilinear form $\<\cdot,\cdot\>$. It may be characterized by $$\<\a_i,\ga_j\> = \de_{ij},\qquad (i,j\in I).$$ Hence, the coefficient of $\a_j$ in $\b$ is equal to: $$\<\b,\ga_j\> = \<\tau^{-1+(p-\xi_{j})/2}(\a_{i'}),\ga_j\>
=\<\tau^{-1}(\a_{i'}),\tau^{(\xi_j-p)/2}(\ga_j)\> =
\<\tau^{-1}(\a_{i'}),\a_i\>=-\<\a_i,\a_{i'}\>.$$ Now $$-\<\a_i,\a_{i'}\> = -\dim({\operatorname{Hom}}(S_i,S_{i'})) + \dim({\operatorname{Ext}}^1(S_i,S_{i'})) = \dim({\operatorname{Ext}}^1(S_i,S_{i'})),$$ and this is equal to 1 if there is an arrow from $i$ to $i'$ in $Q$, and to $0$ otherwise.
We now introduce an algebra $\tL_Q$ defined by a quiver $\tG_Q$ with relations. The vertices of $\tG_Q$ are of two types:
- $w_j(p)$ for every $(j,p) = \varphi^{-1}(\a_i,0)\ (i\in I)$;
- $v_j(p-1)$ for every pair $(j,p)\in \hI_Q$ such that $(j,p-2)\in \hI_Q$.
The arrows of $\tG_Q$ are of three types:
- $a_j(p)\colon w_j(p) \to v_j(p-1)$;
- $b_j(p)\colon v_j(p) \to w_j(p-1)$;
- $\B_{ij}(p)\colon v_i(p)\to v_j(p-1)$ if $j\sim i$.
The relations are: $$a_i(p-1)b_i(p) = \sum_{j\sim i} \epsilon(i,j) \B_{ji}(p-1)\B_{ij}(p).$$ Obviously, as suggested by the notation, the definition of $\tL_Q$ is so that the affine variety $\L^\bullet(V,W^{\bd})$ is nothing but the representation variety of $\tL_Q$ consisting of representations for which the spaces $W_j^\bd(p) = {\mathbb{C}}^{d_i}$ are attached to the vertices $w_j(p)$, and the spaces $V_j(p-1)$ are attached to the vertices $v_j(p-1)$ (we assume that $V_j(p-1)=0$ if $v_j(p-1)$ is not a vertex of $\tG_Q$).
For $i\in I$, we denote by $\veps_i$ the idempotent of $\tL_Q$ associated with the vertex $w_j(p)$ such that $(j,p) = \varphi^{-1}(\a_i,0)$. We endow $I$ with a total ordering such that $i>i'$ if $p>p'$, where as above $\varphi^{-1}(\a_{i'},0) = (j',p')$. It is well-known that if there is an arrow $i\to i'$ in $Q$ then $i>i'$.
\[Lemdiminf1\] For $i\not =i'\in I$, we have $$\dim\left(\veps_{i'} \tL_Q \veps_i\right) =
\left\{
\begin{matrix}
1 & \mbox{if there is a path from $i$ to $i'$ in $Q$,}\\[2mm]
0 &\mbox{otherwise.}
\end{matrix}
\right.$$
If $i<i'$ then $p\le p'$ and clearly $\veps_{i'} \tL_Q \veps_i = 0$. On the other hand if $i<i'$ there can be no path from $i$ to $i'$ in $Q$. Thus the lemma is clear in this case, and we may assume from now on that $i>i'$.
Let $x\in \veps_{i'} \tL_Q \veps_i$, and let $i''\in I$. Then $x\in \veps_{i'} \tL_Q \veps_{i''} \tL_Q \veps_i$ if and only if $x$ belongs to the two-sided ideal of $\tL_Q$ generated by $a_{j''}(p-1)b_{j''}(p)$ where $\varphi(j'',p-1)=(\a_{i''},0)$. This is because $b_{j''}(p)$ is the only arrow entering in $w_{j''}(p-1)$, and $a_{j''}(p-1)$ is the only arrow exiting from $w_{j''}(p-1)$. Note that $\veps_{i'} \tL_Q \veps_{i''} \tL_Q \veps_i \not = 0$ implies that $i>i''>i'$.
Let $\I$ be the the two-sided ideal of $\tL_Q$ generated by all the degree two paths: $$a_{j}(p-1)b_{j}(p), \qquad ((j,p-1) = \varphi^{-1}(\a_i,0),\ (i\in I)).$$ Then, the algebra $\tL_Q/\I$ is defined by the same relations as the graded preprojective algebra $\hL$ of [@L §2.8] (but we have the additional vertices $w_j(p)$ and only a finite set of vertices $v_j(p)$). It follows that $\dim\left(\left(\veps_{i'} \tL_Q \veps_i\right) / \left(\veps_{i'} \I \veps_i\right)\right)$ is equal to the $v_{j'}(p'+1)$-component of the dimension vector of the indecomposable projective $\hL$-module with top $v_j(p-1)$. Now it is well-known that this dimension vector can be read off from the Auslander-Reiten quiver of $Q$. Namely, using our notation, the $v_j(p-1)$-component of the dimension vector of the projective with top $v_i(\xi_i-1)$ is equal to the coefficient of $\a_i$ in the root $\b$ such that $\varphi(j,p) = (\b,0)$. The dimension vectors of the remaining indecomposable projectives are obtained from these particular ones by translation. It follows that we can reformulate Lemma \[Lemdim1\] as follows: $$\label{dimproj}
\dim\left(\left(\veps_{i'} \tL_Q \veps_i\right) / \left(\veps_{i'} \I \veps_i\right)\right) =
\left\{
\begin{matrix}
1 & \mbox{if there is an arrow from $i$ to $i'$ in $Q$,}\\[2mm]
0 &\mbox{otherwise.}
\end{matrix}
\right.$$ In particular if $i'$ is the successor of $i$ in the descending total order defined above, then $\veps_{i'} \I \veps_i = 0$, and we have $$\label{successor}
\dim\left(\veps_{i'} \tL_Q \veps_i\right) =
\left\{
\begin{matrix}
1 & \mbox{if there is an arrow from $i$ to $i'$ in $Q$,}\\[2mm]
0 &\mbox{otherwise.}
\end{matrix}
\right.$$
Assume now that $i>i'$ are such that there is no path from $i$ to $i'$ in $Q$. Then in particular there is no arrow $i\to i'$, so by (\[dimproj\]) we have $$\veps_{i'}\tL_Q\veps_i = \veps_{i'} \I \veps_i
= \sum_{i'<j<i} \left(\veps_{i'}\tL_Q\veps_j\right) \left(\veps_{j}\tL_Q\veps_i\right).$$ For each summand, we have either no path from $i$ to $j$ or no path from $j$ to $i'$. So we can iterate the splitting until we obtain an expression of the form $$\veps_{i'}\tL_Q\veps_i =
\sum_{i'<i_1<\cdots<i_k<i} \left(\veps_{i'}\tL_Q\veps_{i_k}\right)\cdots \left(\veps_{i_1}\tL_Q\veps_i\right),$$ where in the right-hand side each factor $\veps_{k}\tL_Q\veps_{j}$ is such that either we have a path from $j$ to $k$ or $k$ is the successor of $j$ and there is no arrow from $j$ to $k$. Moreover, since we have no path from $i$ to $i'$ each summand contains at least one factor of the second type, which is equal to $0$ by (\[successor\]). Hence we have shown that $\veps_{i'}\tL_Q\veps_i = 0$.
Assume now that there is an arrow $i\to i'$ in $Q$. Then we have as above $$\veps_{i'} \I \veps_i
= \sum_{i'<j<i} \left(\veps_{i'}\tL_Q\veps_j\right) \left(\veps_{j}\tL_Q\veps_i\right),$$ where for each $j$ we have either no path from $i$ to $j$ or no path from $j$ to $i'$ (because the Dynkin diagram is a tree). Thus it follows from above that $\veps_{i'} \I \veps_i = 0$, so by (\[dimproj\]) we get $ \dim\left(\veps_{i'} \tL_Q \veps_i\right) = 1$.
Finally, if there is a path $i\to i_1\to \cdots \to i_k \to i'$ in $Q$, with $k\ge 1$, then there is no arrow from $i$ to $i'$, and by (\[dimproj\]) we have $\veps_{i'}\tL_Q\veps_i = \veps_{i'} \I \veps_i$. Moreover, this path is unique, and arguing as above we can write $$\veps_{i'}\tL_Q\veps_i =
\left(\veps_{i'}\tL_Q\veps_{i_k}\right)\cdots \left(\veps_{i_1}\tL_Q\veps_i\right),$$ where each factor has dimension $1$, so again $ \dim\left(\veps_{i'} \tL_Q \veps_i\right) = 1$.
[*Proof of Proposition \[propgeo1\] — *]{} Let $V$ be a $\hJ$-graded space, and pick $(B,\a,\b)\in \L^\bullet(V,W^\bd)$. As explained above, we can regard $(B,\a,\b)$ as a representation of $\tL_Q$. Choose two vertices $i$ and $i'$ of $Q$, and set as before $$(j, p) = \varphi^{-1}(\a_i,0),
\qquad
(j', p') = \varphi^{-1}(\a_{i'},0).$$ By Lemma \[Lemdiminf1\] we have $\dim(\veps_{i'} \tL_Q \veps_i) \le 1$. Let $\theta_{ii'}$ be a generator of $\veps_{i'} \tL_Q \veps_i$. By the proof of Lemma \[Lemdiminf1\], we can normalize the $\theta_{ii'}$ so that they verify $\theta_{i'i''}\theta_{ii'} = \theta_{ii''}$ for every $i,i',i''\in I$. More precisely, if there is a path $i\to i_1\to \cdots \to i_k\to i'$ in $Q$, then $\theta_{i,i'} = \theta_{i_ki'}\cdots \theta_{ii_1}$, and if there is no path from $i$ to $i'$ then $\theta_{ii'}=0$. Evaluating $\theta_{ii'}$ in the representation $(B,\a,\b)$ we obtain a linear map $\psi_{ii'}\colon W^{\bd}_{j}(p) \to W^{\bd}_{j'}(p')$. The collection of maps $(\psi_{ii'})$ for all arrows $i\to i'$ of $Q$ can be regarded as a representation $\psi$ of $Q$ of dimension vector $\bd$. It follows easily from the definition of the $G_V$-action that $\psi$ depends only on the $G_V$-orbit of $(B,\a,\b)$. Hence the assignment $(B,\a,\b) \mapsto \psi$ induces a morphism of varieties $\Psi_V\colon {{\mathcal M}}^{\bullet}_0(V,W^\bd) \to E_\bd$. Moreover, it follows from the known description of the generators of the coordinate ring of ${{\mathcal M}}^{\bullet}_0(V,W^\bd)$ (see [@Ncl §3.1]) that this coordinate ring is generated by the matrix coefficients of the linear maps $\psi_{ii'}$ for all pairs $(i, i')$. Hence $\Psi_V$ induces a surjective morphism from ${\mathbb{C}}[E_\bd]$ to ${\mathbb{C}}[{{\mathcal M}}^{\bullet}_0(V,W^\bd)]$, thus $\Psi_V$ is a closed immersion. Since for $V$ large enough we have ${{\mathcal M}}^{\bullet}_0(V,W^\bd) = {{\mathcal M}}^{\bullet}_0(W^\bd)$, we obtain a closed immersion $\Psi\colon{{\mathcal M}}^{\bullet}_0(W^\bd) \to E_\bd$. By construction, $\Psi$ commutes with the actions of $G_\bd$ on both varieties.
$$\[email protected]{
&\qquad\qquad\qquad\qquad\qquad&\qquad&&
\\
&&& W_3(5)\ar[d]^{\a_3(5)} &
\\
&&& \ar[lld]^{B_{31}(4)} \ar[ld]^{B_{32}(4)}V_3(4)\ar[rd]^{B_{34}(4)} &
\\
&V_1(3) \ar[rrd]^{B_{13}(3)}& V_2(3)\ar[rd]^{B_{23}(3)}&&\ar[ld]^{B_{43}(3)} V_4(3)
\\
&&&\ar[lld]^{B_{31}(2)}\ar[ld]^{B_{32}(2)}V_3(2)\ar[rd]^{B_{34}(2)}&
\\
&V_1(1)\ar[d]^{\b_1(1)} &V_2(1)\ar[d]^{\b_2(1)}&&\ar[d]^{\b_4(1)} V_4(1)
\\
&W_1(0) &W_2(0)&& W_4(0)
}$$
\[ExaPropGeo1\] [Take $\g$ of type $D_4$. We label the Dynkin diagram so that the central node is numbered $3$, and we choose $\xi_1=\xi_2=\xi_4=4$ and $\xi_3=5$. Thus $Q$ has a sink-source orientation with source $3$ and sinks $1,2,4$. Given a dimension vector $\bd = (d_1,d_2,d_3,d_4)$, the corresponding graded space $W^\bd$ is given by $$W_1^\bd(0)={\mathbb{C}}^{d_1},\quad
W_2^\bd(0)={\mathbb{C}}^{d_2},\quad
W_3^\bd(5)={\mathbb{C}}^{d_3},\quad
W_4^\bd(0)={\mathbb{C}}^{d_4},$$ and the other $W_i(p)$’s are zero (see the Auslander-Reiten quiver of $Q$ in Figure \[FigD\]). An element $(B,\a,\b)$ of $\L(V,W^\bd)$ is represented in Figure \[FigD0\]. The defining equations (\[ADHM\]) read $$\begin{aligned}
B_{13}(3)B_{31}(4) + B_{23}(3)B_{32}(4) + B_{43}(3)B_{34}(4)&=&0, \\[2mm]
B_{31}(2)B_{13}(3)\ =\ B_{32}(2)B_{23}(3)\ =\ B_{34}(2)B_{43}(3)
&=&0.\end{aligned}$$ Thus, $$\b_1(1)B_{31}(2)B_{43}(3)B_{34}(4)\a_3(5) =
- \b_1(1)B_{31}(2)B_{23}(3)B_{32}(4)\a_3(5)$$ and $$\b_1(1)B_{31}(2)B_{13}(3)B_{31}(4)\a_3(5)=0.$$ Hence we can take $$\psi_{31}:= \b_1(1)B_{31}(2)B_{43}(3)B_{34}(4)\a_3(5),$$ and similarly $$\psi_{32}:= \b_2(1)B_{32}(2)B_{13}(3)B_{31}(4)\a_3(5),
\qquad
\psi_{34}:= \b_4(1)B_{34}(2)B_{23}(3)B_{32}(4)\a_3(5).$$ We get a representation $\psi := (\psi_{31},\,\psi_{32},\,\psi_{34})$ of $Q$ on the space $W^\bd$. ]{}
\[propgeo2\] There is a bijection between the set of (nonempty) strata ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W^\bd)$ and the set $I_\bd$ of $G_\bd$-orbits.
Let us first consider a stratum ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W^\bd)$. By §\[quiverNakajima\], the pair $(V,W^\bd)$ is a dominant pair. This means that we have nonnegative integers $a_j\ (1\le j\le r)$ such that $$\label{eqVa}
Y^{W^\bd}A^V = \prod_{j=1}^r Y_{i_j, p_j}^{a_j}.$$ Here for $1\le j \le r$, we have put $(i_j,p_j) = \varphi^{-1}(\b_j,0)$. Indeed, by definition, every dominant commutative monomial of the form $Y^{W^\bd}A^V$ belongs to $\Y_{t,\,Q}$. Moreover we have a natural grading of $\Y_{t,\,Q}$ by the root lattice of $\g$ given by $$\deg(Y_{i_j,p_j}) = \b_j,\qquad (1\le j \le r).$$ It is easy to see that for every $A_{i,s}\in \Y_{t,\,Q}$ we have $\deg A_{i,s} = 0$. Therefore $$\sum_{k=1}^r a_j\b_j = \deg\left(Y^{W^\bd}\right) = \bd.$$ Hence, to every stratum ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W^\bd)$ corresponds an element $\aa$ of $I_\bd$ given by (\[eqVa\]).
Conversely, if $\aa\in I_\bd$, we need to show that $m_\aa:=\prod_{j=1}^r Y_{i_j, p_j}^{a_j}$ can be written in the form (\[eqVa\]) for some nonempty stratum ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W^\bd)$. By [@NJAMS Th. 14.3.2], this is equivalent to the fact that $m_\aa$ appears in the $q$-character of the standard module $M(Y^{W^\bd})$. For $i\in I$ write as in §\[generators\] $(k_i,p_i)=\varphi^{-1}(\a_i,0)$. Then we have by definition of $W^\bd$ $$Y^{W^\bd} = \prod_{i\in I} Y_{k_i,p_i}^{d_i}.$$ By §\[generators\] we know that the $(q,t)$-characters of the fundamental modules $L(Y_{k_i,p_i})\ (i\in I)$ generate $\K_{t,\,Q}$. Hence the simple module $L(m_\aa)$, which is an object of $\CC_Q$, is a composition factor of a standard module of the form $M(\prod_{i\in I} Y_{k_i,p_i}^{e_i})$ for some nonnegative integers $e_i$. But, as before, we must have $$\bd=\deg(m_\aa) = \sum_{i\in I}e_i\deg(Y_{k_i,p_i})=\sum_{i\in I}e_i\a_i,$$ hence $e_i=d_i$ for every $i$. Therefore $m_\aa$ is indeed a weight of $M(Y^{W^\bd})$. This proves the claim.
[ The proof of Proposition \[propgeo2\] shows that ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W^\bd)$ is a nonempty stratum of ${{\mathcal M}}^\bullet_0(W^{\bd})$ if and only if $(V,W^\bd)$ is a dominant pair, a purely combinatorial condition. In general Nakajima [@NJAMS Th. 14.3.2] only shows that this is a necessary condition. In representation-theoretic terms, this means that every dominant monomial of the form $Y^{W^\bd}A^V$ occurs in the $q$-character of the standard module $M(Y^{W^\bd})$. ]{}
$$\def\objectstyle{\scriptstyle}
\[email protected]{
&&& Y_{3,5} &
\\
&Y_{1,4} &Y_{2,4}&&Y_{4,4}
\\
&&&Y_{3,3}&
\\
&Y_{1,2} &Y_{2,2}&&Y_{4,2}
\\
&&& Y_{3,1}&
\\
&Y_{1,0} &Y_{2,0}& & Y_{4,0}
\\
}
\qquad
\def\objectstyle{\scriptstyle}
\[email protected]{
&&& \a_3 &
\\
&\a_1+\a_3 \ar[rru]&\a_2+\a_3\ar[ru]&&\ar[lu]\a_3+\a_4
\\
&&&\ar[llu]\ar[lu]\a_1+\a_2+2\a_3+\a_4\ar[ru]&
\\
&\a_2+\a_3+\a_4\ar[rru] &\a_1+\a_3+\a_4\ar[ru]&&\ar[lu]\a_1+\a_2+\a_3
\\
&&& \ar[llu]\ar[lu]\a_1+\a_2+\a_3+\a_4\ar[ru]&
\\
&\a_1\ar[rru] &\a_2\ar[ru]& & \ar[lu]\a_4
\\
}$$
We continue Example \[ExaPropGeo1\]. There are $12$ positive roots $\b_k$, which we identify with the vertices of the Auslander-Reiten quiver of $Q$ represented in Figure \[FigD\]. The numbering is obtained by reading this graph from top to bottom and left to right: $$\b_1=\a_3,\quad \b_2=\a_1+\a_3,\quad \b_3 = \a_2+\a_3,\quad \b_4 = \a_3+\a_4,
\quad \cdots ,\quad \b_{12}=\a_4.$$ The corresponding generators $Y_{i_k,p_k}$ of $\Y_{t,\,Q}$ can be read at the corresponding place on the left side of Figure \[FigD\]. Let $\bd=(d_1,d_2,d_3,d_4)$ be a dimension vector for $Q$. Then $$Y^{W^{\bd}} = Y_{1,0}^{d_1}Y_{2,0}^{d_2}Y_{3,5}^{d_3}Y_{4,0}^{d_4}.$$ The elements of $I_\bd$ are $12$-tuples $\aa\in{\mathbb{N}}^{12}$ encoding the decompositions of $\bd$ into a sum of positive roots. By Proposition \[propgeo2\], they are in one-to-one correspondence with the dominant monomials of the form $Y^{W^\bd} A^V$. This bijection can be read immediately from Figure \[FigD\].
For example, if $\bd = (1,1,1,1)\equiv \a_1+\a_2+\a_3+\a_4$, the correspondence is: $$\begin{array}{lclcl}
(\a_1)+(\a_2)+(\a_3)+(\a_4) &\leftrightarrow & Y_{1,0}Y_{2,0}Y_{3,5}Y_{4,0}
&\leftrightarrow &1
,\\ [1mm]
(\a_1+\a_3)+(\a_2)+(\a_4) &\leftrightarrow & Y_{1,4}Y_{2,0}Y_{4,0}
&\leftrightarrow &A_{1,1}A_{3,2}A_{2,3}A_{4,3}A_{3,4}
,\\ [1mm]
(\a_2+\a_3)+(\a_1)+(\a_4) &\leftrightarrow & Y_{2,4}Y_{1,0}Y_{4,0}
&\leftrightarrow &A_{2,1}A_{3,2}A_{1,3}A_{4,3}A_{3,4}
,\\ [1mm]
(\a_3+\a_4)+(\a_1)+(\a_2) &\leftrightarrow & Y_{4,4}Y_{1,0}Y_{2,0}
&\leftrightarrow &A_{4,1}A_{3,2}A_{1,3}A_{2,3}A_{3,4}
,\\ [1mm]
(\a_1+\a_2+\a_3)+(\a_4) &\leftrightarrow & Y_{4,2}Y_{4,0}
&\leftrightarrow &A_{1,1}A_{2,1}A_{3,2}^2A_{1,3}A_{2,3}A_{4,3}A_{3,4}
,\\ [1mm]
(\a_1+\a_3+\a_4)+(\a_2) &\leftrightarrow & Y_{2,2}Y_{2,0}
&\leftrightarrow &A_{1,1}A_{4,1}A_{3,2}^2A_{1,3}A_{2,3}A_{4,3}A_{3,4}
,\\ [1mm]
(\a_2+\a_3+\a_4)+(\a_1) &\leftrightarrow & Y_{1,2}Y_{1,0}
&\leftrightarrow &A_{2,1}A_{4,1}A_{3,2}^2A_{1,3}A_{2,3}A_{4,3}A_{3,4}
,\\ [1mm]
(\a_1+\a_2+\a_3+\a_4) &\leftrightarrow & Y_{3,1}
&\leftrightarrow &A_{1,1}A_{2,1}A_{4,1}A_{3,2}^2A_{1,3}A_{2,3}A_{4,3}A_{3,4}
.
\end{array}$$ It is obtained by replacing each root $\b_k$ in a decomposition of $\a_1+\a_2+\a_3+\a_4$ by the corresponding variable $Y_{i_k,p_k}$. The third column gives the monomial $Y^{W^\bd}\left(\prod_{k=1}^r Y_{i_k, p_k}^{a_k}\right)^{-1}$.
We can now state the main result of this section.
\[thmgeo\]
- We have a $G_\bd$-equivariant isomorphism of varieties $\Psi \colon {{\mathcal M}}^\bullet_0(W^\bd)\overset{\sim}{\longrightarrow} E_\bd$.
- ${{\mathcal M}}^\bullet_0(W^\bd)$ is an affine space of dimension $\sum_{i\to j} d_id_j$.
- Lusztig’s perverse sheaves $IC(\O_\aa)$ on $E_\bd$ are the same as Nakajima’s perverse sheaves $IC_{W^{\bd}}(V)$ on ${{\mathcal M}}^\bullet_0(W^\bd)$.
In Proposition \[propgeo1\] we have constructed a $G_\bd$-equivariant closed immersion $\Psi$ of ${{\mathcal M}}^\bullet_0(W^\bd)$ into $E_\bd$. Since each stratum ${{\mathcal M}}^{\bullet\, {\rm reg}}_0(V,W^\bd)$ is $G_{\bd}$-invariant, $\Psi$ maps every stratum to a union of orbits $\O_\aa$. Since $\Psi$ is injective and the number of strata is equal to the number of orbits (Proposition \[propgeo2\]), it follows that $\Psi$ maps each stratum of ${{\mathcal M}}^\bullet_0(W^{\bd})$ to a single $G_\bd$-orbit in $E_\bd$, so every orbit is contained in the image of $\Psi$. Thus $\Psi$ is surjective. Since a surjective closed immersion between reduced schemes is an isomorphism, this proves (a). Claim (b) follows immediately from (a), and claim (c) is again a consequence of Proposition \[propgeo2\], which shows that the stratifications used for defining the perverse sheaves are the same.
\[2remarks\]
(a) By the proof of Proposition \[propgeo2\], for every dominant monomial $m$ in $\Y_{t,\,Q}$ there is a unique pair $(V, W^\bd)$ such that $m = Y^{W^\bd}A^V$. Hence, even if the varieties ${{\mathcal M}}^\bullet_0(W^{\bd})$ involve very particular spaces $W^\bd$, the isomorphisms $\Psi\colon {{\mathcal M}}^\bullet_0(W^{\bd})\overset{\sim}{\longrightarrow} E_\bd$ are enough to identify all the irreducible $(q,t)$-characters of $\CC_{Q}$.
Thus Theorem \[LusztigTh\], Theorem \[NakajimaTh\], and Theorem \[thmgeo\] provide a geometric explanation of part (b) of Theorem \[mainth\]. By comparing convolution diagrams in [@Lu2] and [@VV], it should also be possible to understand in a geometric manner part (a) of Theorem \[mainth\], that is, the multiplicative structure (see [@Ncl §3.5]).
(b) If we take for $Q$ a quiver of type $A$ with all arrows in the same direction, then ${{\mathcal M}}^\bullet_0(W^{\bd})$ is just a space of graded nilpotent endomorphisms as in the Ginzburg-Vasserot construction [@GV] of type $A$ quantum loop algebras (see [@L §2.5.3]). So Theorem \[thmgeo\] becomes tautological in this case.
Acknowledgements {#acknowledgements .unnumbered}
----------------
We thank H. Nakajima for helpful comments and discussions, in particular about the proof of Theorem \[thmgeo\].
[ABCD]{}
, *Representation theory of Artin algebras*, Cambridge studies in advanced mathematics [**36**]{}, Cambridge 1995.
, *Elements of the representation theory of associative algebras. Vol.1*, London Math. Soc. Student Texts [**65**]{}, Cambridge 2006.
, [*String bases for quantum groups of type $A_r$*]{}, in I.M. Gelfand seminar, Adv. Soviet Math. [**16**]{}, pp. 51–89, AMS 1993.
, [*Quantum cluster algebras,*]{} Adv. Math. [**195**]{} (2005), 405–455.
, [*Beyond Kirillov-Reshetikhin modules*]{}, Quantum affine algebras, extended affine Lie algebras, and their applications, Contemp. Math., [**506**]{} (2010), 49–81.
, [*Minimal affinizations of representations of quantum groups: the simply laced case*]{}, J. Algebra [**184**]{} (1996), 1–30.
, *The solution of the quantum $A_1$ $T$-system for arbitrary boundary*, Comm. Math. Phys. [**313**]{} (2012), 329–350.
, [*A new realization of Yangians and quantized affine algebras*]{}, Soviet Math. Doklady [**36**]{} (1988), 212–216.
, [*Combinatorics of $q$-characters of finite-dimensional representations of quantum affine algebras*]{}, Comm. Math. Phys. [**216**]{} (2001), 23–57.
, [*Deformations of $W$-algebras associated to simple Lie algebras*]{}, Commun. Math. Phys. [**197**]{} (1998), 1–32.
, [*The $q$-characters of representations of quantum affine algebras*]{}, Recent developments in quantum affine algebras and related topics, Contemp. Math. [**248**]{} (1999), 163–205.
, *Representations of finite-dimensional algebras*, Encyclopaedia of Mathematical Sciences [**73**]{}, Springer 1997.
, [*Cluster structures on quantum coordinate rings*]{}, Selecta Math. New Series, DOI 10.1007/s00029-012-0099-x, http://link.springer.om/article/10.1007/s00029-012-0099-x (July 2012), arXiv:1104.0531 (2011).
, *Langlands reciprocity for affine quantum groups of type $A_n$*, Internat. Math. Res. Notices **3** (1993), 67–85.
, *Triangulated categories in the representation theory of finite-dimensional algebras*, London Math. Soc. Lecture Notes [**119**]{}, Cambridge 1988.
, [*$t$-analogues des opérateurs d’écrantage associés aux $q$-caractères*]{}, I. M. R. N. [**8**]{} (2003), 451–475.
, [*Algebraic approach to $q,t$-characters*]{}, Adv. Math. [**187**]{} (2004), 1–52.
, [*The Kirillov-Reshetikhin conjecture and solutions of $T$-systems*]{}, J. Reine Angew. Math. [**596**]{} (2006), 63–87.
, [*Simple tensor products*]{}, Invent. Math. [**181**]{} (2010), 649–675.
, [*Cluster algebras and quantum affine algebras*]{}, Duke Math. J. **154** (2010), 265–341.
, [*Periodicities of $T$-systems and $Y$-systems*]{}, Nagoya Math. J.[**197**]{} (2010), 59—174.
, [*On crystal bases of the $q$-analogue of universal enveloping algebras*]{}, Duke Math. J. [**63**]{} (1991), 465–516.
, [*A diagrammatic approach to categorification of quantum groups I*]{}, Represent. Theory [**13**]{} (2009), 309–347.
, [*Functional relations in solvable lattice models. I. Functional relations and representation theory*]{}. Internat. J. Modern Phys. A [**9**]{} (1994), 5215–5266,
, [*Quantum cluster algebras of type A and the dual canonical basis*]{}, arXiv:1101.0580, (2011), 46 p.
, [*Quantum loop algebras, quiver varieties, and cluster algebras*]{}, in Representations of algebras and related topics, EMS Series of Congress Reports, Ed. A. Skowroński, K. Yamagata, 2011, 117-152.
, *Canonical bases arising from quantized enveloping algebras*, J. Amer. Math. Soc. [**3**]{} (1990), 447-498.
, *Quivers, perverse sheaves, and quantized enveloping algebras*, J. Amer. Math. Soc. 4 (1991), no. 2, 365–421.
, *Quiver varieties and finite-dimensional representations of quantum affine algebras*, J. Amer. Math. Soc. **14** (2001), 145–238.
, [*$t$-analogs of $q$-characters of Kirillov-Reshetikhin modules of quantum affine algebras*]{}, Represent. Theory [**7**]{} (2003), 259–274.
, [*Quiver varieties and $t$-analogs of $q$-characters of quantum affine algebras*]{}, Annals of Math. [**160**]{} (2004), 1057–1097.
, [*Quiver varieties and cluster algebras*]{}, Kyoto J. Math. [**51**]{} (2011), 71–126.
, [*Quantum Cluster Variables via Serre Polynomials*]{}, J. Reine Angew. Math., [**668**]{} (2012), 149–190.
, [*Tame algebras and quadratic forms*]{}, Lecture Notes in Math. [**1099**]{}, Springer 1984.
, [*Hall algebras and quantum groups*]{}, Invent. Math. [**101**]{} (1990), 583–592.
, [*Hall algebras revisited*]{}, in Israel Math. Conf. Proc. [**7**]{}, (1993), 171–176.
, [*PBW-bases of quantum groups*]{}, J. Reine Angew. Math. [**470**]{} (1996), 51–88.
, *2-Kac-Moody algebras*, arXiv:0812.5023.
, *AR quivers, exceptional sequences and algorithms in derived Hall algebras*, Science China (2010), doi: 10.1007/s11425-010-0069-8.
, [*Derived Hall algebras*]{}, Duke Math. J. [**135**]{} (2006), 587–615.
, [*Perverse sheaves and quantum Grothendieck rings*]{}, in Studies in memory of Issai Schur, Prog. Math. [**210**]{}, Birkhäuser, 2002, 345–365.
, [*Canonical bases and KLR algebras*]{}, J. Reine Angew. Math. 2011, [**659**]{}, 67–100.
, [*Hall algebras associated to triangulated categories*]{}, Duke Math. J. (2008) [**143**]{}, 357–373.
------------------------------- -------------------------------------------------------------------------
David [Hernandez]{} : Université Paris Diderot-Paris 7,
Institut de Mathématiques de Jussieu - Paris Rive Gauche CNRS UMR 7586,
Bâtiment Sophie Germain, Case 7012,
75205 Paris Cedex 13, France.
email : [[email protected]]{}
\[5mm\] Bernard [Leclerc]{} : Normandie Univ, France
UNICAEN, LMNO F-14032 Caen, France
CNRS UMR 6139, F-14032 Caen, France
Institut Universitaire de France,
email : [[email protected]]{}
------------------------------- -------------------------------------------------------------------------
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We proof here the existence of a topological thick and thin decomposition of any closed definable thick isolated singularity germ in the spirit of the recently discovered metric thick and thin decomposition of complex normal surface singularities of [@BNP]. Our thin zone catches exactly the homology of the family of the links collapsing faster than linearly. Simultaneously we introduce a class of rigid homeomorphisms more general than bi-Lipschitz ones, which map the topological thin zone onto the topological thin zone of its image. As a consequence of this point of view for the class of singularities we consider we exhibit an equivalent description of the notion of separating sets in terms of this fast contracting homology.'
address:
- 'Departamento de Matemática, Universidade Federal do Ceará (UFC), Campus do Picici, Bloco 914, Cep. 60455-760. Fortaleza-Ce, Brasil'
- 'Departamento de Matemática, Universidade Federal do Ceará (UFC), Campus do Picici, Bloco 914, Cep. 60455-760. Fortaleza-Ce, Brasil'
- 'Departamento de Matemática, Universidade Federal do Ceará (UFC), Campus do Picici, Bloco 914, Cep. 60455-760. Fortaleza-Ce, Brasil'
author:
- Lev Birbrair$^1$
- Alexandre Fernandes$^2$
- Vincent Grandjean$^3$
title: Collapsing topology of isolated singularities
---
Introduction
============
A subset of a Riemannian manifold carries two natural metrics: the *outer metric *where the distance between points is measured in the ambient space, and the *inner metric *where the infimum of the lengths of rectifiable curves joining the points considered realizes the distance. Although the underlying topological structures are homeomorphic, the corresponding metric spaces may not be equivalent.****
Real tame or complex analytic subsets share the same local structure, namely they are topological cones over their link [@Loj; @Mil; @vdDMi]. (The adjective *tame *has to be understood as *topologically tame, *as meant in Grothendieck’s *Esquisse d’un programme*[).]{} Since such subsets are always (locally) embedded in a real or complex Euclidean space, they come equipped with the inner metric and with the outer metric. Although the local inner and outer geometric structure of real tame and complex subsets are more delicate to understand than their topology, some tools are available to do so. Following the works of the first two named authors [@Fe1; @Bi2] for curves (see also [@PhTe]) and surfaces definable in an o-minimal structure expanding the field of real numbers, and of Grieser [@Gri] for real analytic surfaces with isolated singularities, the notion of metric type introduced by Valette [@Va1; @Va2] gives a thorough description of the local metric combinatorics of the corresponding subsets. The topology of the family of the links - the family of intersections of a representative of the germ with the Euclidean spheres centered at the origin and of small radii - of tame isolated singularities may be rich, but its metric behavior when the radii tends to $0$ has still a long way to go before being fully understood. Once more some tools are available to deals with such a problem: Birbrair-Brasselet metric homology [@BiBr1; @BiBr2] and Valette vanishing homology [@Va3] provide some insights into this topic. But more is needed.******
The bottom line of the investigation from a metric point of view (inner or outer) is to understand the locally conic geometry of the class of germs and then proceed to higher degenerate metric types. Unlike tame real sets, complex analytic set germs have tangent cones of maximal dimension, consequently most of the (inner and outer) metric type is of conic nature. The principal reason to insist on the locally conic locus of a given singularity set-germ is - as it happens for cones - that the collapsing of the topology of the family of links is not *faster than linearly, *that is not faster than the velocity of the parameter (radius) of the family tending to $0$ (see [@BiBr1; @BiBr2; @Va2] to give meanings to this expression).**
Prior to the paper by the two first named authors [@BF3] was a hazy belief that germs of complex analytic sets could be conic. They found the first example of a complex isolated singularity (a Pham-Brieskorn surface singularity) which cannot be conic at its singular point [@BF3] and did so exhibiting a so called *separating set *which produces a homological obstruction in the family of links. Later, joined by Neumann, they provided more examples of non-conical isolated complex surface singularities [@BF3; @BFN1; @BFN3]. In their very recent work [@BNP], the first named author, Neumann and Pichon completely describe the inner metric structure of complex normal surface singularities. They exhibit a (*metric) *decomposition of the surface germ into thick pieces and thin pieces: The thin pieces catch exactly all the topology of the family of links collapsing onto the origin faster than linearly. These authors do not only give the precise rates of each collapsing, but they also describe very precisely the structures corresponding to each rate and how they are organized relatively to each other by means of a rather elementary combinatorial object. A second remarkable feat of this decomposition is that the complement of any conical neighborhood of the thin zone is an inner-metric cone. Their decomposition is minimal for these two properties. This work rely on the very rich amount of data available for surface complex singularities and topology of $3$-folds, and how all these different points of view fit together.****
Although the aforementioned works relate to the Lipschitz Geometry of tame subsets, the point of view of our paper is slightly more general. Inspired by the quoted paper [@BNP], our starting point was to look for the existence of a real tame avatar of a thick-and-thin-like decomposition for any tame isolated singularity germ. Can such singularity germs have a *topological *thick and thin decomposition, namely: *Does there exist a decomposition into thick zones and thin zones such that the latter one catches the topology of the family of the links collapsing faster than linearly? *\
As we will see our slightly more general point of view on the topology of the family of the links will bring tools to investigate the local metric conicalness. Note also that describing the metric structure of the thin zone in our general framework is an extremely arduous task, since there does not exist any practically usable quantified complexity of such structure, although such formal data exist in the polynomially bounded case [@Va2].****
The present paper proposes a solution to the problem of the existence of a topological thick and thin zone for thick tame isolated singularities. It is organized as follows.
In order to catch the fast contracting topology of the link, we introduce the notion of *fast contracting cycles, *which are cycles supported in the link and contracted in the subset by a tame chain whose support is thin and contains the origin (Definition \[def:admis-van-hom\]). Attached to this classification problem comes the right class of mappings that we call *morphisms, *which are tame continuous mappings which extend as tame and continuous mappings after spherical blowings-up at the source and the target (Definition \[def:morphism\]). The corresponding homeomorphisms will be called *isomorphisms. *The idea of using spherical blowings-up is completely natural when it comes to deal with the tangent directions at the center of the blowing-up, since the tangent link is embedded as the boundary of the strict transform (defined as the intersection of the strict transform with the ambient exceptional divisor). It turns out that the fast contracting cycles form a sub-complex of the usual homology of the link and that any morphism induces a homomorphism between the corresponding subgroups (Proposition \[prop:homom-fc-homol\]). Thanks to these simple notions we can weaken the notion of *normal embedding model *of a (tame) subset [@BiMo] into the notion of *simply embedded model *which is any isomorphic image of a normally embedded model (Definition \[def:SE-mod\]). We then find out necessary and sufficient conditions on a simply embedded model (of a tame closed thick singularity) to be a cone over a topological tame submanifold (Lemma \[lem:loc-conic-homeo-cone\]).**********
In order to find this topological thick and thin composition for tame closed thick isolated singularity our proof rely mostly on a weak version of Federer-Flemming Theorem on the triviality of cycles with small volumes [@FF]. Namely we prove [**Theorem \[thm:FF-polyhedra\].**]{} *Let $\Gamma$ be a compact polyhedron defined as the union of the simplices of a finite set $T$. Let $\xi$ be a definable $k$-cycle in $\Gamma$ such that for any positive real number ${{\varepsilon}}$ small enough there exists a homologous definable $k$-cycle $\xi_{{{\varepsilon}}}$ in $\Gamma$ whose $k$-dimensional volume ${{\rm vol}}_k({{\rm supp}}(\xi_{{{\varepsilon}}}))$ is smaller than ${{\varepsilon}}$. Then the cycle $\xi$ bounds a definable $(k+1)$-chain of $\Gamma$. *As expected consequence of Theorem \[thm:FF-polyhedra\], cones and their isomorphism classes cannot carry any fast contracting homology (Corollary \[cor:cone-no-hom\]).**
Having identified the locus of *non-simple tangent directions *${{\bf S}_{{\bf 0}}^{\rm ns}}X_1$ as the subset of points in the tangent link of a simply embedded model $X_1$ of a closed tame thick isolated singularity where the fast contracting cycles must collapse (Definition \[def:simple-dir\]), and showed that it is tame closed and of codimension at least $1$ in the tangent link of $X_1$, we can prove the first version of our *topological thick and thin decomposition: *[**Theorem \[thm:main\].**]{} *Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin ${{\bf 0}}$ and with connected link at the origin. For any positive ${{\varepsilon}}$ small enough, the inclusion mapping induces the following exact sequence $${\mathcal{H}}_\bullet (X_1\cap {{\rm Cone}}({{\bf S}_{{\bf 0}}^{\rm ns}}X_1, {{\varepsilon}}),{{\bf 0}}) \to {\mathcal{H}}_\bullet (X_1,{{\bf 0}}) \to 0.$$ where ${\mathcal{H}}_\bullet (X_1,{{\bf 0}})$ is the sub-complex corresponding to the fast contracting homology classes. *******
The proof of Theorem \[thm:main\] consists of finding a definable triangulation, which exists, in the simplices of which we can apply Theorem \[thm:FF-polyhedra\] whenever possible.
Since the existence of a topological thick and thin decomposition is now established, we define the topological thin zone in a systematic way (Definition \[def:thin-zone\]) as the image, by the spherical blowing-down mapping, of any *tame neighborhood *(Definition \[def:def-nghb\]), in the strict transform, of the embedding of the non-simple point locus in the boundary of the strict transform as the subset ${{\bf S}_{{\bf 0}}^{\rm ns}}X_1\times 0$. We check that this definition makes sense (Proposition \[prop:defin-neighb\]). Our main result can be rephrased as [**Theorem \[thm:main-bis\].**]{} *Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin and with connected link. The inclusion mapping induces the following exact sequence $${\mathcal{H}}_\bullet ({{\rm Thin}}(X_1),{{\bf 0}}) \to {\mathcal{H}}_\bullet (X_1,{{\bf 0}}) \to 0.$$ where ${{\rm Thin}}(X_1)$ is the topological thin zone of the germ $(X_1,{{\bf 0}})$. *Eventually, we present a few examples to illustrate our point of view. All the (so far) known obstructions to local inner metric conicalness, namely *fast loops, separating sets *and *choking horns, *can be very well described by the fast contracting homology of the topological thin zone. These notions are not only bi-Lipschitz invariant but also isomorphic invariant when it comes to simply embedded models of tame thick isolated singularities. We insist particularly on the case of separating sets. Condition (SC), introduced in Definition \[def:cond-SC\], holds true for simply embedded model of closed tame thick isolated singularities if and only if there exists a separating set. We end the paper by bringing to light the relation of the existence of separating sets with the fast contracting homology of the toplogical thin zone:********
[**Theorem \[thm:sep-set-pure-dim\].**]{} *Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin and with connected link. Suppose the tangent link is of pure dimension $d-1$. There exists a separating set if and only if the surjective homomorphism of the fc-homology groups*
$\iota_*:{\mathcal{H}}_{d-2} ({{\rm Thin}}(X_1),{{\bf 0}}) \to {\mathcal{H}}_{d-2}(X_1,{{\bf 0}})$
induced by the inclusion mapping $\iota:{{\rm Thin}}(X_1) \to X_1$ is not injective. **
Background - Notations {#Section:background}
======================
We present in this section the few definitions and notations we are going to use in the paper. We also expect here to provide most of the material to present this note as rather self-contained.
In order to avoid any confusion about the notion of cone, we recall the following
A subset $C$ of some Euclidean space ${{\mathbb{R}}}^p$ is a cone with vertex the origin if there exists a subset $L$ of the unit sphere ${{\bf S}}^{p-1}$, *the link of the cone, *such that**
$C := {{\rm Cone}}(L) = \{tu: u\in L$, and $t>0\}\cup \{{{\bf 0}}\}$.
[**Notation:**]{} For any non-zero vector $x$ of some Euclidean space, we denote by $\nu (x)$ the oriented direction of the vector $x$
$\nu (x) := \displaystyle{\frac{x}{|x|}}$.
Let $X$ be a subset of some Euclidean space ${{\mathbb{R}}}^p$ whose closure ${{\rm clos}}(X)$ contains the origin ${{\bf 0}}$. The *tangent cone *$T_{{\bf 0}}X$ of $X$ at ${{\bf 0}}$ is the closed cone over the subset ${{\bf S}_{{\bf 0}}}X$, *the tangent link of the subset $X$ at ${{\bf 0}}$ *defined as****
${{\bf S}_{{\bf 0}}}X := \{{{\bf u}}\in {{\bf S}}^{p-1}:$ [there exists]{} $(x_k)_{k\geq 1}\in X\setminus {{\bf 0}}$ [such that]{} $\hfill$
$\hfill x_k \to {{\bf 0}}$ [and]{} $\nu (x_k) \to {{\bf u}}$ [as]{} $k\to+\infty\}$.
As already mentioned in the introduction, any subset $X$ of some Euclidean space can be equipped with two metrics: The *outer metric *measures the distances in the ambient Euclidean space while the *inner metric *measures the distance between two points of $X$ as the infimum of the length of rectifiable curves within $X$ connecting the two points. These two structure of metric spaces on $X$ are not necessarily equivalent metric structures. This fact justify the introduction of the following****
\[def:n-e\] i) A subset $X$ of ${{\mathbb{R}}}^p$ is *normally embedded *if the identity mapping between the metric spaces $(X,{{\rm d_{\rm outer}}})$ and $(X,{{\rm d_{\rm inner}}})$ is bi-Lipschitz.\
ii) Any normally embedded subset $X$ bi-Lipschitz homeomorphic to a subset $X_0$ contained in ${{\mathbb{R}}}^q$ and equipped with the inner metric is called *a normally embedded model *of $X_0$.****
Throughout this paper we are going to work with subsets definable in an o-minimal structure expanding the field of real numbers. For readers not familiar with an o-minimal framework, they can think of the category of the semi-algebraic subsets - which form an o-minimal structure - (see [@vdDMi; @vdD] for definitions and standard properties).
[**Important Note:**]{} We assume for the rest of the paper we are given an o-minimal structure ${\mathcal{M}}$ expanding the field of real numbers. Any future occurrence of the word *definable *will mean definable in this o-minimal structure ${\mathcal{M}}$.**
If $X$ is a compact definable subset of ${{\mathbb{R}}}^p$, *normally embedded *will mean *definably normally embedded *, that is the bi-Lipschitz homeomorphism in Definition \[def:n-e\] is required to also be a definable mapping.****
A compact definable subset $X_0$ of some Euclidean space ${{\mathbb{R}}}^p$ always admits a *normally embedded model, *that is there exists a normally embedded definable subset $X$ of some ${{\mathbb{R}}}^n$ which is definably bi-Lipschitz homeomorphic to the metric space $(X_0,{{\rm d_{\rm inner}}})$ (the definable case works exactly as the semi-algebraic case of [@BiMo] since the existence of a definable pancake decomposition can be proved along the same lines as those of the subanalytic case [@Kur], see also [@Par; @Sh; @Co; @Paw; @Va3]). We introduce now two notions which will be key for the rest of the paper**
\[def:thick-thin\] Let $(X,{{\bf 0}})$ be the germ at the origin of a closed definable subset of some Euclidean space. Let $d$ be the dimension of the set-germ $(X,{{\bf 0}})$.\
i) The subset $X$ is *thick at the origin ${{\bf 0}}$ *if the tangent link ${{\bf S}_{{\bf 0}}}X$ at ${{\bf 0}}$ has dimension $d-1$.\
ii) The subset $X$ is *thin at the origin ${{\bf 0}}$ *if the tangent link ${{\bf S}_{{\bf 0}}}X$ at ${{\bf 0}}$ has dimension strictly smaller than $d-1$.****
Note that the notion of *thick *we propose in Definition \[def:thick-thin\] is not as specific and thus looser than that provided in [@BNP].**
Let $X$ be a definable subset of some Euclidean space ${{{\mathbb{R}}}^n}$. We recall that the germ $(X,x)$ of $X$ at any of its point $x$ is topologically conic: Indeed, there exists ${{\varepsilon}}>0$ such that the intersection $X\cap {{\bf B}}^n(x,r)$, where ${{\bf B}}^n(x,r)$ is the ambient Euclidean open ball of ${{{\mathbb{R}}}^n}$ centered at $x$ of radius $r$, is definably homeomorphic to the cone over the intersection $X\cap {{{\bf S}}^{n-1}}(x,r)$ for any positive radius $r<{{\varepsilon}}$ where ${{{\bf S}}^{n-1}}(x,r)$ is the Euclidean sphere of radius $r$ and centered $x$ ([@Loj; @Mil; @vdDMi; @vdD; @Sh; @Co]).
The *link of *$X$ at $x$ is the topological space homeomorphic to any intersection $X\cap {{{\bf S}}^{n-1}}(x,{{\varepsilon}})$ for any sufficiently small positive ${{\varepsilon}}$. By a minor abuse of language we will also call *link *any intersection $X\cap {{{\bf S}}^{n-1}}(x,{{\varepsilon}})$ when ${{\varepsilon}}$ is small enough.****
Hereafter, every representative of a germ $(X,x)$ is considered into an ambient Euclidean open ball ${{\bf B}}(x,r)$ of radius $r<{{\varepsilon}}.$
\[def:isol-sing\] The germ at the origin ${{\bf 0}}$ of a definable closed subset $X$ of some Euclidean space ${{\mathbb{R}}}^p$ is called *a closed definable isolated singularity, *if $(X\setminus {{\bf 0}},{{\bf 0}})$ is the germ of a $C^1$ submanifold of ${{\mathbb{R}}}^p$.**
Although the notion of isolated singularity is classical, we recall it in the context we are interested in to emphasize that in the following sections we will not only deal with a closed isolated singularity but also and mainly with definably homeomorphic images, allowing the possibility of singular points outside the origin. In particular we will speak very often of a *normally embedded model *and of a *simply embedded model *(see Definition \[def:SE-mod\]) of the given closed definable isolated singularity, which are definably homeomorphic images, with additional properties. Since the link of an isolated singularity is always a $C^1$ submanifold, the link of any definably homeomorphic image will still be a definable topological submanifold We end-up this section recalling the notion of *spherical blowing-up *of the origin of ${{{\mathbb{R}}}^n}$ defined as the following mapping******
$\beta_n : {{{\bf S}}^{n-1}}\times {{\mathbb{R}}}_+ \to {{{\mathbb{R}}}^n}\,$, $\,({{\bf u}},r) \to r{{\bf u}}$.
The *strict transform *of the subset $X$ under the spherical blowing-up $\beta_n$ is the subset of ${{{\bf S}}^{n-1}}\times {{\mathbb{R}}}_+$ defined as**
$X':= {{\rm clos}}(\beta_n^{-1}(X\setminus {{\bf 0}})$.
The *boundary ${{\partial}}X'$ *of the strict transform $X'$ is just the intersection of $X'$ with the ambient exceptional divisor ${{{\bf S}}^{n-1}}\times 0$, namely,**
${{\partial}}X':=X'\cap ({{{\bf S}}^{n-1}}\times 0) = {{\bf S}_{{\bf 0}}}X \times 0$.
We would like to insist on the fact that, since the tangent link ${{\bf S}_{{\bf 0}}}X$ of $X$ at the origin is the basis of the tangent cone $T_{{\bf 0}}X$ of $X$ at the origin, the main interest for us in using a spherical blowing-up is to give substance to the tangent link of $X$, embedded in the strict transform $X'$ of $X$ as the boundary ${{\partial}}X'$. The strict transform $X'$ allows to see simultaneously the geometric object we are interested in ($X$) and its tangent cone at the origin ($T_{{\bf 0}}X$).
Note that when there will be no ambiguity on the ambient Euclidean space we are in, we will just write $\beta$.
Fast contracting cycles and fast contracting homology {#Section:FCCaFCH}
=====================================================
We present here the objects and the morphisms of the classification problem as well as some of their elementary properties that will allow to define in Sections \[Section:MR\] and \[Section:TaToTIS\] the notion of *topological thin zone. ***
We want to classify germs of definable subsets (of some Euclidean space) up to a class of definable homeomorphisms with special properties. The simplest model of germs, up to these special homeomorphisms, we have in mind are just cones. First, we present the objects under our considerations which we want to be preserved by these special homeomorphisms. Only after, we will define precisely what are these special homeomorphisms.
Let $(X,{{\bf 0}})$ be a germ of definable closed subset at the origin ${{\bf 0}}$ of some Euclidean space and with connected link at the origin.
We recall that any singular homology class of any given bounded definable set can be realized by a definable cycle [@Loj; @Har; @vdD; @Sh; @Co].
\[def:admis-van-hom\] A definable singular $k$-dimensional cycle $\xi$ of $X$ *supported in the link *of $X$, that is whose support is contained in $X\setminus {{\bf 0}}$, is said to be *fast-contracting *if there exists a definable $(k+1)$-dimensional chain $\eta$ on $X$ bounded by the cycle $\xi$, whose support is thin at the origin and contains the origin.****
We will shorten the expression *fast contracting cycle supported in the link *into *fc-cycle. *****
We denote ${\mathcal{H}}_k(X,{{\bf 0}})$ the subset of the $k$-th singular homology group of the link $H_k(X\setminus{{\bf 0}})$ consisting only of the homology classes represented by fc-cycles. We observe that ${\mathcal{H}}_k(X,{{\bf 0}})$ is a subgroup of $H_k(X\setminus{{\bf 0}})$.
\[def:van-hom-grp\] The group ${\mathcal{H}}_k(X,{{\bf 0}})$ is called the *$k$-th local homology group of the fc-cycles of $X$ at the origin. ***
Any homology class of the link represented by a fast contracting cycle supported in the link will be called a *fast contracting homology class *and shortened to *fc-homology class. *****
We do not precise the coefficients group since this will not make any change on what we are going to present and demonstrate. What we have in minds are the standard coefficients groups: ${{\mathbb{Z}}}/2, {{\mathbb{Z}}}, {{\mathbb{Q}}}$ and ${{\mathbb{R}}}$.
Having specified the type of properties we want to classify germs of isolated singularities by, we introduce the class of morphisms which goes with our classification problem.
Let $(X,{{\bf 0}})$ and $(Y,{{\bf 0}})$ be two germs of closed definable subsets of ${{{\mathbb{R}}}^n}$ and ${{\mathbb{R}}}^m$ respectively.
\[def:morphism\] A continuous definable mapping $\Phi:(X,{{\bf 0}})\to(Y,{{\bf 0}})$ is called a *morphism *if**
i\) the image $\Phi(X\setminus {{\bf 0}})$ is contained in $Y\setminus{{\bf 0}}$;
ii\) the mapping
$\beta_m^{-1}\circ \Phi\circ \beta_n: X'\setminus {{\partial}}X' \to Y'\setminus {{\partial}}Y'$
extends as a continuous definable mapping $\Phi':X'\to Y'$. In particular the mapping $\Phi'$ induces a continuous definable mapping from the boundary ${{\partial}}X'$ of $X'$ into the boundary ${{\partial}}Y'$ of $Y'$.
An *epi-morphism *is a morphism such that the extension mapping is surjective.**
A *mono-morphism *is a morphism such that the extension mapping is injective.**
A morphism which is an epi-morphism and a mono-morphism admits a continuous definable inverse mapping which extends as an epi-morphism and a mono-morphism. We call such a morphism an *isomorphism. *[**Examples.**]{} 1) Let $S_\alpha:=\{(x,y,z) \in {{\mathbb{R}}}^3: = x^2+y^2 = z^\alpha, z\geq 0\}$ be the $\alpha$-horn, for a real number $\alpha >2$. Let $\pi$ be the orthogonal projection from the $\alpha$-horn $S_\alpha$ to the horizontal plane $\{z=0\}$. Whence $S_\alpha$ is equipped with the outer metric of ${{\mathbb{R}}}^3$, the mapping $\pi$ is definable and Lipschitz but is not a morphism. Its inverse $\pi^{-1}$, which is not Lipschitz, is a morphism but is not an isomorphism.**
2\) Any definable Lipschitz homeomorphism $(X,{{\rm d_{\rm outer}}}) \to (Y,{{\rm d_{\rm outer}}})$ between definable germs embedded in some Euclidean spaces extends as an isomorphism (see below Proposition \[prop:bilip-iso\] and [@BeLy]).
3\) Let $T$ be a closed definable cone with vertex the origin. The homeomorphism $T\ni x \to |x| x \in T$ and its inverse are both isomorphisms. Because of its definability and its fast collapsing to the origin, this homeomorphism admits a differential-like mapping at the origin $T_{{\bf 0}}X \to T_{{\bf 0}}Y$ as in [@BeLy] and is mapping the whole tangent cone $T_{{\bf 0}}X$ onto the null vector. Even though the inverse homeomorphism does not admit a differential-like mapping, it induces a homeomorphism between the tangent links.
4\) We use [@BF3 Example 4.3] to emphasize even more the difference between our notion of isomorphism and definable outer-bi-Lipschitz homeomorphisms. Consider the family of semi-algebraic isolated hypersurface singularities of ${{\mathbb{R}}}^{n-1}\times{{\mathbb{R}}}_+$ defined as:
$H_\lambda : = \{(x,t)\in {{\mathbb{R}}}^{n-1}\times {{\mathbb{R}}}_+: P(x,t)P(x,-t) = t^\lambda\}$ with $\hfill$
$\hfill P(x,t) = -t^2 + (x_1-t)^2 + \sum_{i=2}^{n-1} x_i^2$ and $\lambda\in {{\mathbb{Q}}}_{>2}$.
This family was used when $n=4$ to exhibit a *separating set, *namely the subset $H_\lambda \cap \{x_1 =0\}$, whose existence obstructs the inner metric conicalness of $H_\lambda$ (see Section \[Section:examples\] for the relation of separating set and fast contracting homology). Note that if $\lambda'>\lambda>2$ the subsets $H_{\lambda'}$ and $H_\lambda$ cannot be outer bi-Lipschitz homeomorphic. Indeed, let $\Phi_\alpha:{{\mathbb{R}}}^{n-1}\times {{\mathbb{R}}}_+\to{{\mathbb{R}}}^{n-1}\times {{\mathbb{R}}}_+$ be the definable mapping defined as $(x,t)\to
(t^\alpha x,t^{\alpha+1})$ for $\alpha$ a positive rational number. similarly to example 3), this mapping induces an isomorphism from $H_\lambda$ onto $H_{\lambda'}$ with $\lambda' := \lambda + \alpha(\lambda -2)$.**
As already suggested by the third and fourth example above, the notion of isomorphisms we introduce here is much more general than the notion of definable outer bi-Lipschitz homeomorphisms. As expected, one of the interests of the fc-homology groups lies in the following result:
\[prop:homom-fc-homol\] Let $(X,{{\bf 0}})$ and $(Y,{{\bf 0}})$ be germs of closed definable subsets of some Euclidean spaces. For each $k\leq \dim Y$, a morphism $\Phi:(X,{{\bf 0}})\to(Y,{{\bf 0}})$ induces a homomorphism $\Phi_*: {\mathcal{H}}_k(X,{{\bf 0}})\to{\mathcal{H}}_k(Y,{{\bf 0}})$. If $\Phi$ is an isomorphism then $\Phi_*$ is also an isomorphism of groups.
We recall that the dimension of the image of a definable subset by a definable and continuous mapping is lower than or equal to the dimension of the subset. The image of a definable $k$-chain $\xi$ by the mapping $\Phi$ is a definable $k$-chain. Since the cycle $\xi$ bounds a thin chain $\eta$ whose support contains the origin, its image $\Phi(\xi)$ bounds the image $\Phi (\eta)$. When embedded in the boundary of the strict transform $X'$, the image of the tangent link of the support ${{\rm supp}}(\eta)$ by the mapping $\Phi'$ is of dimension lower than or equal to the dimension of the tangent link of ${{\rm supp}}(\eta)$, thus the image $\Phi(\xi)$ is contracted by a thin chain, namely the image $\Phi (\eta)$.
We can even be more specific requiring an additional information on the collapsing phenomenon of a given fc-homology class taking into account a measure of the collapsing.
\[def:dim-drop\] 1) The *dimension drop at the origin *of a fc-cycle $\xi$ is the maximum of $\dim {{\rm supp}}(\eta) -
\dim T_0 {{\rm supp}}(\eta)$ taken over the definable chains $\eta$ contracting $\xi$ in $X$ whose support contains the origin and is thin.\
2) The *dimension drop of a fc-homology class $[c]$ *is the maximum of the dimension drop at the origin of the fc-cycles homologous to $[c]$.****
For any degree $k = 1,\ldots,\dim X$, and any dimension drop $\delta =1,\ldots,k$, the fc-homology classes of ${\mathcal{H}}_k(X,{{\bf 0}})$ whose dimension drop is larger than or equal to $\delta$ form a subgroup that we denote ${\mathcal{H}}_k^\delta(X,{{\bf 0}})$. We find the following inclusions
${\mathcal{H}}_k^k (X,{{\bf 0}}) \subset \ldots \subset {\mathcal{H}}_k^1 (X, {{\bf 0}}) ={\mathcal{H}}_k (X, {{\bf 0}})$.
Using Proposition \[prop:homom-fc-homol\] and its proof, we refine its statement in the following way
\[prop:homom-fc-homol-degree\] Let $(X,{{\bf 0}})$ and $(Y,{{\bf 0}})$ be germs of closed definable subsets of some Euclidean spaces. For each degree $k\leq \dim (Y,{{\bf 0}})$ and each dimension drop $\delta =1,\ldots,k$, any morphism $\Phi:(X,{{\bf 0}})\to(Y,{{\bf 0}})$ induces a homomorphism $\Phi_*: {\mathcal{H}}_k^\delta(X,{{\bf 0}})\to{\mathcal{H}}_k^\delta(Y,{{\bf 0}})$.
Finally let us remark that when the tangent link of the subset $X$ at the origin has dimension $e-1 < \dim (X,{{\bf 0}}) -1$, then for any degree $k \geq e$ we necessarily find ${\mathcal{H}}_k(X,{{\bf 0}}) = H_k(X\setminus {{\bf 0}})$. In particular if $e= 1$, we find ${\mathcal{H}}_k^k (X,{{\bf 0}}) = {\mathcal{H}}_k(X,{{\bf 0}}) = H_k(X\setminus {{\bf 0}})$ for all $k$.
We check easily that any non-trivial fc-homology class can be realized as Metric homology class [@BiBr1; @BiBr2] and as Vanishing homology class [@Va3].
Simple tangent directions {#Section:STDaBT}
=========================
The aim of this section is multiple. First we characterize the isomorphism equivalence class of a cone over a topological definable submanifold. This will allow to show in Section \[Section:MR\] that cones do not have fast contracting homology (Theorem \[thm:FF-polyhedra\] and Corollary \[cor:cone-no-hom\]). Second, this description result will allow to define the non-simple tangent direction locus, which is precisely the locus of the tangent link over which the fc-cycles will collapse and in a neighborhood of which the thin zone we are looking for is structured.
Let $(X,{{\bf 0}})$ be a closed definable subset of some Euclidean space with connected link at the origin ${{\bf 0}}$.
[**Notation:**]{} The set of points of the tangent link ${{\bf S}_{{\bf 0}}}X$ at which the germ of ${{\bf S}_{{\bf 0}}}X$ is the germ of topological submanifold of maximal dimension is denoted by ${{\bf S}_{{\bf 0}}}^* X$. Note that ${{\bf S}_{{\bf 0}}}^* X$ is open in the tangent link, definable and of maximal dimension.
We begin with the next result in order to have constantly present in mind that the class of definable homeomorphisms extending as isomorphisms we are using contains the definable outer-bi-Lipschitz homeomorphism.
\[prop:bilip-iso\] Let $(X,{{\bf 0}})$ be the germ of a closed definable subset with connected link at the origin ${{\bf 0}}$. Let $\Phi: (X,{{\bf 0}}) \to (Y,{{\bf 0}})$ be a definable bi-Lipschitz homeomorphism for the outer metrics. The homeomorphism $\Phi$ is an isomorphism.
In their beautiful paper, Bernig & Lytchak, by taking the derivative in $t=0$ of a definable curve $\gamma$ of $(X,{{\bf 0}})$ and its image $\Phi\circ \gamma$ of $(Y,{{\bf 0}})$ both starting at the origin, define a definable homogeneous mapping $D_{{\bf 0}}\Phi :T_{{\bf 0}}X \to T_{{\bf 0}}Y$, which also happens to be a bi-Lipschitz homogeneous homeomorphism [@BeLy]. Namely, given a continuous definable curve $\gamma:[0,{{\varepsilon}}[ \to X$ starting at ${{\bf 0}}$ with tangent vector ${{\bf u}}$ at $t=0$, the image vector $D_0\Phi \cdot {{\bf u}}$ is defined simply as the limit $\lim_{t\to0} \frac{{{\rm d}}}{{{\rm d}}t} \Phi\circ \gamma (t)$. The homogeneity and the continuity of the mapping $D_{{\bf 0}}\Phi$ induce a definable homeomorphism ${{\bf S}_{{\bf 0}}}X \to {{\bf S}_{{\bf 0}}}Y$ defined by the normalization of $D_{{\bf 0}}\Phi$
$\nu_{{\bf 0}}\Phi:{{\bf u}}\to \displaystyle{\frac{D_{{\bf 0}}\Phi\cdot{{\bf u}}}{|D_{{\bf 0}}\Phi\cdot{{\bf u}}|}}$
Thus when a sequence $(x_k)_k$ of points of $X$ goes to the origin such that the associated sequence of oriented directions $(\nu(x_k))_k$ goes to ${{\bf u}}$, we can write
$\Phi (x_k) = |x_k| D_{{\bf 0}}\Phi\cdot {{\bf u}}+ o(|x_k|)$.
Assuming that the subsets $X$ and $Y$ are embedded respectively in ${{{\mathbb{R}}}^n}$ and in ${{\mathbb{R}}}^p$, we see that the definable mapping $\beta_p^{-1}\circ\Phi\circ\beta_n$ when restricted to open strict transform $X'\setminus{{\partial}}X'$ is defined as $$\label{eq:lift-phi}
\vspace{4pt}
X'\setminus {{\partial}}X' \ni (r,{{\bf u}}) \to \left( \frac{\Phi(r{{\bf u}})}{|\Phi(r{{\bf u}})|},|\Phi(r{{\bf u}})|\right) \in Y'\setminus {{\partial}}Y'.$$ We just have to check that this mapping extends continuously as the homeomorphism $\nu_{{\bf 0}}\Phi\times0:{{\partial}}X' \to {{\partial}}Y'$ defined as $({{\bf u}},0) \to (\nu_{{\bf 0}}\Phi ({{\bf u}}),0)$ which is immediate from the definition of the mappings $D_{{\bf 0}}\Phi$ and $\nu_{{\bf 0}}\Phi$.
1\) If ${{{\bf S}}^{n-1}}\times {{\mathbb{R}}}_+$ is equipped with the outer metric from the Euclidean metric on ${{\mathbb{R}}}^n\times{{\mathbb{R}}}_+$, then the definable bi-Lipschitz homeomorphism of Proposition \[prop:bilip-iso\] extends as a definable bi-Lipschitz homeomorphism $(X',{{\partial}}X') \to (Y',{{\partial}}Y')$ for the outer metrics.\
2) Proposition \[prop:bilip-iso\] is very specific to isomorphisms induced by an outer definable bi-Lipschitz homeomorphism between closed definable germs. Isomorphisms usually do not even give a well defined mapping between tangent cones as seen in Example 3) and 4) of Section \[Section:FCCaFCH\], even though by definition they induce a homeomorphism between the tangent links.\
3) Any definable bi-Lipschitz homeomorphism $\Phi: (X,{{\rm d_{\rm inner}}})\to (X_0,{{\rm d_{\rm inner}}})$, where $X$ is a normally embedded model of closed definable isolated singularity $X_0$ with connected link, induces a surjective definable homogeneous finite-to-one mapping $D_{{\bf 0}}\Phi:T_{{\bf 0}}X \to T_{{\bf 0}}X_0$ (see [@BeLy]). Thus the mapping $\Phi$ extends as a morphism.
If a normally embedded model $X$ of a closed definable isolated singularity $X_0$ with connected link at the origin ${{\bf 0}}$ is isomorphic to another definable germ $X_1$, the isomorphism induces a definable homeomorphism $X_1 \to X_0$ which extends as a morphism and thus induces a surjective continuous definable finite-to-one mapping ${{\bf S}_{{\bf 0}}}X_1 \to {{\bf S}_{{\bf 0}}}X_0$. These facts and Proposition \[prop:bilip-iso\] justify the introduction of the following - now well defined - notion
\[def:SE-mod\] Let $X$ be a normally embedded model of a closed definable isolated singularity $X_0$ and with connected link at ${{\bf 0}}$.
i\) A subset $X_1$ of some Euclidean space is called *a simply embedded model of $X_0$ at the origin *if it is isomorphic to any normally embedded model of the singularity $X_0$.**
ii\) A definable homeomorphism $X_1\to X_0$ is a *simple embedding morphism *if it is a morphism obtained as the composition of a bi-Lipschitz mapping $(X,{{\rm d_{\rm inner}}}) \to (X_0,{{\rm d_{\rm inner}}})$ from a normally embedded model $X$ of $X_0$ and of an isomorphism $X_1 \to X$.**
A very pleasant property of simply embedded models of an isolated singularity is the following
\[prop:SE-iso-NE\] Any normally embedded model of a simply embedded model of a definable closed isolated singularity with connected link at the origin is isomorphic to the simply embedded model itself.
Let $X_0$ be a closed definable isolated singularity with connected link at the origin. Let $X_1$ be a simply embedded model of $X_0$ and $X$ be a normally embedded model of $X_0$ isomorphic to $X_1$. Let $Y$ be a normally embedded model of $X_1$ and let $\Phi:(Y,{{\rm d_{\rm inner}}})\to (X_1,{{\rm d_{\rm inner}}})$ be a definable bi-Lipschitz homeomorphism. It induces a surjective continuous definable finite-to-one mapping $\phi:{{\bf S}_{{\bf 0}}}Y \to {{\bf S}_{{\bf 0}}}X_1$. Let $h:X_1\to X$ be the isomorphism and let write $h'$ for the homeomorphism $X_1' \to X'$. We recall that the composed mapping $\Psi:= h\circ\Phi$ is a definable homeomorphism and extends as a morphism $\Psi'$. In particular the mapping $\psi'$ induced by the extension $\Phi'$ between the boundaries of the strict transforms of $Y$ and $X$ respectively, namely $\psi' := h'\circ(\phi\times 0) :{{\partial}}Y' \to {{\partial}}X'$, is finite-to-one and surjective. Let $x'$ be a boundary point of $X'$ and let $y_1',y_2'$ be two pre-images of $x'$ by $\psi'$. Let ${\mathcal{V}}$ be a small neighborhood of $x'$ in $X'$ and let ${\mathcal{U}}$ be its pre-image by $\Psi'$. Since the subset $Y'$ is contained in some product ${{\bf S}}^{p-1}\times {{\mathbb{R}}}_+$ we equip it with the outer metric of this product. Let $C_1$ and $C_2$ be two continuous definable curves lying in ${\mathcal{U}}$ with respective starting point $y_1'$ and $y_2'$. We can assume these curve are parametrized as $t\to c_i (t)$ by the height function (projection onto the component ${{\mathbb{R}}}_+$ of the product). Let $M(C_1,C_2)$ be the set of points $z'$ of the neighborhood ${\mathcal{U}}$ such that $z'$ is equidistant to the points of $C_1$ and $C_2$ with the same height as $z'$. This subset is definable, not empty and thus contains a continuous definable curve $C$ such that its starting point lies in the boundary ${\mathcal{U}}\cap {{\partial}}Y'$ and is equidistant to $y_1'$ and $y_2'$. The image $\Psi'(C)$ of $C$ is contained in the open subset ${\mathcal{V}}$ and starts at a point of the boundary ${\mathcal{V}}\cap {{\partial}}X'$. Let $(\delta_k)_k$ be a sequence of positive real numbers decreasing to $0$. Let ${\mathcal{V}}_k$ be the open subset of points of $X'$ with distance to $z'$ is strictly less than $\delta_k$. Let ${\mathcal{U}}_k$ be the pre-image $(\Psi')^{-1}({\mathcal{V}}_k)$. We construct a sequence of definable curves $(C_k)_{k >0}$ of the strict transform $Y'$ with starting point $y_k$ in the boundary ${{\partial}}Y'\cap{\mathcal{U}}_k$, equidistant from $y_1'$ and $y_2'$. We can assume that the sequence $(y_k)_k$ converges to $y'$. Since the continuous definable curve $\Psi'(C_k)$ starts at a point $z_k$ of the boundary ${{\partial}}X'\cap {\mathcal{U}}_k$, the sequence $(z_k)_k$ converges to $z'$ as $k$ goes to $\infty$. Thus $y'$ is a pre-image of $z'$ equidistant from $y_1'$ and $y_2'$. Unless the boundary mapping $\psi'$ is injective, we can construct as finitely many pre-images of $x'$ as we want. Thus $\psi'$ is a homeomorphism, and consequently $X_1$ is isomorphic to any of its normally embedded model.
Any normally embedded model of a closed definable isolated singularity is simply embedded. The converse is not true as seen in the following example in ${{\mathbb{R}}}^3$
$X = {{\rm clos}}(\{x^2+y^2 = z^2, z>0\})$ and $X_1 := {{\rm clos}}(\{y^2z^2 + x^4 = zx^3, z> 0\})$.
These cones are clearly isomorphic but $X_1$ is not normally embedded, since the compact plane curve $\{y^2+x^4=x^3\}$, basis of the cone $X_1$, is not normally embedded. The following result although obvious makes also sense of the notion of isomorphism we have introduced in Section \[Section:FCCaFCH\].
\[prop:SE-universal\] 1) Any two simply embedded models of a same closed definable isolated singularity and with connected link are isomorphic.\
2) If for $i=1,2$, the mappings $\Psi_i:X_i \to X_0$ are simple embedding morphisms of a closed definable isolated singularity $X_0$ of some Euclidean space and with connected link, then the mapping $\Psi_2^{-1}\circ \Psi_1:X_1\to X_2$ is an isomorphism.
Each model is isomorphic to any normally embedded model of the singularity $X_0$. Since the composition of two isomorphisms is again an isomorphism, we get the result. The second point just comes from the definition of a simple embedding morphism.
[**Important note:**]{} from now on we will only deal with definable closed isolated singularities [**thick**]{} at the origin and with connected link.
\[prop:simple-dense\] Let $X$ be a normally embedded model of a closed definable isolated singularity thick at the origin and with connected link. There exists a smallest closed definable subset $\Sigma$ of the tangent link ${{\bf S}_{{\bf 0}}}X$ of codimension at least one such that for each connected component $Z$ of the complement ${{\bf S}_{{\bf 0}}}X \setminus\Sigma$, the topological type of the definable family of germs $((X',z))_{\{z\in Z\times0\}}$ is constant along the component $Z$ and the germ at any point $(z,0)$ of $Z\times 0$ of the pair $(X',{{\partial}}X')$ is the germ of a topological submanifold with boundary.
The classical theorems of topological equisingularity ensures that the definable family of germs $((X',x'))_{\{x'\in{{\partial}}X'\}}$ has finitely many topological types, see [@Har; @vdD; @Sh; @Co]. In particular the compact parameter space ${{\partial}}X'$ is partitioned into finitely many definable pieces along which the topological type of the germs $(X',x')$ is constant (see [@Har; @vdD; @Sh; @Co]). This partitions the tangent link ${{\bf S}_{{\bf 0}}}X$ into finitely many connected definable topological submanifolds $(Z_i)_{i\in I}$ such that along each piece $Z_i$ the topological type of the family of germs $(X',x')_{\{x'\in Z_i\times 0\}}$ is constant. We call the elements of this partition strata. Let $Z$ be a stratum of maximal dimension, then it is open in then tangent link ${{\bf S}_{{\bf 0}}}X$. Since a definable Whitney stratification of the pair $(X',{{\partial}}X')$ exists [@vdDMi; @vdD] and since the subset $X$ is normally embedded, the topological type of the strict transform $X'$ at any point of the stratum $Z\times 0$ is that of a topological submanifold with boundary. We take $\Sigma$ to be the complement of the union of the strata of maximal dimension along which the topological type is constant.
Since a simply embedded singularity is isomorphic to a normally embedded isolated singularity the strict transforms under the respective spherical blowings-up are homeomorphic, thus the topological nature of the germ of the strict transform at a point of the boundary is transported by the extension homeomorphism. We can now introduce the following
\[def:simple-dir\] Let $X_1$ be a simply embedded model of a definable closed isolated singularity $X_0$ thick at the origin and with connected link.
i\) A point of the boundary ${{\partial}}X_1'$ of the strict transform $X_1'$ at which the germ of $X_1'$ is a topological submanifold with boundary is called a *simple point. *We denote by ${{\rm smp}}({{\partial}}X_1')$ the set of simple points of ${{\partial}}X'$ and by ${{\rm ns}}({{\partial}}X_1')$ its complement in ${{\partial}}X_1'$.**
ii\) A tangent direction corresponding to a simple point of the boundary ${{\partial}}X_1'$ is called a *simple tangent direction. *Let ${{\bf S}_{{\bf 0}}^{\rm smp}}X_1$ be the subset of *simple tangent direction of $X_1$ at the origin. *The complement set of the simple tangent directions at the origin in the tangent link, is the *locus of non-simple tangent directions *and is denoted ${{\bf S}_{{\bf 0}}^{\rm ns}}X_1$.******
As already explained isomorphic simply embedded models will have definably homeomorphic simple tangent direction loci.
Since the boundary ${{\partial}}X_1' = {{\bf S}_{{\bf 0}}}X_1 \times 0$ is an embedding of the tangent link ${{\bf S}_{{\bf 0}}}X_1$, we will only work with the tangent link. Note that we obviously find that ${{\rm smp}}({{\partial}}X_1') = {{\bf S}_{{\bf 0}}^{\rm smp}}X_1 \times 0$ and ${{\rm ns}}({{\partial}}X_1') =
{{\bf S}_{{\bf 0}}^{\rm ns}}X_1 \times 0$.
We will see below and in the next two sections that the locus of non-simple tangent directions does carry a lot of information about the fast contracting homology classes. As our first step towards this goal we recall Brown Theorem about the existence of collar neighborhoods in metric spaces, since it will be very useful to define the notion of thick zone. A subset $B$ of a topological space $Y$ is *collared *if there exists a homeomorphism $h:B\times [0,1[ \to {\mathcal{U}}$ onto a neighborhood ${\mathcal{U}}$ of $B$ such that $h(b,0) = b$ for each point $b$ of $B$. A subset $B$ of the topological space $Y$ is *locally collared *if it can be covered by a collection of open subsets $B_i$ of $B$ such that each $B_i$ is collared. The main result of Brown [@Br] is****
\[thm:Brown\] A locally collared subset of a metric space is collared.
Reading the simple and elegant proof of this theorem, we observe that this result can be achieved definably if we start with definable data.
Brown Theorem provides a pleasant property of simply embedded models of a given thick isolated singularity whose tangent link consists only of simple tangent directions:
\[lem:loc-conic-homeo-cone\] Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin and with connected link ${{\bf 0}}$. If the tangent link ${{\bf S}_{{\bf 0}}}X_1$ is a topological submanifold, then it consists only of simple tangent directions if and only if the subset $X_1$ is isomorphic to its tangent cone $T_{{\bf 0}}X_1$.
This simple result is the basis of our main results of Sections \[Section:MR\] and \[Section:TaToTIS\].
If $X_1$ isomorphic to its tangent cone, then its tangent link consists only of simple points.
We equip the strict transform $X_1'$ and its boundary ${{\partial}}X_1'$ with the outer metric of the metric space ${{{\bf S}}^{n-1}}\times {{\mathbb{R}}}_+$ coming from its inclusion in ${{{\mathbb{R}}}^n}\times{{\mathbb{R}}}_+$.
Brown Theorem implies the existence of a positive real number ${{\varepsilon}}$ and of a definable homeomorphism $\Phi:{{\bf S}_{{\bf 0}}}X_1 \times [0,{{\varepsilon}}[ \to {\mathcal{V}}$ where ${\mathcal{V}}$ is an open neighborhood of ${{\partial}}X_1'$ in $X_1'$ and is also definable. Moreover we also know that $\Phi({{\bf u}},0) = ({{\bf u}},0)$. Composing with the blowing-down mappings at the target (of $\Phi$) and at the source (of $\Phi$) we obtain a definable homeomorphism
$\beta\circ\Phi\circ\beta^{-1}:T_{{\bf 0}}X_1 \setminus {{\bf 0}}\to X_1 \setminus{{\bf 0}}$,
which extends continuously and definably at ${{\bf 0}}$.
[**Notation:**]{} The *${{\varepsilon}}$-cone *of a subset $L$ of the unit sphere ${{{\bf S}}^{n-1}}$ is the subset of ${{{\mathbb{R}}}^n}$ defined as**
${{\rm Cone}}(L,{{\varepsilon}}) := {{\rm clos}}(\{x \in {{{\mathbb{R}}}^n}\setminus{{\bf 0}}: {{\rm dist}}(\nu (x),L) < {{\varepsilon}}\})$.
With the same principles as those of the second part of the proof of Lemma \[lem:loc-conic-homeo-cone\], we deduce the following
\[cor:isom-tgt-cone\] Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin and with connected link ${{\bf 0}}$. For a positive real number ${{\varepsilon}}$ small enough, each connected component of the complement $X_1 \setminus {{\rm Cone}}({{\bf S}_{{\bf 0}}^{\rm ns}}X_1,{{\varepsilon}})$ is isomorphic to its tangent cone.
Collapsing homology locus of thick isolated singularities {#Section:MR}
=========================================================
We present here the first version of the main result of this paper: Theorem \[thm:main\] states that the fast contracting homology of a simply embedded model $X_1$ of a closed definable thick isolated singularity $X_0$ with connected link can be realized in a thin zone of $X_1$ organized around the non-simple tangent directions. Prior to getting there, we prove Theorem \[thm:FF-polyhedra\] which guarantees that cones do not have fast contracting homology, fact absolutely essential to show Theorem \[thm:main\]
The next result is a polyhedral version of Federer-Flemming Theorem on the triviality of cycles of small volume in a smooth compact Riemannian manifold [@FF]. By ${{\rm vol}}_k$ is meant the $k$-dimensional volume taken with the outer metric.
\[thm:FF-polyhedra\] Let $\Gamma$ be a compact polyhedron defined as the union of the simplices of a finite set $T$. Let $\xi$ be a definable $k$-cycle in $\Gamma$ such that for any positive ${{\varepsilon}}$ small enough there exists a homologous definable $k$-cycle $\xi_{{{\varepsilon}}}$ in $\Gamma$ whose $k$-dimensional volume ${{\rm vol}}_k({{\rm supp}}(\xi_{{{\varepsilon}}}))$ is smaller than ${{\varepsilon}}$. Then, $\xi$ bounds a definable $(k+1)$-chain of $\Gamma$.
The idea of the proof is the similar to that of the main result of [@BiBr2]. We are going to show that the cycle $\xi$ is homologous to a definable $k$-cycle living in the $(k-1)$-skeleton of the polyhedron $\Gamma$. The dimension of $\Gamma$ is $d$.
First, observe there exists a positive real number $L$ such that: For any simplex $\sigma$ of $T$ and any interior point $x_0$ of $\sigma$, the mapping $$\Phi_{\sigma,x_0}:\sigma\setminus\{x_0\}\rightarrow\partial\sigma$$ defined as the intersection point of the half-line ${{\mathbb{R}}}_+ (x-x_0)$ with the boundary $\partial\sigma$, is a $L$-Lipschitz mapping. The existence of such uniform constant $L$ is due to the fact that the family $(\Phi_{\sigma,x})_{x\in\sigma,\sigma\in T}$ is a semi-algebraic family of piecewise linear mappings over the same compact semi-algebraic source.
Let ${{\varepsilon}}$ be positive and strictly smaller than $$\frac{1}{L^{(d-k)k}}\inf\{{{\rm vol}}_{k}(\sigma) \ : \ \sigma \mbox{ is a $k$-simplex of } T\}.$$ Let $\xi_{{{\varepsilon}}}$ be a definable $k$-cycle in $\Gamma$ homologous to $\xi$ and whose volume ${{\rm vol}}_k({{\rm supp}}(\xi_{{{\varepsilon}}}))$ is smaller than ${{\varepsilon}}$. Let $\sigma$ be a $d$-simplex of $T$. Since the interior of the simplex $\sigma$ is not contained in the support of the cycle $\xi_{{{\varepsilon}}}$, there exists a point $x_0$ in the interior of $\sigma$ such that the support of $\xi_{{\varepsilon}}$ is contained in $\sigma \setminus\{x_0\}$. So, we find $${{\rm vol}}_k(\Phi_{\sigma,x_0}(\xi_{{\varepsilon}}\cap\sigma))\leq L^k {{\rm vol}}_k({{\rm supp}}(\xi_{{\varepsilon}})\cap\sigma).$$ Hence, the $k$-cycle $\xi_{{{\varepsilon}}}$ is homologous to a definable $k$-cycle $\xi_{d-1}$ which lives on the $d-1$-skeleton of the polyhedron $\Gamma$ and whose volume satisfies $${{\rm vol}}_k({{\rm supp}}(\xi_{d-1}))\leq L^k {{\rm vol}}_k({{\rm supp}}(\xi_{{{\varepsilon}}})).$$ If $d-1=k-1$ the proof ends here. If not, since the volume of the support of the cycle $\xi_{d-1}$ is small, we are dealing with a $(d-1)$-simplex $\sigma$ of $T$, whose interior is not contained in the support of $\xi_{d-1}$, hence there exists a point $x_1$ in the interior of $\sigma$ such that the support of $\xi_{d-1}$ is contained in $\sigma\setminus\{x_1\}$. We deduce the following inequality $${{\rm vol}}_k(\Phi_{\sigma,x_1}(\xi_{d-1}\cap\sigma))\leq L^k {{\rm vol}}_k({{\rm supp}}(\xi_{d-1})\cap\sigma).$$ Hence, the cycle $\xi_{d-1}$ is homologous to a definable $k$-cycle $\xi_{d-2}$ which lies in the ($d-2$)-skeleton of $\Gamma$ and satisfies $${{\rm vol}}_k ({{\rm supp}}(\xi_{d-2}))\leq L^k {{\rm vol}}_k({{\rm supp}}(\xi_{d-1}))\leq L^{2k} {{\rm vol}}_k({{\rm supp}}(\xi_{{{\varepsilon}}})).$$ Iterating this process, we conclude that the cycle $\xi$ is homologous to a definable $k$-cycle $\xi_k$ living in the $k$-skeleton of the polyhedron $\Gamma$ and satisfies the following inequality $${{\rm vol}}_k({{\rm supp}}(\xi_k))\leq L^{(d-k)k} {{\rm vol}}_k({{\rm supp}}(\xi_{{{\varepsilon}}})).$$ Since $$L^{(d-k)k} {{\rm vol}}_k({{\rm supp}}(\xi_{{{\varepsilon}}}))<\inf\{{{\rm vol}}_k(\sigma) \ : \ \sigma\in T \ \mbox{is a k-simplex}\},$$ given a $k$-simplex $\sigma$ of $T$, its interior is not contained in the support of the cycle $\xi_k$, thus there exists $x_k$ a point in the interior of $\sigma$ such that the support of $\xi_k$ is contained in $\sigma\setminus\{x_k\}$. Considering the projection mapping $$\Phi_{\sigma,x_k}:\sigma\setminus\{x_k\}\rightarrow\partial\sigma$$ where $\sigma$ is a $k$-simplex of $T$ and $x_k$ is a point in the interior of $\sigma$ such that the support of the cycle $\xi_k$ is contained in $\sigma\setminus\{x_k\}$. We thus obtain a definable $k$-cycle $\xi_{k-1}$ homologous to $\xi$ and contained in the ($k-1$)-skeleton of the polyhedron $\Gamma$. Necessarily $\xi$ was trivial.
We deduce the following two corollaries.
\[cor:FF-defin\] Let $\Gamma$ be a compact definable set of some Euclidean space. Let $\xi$ be a definable $k$-dimensional cycle in $\Gamma$ such that for any positive ${{\varepsilon}}$ there exists a definable $k$-dimensional cycle $\xi_{{{\varepsilon}}}$ in the set $\Gamma$ homologous to $\xi$ and whose volume ${{\rm vol}}_{k}({{\rm supp}}(\xi_{{{\varepsilon}}}))$ is smaller than ${{\varepsilon}}$. Then the cycle $\xi$ bounds a definable $(k+1)$-chain in $\Gamma$.
Let $Z$ be a closed definable subset of ${{{\mathbb{R}}}^n}$ containing the origin ${{\bf 0}}$ and of dimension $d$ at the origin. Let $Z({{\varepsilon}})$ be the intersection $Z\cap {{\bf B}}^n ({{\bf 0}},{{\varepsilon}})$. We recall that the normalized $d$-volume ${{\varepsilon}}^{-d} {{\rm vol}}_d (Z({{\varepsilon}}))$ tends to $0$ as the radius ${{\varepsilon}}$ goes to $0$ if and only if the tangent cone $T_{{\bf 0}}Z$ has dimension strictly less than $d$.
\[cor:cone-no-hom\] A germ of a closed definable subset of some Euclidean space isomorphic to a cone has trivial fast contracting homology.
We will use Theorem \[thm:FF-polyhedra\]. We can assume $Z$ is contained in ${{{\mathbb{R}}}^n}$ and its tangent link ${{\bf S}_{{\bf 0}}}Z$ is connected. We work within the intersection of $Z$ with a closed ambient Euclidean ball ${{\rm clos}}({{\bf B}}^n({{\bf 0}},{{\varepsilon}}))$. Let $\xi$ be a fast contracting $k$-cycle and let $\eta$ be the chain it bounds in $Z$. For a positive small real number $t$ we define $\xi_t$ as the $k$-cycle whose support is the intersection ${{\rm supp}}(\eta) \cap {{\bf S}}_t^{n-1}$, where ${{\bf S}}_t^{n-1}$ is the $(n-1)$-dimensional Euclidean sphere centered at ${{\bf 0}}$ of radius $t$, and the coefficients are those from $\eta$ when the intersection is not empty. Thus $\xi_t$ is a cycle homologous to $\xi$. Since the germ $Z$ is the homeomorphic image of a cone by an isomorphism, the support ${{\rm supp}}(\xi_t)$ is contained in the scaled tangent link $t({{\bf S}_{{\bf 0}}}Z)\subset {{\bf S}}_t^{n-1}$. Since the support of $\eta$ is thin at the origin, the Hausdorff limit at $t=0$ of the rescaled support $\frac{1}{t} {{\rm supp}}(\xi_t)$ tends to a definable subset of dimension at most $k-1$ contained in the tangent link ${{\bf S}_{{\bf 0}}}Z$. We apply Corollary \[cor:FF-defin\] to the tangent link ${{\bf S}_{{\bf 0}}}Z$ and the family of cycles $(\frac{1}{t} \xi_t)_{\{0<t\ll 1\}}$ and get the result.
Note that Corollary \[cor:cone-no-hom\] above gives a very quick proof to the main result [@BFGoS Theorem 2.6] of the paper [@BFGoS].
We can now state the main result of the paper. We recall that ${{\bf S}_{{\bf 0}}^{\rm ns}}X_1$ is the locus of non-simple tangent directions (see Definition \[def:simple-dir\]).
\[thm:main\] Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin and with connected link. For any positive real number ${{\varepsilon}}$ small enough, the inclusion mapping induces the following exact sequence $$\label{eq:main-thm}
{\mathcal{H}}_\bullet (X_1\cap {{\rm Cone}}({{\bf S}_{{\bf 0}}^{\rm ns}}X_1, {{\varepsilon}}),{{\bf 0}}) \to {\mathcal{H}}_\bullet (X_1,{{\bf 0}}) \to 0.
\vspace{4pt}$$
Before getting into the proof let us mention the straightforward following:
\[cor:main-cor\] For any degree $k=1,\ldots,\dim X_1$ and for any dimension drop $\delta = 1,\ldots,k$, for any positive real number ${{\varepsilon}}$ small enough, the inclusion mapping induces the following exact sequence $$\label{eq:main-1}
{\mathcal{H}}_k^\delta (X_1\cap {{\rm Cone}}({{\bf S}_{{\bf 0}}^{\rm ns}}X_1, {{\varepsilon}}),{{\bf 0}}) \to {\mathcal{H}}_k^\delta (X_1,{{\bf 0}}) \to 0.
\vspace{4pt}$$
Theorem \[thm:main\] is a consequence of the next result.
\[lem:homology-in-thin-non-simply\] Let $\xi$ be a $k$-dimensional definable fc-cycle of $X_1$. Then for any ${{\varepsilon}}$ small enough, the homology class of the cycle $\xi$ can be represented by a fast contracting definable cycle supported in the link of ${{\rm Cone}}({{\bf S}_{{\bf 0}}^{\rm ns}}X_1,{{\varepsilon}})$.
For ${{\varepsilon}}$ positive and small enough the topological type of the germ at the origin of the complement $X_1 \setminus {{\rm Cone}}({{\bf S}_{{\bf 0}}^{\rm ns}}X_1, {{\varepsilon}})$ is constant. It consists in finitely many closed connected components each of which is isomorphic to its tangent cone by Corollary \[cor:isom-tgt-cone\].
Let ${\mathcal{U}}_{{\varepsilon}}:= {{\rm Cone}}({{\bf S}_{{\bf 0}}^{\rm ns}}X_1,{{\varepsilon}})$ and ${{\partial}}{\mathcal{U}}_{{\varepsilon}}:= {{\rm clos}}({\mathcal{U}}_{{\varepsilon}})\setminus{\mathcal{U}}_{{\varepsilon}}$. Let $\xi$ be a definable fc-cycle of dimension $k$ bounding a definable $(k+1)$-chain $\eta$ whose support is thin and contains the origin. There exists a definable triangulation of the subset $X_1$ such that $\eta = \sum_{\sigma\in K} a_\sigma \sigma$ and where each of the following intersections ${{\rm supp}}(\eta)\cap(X_1\setminus{\mathcal{U}}_{{\varepsilon}})$, ${{\rm supp}}(\eta)\cap{{\partial}}{\mathcal{U}}_{{\varepsilon}}$ and ${{\rm supp}}(\eta)\cap{{\rm clos}}({\mathcal{U}}_{{\varepsilon}})$ is a finite union of simplices.
We observe that the support of each simplex $\sigma$ of $K$ of dimension $k+1$ is thin at the origin.
Let $K^{{\partial}}$ be the set of simplices of the chain $\eta$ contained in ${{\partial}}{\mathcal{U}}_{{\varepsilon}}$, $K^+$ be the set of simplices of $\eta$ contained in $X_1\setminus {\mathcal{U}}_{{\varepsilon}}$ but not in ${{\partial}}{\mathcal{U}}_{{\varepsilon}}$, and $K^-$ be the set of of simplices of $\eta$ contained in ${{\rm clos}}({\mathcal{U}}_{{\varepsilon}})$ but not in ${{\partial}}{\mathcal{U}}_{{\varepsilon}}$. In particular the chain $\eta$ is presented as the following sum
$\eta = \eta^+ + \eta^- + \eta^\partial,$
where $\eta^* = \sum_{\sigma\in K^*} a_\sigma \sigma$ for $*\in \{-,+,{{\partial}}\}$. We then check that the cycle $\xi$ decomposes as
$\xi = \partial \eta = \xi^+ + \xi^\partial + \xi^-$
with $\xi^* ={{\partial}}\eta^*$ where $*\in \{-,+,{{\partial}}\}$. In particular $\xi^+$, $\xi^\partial$ and $\xi^-$ are fc-cycles at ${{\bf 0}}$. By Corollary \[cor:isom-tgt-cone\] and Corollary \[cor:cone-no-hom\] we deduce that the cycles $\xi^+$ and $\xi^\partial$ are trivial and thus $\xi$ is homologous to $\xi^-$ which is contained in ${\mathcal{U}}_{{{\varepsilon}}'} = {{\rm Cone}}({{\bf S}_{{\bf 0}}^{\rm ns}}X,{{\varepsilon}}')$ for any ${{\varepsilon}}'>{{\varepsilon}}$.
Topological thick and thin decomposition theorem for thick isolated singularities {#Section:TaToTIS}
=================================================================================
As a consequence of the results of Section \[Section:MR\], we will produce a systematic decomposition of the isolated singularity we are working with into a *topological thick part *and a *topological thin part. *As already proved in Section \[Section:MR\], any fc-homology class can be realized in the thin zone.****
As in [@BNP], this decomposition is not unique stricto sensu. It is a one-parameter definable family of open subsets such that the Hausdorff limit at $0$ of the family of tangent cones at the origin is the cone over the locus of the non-simple tangent directions, thus is thin (see Definition \[def:thin-zone\]). We will see that it is consistent under automorphisms and its topology does not depend on the choice of any element of the family.\
In order to provide an appropriate definition of the thin zone, we will review the notion of a definable neighborhood of a compact definable subset of some Euclidean space.
Let $Z$ be a compact definable subset of some Euclidean space.
\[def:def-nghb\] A definable neighborhood of the definable compact subset $Y$ of $Z$ is any neighborhood of $Y$ in $Z$ of the form $f^{-1}([0,{{\varepsilon}}[)$, where $f:Z\to{{\mathbb{R}}}_+$ is a non-negative continuous definable function vanishing exactly on $Y$ and ${{\varepsilon}}$ is a positive real number.
Since $Z$ is embedded in some Euclidean space, the outer distance function on $Z$ to any definable closed subset of $Z$ defines a (definable) family of definable neighborhoods of the given closed subset.
Let $Y$ be a definable compact subset of $Z$ and let ${\mathcal{U}}_{{\varepsilon}}:= f^{-1}([0,{{\varepsilon}}[)$ be a definable neighborhood of $Y$. Since the definable family $(f^{-1}([0,t[))_{t>0}$ admits only finitely many topological type [@vdD; @Co], we can always take ${{\varepsilon}}$ such that for any positive real number ${{\varepsilon}}' < {{\varepsilon}}$, the definable neighborhoods ${\mathcal{U}}_{{\varepsilon}}$ and ${\mathcal{U}}_{{{\varepsilon}}'}$ are (definably) homeomorphic.
The other important property of definable neighborhoods of definable compact subsets is that there are unique, up to a homeomorphism. More precisely
\[prop:defin-neighb\] Let ${\mathcal{U}}_{{\varepsilon}}:= f^{-1}([0,{{\varepsilon}}[)$ and ${\mathcal{V}}_\delta:= g^{-1}([0,\delta[)$ be two definable neighborhoods of $Y$ a definable compact subset of $Z$ for any positive real numbers ${{\varepsilon}},\delta$ small enough. Then ${\mathcal{U}}_{{\varepsilon}}$ and ${\mathcal{V}}_\delta$ are homeomorphic.
Before sketching the proof, let us say that we can have a much more precise statement, but what we need to define the notion of *thin zone *is just what is stated.**
The proof works very much like for the semi-algebraic case [@Dur]. The main new ingredient in this o-minimal setting is that we can definably stratify any continuous definable function as a $C^r$ function on each stratum, for any a-priori prescribed positive integer $r$, such that Thom’s condition $(a_f)$ is satisfied along each stratum [@Loi]. Assume that ${\mathcal{U}}_{{\varepsilon}}$ is contained in ${\mathcal{V}}_\delta$. Let $S := g^{-1}([0,\delta]) \setminus f^{-1}([0,{{\varepsilon}}[)$ and let us define, as in the proof of [@Dur Lemma 1.8], the following continuous definable function $\kappa:S \to [0,1]$
$\displaystyle{\kappa(x) := \frac{f(x) - {{\varepsilon}}}{(f(x) -{{\varepsilon}}) - (g(x) -\delta)}}$
The function $\kappa$ is non negative and vanishes only on $f^{-1}({{\varepsilon}})$. If ${{\varepsilon}}$ and $\delta$ are small enough, then the gradients of the restrictions of the functions of $f$ and $g$ to a given stratum do not point in opposite directions (see [@Dur Lemma 1.8]). Thus the gradient of the restriction of the function $\kappa$ along any stratum does not vanish. We can refine the definable stratification so that $\kappa$ satisfies Thom’s condition $(a_f)$. Using Thom First Isotopy Lemma we conclude that this function induces a locally trivial fibration over its image $[0,1]$. The trivialization is realized by the flow of a vector field, so that the level $f^{-1}({{\varepsilon}})$ is pushed continuously onto the level $g^{-1}(\delta)$. This provides a continuous flow $F: f^{-1}({{\varepsilon}})\times [0,1] \to S$ such that $F(x,0) = x$.
If ${{\varepsilon}}$ was taken small enough to start with, there is also a continuous flow $G : f^{-1}({{\varepsilon}})\times [0,{{\varepsilon}}[ \to
f^{-1} (]0,{{\varepsilon}}])$ such that $f(G(x,t)) = {{\varepsilon}}-t$ and $G(x,0) =x$, and moreover the definable neighborhood $f^{-1} ([0,{{\varepsilon}}[)$ is homeomorphic to the definable neighborhood $f^{-1} ([0,{{\varepsilon}}'[)$, for any positive real number real number ${{\varepsilon}}' <{{\varepsilon}}$.
We deduce from these two flows that the definable neighborhood $f^{-1} ([0,{{\varepsilon}}'[)$ is homeomorphic to the definable neighborhood $g^{-1}([0,\delta[) = S \cup f^{-1} ([0,{{\varepsilon}}[)$ for any positive ${{\varepsilon}}' <{{\varepsilon}}$.
We can now present the definition of the topological thin zone since Proposition \[prop:defin-neighb\] (when the definable neighborhood is chosen small enough) guarantees it is relevant:
\[def:thin-zone\] Let $X_1$ be a simply embedded model of a closed definable isolated singularity $X_0$ of some Euclidean space, thick at the origin and with connected link at the origin. A *topological thin zone *of $X_1$ is the image by the blowing-down mapping $\beta$ of a definable neighborhood in the strict transform $X_1'$ of the non-simple point locus ${{\rm ns}}({{\partial}}X_1') = {{\bf S}_{{\bf 0}}^{\rm ns}}X_1 \times 0$ of the boundary ${{\partial}}X_1'$ of $X_1'$. We will denote it by ${{\rm Thin}}(X_1)$. The complement of a topological thin zone is called a *topological thick zone. *****
The standard family of definable neighborhoods we have in mind is obtained from the restriction to the strict transform $X_1'$ of the Euclidean distance function coming from the inclusions $X_1' \subset {{{\bf S}}^{n-1}}\times {{\mathbb{R}}}_+ \subset{{{\mathbb{R}}}^n}\times {{\mathbb{R}}}_+$.\
Given an isomorphism $\Phi: X_1 \to X_2$ between two simply embedded models, in some Euclidean spaces ${{\mathbb{R}}}^p$ and ${{\mathbb{R}}}^q$ respectively, of a closed definable isolated singularities and thick at the origin, the interest of this definition is that a topological thin zone if mapped onto a topological thin zone. Indeed a topological thin zone of the simply embedded model $X_2$ is defined as $\beta_q(f_2^{-1}([0,{{\varepsilon}}[))$ for a continuous definable function $f_2:X_2'\to {{\mathbb{R}}}_+$ vanishing exactly on $({{\bf S}_{{\bf 0}}^{\rm ns}}X_2) \times 0\cap {{\partial}}X_2$. Defining a continuous definable function $f_1 :X_1' \to {{\mathbb{R}}}_+$ as $f_1 := f_2 \circ \Phi'$, we see that the homeomorphism $\Phi'$ maps the definable neighborhood $f_1^{-1}([0,{{\varepsilon}}[)$ onto the definable neighborhood $f_2^{-1}([0,{{\varepsilon}}[)$, so that $\Phi$ maps the topological thin zone $\beta_p(f_1^{-1}([0,{{\varepsilon}}[))$ onto the topological thin zone $\beta_q(f_2^{-1}([0,{{\varepsilon}}[))$. Thus the decomposition into topological thick zone/topologically thin zone is preserved by isomorphisms. It is as we were expected an invariant decomposition for our class of isomorphisms which, as we have repeatedly said, contains definable bi-Lipschitz homeomorphisms for the outer metrics.
We can now rephrase Theorem \[thm:main\] in terms of the topological thin zone:
\[thm:main-bis\] Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin and with connected link. The inclusion mapping induces the following exact sequence $$\label{eq:main-bis-thm}
{\mathcal{H}}_\bullet ({{\rm Thin}}(X_1),{{\bf 0}}) \to {\mathcal{H}}_\bullet (X_1,{{\bf 0}}) \to 0.
\vspace{4pt}$$
Its expected version with degrees and dimension drops is the following
\[cor:main-cor-bis\] For any degree $k=1,\ldots,\dim X_1,$ and for any dimension drop $\delta = 1,\ldots,k,$ the inclusion mapping induces the following exact sequence $$\label{eq:main-1-bis}
{\mathcal{H}}_k^\delta ({{\rm Thin}}(X_1),{{\bf 0}}) \to {\mathcal{H}}_k^\delta (X_1,{{\bf 0}}) \to 0.
\vspace{4pt}$$
Non-simply embedded thick isolated singularities {#Section:SETIS}
================================================
We are going to have here a short view on what is happening for closed definable isolated singularities of some Euclidean space, thick at the origin and with connected link which are not simply embedded.
Any closed definable isolated singularity $X_0$ thick at the origin with connected link admits a simply embedded model $X_1$ and thus exists a simple embedding morphism $\Phi: X_1\to X_0$. As a consequence of our definitions and of Bernig & Lytchak results [@BeLy], it provides a surjective continuous definable finite-to-one mapping ${{\partial}}X_1' \to {{\partial}}X_0'$. Let $\Phi_i:X_i\to X_0$, $i=1,2$, be two simple embedding morphisms of the closed definable isolated singularity $X_0$ thick at the origin with connected link. By definition of simply embedding morphism, the mapping $\Phi_2^{-1}\circ \Phi_1 : X_1 \to X_2$ is a definable homeomorphism which extends as an isomorphism $X_1' \to X_2'$. A topological thin zone of $X_1$ is then mapped isomorphically onto a topological thin zone of $X_2$. Thus we can introduce the following:
\[def:thinzone-nse\] The *topological thin zone *of $X_0$ is defined as the image of the topological thin zone of any simply embedded model of $X_0$ by the corresponding simple embedding morphism of the simply embedded model.**
As observed just before Definition \[def:thinzone-nse\], defining the topological thin zone this way is really relevant since it does not depend on the simply embedded model of $X_0$. The first consequence is the following
\[prop:homol-thin-same\] Let $\Phi:X_1\to X_0$ be a simply embedding morphism of a closed definable isolated singularity $X_0$ thick at the origin and with connected link. For any degree $k$ and any dimension drop $\delta =1,\ldots,k$, the homomorphism $\Phi_*$ induces an isomorphism ${\mathcal{H}}_k^\delta ({{\rm Thin}}(X_1),{{\bf 0}}) \to {\mathcal{H}}_k^\delta
({{\rm Thin}}(X_0),{{\bf 0}})$.
Since ${{\rm Thin}}(X_0) = \Phi ({{\rm Thin}}(X_1))$ injectivity is immediate. Since the mapping $\Phi$ extends as a morphism, any fc-cycle supported in the thin zone ${{\rm Thin}}(X_1)$ is mapped onto a fc-cycle supported in ${{\rm Thin}}(X_0)$. Since the homeomorphism $\Phi$ induces a surjective continuous definable finite-to-one mapping ${{\bf S}_{{\bf 0}}}X_1 \to {{\bf S}_{{\bf 0}}}X_0$, any fc-cycle supported in ${{\rm Thin}}(X_0)$ and of dimension drop larger than or equal to $\delta$ is mapped by $\Phi^{-1}$ into a fc-cycle supported in ${{\rm Thin}}(X_1)$ of dimension drop larger than or equal to $\delta$, thus we find the surjectivity.
A straightforward corollary of Proposition \[prop:homol-thin-same\] is
The inclusion mapping of the thin zone $\iota:{{\rm Thin}}(X_0)\to X_0$ gives the following exact sequences: $$\label{eq:main-2}
\iota_*:{\mathcal{H}}_k^\delta ({{\rm Thin}}(X_0),{{\bf 0}}) \to {\mathcal{H}}_k^\delta (X_0,{{\bf 0}}) \to 0
\vspace{4pt}$$ for any degree $k$ and any dimension drop $\delta\leq k$.
Let $\Phi:X_1\to X_0$ be a simply embedded morphism of the singularity $X_0$. We obtain a homomorphism $\Phi_*:{\mathcal{H}}_\bullet (X_1,{{\bf 0}})\to {\mathcal{H}}_\bullet (X_0,{{\bf 0}})$. Since the mapping induced on the tangent links ${{\bf S}_{{\bf 0}}}X_1 \to {{\bf S}_{{\bf 0}}}X_0$ is a surjective continuous definable finite-to-one mapping, any fc-cycle of $X_0$ is pulled back onto a cycle of $X_1\setminus {{\bf 0}}$ such that it is contracted by a chain of the simply embedded model $X_1$ whose tangent cone at the origin is of the exact same dimension as the chain contracting the fc-cycle of $X_0$, then the pre-image by the mapping $\Phi$ is a fc-cycle of $X_1$. Thus the homomorphism $\Phi_*:{\mathcal{H}}_k^\delta (X_1,{{\bf 0}}) \to {\mathcal{H}}_k^\delta (X_0,{{\bf 0}})$ is surjective for any degree $k$ and drop dimension $\delta =1,\ldots,k$. Proposition \[prop:homol-thin-same\] allows to conclude.
Examples and applications {#Section:examples}
=========================
This last section presents some examples of fc-homology classes of thick singularities. First we find out that the three known obstructions (so far) to the local inner metric conicalness of complex singularities, namely *fast loops, separating sets *and *choking horns, *are not of metric nature but of something slightly less rigid. Second they illustrate a same phenomenon: they represent fc-homology classes which may be non trivial in the topological thin zone. Proposition \[prop:sep-set\] characterizes completely the existence of separating sets of simply embedded models of closed definable thick isolated singularities with connected link. Last, when the tangent link is of pure dimension Theorem \[thm:sep-set-pure-dim\] asserts that separating sets represents some non trivial fc-homology classes of the topological thin zone but trivial in the link.****
We recall quickly the notions of *fast loops, separating sets and choking horns. *Although these notions were introduced for normally embedded models of semi-algebraic singularities, it is elementary to check that they are all preserved under isomorphisms since each can be defined in terms of properties on the link of some tangent cone.**
\[def:fast-loop-etc\] Let $X$ be a normally embedded model of a closed definable isolated singularity of some Euclidean space with connected link at the origin.
- A *fast loop *is a continuous definable mapping $\gamma: {{\bf S}}^1 \to X$ such that there exists a continuous definable mapping $h: [0,1]\times {{\bf S}}^1 \to X$ such that**
1. $h (0,{{\bf u}}) = {{\bf 0}}$ and $h (1,{{\bf u}}) = \gamma ({{\bf u}})$,
2. the image [*Im*]{}($h$) of the mapping $h$ is thin.
- A *choking horn in $X$ *is the image $H$ of a definable continuous mapping $\phi\colon [0,1]\times {{\bf S}}^p\rightarrow X$, for some positive integer $p$, such that**
1. For any point $(t,{{\bf u}})\in[0,1]\times {{\bf S}}^p$, its image $\phi(t,{{\bf u}})$ lies in the intersection $X_t:=X\cap{{\bf S}}_t^{n-1}$.
2. The tangent cone of the image of $\phi$ at ${{\bf 0}}$ is just a single real half-line.
3. For every $t$ small enough and any chain $\eta_t$ whose support is contained in the subset $X_t$ and such that the boundary of its support ${{\partial}}({{\rm supp}}(\eta)_t)$ coincides with the intersection $H_t:= H\cap {{\bf S}}_t^{n-1}$, there exists a positive real constant $K$ such that the diameter of the support of $\eta_t$ satisfies the inequality ${\rm diam}({{\rm supp}}(\eta)_t)\geq Kt$.
- A separating set $S$ of $X$ is a closed definable subset of $X$ of codimension one, thin at the origin, such that the germ of the complement $X\setminus S$ has at least two thick connected components.
We observe that the both notions of fast loop and choking horn represent a fc-contracting homology class. For isolated complex singularities the presence of either one is an obstruction to metric conicalness [@BF3; @BFN1; @BFN3; @Fe2; @BFGoS]. Note also that the paper [@BNP] considers some special fast loops, called *of the second kind, *which are just choking horns of dimension $p=1$. The main technical result of [@BFGoS] asserts that a subanalytic cone (and thus any subanalytically bi-Lipschitz homeomorphic image), cannot admit choking horns. Since the proof of this result [@BFGoS Theorem 2.6] relies on arguments about the diameters of some family of links, which can be reformulated in terms of dimension of Hausdorff limits of the linearly rescaled family, we observe that with the exact same arguments we actually have the following more general result (which could also deduce from Corollary \[cor:cone-no-hom\]):**
Let $X_1$ be a closed definable subset isomorphic to a cone over a topological submanifold. Then $X_1$ cannot admit a choking horn.
The presence of a separating set from the point of view of the fast contracting homology is slightly more complicated to deal with and it can only occur in thick isolated singularities. According to [@BFN3] the tangent cone at the origin of a separating set separates the tangent cone of the ambient subset. It is a property on the tangent links. Consequently, as we have already suggested, a separating set is a notion invariant by isomorphisms.
The next definition, as we will see, is key to characterize completely the notion of separating set.
\[def:cond-SC\] Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin and with connected link. Let us consider the following condition:\
*(SC): *The locus of non-simple tangent directions ${{\bf S}_{{\bf 0}}^{\rm ns}}X_1$ separates the tangent link ${{\bf S}_{{\bf 0}}}X_1$ and there exists a non empty connected component of the complement ${{\bf S}_{{\bf 0}}}X_1 \setminus {{\bf S}_{{\bf 0}}^{\rm ns}}X_1$ such that its closure intersects with ${{\bf S}_{{\bf 0}}^{\rm ns}}X_1$ in codimension larger than of equal to two.**
The main interest of condition (SC) in relation with separating sets lies in the following
\[lem:sep-set-1\] Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin and with connected link. If Condition $(SC)$ holds true, then $X_1$ admits a separating set.
Let ${\mathcal{U}}$ be a connected component of the simple tangent directions locus ${{\bf S}_{{\bf 0}}^{\rm smp}}X_1 = {{\bf S}_{{\bf 0}}}X_1 \setminus {{\bf S}_{{\bf 0}}^{\rm ns}}X_1$ whose closure intersects with ${{\bf S}_{{\bf 0}}^{\rm ns}}X_1$ in codimension at least two. Let $S$ be a connected component of this intersection. A horn neighborhood of the cone over $S$ is a subset of the form
${\mathcal{U}}(1+a,K) = \{x\in{{{\mathbb{R}}}^n}\setminus{{\bf 0}}: {{\rm d_{\rm outer}}}(\nu(x),S) < K |x|^{1+a}\} \cup \{{{\bf 0}}\}$,
for a positive real number $a$ and some positive constant $K$. There exists a connected component $\mathcal{C}$ of the subset $X_1\setminus{\mathcal{U}}(1+a,K)$ whose tangent cone is exactly ${{\rm clos}}(\mathcal{C})$, for small enough positive real numbers $a$ and for some positive constant $K$. Thus the boundary of the complement $X_1\setminus{\mathcal{U}}(1+a,K)$ is a separating set.
The relation of a separating set with the topological thin zone, its fast contracting homology and the locus of non-simple tangent directions is described in the next result.
\[lem:sep-set-2\] Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin and with connected link. If there exists a separating set $S_1$, then there exists another separating set $S$ contained in the topological thin zone ${{\rm Thin}}(X_1)$ such that for any positive real number $t$ small enough, the intersection $S_t = S\cap {{\bf S}}_t^{n-1}$ gives rise to a fc-cycle which cannot be contracted within the topological thin zone ${{\rm Thin}}(X_1)$. In other words the inclusion mapping $\iota:{{\rm Thin}}(X_1) \to X_1$ induces a surjective and non-injective homomorphism
$\iota_* :{\mathcal{H}}_{d-2} ({{\rm Thin}}(X_1),{{\bf 0}}) \to {\mathcal{H}}_{d-2}(X_1,{{\bf 0}})$.
Moreover condition *(SC) *holds true.**
Since the subset $S_1$ separates $X_1$, let $Y_1$ be one of the thick connected components of the complement $X_1 \setminus S_1$ and let $Y_2$ be the union of the other connected components. Both are open and thick subsets of $X_1$. The intersection of their closures ${{\rm clos}}(Y_1) \cap {{\rm clos}}(Y_2)$ is contained in the separating set $S_1$. We observe that the complement of this intersection $X \setminus ({{\rm clos}}(Y_1) \cap {{\rm clos}}(Y_2))$ coincides with the disjoint union ${{\rm int}}({{\rm clos}}(Y_1)) \sqcup {{\rm int}}({{\rm clos}}(Y_2))$ of the interior of the closures of $Y_1$ and $Y_2$. Moreover the tangent link of ${{\rm clos}}(Y_1) \cap {{\rm clos}}(Y_2)$ cannot contain any interior point of the tangent links of ${{\rm clos}}(Y_1)$ and ${{\rm clos}}(Y)_2$. Thus the tangent link of ${{\rm clos}}(Y_1) \cap {{\rm clos}}(Y_2)$ cannot contain any simple tangent direction, so it is contained in ${{\bf S}_{{\bf 0}}^{\rm ns}}X_1$ the locus of non-simple tangent directions. Since the intersection ${{\rm clos}}(Y_1) \cap {{\rm clos}}(Y_2)$ separates the subset $X_1$ and is contained in ${{\rm clos}}(Y_1)$, it is a separating set. Moreover the subset ${{\rm clos}}(Y_1) \cap {{\rm clos}}(Y_2)$ is contained in the thin zone since its tangent link is contained in ${{\bf S}_{{\bf 0}}^{\rm ns}}X_1$. Last ${{\rm clos}}(Y_1) \cap {{\rm clos}}(Y_2)$ is the boundary of the disjoint union ${{\rm int}}({{\rm clos}}(Y_1)) \sqcup {{\rm int}}({{\rm clos}}(Y_2))$. Defining $S$ as the intersection ${{\rm clos}}(Y_1) \cap {{\rm clos}}(Y_2)$, we deduce that the family $(S_t)_t$, intersections of the subset $S$ with the Euclidean spheres of small radius $t$, are fc-cycles which are trivial in the link $(X_1)_t$ but which cannot be contracted in any sufficiently small conic neighborhood ${{\rm Cone}}({{\bf S}_{{\bf 0}}^{\rm ns}}X_1,{{\varepsilon}})\cap X_1$ of the cone over the locus of non-simple tangent directions, that is the topological thin zone.
It is clear that the intersection of the locus ${{\bf S}_{{\bf 0}}^{\rm ns}}X_1$ with the tangent link of the closure ${{\rm clos}}(Y_1)$ is of codimension at least two since the tangent link of the closure ${{\rm clos}}(Y_1)$ is disconnected from its complement in ${{\bf S}_{{\bf 0}}}X_1$ by the tangent link ${{\bf S}_{{\bf 0}}}S$ of the subset $S$.
We can now state the first main result of this section which is just a juxtaposition of the two previous Lemma.
\[prop:sep-set\] Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space, thick at the origin and with connected link.
There exists a separating set if and only if the locus of non-simple tangent directions ${{\bf S}_{{\bf 0}}^{\rm ns}}X_1$ satisfies condition *(SC). ***
The second main result of this section clearly highlights the connection between separating sets and the fast contracting homology of the topological thin zone.
\[thm:sep-set-pure-dim\] Let $X_1$ be a simply embedded model of a closed definable isolated singularity of some Euclidean space thick at the origin and with connected link. Suppose the tangent link is of pure dimension $d-1$. There exists a separating set if and only the surjective homomorphism of the fc-homology groups
$\iota_*:{\mathcal{H}}_{d-2} ({{\rm Thin}}(X_1),{{\bf 0}}) \to {\mathcal{H}}_{d-2}(X_1,{{\bf 0}})$
induced by the inclusion mapping $\iota:{{\rm Thin}}(X_1) \to X_1$ is not injective
One of the statement is just of Lemma \[lem:sep-set-2\].
Suppose that the homomorphism is not injective. There exists a $(d-2)$-cycle $\xi$ contracted by a chain $\eta$ with thin support and containing the origin. Since the cycle $\xi$ is trivial in the link ${{\bf S}_{{\bf 0}}}X_1$ it separates the link, and thus the support of $\eta$ must separate $X_1$. Let $Y_1$ be a thick connected component of the complement $X_1\setminus{{\rm supp}}(\eta)$. Assume that all the other connected components are thin. Since the cycle $\xi$ is not trivial in the thin zone, it cannot be contracted in the link of the thin zone ${{\rm Thin}}(X_1)$. We deduce that there must be a point in the tangent link which cannot lie in the union of the tangent link of the thin zone ${{\rm Thin}}(X_1)$ with the tangent link of ${{\rm clos}}(Y_1)$. This implies that the tangent link of the closure of the complement $X_1\setminus ({{\rm supp}}(\eta)\cup Y_1)$ is of dimension smaller than or equal to $\dim X_1 -2$ and cannot be contained in the tangent link of the union $Y_1\cup{{\rm Thin}}(X_1)$ (which contains the tangent link of the union $Y_1\cup{{\rm supp}}(\eta)) $. This latter fact contradicts that the tangent link ${{\bf S}_{{\bf 0}}}X_1$ is of pure dimension $\dim X_1 -1$.
Separating sets were introduced and exhibited as objects whose presence denies the local inner metric conicalness of the set-germ at the considered point.
Proposition \[prop:sep-set\] provides an equivalent formulation of the existence of separating set for simply embedded models of thick isolated singularities by means of some topological conditions on the tangent link (Condition (SC)). Theorem \[thm:sep-set-pure-dim\] emphasizes the non-triviality in the topological thin zone of the fast contracting homological nature of separating sets.
These last two results suggest that the notion of separating set could be slightly generalized in order to still detect an obstruction to (local) metric conicalness.
[BirBr2]{} A. Bernig & A. Lytchak, *Tangent spaces and Gromov-Hausdorff limits of definable spaces, J. Reine Angew. Math. [**608**]{} (2007), 1–15. L. Birbrair, *Lipschitz geometry of curves and surfaces definable in O-minimal structures, Illinois J. Math. 52 (2008), no. 4, 1325–1353. L. Birbrair & J.P. Brasselet, *Metric homology, Comm. Pure Appl. Math. 53 (2000), no. 11, 1434-1447. L. Birbrair & J.P. Brasselet, *Metric homology for isolated conical singularities, Bull. Sci. Math. 126 (2002), no. 2, 87–95. L. Birbrair & T. Mostowski, *Normal embeddings of semialgebraic sets, Michigan Math. J. 47 (2000), no. 1, 125–132. L. Birbrair & A. Fernandes *Inner metric geometry of complex algebraic surfaces with isolated singularities, Comm. Pure Appl. Math. 61 (2008), no. 11, 1483–-1494. L. Birbrair & A. Fernandes & V. Grandjean & D. O’Shea, *Choking horns in Lipschitz Geometry of Complex Algebraic Varieties, preprint, 9 pages, available at [http://arxiv.org/pdf/1206.3105.pdf]{} L. Birbrair & A. Fernandes & W. Neumann, *Bi-Lipschitz geometry of weighted homogeneous surface singularities, Math. Ann. 342 (2008), no. 1, 139-–144. L. Birbrair & A. Fernandes & W. Neumann, *Separating sets, metric tangent cone and applications for complex algebraic germs, Selecta Math. 16 (2010) 377–391. L. Birbrair & W. Neumann & A. Pichon, *The thick-thin decomposition and the bilipschitz classification of normal surface singularities, Preprint arXiv:1105.3327, 41 pp. *[Locally flat embedding of topological manifolds]{}, [Annals of Math., [**75**]{} No.2 (1962), 331–341]{} M. Coste, *An introduction to o-minimal geometry, http://perso.univ-rennes1.fr/michel.coste/polyens/OMIN.pdf L. van den Dries, *Tame topology and o-minimal structures. London Mathematical Society Lecture Note Series, 248, Cambridge University Press, Cambridge, 1998. L. van den Dries, & C. Miller, *Geometric categories and o-minimal structures, Duke Math. J. 84 (1996), no. 2, 497-–540. A. Durfee, *Neighborhoods of algebraics sets, Trans. Amer. Math. Soc. 276 (1983), no. 2, 517-–530. A. Fernandes, *Topological equivalence of complex curves and bi-Lipschitz homeomorphisms, Michigan Math. J. 51 (2003), no. 3, 593–606. A. Fernandes, *Fast loops on semi-weighted homogeneous hypersurface singularities, J. Singul. 1 (2010), 85–93. H. Federer & W. H. Fleming, *Normal and integral currents, Ann. of Math. (2) 72 (1960) 458–520. D. Grieser, *Local geometry of singular real analytic surfaces, Trans. Amer. Math. Soc. 355 (2003), no. 4, 1559–1577. R. Hardt, *Semi-algebraic local-triviality in semi-algebraic mappings, Amer. J. Math. 102 (1980), no. 2, 291–-302. K. Kurdyka, *On a subanalytic stratification satisfying a Whitney property with exponent 1, Real algebraic geometry (Rennes, 1991), 316–-322, Lecture Notes in Math., 1524, Springer, Berlin, 1992. T.L. Loi, *Thom stratifications for functions definable in o-minimal structures on $({{\mathbb{R}}},+,·)$, C. R. Acad. Sci. Paris Sér. I Math. 324 (1997), no. 12, 1391–-1394 S. Lojasiewicz, *Triangulation of semi-analytic sets, Ann. Scuola Norm. Sup. Pisa (3) 18 1964 449-–474. J. Milnor, *Singular points of complex hypersurfaces, Annals of Mathematics Studies, No. 61 Princeton University Press, Princeton, N.J. 1968. A. Parusiński, *Lipschitz stratification of definable sets, Ann. Sci. École Norm. Sup. (4) 27 (1994), 661–696. W. Pawłucki, *Lipschitz cell decomposition in o-minimal structures. I, *Illinois J. Math. 52 (2008), no. 3, 1045–-1063. F. Pham and B. Teissier, *Fractions lipschitziennes d’une algèbre analytique complexe et saturation de Zariski, Centre de Mathématiques de l’École Polytechnique (Paris), Juin 1969. M. Shiota, *Geometry of subanalytic and semialgebraic sets. Progress in Mathematics, [**150**]{} Birkhäuser Boston, Inc., Boston, MA, 1997, 431 pp. G. Valette, *Lipschitz triangulations, Illinois J. Math. 49 (2005), no. 3, 953–979. G. Valette, *On metric types that are definable in an o-minimal structure, J. Symbolic Logic 73 (2008), no. 2, 439–447. G. Valette, *Vanishing Homology, Selecta Math. (N.S.) 16 (2010), no. 2, 267–296.********************************
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Abundances of heavier elements (barium and beyond) in many neutron-capture-element-rich halo stars accurately replicate the solar system $r$-process pattern. However, abundances of lighter neutron-capture elements in these stars are not consistent with the solar system pattern. These comparisons suggest contributions from two distinct types of $r$-process synthesis events – a so called main $r$-process for the elements above the $r$-process peak and a weak $r$-process for the lighter neutron-capture elements. We have performed $r$-process theoretical predictions to further explore the implications of the solar and stellar observations. We find that the isotopic composition of barium and the elemental Ba/Eu abundance ratios in $r$-process-rich low metallicity stars can only be matched by computations in which the neutron densities are in the range 23 $\lesssim$ log $n_n$ $\lesssim$ 28, values typical of the main $r$-process. For $r$-process conditions that successfully generate the heavy element pattern extending down to A = 135, the relative abundance of produced in this mass region appears to be at least $\sim$ 90% of the observed solar value. Finally, in the neutron number density ranges required for production of the observed solar/stellar $r$-process-peak (A $\approx$ 200), the predicted abundances of inter-peak element hafnium (Z=72, A $\approx$ 177-180) follow closely those of -peak elements and lead. Hf, observable from the ground and close in mass number to the $r$-process-peak elements, might also be utilized as part of a new nuclear chronometer pair, Th/Hf, for stellar age determinations.'
author:
- |
Karl-Ludwig Kratz, Khalil Farouqi, Bernd Pfeiffer, James W. Truran,\
Christopher Sneden, and John J. Cowan
nocite:
- '[@cam57]'
- '[@bur57]'
- '[@tru02]'
- '[@sne03a]'
- '[@cow04]'
- '[@cow06]'
- '[@kra00; @kra05a; @kra05b]'
- '[@pfe01; @pfe02]'
- '[@mol03]'
- '[@kra06]'
- '[@woo94]'
- '[@tak94]'
- '[@lat77]'
- '[@ros99]'
- '[@fre99a]'
- '[@bus99]'
- '[@tru02]'
- '[@sne03a]'
- '[@cow06]'
- '[@wes00]'
- '[@cow02]'
- '[@hil02]'
- '[@sne03]'
- '[@chr04]'
- '[@iva06]'
- '[@joh01]'
- '[@bar05]'
- '[@was96]'
- '[@qia00]'
- '[@mey00]'
- '[@joh02]'
- '[@aok05]'
- '[@bar05]'
- '[@tra04]'
- '[@kra88; @kra93; @kra05a; @kra06]'
- '[@thi94]'
- '[@pfe97; @pfe01]'
- '[@fre99b]'
- '[@wan04]'
- '[@far05]'
- '[@kra93]'
- '[@tak94]'
- '[@fre99b]'
- '[@far06]'
- '[@kra93]'
- '[@rau04]'
- '[@far06]'
- '[@kra93]'
- '[@cow99]'
- '[@mol90]'
- '[@mol97]'
- '[@mol03]'
- '[@kra98; @kra00; @kra05a; @kra05b]'
- '[@pfe97; @pfe01]'
- '[@kra06]'
- '[@kra93]'
- '[@dob96]'
- '[@pea96]'
- '[@lun03]'
- '[@rik05]'
- '[@pea04]'
- '[@kra00; @kra05a; @kra05b; @kra06]'
- '[@dil03]'
- '[@hil76]'
- '[@mol95]'
- '[@abo95]'
- '[@gor01]'
- '[@gor01]'
- '[@dob96]'
- '[@pea96]'
- '[@aud03]'
- '[@kap89]'
- '[@bur00]'
- '[@sim04]'
- '[@cow06b]'
- '[@arl99]'
- '[@tra04]'
- '[@den06]'
- '[@kra93]'
- '[@pfe97]'
- '[@cow99]'
- '[@kra93]'
- '[@cow99]'
- '[@cow99]'
- '[@far06]'
- '[@pfe02]'
- '[@aud03]'
- '[@kra05a; @kra06]'
- '[@cow04]'
- '[@sne03]'
- '[@iva06]'
- '[@cow06]'
- '[@kra93]'
- '[@fre99b]'
- '[@rau04]'
- '[@far06]'
- '[@sne02]'
- '[@aok03a; @aok03b]'
- '[@and89]'
- '[@hau72]'
- '[@law01]'
- '[@lod03]'
- '[@bur00]'
- '[@sim04]'
- '[@cow06b]'
- '[@arl99]'
- '[@far06]'
- '[@fre99b]'
- '[@cow91]'
- '[@rau00]'
- '[@lam02]'
- '[@joh02]'
- '[@hon04]'
- '[@bar05]'
- '[@wes00]'
- '[@cow02]'
- '[@hil02]'
- '[@sne03]'
- '[@iva06]'
- '[@far06]'
- '[@law06b]'
- '[@cow99]'
- '[@pfe97]'
- '[@cow99]'
- '[@cow99]'
- '[@mol01]'
- '[@far06]'
- '[@mam98]'
- '[@kod75]'
- '[@mey89]'
- '[@how80]'
- '[@hil02]'
- '[@sim04]'
- '[@cow99]'
- '[@cow02]'
- '[@tru02]'
- '[@cow06]'
- '[@iva06]'
- '[@hil02]'
- '[@lod03]'
- '[@ple04]'
- '[@bur57]'
- '[@cam57]'
- '[@far06]'
- '[@sne03; @den03; @law06a; @den06; @law06b; @den05]'
title: 'Explorations of the [*r*]{}-Processes: Comparisons between Calculations and Observations of Low-Metallicity Stars'
---
INTRODUCTION
============
The nature of rapid neutron-capture nucleosynthesis (the ) and its contributions to the abundances of post iron-peak elements (Z $>$ 30) were first delineated in pioneering studies by Cameron (1957) and Burbidge (1957). The details, however, still remain to be worked out (, see Truran 2002; Sneden & Cowan 2003; Cowan & Thielemann 2004; Cowan & Sneden 2006 for recent reviews and discussion). The physics of the involves nuclear masses, $\beta$-decay and neutron-capture () rates, and fission properties of unstable nuclear species far from the region of $\beta$-stability. Laboratory conditions needed to measure properties of such exotic nuclei are very difficult to produce, but much progress has occurred over the past decade. About 50 $\beta$-decay half-lives and 10 nuclear masses have now been measured for the lighter (A $\leq$ 140) isotopes in the production path at neutron freezeout (, Kratz 2000, 2005a,b; Pfeiffer 2001a, 2002; M[ö]{}ller 2003; Kratz 2006). Experimental data for heavier isotopes are not yet available.
Additionally, we do not clearly understand the characteristics of the stellar or supernova environments in which synthesis occurs. Circumstantial evidence for the synthesis of the heavy (A $\ge$ 130) nuclei in some site associated with massive stars – with lifetimes (production timescales) $\tau_{r-process}$ $\leq$ 10$^8$ years – seems compelling. Proposed sites include both an in a high entropy (neutrino driven) wind from a Type II supernova (Woosley 1994; Takahashi, Witti, & Janka 1994) and one occurring in the decompressed ejecta of neutron star mergers (Lattimer 1977; Rosswog 1999; Freiburghaus 1999a). Better understood is the astrophysical site for the slow $n$-captures that synthesize the heaviest nuclei (the so-called “main” component): He-fusion zones of low and intermediate mass asymptotic giant branch stars (see, , Busso, Gallino, & Wasserburg 1999) of significantly longer lifetimes, $\tau_{s-process}$ $\geq$ 10$^9$ years. This difference in the timescales for heavy element enrichment makes the lowest metallicity (=oldest?) halo stars attractive laboratories for empirical insights into synthesis (Truran 2002; Sneden & Cowan 2003; Cowan & Sneden 2006).
Significant numbers of very metal-poor -rich halo giants have been discovered in the past several decades and analyzed with ever-increasing detail and accuracy. We define these stars to be Galactic halo members that have \[Fe/H\] $\lesssim$ –2,[^1] relative overabundance factors of ten or more (, \[Eu/Fe\] $\gtrsim$ +1), and clear evidence for dominance over the (usually indicated by \[Ba/Eu\] $\lesssim$ –0.7). Comprehensive abundance analyses involving 10-40 elements have been published for individual -rich stars by Westin (2000; HD 115444), Cowan (2002; 17), Hill (2002; CS 31082-001), Sneden (2003; CS 22892-052), Christlieb (2004; CS 29497-004) and Ivans (2006; HD 221170).
There also have been recent larger-sample studies of low metallicity -rich giants. Johnson & Bolte (2001) analyzed a set of 22 stars but with generally fewer elements per star: more than 10 elements were detected in four stars, and more than five elements in an additional 13 stars. Employing typically 18 elements per star, Honda (2004) have made fresh analyses of five of the stars studied by previous investigators, and have added two additional ones. Barklem (2005) have presented initial results of a halo-star survey specifically designed to identify and supply initial high-resolution abundance analyses of up to nine elements in new -rich candidates. Over 40 new -rich stars were reported in that paper. More such stars will undoubtedly be discovered as the high-resolution survey work continues.
In all -rich stars studied to date, the abundance distributions for the heavier elements (Z $\geq$ 56, Ba and beyond) are nearly identical. We illustrate this apparently “standard” distribution in Figure \[f1\] for five of these stars. The small star-to-star scatter in observed abundances about the mean pattern is dominated by observational/analytical uncertainties. This uniformity was probably not predictable a priori, since abundances in -rich low metallicity stars reflect contributions from a single supernova or a small number of supernova events, while the solar system (hereafter, SS) abundances result from many generations of supernovae. This suggests that there is a [*unique*]{} astrophysical site that dominates the nucleosynthesis of the heaviest isotopes. The mechanism for the synthesis of the A $\gtrsim$ 130 isotopes (hereafter, the “main” component) must be extremely robust.
However, abundance patterns in the mass regime below A $\approx$ 130 for the solar system and very metal-poor stars do not support such a simple story. Considering data from carbonaceous chondrite meteorites, Wasserburg, Busso, & Gallino (1996) first proposed that abundances of the short-lived and isotopes in primordial SS matter are inconsistent with their having been formed in uniform production together with the heavy radioactivities (specifically, , , and ). Wasserburg argued that these two different mass ranges of nuclei require different timescales for production, and thus require two distinct sites. Qian & Wasserburg (2000) further quantified this idea, distinguishing the two occurrences as: (1) H (high frequency) events that are the main source of heavy nuclei (A $>$ 130) but not ; and (2) L (low frequency) events that are largely responsible for light nuclei (A$<$130) including . These suggestions formed the initial impetus for distinguishing so-called “weak" and “main" components. We caution that this scenario is not universally accepted. For example, a pure origin for the important abundance has been questioned by Meyer & Clayton (2000). Instead, those authors argue that the enhanced abundance of with respect to and in meteorites is due to the injection into the solar nebula of material from outside the edge of the helium-exhausted core of a massive star. If this interpretation is valid, then one observational basis for two sites may not be solid.
Inconsistencies between abundances of lighter and heavier abundances are also found in -rich low metallicity stars. If the solar system abundance curve is scaled to fit the metal-poor-star abundances of elements with Z $\geq$ 56, then several elements in the range 39 $\leq$ Z $\leq$ 50 (, Y, Ag) appear to have significant under-abundances in the stars. This result has been found in many of the studies cited above, and is illustrated in detail in Figures 7 and 8 of Sneden (2003). Additionally, Johnson (2002), Aoki (2005), and Barklem (2005) have investigated the relationship between the abundances of the light Sr-Y-Zr group and the heavier main-component Ba in -dominated stars, finding a very large scatter at lower metallicities. Travaglio (2004) have explored the production of Sr-Y-Zr in relation to the heavier elements over a large stellar metallicity range. These results all provide further evidence for a second distinct (“weak”) site for the synthesis of isotopes below A $\approx$ 130.
In this paper we report updated theoretical computations employing new nuclear data, and use them to explore several aspects of solar system and low-metallicity observational results of elements. The computations are described in §2. The calculated abundances are compared with selected observational data in §3, to address: (a) neutron density constraints imposed by barium isotopic and elemental abundances, (b) the contributions from the two $r$-processes to light abundances, and (c) a new connection between rare-earth and the heaviest stable elements. Finally, in §4 we discuss the new calculations in terms of nuclear chronometers.
[*r*]{}-PROCESS NUCLEOSYNTHESIS CALCULATIONS
============================================
Our calculations were performed under the conditions of the waiting-point assumption or (n,$\gamma$) ${\leftrightharpoons}$ ($\gamma$,n) equilibrium (see Kratz 1988, 1993, 2005a, Kratz 2006; Thielemann 1994; Pfeiffer 1997, 2001a; Freiburghaus 1999b). Although dynamic calculations are feasible today (see , Freiburghaus 1999 and Farouqi 2006), the classical approach can be employed to compare with the astronomical observations - the goal of this paper. Even with the simplistic assumptions of constant neutron number density, and temperature, and instantaneous freezeout the equilibrium models presented here reproduce the solar system abundances well. The classical approach is also largely independent of a stellar model, whereas all recent “more realistic" calculations imply a specific astrophysical environment, , supernovae or neutron-star mergers, which may require quite different astrophysical parameter sets.
More recent large-scale, fully dynamic network calculations within the SN II high-entropy-wind model have shown (Wanajo 2004; Farouqi 2005) that there exists a rather wide range of correlated astrophysical parameters, such as $Y_e$, S, $Y_n$, Y$_{seed}$ and V$_{expansion}$, within which a robust can be performed. However, so far SN models do not yet provide convincingly “unique” astrophysical conditions within the expected parameter space, which would constrain the . In addition, a careful comparison between the classical (, Kratz, 1993) with supernova thermodynamic trajectories given by Takahashi, Witti & Janka (1994), has shown very good agreement (for T$_9$ – $n_n$) between the “waiting-point” equilibrium and dynamic conditions. The assumption of constant $n_n$ and temperature is also a good approximation up until the freezeout, as shown by , Freiburghaus (1999b) and Farouqi (2006). We also note that the neutron freezeout, which occurs on a very rapid timescale (Kratz 1993), does not affect the abundances up to and slightly beyond the peak, and does not have a significant effect on the rare earth or -peak regions (see , Rauscher 2004 and Farouqi 2006).
Our calculations to reproduce the total isotopic SS abundance pattern covered a neutron-density range of 20 $\leq$ log $n_n$ $\leq$ 30 in steps of 0.5 dex, with the neutron exposure times $\tau$($n_n$) and weighting functions $\omega$($n_n$) for each $n_n$-component as given in Kratz (1993) and Cowan (1999). The upper neutron-density limit is constrained by SS abundances. Our results indicate that the highest neutron number densities, log $n_n$ $\geq$ 30, make little contribution to the overall abundances – a result that is confirmed by more detailed network calculations (, Farouqi 2006). More discussion of the astrophysical parameter choices will be given in §2.2. Predicted isotopic abundance trends with neutron density are displayed in Figure \[f2\]. The general shapes of the curves are the same for all isotopes: first, a generally sharp rise to a maximum followed by a more gradual decline – if log $n_n$ $<$ 20 abundance runs were to be displayed in Figure \[f2\], then the rising portions of the curves would be more apparent for the lightest isotopes (A = 100 and 120). The details of the curves are dependent on individual nuclear properties and will not be discussed further.
Of greater interest for this paper are elemental abundances in metal-poor stars. Therefore, as stellar spectroscopy normally can yield only elemental abundances, in Figure \[f3\] we show some examples of the results summed by element. As in Figure \[f2\], predicted abundances are plotted as a function of individual neutron number density log $n_n$. Unsurprisingly, these elemental abundance variations are similar to the isotopic ones, and the same statement about the “missing” parts of the light isotope trends applies to the elemental curves as well. The data of these figures show that lighter elements are produced in bulk at smaller values of log $n_n$ than are the heavier ones.
Nuclear Data Input
------------------
The best agreement with SS abundances is obtained when we employ: [*(a)*]{} the nuclear mass predictions from an extended Thomas Fermi model with quenched shell effects far from stability (, ETFSI-Q; Pearson, Nayak, & Goriely 1996); and [*(b)*]{} the $\beta$-decay properties from QRPA-calculations for the Gamow-Teller(GT) transitions based on the methods described in Möller & Randrup (1990; see also Möller 1997); and the first-forbidden strength contributions from the Gross theory of $\beta$-decay (Möller, Pfeiffer & Kratz 2003).
We have previously tested quite a number of global mass models in our classical calculations (see, , Kratz 1998, 2000, 2005a,b; Pfeiffer 1997, 2001a; Kratz 2006). For our present paper we have chosen to use the “quenched” mass formula ETFSI-Q for the following reasons. In principle, a weakening of the strong N=82 shell below had already been suggested by Kratz (1993) from, at that time, scarce experimental indications in the phase transitional region around A=100. Following the microscopic, self consistent description of “shell quenching” far from stability in the spherical mass model HFB(Hartree-Fock-Bogoliubov)/SkP(Skyrme force) of Dobaczewski (1996), Pearson (1996) implemented this behavior in an algebraic procedure for the shell closures N=50, 82 and 126 into their “unquenched” deformed ETFSI mass formula, hereafter called ETFSI-Q. The big improvement in changing the matter flow at the closed-shell “bottle-necks”, and consequently in reproducing the overall SS abundance pattern by our classical waiting-point calculations, was evident immediately (Kratz 1998; Pfeiffer 1997, 2001a).
Today, several other microscopic HFB models have become available with the same overall quality as the ETFSI-Q approach in terms of the root-mean-square differences between experimentally known and predicted masses (see , Lunney 2003; Rikovska-Stone 2005; Pearson 2004). However, these models do not reproduce well the experimental N = 82 shell-gap behavior around double-magic , where the dominant bottle-neck in the total matter flow occurs (Kratz 2005a,b; Kratz 2006, and further discussion below). Furthermore, the recent HFB models show a rather unphysical chaotic behavior in their $S_{2n}$ systematics, indicating that the treatment of pairing in these approaches is inadequate (see, , Rikovska-Stone 2005, Kratz 2005a,b; Kratz 2006). Summarizing, although ETFSI-Q may not be the ultimate mass model for calculations, it is preferred for our computations because it best reproduces the experimentally measured nuclear-structure quantities of more than 20 isotopes in the A = 130 $r$-abundance peak region.
Our nuclear database has been improved by including recent experimental results and, based on a better understanding of the underlying shell structure, new theoretical predictions. By now, altogether more than 50 isotopes lying in the path under log $n_n$ = 20 freezeout conditions between and have been measured, at least via their $\beta$-decay half-life. Of particular importance are the new spectroscopic data on N $\simeq$ 82 Ag, Cd, In., Sn, and Sb isotopes, which have led to a better understanding of the A $\simeq$ 130 $r$-abundance peak as the major bottle-neck for the matter flow to the heavier elements (see, , Kratz 2000, 2005a,b, 2006). In this context, presumably the measurement of utmost importance is the determination of the high (energy involved in a $\beta$-decay) $Q_{\beta}$ value of N = 82 , which represents the isobaric mass difference between and its daughter (Dillmann 2003; Kratz 2005b). This experimental value is in clear disagreement with predictions from all older “unquenched” global mass models (, the GTNM or Gross Theory of Nuclear Masses, of Hilf 1976; FRDM (Finite Range Droplet Model) of Möller 1995; ETFSI-1 of Aboussir 1995), as well as from the recent series of more microscopic HFB approaches (, Goriely 2001, (HFB-2); Samyn 2004, (HFB-8); Goriely (HFB-9), 2005; Rikovska-Stone 2005). The experimental value is only in agreement with the “quenched” mass models HFB/SkP (Dobaczewski 1996), ETFSI-Q (Pearson 1996) and the latest nuclear-mass evaluation of Audi (2003).
An analysis of the discrepancies between measured and calculated $\beta$-decay properties reveals considerable improvement over the earlier evaluation. For potential progenitor isotopes with half-life T$_{1/2}$ $\leq$ 0.2 s, the total error is now a factor of two; for the $\beta$-delayed neutron emission probability, $P_n$ $\geq$ 1 $\%$, the mean deviation is a factor of three. Moreover, we have performed a careful parameter study of the “robustness” of T$_{1/2}$ and $P_n$ predictions of the N $\simeq$ 126 waiting-point nuclei forming the A $\simeq$ 195 -peak, where no experimental data are available at all. Again, our model predictions lie within a factor of two of the observed abundances – this is in contrast to earlier calculations, often uncertain by an order of magnitude or more, that were dominated by uncertain nuclear physics data for the most neutron-rich nuclei. Taken together, this gives us confidence in the reliability of our nuclear-physics input to the subsequent calculations.
Abundance Fits to the Sun and 22
--------------------------------
As our calculations demonstrate, a range of neutron densities are required to reproduce all major features of the SS meteoritic isotopic abundance distribution. However, we remind the reader that the [*total*]{} meteoritic abundances are generally combinations of both and contributions. Because the proceeds along the valley of stability and depends mostly on neutron capture cross sections that are directly measurable in the lab, its fraction of the total can be estimated. This is done either empirically by fitting a smooth curve to the $N\sigma$ versus mass distribution of isotopes that can only be synthesized in the (K[ä]{}ppeler 1989; Burris 2000; Simmerer 2004; Cowan 2006), or theoretically by computing an abundance set with conditions that correspond to those expected in stellar interior He-fusion zones (Arlandini 1999; Travaglio 2004). Then the chosen abundance set is subtracted from the total SS abundances to yield the set. The resulting abundances may be very accurate for those isotopes with little contribution (, and ), but have significant uncertainties for isotopes where the fraction is dominant (, , the most abundant of the seven naturally-occurring Ba isotopes). This caution should be kept in mind in all SS abundance comparisons, but see also a recent attempt at directly predicting the SS abundances for several rare-earth elements (Den Hartog 2006).
Attempts to fit predicted abundances to the SS values require assumption of a continuous addition of a small number of individual neutron density components, with a varying path related to contour lines of constant neutron separation energies in the range of 4–2 MeV (the latter being determined by the combination of neutron number density and temperature). The number of such components needs to be at least four to match the relative abundances of the three abundance peaks (A = 80, 130, and 195) and abundances in the actinide region (Kratz 1993). Less than 20 components are sufficient to match the detailed SS abundances (Pfeiffer 1997 and Cowan 1999). Adding more components produces no further improvement, given the present uncertainties in both computations and SS abundances.
While such a procedure is both largely site-independent and mainly intended to produce a good fit to solar abundances, it can also provide information regarding the conditions that a “real” site has to fulfill. The fit is performed by adjusting the weight of the individual components (or different neutron separation energies $S_n$($n_n$,$T$)) and the time duration $\tau$ for which these (constant) conditions are experienced, starting with an initial abundance in the Fe-group.[^2] For a given (arbitrary) temperature, $S_n$ is a function of neutron number density $n_n$. The addition weights $\omega$($n_n$) and process durations $\tau$($n_n)$ have a behavior similar to powers of $n_n$. This corresponds to a linear relation in log $n_n$ and is already observed when taking a minimum of three components in order to fit the abundance peaks (Kratz 1993, Cowan 1999). This approach (although only a fit and not a realistic site calculation) is reasonable to the extent that such a continuous dependence on physical conditions can be expected to reflect the range of conditions appropriate to the (yet unidentified) astrophysical site.
The effect of the weighting of the sums of the various components on the abundance fit is shown in the lower panel of Figure \[f4\]. Relatively small weighting factors for the highest neutron number densities and larger weighting values for lighter neutron density components yield good fits to the data. Overall the best fit to this region of the SS abundances (from A $\approx$ 130 to the actinides) is that of log $n_n$ = 20 to 26–27.
For a choice of a typical freezeout temperature (in billions of degrees K) of T$_9$ = 1.35 (see , Cowan 1999), the necessary addition of components (with paths of a corresponding neutron separation energy $S_n$ over a duration $\tau$) can be expressed in the form $\omega$($n_n$) $\simeq$ 8.36$\times$10$^{6}n_n^{-0.247}$ and $\tau$($n_n$) = 6.97$\times$10$^{-2}n_n^{0.062}$ sec. This algebraic/exponential fit is in good agreement with full network calculations (Farouqi 2006) for the high entropy wind scenario in terms of seed nuclei as a function of entropy/neutron-density. When restricting the fit to the mass regions around the peaks (A $\simeq$ 80, 130, 195), where the paths come closest to stability, model extrapolations need not be extended far into unknown territory. Even a significant amount of experimental information is available (Pfeiffer 2002, Audi 2003; Kratz 2005a, Kratz 2006), and is now being utilized in the calculations. These power laws in $n_n$ play roles comparable to that of the assumed exponential addition of neutron exposures in the classical .
The relationship between component abundances computed at single neutron densities to their superposition is illustrated in Figure \[f4\]. In the top panel we show individual, unweighted calculations, ranging from log $n_n$ = 24 to 30, superimposed on the solar isotopic abundances (black dots). We have arbitrarily normalized all curves to match the abundances at A = 195. There are several trends evident in this plot. First, larger neutron number densities (log $n_n$ $\geq$ 25), in general, are required to synthesize the heaviest elements (, the peak and actinide regions). The larger values of $n_n$ are needed to push the so-called “ path” – defining where these nuclei are produced – far enough away from stability, and to very neutron-rich regions, so that the synthesized radioactive nuclei subsequently decay back to high enough mass numbers (, see Figure 1 in Cowan & Thielemann 2004). However, the highest neutron density (log $n_n$ = 30) overproduces the heavy region and the Eu region (A $\approx$ 150). Additionally, the largest neutron densities produce ratios of peak to interpeak abundances that are clearly different than those observed in the solar system. This occurs because such a high value of $n_n$ pushes the path too far into the radioactive region and consequently overproduces the heaviest nuclei. On the other hand the lighter neutron number densities log $n_n$ = 24 do mimic the A $\approx$ 150 range, but are clearly inadequate to synthesize the peak abundances. This reinforces the idea that multiple neutron exposures are required to reproduce the three different SS peaks – one exposure will not be adequate.
We show in Figure \[f5\] the detailed steps in one such typical addition calculation. These calculations include explicitly $\beta$-decays back to pseudo (, long-half life) stability, but do not include $\alpha$-decays from the trans-lead region. In the top panel of this figure, the summation of five individual neutron number density components in the range log $n_n$ = 20–22 have been tuned to reproduce the A = 80-100 SS isotopic abundances. This composition is clearly inadequate to reproduce the abundances of the heavier isotopes. As shown in the second panel of Figure \[f5\], addition of higher neutron number density components (up to log $n_n$ = 24) mostly reproduces the A = 130 abundance peak. Therefore, we estimate that the division between “weak” and “main” density regimes occurs at log $n_n$ $\simeq$ 23.0 $\pm$ 0.5. Likewise, to form the heaviest stable elements, peak, it is necessary to include neutron number density components at least up to log $n_n$ $\simeq$ 26 (Figure \[f5\], third panel). However, even this neutron density is still insufficient to reproduce the abundances of the trans-lead region (including Th and U), which requires log $n_n$ $\simeq$ 28 as shown in the bottom panel of the figure.
In Figure \[f6\] the illustration is reversed. Considering our best-fit weighted distribution (the bottom panels of both Figures \[f5\] and \[f6\]) leads to a few comments. First, the parts of the heaviest stable elements Pb (Z = 82, A = 206–208) and Bi (Z = 83, A = 209) in the SS are principally due to $\alpha$-decays of nuclei along the radioactive decay chains extending through the actinide region. Their observed abundances are reproduced well by our calculations. We show in the subsequent panels (starting at the top with the highest values only) how progressively adding additional lower neutron number density components leads to a better and better fit of the solar system isotopic abundances – specifically, in the second panel at the rare-earth region, the third panel the A = 195 peak and in the bottom panel the complete SS abundance pattern (see Cowan 1999). We note also, as illustrated in Figure \[f5\], that very low neutron densities (log $n_n$ $\simeq$ 20) are insufficient to reproduce the first peak, and very high densities (log $n_n$ = 30) overproduce the peak and the actinides, as shown in Figure \[f4\].
Summing the individual isotopic abundances into elemental abundances also reproduces the observed SS [*elemental*]{} curve from Z = 30–82. Since we and many other investigators (cited in §1) have argued that the abundance pattern in -rich metal-poor stars is also consistent with this SS distribution (Figure \[f1\]), it is not surprising that multiple neutron density ranges are required to adequately match the entire abundance ranges of the stars as well.
As an example, we compare in Figure \[f7\] the empirical SS elemental distribution, theoretical predictions, and abundances of the -rich low metallicity star 22 (Sneden 2003, with updated abundances of Nd, Sm, Gd, Hf, and Pt as indicated in the figure caption.) The SS distribution has been scaled to match the 22 Eu abundance. In the top panel of Figure \[f7\] the “main” abundance calculations, for log $n_n$ $\geq$ 23.0, have also been normalized at the stellar Eu abundance. Good agreement is seen between the calculations and the mean observed abundance levels for stable elements throughout the range 56 $\leq$ Z $\leq$ 82. Comparing the 22 and the scaled solar abundances (for Z = 56-82) we find an average difference of $<$(22)–SS$_{r-only}$$>$ = 0.044 and $\sigma$ = 0.096. The comparison between the stellar and our calculated abundances yields $<$(22)–(main theory)$>$ = 0.15 and $\sigma$ = 0.25. The main as defined here reproduces the 22 data of the heavier elements (Z $>$ 56), including the full A = 130 peak, including iodine. However, both the SS only curve and our calculations (for the “main” ) do not fit the lighter $n$-capture element data (Z $<$ 56) in this star. In particular Y, Mo, Pd and Ag all deviate from both of these curves. These differences in fact have been one of the main supports for the existence of two $r$-processes. We note, however, that the main calculations reproduce the odd-Z, even-Z abundance staggering in both light and heavy elements in 22. Further, the calculated abundances fall off significantly with respect to the SS $r$-only abundances at the lower atomic numbers, and do not seriously clash with the four abundances for which only upper limits have been determined: Ge, Ga, Cd, and Sn.
Elemental abundance predictions of the “weak” (log $n_n$ $<$ 23) are overlayed with the observed abundances of 22 in bottom panel of Figure \[f7\]. These predictions only represent the light “missing part” of the solar abundances (similar to that of the top panel of Figure \[f5\]). The scaling to the 22 light abundances is approximate, for display purposes only. Consistency between predicted and observed abundances of these elements can be achieved under the weak conditions with little production of the heavier elements (5–10% of the peak and essentially 0% for Ba and beyond). As can be seen in the figure, the weak predictions do not contribute to the heavier (Z$>$ 56) $n$-capture element abundances. In other words the similar abundance pattern (for the elements with Z $\ge$ 56) seen in the -rich, metal-poor stars (Figure \[f1\]) does not need or have any contribution from the weak . The figure also confirms our results from above - that the SS abundance peak can be reproduced with a low $n_n$ ($<$ 10$^{23}$). The general conclusions about a separation in abundance levels (and the associated synthesis conditions) between the lighter and heavier elements are not new, and have been discussed previously in the literature (, Sneden 2003; Ivans 2006; Cowan & Sneden 2006). However, our new calculations suggest that the separation occurs below the element iodine, which appears to be formed along with Ba in the A = 130 abundance peak. We note in this regard that there are experimental nuclear data in the the path in the mass range from to . For these isotopes where the data might not be well defined (leading to large uncertainties in the residuals), the experimental data might be used in the future to directly predict, or at least constrain, the actual SS abundances. Such data could be employed to quantify where in mass (atomic number) the possible separation between the main and weak (if one exists) occurs.
DETAILED ABUNDANCE SIGNATURES OF THE $r$-PROCESS
================================================
In this section we consider some additional abundance clues from the details of our calculations, beyond considerations of the overall fits to solar and stellar abundance distributions. In the near future some isotopic, as well as additional (yet unobserved) elemental, abundance ratios will become available for certain stars. Such ratios will provide increasingly stringent constraints on both $s$- and nucleosynthesis contributions. Our calculations predict isotopic abundance ratios for the light-heavy elements – up to the heavy wing of the A=130 peak which includes for example the Ba isotopes. At least within the high entropy model, the freezeout is very fast; non-equilibrium-captures do not play a significant role; $P_n$ values are small (Kratz 1993; Freiburghaus 1999; Rauscher 2004; Farouqi 2006). Hence, the final abundance pattern of the [*whole*]{} A = 130 peak, including the Ba isotopes, is mainly determined by the initial “progenitor abundances” with their odd-Z/even-Z staggering smoothed out by delayed neutron emission during the decay back to stability. These calculations should not be viewed as necessarily tightly constraining those isotopic ratios, but instead predicting a reliable range of values as a function of neutron number densities or entropy.
Barium Isotopic Fractions
-------------------------
An additional probe of the heavy element pattern and its range in metal-poor stars is provided by the (limited) isotopic abundance information that has been reported in the literature. The isotopes of Eu in metal-poor halo stars have been investigated by Sneden (2002) and Aoki (2003a,b). For several -rich stars, the isotopic abundance fraction ratios, $f(\iso{Eu}{151})/f(\iso{Eu}{153})$ = N()/N() = 1.0 $\pm$ 0.1, are in good agreement with the meteoritic fraction, $f(\iso{Eu}{151})/f(\iso{Eu}{153})$ = 0.478/0.522 = 0.916 (Anders & Grevesse 1989). Estimates of the photospheric Eu isotopic ratio (Hauge 1972, Lawler 2001) are also in accord with these values.
Barium has five naturally-occurring isotopes, A = 134–138, which are produced in substantially different amounts in the $s$- and .[^3] Even-Z isotopes and cannot be reached by the . As a result the abundance ratio of odd-A isotopes to the total, defined as $f_{odd}$ $\equiv$ \[N()+N()\]/N(Ba), is larger in than nucleosynthesis events; $f^r_{odd}$ $>$ $f^s_{odd}$. In principle observation should be able to assess $f_{odd}$ in stars, for the odd-Z isotopes and have hyperfine-split line substructures that the even-Z isotopes lack. Therefore, careful measurement of the line broadening of lines should indicate the relative $r$-/$s$-process contributions to Ba production.
Unfortunately even the SS value of $f^r_{odd}$ for Ba, which can be determined from meteoritic studies, is not known to high accuracy. This stems from the overall dominance of barium in SS material. The contributions to the barium isotopes are small, thus any isotopic abundance uncertainties are magnified in the estimates of their contributions. Lodders (2003) has given a new empirical estimate of the and contributions to each barium isotope, and with those data we compute a value of $f^r_{odd}$ = 0.72. The Lodders isotopic breakdowns are very similar to those that were used by Burris (2000) and Simmerer (2004) to compute total solar elemental abundances of elements (see also the isotopic breakdowns in Cowan 2006). Arlandini (1999) subtracted model calculations from the solar isotopic abundances to yield their contributions. From their data we compute $f^r_{odd}$ = 0.46.
These independent assessments of $f^r_{odd}$ for Ba in the solar system are in agreement mainly because the uncertainties in each are large, about $\pm$0.2. Lacking a clear indication of which value to adopt, we simply average them to adopt $f^r_{odd}$ $\approx$ 0.60 $\pm$ 0.20. In contrast $f^s_{odd}$ is tightly constrained between values of 0.09 to 0.11 in all of the estimates, but this again is due to the dominance of Ba synthesis in SS material. From these independent assessments we adopt $f^s_{odd}$ = 0.10 $\pm$ 0.02.
In the Table \[tab1\] we list our theoretical predictions for Ba isotopic abundances, and the resulting $f^r_{odd}$ values. These calculations are for single fixed neutron number density conditions for 20 $<$ log $n_n$ $<$ 30. We show the trend of $f^r_{odd}$ with neutron number density in Figure \[f8\], indicating also in the figure the SS values of $f^r_{odd}$ and $f^s_{odd}$. In addition to the theoretical calculations adopted here, some full dynamic network calculations have been performed (Farouqi 2006) in the context of the high-entropy supernova model – with post processing during the non-equilibrium neutron-freezeout phase. These computations have been made to check or constrain the more simplistic, but detailed, parametric predictions from the waiting-point approximation employed here. The Farouqi calculations employ an extension of the dynamic code (Freiburghaus 1999b; Cowan, Thielemann & Truran 1991), but include cross sections from the NON-SMOKER code (Rauscher & Thielemann 2000). For a range of conditions (electron abundance $Y_e$ from 0.41–0.49, entropy $S$ of 200-300 k$_b$/baryon) these dynamic calculations predict a range of $f^r_{odd}$ values of 0.48 to 0.51, thus overlapping with the waiting-point approximation values.
Few assessments of $f_{odd}$ in metal-poor stars have been attempted. The wavelength shifts due to isotopic mass differences and hyperfine splits for the two odd-A barium isotopes are comparable to the line widths (due to thermal and microturbulent broadening) in stellar spectra, making isotopic abundance determinations very challenging. Recently, Lambert & Allende Prieto (2002), following an earlier investigation by Magain (1995), observed the 4554 Å line in the halo subgiant HD 140283 (\[Fe/H\] = –2.4), at very high spectral resolution and signal-to-noise. Detectable excess breadth of this line compared with single-component absorption features would be due mainly to the hyperfine splitting of the odd-A isotopes. Lambert & Allende Prieto detected the excess broadening at a marginal level, and their line profile analysis yielded a combined fractional abundance of the odd-A isotopes $f_{odd}$ = 0.30 $\pm$ 0.21.
The HD 140283 result is shown in Figure \[f8\]. Within large uncertainties, $f_{odd}$(HD 140283) $\sim$ $f^r_{odd}$(s.s.). Frustratingly, the range in isotopic values for HD 140283 does not permit exclusion of the SS value. Given that the HD 140283 result is marginally more consistent with an than an origin for that star, we examine the implications of our calculations. As indicated in Table \[tab1\] and Figure \[f8\], the lowest neutron number densities, log $n_n$ $\lesssim$ 22, produce very small amounts of Ba. Since these low densities also yield $f^r_{odd}$ values at odds with SS values, they are excluded from consideration here. Likewise, the highest neutron number densities, log $n_n$ $>$ 28, predict Ba isotopic fractions that are much larger than the values permitted by the HD 140283 measurement. Averaging the isotopic abundances in the range 23 $\leq$ log $n_n$ $\leq$ 28 yields $f^r_{odd}$ = 0.44, very similar to the Arlandini (1999) value and that found from the preliminary dynamic calculations. The meager Ba isotopic abundance data do not permit a more sharply defined neutron density range.
Barium and Europium $r$-Process-Rich Metal-Deficient Stars
----------------------------------------------------------
More progress can be made by considering Ba/Eu elemental abundance ratios, which have been used for decades to estimate the relative influence of and contributions to stellar abundances (, see Spite & Spite 1978). For SS material, log $\epsilon_{r}$(Ba) = +1.446 and log $\epsilon_{r}$(Eu) = +0.494, or log $\epsilon_{r,s.s}$(Ba/Eu) = +0.952. For HD 140283 Gratton & Sneden (1994) reported \[Ba/Fe\] = $-$0.64 $\pm$ 0.06 and \[Eu/Fe\] = +0.09 $\pm$ 0.01, yielding \[Ba/Fe\] = $-$0.73 $\pm$ 0.06, which translates to $<$log $\epsilon_{obs}$(Ba/Eu)$>$ = +0.92 $\pm$ 0.06 using their solar Ba and Eu abundances. This abundance ratio is obviously in good agreement with log $\epsilon_{r,SS}$(Ba/Eu), strengthening the suggested attribution of $f_{odd}$(Ba) to an -only synthesis.
Many very low metallicity stars exhibit similar Ba/Eu ratios. We surveyed the literature to estimate these spectroscopic values. In Figure \[f9\] we plot \[Ba/Eu\] as a function of \[Fe/H\] from several recent halo-star comprehensive abundance analyses: Johnson (2002), Honda (2004), and Barklem (2005). We added points from detailed studies of a few extremely -rich stars (Westin 2000, Cowan 2002, Hill 2002, Sneden 2003, and Ivans 2006). This is a representative (but not complete) list of Ba/Eu abundance studies in this metallicity domain. A large majority of the stars considered in the cited studies exhibit $<$log $\epsilon_{obs}$(Ba/Eu)$>$ $\simeq$ +1.00. Some points scatter to significantly larger values ($\gtrsim$+1.4) of this quantity, clearly indicative of substantial contributions to Ba in those stars. The envelope containing essentially the entire star-to-star scatter of the Ba/Eu ratios observed in -rich stars is $\approx$0.35 in width, or $<$log $\epsilon_{obs}$(Ba/Eu) $>$ = +1.00 $\pm$ 0.17
In Figure \[f10\] we show our calculated Ba/Eu abundance ratios plotted as a function of neutron number density. To this figure we add a point representing the value for HD 140283, and a color band representing the range in values for -rich low metallicity stars. The neutron number density range containing theoretical predictions and stellar observations is substantially less than that inferred from consideration of Ba isotopic abundances alone. This figure, together with log $\epsilon_{r,s.s}$(Ba/Eu) = +0.95, suggests that 24 $\lesssim$ log $n_n$ $<$ 28, with fairly well-defined boundaries on both ends of the allowable neutron number density range. Densities beyond this domain appear unable to reproduce $f_{odd}$(Ba) and log $\epsilon$(Ba/Eu) of metal-poor -rich stars, and by extension the relative abundances of most rare-earth elements in these stars.
The Iodine–Barium Connection
----------------------------
The existence of two distinct synthesis sites began with the Wasserburg (1996) critical assessment of and meteoritic abundance levels. In the preceding subsections we demonstrated that computations with an addition of neutron densities in the range 24 $\lesssim$ log $n_n$ $<$ 28 yields both barium isotopic and Ba/Eu elemental abundance ratios that are compatible with the solar abundances.
Computed abundances of Sr (Z = 38), I (53), Hf (72), and Ba (56), are shown in the top panel of Figure \[f11\] and several abundance ratios among these elements are shown in the bottom panel. These are plotted as a function of the neutron density range defined in the following way. Beginning at the highest neutron density log $n_n$ = 30, successively smaller neutron density steps represent the accumulation of the weighted abundances resulting from that particular neutron density [*plus*]{} the total of all higher densities. The highest values of log $n_n$ contribute only trace amounts of the solar and stellar abundances, which then grow as smaller neutron density contributions are added into the sums. The abundance curve of the heavy element Hf “saturates” at its approximate solar value most quickly (at log $n_n$ $\simeq$ 25.5) in agreement with the overall trend shown in Figure \[f6\], while the light element Sr does not reach its SS $r$-process value until log $n_n$ $\simeq$ 21.5.
What is more important in Figure \[f11\] is the close tracking between I and Ba in our computations. These neighboring elements correlate over a wide neutron density range – the slope of their abundance ratio shown in the bottom panel of Figure \[f11\] is essentially flat. The situation for I and “distant” Hf ($\Delta$Z = 19) is somewhat different. Here, large differences (factors of two or more) occur for “high” neutron densities, log $n_n$ $\gtrsim$ 25. These also are conditions where the Ba abundance is no longer solar. For conditions of a “main” with full solar I, Ba and Hf $r$-abundances (, 23 $<$ log $n_n$ $<$ 28), iodine at the top of the A = 130 N${_r,\odot}$ peak and hafnium at the onset of the A = 195 N${_r,\odot}$ peak are “coupled” (\[I/Hf\]$_{r,\odot}$ $\simeq$ 0.85). Thus, for the conditions in the main I, Ba and Hf appear to be synthesized together, and in solar proportions. Sizable abundance “decouplings” of these three elements only occur at the highest neutron densities, log $n_n$ $\gtrsim$ 25, when \[I/Hf\] and \[Hf/Ba\] do not retain their solar abundance ratios.
For “main” conditions, 23 $<$ log $n_n$ $<$ 28, we estimate that both the abundance level of iodine and the (I/Hf) ratio are approximately 90% of their SS values. In the context of our model, this conclusion seems quite robust and results from two primary factors. First, from nuclear-structure arguments there is a bottle-neck behavior in the flow at the N = 82 shell-closure. Thus, full-solar $r$-process Ba production is accompanied by the full-solar A = 130 abundance peak synthesis – at least at its top. Second, the “classical” reaches (or enters) the N = 82 shell at [*lower*]{} Z than the top of the peak ( or ). Our parameterized waiting-point predictions, as well as the preliminary Farouqi (2006) dynamic calculations, show that N = 82 is already reached in the Tc (Z = 43) isotopic chain – in the very extreme even in the Zr (Z = 40) chain. Hence, the “dividing line” between the two $r$-processes appears to fall well below the , because of the N = 82 shell closure far from stability. Instead, is included with the “rising wing” of the A = 130 -abundance peak from about A = 125 upwards.
Hafnium and the Third $R$-Process Peak
--------------------------------------
Hafnium (Z = 72) is the next element beyond the rare-earth group, and it may serve as an important link between those elements and the -peak. For practical abundance determinations, we note that Hf, like all of the rare earths and the radioactive elements Th and U, is detectable in metal-poor stars via absorption lines arising from low excitation states of the first ion (see Lawler 2006b). Therefore, observed ratios of Hf/Ba, Hf/Eu, Hf/Th, etc., are very insensitive to uncertainties in stellar atmospheric parameters T$_{\rm eff}$ and log $g$. This is a happier situation than that existing for ratios involving -peak elements, such as Eu/(Os,Ir,Pt) or Th/(Os,Ir,Pt), because the -peak elements are detectable only via their neutral species. Atmospheric parameter errors are echoed directly into such ratios. Additionally, the dominant transitions of -peak neutral species lie in the UV ($\lambda$ $<$ 4000 Å), making detections and reliable abundance determinations for these elements difficult.
In Figure \[f12\] we show abundance variations of Eu, Hf, Pt, Pb, and Th as a function of neutron density range, in the same fashion as was done for lighter elements in Figure \[f11\]. Variations in abundance ratios of these elements (shown in the bottom panel of Figure \[f12\]) are small for neutron densities log $n_n$ $<$ 28. Note in particular the near constancy of the Hf/Th ratio over the entire neutron density range of the . Our calculations suggest that spectroscopists should invest effort in the determination of Hf abundances (using such transitions as those at 4093.15 and 3918.09 Å) to see if the ratios of Hf/Eu and Hf/Th are constant in -rich metal-poor stars. If so, this may strengthen the use of Th/Hf in cosmochronometry studies.
**NUCLEAR CHRONOMETERS**
========================
In this paper we have emphasized the robustness of the production of heavy nuclei in the mass range from A $\simeq$ 130 through the actinide region, as reflected in the observed abundance patterns for metal-deficient -enriched field halo stars. Our ability to identify this pure pattern is a consequence of the fact that one can identify low metallicity stars for which no significant contamination from longer lived asymptotic giant branch stars (Busso 1999) has yet occurred. An important further measure of this robustness is provided by the general consistency of the stellar ages obtained with chronometers. We will examine specifically both the case of the classical thorium/uranium actinide chronometer and that of the thorium/europium chronometer.
We consider again the abundance distributions of five “pure” stars shown in Figure \[f1\], concentrating now on the the actinide chronometers (hereafter simply called Th), , and (which can be formed [*only*]{} in the ). We have computed abundances using various global mass models, constraining all of them to reproduce the SS stable pattern, with special emphasis on fitting the yields of the peak and extrapolating to masses A $\simeq$ 250. In Table \[tab2\] we summarize these computations. Column 2 lists abundance ratios derived from calculations described in Cowan (1999). Columns 3 and 4 are these ratios from new calculations using Fe seeds that yield the best overall fit to the stable abundance data for masses A $>$ 83 (“fit1”, considering all isotopes), and for masses A $>$ 125 (“fit2”, restricted to only those isotopes matched by the main ). Column 5 employs similar calculations using a Zr seed that best match the A $>$ 125 nuclei. Finally, column 6 lists the Th abundance from the fit2 calculation to the present-day observed SS elemental abundances.
Table \[tab2\]’s Th/U production ratios lie in the range 1.475 $\leq$ Th/ $\leq$ 1.805. The present computations suggest Th/ $\simeq$ 1.5 for this important chronometer pair. The new lower values result from improved nuclear data. Specifically, there is a significant change in the -matter flow through the A $\simeq$ 130 bottle-neck region of the N$_{r,\odot}$ peak, which continues to affect the build-up of the third r-peak and the formation of the heaviest r-elements. Therefore slightly less material is shifted beyond Bi than in our earlier approach (Pfeiffer 1997, Cowan 1999), yielding somewhat lower Th/, /, and Th/-peak abundance ratios.
We have taken into account in a crude manner the main fission modes, spontaneous and neutron-induced fission, in the calculations. We make the simplifying assumption that everything beyond mass number 256 (, ) fissions completely, analogously to what we have done previously (, Cowan 1999). Furthermore, we have verified again that the known nuclei undergoing spontaneous fission for A $<$ 256 make no significant contribution to the mass range 232 $<$ A $<$ 255. For cases below the A = 256 region, new multi-dimensional fission barrier calculations (Möller 2001, and private communication 2006) show that spontaneous and neutron-induced fission have no effect, since the barrier heights of isotopes in the path (with $S_n$ $<$2 MeV) are $>$8 MeV. Also in the $\beta$-decay back to the valley of the mass parabola, the isobars still have fission barriers of 6-7 MeV. Hence, at least from the above model predictions, spontaneous and neutron-induced fission should be negligible. Our (site independent) waiting point approximation calculations, as well as dynamical network calculations (Farouqi 2006 for the SN II high-entropy wind scenario) indicate that the amount of matter actually involved in fission in the beyond mass number 256 is of the order of 1-5% of the total amount of matter. The only minor effect in this mass region instead comes from $\beta$-delayed fission.
However, most all of these calculations up to now neglect possible effects from $\beta$-delayed fission ($\beta$df), since they are generally believed to be unimportant. Nevertheless, we have tried to estimate the $\beta$df-rates of 20 potential $\beta$df-candidates between and using the above QRPA(Quasiparticle Random Phase Approximation) model for GT and GT+ff strength functions and the recent fission barriers of Mamdouh (1998) (based on the ETFSI-1 mass model) by following the simple approach outlined by Kodama & Takahashi (1975). We then have compared these results with the “complete-damping” and “WKB barrier penetration” calculations of $\beta$df-rates by Meyer (1989), who used the same QRPA model for GT strength functions and the (systematically lower) fission barriers of Howard & Möller (1980). This comparison convinces us that the effect of the $\beta$df-mode on the final Th and U abundances is small but may not be completely negligible. It may reduce the final Th abundance by about 8% and the final U abundance by approximately 3%. This would reduce our “fit2” Th/U abundance ratio from 1.568 (see Table \[tab2\]) to 1.479, which would still lie within the range of acceptable ratios. On the other hand, one might think of including the potential $\beta$df effects in the uncertainties of the $r$-chronometer ages. For example, this would add a further uncertainty of 0.55 Gy to the Th/Eu age of a star like 22, or to the Th/U age of stars like CS 31082-001 or 17.
**The Th/U Chronometer**
------------------------
We first examine the case for the Th/U chronometer.[^4] This chronometer ratio is considered ideal as the elements are -only and near each other in nuclear mass number. As we have noted, however, the one optically available uranium line is very weak in stellar spectra and blended with molecular lines, making its detection very difficult. Furthermore, for the Th/U ages to be meaningful they need to be determined in the context of (or constrained by) the abundances of other stable elements, particularly the -peak elements and Pb and Bi. We consider the conditions we have explored for which we obtain the best overall agreement with the observed patterns in halo stars. The mass model chosen for this purpose is the ETFSI-Q, yielding the ratios shown in Table \[tab2\].
There are only three halo stars for which there exist observational determinations of both the thorium and the uranium abundances, along with detailed abundances of many stable elements. The observed Th/U ratio for CS 31082-001 (Hill 2002) is Th/U = 8.7 (log(Th/U) = 0.94 $\pm$ 0.11), while that for 17 is Th/U = 7.6 (log(Th/U) = 0.88 $\pm$ 0.10). For these abundance ratios and our production ratio (Th/U)$_0$ = 1.557, the ages for the two halo stars CS 31082-001 and 17 are, respectively, 16.2 Gyr and 14.9 Gyr, both having uncertainties of approximately $\pm$3–3.5 Gyr arising from observational uncertainties. Very recently a new uranium detection has been made in HE 1523-0901 by Frebel et al. (2007), who find Th/U = 7.24 (log(Th/U) = 0.86 $\pm$ 0.15) for this star.
**The Th/Eu Chronometer**
-------------------------
An alternative to the Th/U chronometer is the ratio of the abundance of the long lived radioactive Th nucleus to the abundance of the stable product europium. Eu is formed almost entirely in the (, Simmerer 2004 and references therein) and is readily observable from the ground. The Th/U chronometer pair may be more robust and intrinsically more accurate, but determinations of both the Th and Eu abundances are available for many more halo stars than Th and U. It is important, therefore, to quantify the use of this Th/Eu chronometer and the reliability of age determinations resulting from it.
The production ratio (Th/Eu)$_0$ from our current study is 0.530 (Table \[tab2\]). This value is slightly higher than previous values used in our age calculations (, Cowan 1999) and results from better nuclear data and better fits to the stellar and solar abundance data (see discussion above). The observed Th/Eu ratio for 17 is 0.309 (Cowan 2002), leading to an implied age of 10.9 Gyr. Previously determined Th/Eu ages for other -rich halo stars (Truran 2002; Cowan & Sneden 2006; Ivans 2006) have in general been consistent and found to lie in the range $\simeq$ 10–15 Gyr. However, the observed abundance ratio for the star CS 31082-001 (Hill 2002) is significantly higher: Th/Eu = 0.603 (log $\epsilon$(Th/Eu) = $-$0.22 $\pm$ 0.07), which for our production ratio yields a very low age. This results from high U and Th abundances in this star relative to the abundances of elements in the range from Ba to the abundance peak, as determined by Hill . However, the lead abundance – resulting from the decay of Th and U – can provide a strong constraint on the abundance values and insight into the synthesis mechanisms of these radioactive actinides.
**The Actinide/Lead Abundance Ratio in Halo Stars**
---------------------------------------------------
Can the apparent single-valued abundance pattern in the elements Ba–Pt be expected to extend beyond the Pt–Pb peak? A further constraint on the robustness of the in the regime from the peak through the actinides can be provided by observations of the Th, U and Pb abundances in metal-poor stars.
The actinide chronometers Th, , and decay directly into the lead isotopes , , and , respectively. The lead abundances, therefore, provide a measure of the abundance levels of the chronometer nuclei. Consider the following quantitative measures. The Th/Pb ratio at the time of SS formation (Lodders 2003) is (Th/Pb) = 0.04399/1.4724 = 0.02988. The Th/Pb ratio in our calculations is significantly higher: (Th/Pb)$_{r-process}$ = 0.095. This implies first that the is responsible for of order 30% ($\approx$ 0.03/0.095) of the SS abundances of the heavy lead isotopes , , and . The Th and U chronometer abundances in extremely metal-poor -rich halo stars, therefore, also provide a measure of the expected level of Pb in these same stars. These lead levels, since they are the products of the decays of nuclei Th, , and , can in turn be utilized to constrain the ages of these stars as well. In this context, the high levels of Th and U cited for the halo star CS 31082-001 are inconsistent with the abundances of the platinum peak isotopes and the limits on the lead abundance for this star (Plez 2004). In contrast, the adoption of the upper limit on the Pb abundance for the star 17 (Cowan 2002) yields log $\epsilon$(Th/Pb) $>$ $-$1.48, and thus Th/Pb $>$ 0.033 – this value seems consistent with the abundances of the -peak elements and the expected production ratio. Also consistent are the Th/Pb ratios of 0.024 to 0.042 (depending upon adopting one of the very uncertain lead values) that have been reported for 22 (Sneden 2003). More work needs to be done to understand specifically how CS 31082-001 can have such an overabundance of the actinide chronometers Th and U with a correspondingly low Pb abundance. More generally it needs to be determined if there are other such stars.
Abundances of lead in the metal-poor halo stars, while difficult to derive and ideally requiring space-based observations, offer the promise of more refined (and constraining) chronometric age determinations.
CONCLUSIONS
===========
The rapid process is understood to be responsible for the synthesis of approximately half of all of the isotopes present in solar system matter in the mass region from approximately zinc through the actinides. While the general features of this process were identified in the classic papers by Burbidge (1957) and Cameron (1957), our current understanding of the remains incomplete. We have yet to cleanly identify which of the proposed astrophysical sites contribute significantly to the observed abundance pattern and we have yet to reconcile the apparent duplicity of sites with extant models for the operation of the in diverse astronomical environments. In this paper we have explored implications from parameterized waiting-point synthesis calculations.
Various observations suggest that contributions from two different astronomical sites or environments are required, for the mass regimes A $\lesssim$ 130 and A $\gtrsim$ 130 respectively. Here, we have tried to identify the mass number that represents the dividing line between these two (“weak” and “main”) contributions. In this context, our discussions in this paper and associated calculations lead to the following conclusions:
- The combined elemental and isotopic data on low metallicity -rich stars confirms a robust pattern extending over the (A $>$ 130-140) isotopic domain.
- We are able to reproduce the total SS abundances with a superposition of neutron number densities ranging from log $n_n$ = 20–28. Our calculations indicate that smaller neutron number densities, (log $n_n$ = 20–22) that characterize the weak , are required to reproduce the A = 80-100 SS isotopic abundances. We estimate that the division between “weak” and “main” density regimes occurs at log $n_n$ $\simeq$ 23.0 $\pm$ 0.5. More sophisticated network calculations (Farouqi 2006) appear to bear out our general results.
- The calculations that successfully generate the element pattern extending down to A=135 indicate that the production of is at a level $\sim$90% of its solar value, relative to the Ba-Pb region. Our calculations suggest that the / production ratio is quite compatible with the anticipated abundance pattern. In this context, our results imply that the dividing line (in mass number) between the “weak” and “main” components must necessarily fall below . We note, however, that observational limitations, so far, prevent, an exact identification of the elemental atomic number where the break occurs between the “main” and “weak” . (Iodine, for example, has not yet been observed in a metal-poor halo star.)
- We find that the isotopic fractions of barium, and the elemental Ba/Eu abundance ratios in -rich low metallicity stars can only be matched by computations in which the neutron densities are in the range 23 $\lesssim$ log $n_n$ $\lesssim$ 28, values typical of the main . For the main our calculations predict a solar value of the Ba isotopic ratio, $f_{odd}$ $\equiv$ \[N()+N()\]/N(Ba) $\simeq$ 0.47–0.50. The observed value of $f_{odd}$ = 0.3 $\pm$ 0.21 in one metal-poor halo star is consistent with the SS ratio. While the uncertainty is large and does not rule out some slight production, the elemental ratio of \[Ba/Eu\] in this star is in agreement with the SS only ratio. Since this star’s Ba/Eu elemental abundance ratio is also confirmed by our calculations, an -only origin for its Ba isotopes is strongly indicated.
- In the neutron number density ranges required for production of the observed solar/stellar -peak (A $\approx$ 200), the predicted abundances of inter-peak element hafnium (Z = 72, A $\approx$ 180) follow closely those of -peak elements (osmium through platinum) and lead. This suggests that abundance comparisons of hafnium to both rare-earth and -peak elements can shed further light on claims of invariance in the entire heavy end of the abundance pattern. Hafnium, observable from the ground and close in mass number to the -peak elements, could be utilized as a new nuclear chronometer pair Th/Hf for age determinations, particularly for cases where U is not detected. In the context of the calculations that reproduce the stable SS abundances, we have determined several important chronometric production ratios including Th/U, Th/Eu, Th/Pt and Th/-peak elements and Th/Hf. For example, the present computations suggest Th/ $\simeq$ 1.5 for this chronometer pair. These newly predicted chronometric ratios can then be employed to determine ages in stars where Th or U have been detected.
We thank Roberto Gallino, Peter M[ö]{}ller, Anna Frebel, and Ulli Ott for useful discussions, and an anonymous referee for helpful suggestions. This work has been supported in part by the Deutsche Forschungsgemeinschaft (DFG) under contract KR 806/13-1, and the Helmholtz Gemeinschaft under grant VH-VI-061 and by GSI (Univ. Mainz F+E-Vertrag (MZ/KLK)). Support was also provided by the National Science Foundation under grants AST 03-07279 (J.J.C.), AST 03-07495 (C.S.), and the Physics Frontier Center (JINA) PHY 02-16783 (J.W.T.), by the DOE under contract B523820 to the ASCI Alliances Center for Astrophysical Flashes (J.W.T.), and at the Argonne National Laboratory, which is operated under contract No. W-31-109-ENG-38 (J.W.T).
Aboussir, Y., Pearson, J. M., Dutta, A. K., and Tondeur, F. 1995, ADNDT, 61, 127
Anders, E., & Grevesse, N. 1989, , 53, 197
Aoki, W., 2003a, , 592, L67
Aoki, W., 2005, , 632, 611
Aoki, W., Honda, S., Beers, T. C., & Sneden, C. 2003b, , 586, 506
Arlandini, C., K[ä]{}ppeler, F., Wisshak, K., Gallino, R., Lugaro, M., Busso, M., & Straniero, O. 1999, , 525, 886
Audi, G., Bersillon, O., Blachot, J., & Wapstra, A. H. 2003, Nuc. Phys. A, 729, 3A
Barklem, P. S., 2005, , 439, 129
Burbidge, E.M., Burbidge, R.R., Fowler, W.A., & Hoyle, F. 1957, Rev. Mod. Physics, 29, 547
Burris, D. L., Pilachowski, C. A., Armandroff, T. A., Sneden, C., Cowan, J. J., & Roe, H. 2000, , 544, 302
Busso, M., Gallino, R., & Wasserburg, G.J. 1999, , 37, 239
Cameron, A.G.W. 1957, Chalk River Report CRL-41
Christlieb, N., 2004, , 428, 1027
Cowan, J. J., Sneden, C., Den Hartog, E. A., & Collier, J. 2006, in NASA Laboratory Astrophysics Workshop, in press
Cowan, J. J., Pfeiffer, B., Kratz, K.-L., Thielemann, F.-K., Sneden, C., Burles, S., Tytler, D., & Beers, T. C. 1999, , 521, 194
Cowan, J. J., 2002, , 572, 861
Cowan, J. J., & Sneden, C. 2006, , 440, 1151
Cowan, J. J., & Thielemann, F.-K., 2004, Phys. Today, 57, 47
Cowan, J. J., Thielemann, F.-K., & Truran, J. W. 1991, , 208, 267
Den Hartog, E. A., Lawler, J. E., Sneden, C., & Cowan, J. J. 2003, , 148, 543
Den Hartog, E. A., Herd, T. M., Lawler, J. E., Sneden, C., & Cowan, J. J., & Beers, T. C. 2005, ApJ, 619, 639
Den Hartog, E. A., Lawler, J. E., Sneden, C., & Cowan, J. J. 2006, ApJS, in press
Dillmann, I., et al. 2003, Phys. Rev. Lett., 91, 162503
Dobaczewski, J., et al. 1996, Phys. Rev. Lett. 72, 981
Farouqi, K., Freiburghaus, C., Kratz, K.-L., Pfeiffer, B., Rauscher, T., & Thielemann, F.-K. 2005, Nuc. Phys. A, 758, 631
Farouqi, K., Kratz, K.-L., Pfeiffer, B., Rauscher, T., & Thielemann, F.-K. 2006, AIP Conf. Proc. 819: Capture Gamma-Ray Spectroscopy and Related Topics, 819, 419
Frebel, A., Christlieb, N., Norris, J. E., Thom, C., Beers, T. C., & Rhee, J. 2007, ApJ, in press
Freiburghaus, C., Rosswog, S., & Thielemann, F.-K. 1999a, , 525, L121
Freiburghaus, C., Rembges, J.-F., Rauscher, T., Thielemann, F.-K., Kratz, K.-L., Pfeiffer, B., & Cowan, J. J. 1999b, , 516, 381
Goriely, S., Tondeur, F., Pearson, J. M. 2001, ADNDT, 77, 31
Goriely, S., Samyn, M., Pearson, J. M., & Onsi, M. 2005, Nuc. Phys. A, 750, 425
Gratton, R., & Sneden, C. 1994, , 287, 927
Hauge, [Ø]{}. 1972, , 27, 286
Hill, V. 2002, , 387, 560
Hilf, E. R., et al. 1976, CERN-Rep. 76-13, 142
Honda, S., Aoki, W., Kajino, T., Ando, H., Beers, T. C., Izumiura, H., Sadakane, K., & Takada-Hidai, M. 2004, , 607, 474
Howard, W. M., & M[ö]{}ller, P. 1980, Atomic Data and Nuclear Data Tables, 25, 219
Ivans, I. I., Simmerer, J., Sneden, C., Lawler, J. E., Cowan, J. J., Gallino, R., & Bisterzo, S. 2006, , 645, 613
Johnson, J. A. 2002, , 139, 219
Johnson, J. A., & Bolte, M. 2001, , 554, 888
K[ä]{}ppeler, F., Beer, H., & Wisshak, K. 1989, Rep. Prog. Phys., 52, 945
Kautzsch, T., et al. 2000, European Physical Journal A, 9, 201
Kodama, T., & Takahashi, K. 1975, Nuclear Physics A, 239, 489
Kratz, K.-L. 2001, Nuclear Physics A, 688, 308
Kratz, K.-L. 2006, AIP Conf. Proc. 819: Capture Gamma-Ray Spectroscopy and Related Topics, 819, 409
Kratz, K.-L., Bitouzet, J.-P., Thielemann, F.-K., Möller, P., & Pfeiffer, B. 1993, , 403, 216
Kratz, K.-L., Ostrowski, A. N., & Pfeiffer, B. 2005a, AIP Conf. Proc. 769: International Conference on Nuclear Data for Science and Technology, 769, 1356
Kratz, K.-L., Pfeiffer, Arndt, O., Hennrich, S., W[ö]{}hr, and the ISOLDE / IS333, IS378, IS393 Collaborations 2005b, Proc. ENAM 2004, Eur. Phys. J., A25, 633
Kratz, K.-L., Pfeiffer, B., & Thielemann, F.-K. 1998, Nuc. Phys. A, 630, 352
Kratz, K.-L., Pfeiffer, B., Thielemann, F.-K., & Walters, W. B. 2000, Hyperfine Interactions, 129, 185
Kratz, K.-L., Thielemann, F.-K., Hillebrandt, W., Möller, P., Harms, V., & Truran, J. W. 1988, J. Phys. G, 14, 331
Lambert, D.L. & Allende Prieto, C. 2002, , 335, 325
Lattimer, J.M., Mackie, F., Ravenhall, D.G., & Schramm, D.N. 1977, , 213, 225
Lawler, J. E., Den Hartog, E. A., Sneden, C., & Cowan, J. J. 2006a, ApJS, 162, 227
Lawler, J. E., Den Hartog, E. A., Labby, Z. E., Sneden, C., Cowan, J. J., & Ivans, I. I. 2006b, ApJ, in press
Lawler, J. E., Wickliffe, M. E., den Hartog, E. A., & Sneden, C. 2001, , 563, 1075
Lodders, K. 2003, , 591, 1220
Lunney, D., Pearson, J. M., & Thibault, C. 2003, Rev. Mod. Phys. 75, 1021
Magain, P. 1995, , 297, 686
Mamdouh, A., Pearson, J. M., Rayet, M., & Tondeur, F. 1998, Nuclear Physics A, 644, 389
Meyer, B., & Clayton, D. D. 2000, Sp. Sci. Rev., 92, 133
Meyer, B. S., Howard, W. M., Mathews, G. J., Takahashi, K., M[ö]{}ller, P., & Leander, G. A. 1989, , 39, 1876
M[ö]{}ller, P., Nix, J. R., & Kratz, K.-L. 1997, At. Data Nucl. Data Tables, 66, 131
Möller, P., Nix, J. R., Myers, W. D., & Swiatecki, W. J. 1995, ADNDT, 59, 185
M[ö]{}ller, P., & Randrup, J. 1990, , 514, 1
M[ö]{}ller, P., Pfeiffer, B., & Kratz, K.-L. 2003, , 67, 055802
M[ö]{}ller, P., Madland, D. G., Sierk, A. J., & Iwamoto, A. 2001, , 409, 785
Pearson, J. M. 2004, The r-Process: The Astrophysical Origin of the Heavy Elements and Related Rare Isotope Accelerator Physics (Singapore: World Scientific), p. 43
Pearson, J. M., Nayak, R. C., & Goriely, S. 1996, Phys. Lett., B387, 455
Pfeiffer, B., Kratz, K.-L., & M[ö]{}ller, P. 2002, Prog. Nuc. Energy, 41, 39
Pfeiffer, B., Kratz, K.-L., & Thielemann, F.-K. 1997, Z. Phys. A, 357, 23
Pfeiffer, B., Kratz, K.-L., Thielemann, F.-K., & Walters, W. B. 2001, Nucl. Phys. A, 693, 282
Plez, B., et al. 2004, , 428, L9
Qian, Y.Z., & Wasserburg, G. J. 2000, , 333, 77
Rauscher, T. 2004, in The r-Process: The Astrophysical Origin of the Heavy Elements and Related Rare Isotope Accelerator Physics (Singapore: World Scientific), p. 63
Rauscher, T., & Thielemann, F.-K. 2000, ADNDT 75, 1
Rikovska-Stone, J. 2005, JPhG, 31, R.211
Rosswog, S., Liebendorfer, M., Thielemann, F.-K., Davies, M.B., Benz, W., & Piran, T. 1999, , 341, 499
Samyn, M., Goriely, S., Bender, M., & Pearson, J. M. 2004, Phys. Rev. C, 70, 044309
Simmerer, J., Sneden, C., Cowan, J. J., Collier, J., Woolf, V., & Lawler, J. E. 2004, , 617, 1091
Sneden, C., & Cowan, J. J. 2003, Science, 299, 70
Sneden, C., 2002, , 566 L25
Sneden, C., 2003, , 591, 936
Spite, M., & Spite, F. 1978, , 67, 23
Takahashi, K., Witti, J., & Janka, H.-T. 1994, , 286, 857
Thielemann, F.-K., Kratz, K.-L., Pfeiffer, B., Rauscher, T., van Wormer, L., & Wiescher, M. 1994, Nucl. Phys. A, 570, 329
Travaglio, C., Gallino, R., Arnone, E., Cowan, J., Jordan, F., & Sneden, C. 2004, , 601, 864
Truran, J. W., Cowan, J. J., Pilachowski, C. A., & Sneden, C. 2002, , 114, 1293
Wanajo, S., Goriely, S., Samyn, M., & Itoh, N. 2004, , 606, 1057
Wasserburg, G.J., Busso, M., & Gallino, R. 1996, , 466, L109
Westin, J., Sneden, C., Gustafsson, B., & Cowan, J.J. 2000, , 530, 783
Woosley, S.E., Wilson, J.R., Mathews, G.J., Hoffman, R.D., & Meyer, B.S. 1994, , 433, 229
[^1]: We adopt the usual spectroscopic notations that \[A/B\] $\equiv$ log$_{\rm 10}$(N$_{\rm A}$/N$_{\rm B}$)$_{\rm star}$ – log$_{\rm 10}$(N$_{\rm A}$/N$_{\rm B}$)$_{\odot}$, and that log $\epsilon$(A) $\equiv$ log$_{\rm 10}$(N$_{\rm A}$/N$_{\rm H}$) + 12.0, for elements A and B. Also, metallicity will be assumed here to be equivalent to the stellar \[Fe/H\] value.
[^2]: We have also performed computations that employ a “Zr seed” beyond Z = 40, N = 50, and the results are similar.
[^3]: Two other very minor but stable isotopes, and , are products of the $p$-process and not relevant to the present study.
[^4]: For this purpose we approximate the uranium abundance as that of $^{238}$U, given the significantly shorter half-life of $^{235}$U ($\tau_{1/2}$ = 7.038x10$^8$ years) compared to $^{238}$U ($\tau_{1/2}$ = 4.468x10$^9$ years). The thorium ($^{232}$Th) half-life is $\tau_{1/2}$ = 1.405x10$^{10}$ years.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove a variety results on tensor product factorizations of finite dimensional Hopf algebras (more generally Hopf algebras satisfying chain conditions in suitable braided categories). The results are analogs of well-known results on direct product factorizations of finite groups (or groups with chain conditions) such as Fitting’s Lemma and the uniqueness of the Krull-Remak-Schmidt factorization. We analyze the notion of normal (and conormal) Hopf algebra endomorphisms, and the structure of endomorphisms and automorphisms of tensor products. The results are then applied to compute the automorphism group of the Drinfeld double of a finite group in the case where the group contains an abelian factor. (If it doesn’t, the group can be calculated by results of the first author.)'
address: |
Institut de Math[é]{}matiques de Bourgogne, UMR 5584 du CNRS\
Universit[é]{} de Bourgogne\
Facult[é]{} des Sciences Mirande\
9 avenue Alain Savary\
BP 47870 21078 Dijon Cedex\
France
author:
- Marc Keilberg
- Peter Schauenburg
bibliography:
- 'andere.bib'
- 'arxiv.bib'
- 'eigene.bib'
- 'mathscinet.bib'
title: On tensor factorizations of Hopf algebras
---
[^1]
Introduction {#introduction .unnumbered}
============
The larger part of this paper is concerned with general results on Hopf algebras in braided categories generalizing well-known results from the theory of finite groups (or groups with chain conditions), such as Fitting’s lemma, the Krull-Remak-Schmidt decomposition, and a description of endomorphisms and automorphisms of products of Hopf algebras. The last section deals with the description of the automorphism group of the Drinfeld double ${\mathcal D}(G)$ of a finite group $G$. This last problem was the starting point of our work.
In the case that $G$ has no non-trivial abelian direct factors, a complete description of the automorphisms was given in [@K14]. The case when $G$ has such an abelian factor was left open. We will write such a group as $G=C\times H$, where $H$ has no non-trivial abelian direct factors and $C$ is abelian. In this case we naturally have that ${\mathcal D}(G)\cong {\mathcal D}(C){\otimes}{\mathcal D}(H)$ is a tensor product of Hopf algebras.
Thus, we are naturally led to analyze endomorphisms and automorphisms of a tensor product of two Hopf algebras. In [@BidCurMcC:ADPFG; @Bid:ADPFG2] an analysis of the automorphisms of direct products of groups was provided. The basic idea is to describe such automorphisms by a matrix of morphisms between the factors. The machinery of normal group endomorphisms and Fitting’s lemma then allows one to deduce conditions on the various morphisms from conditions on the factors. For example, when the two factors have no common direct factors, then the diagonal terms of the matrix have to be automorphisms. In \[sec:autom-tens-prod\] we derive suitably analogous results for tensor product Hopf algebras. Before this can be done, however, we have to carry over to our Hopf algebraic setting some basic notions and classical results from group theory. In \[sec:comm-cocomm-morph\] we develop the terminology of commuting morphisms (for groups these are just morphisms whose images commute) and dually of cocommuting morphisms, and in \[sec:normal-endomorphisms\] the notions of normal and conormal Hopf endomorphisms. The analog of Fitting’s lemma which will produce tensor product decompositions from binormal endomorphisms and thus, under suitable circumstances, common tensor factors from certain endomorphisms of tensor products, will be proved in \[sec-fitting\]. An important application of Fitting’s lemma in group theory is the uniqueness of the Krull-Remak-Schmidt decomposition, which we prove in \[sec:krull-remak-schmidt\]. Extensions of the Krull-Remak-Schmidt decomposition were studied previously in [@Bur:CFPHA] for decompositions of semisimple Hopf algebras into simple semisimple tensor factors. By contrast our techniques make no use of semisimplicity but only of chain conditions. It is also worth noting that the Krull-Remak-Schmidt result shows that our results are specific to Hopf algebras and cannot be readily generalized to finite or even fusion tensor categories. In fact Müger [@Mue:SMC] gives an example where the factors in the decomposition of a fusion category into prime factors are not unique.
In fact the above results on the structure theory of finite dimensional Hopf algebras over a field ${\mathbbm k}$ will be developed in greater generality for Hopf algebras in braided abelian tensor categories that fulfill chain conditions on Hopf subalgebras and quotient Hopf algebras. Apart from the fact that the results will thus immediately apply to objects like super-Hopf algebras, for some purposes the categorical setting is simply very natural, since it allows treating mutually dual notions like normality and conormality or ascending and descending chain conditions on the same footing. If the braiding of the base category is not a symmetry, then some of our basic objects of study may be hard to come by: It is well-known that the tensor product of two Hopf algebras in a braided monoidal category can only be formed if the two factors are “unbraided”, that is, the braiding between them behaves like a symmetry. On the other hand, some of our results imply that tensor product decompositions have to exist in certain situations. Thus these results also imply that the braiding has to be “partially trivial”. For example, if non-nilpotent normal endomorphisms of a Hopf algebra exist, they have to be isomorphisms by Fitting’s lemma unless the braiding is partially trivial. An automorphism between a tensor product of nonisomorphic Hopf algebras (necessarily “unbraided” between each other) has to induce automorphisms on the factors, unless the braiding is partially trivial on one of the factors.
deals with some technical issues raised by our categorical framework. In preparation for Fitting’s lemma we decompose a Hopf algebra with chain conditions, for which a Hopf algebra endomorphism is given, into a Radford biproduct (in the generalized braided version due to Bespalov and Drabant [@BesDra:HBMCMBC]). A technical result on (co)invariants under Hopf algebra endomorphisms has some bearing on the notions of epimorphisms and monomorphisms studied notably for infinite dimensional Hopf algebras in [@Chi:EMHA].
In \[sec:appl-doubl-groups\] we present the application of the general results on the structure of finite Hopf algebras and their automorphisms to the study of automorphisms of Drinfeld doubles of groups. Letting $G=C\times H$ as before, taking the field to be the complex numbers, and defining $\widehat{H}$ to be the group of linear characters of $H$, then under the isomorphisms ${\mathcal D}(C)\cong\mathbb{C}(\widehat{C}\times C)$ and $\widehat{C}\times C\cong C^2$ the result can be stated as $${\operatorname{Aut}}({\mathcal D}(C\times H)) \cong \ \begin{pmatrix} \operatorname{Aut}(C^2)&{\operatorname{Hom}}({\mathcal D}(H),\mathbb{C}C^2)\\
\operatorname{Hom}(C^2,\widehat{H}\times Z(H))&{\operatorname{Aut}}({\mathcal D}(H))\end{pmatrix}.$$ The only term not explicitly determined by [@K14] or standard methods for finite abelian groups is ${\operatorname{Hom}}({\mathcal D}(H),\mathbb{C}C^2)$. In this case the morphisms can be described entirely in terms of group homomorphisms and central subgroups of $G$ satisfying certain relations [@ABM:CBPHA; @K14], so the description is not a significant problem. In \[ex:dihedrals\] we completely describe ${\operatorname{Aut}}({\mathcal D}(D_{2n}))$ where $D_{2n}$ is the dihedral group of order $2n$, for the case $n\equiv 2\bmod 4$ and $n>2$. This is precisely when there is an isomorphism $D_{2n}\cong {\mathbb Z}_2\times D_n$. From this we can easily provide a formula for the order of ${\operatorname{Aut}}({\mathcal D}(D_{2n}))$. In particular we find that ${\operatorname{Aut}}({\mathcal D}(D_{12}))$ has order $1152=2^7 3^2$.
Preliminaries and notation
==========================
Throughout the paper, ${\mathcal B}$ is an abelian braided tensor category with braiding ${\tau}$; we will assume that ${\mathcal B}$ is strict, backed up by the well-known coherence theorems. Algebras, coalgebras, bialgebras, Hopf algebras are in ${\mathcal B}$. All undecorated ${\operatorname{Hom}}$s, ${\operatorname{End}}$s, etc. will be for morphisms of Hopf algebras or groups, as appropriate. We will use the following graphical notations to do computations in ${\mathcal B}$: The braiding is $${\tau}_{VW}=\gbeg23\got1V\got1W\gnl\gbr\gnl\gob1W\gob1V\gend\qquad
\text{ and }\qquad
{\tau}_{VW}{^{-1}}=\gbeg23\got1W\got1V\gnl\gibr\gnl\gob1V\gob1W\gend.$$ We shall say that the objects $V$ and $W$ are *unbraided* if ${\tau}_{VW}={\tau}_{WV}{^{-1}}$.
Multiplication and unit of an algebra $A$, and comultiplication and counit of a coalgebra $C$ are $$\nabla_A=\gbeg 23\got1A\got1A\gnl\gmu\gnl\gob2A\gend,
\qquad
\eta_A=\gbeg13\gnl\gu1\gnl\gob1A\gend,
\qquad
\Delta_C=\gbeg23\got2C\gnl\gcmu\gnl\gob1C\gob1C\gend,
\qquad
{\varepsilon}_C=\gbeg13\got1C\gnl\gcu1\gnl\gend.$$ The antipode of a Hopf algebra and, if it exists, its inverse are $$S=\gbeg15\got1H\gnl\gcl1\gnl\gmp+\gnl\gcl1\gnl\gob1H\gend\qquad\text{ and
}\qquad
S{^{-1}}=\gbeg15\got1H\gnl\gcl1\gnl\gmp-\gnl\gcl1\gnl\gob1H\gend.$$
In order to have a straightforward notion of Hopf subalgebra and quotient Hopf algebra of a given Hopf algebra, we shall assume that tensor products in ${\mathcal B}$ are exact.
An object in ${\mathcal B}$ satisfies the ascending chain condition on subobjects if and only if it satisfies the descending chain condition on quotient objects, by which we understand the descending chain condition on subobjects in the opposite category. For Hopf algebras we will use the descending chain conditions on Hopf-subalgebras and on quotient Hopf algebras. This is done since Hopf algebras which are artinian as algebras are finite dimensional [@LiuZh07:AHFD]. When a Hopf algebra satisifies the descending chain conditions on both Hopf-subalgebras and quotient Hopf algebras, we simply say that it satisfies both chain conditions.
If $f\colon H\to G$ is a Hopf algebra morphism, we define the right and left $f$-coinvariant subobjects of $H$ as being the equalizers $$\begin{gathered}
\xymatrix{0\ar[r]&{{H}{^{\operatorname{co} f}}}\ar[r]&H\ar@<0.5ex>[rr]^-{(H{\otimes}f)\Delta}\ar@<-0.5ex>[rr]_-{H{\otimes}\eta}&&H{\otimes}G}\\
\xymatrix{0\ar[r]&{{{^{\operatorname{co} f}}{H}}}\ar[r]&H\ar@<0.5ex>[rr]^-{(f{\otimes}H)\Delta}\ar@<-0.5ex>[rr]_-{\eta{\otimes}H}&&G{\otimes}H}\end{gathered}$$ And dually, the left and right invariant quotients by coequalizers $$\begin{gathered}
\xymatrix{H{\otimes}G\ar@<0.5ex>[rr]^-{\nabla(f{\otimes}G)}\ar@<-0.5ex>[rr]_-{{\varepsilon}{\otimes}G}&&G\ar[r]&H\backslash G\ar[r]&0}\\
\xymatrix{G{\otimes}H\ar@<0.5ex>[rr]^-{\nabla(G{\otimes}f)}\ar@<-0.5ex>[rr]_-{G{\otimes}{\varepsilon}}&&G\ar[r]&G/H
\ar[r]&0}\end{gathered}$$
We note that the coinvariant subobjects are subalgebras of $H$, and the invariant quotients are quotient coalgebras of $G$.
We will say a Hopf algebra is abelian if it is both commutative and cocommutative. In the category of vector spaces over a field ${\mathbbm k}$ of characteristic zero, such Hopf algebras are precisely group algebras of abelian groups, up to a separable field extension [@Mon:HAAR Theorem 2.3.1]. We will say a Hopf algebra is non-abelian when it is not abelian.
Commuting and Cocommuting morphisms {#sec:comm-cocomm-morph}
===================================
In this section, we formulate an obvious commutation condition for morphisms to an algebra (for ordinary algebras it just means that elements in the respective images commute) and its dual, and we collect equally obvious consequences that will be useful in later calculations. We note that for each and every fact on Hopf algebras in a braided category there is a dual fact. We will not always state, but still freely use the duals of our statements
Let $A$ be an algebra, $V,W\in{\mathcal B}$, and $f\colon V\to A,g\colon W\to A$ morphisms in ${\mathcal B}$. We say that $f$ and $g$ multiplication commute and write $f{\curlyvee}g$ if $\nabla(g{\otimes}f)=\nabla(f{\otimes}g){\tau}(=\nabla{\tau}(g{\otimes}f))$, or graphically$${ \gbeg23
\gcl1\gcl1\gnl
\gbmp{g}\gbmp{f}\gnl
\gmu\gend
=
\gbeg23
\gbr\gnl
\gbmp{f}\gbmp{g}\gnl
\gmu\gend}=\gbeg23\gbmp g\gbmp f\gnl\gbr\gnl\gmu\gend.$$
Dually, two morphisms $f\colon C\to V$ and $g\colon C\to W$ from a coalgebra $C$ in ${\mathcal B}$ comultiplication commute, or cocommute for short, and write $f{\curlywedge}g$ if $${ \gbeg23
\gcmu\gnl
\gbmp{g}\gbmp{f}\gnl
\gcl1\gcl1\gend
=
\gbeg23
\gcmu\gnl
\gbmp{f}\gbmp{g}\gnl
\gbr\gend}=\gbeg23\gcmu\gnl\gbr\gnl\gbmp g\gbmp f\gend.$$
We say that $f,g$ bicommute if both $f{\curlyvee}g$ and $f{\curlywedge}g$.
If $A$ and $B$ are algebras in ${\mathcal B}$, then the natural maps $f\colon
A\to A{\otimes}B$ and $g\colon B\to A{\otimes}B$ satisfy $f{\curlyvee}g$, but they only satisfy $g{\curlyvee}f$ if $A$ and $B$ are unbraided: In fact $$\begin{aligned}
\gbeg25
\got1A\got1B\gnl
\gcl1\gcl1\gnl
\gbmp f\gbmp g\gnl
\gmu\gnl
\gob2{A{\otimes}B}\gend
&=
\gbeg25
\got1A\got1B\gnl
\gcl3\gcl3\gnl\gnl\gnl
\gob1A\gob1B\gend
&\text{and}&&
\gbeg25
\got1B\got1A\gnl
\gcl1\gcl1\gnl
\gbmp g\gbmp f\gnl
\gmu\gnl
\gob2{A{\otimes}B}\gend
&=
\gbeg25
\got1B\got1A\gnl
\gcl1\gcl1\gnl
\gbr\gnl
\gcl1\gcl1\gnl
\gob1A\gob1B\gend.\end{aligned}$$
\[usefulsimplecommutationfacts\] Let $A$ be an algebra, $C$ a coalgebra, and $U,V,W,X,Y$ objects in ${\mathcal B}$.
1. Let $f\colon U\to A$, $g\colon V\to A$ and $h\colon W\to A$.
1. If $f{\curlyvee}g$ and $f{\curlyvee}h$, then $f{\curlyvee}(\nabla(g{\otimes}h))$.
2. If $f{\curlyvee}h$ and $g{\curlyvee}h$, then $(\nabla(f{\otimes}g)){\curlyvee}h$.
3. If $f{\curlyvee}g$, then $fa{\curlyvee}gb$ for any $a\colon X\to U$ and $b\colon Y\to V$.\[compositescommute\]
2. Let $f,g,h\colon C\to A$.
1. If $f{\curlyvee}g$ and $f{\curlyvee}h$ then $f{\curlyvee}(g*h)$.
2. If $f{\curlyvee}h$ and $g{\curlyvee}h$ then $(f*g){\curlyvee}h$.
3. If $f{\curlyvee}g$ and $g{\curlywedge}f$, then $f*g=g*f$.
3. Let $f,g\colon C\to A$.
1. If $C$ is a bialgebra, $f,g$ are algebra morphisms, and $f{\curlyvee}g$, then $f*g$ is an algebra morphism.
2. If $A,C$ are bialgebras, $f,g$ are bialgebra morphisms, $f{\curlyvee}g$ and $f{\curlywedge}g$, then $f*g$ is a bialgebra morphism.
3. If $A$ is a bialgebra, $C$ a Hopf algebra, and $f,g$ are unital coalgebra morphisms, then $f{\curlywedge}g$ $\iff$ $f*g$ is a coalgebra morphism.
Note that $f{\curlyvee}g$ is not necessarily equivalent to $g{\curlyvee}f$ in the braided setting. The first part of the following result says, however, that the two properties are equivalent for Hopf algebras with sufficiently well-behaved antipodes. On the other hand, the second part says that if both properties are fulfilled then either the braiding is close to being a symmetry, or the morphisms are close to being trivial.
Let $H,K,$ and $A$ be Hopf algebras, and $f\colon H\to A$, $g\colon K\to A$ Hopf algebra morphisms.
1. If $f{\curlyvee}g$, and if the antipode of $A$ is a monomorphism or the antipodes of $H$ and $K$ are epimorphisms, then $g{\curlyvee}f$.
2. If $f{\curlyvee}g$ and $g{\curlyvee}f$, then $$\gbeg25
\got1H\got1K\gnl
\gcl1\gcl1\gnl
\gbmp f\gbmp g\gnl
\gbr\gnl
\got1A\got1A\gend
=
\gbeg25
\got1H\got1K\gnl
\gcl1\gcl1\gnl
\gbmp f\gbmp g\gnl
\gibr\gnl
\got1A\got1A\gend$$
For the first claim, we calculate $$\gbeg26
\got1H\got1K\gnl
\gcl1\gcl1\gnl
\gmp+\gmp+\gnl
\gbmp f\gbmp g\gnl
\gmu\gnl
\gob2A\gend
=
\gbeg26
\got1H\got1K\gnl
\gcl1\gcl1\gnl
\gbmp f\gbmp g\gnl
\gmp+\gmp+\gnl
\gmu\gnl
\gnl
\gob2A\gend
=
\gbeg37
\got1H\gvac1\got1K\gnl
\gcl1\gvac1\gcl1\gnl
\gbmp f\gvac1\gbmp g\gnl
\gibbr31\gnl
\gwmu3\gnl
\gvac1\gmp+\gnl
\gob3A\gend
=
\gbeg36
\got1H\gvac1\got1K\gnl
\gcl1\gvac1\gcl1\gnl
\gbmp f\gvac1\gbmp g\gnl
\gwmu3\gnl
\gvac1\gmp+\gnl
\gob3A\gend
=
\gbeg27
\got1H\got1K\gnl
\gcl1\gcl1\gnl
\gbmp f\gbmp g\gnl
\gmp+\gmp+\gnl
\gbr\gnl
\gmu\gnl
\gob2A\gend
=
\gbeg27
\got1H\got1K\gnl
\gcl1\gcl1\gnl
\gmp+\gmp+\gnl
\gbmp f\gbmp g\gnl
\gbr\gnl
\gmu\gnl
\gob2A\gend,$$ which implies $g{\curlyvee}f$ if the antipodes of $H$ and $K$ are epimorphisms. A similar argument shows the same if the antipode of $A$ is a monomorphism.
We now turn to the second claim.
First, we note that $$\begin{gathered}
\gbeg46
\got2H\got2K\gnl
\gcmu\gcmu\gnl
\gcl1\gibr\gcl1\gnl
\gbmp f\gbmp g\gbmp f\gbmp g\gnl
\gmu\gmu\gnl
\gob2A\gob2A\gend
=
\gbeg47
\got2H\got2K\gnl
\gcmu\gcmu\gnl
\gcl1\gibr\gcl1\gnl
\gbmp f\gbmp g\gbmp f\gbmp g\gnl
\gibr\gibr\gnl
\gmu\gmu\gnl
\gob2A\gob2A\gend
=
\gbeg58
\gvac1\got1H\gvac1\got1K\gnl
\gvac1\gcl1\gvac1\gcl1\gnl
\gvac1\gbmp f\gvac1\gbmp g\gnl
\gwcmh324\gvac{-1}\gwcmh324\gnl
\gcn1122\gibbrh2124\gcn0122\gnl
\gibbrh2124\gibbrh2124\gnl
\gwmuh324\gvac{-1}\gwmuh324\gnl
\gvac1\gob1A\gvac1\gob1A\gend
=
\gbeg59
\gvac1\got1H\gvac1\got1K\gnl
\gvac1\gcl1\gvac1\gcl1\gnl
\gvac1\gbmp f\gvac1\gbmp g\gnl
\gibbrh4237\gnl
\gnl
\gwcmh324\gvac{-1}\gwcmh324\gnl
\gcn1122\gbbrh2124\gcn0122\gnl
\gwmuh324\gvac{-1}\gwmuh324\gnl
\gvac1\gob1A\gvac1\gob1A\gend
=\\\\=
\gbeg27
\got1H\got1K\gnl
\gcl1\gcl1\gnl
\gbmp f\gbmp g\gnl
\gibr\gnl
\gmu\gnl
\gcmu\gnl
\gob1A\gob1A\gend
=
\gbeg26
\got1H\got1K\gnl
\gcl1\gcl1\gnl
\gbmp f\gbmp g\gnl
\gmu\gnl
\gcmu\gnl
\gob1A\gob1A\gend
=
\gbeg57
\gvac1\got1H\gvac1\got1K\gnl
\gvac1\gcl1\gvac1\gcl1\gnl
\gvac1\gbmp f\gvac1\gbmp g\gnl
\gwcmh324\gvac{-1}\gwcmh324\gnl
\gcn1122\gbbrh2124\gcn0122\gnl
\gwmuh324\gvac{-1}\gwmuh324\gnl
\gvac1\gob1A\gvac1\gob1A\gend
=
\gbeg46
\got2H\got2H\gnl
\gcmu\gcmu\gnl
\gcl1\gbr\gcl1\gnl
\gbmp f\gbmp g\gbmp f\gbmp g\gnl
\gmu\gmu\gnl
\gob2A\gob2A\gend\end{gathered}$$ In other words $$\gbeg46
\got2H\got2H\gnl
\gcmu\gcmu\gnl
\gcl1\gdnot X\glmptb\grmptb\gcl1\gnl
\gbmp f\gbmp g\gbmp f\gbmp g\gnl
\gmu\gmu\gnl
\gob2A\gob2A\gend$$ does not depend on the choice of $X\in\left\{\gbeg21\gbr\gend,\gbeg21\gibr\gend\right\}$. But since $f\circ S$ and $g\circ S$ are convolution inverse to $f$ and $g$, respectively, we have $$\gbeg68
\got3H\got3K\gnl
\gwcmh314\gwcmh325\gnl
\gcl1\gcmu\gcmu\gcl1\gnl
\gmp+\gcl1\gdnot X\glmptb\grmptb\gcl1\gmp+\gnl
\gbmp f\gbmp f\gbmp g\gbmp f\gbmp g\gbmp g\gnl
\gcl1\gmu\gmu\gcl1\gnl
\gwmuh314\gwmuh325\gnl
\gob3A\gob3A\gend
=
\gbeg68
\got3H\got3K\gnl
\gwcmh325\gwcmh314\gnl
\gcmu\gcl1\gcl1\gcmu\gnl
\gmp+\gcl1\gdnot X\glmptb\grmptb\gcl1\gmp+\gnl
\gbmp f\gbmp f\gbmp g\gbmp f\gbmp g\gbmp g\gnl
\gmu\gcl1\gcl1\gmu\gnl
\gwmuh325\gwmuh314\gnl
\gob3A\gob3A\gend
=
\gbeg46
\got2H\got2K\gnl
\gcmu\gcmu\gnl
\gcu1\gdnot X\glmptb\grmptb\gcu1\gnl
\gu1\gbmp g\gbmp f\gu1\gnl
\gmu\gmu\gnl
\gob2A\gob2A\gend
=
\gbeg26
\got1H\got1K\gnl
\gcl1\gcl1\gnl
\gdnot X\glmptb\grmptb\gnl
\gbmp g\gbmp f\gnl
\gcl1\gcl1\gnl
\gob1A\gob1A\gend$$ That the latter expression does not depend on the choice of $X$ is the claim.
As special cases one recovers two known facts that show how badly usual Hopf algebra constructions behave in a “truly braided” tensor category: A Hopf algebra cannot be commutative (or cocommutative) as a (co)algebra in $\mathcal B$ unless the braiding on the Hopf algebra is an involution [@Sch:BHABC], and the tensor product of two Hopf algebras cannot be a Hopf algebra unless the two factors are unbraided.
Normal endomorphisms {#sec:normal-endomorphisms}
====================
Recall that the left adjoint action and the left coadjoint coaction of a Hopf algebra $H$ on itself are $$\begin{aligned}
\gbeg25
\got1H\got1H\gnl
\gcl1\gcl1\gnl
\gdnot{\operatorname{ad}}\glmpt\grmptb\gnl
\gvac1\gcl1\gnl
\gvac1\gob1H\gend
&=
\gbeg37
\got2H\got1H\gnl
\gcmu\gcl1\gnl
\gcl3\gbr\gnl
\gvac1\gcl1\gmp+\gnl
\gvac1\gmu\gnl
\gwmuh314\gnl
\gob3H\gend
&
\gbeg25
\gvac1\got1H\gnl
\gvac1\gcl1\gnl
\gdnot{\operatorname{coad}}\glmpb\grmptb\gnl
\gcl1\gcl1\gnl
\gob1H\gob1H\gend
&=
\gbeg37
\got3H\gnl
\gwcmh314\gnl
\gcl3\gcmu\gnl
\gvac1\gcl1\gmp+\gnl
\gvac1\gbr\gnl
\gmu\gcl1\gnl
\gob2H\gob1H\gend\end{aligned}$$
We note that the adjoint action is characterized by a twisted commutativity condition: $$\label{eq:3}
\gbeg36
\got 2H\got1H\gnl
\gcmu\gcl1\gnl
\gcl1\gbr\gnl
\gdnot{\operatorname{ad}}\glmpt\grmptb\gcl1\gnl
\gvac1\gmu\gnl
\gvac1\gob2H\gend
=
\gbeg25
\got1H\got1H\gnl
\gcl1\gcl1\gnl
\gmu\gnl
\gcn2122\gnl
\gob2H\gend$$
Let $f\colon H\to H$ be a morphism in ${\mathcal B}$, with $H$ a Hopf algebra.
1. $f$ is normal if it is left $H$-linear with respect to the adjoint action.
2. $f$ is conormal if it is left $H$-colinear with respect to the coadjoint coaction.
3. $f$ is binormal if it is both normal and conormal.
For group algebras considered in the category of $\mathbb{C}$-vector spaces, the definition of a normal morphism agrees with the one used in group theory [@Rot:Book]. Since group algebras are cocommutative, every group endomorphism is trivially conormal. We will be primarily concerned with normal algebra morphisms, conormal coalgebra morphisms, and binormal bialgebra morphisms.
Let $f\colon H\to H$ be an endomorphism of the Hopf algebra $H$.
1. The following are equivalent:
1. $f$ is normal.\[equiv:normal\]
2. $f{\curlyvee}((fS)*\operatorname{\operatorname{id}}_H)$.\[equiv:comm\]
3. $(fS)*\operatorname{\operatorname{id}}_H$ is an algebra morphism.\[equiv:algmor\]
2. The following are equivalent:
1. $f$ is binormal.
2. $f{\curlywedge}((fS)*\operatorname{\operatorname{id}}_H)$ and $f{\curlyvee}((fS)*\operatorname{\operatorname{id}}_H)$.
3. $(fS)*\operatorname{\operatorname{id}}_H$ is a bialgebra morphism
We only show the first part. For the equivalence of \[equiv:comm\] and \[equiv:algmor\] we apply the bijection $${\mathcal B}(H{\otimes}H,H)\ni T\mapsto
\gbeg35
\got1H\got2H\gnl
\gcl1\gcmu\gnl
\gdnot T\glmpt\grmptb\gmp+\gnl
\gvac1\gmu\gnl
\gvac1\gob2H\gend
\in{\mathcal B}(H{\otimes}H,H)$$ to the two sides of the equation expressing multiplicativity of $g:=fS*\operatorname{\operatorname{id}}$. We get $$\gbeg39
\got1H\got2H\gnl
\gcl1\gcmu\gnl
\gmu\gcl1\gnl
\gcmu\gmp+\gnl
\gmp+\gcl2\gcl3\gnl
\gbmp f\gnl
\gmu\gnl
\gwmuh325\gnl
\gob3H\gend
=
\gbeg5{10}
\got3H\got1H\gnl
\gvac1\gcl1\gvac1\gcmu\gnl
\gwcmh324\gvac{-1}\gwcmh324\gvac{-1}\gcn1511\gnl
\gcn1122\gbbrh2124\gcn0122\gnl
\gwmuh324\gvac{-1}\gwmuh324\gnl
\gvac1\gmp+\gvac1\gcl2\gnl
\gvac1\gbmp f\gnl
\gvac1\gwmu3\gmp+\gnl
\gvac2\gwmu3\gnl
\gvac2\gob3H\gend
=
\gbeg59
\got2H\got3H\gnl
\gcmu\gwcmh314\gnl
\gmp+\gcl2\gmp+\gcmu\gnl
\gbmp f\gvac1\gbmp f\gcl1\gmp+\gnl
\gcl1\gbr\gmu\gnl
\gbr\gwmuh314\gnl
\gmu\gvac1\gcl1\gnl
\gwmuh527\gnl
\gob5H
\gend
=
\gbeg26
\got1H\got1H\gnl
\gcl1\gmp+\gnl
\gbr\gnl
\gbmp f\gbmp g\gnl
\gmu\gnl
\gob2H\gend$$ and $$\gbeg36
\got1H\got2H\gnl
\gcl1\gcmu\gnl
\gbmp g\gbmp g\gmp+\gnl
\gmu\gcl1\gnl
\gwmuh325\gnl
\gob3H\gend
=
\gbeg36
\got1H\got2H\gnl
\gcl1\gcmu\gnl
\gbmp g\gbmp g\gmp+\gnl
\gcl1\gmu\gnl
\gwmuh314\gnl
\gob3H\gend
=
\gbeg25
\got1H\got1H\gnl
\gcl1\gmp+\gnl
\gbmp g\gbmp f\gnl
\gmu\gnl
\gob2H\gend$$ that is, the two sides of \[equiv:comm\], up to composition with the isomorphism $H{\otimes}S$.
For the equivalence of \[equiv:normal\] and \[equiv:comm\], we apply the bijection $${\mathcal B}(H{\otimes}H,H)\ni T\mapsto
\gbeg48
\got3H\got1H\gnl
\gwcmh314\gcl2\gnl
\gmp+\gcmu\gnl
\gbmp f\gcl1\gbr\gnl
\gcl2\gdnot T\glmptb\grmpt\gcl1\gnl
\gvac1\gwmu3\gnl
\gwmu3\gnl
\gob3H\gend
\in{\mathcal B}(H{\otimes}H,H)$$ to the two sides of \[equiv:normal\] to get $$\gbeg5{12}
\got3H\gvac1\got1H\gnl
\gwcm3\gvac1\gcl2\gnl
\gcl2\gwcmh325\gnl
\gvac1\gcmu\gbr\gnl
\gmp+\gcl3\gbr\gcl5\gnl
\gbmp f\gvac1\gcl1\gmp+\gnl
\gcl4\gvac1\gmu\gnl
\gvac1\gwmuh314\gnl
\gvac2\gbmp f\gnl
\gvac2\gwmu3\gnl
\gwmu4\gnl
\gob4H\gend
=
\gbeg5{11}
\got4H\got1H\gnl
\gwcmh426\gcl2\gnl
\gcmu\gcmu\gnl
\gmp+\gcl3\gcl1\gbr\gnl
\gcl2\gvac1\gbr\gcl3\gnl
\gvac2\gcl1\gmp+\gnl
\gbmp f\gbmp f\gbmp f\gbmp f\gnl
\gmu\gcl1\gmu\gnl
\gwmuh325\gcn2122\gnl
\gvac1\gwmuh416\gnl
\gvac1\gob4H\gend
=
\gbeg25
\got1H\got1H\gnl
\gbr\gnl
\gbmp f\gbmp g\gnl
\gmu\gnl
\gob2H\gend$$ and $$\gbeg5{12}
\got3H\gvac1\got1H\gnl
\gwcm3\gvac1\gcl2\gnl
\gcl4\gwcmh325\gnl
\gvac1\gcmu\gbr\gnl
\gvac1\gcl4\gcl1\gbmp f\gcl6\gnl
\gvac2\gbr\gnl
\gmp+\gvac1\gcl1\gmp+\gnl
\gbmp f\gvac1\gmu\gnl
\gcl1\gwmuh314\gnl
\gwmu3\gnl
\gvac1\gwmu4\gnl
\gvac1\gob4H\gend
=
\gbeg5{10}
\got4H\got1H\gnl
\gwcmh426\gcl2\gnl
\gcmu\gcmu\gnl
\gcl1\gcl3\gcl1\gbr\gnl
\gmp+\gvac1\gbr\gcl2\gnl
\gbmp f\gvac1\gbmp f\gmp+\gnl
\gmu\gcl1\gmu\gnl
\gwmuh325\gcn2122\gnl
\gvac1\gwmuh416\gnl
\gvac1\gob4H
\gend
=
\gbeg25
\got1H\got1H\gnl
\gcl1\gcl1\gnl
\gbmp g\gbmp f\gnl
\gmu\gnl
\gob2H\gend$$ which are the two sides of \[equiv:normal\].
Epic or monic endomorphisms {#sec:endom-that-are}
===========================
We recall Radford’s theorem on Hopf algebras with a projection [@Rad:SHAP], which was generalized to a categorical setting even more general than the one in the present paper by Bespalov and Drabant [@BesDra:HBMCMBC]:
Let $H$ be a Hopf algebra, and $\pi$ an idempotent Hopf algebra endomorphism of $H$. Then $H\cong {\operatorname{Im}}(\pi){\otimes}{\operatorname{Im}}(p)$, where $p=(\pi\circ S)*\operatorname{\operatorname{id}}_H$ is an idempotent endomorphism of the object $H$ in ${\mathcal B}$ (but not necessarily a Hopf endomorphism). $B:={\operatorname{Im}}(p)$ is a subalgebra and a quotient coalgebra of $H$. The algebra structure of ${\operatorname{Im}}(\pi){\otimes}B$ is a semidirect product with respect to a certain action of $K={\operatorname{Im}}(\pi)$ on $B$, and the coalgebra structure is the cosemidirect product with respect to a certain coaction.
Moreover ${\operatorname{Im}}(p)\cong{{{^{\operatorname{co} \pi}}{H}}}\cong {{\pi}\backslash{H}}$.
Only the last statement is not in [@BesDra:HBMCMBC], who avoid using coinvariant subobjects altogether to generalize [@Rad:SHAP] to categories that might not have equalizers. We check the first isomorphism: We find $$\gbeg37
\got3H\gnl
\gvac1\gcl1\gnl
\gvac1\gbmp p\gnl
\gwcm3\gnl
\gbmp\pi\gvac1\gcl2\gnl
\gcl1\gnl
\gob1H\gvac1\gob1H\gend
=
\gbeg29
\got2H\gnl
\gcmu\gnl
\gmp+\gcl2\gnl
\gbmp\pi\gnl
\gmu\gnl
\gcmu\gnl
\gbmp\pi\gcl2\gnl
\gcl1\gnl
\gob1H\gob1H\gend
=
\gbeg5{10}
\got5H\gnl
\gvac1\gwcm3\gnl
\gvac1\gmp+\gvac1\gcl2\gnl
\gvac1\gbmp\pi\gnl
\gwcmh324\gvac{-1}\gwcmh324\gnl
\gcn2122\gvac{-1}\gbbrh2124\gcn2122\gnl
\gwmuh324\gvac{-1}\gwmuh324\gnl
\gvac1\gbmp\pi\gvac1\gcl2\gnl
\gvac1\gcl1\gnl
\gvac1\gob1H\gvac1\gob1H
\gend
=
\gbeg49
\got4H\gnl
\gwcmh426\gnl
\gcmu\gcmu\gnl
\gbr\gcl2\gcl4\gnl
\gmp+\gmp+\gnl
\gbmp\pi\gbr\gnl
\gbmp\pi\gbmp\pi\gbmp\pi\gnl
\gmu\gmu\gnl
\gob2H\gob2H\gend
=
\gbeg4{10}
\got3H\gnl
\gwcm3\gnl
\gcl2\gwcmh325\gnl
\gvac1\gcmu\gcl5\gnl
\gbr\gcl2\gnl
\gmp+\gmp+\gnl
\gcl1\gbr\gnl
\gbmp\pi\gbmp\pi\gbmp\pi\gnl
\gmu\gmu\gnl
\gob2H\gob2H\gend
=
\gbeg36
\gvac1\got2H\gnl
\gvac1\gcmu\gnl
\gvac1\gmp+\gcl2\gnl
\gu1\gbmp\pi\gnl
\gcl1\gmu\gnl
\gob1H\gob2H\gend$$ and if some morphism $t\colon T\to H$ satisfies $(\pi{\otimes}\operatorname{\operatorname{id}}_H)\Delta
t=\eta{\otimes}t$, then $$pt
=
\gbeg38
\got3T\gnl
\gvac1\gcl1\gnl
\gvac1\gbmp t\gnl
\gwcm3\gnl
\gbmp\pi\gvac1\gcl2\gnl
\gmp+\gnl
\gwmu3\gnl
\gob3H\gend
=
\gbeg25
\gvac1\got1T\gnl
\gvac1\gcl1\gnl
\gu1\gbmp t\gnl
\gmu\gnl
\gob2H\gend
=t$$
\[Radford\] Let $H$ be a Hopf algebra, and $f$ a Hopf algebra endomorphism of $H$.
Assume that $H$ satisfies both chain conditions.
Then there is $n\in{\mathbb N}$ such that $H\cong {\operatorname{Im}}(f^n){\otimes}{{{^{\operatorname{co} f^n}}{H}}}$ is a Radford biproduct.
Consider the epi-mono factorization $f=(H\xrightarrow eB\xrightarrow
mH)$, where we identify $B={\operatorname{Im}}(f)={\operatorname{Coim}}(f)$. Then the endomorphism $t=em$ of $B$ satisfies $mt=fm$ and $te=ef$. The chain conditions on $H$ imply that the ascending chain of the kernels of $f^n$ and the descending chain of the images, hence the ordered chain of quotient objects formed by the cokernels of $f^n$ stablilize. Then, replacing $f$ by a suitable power $f^n$, we can assume that $t$ is an isomorphism. Then $\pi=mt{^{-1}}e$ is an idempotent endomorphism of $H$, since $\pi^2=mt{^{-1}}emt{^{-1}}e=mt{^{-1}}tt{^{-1}}e=mt{^{-1}}e=\pi$.
Thus $H\cong {\operatorname{Im}}(\pi){\otimes}{{{^{\operatorname{co} \pi}}{H}}}$ is a Radford biproduct. Moreover, ${\operatorname{Im}}(\pi)={\operatorname{Im}}(f)$, and ${{{^{\operatorname{co} \pi}}{H}}}={{{^{\operatorname{co} f}}{H}}}$.
Let $H$ be a Hopf algebra in ${\mathcal B}$ that satisfies both chain conditions, and $f$ a Hopf algebra endomorphism of $H$.
1. If the left or right $f$-coinvariants of $H$ are trivial, then $f$ is a monomorphism in ${\mathcal B}$.
2. If the left or right $f$-invariant quotient of $H$ is trivial, then $f$ is an epimorphism in ${\mathcal B}$.
We prove the first part. By \[Radford\], $H\cong {\operatorname{Im}}(f^n){\otimes}{{{^{\operatorname{co} f^n}}{H}}}$ is a Radford biproduct for some $n$. If ${{{^{\operatorname{co} f^n}}{H}}}$ were trivial without $f^n$ being monic, it would follow that $H$ is isomorphic to a proper quotient of itself, contradicting the chain conditions. Now assume for some $m>1$ that ${{{^{\operatorname{co} f^m}}{H}}}$ is nontrivial. Let $j\colon{{{^{\operatorname{co} f^m}}{H}}}\to H$ be the inclusion. By exactness of tensor products in ${\mathcal B}$, we have an equalizer $$\xymatrix{0\ar[r]&{{{^{\operatorname{co} f}}{H}}}{\otimes}H\ar[r]&H\ar@<0.5ex>[rr]^-{(f{\otimes}H)\Delta{\otimes}H}\ar@<-0.5ex>[rr]_-{\eta{\otimes}H{\otimes}H}&&H{\otimes}H{\otimes}H}$$ and by the calculation $$\begin{gathered}
((f^{m-1}{\otimes}H)\Delta{\otimes}H)(f{\otimes}H)\Delta j
=(f^m{\otimes}f{\otimes}H)(\Delta{\otimes}H)\Delta j
\\=(f^m{\otimes}(f{\otimes}H)\Delta)\Delta j
=\eta{\otimes}(f{\otimes}H)\Delta j
\end{gathered}$$ we see that $(f{\otimes}H)\Delta j$ factors through this equalizer. We conclude that if $(f{\otimes}H)\Delta j$ were not trivial, then it would follow that ${{{^{\operatorname{co} f^{m-1}}}{H}}}{\otimes}H$ is not ${I}{\otimes}H$, which implies ${{{^{\operatorname{co} f^{m-1}}}{H}}}$ is nontrivial. We can conclude by induction that ${{{^{\operatorname{co} f}}{H}}}$ is nontrivial after all.
Let $f\colon H\to G$ be a Hopf algebra homomorphism in ${\mathcal B}$.
1. Clearly, if $f$ is a monomorphism in ${\mathcal B}$, then it is a monomorphism in ${\operatorname{\underline{HopfAlg}}(\mathcal B)}$.
2. If $f$ has trivial left or right coinvariants, then $f$ is a monomorphism in ${\operatorname{\underline{Coalg}}(\mathcal B)}$.
3. If $f$ is normal, and a monomorphism in ${\operatorname{\underline{HopfAlg}}(\mathcal B)}$, then $f$ has trivial left and right coinvariants.
Thus the preceding result shows that normal endomorphisms of a Hopf algebra in ${\mathcal B}$ are monic (epic) if and only if they are so considered as morphisms in ${\mathcal B}$.
If $C$ is a coalgebra and $g,h\colon C\to H$ are coalgebra morphisms with $fg=fh$, then $$\begin{gathered}
\gbeg29
\got2C\gnl
\gcmu\gnl
\gbmp g\gbmp h\gnl
\gcl1\gmp+\gnl
\gmu\gnl
\gcmu\gnl
\gcl2\gbmp f\gnl
\gvac1\gcl1\gnl
\gob1H\gob1G\gend
=
\gbeg5{10}
\got5C\gnl
\gvac1\gwcm3\gnl
\gvac1\gbmp g\gvac1\gbmp h\gnl
\gvac1\gcl1\gvac1\gmp+\gnl
\gwcmh324\gvac{-1}\gwcmh324\gnl
\gcn1122\gbbrh2124\gcn1122\gnl
\gwmuh324\gvac{-1}\gwmuh324\gnl
\gvac1\gcl2\gvac1\gbmp f\gnl
\gvac3\gcl1\gnl
\gvac1\gob1H\gvac1\gob1G\gend
=
\gbeg49
\got4C\gnl
\gwcmh426\gnl
\gcmu\gcmu\gnl
\gcl3\gcl1\gbr\gnl
\gvac1\gbr\gbmp h\gnl
\gvac1\gmp+\gbmp g\gmp +\gnl
\gbmp g\gbmp h\gbmp f\gbmp f\gnl
\gmu\gmu\gnl
\gob2H\gob2G\gend
=
\gbeg49
\got4C\gnl
\gwcmh426\gnl
\gcmu\gcmu\gnl
\gcl3\gcl1\gbr\gnl
\gvac1\gbr\gbmp g\gnl
\gvac1\gmp+\gbmp g\gmp +\gnl
\gbmp g\gbmp h\gbmp f\gbmp f\gnl
\gmu\gmu\gnl
\gob2H\gob2G\gend
=\\\\=
\gbeg4{11}
\got4C\gnl
\gwcmh416\gnl
\gcl6\gwcmh425\gnl
\gvac1\gcmu\gcl2\gnl
\gvac1\gbmp g\gbmp g\gnl
\gvac1\gcl1\gbr\gnl
\gvac1\gbr\gmp+\gnl
\gvac1\gbmp h\gbmp f\gbmp f\gnl
\gbmp g\gmp+\gmu\gnl
\gmu\gcn2122\gnl
\gob2H\gob2G
\gend
=
\gbeg36
\got2C\gnl
\gcmu\gnl
\gbmp g\gbmp h\gnl
\gcl1\gmp+\gu1\gnl
\gmu\gcl1\gnl
\gob2H\gob1G\gend
\end{gathered}$$ and thus, if ${{H}{^{\operatorname{co} f}}}$ is trivial, $g*Sh=\eta{\varepsilon}$, whence $g=h$. If $f$ is normal, then the coinvariants are a Hopf subalgebra.
In general it is false that monic is equivalent to trivial coinvariants, or that epic is equivalent to trivial invariants. In finite dimensions these concepts agree by the Nichols-Zoeller theorem [@NicZoe:HAFT; @Scha:NZTHACYDM]. In infinite dimensions, however, counterexamples are known [@Chi:EMHA].
Let $H$ be a Hopf algebra in ${\mathcal B}$ that satisfies both chain conditions. Assume further that the braiding ${\tau}_{HH}$ has finite order. Then the antipode of $H$ is an automorphism in ${\mathcal B}$.
Depict the iterates of the antipode by $$S^m=\gbeg13\gcl1\gnl\gmp m\gnl\gcl1\gend$$ One has $$\gbeg24
\gcl1\gcl1\gnl
\gmp m\gmp m\gnl
\gdnot{\tau^m}\glmptb\grmptb\gnl
\gmu\gend
=
\gbeg33
\gwmuh324\gnl
\gvac1\gmp m\gnl
\gvac1\gcl1\gend$$ Using this, we can show inductively that the coinvariants of $H$ under an iterate of the antipode are trivial as follows: Let $t\colon T\to H$ be a morphism factoring through ${{{^{\operatorname{co} S^{2n}}}{H}}}$, i. e. $(S^{2n}{\otimes}H)\Delta t=\eta{\otimes}t$. We will show that $(S^m{\otimes}H)\Delta t=\eta{\otimes}t$ for any $m$, whence (taking $m=0$) $t=\eta{\varepsilon}t$.
Assume $(S^{m+1}{\otimes}H)\Delta t=\eta{\otimes}t$, or pictorially $$\gbeg38
\got3T\gnl
\gvac1\gcl1\gnl
\gvac 1\gbmp t\gnl
\gwcm 3\gnl
\gmp +\gvac1\gcl3\gnl
\gmp m\gnl
\gcl1\gnl
\gob1H\gvac1\gob1H\gend
=
\gbeg25
\gvac1\got1T\gnl
\gvac1\gcl1\gnl
\gu1\gbmp t\gnl
\gcl1\gcl1\gnl
\gob1H\gob1H\gend$$ Then $$\gbeg25
\gvac1\got1T\gnl
\gvac1\gcl1\gnl
\gu1\gbmp t\gnl
\gcl1\gcl1\gnl
\gob1H\gob1H\gend
=
\gbeg3{11}
\got3T\gnl
\gvac1\gcl1\gnl
\gvac1\gbmp t\gnl
\gwcmh325\gnl
\gcmu\gcl6\gnl
\gmp+\gcl1\gnl
\gmu\gnl
\gcn2123\gnl
\gvac1\gmp m\gnl
\gvac1\gcl1\gnl
\gvac1\gob1H\gob1H\gend
=
\gbeg39
\got3T\gnl
\gvac1\gcl1\gnl
\gvac1\gbmp t\gnl
\gwcmh314\gnl
\gmp+\gcmu\gnl
\gmp m\gmp m\gcl3\gnl
\gdnot{{\tau}^m}\glmptb\grmptb\gnl
\gmu\gnl
\gob2H\gob1H\gend
=
\gbeg39
\got3T\gnl
\gvac1\gcl1\gnl
\gvac1\gbmp t\gnl
\gcn3134\gnl
\gvac1\gcmu\gnl
\gu1\gmp m\gcl3\gnl
\gdnot{{\tau}^m}\glmptb\grmptb\gnl
\gmu\gnl
\gob2H\gob1H\gend
=
\gbeg38
\got3T\gnl
\gvac1\gcl1\gnl
\gvac 1\gbmp t\gnl
\gwcm 3\gnl
\gmp m\gvac1\gcl2\gnl
\gcl1\gnl
\gob1H\gvac1\gob1H\gend$$
Since the braiding on $H$ has finite order by assumption, some even power of the antipode is a Hopf algebra endomorphism of $H$. Therefore that even power of the antipode is a monomorphism in ${\mathcal B}$. By the dual reasoning it is also an epimorphism, and therefore $S$ itself is an automorphism in ${\mathcal B}$.
Fitting’s lemma {#sec-fitting}
===============
Let $H$ be a Hopf algebra, and $f$ a Hopf algebra endomorphism of $H$.
Assume that $H$ satisfies both chain conditions, so that there is an $n\in{\mathbb N}$ such that $H\cong {\operatorname{Im}}(f^n){\otimes}{{{^{\operatorname{co} f^n}}{H}}} $ is a Radford biproduct.
If $f$ is normal, the action of ${\operatorname{Im}}(f^n)$ on ${{{^{\operatorname{co} f^n}}{H}}}$ is trivial, so that, as an algebra, $H$ is the tensor product of ${\operatorname{Im}}(f^n)$ and ${{{^{\operatorname{co} f^n}}{H}}}$ in ${\mathcal B}$. Similarly if $f$ is conormal, then the coalgebra $H$ is a tensor product of coalgebras in ${\mathcal B}$. In particular, if $f$ is binormal then ${\operatorname{Im}}(f^n)$ and ${{{^{\operatorname{co} f^n}}{H}}}$ are unbraided Hopf algebras in ${\mathcal B}$, and $H$ is isomorphic to their tensor product.
We continue the proof of \[Radford\], assuming that ${\operatorname{Ker}}(f^2)={\operatorname{Ker}}(f)$ and ${\operatorname{Coker}}(f^2)={\operatorname{Coker}}(f)$ after replacing $f$ by a power of $f$. We now add the observation that normality of $f$ implies that $p=(f*(Sf*\operatorname{\operatorname{id}}_H))p=(Sf*\operatorname{\operatorname{id}}_H)p$. Therefore $f{\curlyvee}p$, and dually $f{\curlywedge}p$ if $f$ is conormal. This in turn implies that the Radford biproduct is just an ordinary tensor product algebra or tensor product coalgebra, as appropriate.
Let $H$ be a Hopf algebra. If $H\cong A{\otimes}B$ for two Hopf algebras $A$ and $B$, then we say that $A$ is a tensor factor of $H$. We note that this implies that $A$ and $B$ are unbraided.
We say that $H$ is tensor indecomposable if it does not have a nontrivial tensor factor. An endomorphism $f$ of $H$ is nilpotent if there is $n\in{\mathbb N}$ such that $f^n=\eta{\varepsilon}$.
If $H$ is a tensor indecomposable Hopf algebra satisfying both chain conditions, then every binormal endomorphism of $H$ is nilpotent or an automorphism.
Krull-Remak-Schmidt {#sec:krull-remak-schmidt}
===================
Of course, a Hopf algebra satisfying both chain conditions can be (inductively) decomposed as a tensor product of indecomposable Hopf subalgebras. We shall now show that the Hopf algebraic analog of the Krull-Remak-Schmidt theorem asserting the uniqueness of such a decomposition also holds. A version of this for completely reducible semisimple Hopf algebras was established in [@Bur:CFPHA]. In general, it cannot be hoped that this result has a categorical version. In [@Mue:SMC] it was shown that a non-degenerate fusion category factorizes into a product of prime ones, but that this was generally not unique. Therefore, such decompositions are rather specific to Hopf algebras.
Let $f,g$ be bicommuting, binormal endomorphisms of a tensor indecomposable Hopf algebra $H$.
If $f$ and $g$ are nilpotent, then so is $f*g$.
Otherwise $f*g$ is a normal automorphism, and after composing $f$ and $g$ with its inverse, we can assume that $f*g=\operatorname{\operatorname{id}}_H$. In particular $f$ composition commutes with $g$. Then one can show by induction that $ \id_H=(f*g)^n
$ is a convolution product of terms of the form $f^kg^{n-k}$ for $0\leq k\leq n$ (in fact this is a binomial formula with binomial coefficients, but writing it is cumbersome because addition is replaced with convolution products). If $f^m=\eta{\varepsilon}=g^m$, this implies $(f*g)^{2m}=\eta{\varepsilon}$, since each term contains an $m$-th power of either $f$ or $g$.
Let $H$ and $H_1\dots H_k$ be Hopf algebras in ${\mathcal B}$. Decomposing $H$ as a tensor product Hopf algebra $$H\cong H_1{\otimes}H_2{\otimes}\dots{\otimes}H_k$$ amounts to specifying a system of injections $\iota_i\colon H_i\to H$ and projections $\pi_i\colon H\to H_i$, all of them Hopf algebra morphisms, which commute and cocommute pairwise, and satisfy $\pi_i\iota_i=\operatorname{\operatorname{id}}_{H_i}$, $\pi_i\iota_j=\eta{\varepsilon}$ if $i\neq j$, and $\iota_1\pi_1*\iota_2\pi_2*\dots*\iota_k\pi_k=\operatorname{\operatorname{id}}_H$. The isomorphisms between $H$ and the tensor product are then given by $$H\xrightarrow{\Delta^{(k-1)}}H^{{\otimes}k}\xrightarrow{\pi_1{\otimes}\dots\pi_k}H_1{\otimes}\dots{\otimes}H_k$$ and $$H_1{\otimes}\dots{\otimes}H_k\xrightarrow{\iota_1{\otimes}\dots{\otimes}\iota_k}H^{{\otimes}k}\xrightarrow{\nabla^{(k-1)}}H.$$ Note that the $H_i$ need to be pairwise unbraided for the tensor product Hopf algebra to make sense.
Let $H$ be a Hopf algebra in ${\mathcal B}$, and let $$\begin{aligned}
H&=H_1{\otimes}H_2{\otimes}\dots {\otimes}H_k\\
&=G_1{\otimes}G_2{\otimes}\dots {\otimes}G_\ell
\end{aligned}$$ be two tensor decompositions of $H$ in tensor indecomposable factors.
Then $k=\ell$, and $H_i\cong G_i$ after a suitable permutation of the indices.
Moreover, if $$\xymatrix{H_i\ar@<.5ex>[r]^{\iota_i}&H\ar@<.5ex>[l]^{\pi_i}\ar@<.5ex>[r]^{p_j}&G_j\ar@<.5ex>[l]^{q_j}}$$ denote the systems of injections and projections going with the decompositions into tensor factors, then the factors can be so numbered that for any $1\leq m\leq k$ $$H\xrightarrow{\Delta^{(k-1)}}H^{{\otimes}k}\xrightarrow{\pi_1{\otimes}\dots{\otimes}\pi_m{\otimes}p_{m+1}{\otimes}\dots{\otimes}p_k}H_1{\otimes}\dots H_m{\otimes}G_{m+1}{\otimes}\dots{\otimes}G_k\label{eq:2}$$ and $$H_1{\otimes}\dots H_m{\otimes}G_{m+1}{\otimes}\dots{\otimes}G_k\xrightarrow{\iota_1{\otimes}\dots{\otimes}\iota_m{\otimes}q_{m+1}{\otimes}\dots{\otimes}q_k}H^{{\otimes}k}\xrightarrow{\nabla^{(k-1)}}H\label{eq:1}$$ are isomorphisms.
There is nothing to show if one of the decompositions consists of only one factor. Otherwise we consider $$\operatorname{\operatorname{id}}_{H_1}=\pi_1\iota_1=\pi_1(q_1p_1*\dots*q_\ell p_\ell)\iota_1
=\pi_1q_1p_1\iota_1*\dots*\pi_1q_\ell p_\ell\iota_1.$$ Since $H_1$ is indecomposable, and the terms in the last convolution product are bicommuting binormal endomorphisms, we know that one of $\pi_1q_jp_j\iota_1$ is an isomorphism. Without loss of generality we assume this happens for $j=1$, and that $\pi_1q_1p_1\iota_1=\operatorname{\operatorname{id}}_{H_1}$. It follows that $\pi_1q_1$ and $p_1\iota_1$ are mutually inverse isomorphisms between $H_1$ and $G_1$. Now put $f=q_2p_2*\dots*q_\ell p_\ell$ and $t=\iota_1\pi_1q_1p_1*f$. Since $p_1t=p_1\iota_1\pi_1q_1p_1=p_1$, we have ${{H}{^{\operatorname{co} t}}}\subset{{H}{^{\operatorname{co} p_1t}}}={{H}{^{\operatorname{co} p_1}}}$. Thus, for $j\colon{{H}{^{\operatorname{co} t}}}\to H$ the inclusion, we find $$\Delta j=(H{\otimes}q_1p_1*\dots*q_\ell p_\ell)\Delta j
=(H{\otimes}f)\Delta j
=(H{\otimes}t)\Delta j
=(H{\otimes}\eta)j,$$ and therefore ${{H}{^{\operatorname{co} t}}}$ is trivial. We conclude that $t$ is an automorphism of $H$.
Write $\tilde\pi\colon H\to H_2{\otimes}\dots{\otimes}H_k=:\tilde H$ and $\tilde\iota\colon \tilde H\to H$ for the natural projection and injection morphisms, and similarly for $\tilde
p\colon H\to \tilde G$, $\tilde q\colon \tilde G\to H$. Since $tq_1=\iota_1\pi_1q_1$, we have $\tilde\pi tq_1=\eta{\varepsilon}$, and thus $\tilde\pi t=\tilde\pi
t\tilde q\tilde p$ and $\tilde\pi t\tilde q\tilde
pt{^{-1}}\tilde\iota=\tilde\pi\tilde\iota=\operatorname{\operatorname{id}}_{\tilde H}$. It follows that $\tilde\pi t\tilde q$ and $\tilde p t{^{-1}}\tilde\iota$ are mutually inverse isomorphism between $\tilde G$ and $\tilde H$.
Thus, by an inductive argument we have $k=\ell$, and we can rearrange the indices to get $H_i\cong G_i$ for all $i$.
Note further that the automorphism $t$ above is the composition of the isomorphism $$H\xrightarrow{\Delta^{(k-1)}}H^{{\otimes}k}\xrightarrow{p_1{\otimes}\dots{\otimes}p_k}G_1{\otimes}\dots{\otimes}G_k\xrightarrow{\pi_1q_1{\otimes}G_2{\otimes}\dots{\otimes}G_k}H_1{\otimes}G_2{\otimes}\dots{\otimes}G_k$$ with the morphism $$H_1{\otimes}G_2{\otimes}\dots{\otimes}G_k\xrightarrow{\iota_1{\otimes}q_2{\otimes}\dots{\otimes}q_k}H^{{\otimes}k}\xrightarrow{\nabla^{(k-1)}}H,$$ whence the latter is an isomorphism. Again by an inductive argument, we get that is an isomorphism; the reasoning for is similar.
Endomorphisms of tensor products {#sec:endo-tens-prod}
================================
Let $H$ and $K$ be two Hopf algebras in ${\mathcal B}$, unbraided so that one can form the tensor product bialgebra $H{\otimes}K$. Let $A$ be an algebra in ${\mathcal B}$. It is well-known that there is a bijection $${\operatorname{\underline{Alg}}}(H{\otimes}K,A)\cong\{(a,b)\in {\operatorname{\underline{Alg}}}(H,A)\times{\operatorname{\underline{Alg}}}(K,A)|a{\curlyvee}b\}.$$ In fact, a pair $(a,b)$ of commutation commuting algebra morphisms induces $f=\nabla_A(a{\otimes}b)$, and $$\gbeg56
\got2{H{\otimes}K}\gvac1\got2{H{\otimes}K}\gnl
\gcn2122\gvac1\gcn2122\gnl
\gwmuh528\gnl
\gvac2\gbmp{f}\gnl
\gvac2\gcl1\gnl
\gvac2\gob1A\gend
=
\gbeg66
\got1H\gvac1\got1K\got1H\gvac1\got1K\gnl
\gcl1\gvac1\gbr\gvac1\gcl1\gnl
\gwmu3\gwmu3\gnl
\gvac1\gbmp a\gvac2\gbmp b\gnl
\gvac1\gwmu4\gnl
\gvac1\gob4A\gend
=
\gbeg46
\got1H\got1K\got1H\got1K\gnl
\gcl1\gbr\gcl1\gnl
\gbmp a\gbmp a\gbmp b\gbmp b\gnl
\gmu\gmu\gnl
\gwmuh426\gnl
\gob4A\gend
=
\gbeg46
\got1H\got1K\got1H\got1K\gnl
\gcl1\gcl1\gcl1\gcl1\gnl
\gbmp a\gbmp b\gbmp a\gbmp b\gnl
\gmu\gmu\gnl
\gwmuh426\gnl
\gob4A\gend$$ shows that $f$ is multiplicative. Conversely, given $f\colon H{\otimes}K\to
A$ define $a=f(H{\otimes}\eta)$ and $b=f(\eta{\otimes}K)$. Then, with $T:=H{\otimes}K$: $$\gbeg25
\got1H\got1K\gnl
\gcl1\gcl1\gnl
\gbmp a\gbmp b\gnl
\gmu\gnl
\gob2A\gend
=
\gbeg45
\got1H\gvac2\got1K\gnl
\gcl1\gu1\gu1\gcl1\gnl
\gdnot f\glmpt\grmptb\gdnot f\glmptb\grmpt\gnl
\gvac1\gmu\gnl
\gob4A\gend
=
\gbeg67
\got1H\gvac4\got1K\gnl
\gcl2\gvac1\gu1\gu1\gvac1\gcl2\gnl
\gvac2\gbr\gnl
\gwmu3\gwmu3\gnl
\gvac1\glmpt\gdnot f\gcmpb\gcmp\grmpt\gnl
\gvac2\gcl1\gnl
\gvac2\gob1A\gend
=f$$ and $$\gbeg25
\got1K\got1H\gnl
\gcl1\gcl1\gnl
\gbmp b\gbmp a\gnl
\gmu\gnl
\gob2A\gend
=
\gbeg45
\gvac1\got1K\got1H\gnl
\gu1\gcl1\gcl1\gu1\gnl
\gdnot f\glmpt\grmptb\gdnot f\glmptb\grmpt\gnl
\gvac1\gmu\gnl
\gob4A\gend
=
\gbeg66
\gvac2\got1K\got1H\gnl
\gu1\gvac1\gbr\gvac1\gu1\gnl
\gwmu3\gwmu3\gnl
\gvac1\glmpt\gdnot f\gcmpb\gcmp\grmpt\gnl
\gvac2\gcl1\gnl
\gvac2\gob1A\gend
=
\gbeg25
\got1K\got1H\gnl
\gbr\gnl
\gbmp a\gbmp b\gnl
\gmu\gnl
\gob2A\gend$$
Assume that $A$ is a bialgebra in ${\mathcal B}$, and $a,b,f$ are as above. Then $f$ is a bialgebra homomorphism if and only if $a$ and $b$ are.
Dually, for a coalgebra $C$ in ${\mathcal B}$, a bijection $$\left\{(a,b)\in{\operatorname{\underline{Coalg}}}(C,H)\times{\operatorname{\underline{Coalg}}}(C,H)\left|
a{\curlywedge}b
\right.\right\}\longrightarrow{\operatorname{\underline{Coalg}}}(C,H{\otimes}K)$$ is given by $(a,b)\mapsto (a{\otimes}b)\Delta$, and it induces bijections on the subsets containing (pairs of) bialgebra maps.
Putting the above together, one obtains a bijection between ${\operatorname{End}}(H{\otimes}K)$ and $$\left\{(a,b,c,d)\left|\begin{array}{l}a\in{\operatorname{End}}(H),\
b\in{\operatorname{Hom}}(K,H),\\
c\in{\operatorname{Hom}}(H,K),\
d\in{\operatorname{End}}(K),\\
a{\curlyvee}b,c{\curlyvee}d,
a{\curlywedge}c,b{\curlywedge}d
\end{array}
\right.\right\}$$ with the endomorphism of $H{\otimes}K$ corresponding to a quadruple of Hopf algebra map “components” given by $$\begin{pmatrix}a&b\\c&d\end{pmatrix}
:=
\gbeg46
\got2H\got2K\gnl
\gcmu\gcmu\gnl
\gcl1\gbr\gcl1\gnl
\gbmp a\gbmp b\gbmp c\gbmp d\gnl
\gmu\gmu\gnl
\gob2H\gob2K\gend.$$
Consider a second endomorphism $g$ of $H{\otimes}K$ dissected analogously into a matrix $ \begin{pmatrix}
a'&b'\\c'&d''
\end{pmatrix}$ of Hopf algebra endomorphisms. Then it is straightforward to check that $gf$ corresponds to $ \begin{pmatrix}
a'a*b'c&a'b*b'd\\
c'a*d'c&c'b*d'd
\end{pmatrix}.$
Let $H$ and $K$ be as above, and $f\in{\operatorname{End}}(H{\otimes}K)$ described by a matrix $
\begin{pmatrix}
a&b\\c&d
\end{pmatrix}$. Assume that the antipodes of $H$ and $K$ are automorphisms in ${\mathcal B}$.
Then $f$ is normal if and only if $a$ and $c$ are normal, $b{\curlyvee}\operatorname{\operatorname{id}}_H$, and $d{\curlyvee}\operatorname{\operatorname{id}}_K$; a similar characterization holds for conormal endomorphisms.
We fix projections and injections for the tensor product $P:=H{\otimes}K$: $$\xymatrix{H\ar@<.5ex>[r]^{\iota_H}&P\ar@<.5ex>[l]^{\pi_H}\ar@<.5ex>[r]^{\pi_K}&K\ar@<.5ex>[l]^{\iota_K}}$$ First assume that $f$ is normal. Since $f{\curlyvee}(fS*\operatorname{\operatorname{id}}_P)$, $a=\pi_Hf\iota_H$ commutes with $\pi_H(fS*\operatorname{\operatorname{id}}_P)\iota_H=aS*\operatorname{\operatorname{id}}_H$. Similarly $c$ is normal. Using we have $$\gbeg25
\got1H\got1K\gnl
\gbr\gnl
\gbmp b\gcl1\gnl
\gmu\gnl
\gob2H\gend
=
\gbeg5{11}
\got3H\got1K\gnl
\gvac1\gcl1\gvac1\gcl1\gnl
\gvac1\gbmp{\iota_H}\gvac1\gbmp{\iota_K}\gnl
\gwcm3\gcl1\gnl
\gcl1\gvac1\gbr\gnl
\glmpt\gnot{\operatorname{ad}}\gcmp\grmptb\gcn2213\gnl
\gvac2\gbmp f\gnl
\gvac2\gwmu3\gnl
\gvac3\gbmp{\pi_H}\gnl
\gvac3\gcl1\gnl
\gvac3\gob1H\gend
=
\gbeg38
\got1H\gvac1\got1K\gnl
\gcl1\gvac1\gcl1\gnl
\gbmp{\iota_H}\gvac1\gbmp{\iota_K}\gnl
\gcl1\gvac1\gbmp f\gnl
\gwmu3\gnl
\gvac1\gbmp{\pi_H}\gnl
\gvac1\gcl1\gnl
\gob3H\gend
=
\gbeg25
\got1H\got1K\gnl
\gcl2\gcl1\gnl
\gvac1\gbmp b\gnl
\gmu\gnl
\gob2H\gend$$ so that $b{\curlyvee}\operatorname{\operatorname{id}}_H$. Similarly $d{\curlyvee}\operatorname{\operatorname{id}}_K$.
Now suppose that the stated normality and commutation conditions on $a,b,c,d$ hold. Writing $\hat a=\iota_Ha\pi_H,\hat b=\iota_Hb\pi_K$ etc. we can write $f=\hat a*\hat b*\hat c*\hat d$ as a convolution product of four commuting and cocommuting endomorphisms of $P$. We are claiming that this product commutes with $$fS*\operatorname{\operatorname{id}}_P=\hat aS*\hat bS*\hat cS*\hat dS*\iota_K\pi_K*\iota_H\pi_H
=\hat bS*\hat cS*\hat dS*\iota_K\pi_K*\hat aS*\iota_H\pi_H.$$ (the last equality using that $\hat a$ bicommutes with $\hat b,\hat c,\hat d,$ and $\iota_K$.) Now $\hat a$ commutes with $\hat aS*\iota_H\pi_H$ since $a$ is normal, with $\hat bS$ since $a{\curlyvee}b$, and with $\iota_K\pi_K$ and $\hat cS$ since $\iota_H{\curlyvee}\iota_K$. The next factor $\hat b$ commutes with $\hat aS$, $\hat bS$, and $\iota_H\pi_H$ since $b{\curlyvee}\operatorname{\operatorname{id}}_H$, and it commutes with $\hat cS,\hat dS,$ and $\iota_K\pi_K$, since $\iota_K{\curlyvee}\iota_H$. Similar arguments deal with the convolution factors $\hat c$ and $\hat d$.
Similarly, an endomorphism $f$ of a tensor product of several pairwise unbraided Hopf algebras $H_1,\dots,H_k$ can be described by a matrix $(v_{ij})$ of Hopf algebra homomorphisms between the factors. By inductive arguments one can show that $f$ is normal iff all the diagonal terms are normal, and the off-diagonal terms commute with the identities on their targets.
An interesting case arises when there are no nontrivial homomorphisms $H_i\to H_j$ commuting with $\operatorname{\operatorname{id}}_{H_j}$. In this case any normal endomorphism preserves the decomposition into tensor factors. One can deduce from this that the Krull-Remak-Schmidt decomposition is unique in a stronger sense than up to permutation and isomorphism; in the original case of decompositions of groups the uniqueness of the subgroups in a direct decomposition into directly indecomposable factors follows as stated in Remak’s thesis [@Rem:ZEGDUF].
Automorphisms of tensor products {#sec:autom-tens-prod}
================================
We consider now the automorphisms of tensor products of Hopf algebras. These are the natural extensions of the corresponding results in group theory [@BidCurMcC:ADPFG; @Bid:ADPFG2].
Throughout this section we let $H$ and $K$ be unbraided Hopf algebras, so that we can form the tensor product $H{\otimes}K$, and we assume that the antipodes of $H$ and $K$ are automorphisms in ${\mathcal B}$.
Identify endomorphisms of $H{\otimes}K$ with matrices of Hopf algebra homomorphisms as in \[sec:endo-tens-prod\]. Let $ \begin{pmatrix}
a&b\\c&d
\end{pmatrix}\in\Hopfend(H\ot K).$ If $a$ is an automorphism, then by \[compositescommute\] of \[usefulsimplecommutationfacts\] the condition $a{\curlyvee}b$ implies $\operatorname{\operatorname{id}}_H{\curlyvee}b$ (and $x{\curlyvee}b$ for any $x\colon X\to H$). Similarly $\operatorname{\operatorname{id}}_K{\curlywedge}c$, and, if $d$ is also an automorphism, $b{\curlywedge}\operatorname{\operatorname{id}}_H$ and $c{\curlyvee}\operatorname{\operatorname{id}}_K$.
Define $$\mathcal{A} = \begin{pmatrix}
{\operatorname{Aut}}(H)&{\operatorname{Hom}^c}(K,H)\\
{\operatorname{Hom}^c}(H,K)&{\operatorname{Aut}}(K)
\end{pmatrix},$$ where ${\operatorname{Hom}^c}(K,H):=\{b\in{\operatorname{Hom}}(K,H)|b{\curlyvee}\operatorname{\operatorname{id}}_H\text{ and
}b{\curlywedge}\operatorname{\operatorname{id}}_H\}$. This is easily seen to be an abelian group under convolution product. Indeed, the image of any such morphism determines an abelian Hopf sub-algebra of $H$. Note that $b{\curlyvee}\operatorname{\operatorname{id}}_H$ $\iff$ $b{\curlyvee}\alpha$ for some/all $\alpha\in{\operatorname{Aut}}(H)$, and similarly $b{\curlywedge}\operatorname{\operatorname{id}}$ $\iff$ $b{\curlywedge}\alpha$ for some/all $\alpha\in{\operatorname{Aut}}(H)$.
Consider an automorphism $f$ of $H{\otimes}K$, and its decomposition as a matrix $\begin{pmatrix}
a&b\\c&d
\end{pmatrix}$ of Hopf algebra homomorphisms as in \[sec:endo-tens-prod\]. Let $f{^{-1}}$ correspond in the same way to a matrix $ \begin{pmatrix}
a'&b'\\c'&d'
\end{pmatrix}.$ Then we have $\operatorname{\operatorname{id}}_K=({\varepsilon}{\otimes}K)f{^{-1}}f(\eta{\otimes}K)=c'b*d'd$. Since $c'{\curlyvee}d'$ and $b{\curlywedge}d$, we have that $c'b{\curlyvee}d'd=(c'bS)*\operatorname{\operatorname{id}}_K$ and $c'b{\curlywedge}(c'bS)*\operatorname{\operatorname{id}}_K$. In other words, $c'b$ is a binormal endomorphism of $K$. In the same way $bc'$ is a binormal endomorphism of $H$. If we assume both chain conditions on $H$ and $K$, then for sufficiently large $n$, $b$ and $c'$ induce mutually inverse isomorphisms between the images of $(c'b)^n$ and $(bc')^n$. Thus, using Fittings Lemma, the image of $(c'b)^n$ is a common tensor factor of $H$ and $K$.
This gives part of the following result.
Suppose that $H$ and $K$ satisfy both chain conditions. Then $\mathcal{A}\subseteq {\operatorname{Aut}}(H{\otimes}K)$ if and only if $H$ and $K$ have no non-trivial common abelian direct tensor factors. On the other hand, ${\operatorname{Aut}}(H{\otimes}K)=\mathcal{A}$ if and only if $H$ and $K$ have no non-trivial common direct tensor factors.
If $H$ and $K$ have a common non-trivial direct tensor factor, then permutations of this factor in $H{\otimes}K$ are automorphisms of $H{\otimes}K$ not contained in $\mathcal{A}$.
By the preceeding remarks, to show ${\operatorname{Aut}}(H{\otimes}K)\subseteq \mathcal{A}$ it remains to prove that the common tensor factor in $H{\otimes}K$ that we found is necessarily nontrivial if $d$ is not an automorphism. A similar argument will apply to show that $a$ is an automorphism, and the commutation and cocommutation conditions for the components of an endomorphism will be equivalent to the off-diagonal terms (co)commuting with the identity instead of the automorphisms on the diagonal.
Thus suppose that $d$ is not an automorphism. Then we can assume without loss of generality that the right $d$-coinvariant subobject $D$ of $H$ is nontrivial. If $\iota\colon D\to H$ is the inclusion, then $c'b\iota=\nabla(c'b{\otimes}\eta)\iota=\nabla(c'b{\otimes}d'd)\Delta\iota=(c'b*d'd)\iota=\iota$, hence $(c'b)^n\iota=\iota$ for all $n$, and the image of $(c'b)^n$ is nontrivial as desired.
The desired equality in the second part will then hold once we have proven the first equivalence.
To this end we first consider the forward direction by contrapositive. Suppose that $H$ and $K$ have a common abelian direct tensor factor $L$, and write $H=H'{\otimes}L$ and $K=K'{\otimes}L$. Since $L$ is abelian its antipode $S_L$ is a Hopf endomorphism of $L$. Taking $a=\operatorname{\operatorname{id}}_H,d=\operatorname{\operatorname{id}}_K$, $b=\eta_{K'}\varepsilon_{K'}{\otimes}S_L$ and $c=\eta_{H'}\varepsilon_{H'}{\otimes}S_L$ we find that $\psi=\begin{pmatrix} a&b\\c&d \end{pmatrix}\in\mathcal{A}$. However, $L$ is a sub-object of the right $\psi$-coinvariant subobject, whence $\psi\not\in{\operatorname{Aut}}(H{\otimes}K)$.
For the remaining direction, assume that $f=
\begin{pmatrix}
a&b\\c&d
\end{pmatrix}$ belongs to $\mathcal{A}$; in particular $f$ is a Hopf algebra endomorphism of $H{\otimes}K$. After multiplying with the obvious automorphism $
\begin{pmatrix}
a{^{-1}}&\eta{\varepsilon}\\\eta{\varepsilon}&d{^{-1}}\end{pmatrix}$ of $H{\otimes}K$ we may assume that $a=\operatorname{\operatorname{id}}_H$ and $d=\operatorname{\operatorname{id}}_K$. Now consider $g=
\begin{pmatrix}
\operatorname{\operatorname{id}}_H&bS\\cS&\operatorname{\operatorname{id}}_K
\end{pmatrix}$, another Hopf endomorphism of $H{\otimes}K$. One computes $gf=
\begin{pmatrix}
\operatorname{\operatorname{id}}*bcS&b*bS\\cS*c&cbS*\operatorname{\operatorname{id}}\end{pmatrix}
=
\begin{pmatrix}
\operatorname{\operatorname{id}}*bcS&\eta{\varepsilon}\\\eta{\varepsilon}&cbS*\operatorname{\operatorname{id}}\end{pmatrix}.$ By the chain conditions on $H$ and $K$, for $n$ sufficiently large $b$ and $c$ induce mutually inverse isomorphisms between the images of $(bc)^n$ and $(cb)^n$. Fitting’s lemma implies that these isomorphic images are an abelian common tensor factor of $H$ and $K$. It can only be trivial if $bc$ and $cb$ are nilpotent, in which case $\operatorname{\operatorname{id}}*bcS$ and $cbS*\operatorname{\operatorname{id}}$ are automorphisms. In the latter case, $f$ was an automorphism.
These results have obvious extensions to more than two factors by induction, which we leave to the reader. The results, however, do not cover the case of a repeated tensor factor. For a given Hopf algebra $H$ in ${\mathcal B}$ we can form the unbraided iterated tensor product $H^{{\otimes}n}=H{\otimes}\cdots{\otimes}H$ for $n\in{\mathbb N}$ precisely when $H$ is in a (sub)category where the braiding is a symmetry.
Let $H$ be a tensor indecomposable non-abelian Hopf algebra satisfying both chain conditions in ${\mathcal B}$, and suppose the braiding of ${\mathcal B}$ is a symmetry. Fix $n\in{\mathbb N}$, and let $\mathcal{A}_n$ denote those $(\alpha_{ij})\in{\operatorname{End}}(H^{{\otimes}n})$ such that $\alpha_{ii}\in{\operatorname{Aut}}(H)$ and $\alpha_{ij}\in{\operatorname{End}}(H)$ for all $i$ and $j\neq i$. Then $${\operatorname{Aut}}(H^{{\otimes}n}) \cong \mathcal{A}_n\rtimes S_n.$$
By assumptions on $H$, $\mathcal{A}_n\subseteq{\operatorname{Aut}}(H^{{\otimes}n})$. The group $S_n$ acts on $H^{{\otimes}n}$ by permuting factors, and so acts on ${\operatorname{Aut}}(H^{{\otimes}n})$ by permuting columns. Conjugating by this action sends $\mathcal{A}_n$ to itself. We need only show that every automorphism is a column permutation of an element of $\mathcal{A}_n$.
So let $(\alpha_{ij})\in{\operatorname{Aut}}(H^{{\otimes}n})$, with inverse $(\alpha_{ij}')$. Then for all $i$ we have $\alpha_{i1}\alpha_{1i}'*\cdots*\alpha_{in}\alpha_{ni}'=\operatorname{\operatorname{id}}$. Since the $\alpha_{ik}\alpha_{ki}'$ are all binormal endomorphisms the notation is unambiguous, and the terms of the convolution product can be arbitrarily reordered. Moreover, since $H$ is indecomposable we may conclude that one of the $\alpha_{ik}\alpha_{ki}'$ is an automorphism. In particular for all $i$ there is a $k$ such that $\alpha_{ik}$ is an epimorphism and $\alpha_{ki}'$ is a monomorphism. By the chain conditions it follows that $\alpha_{ik}$ and $\alpha_{ki}'$ are both automorphisms. Since $H$ is non-abelian there is at most one such $k$ for any given $i$. This completes the proof.
Application to doubles of groups {#sec:appl-doubl-groups}
================================
For this section we work in the category of vector spaces over a field ${\mathbbm k}$. Throughout this section $G,H,K,C$ will all be finite groups. For any group $G$ let $\widehat{G}$ be the group of group-like elements of ${{\mathbbm k}^{G}}$, the dual of the group algebra ${\mathbbm k}G$. Note that $\widehat{G}$ is precisely the ${\mathbbm k}$-linear characters of $G$. We also define $\Gamma_G=\widehat{G}\times G$. We denote the conjugation action of $G$ on ${{\mathbbm k}^{G}}$ and ${\mathbbm k}G$ both by $\rightharpoonup$; e.g. $g\rightharpoonup x = gxg^{-1}$ for all $g,x\in G$. We will be concerned with ${\mathcal D}(G)$, the Drinfeld double of a finite group $G$. As a coalgebra ${\mathcal D}(G)={{{{\mathbbm k}^{G}}}{^{\operatorname{co} }}}{\otimes}{\mathbbm k}G$, and the algebra structure is given by having $G$ act on ${{{{\mathbbm k}^{G}}}{^{\operatorname{co} }}}$ by the conjugation action. We note that $\Gamma_G$ is the group of group-like elements of ${\mathcal D}(G)$. See [@DijPasRoc:QHAGCOM; @Mon:HAAR] for further details on the construction and properties of this Hopf algebra.
In [@K14] the first author gave a complete description of ${\operatorname{Aut}}({\mathcal D}(G))$ whenever $G$ has no non-trivial abelian direct factors. Such a group is said to be purely non-abelian. When $G$ is abelian we have ${\mathcal D}(G)={{\mathbbm k}^{G}}{\otimes}{\mathbbm k}G$, an abelian Hopf algebra, and the determination of ${\operatorname{Aut}}({\mathcal D}(G))$ is then straightforward. Indeed, under mild assumptions on ${\mathbbm k}$ we have ${\mathcal D}(G) \cong {\mathbbm k}(G\times G)$. Subsequently in this case ${\operatorname{Aut}}({\mathcal D}(G))$ can be computed by classical methods in group theory [@Sho:AAG]. We note that the structure of such an automorphism group has been of more recent interest [@BidCur:AFAG; @HilRhe:AFAG]. It is the goal of this section to complete the description of ${\operatorname{Aut}}({\mathcal D}(G))$ when $G$ has an abelian direct factor.
So suppose that $G=C\times H$ with $C$ abelian. Then ${\mathcal D}(G)\cong{\mathcal D}(C)\otimes{\mathcal D}(H)$. Since ${\mathcal D}(C)$ is an abelian Hopf algebra the results of the previous section can be applied whenever ${\mathcal D}(H)$ has no abelian direct tensor factors. We will proceed to show this happens precisely when $H$ is purely non-abelian.
We have the following description of ${\operatorname{Hom}}({\mathcal D}(G),{\mathcal D}(K))$.
The elements of ${\operatorname{Hom}}({\mathcal D}(G),{\mathcal D}(K))$ are in bijective correspondence with matrices $\begin{pmatrix}
u&r\\p&v
\end{pmatrix}$ where $u\colon {{\mathbbm k}^{G}}\to{{\mathbbm k}^{K}}$ is a morphism of unitary coalgebras; $p\colon {{\mathbbm k}^{G}}\to{\mathbbm k}K$ is a morphism of Hopf algebras; and $r\colon G\to\widehat{K}$ and $v\colon G\to K$ are group homomorphisms. These are all subject to the following compatibility relations, for all $a,b\in{{\mathbbm k}^{G}}$ and $g\in G$:
1. $u(g\rightharpoonup a)=v(g)\rightharpoonup u(a)$, from which it follows that $u^*v$ is normal;
2. $u{\curlywedge}p$;
3. $u(ab)=u(a_{(1)})(p(a_{(2)})\rightharpoonup u(b))$;
4. $p(g\rightharpoonup a)=v(g)\rightharpoonup p(a)$.
The morphism is defined by $$a\# g\mapsto u(a_{(1)})r(g)\# p(a_{(2)})v(g).$$ Composition of such morphisms is given by matrix multiplication, as in \[sec:endo-tens-prod\].
The morphism $p$ is uniquely determined by an isomorphism ${{\mathbbm k}^{A}}\cong{\mathbbm k}B$, where $A$ is an abelian normal subgroup of $G$ and $B$ is an abelian subgroup of $K$. In particular we must have ${{\mathbbm k}^{A}}\cong {\mathbbm k}\widehat{A}$. For the remainder of this section any use of $A,B$ refers to these subgroups. We note that the last relation says $p{\curlyvee}v$ if and only if $A\leq Z(G)$, or equivalently $p$ is cocentral: $p{\curlywedge}\operatorname{\operatorname{id}}$.
By convention we implicitly identify any element of ${\operatorname{Hom}}({\mathcal D}(G),{\mathcal D}(K))$ or ${\operatorname{Hom}}({{\mathbbm k}^{G}\otimes\kkG},{{\mathbbm k}^{K}\otimes\kkK})$ with its quadruple of components $(u,r,p,v)$, or equivalently as a matrix $\begin{pmatrix} u&r\\p&v \end{pmatrix}$.
The following is then immediate.
A morphism $\psi\in{\operatorname{Hom}}({\mathcal D}(G),{\mathcal D}(K))$ is canonically an element of ${\operatorname{Hom}}({{\mathbbm k}^{G}\otimes\kkG},{{\mathbbm k}^{K}\otimes\kkK})$ precisely when $p{\curlyvee}v$ and $u$ is a morphism of Hopf algebras.
On the other hand, $\phi\in{\operatorname{Hom}}({{\mathbbm k}^{G}\otimes\kkG},{{\mathbbm k}^{K}\otimes\kkK})$ is canonically an element of ${\operatorname{Hom}}(D(G),{\mathcal D}(K))$ precisely when $u^*\circ v$ is normal and $A\leq Z(G)$.
In the first case we call such a morphism untwistable, and in the second we call it twistable. Clearly any untwistable morphism is also twistable, and vice versa. The distinction is simply in the algebra structures we start with.
Now since ${{\mathbbm k}^{G}\otimes\kkG}$ and ${{\mathbbm k}^{K}\otimes\kkK}$ are canonically self-dual any morphism $\psi\in{\operatorname{Hom}}({{\mathbbm k}^{G}\otimes\kkG},{{\mathbbm k}^{K}\otimes\kkK})$ yields a dual morphism $\psi^*\in{\operatorname{Hom}}({{\mathbbm k}^{K}\otimes\kkK},{{\mathbbm k}^{G}\otimes\kkG})$ with components $(v^*,r^*,p^*,u^*)$. The following is then clear.
Both $\psi\in{\operatorname{Hom}}({{\mathbbm k}^{G}\otimes\kkG},{{\mathbbm k}^{K}\otimes\kkK})$ and $\psi^*$ are twistable if and only if the following all hold
1. $u^* v$ is normal;
2. $v u^*$ is normal;
3. $A\leq Z(G)$;
4. $B\leq Z(K)$.
In this case we may canonically view $\psi\in{\operatorname{Hom}}({\mathcal D}(G),{\mathcal D}(K))$ and $\psi^*\in{\operatorname{Hom}}({\mathcal D}(K),{\mathcal D}(G))$.
In [@K14] a morphism $\psi=(u,r,p,v)\in{\operatorname{Hom}}({\mathcal D}(G),{\mathcal D}(K))$ was said to be flippable if also $(v^*,r^*,p^*,u^*)\in{\operatorname{Hom}}({\mathcal D}(K),{\mathcal D}(G))$. This is equivalent to saying that $\psi$ is untwistable and the corresponding dual $\psi^*$ is twistable. In particular the Corollary gives a complete description of the flippable elements of ${\operatorname{Hom}}({\mathcal D}(G),{\mathcal D}(K))$, and shows that ’flipping’ an element of ${\operatorname{Hom}}({\mathcal D}(G),{\mathcal D}(K))$ can naturally be described as dualizing the morphism.
For any group $G$, ${\operatorname{Aut}}({\mathcal D}(G))$ is canonically a subgroup of ${\operatorname{Aut}}({{\mathbbm k}^{G}\otimes\kkG})$ which is closed under dualization.
Follows from the preceeding corollary, \[sec:autom-tens-prod\], and the properties of ${\operatorname{Aut}}({\mathcal D}(G))$ established in [@K14].
We now show that the act of untwisting a morphism is fairly well-behaved whenever the image is commutative.
\[prop:algebra-change-preserves\] Let $\psi\in{\operatorname{Hom}}({\mathcal D}(G),{\mathcal D}(K))$ be untwistable. For convenience, let $\psi'=\psi\in{\operatorname{Hom}}({{\mathbbm k}^{G}\otimes\kkG},{{\mathbbm k}^{H}\otimes\kkH})$. Then the following all hold.
1. If $G=H$, then $\psi$ is conormal if and only if $\psi'$ is conormal.\[prop:conormal-preserved\]
2. If $\psi$ has a commutative image then $\psi'$ has commutative image.
3. If $\psi$ has commutative image and $G=H$, then $\psi'$ is normal if and only if $v$ is normal and $B\leq Z(G)$.
4. If $\psi$ has commutative image and $G=H$, then $\psi$ is normal if and only if $\psi'$ is normal and $G$ acts trivially on $\operatorname{Img}(u)$.
5. If $\psi$ has commutative image and $G=H$, then $\psi$ is conormal.
We first prove \[prop:conormal-preserved\], as it is the only one that does not suppose that $\psi$ has commutative image. To this end we compute $$\begin{aligned}
a_{(3)}\# g \cdot S(a_{(1)}\# g) \otimes a_{(2)}\# g &= a_{(3)}\#g\cdot (g^{-1}\rightharpoonup S(a_{(1)})\#g^{-1})\otimes a_{(2)}\# g\nonumber\\
&= a_{(3)}S(a_{(1)})\#1 \otimes a_{(2)}\# g;\label{eq:double-conormal-base}\\
a_{(3)}\otimes g \cdot S(a_{(1)}\otimes g) \otimes a_{(2)}\otimes g
&= a_{(3)}S(a_{(1)})\otimes 1\otimes a_{(2)}\otimes g\label{eq:tensor-conormal-base}.
\end{aligned}$$ The claim then follows.
For the remainder of the proof, suppose that $\psi$ is untwistable and has commutative image. By checking the commutativity condition we can easily determine the following facts: $p{\curlyvee}v$; $v$ has abelian image, or equivalently $v{\curlyvee}v$; $u(a(g\rightharpoonup b))=u((h\rightharpoonup a)b)=u(ab)$ for all $g,h\in G$ and $a,b\in{{\mathbbm k}^{G}}$, which implies $u{\curlywedge}\operatorname{\operatorname{id}}$. In particular, $u(g\rightharpoonup a)=u(a)$ and $p(g\rightharpoonup a)=p(a)$ for all such $a,g$.
The first part is then immediate, as we have $\psi(a\#g \cdot b\# h) = \psi'(a\otimes g \cdot b\otimes h)$. Another way of saying this is that when $\psi$ has commutative image we may compute products in either the double or the tensor product without affecting the result. Furthermore $\psi((g\rightharpoonup a)\# g)=\psi(a\# g)$ for all appropriate $a,g$, and so $$\psi(S(a\# g)) = \psi(g^{-1}\rightharpoonup S(a)\# g^{-1})=\psi(S(a)\# g^{-1})=\psi'(S(a{\otimes}g)).$$ Thus we may perform all computations with $\psi$ in either ${\mathcal D}(G)$ or ${{\mathbbm k}^{G}\otimes\kkG}$ as we desire.
The last part of the result follows from \[eq:double-conormal-base,eq:tensor-conormal-base\] and $u{\curlywedge}\operatorname{\operatorname{id}}$. We need only prove the parts concerning normality of $\psi,\psi'$.
To determine when $\psi'$ is normal, we first note that by commutativity we have $$\psi'(a_{(2)}\otimes g\cdot b\otimes h \cdot S(a_{(1)})\otimes g^{-1}) = a(1)\psi'(b\otimes h).$$ On the other hand, $$a_{(2)}\otimes g \cdot \psi'(b\otimes h) \cdot S(a_{(1)})\otimes g^{-1} = a(1) r(h)u(b_{(1)})\otimes g p(b_{(1)}) v(h) g^{-1}.$$ $\psi'$ is normal precisely when these two expressions are the same, and we easily find this is equivalent to $B\leq Z(H)$ and $v$ normal.
Finally, we determine when $\psi$ is normal. By previous remarks, we have
$$\begin{aligned}
\psi(a_{(2)}\#g \cdot b\# h \cdot S(a_{(1)}\#g)) &= a(1) \psi(b\# ghg^{-1})\nonumber\\
&= a(1) r(h)u(b_{(1)})\# p(b_{(2)})v(ghg^{-1})\label{eq:psi-normal-inside}.
\end{aligned}$$
On the other hand, we have $$\begin{aligned}
&a_{(2)}\#g \cdot \psi(b\# h) \cdot \left(g^{-1}\rightharpoonup Sa_{(1)}\# g^{-1}\right)\nonumber\\& \qquad= r(h) a_{(2)} \left(g\rightharpoonup u(b_{(1)})\right) (g p(b_{(2)})v(h)g^{-1}\rightharpoonup S(a_{(1)}))\# g p(b_{(3)})v(h)g^{-1}.
\end{aligned}$$ Applying ${\varepsilon}\#\operatorname{\operatorname{id}}$ to both expressions we get $$a(1)p(b)v(ghg^{-1})$$ for the first and $$a(1) g p(b) v(h) g^{-1}$$ for the second. These are equal for all $a,b,g,h$ if and only if $B\leq Z(H)$ and $v$ is normal; equivalently, $\psi'$ is normal. Note that if $v$ is normal and has abelian image, then its image is in fact central. Therefore $g p(b)v(h)g^{-1}\rightharpoonup S(a) = S(a)$ precisely when $\psi'$ is normal. Subsequently the previous equation simplifies to $$a(1) r(h)(g\rightharpoonup u(b_{(1)}))\# p(b_{(2)})v(ghg^{-1}).$$ Comparing with \[eq:psi-normal-inside\] completes the proof.
Any commutative direct tensor factor of ${\mathcal D}(G)$ is also a commutative direct tensor factor of ${{\mathbbm k}^{G}\otimes\kkG}$.
Suppose $L$ is a commutative Hopf sub-algebra of ${\mathcal D}(G)$ such that ${\mathcal D}(G)=M\otimes L$ for some Hopf sub-algebra $M$. We then have a projection $\pi\colon{\mathcal D}(G)\to L$ with associated right inverse the imbedding $i\colon L\hookrightarrow{\mathcal D}(G)$.
The morphism $i\pi$ is an endomorphism of ${\mathcal D}(G)$. Since the image is central in ${\mathcal D}(G)$ it is easily seen to be untwistable and binormal. Therefore $i\pi$ is canonically a twistable, binormal, idempotent endomorphism of ${{\mathbbm k}^{G}\otimes\kkG}$ with image $L$. By Fitting’s lemma we conclude that $L$ is also a direct tensor factor of ${{\mathbbm k}^{G}\otimes\kkG}$.
Since ${{\mathbbm k}^{G}}$ is commutative we see that the converse will only hold when $G$ is abelian. Indeed since ${\mathcal D}(G)$ is quasitriangular any commutative direct tensor factor of ${\mathcal D}(G)$ is necessarily abelian.
The lemma gives one part of the following.
Let $G$ be a finite group. Then the following are equivalent.
1. $G$ is purely non-abelian.
2. ${\mathbbm k}G$ is purely non-abelian.
3. ${{\mathbbm k}^{G}}$ is purely non-abelian.
4. ${{\mathbbm k}^{G}\otimes\kkG}$ is purely non-abelian.
5. ${\mathcal D}(G)$ is purely non-abelian.
Indeed, ${{\mathbbm k}^{G}\otimes\kkG}$ and ${\mathcal D}(G)$ have the same abelian direct tensor factors.
Since the dual of an abelian Hopf algebra is again abelian, the equivalence of the second and third is immediate. By Krull-Remak-Schmidt, any abelian indecomposable factor of ${{\mathbbm k}^{G}\otimes\kkG}$ is isomorphic to an abelian indecomposable factor of either ${\mathbbm k}G$ or ${{\mathbbm k}^{G}}$. Thus the fourth is equivalent to the second and third. Since any Hopf sub-algebra of ${\mathbbm k}G$ is a subgroup algebra, the first and second are equivalent. By the lemma the fourth implies the fifth. To prove the fifth implies the fourth, we need only show that any abelian factor of ${{\mathbbm k}^{G}\otimes\kkG}$ yields an abelian factor of ${\mathcal D}(G)$.
So let $L$ be an abelian tensor factor of ${{\mathbbm k}^{G}\otimes\kkG}$ with associated projection $\pi$ and inclusion $i$. We wish to show that $i\pi$ is canonically a binormal endomorphism of ${\mathcal D}(G)$. Writing $i\pi=\begin{pmatrix} u&r\\p&v \end{pmatrix}$, the properties of ${\operatorname{End}}({\mathcal D}(G))$ and commutativity of the image easily imply the following: $p{\curlyvee}v$, $v{\curlyvee}v$, $u{\curlywedge}p$, $p{\curlyvee}\operatorname{\operatorname{id}}$, $v{\curlyvee}\operatorname{\operatorname{id}}$. In particular, $v$ and $p$ have central image, and $v$ is a (bi)normal group homomorphism. Since $(i\pi)^*$ is also an idempotent endomorphism with abelian image we similarly conclude that $p^*$ and $u^*$ have central image, and that $u^*$ is a (bi)normal group homomomorphism. Centrality of the image of $u^*$ (indeed, that $u^*$ has abelian image and is thus a class function) implies that $G$ acts trivially on the image of $u$. Applying the proposition we conclude that $i\pi$ is canonically a binormal endomorphism of ${\mathcal D}(G)$ with image $L$. Fitting’s lemma then implies that $L$ is a direct tensor factor of ${\mathcal D}(G)$, as desired.
This completes the proof.
Thus for $G=C\times H$ with $C$ abelian and $H$ purely non-abelian we conclude that ${\mathcal D}(C)$ and ${\mathcal D}(H)$ have no common direct tensor factors. Therefore we may apply the results of the previous section to obtain the following.
Let $G=C\times H$, where $C,H$ are finite groups with $C$ abelian and $H$ purely non-abelian. Then $${\operatorname{Aut}}({\mathcal D}(G)) = \begin{pmatrix}
{\operatorname{Aut}}({\mathcal D}(C))&{\operatorname{Hom}^c}({\mathcal D}(H),{\mathcal D}(C))\\
{\operatorname{Hom}^c}({\mathcal D}(C),{\mathcal D}(H))&{\operatorname{Aut}}({\mathcal D}(H))
\end{pmatrix}.$$
The determination of the ${\operatorname{Hom}^c}$ terms remains a computational problem, but all of the components of these morphisms are guaranteed to be morphisms of Hopf algebras, and so determined by group homomorphisms. Note that for ${\operatorname{Hom}^c}({\mathcal D}(H),{\mathcal D}(C))$ we have a commutative image, as considered in \[prop:algebra-change-preserves\]. Whenever the field is such that ${\mathcal D}(C)$ is just a group algebra then the situation is further simplified. In this case ${\operatorname{Hom}^c}({\mathcal D}(C),{\mathcal D}(H))={\operatorname{Hom}}(\Gamma_C,Z(\Gamma_H))$, a group of morphisms between abelian groups.
\[ex:dihedrals\] Consider a field ${\mathbbm k}$ of characteristic not $2$. For $n\geq 3$ let $G=D_{2n}$ be the dihedral group of order $2n$, and suppose that $n\equiv 2\bmod 4$. The group $G$ has an abelian direct factor precisely under this assumption on $n$, in which case $G\cong{\mathbb Z}_2\times D_n$. So we take $C={\mathbb Z}_2$ and $H=D_n$, and note $\Gamma_C\cong{\mathbb Z}_2^2$ and $\Gamma_H\cong {\mathbb Z}_2\times D_n$. It is also well-known that $\operatorname{Aut}(\Gamma_C)\cong S_3$. By [@K14] we have ${\operatorname{Aut}}({\mathcal D}(D_{n}))\cong {\mathbb Z}_2\times \operatorname{Aut}(D_n)\cong {\mathbb Z}_2\times\operatorname{Hol}({\mathbb Z}_{n/2})$. Here $\operatorname{Hol}({\mathbb Z}_n)={\mathbb Z}_n\rtimes\operatorname{Aut}({\mathbb Z}_n)$ is the holomorph of ${\mathbb Z}_n$, a group of order $n\phi(n)$, where $\phi$ is the Euler totient function.
We have $Z(\Gamma_H)\cong{\mathbb Z}_2$, from which it follows that ${\operatorname{Hom}}(\Gamma_C,Z(\Gamma_H))\cong {\mathbb Z}_2^2$ as groups. We claim that $${\operatorname{Hom}^c}({\mathcal D}(H),{\mathcal D}(C))\cong{\mathbb Z}_2^2$$ as well. Let $(u,r,p,v)\in{\operatorname{Hom}}({\mathcal D}(H),{\mathcal D}(C))$. The abelian normal subgroups of $D_n$ all have odd order, so $p$ is necessarily trivial. By normality of $u^*\circ v$, we have $u^*(b^{v(x)}) = u^*(b) = u^*(b)^x$ for all $x\in D_n$ and $b\in {\mathbb Z}_2$. Since no order 2 subgroup of $D_n$ is normal we conclude that $u^*$ is trivial. From this we can then easily check that ${\operatorname{Hom}}({\mathcal D}(H),{\mathcal D}(C))={\operatorname{Hom}^c}({\mathcal D}(H),{\mathcal D}(C))$. Since there are two possible homomorphisms $v\colon D_n\to{\mathbb Z}_2$, and two possible homomorphisms $r\colon D_n\to\widehat{{\mathbb Z}_2}$, all of which satisfy the necessary compatibilities, it quickly follows that ${\operatorname{Hom}^c}({\mathcal D}(H),{\mathcal D}(C))\cong{\mathbb Z}_2^2$ as desired.
As a consequence, $|{\operatorname{Aut}}({\mathcal D}(D_{2n}))| = 2^{5}\cdot 3\cdot n\cdot\phi(n/2)$ whenever $n\equiv 2\bmod 4$. For $n=6$ the order is $1152=2^7\cdot 3^2$. The description and order of ${\operatorname{Aut}}({\mathcal D}(D_{2n}))$ for $n\not\equiv 2\bmod 4$ is given in [@K14].
[^1]: Research partially supported through a FABER Grant by the *Conseil régional de Bourgogne*
| {
"pile_set_name": "ArXiv"
} |
---
address:
- '$^+$P.L. Kapitza Institute for Physical Problems of RAS, 119334 Moscow, Russia'
- '$^*$Metallurg Engineering Ltd., 11415 Tallinn, Estonia'
- '$^\times$Moscow Institute of Physics and Technology, 141700 Dolgoprudny, Russia'
- '$^\circ$National Research University Higher School of Economics, 101000 Moscow, Russia'
author:
- 'V.V.Dmitriev$^+$[^1],M.S.Kutuzov$^*$,A.A.Soldatov$^{+,\times}$,A.N.Yudin$^{+,\circ}$'
title: NMR shifts in $^3$He in aerogel induced by demagnetizing fields
---
Introduction
============
As it was shown by C. Kittel [@kit48], demagnetizing fields may result in an additional frequency shift in magnetic resonance experiments at large values of the sample magnetization. The spin susceptibility of liquid $^3$He is small, and in this case Kittel shifts in bulk samples were observed in experiments with spin polarized $^3$He [@Tast; @Cand] or with thin $^3$He films [@Bozler; @free88]. In the latter case the shift is due to the presence of few ($\sim2$) atomic layers of solid paramagnetic $^3$He adsorbed on the surface. In result, the overall nuclear magnetic resonance (NMR) signal from $^3$He, containing both liquid and solid components, is observed as a single NMR line (due to fast spin exchange mechanism [@free88]) with the frequency shift as weighted average of those in liquid and solid fractions of $^3$He. Solid layers follow the Curie-Weiss law [@aho78; @saul05; @coll09] and their magnetization (as well as the Kittel shift) at low temperatures may be large. The Kittel shift also may be observable in normal $^3$He confined in different nanostructures, e.g., in aerogels consisting of nanostrands. In globally isotropic aerogel the average shift is zero, but it may appear in the anisotropic sample.
Here we study the Kittel effect in pure liquid $^3$He in two different globally anisotropic nanostructures called below nematic and planar aerogels. Nematic aerogel consists of nearly parallel strands, while in planar aerogel the strands are uniformly distributed in a plane perpendicular to the symmetry axis. The solid $^3$He adsorbed on the strands can be considered as a system of independent cylindrical surfaces oriented either along one direction (in nematic aerogel) or chaotically distributed in the plane (in planar aerogel). We use this model to calculate the Kittel frequency shifts in $^3$He for both cases and to interpret our experimental results.
Theoretical model
=================
The total magnetic susceptibility of $^3$He in aerogel is a sum of the susceptibilities of liquid ($\chi_l$) and solid ($\chi_s$) $^3$He: $$\label{susc}
\chi=\chi_l+\chi_s=\chi_l+\frac{C_s}{T-\Theta},$$ where $C_s$ is the Curie constant, $\Theta$ is the Curie temperature of solid $^3$He, $\chi_i\equiv M_i^{aero}/H$, $\bf H$ is an external magnetic field, $M_i^{aero}$ is a total magnetic moment of liquid ($i=l$) or solid ($i=s$) $^3$He per unit volume of the *aerogel* sample. $\chi_l$ is temperature independent in normal liquid $^3$He and may only decrease with temperature in superfluid $^3$He [@VW], so at low temperatures the NMR signal from solid $^3$He can prevail over that from liquid $^3$He.
In experiments with pure $^3$He in aerogel the common NMR line has the following frequency shift [@free88]: $$\label{fastexch}
\Delta\omega^\prime=\frac{\chi_l\Delta\omega_l+\chi_s\Delta\omega_s}{\chi_l+\chi_s},$$ where $\Delta\omega_l$ is a frequency shift in liquid $^3$He, $\Delta\omega_s$ is a Kittel frequency shift in solid $^3$He. Here all the shifts are measured from the Larmor frequency $\omega_L=\gamma H$, where $\gamma$ is the gyromagnetic ratio of $^3$He.
In the first approximation the solid $^3$He adsorbed on the strands of nematic and planar aerogels is a combination of cylindrical surfaces. The frequency shift in solid $^3$He at a separate strand [@Cand] is $$\label{kitstr}
\Delta\omega_s=\pi\gamma M_s^{cyl}\left(2-3\sin^2\varphi\right),$$ where $M_s^{cyl}$ is a magnetization of the solid $^3$He on a cylinder surface, $\varphi$ is an angle between $\bf H$ and the cylinder axis. In nematic aerogel strands are almost parallel to one another, so the mean frequency shift in $^3$He adsorbed on nematic aerogel is given by $$\label{kitnem}
\Delta\omega_s=\pi\gamma\frac{\chi_s}{s_V\delta}H\left(2-3\sin^2\varphi\right),$$ where $s_V$ is an effective surface area per unit volume of the aerogel, $\delta$ is a thickness of solid $^3$He layers. The shift is positive for $\varphi=0$, negative for $\varphi=\pi/2$, while the ratio of the corresponding absolute values is 2.
In planar aerogel strands are parallel to the distinguished plane. After averaging over angular distribution of the non-interacting strands in the plane we get the following value of the frequency shift: $$\label{kitpl}
\Delta\omega_s=\pi\gamma\frac{\chi_s}{s_V\delta}H\left(\frac{3}{2}\sin^2\varphi-1\right),$$ where $\varphi$ is an angle between $\bf H$ and the normal to the plane. In contrast to the case of nematic aerogel the shift is negative for $\varphi=0$, positive for $\varphi=\pi/2$, while the ratio of the corresponding absolute values is still 2.
Samples and methods
===================
In the experiments as a nematic aerogel we have used a nanomaterial called nafen [@anf] produced by ANF Technology Ltd. It consists of Al$_2$O$_3$ strands which are oriented along the same direction, have diameters $d\approx9$nm [@asad15] and length $\sim1$cm. We had two samples of nafen: nafen-243 with overall density $\rho=243$mg/cm$^3$, porosity $p=93.9$%, characteristic separation of strands $\ell\approx40$nm and nafen-910 with $\rho=910$mg/cm$^3$, $p=78$%, $\ell\approx20$nm. The sample of nafen-910 was obtained from nafen with density of 72mg/cm$^3$ by a technique described in Ref. [@vol17].
The sample of planar aerogel was produced from an aluminum silicate (mullite) nematic aerogel consisting of strands with $d\approx10$nm (see Ref. [@dmit18]). It is a fibrous network in the plane with $p=88$%, $\rho=350$mg/cm$^3$, and with characteristic lengths of separate strands of $\sim1$$\mu$m which is much bigger than their diameters.
The spin diffusion measurements in normal $^3$He confined by these nanostructures [@dmit18; @dmit15] confirm their strong anisotropy.
Samples of nafen had a form of cuboid with a side 4mm, the sample of planar aerogel was a stack of three plates with thickness $\approx1$mm and sizes $4\times4$mm. They were placed in the separate cells of our experimental chamber with a filling factor $\sim85$%. The experimental chamber was made from Stycast-1266 epoxy resin and was similar to that described in Ref. [@ask12].
![Fig.. The total magnetic susceptibility of pure $^3$He in nafen-243 normalized to $\chi_l$ in normal $^3$He. $\varphi=0$, $P=7.1$bar, $\omega_L/(2\pi)=880$kHz. The solid curve is a fit to Eq. . $\Theta=0.37$mK, $\chi_s/\chi_l\approx2.7$ at $T=T_c=1.643$mK[]{data-label="susc_naf243"}](f1.eps){width="\columnwidth"}
![Fig.. The original frequency shifts of pure $^3$He in nafen-243 in lower ($\omega_L/(2\pi)=361$kHz, circles, left $y$-axis) and higher ($\omega_L/(2\pi)=880$kHz, triangles, right $y$-axis) magnetic fields. $\varphi=0$, $P=7.1$bar. The arrow indicates the superfluid transition temperature of $^3$He in aerogel $T_{ca}\approx0.65T_c$. Solid curves are fits to Curie-Weiss law at $T>T_{ca}$[]{data-label="df_naf243"}](f2.eps){width="\columnwidth"}
Experiments were carried out using linear continuous wave NMR in magnetic fields 24–402Oe (corresponding NMR frequencies are 78–1303kHz) at pressures s.v.p.–29.3bar. The purity of used $^3$He was about 0.01% in experiments with nafen and 0.07% in experiments with planar aerogel, which for the surfaces inside the experimental chamber dominated by the heat exchanger with area of $\approx40$m$^2$ corresponds to preplating of aerogel strands with $\sim0.1$ and $\sim0.6$ atomic layers of $^4$He respectively. We were able to rotate $\bf H$ by any angle $\varphi$ defined in the previous section. Additional gradient coils were used to compensate the magnetic field inhomogeneity. The necessary temperatures were obtained by a nuclear demagnetization cryostat and measured by a quartz tuning fork. Below the superfluid transition temperature ($T_c$) of bulk $^3$He the fork was calibrated by Leggett frequency measurements in bulk $^3$He-B. Above $T_c$ the temperature was determined in assumption that the resonance linewidth of the fork in normal $^3$He is inversely proportional to the temperature [@fork].
Results and discussions
=======================
In all samples the measured magnetic susceptibility, determined from the intensity of the NMR absorption line, has a clear Curie-Weiss behavior (Fig. \[susc\_naf243\]). Due to the fast exchange between liquid and solid $^3$He atoms we observe a single NMR line. In Fig. \[df\_naf243\] examples of temperature dependencies of the frequency shift in $^3$He in nafen-243 are shown. In high magnetic field the shift (triangles) is mostly determined by the Kittel shift from the surface solid $^3$He, while at lower field the kink is observed on the data (circles) indicating a transition to superfluid $^3$He. At $T<T_c$ the magnetic susceptibility of solid $^3$He in the sample $\chi_s\gg\chi_l$, so when liquid $^3$He in aerogel is normal ($\Delta\omega_l=0$) from Eqs. (\[susc\],\[fastexch\],\[kitnem\]) it follows that $\Delta\omega^\prime\approx\Delta\omega_s\propto1/(T-\Theta)$ (solid lines in Fig. \[df\_naf243\]). In superfluid $^3$He the frequency shift is usually inversely proportional to the magnetic field $\Delta\omega_l\propto1/H$ [@VW], while in the solid $^3$He on the aerogel strands $\Delta\omega_s\propto H$ according to Eqs. (\[kitnem\],\[kitpl\]). Therefore, low magnetic fields in NMR experiments allow to get rid of the Kittel effect originating from anisotropy of the aerogel and to investigate purely superfluid properties of $^3$He in aerogel. On the other hand, in high magnetic fields the superfluid frequency shift can be significantly suppressed with respect to the Kittel shift.
![Fig.. The Kittel shift in solid $^3$He in nafen-910 versus $\chi_s/\chi_l$ at $P=7.1$bar for $\varphi=0$ (circles) and $\varphi=\pi/2$ (triangles) (at $P=29.3$bar for $\varphi\approx55^\circ$ in the inset). $\omega_L/(2\pi)=361.4$kHz ($\omega_L/(2\pi)=78.4$kHz in the inset). The dependencies are implicit functions of $T$, e.g., triangles correspond to temperatures from 23mK down to $\sim0.6$mK. Solid lines are linear fits with the ratio of slopes $\approx1.9$. Dashed lines are theoretical predictions (see text)[]{data-label="df_naf910"}](f3.eps){width="\columnwidth"}
![Fig.. The Kittel shift in solid $^3$He in planar aerogel versus $\chi_s/\chi_l$ at s.v.p. ($P\approx0$bar) for $\varphi=0$ (circles), $\varphi=\pi/2$ (triangles), and $\varphi\approx55^\circ$ (squares). Triangles and squares ($\omega_L/(2\pi)=588.6$kHz) are recalculated to match the Larmor frequency of circles ($\omega_L/(2\pi)=1303.2$kHz). The dependencies are implicit functions of $T$, e.g., circles correspond to temperatures from 14mK down to $\sim0.5$mK. Solid lines are linear fits with the ratio of slopes $\approx2.1$. Dashed lines are theoretical predictions (see text)[]{data-label="df_planar"}](f4.eps){width="\columnwidth"}
The Kittel effect is more clearly manifested in nafen-910 which is denser than nafen-243. Superfluidity of $^3$He in presence of solid $^3$He on the aerogel strands is completely suppressed in nafen-910 [@mag], so using Eq. we can determine $\Delta\omega_s$ from measurements of $\Delta\omega^\prime$ down to the lowest attained temperatures (see Fig. \[df\_naf910\]). The shift is positive in the magnetic field parallel to the nafen anisotropy axis ($\varphi=0$) and negative in the transverse direction of the field ($\varphi=\pi/2$). The absolute value of the ratio of the corresponding shifts is $\approx1.9$ which is in a good agreement with Eq. , while the data in the inset (squares) demonstrate the absence of the shift at $\sin^2\varphi=2/3$.
To estimate the expected value of $\Delta\omega_s$ using Eq. , we need to know values of $\chi_s$, $s_V$, and $\delta$. First, $\chi_s$ can be found from measurements of the total magnetization of the sample that allows to determine $\chi_s/\chi_l$. We note that $\chi_l=p\chi_l^{bulk}\approx4.25\cdot10^{-8}$emu, where $p=0.78$ is the porosity of nafen-910 and $\chi_l^{bulk}=5.45\cdot10^{-8}$emu is the magnetic susceptibility in bulk normal $^3$He at $P=7.1$bar. Second, $s_V$ can be estimated in the assumption that the nafen strands are ideal cylinders. In this case $s_V=\frac{4}{d}\frac{\rho}{\rho_0}\approx102$m$^2$/cm$^3$, where $d\approx9$nm is a strand diameter, $\rho=910$mg/cm$^3$ is nafen-910 density, and $\rho_0=3.95$g/cm$^3$ is Al$_2$O$_3$ density. Third, the solid $^3$He on the surfaces at $P=7.1$bar has a coverage of $\approx1.6$ atomic layers [@sch87]. The thickness of one layer may be estimated as $3.5$ (the lattice period of $^3$He crystal at low temperatures) that gives $\delta\approx5.5$. In result, theoretical predictions according to Eq. are plotted in Fig. \[df\_naf910\] as dashed lines which are surprisingly close to the experimental results, considering a rather bad accuracy in estimations of $\chi_s$, $s_V$, and $\delta$ ($\pm20$%).
The frequency shift measurements in solid $^3$He in planar aerogel (see Fig. \[df\_planar\]) are also in agreement with the theory (Eq. ). The shift for $\varphi=0$ is negative and $\approx2.1$ larger than that for the field along the plane ($\varphi=\pi/2$) which is positive. At $\sin^2\varphi=2/3$ the shift is zero as expected from Eq. . Estimation of the shift values in planar aerogel using Eq. gives the same order of magnitude as in experiments. In this case $\chi_l=p\chi_l^{bulk}\approx3.11\cdot10^{-8}$emu (where $p=0.88$ and $\chi_l^{bulk}=3.53\cdot10^{-8}$emu at s.v.p.), $s_V=\frac{4}{d}\frac{\rho}{\rho_0}\approx47$m$^2$/cm$^3$ (where $d\approx10$nm, $\rho=350$mg/cm$^3$ is planar aerogel density, and $\rho_0\approx3$g/cm$^3$ is mullite density), and $\delta\approx2.8$ at s.v.p. for 0.8 atomic layers of solid $^3$He [@sch87] (here we assume that $\sim0.6$ atomic layers of solid $^3$He is replaced by $^4$He [@free90] due to a rather “dirty” $^3$He with 0.07% of $^4$He used in the experiment).
Conclusions
===========
We have observed NMR shifts due to the Kittel effect in $^3$He confined in aerogel-like nanostructures with different types of the global anisotropy and demonstrated that values of the shift well agree with the theoretical expectations. At ultralow temperatures even in moderate magnetic fields these shifts may be large enough to mask the $^3$He superfluid transition but can be avoided by using lower magnetic fields or by choosing the proper angle between the axis of the anisotropy and the magnetic field.
This work was supported by grant of the Russian Science Foundation (project \#18-12-00384).
[99]{} C. Kittel, Phys. Rev. [**73**]{}, 155 (1948). G. Tastevin, J. Low Temp. Phys. [**89**]{}, 317 (1992). D. Candela, M.E. Hayden, and P.J. Nacher, Phys. Rev. Lett. [**73**]{}, 2587 (1994). H.M. Bozler and D.M. Bates, Phys. Rev. B [**27**]{}, 6992 (1983). M.R. Freeman, R.S. Germain, E.V. Thuneberg, and R.C. Richardson, Phys. Rev. Lett. [**60**]{}, 596 (1988). A.I. Ahonen, T.A. Alvesalo, T. Haavasoja, and M.C. Veuro, Phys. Rev. Lett. [**41**]{}, 494 (1978). J.A. Sauls, Yu.M. Bunkov, E. Collin, H. Godfrin, and P. Sharma, Phys. Rev. B [**72**]{}, 024507 (2005). E. Collin, S. Triqueneaux, Yu.M. Bunkov, and H. Godfrin, Phys. Rev. B [**80**]{}, 094422 (2009). D. Vollhardt and P. Wölfle, *The Superfluid Phases of Helium 3* (Taylor & Francis, London, 1990). <http://www.anftechnology.com> V.E. Asadchikov, R.Sh. Askhadullin, V.V. Volkov, V.V. Dmitriev, N.K. Kitaeva, P.N. Martynov, A.A. Osipov, A.A. Senin, A.A. Soldatov, D.I. Chekrygina, and A.N. Yudin, JETP Lett. [**101**]{}, 556 (2015). V.V. Volkov, V.V. Dmitriev, D.V. Zolotukhin, A.A. Soldatov, A.N. Yudin, Instrum. Exp. Tech. [**60**]{}, 737 (2017). V.V. Dmitriev, M.S. Kutuzov, L.A. Melnikovsky, B.D. Slavov, A.A. Soldatov, and A.N. Yudin, JETP Lett. [**108**]{}, Issue 11, (2018). V.V. Dmitriev, L.A. Melnikovsky, A.A. Senin, A.A. Soldatov, and A.N. Yudin, JETP Lett. [**101**]{}, 808 (2015). R.Sh. Askhadullin, V.V. Dmitriev, D.A. Krasnikhin, P.N. Martinov, A.A. Osipov, A.A. Senin, and A.N. Yudin, JETP Lett. [**95**]{}, 326 (2012). R. Blaauwgeers, M. Blažková, M. Človečko, V.B. Eltsov, R. de Graaf, J. Hosio, M. Krusius, D. Schmoranzer, W. Schoepe, L. Skrbek, P. Skyba, R.E. Solntsev, and D.E. Zmeev, J. Low Temp. Phys. [**146**]{}, 537 (2007). V.V. Dmitriev, A.A. Soldatov, and A.N. Yudin, Phys. Rev. Lett. [**120**]{}, 075301 (2018). A. Schuhl, S. Maegawa, M.W. Meisel, and M. Chapellier, Phys. Rev. B [**36**]{}, 6811 (1987). M.R. Freeman and R.C. Richardson, Phys. Rev. B [**41**]{}, 11011 (1990).
[^1]: e-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this note we discuss a parameter $\sigma$ on weighted $k$-element multisets of $[n]= \{1,\dots ,n\}$. The sums of weighted $k$-multisets relate $k$-subsets, $k$-multisets, as well as special instances of (truncated) interpolated multiple zeta values, as introduced by Yamamoto [@Y]. We study properties of this parameter using symbolic combinatorics. We (re)derive and extend certain identities for $\ztt(\{m\}_k)$, as well as $\ztt_n(\{m\}_k)$. Moreover, we introduce random variables on the $k$-element multisets and derive their distributions, as well as limit laws for $k$ or $n$ tending to infinity.'
author:
- Markus Kuba
title: 'On multisets, interpolated multiple zeta values and limit laws'
---
Introduction
============
The [*multiple zeta values*]{} were introduced by Hoffman [@H92] and Zagier [@Z]. They are defined as $$\zt(i_1,\dots,i_k)=\sum_{\ell_1>\cdots>\ell_k\ge 1}\frac1{\ell_1^{i_1}\cdots \ell_k^{i_k}},$$ with admissible indices $(i_1,\dots,i_k)$ satisfying $i_1\ge 2$, $i_j\ge 1$ for $2\le j\le k$. Their [*truncated*]{} counterparts, sometimes called multiple harmonic sums, are given by $$\zt_n(i_1,\dots,i_k)=\sum_{n\ge \ell_1>\cdots>\ell_k\ge 1}\frac1{\ell_1^{i_1}\cdots \ell_k^{i_k}}.$$ We refer to $i_1 +\dots + i_k$ as the weight of this multiple zeta value, and $k$ as its depth. For a comprehensive overview as well as a great many pointers to the literature we refer to the survey of Zudilin [@Zu].
An important variant of the (truncated) multiple zeta values are the so-called multiple zeta [*star values*]{}, where equality is allowed: $$\begin{split}
\zts(i_1,\dots,i_k) & = \sum_{\ell_1\ge \cdots\ge \ell_k\ge 1}\frac1{\ell_1^{i_1}\cdots \ell_k^{i_k}} \quad \text{and}\\
\zts_{n}(i_1,\dots,i_k) & =\sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}\frac1{\ell_1^{i_1}\cdots \ell_k^{i_k}}.
\end{split}$$ On the combinatorial side, it is well known that two special cases, namely the series $\zt_{n}(\{1\}_k)$ and $\zts_{n}(\{1\}_k)$ occur in a multitude of different places. See for example [@KP] and the references therein for different representation of the two series. The value $\zt_{n}(\{1\}_k)$ is closely related to the Stirling numbers of the first kind, and $\zts_{n}(\{1\}_k)$ is related to alternating binomial sums. For non-truncated series $\zt$ and $\zts$, Yamamoto [@Y] introduced a generalization of both versions called [*interpolated multiple zeta values*]{}. Noting that, $$\zts(i_1,\dots,i_k)=\sum_{\circ = \text{``},\text{''} \text{or} \, \text{``}+\text{''}}\zt(i_1\circ i_2 \dots \circ i_k),$$ let the parameter $\sigma$ denote the number of plus in the expression $i_1\circ i_2 \dots \circ i_k$. Yamamoto defines $$\label{EqYamamoto}
\ztt(i_1,\dots,i_k) = \sum_{\circ = \text{``},\text{''} \text{or} \, \text{``}+\text{''}}t^{\sigma}\zt(i_1\circ i_2 \dots \circ i_k).$$ Thus, the series $\ztt(i_1,\dots,i_k)$ interpolates between multiple zeta values, case $t=0$, and multiple zeta star values, case $t=1$. It turned out that the interpolated series satisfies many identities generalizing or unifying earlier result for multiple zeta and zeta star values, see for example Yamamoto [@Y], Hoffman and Ihara [@HI] or Hoffman [@H2000; @H2018]. In particular, a so-called quasi-shuffle product [@H2000; @H2018; @HI; @Y], sometimes also called stuffle product, can be defined for the interpolated zeta values. Hoffman and Ihara used an algebra framework, which leads amongst others to expressions for interpolated multiple zeta values $\ztt(\{m\}_{k})$ in terms of Bell polynomials and ordinary single argument zeta values, $m\ge 1$. We discuss and (re-)derive in this work the results for $\ztt_{n}(\{m\}_{k})$ and $\ztt(\{m\}_{k})$ using symbolic combinatorics.
We study in particular the interpolated truncated multiple zeta values: we are interested in the truncated series $$\begin{aligned}
\label{EqYamamotoOrdered}
\ztt_{n}(\{1\}_{k}) & =\sum_{\circ = \text{``},\text{''} \text{or} \, \text{``}+\text{''}}t^{\sigma}\zt_{n}(\underbrace{1 \circ 1 \circ \dots \circ 1}_{k}) $$ and their properties. Note that $\ztt_{n}(\{1\}_{k})$ interpolates between $\zt_n(\{1\}_k)$ and $\zts_n(\{1\}_k)$, $t=0$ and $t=1$, respectively.
On the other hand, classical combinatorial problems are the enumeration of $k$-element multisets, short [*$k$-multisets*]{}, of $[n]=\{1,\dots,n\}$, leading to the multiset coefficients, $\binom{n+k-1}{k}$ and the enumeration of $k$-element sets, short [*$k$-subsets*]{}, of $[n]=\{1,\dots,n\}$ leading to the binomial coefficients, $\binom{n}{k}$.
The goal of this work is to unify these two topics, namely interpolated (truncated) multiple zeta values and the enumeration of $k$-subsets and $k$-multisets. We study $k$-[*multisets*]{} of $[n]$ and introduce a [*weighted enumeration*]{} of $k$-multisets of $[n]$ in terms of a given weight sequence $\boldsymbol{a}=(a_j)_{j\in{\ensuremath{\mathbb{N}}}}$. Our weighted enumeration also takes into account an additional parameter $\sigma=\sigma_{n,k}$. It is defined similarly to Yamamoto [@Y]: $\sigma$ counts the number of equalities in the nested sum representation, corresponding to the number of plus signs in or . The weighted enumeration also has an [*algebraic interpretation*]{}, connecting elementary symmetric functions $e_k$ and complete symmetric functions $h_k$; moreover power sum $p_ks$ also appear in an alternative representation.
In the next section we introduce our weighted enumeration of $k$-multisets, given an arbitrary sequence $(a_j)_{j\in {\ensuremath{\mathbb{N}}}}$ of positive reals, $a_j>0$. Closely related ideas have been considered by Vignat and Takhare [@VW], who considered (non-truncated) generalized multiple zeta values and infinity products, with $(a_j)_{j\in {\ensuremath{\mathbb{N}}}}$ given by the zeros of certain special functions. Moreover, Bachmann very briefly discussed related ideas at the end of his article [@B].
We study two aspects of the weighted enumerations: first, its relation to $k$-sets and $k$-multisets and to the interpolated (truncated) multiple zeta function, second, the probabilistic aspects of $\sigma$. Introducing suitably defined random variables $S_{n,k}$, we study its distribution as well as limit laws for $k$ or $n$ tending to infinity. In particular, we show that the parameter $\sigma$ in $k$-multisets follows a hypergeometric distribution and derive Poisson and Gaussian limit laws. For $\zts_n(\{1\}_k)$ we show that also different limit laws occur. In a later section we introduce refinements of the parameter $\sigma$ an also discuss related random variables. It turns out that the random variables $S_{n,k}$, its refinements and its limit laws are closely related to many quantities and techniques studied earlier in combinatorial probability theory: maxima in hypercubes [@Bai; @Hwang2004], unsuccessfull seach in binary search tree [@DobSmy1996], descendants in increasing trees [@KP2006], edge-weighted increasing trees [@KP2008; @KW], distances in increasing trees [@Dob1996; @DobSmy1996], leaf-isolation procedures in random trees [@KP2008b], as well as asymptotics of the Poisson distribution [@DobSmy1996; @Hwang1999].
Throughout this work use the notation $H_n=\zt_n(1)$ for the $n$th harmonic number, $H_n^{(j)}=\zt_n(j)$ for the $n$th generalized harmonic number of order $j$. We denote with $\mathcal{N}(0,1)$ a standard normal distributed random variable, with $\operatorname{\text{Po}}(\lambda)$ a Poisson distributed random variable with parameter $\lambda$, and with $\operatorname{\text{Hypergeo}}(N,K,n)$ a hypergeometric distributed random variable with parameters $N,K,n$. Moreover, we denote with ${\ensuremath{x^{\underline{k}}}}$ the $k$th falling factorial, ${\ensuremath{x^{\underline{k}}}}=x(x-1)\dots(x-(k-1))$, $k\in{\ensuremath{\mathbb{N}}}$, with ${\ensuremath{x^{\underline{0}}}}=1$. Furthermore, we use the abbreviation $E_x$ for the evaluation operator at $x=1$. Moreover, we denote by $X {\ensuremath{\stackrel{(d)}=}}Y$ the equality in distribution of the random variables $X$ and $Y$, and by $X_{n} {\ensuremath{\xrightarrow{(d)}}}X$ convergence in distribution of the sequence of random variables $X_{n}$ to a random variable $X$. We denote with $P_O(k)$ the set of ordered partitions of the integer $k$, often also called compositions of the integer $k$, and ${\ensuremath{\mathcal{L}}}(\mathbf{p})$ the length of a partition $\mathbf{p}$, defined as the number of its summands. E.g., for $k=3$ we have $P_O(3)=\{(1,1,1), (2,1), (1,2),(3)\}$ and $$P_O(4)=\{(1,1,1,1), (1,1,2), (1,2,1),(2,1,1),(1,3),(3,1),(4)\}.$$ Concerning length of partitions, we have ${\ensuremath{\mathcal{L}}}((1,1,1,1))=4$ and ${\ensuremath{\mathcal{L}}}((4))=1$.
Auxiliary results about probability distributions\[SubProb\]
------------------------------------------------------------
For the reader’s convenience we collect a few basic facts about the probability distributions, mentioned before, appearing later in our analysis.
A beta-distributed random variable $Z {\ensuremath{\stackrel{(d)}=}}\beta(\alpha,\beta)$ with parameters $\alpha,\beta>0$ has a probability density function given by $f(x)=\frac1{B(\alpha,\beta)}x^{\alpha-1}(1-x)^{\beta-1}$, where $B(\alpha,\beta)=\frac{\Gamma(\alpha)\Gamma(\beta)}{\Gamma(\alpha+\beta)}$ denotes the Beta-function. The (power) moments of $Z$ are given by $${\ensuremath{\mathbb{E}}}(Z^s)=\frac{\prod_{j=0}^{s-1}(\alpha+j)}{\prod_{j=0}^{s-1}(\alpha+\beta+j)}
=\frac{{\ensuremath{(\alpha+s-1)^{\underline{s}}}}}{{\ensuremath{(\alpha+\beta+s-1)^{\underline{s}}}}},\quad s\ge 1.$$ The beta-distribution is uniquely determined by the sequence of its moments. In this work we will discuss a beta-distributed random variable $Z {\ensuremath{\stackrel{(d)}=}}\beta(1,n-1)$, with moments given by ${\ensuremath{\mathbb{E}}}(Z^s)=\frac{s!}{{\ensuremath{(n+s-1)^{\underline{s}}}}}$, for $s\ge 1$.
An exponentially distributed random variable $Z=\operatorname{\text{Exp}}(1)$ with parameter one has density $f(x)=e^{-x}$, $x\ge 0$ and power moments ${\ensuremath{\mathbb{E}}}(Z^s)=s!$.
A Bernoulli distributed random variable $B{\ensuremath{\stackrel{(d)}=}}\operatorname{\text{Be}}(p)$ with parameter $p\in[0,1]$ has probability mass function $${\ensuremath{\mathbb{P}}}\{B=1\}=p,\quad {\ensuremath{\mathbb{P}}}\{B=0\}=1-p.$$ The sum $Z_n=\bigoplus_{j=1}^{n}B_j$ of $n$ independent Bernoulli distributed random variables $B_j=\operatorname{\text{Be}}(p_j)$ has probability generating function $${\ensuremath{\mathbb{E}}}(v^{Z_n})=\prod_{j=1}^{n}{\ensuremath{\mathbb{E}}}(t^{B_j})=\prod_{j=1}^{n}\big(1+(v-1)p_j\big).$$ The factorial moments ${\ensuremath{\mathbb{E}}}({\ensuremath{Z_n^{\underline{s}}}})$ of $Z_n$ are given by $${\ensuremath{\mathbb{E}}}({\ensuremath{Z_n^{\underline{s}}}})=E_v\frac{\partial^s}{\partial v^s}{\ensuremath{\mathbb{E}}}(v^{Z_n})
= s!\cdot\sum_{n \ge \ell_1> \cdots> \ell_s\ge 1}p_{\ell_1}p_{\ell_2}\dots p_{\ell_s}.$$ Under the assumption that the probabilities $p_j$ tend to zero fast enough, we may define $Z=Z_\infty$ as the sum of infinitely many Bernoulli random variable: $$Z=\bigoplus_{j=1}^{\infty}B_j,$$ with factorial moments formally given by $s!$ times a multiple series: $${\ensuremath{\mathbb{E}}}({\ensuremath{Z^{\underline{s}}}})= s!\cdot\sum_{\ell_1> \cdots> \ell_s\ge 1}p_{\ell_1}p_{\ell_2}\dots p_{\ell_s}.$$
A Poisson distributed random variable $Z{\ensuremath{\stackrel{(d)}=}}\operatorname{\text{Po}}(\lambda)$, $\lambda >0$, has probability mass function and factorial moments given by $${\ensuremath{\mathbb{P}}}\{Z=j\}=\frac{\lambda^j}{j!}\cdot e^{\lambda},\,j\ge 0,\quad {\ensuremath{\mathbb{E}}}({\ensuremath{Z^{\underline{s}}}})=\lambda^s,\,s\ge 1.$$
A hypergeometric distributed random variable $Z{\ensuremath{\stackrel{(d)}=}}\operatorname{\text{Hypergeo}}(N,K,n)$ with parameters $N,K,n$ has probability mass function and factorial moments given by $${\ensuremath{\mathbb{P}}}\{Z=j\}=\frac{\binom{K}{j}\binom{N-K}{n-j}}{\binom{N}{n}},0\le j\le n\quad
{\ensuremath{\mathbb{E}}}({\ensuremath{Z^{\underline{s}}}})=\frac{{\ensuremath{K^{\underline{s}}}}{\ensuremath{n^{\underline{s}}}}}{{\ensuremath{N^{\underline{s}}}}},\,s\ge 1.$$ In particular, $${\ensuremath{\mathbb{E}}}(Z)=n\cdot\frac{ K}{N},\quad {\ensuremath{\mathbb{V}}}(Z)=n\cdot \frac{K(N-K)(N-n)}{N^2(N-1)}.$$ Moreover, the following normal limit law can be deduced (see Nicholson [@N] or Feller [@F]).
Let $Z{\ensuremath{\stackrel{(d)}=}}\operatorname{\text{Hypergeo}}(N,K,n)$ denote a hypergeometric distributed random variable. Under the assumption $\min\{K,N\}\to\infty$ and $n$ such that ${\ensuremath{\mathbb{E}}}(Z),{\ensuremath{\mathbb{V}}}(Z)\to\infty$, the random variable $Z$, centered and normalized, is asymptotically standard normal distributed: $$\frac{Z-{\ensuremath{\mathbb{E}}}(Z)}{\sqrt{{\ensuremath{\mathbb{V}}}(Z)}}{\ensuremath{\xrightarrow{(d)}}}\mathcal{N}(0,1).$$
Interpolated weighted Multisets and parameter sigma
===================================================
Parameter sigma
---------------
We consider $k$-multisets of $[n]= \{1,\dots ,n\}$. Strictly speaking, we do not directly consider multisets. Instead, we induce an increasing order on the distinct elements in the multisets and consider the ordered sequences. For example, the unordered multiset $\{1,1,4,2,2,6,2\}$ corresponds to the increasing sequence $(1,1,2,2,2,4,6)$. Let ${\ensuremath{\mathcal{M}}}_{n,k}$ denote the set $${\ensuremath{\mathcal{M}}}_{n,k}=\{\vec{\ell}=(\ell_{k},\ell_{k-1}\dots,\ell_1)\in{\ensuremath{\mathbb{N}}}^k\colon 1\le \ell_k\le \dots \le \ell_2\le \ell_1 \le n\},$$ such that each $\vec{\ell}$ corresponds to a $k$-multiset of $[n]$. In a slight abuse of notation, we denote with $| \vec{\ell} |=k$ the length of $\vec{\ell}$, in other words the cardinality of the corresponding multiset. The cardinality of ${\ensuremath{\mathcal{M}}}_{n,k}$ is given by the multiset coefficients, counting the number of $k$-multisets of $[n]$: $$|{\ensuremath{\mathcal{M}}}_{n,k}|=\sum_{\ell_1=1}^n \sum_{\ell_2=1}^{\ell_1}\dots \sum_{\ell_k=1}^{\ell_{k-1}}1=\binom{n+k-1}{k}.$$ Closely related is the set $${\ensuremath{\mathcal{S}}}_{n,k}=\{\vec{\ell}=(\ell_k,\dots,\ell_1)\in{\ensuremath{\mathbb{N}}}^k\colon 1\le \ell_k< \dots < \ell_2< \ell_1 \le n\},$$ whose cardinality is given by binomial coefficients, counting the number of $k$-sets of $[n]$: $$|{\ensuremath{\mathcal{S}}}_{n,k}|=\sum_{\ell_1=1}^n \sum_{\ell_2=1}^{\ell_1-1}\dots \sum_{\ell_k=1}^{\ell_{k-1}-1}1=\binom{n}{k}.$$
In the following we introduce the parameter $\sigma=\sigma_{n,k}$ for elements $\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}$.
Let $\vec{\ell}\in {\ensuremath{\mathcal{M}}}_{n,k}$. Then $\sigma=\sigma_{n,k}(\vec{\ell})$ is defined as the number equal values in $\vec{\ell}$, or in other words $$\sigma(\vec{\ell}) =\sigma(\ell_1,\dots,\ell_k) = |\{1 \le j \le k - 1\colon\,\, \ell_j = \ell_{j+1} \}|.$$
Given $\vec{\ell}=(1,1,2,2,2,4,6)\in{\ensuremath{\mathcal{M}}}_{n,7}$ then $\sigma(\vec{\ell})=3$. Given $\vec{\ell}=(5,5,5,5)\in{\ensuremath{\mathcal{M}}}_{n,3}$ then $\sigma(\vec{\ell})=3$.
The parameter $\sigma$ relates ${\ensuremath{\mathcal{M}}}_{n,k}$ and ${\ensuremath{\mathcal{S}}}_{n,k}$: $${\ensuremath{\mathcal{S}}}_{n,k}=\{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}\colon \sigma(\vec{\ell})=0\}.$$
Next we define multiplicative weights for $\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}$.
Given an arbitrary sequence weight sequence $\boldsymbol{a}=(a_{j})_{j\in {\ensuremath{\mathbb{N}}}}$ of positive reals, the weight $w(\vec{\ell})=w(\vec{\ell},\boldsymbol{a})$ of $\vec{\ell}=(\ell_1,\dots,\ell_k)\in{\ensuremath{\mathcal{M}}}_{n,k}$ is defined in terms of $\boldsymbol{a}$ as $$w(\vec{\ell})=\prod_{j=1}^{k}a_{\ell_{j}}.$$
Using the two definitions stated before, we are ready to define our weighted enumerations of ${\ensuremath{\mathcal{M}}}_{n,k}$.
Given the weight sequence $\boldsymbol{a}=(a_{j})_{j\in {\ensuremath{\mathbb{N}}}}$. Let $\theta_{n;k}(t)=\theta_{n;k}(t;\boldsymbol{a})$ be defined as $$\theta_{n;k}(t)=\sum_{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}}w(\vec{\ell})t^{\sigma(\vec{\ell})}
=\sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}a_{\ell_1}a_{\ell_2}\dots a_{\ell_k}t^{\sigma(\vec{\ell})}.$$
Note that we use the convention $\theta_{n;0}(t)=1$ in the boundary case $k=0$.
The construction is also well defined for $${\ensuremath{\mathcal{M}}}_{\infty,k}=\{\vec{\ell}=(\ell_k,\dots,\ell_1)\in{\ensuremath{\mathbb{N}}}^k\colon 1\le \ell_k\le \dots \le \ell_2\le \ell_1 <\infty\},$$ as long as the weight sequence $\boldsymbol{a}=(a_{j})_{j\in {\ensuremath{\mathbb{N}}}}$ is chosen in such a way that $\theta_{\infty;k}(t)=\sum_{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{\infty,k}}w(\vec{\ell})t^{\sigma(\vec{\ell})}$ converges. This construction is very closely related to the generalized multiple zeta values of Vignat and Wakhare [@VW].
\[RemarkSym\] Regarding the weight sequence $\boldsymbol{a}=(a_{j})_{j\in {\ensuremath{\mathbb{N}}}}$ as variables, the values $\theta_{n;k}(t)$ can also be interpreted in an algebraic way. This is no surprise, due to the well known relationship between the multiple zeta values and quasi-symmetric functions [@H2000; @H2018; @HI; @VW; @Z]. They relate complete elementary symmetric polynomials $h_k$ and elementary symmetric polynomials $e_k$: $\theta_{n;k}(t)=\theta_{n;k}(t;\boldsymbol{a})$ at $t=0$ or $t=1$ is given by $$\theta_{n;k}(0;\boldsymbol{a})=e_k(a_1,\dots,a_n)=\sum_{n \ge \ell_1> \cdots> \ell_k\ge 1}a_{\ell_1}a_{\ell_2}\dots a_{\ell_k},$$ and $$\theta_{n;k}(1;\boldsymbol{a})=h_k(a_1,\dots,a_n)=\sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}a_{\ell_1}a_{\ell_2}\dots a_{\ell_k}.$$ We will see later that the power sums $p_k$ on $n$ variables, $$p_k(a_1,\dots,a_n)=A_n(k)=\sum_{m=1}^{n}a_m^k,$$ also appear in an alternative representation of $\theta_{n;k}(t)$.
\[ExampleMultiset\] Given the weight sequence $\boldsymbol{a}=(1)_{j\in{\ensuremath{\mathbb{N}}}}$ we obtain $$\theta_{n;k}(t)=\sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}t^{\sigma(\vec{\ell})},$$ interpolating between $k$-sets and multisets of $[n]$. The special values $\theta_{n;k}(0)$ and $\theta_{n;k}(1)$ enumerate $k$-sets and multisets of $[n]$: $$\theta_{n;k}(0)=|{\ensuremath{\mathcal{S}}}_{n,k}|=\binom{n}{k},\quad \theta_{n;k}(1)=|{\ensuremath{\mathcal{M}}}_{n,k}|=\binom{n+k-1}{k}.$$
Given the weight sequence $\boldsymbol{a}=(j)_{j\in{\ensuremath{\mathbb{N}}}}$ we obtain $$\theta_{n;k}(t)=\sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}\ell_1\cdots \ell_k\cdot t^{\sigma(\vec{\ell})}.$$ The special values are given in terms of the Stirling numbers of the first and second kind [@GraKnuPa]: $$\theta_{n;k}(0)={\genfrac{ [ }{ ] }{0pt}{}{n+1}{n+1-k}}=n!\cdot\zt_{n}(\{1\}_{n+k}),\quad \theta_{n;k}(1)={\genfrac{ \{ }{ \} }{0pt}{}{n+k}{n}}.$$
Given the weight sequence $\boldsymbol{a}=(\frac1{j^{m}})_{j\in{\ensuremath{\mathbb{N}}}}$, $m>0$ we obtain a special instance of the truncated interpolated multiple zeta values, $$\theta_{n;k}(t)=\ztt_n(\{m\}_k)=\sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}\frac{t^{\sigma(\vec{\ell})}}{\ell_1^{m}\cdots \ell_k^{m}},$$ such that $$\theta_{n;k}(0)=\zt_n(\{m\}_k),\qquad \theta_{n;k}(1)=\zts_n(\{m\}_k).$$
Ordered partitions and Interpolated multiple zeta functions
-----------------------------------------------------------
Let $\vec{\ell}=(\ell_1,\dots,\ell_k)\in{\ensuremath{\mathcal{M}}}_{n,k}$. We can associate an ordered partition $\mathbf{p}\in P_O(k)$ to $\vec{\ell}$ as follows: there exist integers $1\le s\le k$ and $n\ge j_1>j_2>\dots>j_s\ge 1$ with multiplicities $r_1,\dots,r_s\in {\ensuremath{\mathbb{N}}}_0$ with $\sum_{i=1}^{s}r_i=k$ such that $$\vec{\ell}=\vec{j}=(j_s^{r_s},\dots,j_1^{r_1}).$$ Here, $j_i^{r_i}$ denotes $j_i$ exactly $r_i$ times. We refer to the ordered partition $\mathbf{p}=\mathbf{p}(\vec{j})=(r_1,\dots,r_s)\in P_O(k)$ as the shape of $\vec{j}$. The shape, as a map, is a surjection from ${\ensuremath{\mathcal{M}}}_{n,k}$ to $P_O(k)$. Note that $$\sigma(\vec{j})=\sum_{i=1}^{s}(r_i-1)=k-s=k-{\ensuremath{\mathcal{L}}}(\mathbf{p}).$$ We may write $$\theta_{n;k}(t)=\sum_{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}}a_{\ell_1}\cdot\dots\cdot a_{\ell_k} t^{\sigma(\vec{\ell})}
=\sum_{\mathbf{p}\in P_O(k)}\sum_{\substack{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}\\\text{shape}(\vec{\ell})=\mathbf{p} } } a_{\ell_1}\cdot\dots\cdot a_{\ell_k} t^{\sigma(\vec{\ell})}$$ Thus, for $\boldsymbol{a}=(\frac1{j^{m}})_{j\in{\ensuremath{\mathbb{N}}}}$ we get $$\begin{split}
\theta_{n;k}(t)
&= \sum_{\mathbf{p}\in P_O(k)}\sum_{\substack{\vec{j}\in{\ensuremath{\mathcal{M}}}_{n,k}\\\text{shape}(\vec{j})=(r_1,\dots,r_s)=\mathbf{p} } } \frac1{j_1^{m\cdot r_1}}\cdot\dots\cdot \frac1{j_s^{ m\cdot r_s }} t^{\sigma(\vec{j})}\\
&= \sum_{\mathbf{p}\in P_O(k)}\sum_{\substack{n\ge j_1>j_2>\dots>j_s\ge 1\\\text{shape}(\vec{j})=(r_1,\dots,r_s)=\mathbf{p} } } \frac1{j_1^{m\cdot r_1}}\cdot\dots\cdot \frac1{j_s^{ m\cdot r_s }} t^{\sigma(\vec{j})}\\
&=\sum_{\mathbf{p}\in P_O(k)}\zt_n(m\cdot r_1,\dots,m\cdot r_s)t^{k-s}=\sum_{\mathbf{p}\in P_O(k)}\zt_n(m\cdot \mathbf{p})t^{k-{\ensuremath{\mathcal{L}}}(p)}.
\end{split}$$
Generating functions and Bell polynomials
=========================================
Let ${\ensuremath{\mathcal{M}}}_n$ denote all multisets of $[n]$, ${\ensuremath{\mathcal{M}}}_n=\bigcup_{k=1}^{\infty}{\ensuremath{\mathcal{M}}}_{n,k}$. In order to gain more insight into $\theta_{n;k}(t)$ we use symbolic combinatorics and study the generating function $\Theta_{n}(z,t)$, defined by $$\Theta_{n}(z,t)=\sum_{k\ge 0}\theta_{n;k}(t)z^k=\sum_{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n}}w(\vec{\ell})t^{\sigma(\vec{\ell})}z^{|\vec{\ell}|}.$$
\[ThmGF\] The generating function $\Theta_{n}(z,t)$ is given by $$\Theta_{n}(z,t)=\prod_{m=1}^n\Big(1+\frac{a_m z}{1-a_m z t}\Big).$$ Moreover, let $A_n(j)$ denote the power sums $A_n(j)=\sum_{m=1}^{n}a_m^j$. Then, $$\Theta_{n}(z,t)=\exp\Big(\sum_{j=1}^{\infty}\frac{z^j}{j}A_n(j)\big( t^j-(t-1)^j\big)\Big).$$
In special cases it is possible to derive the complete generating function $T(x,z,t)=\sum_{n\ge 0}\Theta_n(z,t)x^n=\sum_{n\ge 0}\sum_{k\ge 0}\theta_{n;k}(t) x^n z^k$. For $(a_j)=(1)$ we get $$T(x,z,t)=\frac{1-zt}{(1-zt)(1-x)-zx}.$$ For $(a_j)=(\frac1j)$ we get an ordinary hypergeometric function $$T(x,z,t)={}_2F_1(1,1-z(t-1),1-z t;x).$$
We use the symbolic constructions from analytic combinatorics, see Flajolet and Sedgewick [@FS2009]. Let $\mathcal{Z}_m=\{m\}$ be a combinatorial class of size one, $1\le m\le n$. Due to the sequence construction we can describe the class of multisets $\mathcal{B}_m$ of $\mathcal{Z}_m$ as follows $$\mathcal{B}_m=\text{SEQ}(\mathcal{Z}_m)=\{\epsilon\} + \mathcal{Z}_m+ \mathcal{Z}_m\times \mathcal{Z}_m+\mathcal{Z}_m\times \mathcal{Z}_m\times \mathcal{Z}_m+\dots;$$ Thus, the generating function $$B_m(z)=\sum_{\beta\in\mathcal{B}_m}w(\beta)t^{\sigma(\beta)}z^{|\beta|}= 1+\sum_{\epsilon\neq \beta\in\mathcal{B}_m}w(\beta)t^{|\beta|-1}z^{|\beta|}$$ is given by $$B_m(z)=1+\sum_{j=1}^{\infty}t^{j-1}a_m^j z^{j}=1+\frac{a_mz}{1-a_m t z}.$$ All multisets ${\ensuremath{\mathcal{M}}}_n=\bigcup_{k=1}^{\infty}{\ensuremath{\mathcal{M}}}_{n,k}$ of $[n]$ can be combinatorially described by $${\ensuremath{\mathcal{M}}}_n=\mathcal{B}_1\times \mathcal{B}_2\times\dots \times \mathcal{B}_n.$$ Hence, the generating function $\Theta_n(z,t)$ is given by $$\Theta_n(z,t)=\prod_{m=1}^{n}B_m(z)=\prod_{m=1}^{n}\left(1+\frac{a_mz}{1-a_m t z}\right).$$
As a consequence of the theorem before, we obtain alternative expressions for the values $\theta_{n;k}(t)$.
\[PropGF\] The values $\theta_{n;k}(t)$ can be expressed in terms of values $\theta_{n;k}(0)$ and $\theta_{n;k}(1)$: $$\theta_{n;k}(t)=\sum_{j=0}^{k}t^j\theta_{n,j}(1)\cdot (1-t)^{k-j}\theta_{n,k-j}(0).
$$ Moreover, the values $\theta_{n;k}(t)$ can be expressed in terms of Bell polynomials and the power sums $A_n(j)=\sum_{m=1}^{n}a_m^j$: $$\begin{split}
\theta_{n;k}(t)&=\frac1{k!} B_k(0!A_n(1)\big( t^1-(t-1)^1\big),\dots,(k-1)!A_n(k)\big( t^k-(t-1)^k\big))\\
&=\sum_{m_1+2m_2+\dots=k}\frac{1}{m_1!m_2!\dots}\Big(\frac{A_n(1)( t^1-(t-1)^1)}{1}\Big)^{m_1} \Big(\frac{A_n(2)( t^2-(t-1)^2)}{2}\Big)^{m_2}\dots.
\end{split}$$ Let $\alpha_n(j;t)=A_n(j)( t^j-(t-1)^j)$: the values $\theta_{n;k}(t)$ are given by $$\theta_{n;k}(t)=\frac{1}{k!}\left|
\begin{matrix}
\alpha_n(1;t) & -1 & 0 & \dots &0\\
\alpha_n(2;t) & \alpha_n(1;t) & -2 & \dots &0\\
\hdots & \hdots & \hdots & \vdots &\hdots\\
\alpha_n(k-1;t) & \alpha_n(k-2;t)& \alpha_n(k-3;t) & \dots &-(k-1)\\
\alpha_n(k;t) & \alpha_n(k-1;t) & \alpha_n(k-2;t) & \dots &\alpha_n(1;t)\\
\end{matrix}
\right|.$$
For $t=\frac12$ we solely sum over the odd indices $m_1,m_3,\dots$, since only $m_2=m_4=\dots=0$ lead to a positive contribution. We obtain $$\begin{split}
\theta_{n;k}(\frac12)&=\frac1{k!} B_k(0!A_n(1) 2\cdot \frac1{2^1},0,2!A_n(1) 2\cdot \frac1{2^3}\dots,(k-1)!A_n(k)\big( \frac1{2^k}-(-\frac12)^k\big))\\
&=\sum_{m_1+3m_3+\dots=k}\frac{2^{m_1+m_3+\dots}}{2^{1\cdot m_1+3\cdot m_3+\dots}m_1!m_3!\dots}\Big(\frac{A_n(1)}{1}\Big)^{m_1} \Big(\frac{A_n(3)}{3}\Big)^{m_3}\dots.
\end{split}$$
A direct byproduct of this result is an expression for $\ztt_n(\{m^k\})$. We state the formula for $\ztt_n(\{1^k\})$, generalizing the already known results for $t=0$, $\zt_n(\{1\}_k)$ and $t=1$, $\zts_n(\{1\}_k)$; see for example [@KP] and the references therein.
\[CorollZeta\] Given the weight sequence $\boldsymbol{a}=(\frac1{j})_{j\in{\ensuremath{\mathbb{N}}}}$ we obtain for $\theta_{n;k}(t)=\ztt_n(\{1\}_k)$ the results $$\ztt_n(\{1\}_k)=\sum_{j=0}^{k}t^j\zts_n(\{1\}_j)\cdot (1-t)^{k-j}\zt_n(\{1\}_{k-j}).$$ as well as $$\begin{split}
\ztt_n(\{1\}_k)&=\frac1{k!} B_k(0!H_n^{(1)}\big( t^1-(t-1)^1\big),\dots,(k-1)!H_n^{(k)}\big( t^k-(t-1)^k\big))\\
&=\sum_{m_1+2m_2+\dots=k}\frac{1}{m_1!m_2!\dots}\Big(\frac{H_n^{(1)}( t^1-(t-1)^1)}{1}\Big)^{m_1} \Big(\frac{H_n^{(2)}( t^2-(t-1)^2)}{2}\Big)^{m_2}\dots.
\end{split}$$ For $t=\frac12$ we have a truncated analog of the formula of Hoffman and Ihara [@HI eqn. $(41)$]: $$\begin{aligned}
\zt^{\frac12}_n(\{1\}_k)
&=\frac{1}{2^k k!}(B_{k}(0!\cdot 2{\ensuremath{H_{n}^{(1)}}},0,2!\cdot 2{\ensuremath{H_{n}^{(3)}}},0,\dots)\\
&=\sum_{m_1+3m_3+5m_5\dots=k}\frac{2^{m_1+m_3+m_5+\dots}}{2^{k} m_1!m_3!\dots}\Big(\frac{{\ensuremath{H_{n}^{(1)}}}}{1}\Big)^{m_1} \Big(\frac{{\ensuremath{H_{n}^{(3)}}}}{3}\Big)^{m_3}\dots.\end{aligned}$$
First, we turn to the expression for $\theta_{n,k}(t)$. We have $$\begin{aligned}
\Theta_{n}(z,t)&
=\prod_{m=1}^n\Big(1+\frac{a_m z}{1-a_m z t}\Big)
=\prod_{m=1}^n\frac{1+a_m z(1-t)}{1-a_m z t}\\
&=\bigg(\sum_{j\ge 0 }(zt)^j\theta_{n,j}(1)\bigg)\cdot
\bigg(\sum_{j\ge 0 }z^j(1-t)^j\theta_{n,j}(0)\bigg).\end{aligned}$$ Thus, extraction of coefficients gives the stated result: $$\begin{aligned}
\theta_{n,k}(t)&
=[z^k]\bigg(\sum_{j\ge 0 }(z t)^j\theta_{n,j}(1)\bigg)\cdot
\bigg(\sum_{j\ge 0 }z^j(1-t)^j\theta_{n,j}(0)\bigg)\\
&=\sum_{j=0}^{k}t^j\theta_{n,j}(1)\cdot (1-t)^{k-j}\theta_{n,k-j}(0).\end{aligned}$$
Next, we use the $\exp-\log$ representation to obtain a different expression for $\Theta_{n}(z,t)$. $$\begin{split}
\Theta_{n}(z,t)&=\prod_{m=1}^n\Big(1+\frac{a_m z}{1-a_m z t}\Big)= \prod_{m=1}^n\exp\Big(\log\big(1+\frac{a_m z}{1-a_m z t}\big)\Big)\\
&=\exp\Big(\sum_{m=1}^n\log\big(1+\frac{a_m z}{1-a_m z t}\big)\Big)=\exp\Big(\sum_{m=1}^n\log\big(\frac{1-a_m z (t-1)}{1-a_m z t}\big)\Big)\\
&=\exp\Big(\sum_{m=1}^n\log\big(1-a_m z (t-1)\big) - \sum_{m=1}^n\log\big(1-a_m z t\big)\Big).
\end{split}$$ Expansion of the two logarithm functions gives $$\begin{split}
\Theta_{n}(z,t)&=\exp\Big(\sum_{m=1}^n\sum_{j=1}^{\infty}\frac{a_m^j z^j}{j}\big( t^j-(t-1)^j\big)\Big)\\
&=\exp\Big(\sum_{j=1}^{\infty}\frac{z^j}{j}A_n(j)\big( t^j-(t-1)^j\big)\Big).
\end{split}$$ Thus, it follows that $\theta_{n;k}(t)$ can be expressed in terms of the complete Bell polynomials $B_n(x_1,\dots,x_n)$, which are defined via $$\exp\Big(\sum_{\ell\ge 1}\frac{z^{\ell}}{\ell!}x_\ell\Big)
= \sum_{j\ge 0}\frac{B_j(x_1,\dots,x_j)}{j!}z^j,$$ evaluated at $x_\ell=(\ell-1)!A_n(\ell)\big( t^\ell-(t-1)^\ell\big)$.
Following Hoffman [@H2017], the determinant can be obtained using the theory of symmetric functions. Let $x_1,x_2,\dots$ denote variables of degree one. As in Remark let $h_k$ denote the complete symmetric functions and $p_k$ the power sums. There exists polynomials $Q_k$ such that $$h_k=Q_k(p_1,\dots,p_k).$$ Let $H(z)$ and $P(z)$ denote the generating functions of $h_k$ and $p_k$: $$H(z)=\sum_{k=0}^{\infty}h_k z^k=\prod_{i\ge 1}\frac{1}{1-z x_i},
\quad
P(z)=\sum_{k=1}^{\infty}p_k z^k=\sum_{i\ge 1}\frac{x_i}{1-z x_i}.$$ Then, $$H(z)=\exp\left(\int_{0}^{z}P(v)dv \right)=\exp\left(-\sum_{i\ge 1}\ln(1-z x_i)\right)
=\exp\left(\sum_{j\ge 1}z^j \cdot \frac{p_j}j\right).$$ This gives the Bell polynomial type expression for the $Q_k$: $$Q_k(p_1,\dots,p_k)=\sum_{m_1+2m_2+\dots =k}\frac{1}{m_1!m_2!\dots}
\Big(\frac{p_1}{1}\Big)^{m_1}\Big(\frac{p_2}{2}\Big)^{m_2}\dots .$$ On the other hand, MacDonald [@MacDo] gives the determinant $$Q_k(y_1,\dots,y_k)=\frac{1}{k!}\left|
\begin{matrix}
y_1 & -1 & 0 & \dots &0\\
y_2 & y_1 & -2 & \dots &0\\
\hdots & \hdots & \hdots & \vdots &\hdots\\
y_{k-1} & y_{k-2} & y_{k-3} & \dots &-(k-1)\\
y_{k} & y_{k-1} & y_{k-2} & \dots &y_1\\
\end{matrix}
\right|,$$ which proves the stated result.
Further generalizations
=======================
Refinements of the parameter sigma
----------------------------------
Let $\vec{\ell}\in {\ensuremath{\mathcal{M}}}_{n,k}$. Then $\sigma^{(i)}(\vec{\ell})=\sigma^{(i)}_{n,k}(\vec{\ell})$ is defined as the number equal signs in $\vec{\ell}$ stemming from numbers $i\in{\ensuremath{\mathbb{N}}}$. Then, as functions acting from ${\ensuremath{\mathcal{M}}}_{n,k}$ to ${\ensuremath{\mathbb{N}}}_0$: $$\sigma=\sum_{i=1}^{n}\sigma^{(i)}.$$
Given $\vec{\ell}=(1,1,2,2,2,4,6)\in{\ensuremath{\mathcal{M}}}_{n,7}$ then $\sigma^{(1)}=1$, $\sigma^{(2)}=2$ and $\sigma^{(i)}=0$ $i>2$.
In the following we use the vector notation $\vec{t}^{\vec{\sigma}(\vec{\ell})}=t_1^{\sigma_1(\vec{\ell})}\dots t_n^{\sigma_n(\vec{\ell})}$.
Given the weight sequence $\boldsymbol{a}=(a_{j})_{j\in {\ensuremath{\mathbb{N}}}}$. Let $\theta_{n;k}(\vec{t})=\theta_{n;k}(\vec{t};\boldsymbol{a})$ be defined as $$\theta_{n;k}(\vec{t})=\sum_{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}}w(\vec{\ell})\vec{t}^{\vec{\sigma}(\vec{\ell})}
:=\sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}a_{\ell_1}a_{\ell_2}\dots a_{\ell_k}t_1^{\sigma^{(1)}(\vec{\ell})}\dots t_n^{\sigma^{(n)}(\vec{\ell})}.$$
The generating function $\Theta_{n}(z,\vec{t})=\sum_{k\ge 0}\theta_{n;k}(\vec{t})z^k=\sum_{\ell\in{\ensuremath{\mathcal{M}}}_n}w(\vec{\ell})\vec{t}^{\vec{\sigma}(\vec{t})}z^{|\vec{\ell}|}$ can readily be obtained using the symbolic methods.
The generating function $\Theta_{n}(z,\vec{t})$ is given by $$\begin{split}
\Theta_{n}(z,\vec{t})&=\prod_{m=1}^n\Big(1+\frac{a_m z}{1-a_m z t_m}\Big)\\
&=\exp\Big(\sum_{j=1}^{\infty}\frac{z^j}{j}\sum_{m=1}^{n}a_m^j\big( t_m^j-(t_m-1)^j \big)\Big).
\end{split}$$
For even indices $m$ let $t_m=t_E$, whereas for odd indices let $t_m=t_O$. Then, $$\theta_{n;k}(t_E,t_O)=\sum_{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}}w(\vec{\ell})t^{\sigma^{E}(\vec{\ell})}t^{\sigma^{O}(\vec{\ell})}$$ The generating function $$\Theta_{n}(z,t_E,t_O)=\exp\Big(\sum_{j=1}^{\infty}\frac{z^j}{j}\Big[\big(t_O^j-(t_O-1)^j \big)\sum_{\substack{m=1\\ m\text { odd}}}^{n}a_m^j+\big(t_E^j-(t_E-1)^j \big)\sum_{\substack{m=1\\ m\text{ even}}}^{n}a_m^j\Big).$$ We consider for example $a_m=\frac{1}m$, and assume the $n=2N$. Then, $$\sum_{\substack{m=1\\ m\text {even}}}^{n}a_m^j=\frac1{2^j}\sum_{m=1}^{N}\frac{1}{m^j}=\frac1{2^j}H_N^{(j)}$$ and $$\sum_{\substack{m=1\\ m\text{ odd}}}^{n}a_m^j=\sum_{m=1}^{N}\frac{1}{(2m-1)^j}=H_{2N}^{(j)}-\frac1{2^j}H_N^{(j)}.$$ Thus, $$\Theta_{2N}(z,t_E,t_O)=\exp\Big(\sum_{j=1}^{\infty}\frac{z^j}{j}\Big[\big(t_O^j-(t_O-1)^j \big)(H_{2N}^{(j)}-\frac1{2^j}H_N^{(j)})
+\big(t_E^j-(t_E-1)^j \big)\frac1{2^j}H_N^{(j)}\Big).$$
Infinite multisets
------------------
In another direction the generating function identities can be extended to infinity multisets. Let $${\ensuremath{\mathcal{M}}}_{\infty,k}=\{\vec{\ell}=(\ell_1,\dots,\ell_k)\in{\ensuremath{\mathbb{N}}}^k\colon \infty>\ell_1 \ge \ell_2\ge \dots \ge \ell_1\ge 1\},$$ and ${\ensuremath{\mathcal{M}}}_{\infty}=\bigcup_{k=1}^{\infty}{\ensuremath{\mathcal{M}}}_{\infty,k}$. The formal expressions in the definition of $\theta_{\infty;k}(t)=\sum_{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{\infty,k}}a_{\ell_1}\dots a_{\ell_k}t^{\sigma(\vec{\ell})}$ are well defined if the weight sequence $\boldsymbol{a}$ does not grow too fast.
The generating function $\Theta_{\infty}(z,t)=\sum_{k\ge 0}\theta_{\infty;k}(t)z^k$ satisfies $$\begin{split}
\Theta_{\infty}(z,t)=\prod_{m=1}^\infty\Big(1+\frac{a_m z}{1-a_m z t}\Big)=\exp\Big(\sum_{j=1}^{\infty}\frac{z^j}{j}A_\infty(j)\big( t^j-(t-1)^j\big)\Big),
\end{split}$$ where $A_\infty(j)$ denotes the power sums $A_\infty(j)=\sum_{m=1}^{\infty}a_m^j$.
Special choices of $\boldsymbol{a}$, e.g., $\boldsymbol{a}=(\frac1{m^s})$, then directly leads to certain previously considered non-truncated multiple zeta values, see [@VW; @HI; @H2018] for general results.
Let $\boldsymbol{a}=(\frac1{j^m})$ with $m\ge 2$, then $$\ztt(\{m\}_k)=\sum_{j=0}^{k}t^j\zts(\{m\}_j)\cdot (1-t)^{k-j}\zt(\{m\}_{k-j}).$$ Alternatively, there is an expression in terms of Bell polynomials and ordinary zeta values: $$\ztt(\{m\}_k)=\frac1{k!} B_k(0!\zt(m)\big( t^1-(t-1)^1\big),\dots,(k-1)!\zt(m\cdot k)\big( t^k-(t-1)^k\big)).$$
Taking into account the refinement of $\sigma$, we may consider a refined generating function $\Theta_{\infty}(z,\vec{t})=\sum_{k\ge 0}\theta_{\infty;k}(\vec{t})z^k$; we obtain: $$\begin{split}
\Theta_{\infty}(z,\vec{t})=\prod_{m=1}^\infty\Big(1+\frac{a_m z}{1-a_m z t_m}\Big)
=\exp\Big(\sum_{j=1}^{\infty}\frac{z^j}{j}\sum_{m=1}^{\infty}a_m^j\big( t_m^j-(t_m-1)^j \big)\Big).
\end{split}$$ This leads to multi-interpolated versions of identities for $\ztt(\{m\}_k)$.
Sums of multisets: partitions
-----------------------------
Yet another direction would be to introduce a new parameter $p=p_{n,k}$ on ${\ensuremath{\mathcal{M}}}_{n,k}$, with $p$ the one-norm of $\vec{\ell}$. It measures the numbers $p(\vec{\ell})$ the partitions $\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}$ induce: $$p(\vec{\ell})=\Vert \vec{\ell}\Vert_1=\sum_{i=1}^{k}\ell_k.$$ Then, $$\theta_{n;k}(t,q)=\sum_{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}}w(\vec{\ell})q^{p(\vec{\ell})}t^{\sigma(\vec{\ell})}
=\sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}a_{\ell_1}a_{\ell_2}\dots a_{\ell_k}q^{\sum_{i=1}^{k}\ell_k} t^{\sigma(\vec{\ell})}.$$ Note that this can alternatively be achieved by a change of weights $a_m\mapsto a_m\cdot q^m$. Then, the generating functions $\Theta_{n}(z,t,q)=\sum_{k\ge 0}\theta_{n;k}(t,q)$ and $\Theta_{\infty}(z,t,q)=\sum_{k\ge 0}\theta_{\infty;k}(t,q)$ satisfy $$\Theta_{n}(z,t,q)=\prod_{m=1}^n\Big(1+\frac{a_m q^m z}{1-a_m q^m z t}\Big),\quad
\Theta_{\infty}(z,t,q)=\prod_{m=1}^\infty\Big(1+\frac{a_m q^m z}{1-a_m q^m z t}\Big).$$ It is then possible to study the distribution of $p$, or the joint distribution of $p$ and $\sigma$ on ${\ensuremath{\mathcal{M}}}_{n,k}$.
Setting $t=1$ we the generating function $\Theta_{n}(z,1,q)=\sum_{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n}}w(\vec{\ell})q^{p(\vec{\ell})}z^{|\vec{\ell}|}$ simplifies to $$\prod_{m=1}^n\Big(\frac{1}{1-a_m q^m z}\Big).$$
Moreover, setting $a_m=z=t=1$ in $\Theta_{\infty}(z,t,q)$ leads us directly to the generating function of the partition function: $$\Theta_{\infty}(1,1,q)=\sum_{\vec{\ell}\in{\ensuremath{\mathcal{M}}}_\infty}q^{p(\vec{\ell})}=\prod_{m=1}^{\infty}\frac{1}{1-q^m}.$$
Distributions and limit laws
============================
Before, we treated the variable $t$ as a parameter, mainly between zero and one, interpolating between weighted $k$-subsets and $k$-multisets or (truncated) multiple zeta values and their star counterparts. Here, we are interested in the distribution of $\sigma$ and introduce a random variable $S_{n,k}$. In the following we denote with $[z^j]$ the extraction of coefficients operator, $[z^j]f(x)=\sum_{k=0}^{\infty}f_k \cdot z^k=f_j$, $j\in{\ensuremath{\mathbb{N}}}_0$.
Given the weight sequence $\boldsymbol{a}=(a_{j})_{j\in {\ensuremath{\mathbb{N}}}}$. The random variable $S_{n,k}$ counts the number of equal signs in an element $\vec{\ell}\in{\ensuremath{\mathcal{M}}}_{n,k}$. Its probability mass function is defined in terms of $\theta_{n;k}(t)$: $${\ensuremath{\mathbb{P}}}\{S_{n,k}=j\}:=\frac{[t^j]\theta_{n;k}(t)}{\theta_{n;k}(1)},\quad 0\le j\le k-1;$$ equivalently, the probability generating function ${\ensuremath{\mathbb{E}}}(t^{S_{n,k}})$ is given by $${\ensuremath{\mathbb{E}}}(t^{S_{n,k}})=\frac{\theta_{n;k}(t)}{\theta_{n;k}(1)}.$$
Note the the actual support of $S_{n,k}$ is the range $\max\{0,k-n\}\le j \le k-1$, since for $k \ge n$ at most $n$ sums out of a total of $k$ can have unequal summands.
The expected value $${\ensuremath{\mathbb{E}}}(S_{n,k})=\frac{\sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}\sigma(\vec{\ell})w(\vec{\ell})}{\sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}w(\vec{\ell})}$$ can be obtained by extracting of coefficients: $$\label{E}
{\ensuremath{\mathbb{E}}}(S_{n,k})=\frac{\theta_{n;k}'(1)}{\theta_{n;k}(1)}=\frac{[z^k]E_t\frac{\partial}{\partial t}\Theta_{n}(z,t)}{\theta_{n;k}(1)}.$$ Here $E_t$ denotes the evaluation operator at $t=1$. The variance is determined via the second factorial moment ${\ensuremath{\mathbb{E}}}({\ensuremath{S_{n,k}^{\underline{2}}}})$, $$\label{V}
{\ensuremath{\mathbb{V}}}(S_{n,k})={\ensuremath{\mathbb{E}}}({\ensuremath{S_{n,k}^{\underline{2}}}})+{\ensuremath{\mathbb{E}}}(S_{n,k})-{\ensuremath{\mathbb{E}}}(S_{n,k})^2,$$ with $${\ensuremath{\mathbb{E}}}({\ensuremath{S_{n,k}^{\underline{2}}}})=\frac{\theta_{n;k}''(1)}{\theta_{n;k}(1)}=\frac{[z^k]E_t\frac{\partial^2}{\partial t^2}\Theta_{n}(z,t)}{\theta_{n;k}(1)}.$$
Interpolated k-element multisets
--------------------------------
We choose the weight sequence $\boldsymbol{a}=(1)_{j\in{\ensuremath{\mathbb{N}}}}$ and obtain the distribution of $S_{n,k}$
The probability mass function of the random variable $S_{n,k}$ in random $k$ multisets of $[n]$, is given by $${\ensuremath{\mathbb{P}}}\{S_{n,k}=j\}=\frac{\binom{n}{k-j}\binom{k-1}{j}}{\binom{n+k-1}k},\quad 0\le j\le k-1.$$ Thus, $S_{n,k}{\ensuremath{\stackrel{(d)}=}}\operatorname{\text{Hypergeo}}(n+k-1,k-1,k)$ follows a hypergeometric distribution. Its expected value and variance are given by $${\ensuremath{\mathbb{E}}}(S_{n,k})=\frac{k(k-1)}{n+k-1},\quad{\ensuremath{\mathbb{V}}}(S_{n,k})=\frac{k(k-1)n(n-1)}{(n+k-1)^2(n+k-2)}.$$ Moreover, the factorial moments are given by $${\ensuremath{\mathbb{E}}}({\ensuremath{S_{n,k}^{\underline{s}}}})=\frac{{\ensuremath{k^{\underline{s}}}}\cdot{\ensuremath{(k-1)^{\underline{s}}}}}{{\ensuremath{(n+k-1)^{\underline{s}}}}}, \quad s\ge 1.$$
The probability mass function is readily obtain by extraction of coefficients. First, we note that $$\Theta_{n}(z,t)=\prod_{m=1}^n\Big(1+\frac{z}{1- z t}\Big)=\Big(\frac{1-z t+z}{1-z t}\Big)^n.$$ Then, $$\begin{split}
{\ensuremath{\mathbb{P}}}\{S_{n,k}=j\}&=\frac{[t^j]\theta_{n,k}(t)}{\theta_{n,k}(1)}=\frac{[z^k t^j]\Theta_{n,k}(z,t)}{\binom{n+k-1}{k}}
=\frac1{\binom{n+k-1}{k}}[z^{k-j} (zt)^j]\Big(\frac{1-zt+z}{1-z t}\Big)^n\\
&=\frac{\binom{n}{k-j}}{\binom{n+k-1}{k}}[(zt)^j]\frac1{(1-zt)^{k-j}}=\frac{\binom{n}{k-j}\cdot \binom{k-1}{j}}{\binom{n+k-1}{k}}.
\end{split}$$ Alternatively, observe elementary that $j$ equal signs are distributed amongst $k-1$ places, leading to $\binom{k-1}{j}$, and that the remaining $k-j$ sums give a factor $$\sum_{n \ge \ell_1> \cdots>\ell_{k-j}\ge 1} 1 = |{\ensuremath{\mathcal{S}}}_{n,k-j}|=\binom{n}{k-j}.$$ The expected value, the variance and the factorial moments of the hypergeometric distribution are well known, or directly obtained using the Vandermonde identity.
The limit laws for $\max\{n,k\}\to\infty$ are given by three different distributions:
1. Degenerate case: for $n\to\infty$ and $k=o(\sqrt{n})$: $S_{n,k}\to 0$; similarly, for $k\to\infty$ and $n=o(\sqrt{k})$: $S_{n,k}-k+n\to 0$.
2. Poisson range: for $k\to\infty$ and $n\sim c \sqrt{k}$: $S_{n,k}-k+n\to \operatorname{\text{Po}}(c^2)$; similarly, for $k,n\to\infty$ and $k\sim c\cdot \sqrt{n}$ with $c>0$: $S_{n,k}\to \operatorname{\text{Po}}(c^2)$.
3. Normal range: for $k,n\to\infty$ and $\sqrt{n}\ll k\ll n^2$: $$\frac{S_{n,k}-{\ensuremath{\mathbb{E}}}(S_{n,k})}{\sqrt{{\ensuremath{\mathbb{V}}}(S_{n,k})}}\to \mathcal{N}(0,1).$$
First we note that the expected value of the random variable $S_{n,k}$, counting the number of equal signs in random $k$ multisets of $[n]$, has the following behavior: $${\ensuremath{\mathbb{E}}}(S_{n,k})\sim
\begin{cases}
\frac{k^2}n, \quad n\to\infty \text{ and }k=o(\sqrt{n}),\\
c, \quad k,n\to\infty \text{ and }k\sim c\cdot \sqrt{n},\\
k-n+c^2,\quad k\to\infty \text{ and }n\sim c \sqrt{k},\\
k-n+\frac{n^2}{k},\quad k\to\infty \text{ and }n=o(\sqrt{k}),\\
\end{cases}$$ The Poisson limit law for $k\sim c\cdot \sqrt{n}$ follows directly from the asymptotic expansion of the factorial moments: $${\ensuremath{\mathbb{E}}}({\ensuremath{S_{n,k}^{\underline{s}}}})=\frac{{\ensuremath{k^{\underline{s}}}}\cdot{\ensuremath{(k-1)^{\underline{s}}}}}{{\ensuremath{(n+k-1)^{\underline{s}}}}}
\sim (c^2)^s.$$ Thus, by the method of moments the (factorial) moments $S_{n,k}$ converge to the (factorial) moments of a Poisson distributed random variable with parameter $\lambda=c^2$. In the range $k\to\infty$ and $n\sim c \sqrt{k}$ the Poisson limit law is indicated by the asymptotic expansions of ${\ensuremath{\mathbb{E}}}(S_{n,k})$ and ${\ensuremath{\mathbb{V}}}(S_{n,k})$. We obtain elementarily $${\ensuremath{\mathbb{P}}}\{S_{n,k}=k-n+\ell\}=\frac{\binom{n}{n-\ell}\binom{k-1}{k-n+\ell}}{\binom{n+k-1}k}.$$ By our assumption $n\sim c \sqrt{k}$ so $\binom{n}{n-\ell}\sim (nc)^{\ell}/\ell!$. We use the asymptotic expansion of the factorials, $$n!\sim \frac{n^n}{e^n}\sqrt{2\pi n}$$ as well as a precise expansion of the terms $$(n+k-1)^{n+k-1}, \quad (k-n+\ell)^{k-n+\ell},$$ using the $\exp-\log$ representation. This gives the desired result $${\ensuremath{\mathbb{P}}}\{S_{n,k}=k-n+\ell\}\sim \frac{n^{2\ell} (c^2)^{\ell}}{\ell! k^{\ell}}e^{-c^2},\quad \ell\ge 0.$$ The normal limit laws for the hypergeometric distribution are classical and well known, see Nicholson [@N] or Feller [@F] and the discussion in the introduction; the proofs omitted.
Interpolated truncated multiple zeta values
-------------------------------------------
We consider the weight sequence $\boldsymbol{a}=(1/j)_{j\in{\ensuremath{\mathbb{N}}}}$ and the distribution of $S_{n,k}$, defined by the probability generating function $${\ensuremath{\mathbb{E}}}(t^{S_{n,k}})=\frac{\theta_{n,k}(t)}{\theta_{n,k}(1)}
=\frac{\ztt_n(\{1\}_k)}{\zts_n(\{1\}_k)}.$$ For example, the boundary values are given by $${\ensuremath{\mathbb{P}}}\{S_{n,k}=0\}=\frac{\zt_n(\{1\}_k)}{\zts_n(\{1\}_k)},\quad {\ensuremath{\mathbb{P}}}\{S_{n,k}=k-1\}=\frac{H_n^{(k)}}{\zts_n(\{1\}_k)}.$$ We derive the expected value $${\ensuremath{\mathbb{E}}}(S_{n,k})=\frac{1}{\zts_n(\{1\}_k)} \sum_{n \ge \ell_1\ge \cdots\ge \ell_k\ge 1}\frac{\sigma(\vec{\ell})}{\ell_1^{m}\cdots \ell_k^{m}}$$ using .
The expected value is given by the following exact expression: $${\ensuremath{\mathbb{E}}}(S_{n,k})=\frac{1}{\zts_n(\{1\}_k)}\sum_{\ell=0}^{k-2}H_n^{(\ell+2)}\zts_n(\{1\}_{k-\ell-2}).$$
Note that the variance can be obtained using the second factorial moment , but it is more involved. The limit laws for $S_{n,k}$ are technically much more involved compared to the case $\boldsymbol{a}=(1)_{j\in{\ensuremath{\mathbb{N}}}}$. The cases of either $n$ fixed and $k\to\infty$ or $n\to\infty$ and $k$ fixed are not too difficult to analyze. In contrast, for both $n,k\to\infty$ the asymptotic expansions of the denominator $\zts_n(\{1\}_k)$ require a more precise analytic combinatorial analysis; see the works of Hwang [@Hwang1994; @Hwang1998; @Hwang2004] and Bai et al. [@Bai]. The case of $n$ and $k\to\infty$ indicates that a normal limit should appear. Note that the limit law, given by the sum of independent Bernoulli random variables, also appeared in the analysis of algorithms literature; amongst others, in the the unsuccessful search in binary search trees [@DobSmy1996], the degree of the root in increasing trees [@DobSmy1996], Distances in increasing trees [@Dob1996; @DobSmy1996] or edge-weights in increasing trees [@KP2008; @KW].
For $k\to\infty$ and $n$ fixed, $$k-S_{n,k}{\ensuremath{\xrightarrow{(d)}}}D_n=B_1\oplus B_2 \oplus \dots \oplus B_n,$$ where $B_j=\operatorname{\text{Be}}(\frac{1}j)$ denote independent Bernoulli-distributed random variables; hence the probability mass function of $D$ is given by $${\ensuremath{\mathbb{P}}}\{D=\ell\}=\frac{\zt_{n-1}(\{1\}_{\ell-1})}{n},\quad 1\le \ell\le n.$$
For $n\to\infty$ and $k$ fixed the random variable $S_{n,k}$ degenerates: $S_{n,k} {\ensuremath{\xrightarrow{(d)}}}0$.
The growth range of the degenerate case can be extended and made more precise using the results of Hwang [@Hwang1995] for $\zt_n(\{1\}_k$, stated in terms of the unsigned Stirling numbers of the first kind, and his results [@Hwang2004] (see also Bai et al. [@Bai]) for $\zts_n(\{1\}_k$, related to the expected number of maxima in hypercubes.
It is well known that the sum of the limit law of $D_n$, normalized and centralized, for $n\to\infty$ is asymptotically normal: $$\frac{D_n-\log n}{\sqrt{\log n}}{\ensuremath{\xrightarrow{(d)}}}\mathcal{N}(0,1).$$ See for example Dobrow and Smythe for an approach using Poisson approximation [@DobSmy1996] or Hwang [@Hwang1999]. Thus, we expect that the random variable $S_{n,k}$ is asymptotically normal in a certain range of $n,k$ both tending to infinity. Note further, that our result above may shed light on the nature of the asymptotic expansions of $\zts_n(\{1\}_k)$ derived earlier by Hwang [@Hwang1994; @Hwang1998; @Hwang2004]: informally speaking, we observe that $k-S_{n,k}~\sim D_n$, or $S_{n,k}\sim k-D_n$, where $D_n$ is concentrated around $\log n$. Thus $k-\log n$ should govern the behavior of $S_{n,k}$ and also its total mass $\zts_n(\{1\}_k)$.
Our starting point is the generating function $$\Theta_n(z,t)=\prod_{m=1}^n\Big(1+\frac{\frac{z}m}{1-\frac{z t}m}\Big).$$ Differentiating with respect to $t$ gives $$\begin{split}
\frac{\partial}{\partial t}\Theta_n(z,t)
&=\Theta_n(z,t)\cdot\sum_{j=1}^{n}\frac{z^2}{j^2(1-\frac{zt}{j})^2}\cdot
\frac{1}{1+\frac{\frac{z}j}{1-\frac{z t}j}}\\
&=\Theta_n(z,t)\cdot \sum_{j=1}^{n}\frac{z^2}{j^2(1-\frac{zt}{j})(1-\frac{zt}{j}+\frac{z}{j})}.
\end{split}$$ Evaluation at $t=1$ gives $$E_t\frac{\partial}{\partial t}\Theta_n(z,t)=
\Theta_n(z,1)\cdot \sum_{j=1}^{n}\frac{z^2}{j^2(1-\frac{z}{j})}.$$ Extraction of coefficients leads to the stated result: $${\ensuremath{\mathbb{E}}}(S_{n,k})=\frac{[z^k]E_t\frac{\partial}{\partial t}\Theta_n(z,t)}{[z^k]\Theta_n(z,1)}
=\frac{\sum_{j=1}^{n}[z^{k-2}]\frac{\Theta_n(z,1)}{1-\frac{z}j}}{\zts_n(\{1\}_k)},$$ noting that $[z^\ell]\Theta_n(z,1)=\zts_n(\{1\}_\ell)$.
For $n\to\infty$ and $k$ fixed we study $${\ensuremath{\mathbb{P}}}\{S_{n,k}=0\}=\frac{\zt_n(\{1\}_k}{\zts_n(\{1\}_k}.$$ We use the representation given in Corollary \[CorollZeta\] for $t=0$ and $t=1$. We require the asymptotic expansion of the harmonic numbers $H_n$ for $n\to\infty$: $$H_{n}=\log n +\gamma
+\frac{1}{2n}-\frac{1}{12n^{2}}+\mathcal{O}\Bigl(\frac{1}{n^{4}}\Bigr).$$ Asymptotically, the summand $m_1=k$ dominates and gives for both $t=0$ and $t=1$ the same value $\frac{1}{k!}\log^k_n$. Consequently, ${\ensuremath{\mathbb{P}}}\{S_{n,k}=0\}\sim \frac{\log^k n}{\log^k n}=1$ for large $n$.
For $k\to\infty$ and $n$ fixed we use singularity analysis [@FS] and derive the asymptotic equivalent of ${\ensuremath{\mathbb{P}}}\{S_{n,k}=k-j\}$, $1\le j\le n$. We have $${\ensuremath{\mathbb{P}}}\{S_{n,k}=k-j\}=\frac{1}{\zts_n(\{1\}_k)}[z^{k}t^{k-j}]\Theta_n(z,t)=\frac{1}{\zts_n(\{1\}_k)}[z^{k}t^{k-j}]\prod_{m=1}^{n}\Big(1+\frac{\frac{z}m}{1-\frac{z t}m}\Big).$$ Since $z^{k}t^{k-j}=u^{k-j}z^j$ for $u=z t$ we get $${\ensuremath{\mathbb{P}}}\{S_{n,k}=k-j\}=\frac{1}{\zts_n(\{1\}_k)}[z^{j}u^{k-j}]\prod_{m=1}^{n}\Big(1+\frac{\frac{z}m}{1-\frac{u}m}\Big).$$ The product has a dominant singularity is at $u=1$ and can be written as $$\prod_{m=1}^{n}\Big(1+\frac{\frac{z}m}{1-\frac{u}m}\Big)=\frac1{1-u}\cdot R_n(u,z),
$$ with $R_n(u,z)$ analytic inside a circle of radius $2$. Consequently, for $k\to\infty$ $$[u^{k-j}]\prod_{m=1}^{n}\Big(1+\frac{\frac{z}m}{1-\frac{u}m}\Big)\sim R_n(1,z)= n\cdot \prod_{m=1}^{n}\big(1+\frac{z-1}{m}\big).$$ It remains to derive the asymptotic expansion of $\zts_n(\{1\}_k)$. We can use again singularity analysis as before; alternatively, the well known [@FS; @Bai; @KP] binomial sum representation directly gives us the desired result: $$\zts_n(\{1\}_k)=\sum_{\ell=1}^{n}\binom{n}{\ell}\frac{(-1)^{\ell+1}}{\ell^k}\sim n.$$ Finally, combining our results gives $${\ensuremath{\mathbb{P}}}\{S_{n,k}=k-j\}\sim[z^{j}]\prod_{m=1}^{n}\big(1+\frac{t-1}{m}\big),\quad 1\le j\le n.$$ The product is exactly the probability generating function of the independent Bernoulli random variables with success probability $\frac1m$. Extraction of coefficients directly lead to the stated result, using the representation of $\zt_n(\{1\}_k)$ given in[@KP].
Interpolated truncated multiple zeta values - only twos
-------------------------------------------------------
We consider the weight sequence $\boldsymbol{a}=(1/j^2)_{j\in{\ensuremath{\mathbb{N}}}}$ and the distribution of $S_{n,k}$, defined by the probability generating function $${\ensuremath{\mathbb{E}}}(t^{S_{n,k}})=\frac{\theta_{n,k}(t)}{\theta_{n,k}(1)}=\frac{\ztt_n(\{2\}_k)}{\zts_n(\{2\}_k)}.$$ For example, the boundary are given by $${\ensuremath{\mathbb{P}}}\{S_{n,k}=0\}=\frac{\zt_n(\{2\}_k)}{\zts_n(\{2\}_k)},\quad {\ensuremath{\mathbb{P}}}\{S_{n,k}=k-1\}=\frac{H_n^{(2k)}}{\zts_n(\{2\}_k)}.$$ We obtain the following limit laws.
The limit laws for $S_{n,k}$, $\max\{n,k\}\to\infty$, are given by three different distributions:
1. For $n\to\infty$ and $k$ fixed: $S_{n,k}{\ensuremath{\xrightarrow{(d)}}}S{\infty,k}$, with $${\ensuremath{\mathbb{P}}}\{S_{\infty,k}=j\}= \frac{1}{\zts(\{2\}_k)}\cdot\sum_{\substack{\mathbf{p}\in P_O(k)\\ {\ensuremath{\mathcal{L}}}(\mathbf{p})=k-j} }\zt(2\cdot\mathbf{p}),\quad 0\le j\le k-1.$$
2. For $k\to\infty$ and $n$ fixed: $k-S_{n,k}{\ensuremath{\xrightarrow{(d)}}}D_n$. The random variable $D_n$ is given by the sum of independent Bernoulli random variables: $$D_n=B_1\oplus B_2 \oplus \dots \oplus B_n,$$ where $B_j=\operatorname{\text{Be}}(\frac{1}{j^2})$ denote independent Bernoulli-distributed random variables. Moreover, ${\ensuremath{\mathbb{E}}}({\ensuremath{D_n^{\underline{s}}}})=s!\zt_{n-1}(\{2\}_s)$, $s\ge 1$.
3. For $k,n\to\infty$: $k-S_{n,k}{\ensuremath{\xrightarrow{(d)}}}D$; $$D=B_1\oplus B_2 \oplus \dots = \bigoplus_{m=1}^{\infty}B_m.$$ Here $B_j=\operatorname{\text{Be}}(\frac{1}{j^2})$, $j\ge 1$, denote independent Bernoulli-distributed random variables. Moreover, ${\ensuremath{\mathbb{E}}}({\ensuremath{D^{\underline{s}}}})=s!\zt(\{2\}_s)$, $s\ge 1$.
The random variable $D$ is exactly the limit law of the number of cuts in a recursive tree to isolate a leaf [@KP2008b]; modified weight sequences lead to other families of increasing trees. Moreover, $D_n$ is closely related the the distribution of the number of cuts in a tree of size $n$.
Similar to we have, $$\ztt_{n}(\{2\}_{k}) =\sum_{\circ = \text{``},\text{''} \text{or} \, \text{``}+\text{''}}t^{\sigma}\zt_{n}( \underbrace{2 \circ 2 \circ \dots \circ 2}_{k}) = \sum_{\mathbf{p}\in P_O(k)}t^{k-{\ensuremath{\mathcal{L}}}(\mathbf{p})}\zt_{n}(2\cdot \mathbf{p}).$$ By taking the limit $n\to\infty$, $k$ being fixed, and extraction of coefficients we directly obtain the stated result.
For $k\to\infty$ and arbitrary $n$ we use singularity analysis [@FS] and derive the asymptotic equivalent of ${\ensuremath{\mathbb{P}}}\{S_{n,k}=k-j\}$, $1\le j$. We have $${\ensuremath{\mathbb{P}}}\{S_{n,k}=k-j\}=\frac{1}{\zts_n(\{2\}_k)}[z^{k}t^{k-j}]\Theta_n(z,t)=\frac{1}{\zts_n(\{2\}_k)}[z^{k}t^{k-j}]\prod_{m=1}^{n}\Big(1+\frac{\frac{z}{m^2}}{1-\frac{z t}{m^2}}\Big).$$ Setting as before $u=z t$, we get $${\ensuremath{\mathbb{P}}}\{S_{n,k}=k-j\}=\frac{1}{\zts_n(\{2\}_k)}[z^{j}u^{k-j}]\prod_{m=1}^{n}\Big(1+\frac{\frac{z}{m^2}}{1-\frac{u}{m^2}}\Big).$$laenge The product has a dominant singularity is at $u=1$ and can be written as $$\prod_{m=1}^{n}\Big(1+\frac{\frac{z}{m^2}}{1-\frac{u}{m^2}}\Big)=\frac1{1-u}\cdot R_n(u,z),
$$ with $R_n(u,z)$ analytic inside a circle of radius $2$. Consequently, for $k\to\infty$ $$[u^{k-j}]\prod_{m=1}^{n}\Big(1+\frac{\frac{z}{m^2}}{1-\frac{u}{m^2}}\Big)\sim R_n(1,z) = \frac{2n}{n+1}\cdot \prod_{m=1}^{n}\big(1+\frac{z-1}{m^2}\big).$$ The asymptotic expansion of $\zts_n(\{2\}_k)$. We can use again singularity analysis as before and get $$\zts_n(\{2\}_k)\sim \frac{2n}{n+1}.$$ Finally, combining our results gives $${\ensuremath{\mathbb{P}}}\{S_{n,k}=k-j\}\sim[z^{j}]\prod_{m=1}^{n}\big(1+\frac{t-1}{m^2}\big),\quad 1\le j\le n.$$ The product is exactly the probability generating function of the independent Bernoulli random variables with success probability $\frac1{m^2}$.
Refined decomposition of sigma, random vectors and marginals
------------------------------------------------------------
Similarly to the refinements of the parameter $\sigma$ into $\sigma^{(i)}$, $1\le i\le n$, the random variable $S_{n,k}$ can be decomposed into $S_{n,k}^{(i)}$, $$S_{n,k}=\sum_{i=1}^{n}S_{n,k}^{(i)}.$$ Given the weight sequence $\boldsymbol{a}=(a_{j})_{j\in {\ensuremath{\mathbb{N}}}}$, the joint distribution of the random vector $\mathbf{S}_{n,k}=(S_{n,k}^{(1)},\dots, S_{n,k}^{(n)})$ is determined by $${\ensuremath{\mathbb{P}}}\{\mathbf{S}_{n,k}=\vec{j}\}=[\vec{t}^{\vec{j}}]\frac{\theta_{n;k}(\vec{t})}{\theta_{n;k}(\vec{1})}.$$
Another possibility is to distinguish only between odd and even indices, such that $S_{n,k}=S_{n,k}^{(O)}+S_{n,k}^{(E)}$, and to study the bivariate probability generating function $${\ensuremath{\mathbb{E}}}(t_O^{S_{n,k}^{(O)}}t_E^{S_{n,k}^{(E)}})=\frac{\theta_{n;k}(t_O,t_E)}{\theta_{n;k}(1,1)}.$$
For interpolated multisets, weight sequence $\boldsymbol{a}=(1)_{j\in{\ensuremath{\mathbb{N}}}}$, the random vector $\mathbf{S}_{n,k}$ has the probability generating function $${\ensuremath{\mathbb{E}}}(\mathbf{t}^{\mathbf{S}_{n,k}})
=\frac{1}{\binom{n+k-1}{k}}[z^k]\prod_{m=1}^{n}\Big(1+\frac{z}{1- z t_m}\Big).$$
For truncated zeta values, weight sequence $\boldsymbol{a}=(1/j)_{j\in{\ensuremath{\mathbb{N}}}}$, the random vector $\mathbf{S}_{n,k}$ has the probability generating function $${\ensuremath{\mathbb{E}}}(\mathbf{t}^{\mathbf{S}_{n,k}})
=\frac{1}{\zts(\{1\}_k)}[z^k]\prod_{m=1}^{n}\Big(1+\frac{\frac{z}m}{1- \frac{z t_m}m}\Big).$$
In contrast to the multiset case, the marginals $S_{n,k}^{(i)}$ are not exchangeable anymore. The probability generating functions are given $${\ensuremath{\mathbb{E}}}(t^{S_{n,k}^{(i)}})
=\frac{1}{\zts(\{1\}_k)}[z^k](1-\frac{z}{i})\Big(1+\frac{\frac{z}i}{1- \frac{z t}i}\Big)\cdot\prod_{m=1}^{n}\Big(\frac{1}{1-\frac{z}m}\Big).$$ It is expected, that the marginals are asymptotically independent, at least in some growth range of $k$ and $n$.
The marginals $S_{n,k}^{(i)}$ are identically distributed, but not independent; the sequence $\mathbf{S}_{n,k}$ is exchangeable. The probability mass function is given by $${\ensuremath{\mathbb{P}}}\{S_{n,k}^{(i)}=j\}=
\begin{cases}
\frac{\binom{n-2+k}{k}+\binom{n-3+k}{k}}{\binom{n+k-1}{k}}, \quad j=0,\\
\frac{\binom{n-3+k-j}{k-j-1}}{\binom{n+k-1}{k}},\quad 1\le j \le k-1.
\end{cases}$$ The factorial moments, $s\ge 1$, are given by $${\ensuremath{\mathbb{E}}}({\ensuremath{S_{n,k}^{\underline{s}}}})=s!\cdot\frac{{\ensuremath{k^{\underline{s+1}}}}}{(n+k-1){\ensuremath{(n+s-1)^{\underline{s}}}}}.$$
We obtain the following limit laws:
- For $k\to \infty$ and $n$ fixed the normalized random variable $\frac{S_{n,k}}{k}$ converges to a Beta-distributed random variable $\frac{S_{n,k}}{k}{\ensuremath{\xrightarrow{(d)}}}\operatorname{\text{B}}(1,n-1)$.
- For $k/n\to \infty$ and $k,n\to \infty$ the normalized random variable $\frac{n}{k}S_{n,k}$ converges to a standard exponentially distributed random variable, $\frac{n}{k}S_{n,k}{\ensuremath{\xrightarrow{(d)}}}\operatorname{\text{Exp}}(1)$.
- For $k/n\to c>0$ and $k,n\to \infty$ the random variable $S_{n,k}$ converges to a (modified) geometric distribution $G$: $${\ensuremath{\mathbb{P}}}\{G=j\}=\frac{c^{j+1}}{(1+c)^{j+2}}, \quad j\ge 1\quad, {\ensuremath{\mathbb{P}}}\{G=0\}=\frac{1}{1+c}+\frac{c}{(1+c)^2}.$$
- For $k/n\to 0$ and $n\to \infty$ the random variable $S_{n,k}$ degenerates, $S_{n,k}{\ensuremath{\xrightarrow{(d)}}}0$.
The random variable $S_{n,k}^{(i)}$ is closely related to the number of descendants $D_{n+k,k}$ of node labelled $k$ in a recursive trees of size $n+k$ [@KP2006]. Thus, $S_{n,k}^{(i)}$ can be described in terms of a Polya-urn model and the asymptotics for $k\to \infty$ and $n$ fixed can be refined to almost sure convergence; moreover a classical result is available in terms of a martingale tail sum [@HH80].
Let $1\le i\le n$. The probability generating function of the marginal distribution $S_{n,k}^{(i)}$ is given by $${\ensuremath{\mathbb{E}}}(t^{S_{n,k}^{(i)}})
=\frac{1}{\binom{n+k-1}{k}}[z^k]\frac{1}{(1-z)^{n-1}}\cdot\Big(1+\frac{z}{1- z t}\Big).$$ By the structure of the probability generating function, the random variable are identically distributed and the random vector is exchangable. Thus, we can directly obtain the expected value from ${\ensuremath{\mathbb{E}}}(S_{n,k})$. $${\ensuremath{\mathbb{E}}}(S_{n,k}^{(i)})=\frac{1}n{\ensuremath{\mathbb{E}}}(S_{n,k})=\frac{k(k-1)}{n(n+k-1)}.
$$ Additionally, the probability mass function is obtained readily by extraction of coefficients. Additionally, all factorial moments, $s\ge 1$, can be obtained in a straight forward way: $$\begin{split}
{\ensuremath{\mathbb{E}}}({\ensuremath{S_{n,k}^{\underline{s}}}})
&=[z^k]E_t\frac{\partial^s}{\partial t^s}{\ensuremath{\mathbb{E}}}(t^{S_{n,k}^{(i)}})
=\frac{s!}{\binom{n+k-1}{k}}[z^k]\frac{z^{s+1}}{(1-z)^{n+s}}\\
&=\frac{s!\binom{n+k-2}{k-s-1}}{\binom{n+k-1}{k}}
=s!\cdot\frac{{\ensuremath{k^{\underline{s+1}}}}}{(n+k-1){\ensuremath{(n+s-1)^{\underline{s}}}}}.
\end{split}$$
We use the method of moments and derive asymptotic expansions of the factorial moments. For $n$ fixed and $k\to\infty$ we get $${\ensuremath{\mathbb{E}}}({\ensuremath{S_{n,k}^{\underline{s}}}})=s!\cdot\frac{{\ensuremath{k^{\underline{s+1}}}}}{(n+k-1){\ensuremath{(n+s-1)^{\underline{s}}}}}
\sim \frac{s!}{{\ensuremath{(n+s-1)^{\underline{s}}}}}\cdot k^s.$$ Consequently, the power moments of $S_{n,k}/k$ are asymptotically given by $\frac{s!}{{\ensuremath{(n+s-1)^{\underline{s}}}}}$, which proves the Beta limiting distributions.
For $n\to \infty$ we observe that $${\ensuremath{\mathbb{E}}}({\ensuremath{S_{n,k}^{\underline{s}}}})\sim s!\cdot \lambda_{n,k}^s\cdot \frac{k}{n+k},\quad\text{with } \lambda_{n,k}=\frac{k}{n}.$$ Thus, the factorial moments are almost of mixed Poisson type [@KP2016] with standard exponential mixing distribution; the additional factor $\frac{k}{n+k}$ can be explained by the definition of the parameter $\sigma$, which influences the discrete limit case. This directly leads to the stated limit laws using Lemma 2 of [@KP2016]. Alternatively, the discrete limit for $k/n\to c$ can be directly obtained as follows: $${\ensuremath{\mathbb{P}}}\{S_{n,k}^{(i)}=j\}=\frac{\binom{n-3+k-j}{k-j-1}}{\binom{n+k-1}{k}}
=\frac{{\ensuremath{k^{\underline{j+1}}}}\cdot (n-1)}{{\ensuremath{(n+k-1)^{\underline{j+2}}}}}
\sim \frac{c^{j+1}}{(1+c)^{j+2}},\quad j\ge 1.$$ The remaining case $j=0$ is treated in a similar way. For $k/n\to 0$ we observe that $${\ensuremath{\mathbb{P}}}\{S_{n,k}^{(i)}=0\}=\frac{n-1}{n+k-1}+\frac{k(n-1)}{{\ensuremath{(n+k-1)^{\underline{2}}}}}\sim 1.$$
Summary and Outlook
===================
We introduced a parameter $\sigma$ on weighted $k$-element multisets and studied properties of it using symbolic combinatorics. This allows to prove several relations for truncated interpolated multiple zeta values $\ztt_n(\{m\}_k)$, as well as reproving identities for truncated multiple zeta values $\ztt(\{m\}_k)$. Introducing refined enumeration leads to new refinements of previous identities. Interpreting the parameter $\sigma$ as a random variable $S_{n,k}$ leads to several different limit laws, depending on the considered weight sequences and the growth of $n$ and $k$.
It is of interest to complete the analysis of $S_{n,k}$ in the case $\boldsymbol{a}=(1/j)$; we will report on this elsewhere. Similar to the random variable $S_{n,k}$, defined in terms of the parameter $\sigma_{n,k}$, one may also study the distribution of the parameter $p=p_{n,k}$ or the joint distribution of $p$ and $\sigma$. Moreover, it is certainly of interest to study the distribution of $S_{n,k}$ for other interesting sequences, compare with Vignat and Wakhare [@VW] or Hoffman and Mező [@HM].
[9]{} H. Bachmann, Interpolated Schur multiple zeta values, submitted. Z.-D. Bai, L. Devroye, H. -K. Hwang, T. -H. Tsai, Maxima in hypercubes, *Random Structures and Algorithms* [**27**]{}, 290–309, 2005. K. Dilcher, Some q-series identities related to divisor functions, *Discrete Math.* [**145**]{}, 83–93, 1995. R. Dobrow, On the distribution of distances in recursive trees, *Journal of Applied Probability* 33, 749-757, 1996. R. Dobrow and R. Smythe, Poisson approximations for functionals of random trees, *Random Structures & Algorithms* 9, 79-92, 1996. W. Feller, *Intoduction to Probability Theory and its Applications*, John Wiley and Sons, Third Edition, 1968. P. Flajolet and R. Sedgewick, *Analytic Combinatorics*. Cambridge Univ. Press, Cambridge, UK, 2009. P. Flajolet and R. Sedgewick, Mellin transforms and asymptotics: Finite differences and Rice’s integrals, *Theoretical Computer Science* [**144**]{}, 101–124, 1995. R. L. Graham, D. E. Knuth and O. Patashnik, *Concrete Mathematics*, Addison-Wesley, 1994. P. Hall, C. C. and Heyde, *Martingale limit theory and its application. Probability and Mathematical Statistics*, Academic Press, New York-London, 1980. M. E. Hoffman, Multiple harmonic series, *Pacific J. Math.* [**152**]{}, 275-290, 1992. M. E. Hoffman, Quasi-shuffle products, *Journal of Algebraic Combinatorics* 11, 49–68, 2000. M. E. Hoffman, Harmonic-number summation identities, symmetric functions, and multiple zeta values, *Ramanujan J.* [**42**]{}, 501–526, 2017. M. E. Hoffman, Quasi-shuffle algebras and applications, submitted. M. E. Hoffman and K. Ihara, Quasi-shuffle products revisited, *Journal of Algebra* [**481**]{}, 293–326, 2017. I. Mező and M. E. Hoffman, Zeros of the digamma function and its Barnes G-function analogue, *Integral Transforms and Special Functions* 28, 846–858, 2017. H.-K. Hwang, Théorèmes limites pour les structures combinatoires et les fonctions arithmétiques. Thèse, Ecole polytechnique, 1994. H.-K. Hwang, Asymptotic expansions for the Stirling numbers of the first kind, *Journal of Combinatorial Theory, Series A*, 71:2, 343–351, 1995. H.-K. Hwang, A Poisson $\ast$ geometric convolution law for the number of components in unlabelled combinatorial structures. *Combinatorics, Probability and Computing*, 7:1, 89–110, 1998. H.-K. Hwang, Asymptotics of Poisson approximation to random discrete distributions: an analytic approach, *Advances in Applied Probability*, 31, 448-491, 1999. H.-K. Hwang, Phase changes in random recursive structures and algorithms, *Proceedings of the Workshop on Probability with Applications to Finance and Insurance*, edited by T. L. Lai, H. Yang and S. P. Yung, World Scientific, 82–97, 2004. M. Kuba and A. Panholzer, Descendants in increasing trees, *Electronic Journal of Combinatorics* 13 (1), Paper 8, 2006. M. Kuba and A. Panholzer, On edge-weighted recursive trees and inversions in random permutations. *Discrete Mathematics*, Volume 308, Issue 4, 529-540, 2008. M. Kuba and A. Panholzer, Isolating a leaf in rooted trees via random cuttings, *Annals of Combinatorics* 12, 81–99, 2008. M. Kuba and A. Panholzer, On Moment Sequences and Mixed Poisson Distribution, *Probability Surveys* 13, 89–155, 2016. M. Kuba and A. Panholzer, A Note on Harmonic number identities, Stirling series and multiple zeta values, *International journal of number theory*, to appear. M. Kuba and S. Wagner, Deterministic edge-weights in increasing tree families, *Combinatorics, Probability and Computing*, Volume 19, issue 01, pp. 99-119, 2010. I. G. MacDonald, *Symmetric Functions and Hall Polynomials*, 2nd ed., Clarendon Press, Oxford, 1995. W. L. Nicholson, On the Normal Approximation to the Hypergeometric Distribution, *The Annals of Mathematical Statistics* 27(2), 1956. C. Vignat and T. Wakhare, Multiple zeta values for classical special functions, submitted. S. Yamamoto, Interpolation of multiple zeta and zeta-star values, *J. Algebra* [**385**]{}, 102–114, 2013. D. Zagier, Values of zeta functions and their applications, *First European Congress of Mathematics* (Paris, 1992), vol. II (A. Joseph et al., eds.), (*Progr. Math.*, vol. 120) Birkhäuser, Boston 1994, pp. 497–512. W. Zudilin, Algebraic relations for multiple zeta values, *Russian Math. Surveys* 58:1 1–29, 2003.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the dependence of galaxy clustering on luminosity and spectral type using the 2dF Galaxy Redshift Survey (2dFGRS). Spectral types are assigned using the principal component analysis of Madgwick et al. We divide the sample into two broad spectral classes: galaxies with strong emission lines (‘late-types’), and more quiescent galaxies (‘early-types’). We measure the clustering in real space, free from any distortion of the clustering pattern due to peculiar velocities, for a series of volume-limited samples. The projected correlation functions of both spectral types are well described by a power law for transverse separations in the range 2$<$($\sigma/h^{-1}$Mpc)$<$15, with a marginally steeper slope for early-types than late-types. Both early and late types have approximately the same dependence of clustering strength on luminosity, with the clustering amplitude increasing by a factor of $\sim$2.5 between $L^*$ and 4$L^*$. At all luminosities, however, the correlation function amplitude for the early-types is $\sim$50% higher than that of the late-types. These results support the view that luminosity, and not type, is the dominant factor in determining how the clustering strength of the whole galaxy population varies with luminosity.'
author:
- |
\
$^{1}$Department of Physics, University of Durham, South Road, Durham DH1 3LE, UK\
$^{2}$School of Physics & Astronomy, University of Nottingham, Nottingham NG7 2RD, UK\
$^{3}$Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK\
$^{4}$Department of Physics & Astronomy, Johns Hopkins University, Baltimore, MD 21218-2686, USA\
$^{5}$Anglo-Australian Observatory, P.O. Box 296, Epping, NSW 2121, Australia\
$^{6}$Research School of Astronomy & Astrophysics, The Australian National University, Weston Creek, ACT 2611, Australia\
$^{7}$Astrophysics Research Institute, Liverpool John Moores University, Twelve Quays House, Birkenhead, L14 1LD, UK\
$^{8}$Department of Astrophysics, University of New South Wales, Sydney, NSW 2052, Australia\
$^{9}$Department of Physics, University of Oxford, Keble Road, Oxford OX1 3RH, UK\
$^{10}$School of Physics and Astronomy, University of St Andrews, North Haugh, St Andrews, Fife, KY6 9SS, UK\
$^{11}$Department of Astronomy, California Institute of Technology, Pasadena, CA 91125, USA\
$^{12}$Department of Physics, University of Leeds, Woodhouse Lane, Leeds, LS2 9JT, UK\
$^{13}$Institute for Astronomy, University of Edinburgh, Royal Observatory, Blackford Hill, Edinburgh EH9 3HJ, UK\
title: ' The 2dF Galaxy Redshift Survey: The dependence of galaxy clustering on luminosity and spectral type.'
---
methods: statistical - methods: numerical - large-scale structure of Universe - galaxies: formation
Introduction {#sec:intro}
============
One of the major goals of large redshift surveys like the 2 degree Field Galaxy Redshift Survey (2dFGRS) is to make an accurate measurement of the spatial distribution of galaxies. The unprecedented size of the 2dFGRS makes it possible to quantify how the clustering signal depends on intrinsic galaxy properties, such as luminosity or star formation rate.
The motivation behind such a program is to characterize the galaxy population and to provide constraints upon theoretical models of structure formation. In the current paradigm, galaxies form in dark matter haloes that are built up in a hierarchical way through mergers or by the accretion of smaller objects. The clustering pattern of galaxies is therefore determined by two processes: the spatial distribution of dark matter haloes and the manner in which dark matter haloes are populated by galaxies (Benson 2000b; Peacock & Smith 2000; Seljak 2000; Berlind & Weinberg 2002). The evolution of clumping in the dark matter has been studied extensively using N-body simulations of the growth of density fluctuations via gravitational instability (e.g. Jenkins 1998; 2001). With the development of powerful theoretical tools that can follow the formation and evolution of galaxies in the hierarchical scenario, the issue of how galaxies are apportioned amongst dark matter haloes can be addressed, and detailed predictions of the clustering of galaxies are now possible (Kauffmann, Nusser & Steinmetz 1997; Kauffmann 1999; Benson 2000a,b; Somerville 2001).
The first attempt to quantify the difference between the clustering of early and late-type galaxies was made using a shallow angular survey, the Uppsala catalogue, with morphological types assigned from visual examination of the photographic plates (Davis & Geller 1976). Elliptical galaxies were found to have a higher amplitude angular correlation function than spiral galaxies. In addition, the slope of the correlation function of ellipticals was steeper than that of spiral galaxies at small angular separations. More recently, the comparison of clustering for different types has been extended to three dimensions using redshift surveys. Again, similar conclusions have been reached in these studies, namely that ellipticals have a stronger clustering amplitude than spirals (Lahav & Saslaw 1992; Santiago & Strauss 1992; Iovino 1993; Hermit 1996; Loveday 1995; Guzzo 1997; Willmer 1998).
The subjective process of visual classification can now be superseded by objective, automated algorithms to quantify the shape of a galaxy. One recent example of such a scheme can be found in Zehavi (2002), who measured a “concentration parameter” for $30\,000$ galaxy images from the Sloan Digital Sky Survey, derived from the radii of different isophotes. Again, based upon cuts in the distribution of concentration parameter, early-types are found to be more clustered than late-types.
In this paper, we employ a different method to classify galaxies, based upon a principal component analysis (PCA) of galaxy spectra, which is better suited to the 2dFGRS data (Madgwick 2002). This technique has a number of attractive features. First, the PCA approach is completely objective and reproducible. An equivalent analysis can, for example, be applied readily to spectra produced by theoretical models of galaxy formation or to spectra obtained in an independent redshift survey. Secondly, the PCA can be applied over the full magnitude range of the 2dFGRS, whenever the spectra are of sufficient signal to noise (see Section \[subsec:eta\_class\]). For the 2dFGRS, the image quality is adequate to permit a visual determination of morphological type only for galaxies brighter than $b_{\rm J}\,\simeq\,17$, which comprise a mere $5\%$ of the spectroscopic sample.
Two previous clustering studies have used spectral information to select galaxy samples. Loveday, Tresse & Maddox (1999) grouped galaxies in the Stromlo-APM redshift survey into three classes based upon the equivalent width of either the or OII lines, and found that galaxies with prominent emission lines display weaker clustering than more quiescent galaxies. Tegmark & Bromley (1999) measured the relative bias between different spectral classes in the Las Campanas redshift survey (Shectman 1996), using a classification derived from PCA analysis (Bromley 1998), and also found that early spectral types are more strongly clustered than late spectral types. (See also Blanton 2000 for a revision of Tegmark & Bromley’s analysis, which takes into account the effect of errors in the survey selection function.)
Here, we use the 2dFGRS survey to measure the dependence of galaxy clustering jointly on luminosity and spectral type, adding an extra dimension to the analysis carried out by . Previously, a pioneering study of bivariate galaxy clustering, in terms of luminosity and morphological type, was carried out using the Stromlo-APM redshift survey (Loveday 1995). To place the analysis presented here in context, the samples that we consider cover a larger volume and, despite being volume-limited (see Section \[sec:vol.lim\]), typically contain over an order of magnitude more galaxies than those available to Loveday
We give a brief overview of the 2dFGRS in Section \[sec:data\], along with details of the spectral classification and an explanation of how the samples used in the clustering analysis were constructed. The estimation of the redshift space correlation function and its real space counterpart, the projected correlation function, are outlined in Section \[sec:meth\]. A brief overview of the clustering of 2dFGRS galaxies in redshift space, selected by luminosity and spectral type, is given in Section \[sec:redspace\]; a more detailed analysis of the redshift space clustering can be found in Hawkins (2002). We present the main results of the paper in Section \[sec:res\] and conclude in Section \[sec:conc\].
The data {#sec:data}
========
The 2dFGRS sample
-----------------
Detailed descriptions of the construction of the 2dFGRS and its properties are given by Colless (2001). In summary, galaxies are selected down to a magnitude limit of $b_{\rm J}\approx\,19.45$ from the APM Galaxy Survey (Maddox 1990a,b, 1996, 2002). The sample considered in this paper consists of over $160\,000$ redshifts measured prior to May 2001. We focus our attention on the two large contiguous volumes of the survey, one centred on the Southern Galactic Pole (hereafter SGP) and the other close to the direction of the Northern Galactic Pole (NGP).
Spectral classification of 2dFGRS galaxies {#subsec:eta_class}
------------------------------------------
The spectral properties of 2dFGRS galaxies are characterized using the principal component analysis (PCA) described by Madgwick (2002). This analysis makes use of the spectral information in the rest-frame wavelength range 3700Å to 6650Å, thereby including all the major optical diagnostics between OII and . For galaxies with $z>0.15$, sky absorption bands contaminate the line. Since this can affect the stability of the classification, we restrict our analysis to galaxies with $z<0.15$ following Madgwick et al. (2002).
The 2dFGRS spectra are classified by a single parameter, , which is a linear combination of the first and second principal components. This combination has been chosen specifically to isolate the relative strength of emission and absorption lines present in each galaxy’s spectrum, thereby providing a diagnostic which is robust to the instrumental uncertainties that affect the calibration of the continuum. Physically, this parameter is related to the star formation rate in a galaxy, as is apparent from the tight correlation of with the equivalent width of in emission line galaxies (Bland-Hawthorn 2002). In this paper, we divide the 2dFGRS sample into two broad, distinct classes: galaxies with spectra for which the PCA returns $\eta<-1.4$, and which we refer to, for the sake of brevity, as early-type, and galaxies with $\eta>-1.4$, which we call late-type. The distribution of for 2dFGRS spectra displays a shoulder feature at this value (see Fig. 4 of Madgwick 2002).
The spectral type of a galaxy, as given by the value of , clearly depends upon its physical properties and is therefore a useful and effective way in which to label galaxies. Nevertheless, it is still instructive to see how well, if at all, correlates with the morphogical type assigned in a subjective fashion from a galaxy image. Madgwick (2002) show that there is a reasonable correspondence between and morphological type, using high signal-to-noise spectra and photometry taken from Kennicutt (1992); $\eta\simeq-1.4$ approximately delineates the transition between early and late morphological types in . We revisit the comparison between classifications based on spectral and morphological types in Fig. \[fig:jon.eta\], this time using 2dFGRS spectra and UK Schmidt images. The horizontal axis shows the morphological type assigned to a subset of bright APM galaxies by Loveday (1996). Although there is a substantial amount of scatter in the values of spectra that lie within a given morphological class, it is reassuring to see that the median does correlate with morphological class. Moreover, the median values match up well with the broad division that we employ to separate early and late types. Galaxies denoted “early-type” on the basis of their morphology have a median that is smaller than our fiducial value of $\eta=-1.4$ and vice-versa for late-types. In practice, for the samples analysed in this paper, the correspondence between morphological type and spectral class will be better than suggested by Fig. \[fig:jon.eta\]. This is because the sample used in the comparison in Fig. \[fig:jon.eta\] consists of nearby extended galaxies, and so the distribution of spectral types is distorted somewhat by aperture effects (see e.g. Kochanek, Pahre & Falco 2002; Madgwick et al. 2002). This effect arises because the fibres used to collect the galaxy spectra are of finite size (subtending 2” on the sky). For this reason, when we measure the spectrum of a nearby galaxy it is possible that a disproportionate amount of light will be sampled from the bulge, thereby making the galaxy appear systematically redder or “earlier" in type. We find that this effect is only significant for the most nearby galaxies ($z < 0.05$) and should be completely negligible beyond $z\sim0.1$ (Madgwick et al. 2002).
Sample Selection {#sec:compl}
----------------
In order to construct an optimal sample for the measurement of the two point correlation function, we select regions with high completeness in terms of measured redshifts, using a redshift completeness mask for the 2dFGRS, similar to the one described in Colless (2001; see also Norberg 2002). Such a mask is required because of the tiling strategy adopted to make the best use of the allocated telescope time, along with the fact that the survey is not yet finished. An additional consideration is the success rate with which spectral types have been assigned to galaxies, which depends upon the signal-to-noise ratio of the galaxy spectrum.
In Fig. \[fig:eta.compl\], we show histograms of the spectral classification success rate for two different ranges of field completeness, $c_{\rm F}$, which is defined as the ratio of the number of measured redshifts in a given 2dF field to the number of targets. The spectral classification success rate has two contributions. The first of these is the redshift completeness, shown by the dotted curve. This incompleteness arises because we do not always succeed in measuring a redshift for a targeted object. The redshift incompleteness is necessarily small for the high completeness fields contributing to the histograms. The second contribution is the spectral classification completeness. Galaxies do not recieve a spectral classification when a redshift is measured with $z\le0.15$ (and is therefore within the redshift range over which the PCA can be carried out), but the spectrum has too small a signal-to-noise ratio for the PCA to be applied successfully (typically ${\rm S}/{\rm N}<10$). The spectral classification success rate is given by the product of these two contributions. Our model for this effect, plotted as the solid curves in each panel of Fig. \[fig:eta.compl\], is in good agreement with the success rate realised in the 2dFGRS, shown by the histograms.
Rather than weight the data to compensate for a spectral classification success rate below $100\%$, we instead modulate the number of unclustered or random points laid down in each field in the clustering analysis to take into account the varying success rate. We have conducted a number of tests in which we varied the completeness thresholds used, adopted different weighting schemes using samples of higher completeness, and we have also compared our results with those from Norberg (2001), whose samples are not subject to spectral classification incompleteness. The results of these tests confirm that our clustering measurements are robust to changes to the details of how we model the incompleteness; this is largely due to our practice of restricting the analysis to high completeness fields. Excluding areas below our relatively high sector completeness threshold (see Colless 2001 for a definition), we estimate that the effective solid angle used in the SGP region is $\sim$ 380 , and in the NGP 250 .
Constructing a volume-limited sample {#sec:vol.lim}
------------------------------------
We analyse a series of volume-limited samples drawn from the 2dFGRS, following the strategy adopted by . The chief advantage of this approach is simplicity; the radial distribution of galaxies is uniform apart from modulations in space density due to clustering. Therefore, the complication of modelling the radial selection function in a flux-limited survey is avoided. This is particularly appealing for the current analysis, as separate selection functions would be required for each class of spectral type studied, since Madgwick (2002) have demonstrated that galaxies with different spectral types have different luminosity functions.
The disadvantage of using volume-limited samples is that a large fraction of galaxies in the flux-limited catalogue do not satisify the selection criteria. As point out, a volume-limited sample specified by a range in absolute magnitude has both a lower () and an upper redshift cut (), because the flux-limited catalogue has, in practice, bright and faint apparent magnitude limits. This seemingly profligate use of galaxy redshifts was a serious problem for previous generations of redshifts surveys. This is not, however, the case for the 2dFGRS, which contains sufficient galaxies to permit the construction of large volume-limited samples defined both by luminosity and spectral type. As we demonstrate in section \[sec:res\], the volume-limited samples we analyse are large enough, both in terms of volume and number of galaxies, to give extremely robust clustering measurements.
To construct a volume-limited sample, it is necessary to estimate the absolute magnitude that each galaxy would have at $z=0$. This requires assumptions about the variation of galaxy luminosity with wavelength and with redshift, or equivalently, with cosmic time. We make use of the class dependent $k$-corrections derived by Madgwick (2002). The mean weighted $k$-corrections are given by the following expressions: $$\begin{aligned}
k(z) &=& 2.6 z + 4.3 z^2 \,\,\, ({\rm early-types}) \\
k(z) &=& 1.5 z + 2.1 z^2 \,\,\, ({\rm late-types}) \\
k(z) &=& 1.9 z + 2.7 z^2 \,\,\, ({\rm full\,\,sample}). \end{aligned}$$ These $k$-corrections have the appeal that they are extracted directly in a self-consistent way from 2dFGRS spectra. However, no account is taken of evolution in the galaxy spectrum. The explicit inclusion of evolution could lead to the ambiguous situation whereby a galaxy’s spectral type changes with redshift. We have checked that our results are, in fact, insensitive to the precise choice of $k$-correction, comparing clustering results obtained with the spectral type dependent $k$-corrections given above with those obtained when a global $k+e$-correction (i.e. making an explicit attempt to account for galaxy evolution, albeit in an average sense) is applied (as in Norberg 2001).
Since the $k$-corrections are class dependent, the and values corresponding to a given absolute magnitude range are also slightly class dependent. Hence, the volumes defining the samples for two different spectral classes for the same bin in absolute magnitude will not coincide exactly. In addition to this subtle class dependent definition of the volumes, the values of and vary slightly with position on the sky. This is due to revisions made to the map of galactic extinction (Schlegel, Finkbeiner & Davis 1998) and to the CCD recalibration of APM plate zero-points since the definition of the original input catalogue.
Finally, throughout the paper, we adopt an $\Omega_{0}=0.3$, $\Lambda_{0}=0.7$ cosmology to convert redshift into comoving distance. The relative clustering strength of our samples is insensitive to this choice.
Estimating the two-point correlation function {#sec:meth}
=============================================
The galaxy correlation function is estimated on a two dimensional grid of pair separation parallel ($\pi$) and perpendicular ($\sigma$) to the line-of-sight. To estimate the mean density of galaxy pairs, a catalogue of randomly positioned points is generated with the same angular distribution and the same values of and as the data, taking into account the completeness of the survey as a function of position on the sky, as described in Section \[sec:compl\]. The correlation function is estimated using $$\xi_{\rm{H}} = \frac{DD\,RR}{DR^2}\,-\,1\, ,
\label{eq:ham}$$ where $DD$, $DR$ and $RR$ are the numbers of data-data, data-random and random-random pairs respectively in each bin (Hamilton 1993). This estimator does not require an explicit estimate of the mean galaxy density. We have also cross-checked our results using the estimator proposed by Landy & Szalay (1993): $$\xi_{\rm{LS}} = \frac{ DD - 2DR + RR }{RR},
\label{eq:ls}$$ where, this time $DD$, $DR$ and $RR$ are the suitably normalised numbers of data-data, data-random and random-random pairs. We find that the two estimators give the same results over the range of pair separations in which we are interested.
The clustering pattern of galaxies is distorted when radial positions are inferred from redshifts, as expected in the gravitational instability theory of structure formation (e.g. Kaiser 1987; Cole, Fisher & Weinberg 1994). Clear evidence for this effect is seen in the shape of the two point correlation function when plotted as $\xi(\sigma,\pi)$, as demonstrated clearly for galaxies in the 2dFGRS by Peacock (2001) and for groups of galaxies in the Zwicky catalogue by Padilla (2001). After giving a brief flavour of the clustering of 2dFGRS galaxies in redshift space in Section \[sec:redspace\], we focus our attention on clustering in real space in the remainder of the paper. The clustering signal in real space is inferred by integrating $\xi(\sigma,\pi)$ in the $\pi$ direction (i.e. along the line of sight): $$\frac{\Xi(\sigma)}{\sigma} = \frac{1}{\sigma} \int_{-\infty}^{\infty}
\xi(\sigma,\pi) {\rm d}\pi.
\label{eq:chiofsigma}$$ For the samples that we consider, the integral converges by pair separations of $\pi\,\ge\,50\,$. The projected correlation function can then be written as an integral over the spherically averaged real space correlation function, $\xi(r)$, $$\frac{\Xi(\sigma)}{\sigma} = \frac{2}{\sigma} \int_{\sigma}^{\infty} \xi(r)
\frac{r{\rm d}r}{\left(r^{2}-\sigma^{2}\right)^{1/2}},
\label{eq:projxi}$$ (Davis & Peebles 1983). If we assume that the real space correlation function is a power law (which is a fair approximation for APM galaxies out to separations around $r\sim10\,h^{-1}\,$Mpc, see e.g. Baugh 1996), then Eq. \[eq:projxi\] can be written as $$\frac{\Xi(\sigma)}{\sigma} = \left(\frac{r_{0}}{\sigma} \right)^{\gamma}
\frac{\Gamma(1/2)\Gamma( [\gamma-1]/2)}
{\Gamma(\gamma/2)} = \left(\frac{r_{0}}{\sigma} \right)^{\gamma} A(\gamma),
\label{eq:pow}$$ where $\Gamma(x)$ is the usual Gamma function, and we have used $\xi(r) = (r_{0}/r)^{\gamma}$, where $r_{0}$ is the real space correlation length and $\gamma$ is equal to the slope of the projected correlation function $\Xi(\sigma)/\sigma$. As we demonstrate in Section \[sec:xi.s\], the projected correlation function is well described by a power law. We study a range of samples containing different numbers of galaxies and covering different volumes of the Universe. It is imperative to include sampling fluctuations when estimating the errors on the measured correlation function, to allow a meaningful comparison of the results obtained from different samples. This contribution to the errors has often been neglected in previous work. Following , we employ a sample of 22 mock 2dFGRS catalogues drawn from the $\Lambda$CDM Hubble Volume simulation (Evrard 2002) to estimate the error bars on the measured correlation functions. The construction of these mock catalogues is explained in Baugh (2002, in preparation; see also Cole 1998 and Norberg 2002). These catalogues have the same selection criteria and the same clustering amplitude as measured for galaxies in the flux-limited 2dFGRS. We have experimented with ensembles of mock catalogues constructed with different clustering strengths to ascertain how best to assign error bars when the measured clustering has a different amplitude from that of our fiducial sample of 22 2dFGRS mocks. We found that the error bars obtained directly by averaging over a test ensemble of mocks are reproduced most closely by using the scaled fractional [*rms*]{} scatter derived from the fiducial ensemble of 22 mocks, rather than by taking the absolute error.
Clustering in Redshift Space {#sec:redspace}
============================
In this section we give a brief overview of the clustering of 2dFGRS galaxies in redshift space, for samples selected by luminosity and spectral type. First, in Section \[sec:cone\], we give a qualitative impression of the clustering differences by plotting the spatial distribution of galaxies in volume-limited samples. Then we quantify these differences by measuring the spherically averaged correlation function, $\xi(s)$. A more comprehensive analysis of the clustering of 2dFGRS galaxies in redshift space will be presented by Hawkins (2002).
Spatial distribution of 2dFGRS galaxies {#sec:cone}
---------------------------------------
It is instructive to gain a visual impression of the spatial distribution of 2dFGRS galaxies before interpreting the measured correlation functions. In Fig. \[fig:coneplot\], we show the spatial distribution of galaxies in two ranges of absolute magnitude: in panel (a) we show a sample of faint galaxies ($-18.0\,\ge\,M_{b_{\rm J}}-5\log_{10}\,h\,\ge\,-19.0$) and in panel (b) a sample of bright galaxies ($-20.0\,\ge\,M_{b_{\rm J}}-5\log_{10}\,h\,\ge\,-21.0$). Within each panel, early and late type galaxies, as distinguished by their spectral types, are plotted with different symbols; the positions of early-types are indicated by circles and the late-types are marked by stars. For clarity, we show only a three degree slice in declination cut from the SGP region and we have sparsely sampled the galaxies, so that the space densities of the two spectral classes are equal. In order to expand the scale of the plot, the range of redshifts shown is restricted, taking a subset of the full volume-limited sample in each case. (Note also that the redshift ranges differ between the two panels.)
A hierarchy of structures is readily apparent in these plots, ranging from isolated objects, to groups of a handful of galaxies and on through to rich clusters containing over a hundred members. It is interesting to see how structures are traced by galaxies in the different luminosity bins by comparing common structures between the two panels. For example, the prominent structure (possibly a supercluster of galaxies) seen at $\alpha\,\simeq\,0^{\rm h}$ and $z\,\simeq\,0.061$ is clearly visible in both panels. The same is true for the overdensity seen at $\alpha\,\simeq\,03^{\rm h}15'$ at $z\,\simeq\,0.068$.
This is the first time that a large enough survey has been available, both in terms of the volume spanned and the number of measured redshifts, to allow a comparison of the clustering of galaxies of different spectral types and luminosities in representative volume-limited samples, without the complication of the strong radial gradient in number density seen in flux-limited samples.
It is apparent from a comparison of the distribution of the different spectral types in Fig. \[fig:coneplot\](a), that the faint early-type galaxies tend to be grouped into structures on small scales whereas the faint late-types are more spread out. One would therefore anticipate that the early-types should have a stronger clustering amplitude than the late-types, an expectation that is borne out in Section \[sec:xi.s\].
In Fig. \[fig:coneplot\](b), the distinction between the distribution of the spectral types is less apparent. This is partly due to the greater importance of projection effects in the declination direction, as the cone extends to a greater redshift than in Fig. \[fig:coneplot\](a). However, close examination of the largest structures suggests that early-types are more abundant in them than late-types, again implying a stronger clustering amplitude.
$\xi(s)$ as function of luminosity and spectral type {#sec:xi.s}
----------------------------------------------------
In Fig. \[fig:xis\], we show the spherically averaged redshift space correlation function, $\xi(s)$, as a function of luminosity and spectral type. Results are shown for samples selected in bins of width one magnitude, as indicated by the legend in each panel. The top panel shows the correlation functions of all galaxies that have been assigned a spectral type, the middle panel shows results for galaxies classified as early-types ($\eta<-1.4$) and the bottom panel shows results for late-types ($\eta>-1.4$). Note that, at present, there are insufficient numbers of late-type galaxies to permit a reliable measurement of the correlation function for the brightest magnitude bin $-21.0\,\ge\,M_{b_{\rm J}}-5\log_{10}\,h\,\ge\,-22.0$.
Several deductions can be made immediately from Fig. \[fig:xis\]. In all cases, the redshift space correlation function is well described by a power-law [*only*]{} over a fairly limited range of scales. The correlation functions of early-type galaxies are somewhat steeper than those of late-types. However, the main difference is that the early-type galaxies have a stronger clustering amplitude than the late-type galaxies. The correlation length, defined here as the pair separation for which $\xi(s_{0})\,=\,1$, varies for early-types from $s_{0}\,=\,7.1\,\pm\,0.7\,$ for galaxies with absolute magnitudes around $M_{b_{\rm J}}-5\log_{10}\,h\,\sim\,-19.5$ to $s_{0}\,=\,8.9\,\pm\,0.7\,$ for the brightest sample with $-21.0\,\ge\,M_{b_{\rm J}}-5\log_{10}\,h\,\ge\,-22.0$. The faintest early-types, with magnitudes $-17.5\,\ge\,M_{b_{\rm J}}-5\log_{10}\,h\,\ge\,-18.5$, display a clustering amplitude that is similar to that of the brightest early-types. However, the measurement of the correlation function for this faint sample is relatively noisy, as the volume in which galaxies are selected is small compared with the volumes used for brighter samples. The late-type galaxies show, by contrast, little change in clustering amplitude with increasing luminosity, with a redshift space correlation length of $s_{0}\,=\,5.6\,\pm\,0.6\,$. Only a slight steepening of the redshift space correlation function is apparent with increasing luminosity, until the brightest sample, which displays a modest increase in the redshift space correlation length. The correlation lengths of all our samples of early-type galaxies are larger than those of late-type galaxies.
Clustering in real space {#sec:res}
========================
Robustness of clustering results {#sec:res.2mag}
--------------------------------
The approach adopted to study the dependence of galaxy clustering on luminosity relies upon being able to compare correlation functions measured in different volumes. It is important to ensure that there are no systematic effects, such as significant sampling fluctuations, that could undermine such an analysis. In we demonstrated the robustness of this approach in two ways. First, we constructed a volume-limited sample defined using a broad magnitude range, that could be divided into co-spatial subsamples of galaxies in different luminosity bins, i.e. subsamples within the same volume and therefore subject to the same large-scale structure fluctuations. A clear increase in clustering amplitude was found for the brightest galaxies in the volume, establishing the dependence of clustering on galaxy luminosity (see Fig. 1a of Norberg 2001). Secondly, we demonstrated that measuring the correlation function of galaxies in a fixed luminosity bin, but using samples taken from different volumes, gave consistent results (see Fig. 1b of Norberg 2001).
In this section we repeat these tests. The motivation for this exercise is that the samples considered here contain fewer galaxies than those used by as only galaxies with $z<0.15$ are suitable for PCA spectral typing, and because the samples are more dilute as they have been selected on the basis of spectral type as well as luminosity. In Fig. \[fig:xr.test\](a) we plot the projected correlation function of late-type galaxies in a fixed absolute magnitude bin ($-19.0\,\ge\,M_{b_{\rm J}}-5\log_{10}\,h\,\ge\,-20.0$), but measured for samples taken from volumes defined by different and . The clustering results are in excellent agreement with one another. In Fig. \[fig:xr.test\](b), we compare the projected correlation function of late-type galaxies in different absolute magnitude ranges but occupying the same volume. A clear difference in the clustering amplitude is seen. We have also performed these tests for early-type galaxies and arrived at similar conclusions.
As an additional test, we also show in Fig. \[fig:xr.test\] the correlation function measured in what we refer to as the [*optimal sample*]{} for a given magnitude bin. The optimal sample contains the maximum number of galaxies for the specified magnitude bin. The correlation functions of galaxies in optimal samples are shown by thin solid lines in both panels and are in excellent agreement with the other measurements shown.
Projected correlation function {#sec:res.xir}
------------------------------
Fig. \[fig:xr.tot\] shows how the real space clustering of galaxies of different spectral type depends on luminosity. We use the optimal sample for each magnitude bin, i.e. the volume-limited sample with the maximum possible number of galaxies, the properties of which are listed in Tables \[tab:eta0\] (early & late types together), \[tab:eta1\] (early-types only) and \[tab:eta6\] (late-types only).
The top panel of Fig. \[fig:xr.tot\] confirms the results found by , namely that the clustering strength of the full sample increases slowly with increasing luminosity for galaxies fainter than , and then shows a clear, strong increase for galaxies brighter than . (We take to be $M_{b_{\rm J}}-5\log_{10}\,h\,\simeq\,-19.7$, following Folkes 1999 and Norberg 2002.) Furthermore, the projected correlation functions are well described by a power law with a slope that is independent of luminosity. The middle panel of Fig. \[fig:xr.tot\] shows the projected correlation function of early-type galaxies for different absolute magnitude ranges. The clustering amplitude displays a non-monotonic behaviour, with the faintest sample having almost the same clustering strength as the brightest sample. The significance of this result for the faintest galaxies will be discussed further in the next section. Early-type galaxies with $M_{b_{\rm J}}-5\log_{10}\,h\,\simeq\,-19.5$, display weaker clustering than the faint and bright samples. The bottom panel of Fig. \[fig:xr.tot\] shows the real space clustering of late-type galaxies as a function of luminosity. In this case, the trend of clustering strength with luminosity is much simpler. There is an increase in clustering amplitude with luminosity, and also some evidence that the projected correlation function of the brightest subset is steeper than that of the other late-type samples. In general, for the luminosity ranges for which a comparison can be made, the clustering strength of early-type galaxies is always stronger than that of late-types.
The comparison of the correlation functions of different samples is made simpler if we divide the curves plotted in Fig. \[fig:xr.tot\] by a fiducial correlation function. As a reference sample we choose all galaxies that have been assigned a spectral type, with absolute magnitudes in the range $-19.5\,\ge\,M_{b_{\rm J}}-5\log_{10}\,h\,\ge\,-20.5$ (the short-dashed line in the top panel of Fig. \[fig:xr.tot\]). In Fig. \[fig:ratio\], we plot the ratio of the correlation functions shown in the panels of Fig. \[fig:xr.tot\], to the reference correlation function, with error bars obtained from the mock catalogues. The trends reported above for the variation of clustering strength with luminosity and spectral type are now clearly visible (see Fig. \[fig:ratio\]), particularly the difference in clustering amplitude between early-types and late-types. In the upper and lower panels, the ratios of correlation functions are essentially independent of $\sigma$, indicating that a single power-law slope is a good description over the range of scales plotted. The one exception is the brightest sample of late-type galaxies, which shows some evidence for a steeper power-law. In the middle panel, the ratios for early-type galaxies show tentative evidence for a slight steepening of the correlation function at small pair separations, $\sigma<2h^{-1}\,$Mpc, which is most pronounced for the brightest sample.
Real space correlation length {#sec:xi.r0}
-----------------------------
In the previous subsection, we demonstrated that the projected correlation functions of galaxies in the 2dFGRS have a power-law form with a slope that varies little as the sample selection is changed, particularly for pair separations in the range $2.0\,\le\,\sigma/(\,h^{-1}\,{\rm Mpc})\,\le\,15.0$. To summarize the trends in clustering strength found when varying the spectral type and luminosity of the sample, we fit a power-law over this range of scales. We follow the approach, based on Eq. \[eq:pow\], used by , who performed a $\chi^{2}$ minimisation to extract the best fitting values of the parameters in the power-law model for the real space correlation function: the correlation length, $r_0$, and the power law slope, $\gamma$. As pointed out by , a simple $\chi^2$ approach does not give reliable estimates of the errors on the fitted parameters because of the correlation between the estimates of the correlation function at differing pair separations. We therefore use the mock 2dFGRS catalogues to estimate the errors on the fitted parameters in the following manner. The best fitting values of $r_0$ and $\gamma$ are found for each mock individually, using the $\chi^2$ analysis. The estimated error is then taken to be the fractional [*rms*]{} scatter in the fitted parameters over the ensemble of mock catalogues. The best fitting parameters for each 2dFGRS sample are listed in Tables \[tab:eta0\] (early & late types together), \[tab:eta1\] (early-types only) and \[tab:eta6\] (late-types only). The results for the correlation length are plotted in the top panel of Fig. \[fig:r0\].
The correlation lengths estimated for the full sample with assigned spectral types (shown by the open squares in Fig. \[fig:r0\]) are in excellent agreement with the results of (shown by the filled circles). The bright samples constructed by have values in excess of the limit of $=0.15$ enforced upon the samples analysed in this paper by the PCA. The bright samples used in this paper therefore cover smaller volumes than those used by Norberg and so the error bars are substantially larger. A cursory inspection of Fig. \[fig:r0\] would give the misleading impression that we find weaker evidence for an increase in correlation length with luminosity. It is important to examine this plot in conjunction with Tables \[tab:eta0\] to \[tab:eta6\], which reveals that there is significant overlap in the volumes defined by the four brightest magnitude slices, because of the common $=0.15$ limit.
In this case, the error bars inferred directly from the mocks do not take into account that our samples are correlated. The errors fully incorporate cosmic variance, [*i.e.*]{} the variance in clustering signal expected when sampling a given volume placed at different, independent locations in the Universe. The volumes containing the four brightest samples listed in Table \[tab:eta1\] contain long-wavelength fluctuations in common and so clustering measurements from these different volumes are subject to a certain degree of coherency. The clearest way to show this is by calculating the difference between the correlation lengths fitted to two samples. The error on the difference, derived using the mock catalogues, has two components: the first comes from adding the individual errors in quadrature; the second is from the correlation of the samples. For correlated samples, this second term is negative and, therefore lowers the estimated error on the difference. This is precisely what is seen in the lower panel of Fig. \[fig:r0\], where we plot $\Delta\,r_0^\eta\,=\,r_{0}^\eta({\rm M_{b_{\rm J}}})
- r_{0}^\eta({\rm M_{b_{\rm J}}^{\rm ref}})$ for each spectral type as a function of absolute magnitude, with error bars taking into account the correlation of the samples. For all samples brighter than , the increase in clustering length with luminosity is clear.
There is a suggestion, in the top panel of Fig. \[fig:r0\], of a non-monotonic dependence of the correlation length on luminosity for early-type galaxies. The evidence for this behaviour is less apparent on the $\Delta\,r_0^\eta$ panel, where a difference in $r_0$ for the faintest sample is seen at less than the 2$\sigma$ level. These volumes are small compared to those defining the brighter samples. Furthermore, when analyzing the faintest sample for SGP and NGP separately, our results are somewhat sensitive to the presence of single structures in each region (at $\alpha\,\simeq\,0^{\rm h}$ and $z\,\simeq\,0.061$ for the SGP, as shown in Fig. \[fig:coneplot\](a)). Thus we conclude that the upturn at faint magnitudes in the correlation length of early-types is not significant.
The projected correlation function of early-type galaxies brighter than is well fitted by a power-law real space correlation function, with a virtually constant slope of $\gamma\,\simeq\,1.87\,$ and a correlation length which increases with luminosity, from $r_0\,=\,5.7\,\pm\,0.6\,$ for galaxies to $r_0\,=\,9.7\,\pm\,1.2\,$ for brighter galaxies ($M_{b_{\rm J}}-5\log_{10}\,h\,\simeq\,-21.2$). This represents an increase in clustering strength by a factor of $2.7$, as seen in Fig. \[fig:ratio\]. The projected correlation functions of late-type galaxies are also consistent with a power-law in real space, with an essentially constant slope. There is a very weak trend for $\gamma$ to increase with luminosity, although at little more than the $1 \sigma$ level. Ignoring this effect, the fitted slope of the late-type correlation function is $\gamma\,\simeq\,1.76\,$. The correlation length increases with luminosity from a value of $r_0\,=\,3.7\,\pm\,0.8\,$ for faint galaxies ($M_{b_{\rm J}}-5\log_{10}\,h\,\simeq\,-18.4$) to $r_0\,=\,6.3\,\pm\,1.0\,$ for bright galaxies ($M_{b_{\rm J}}-5\log_{10}\,h\,\simeq\,-20.7$), a factor of $2.5$ increase in clustering strength. It should be possible to extend the analysis for late-type galaxies beyond $M_{b_{\rm J}}-5\log_{10}\,h\,\simeq\,-21$ when the 2dFGRS is complete.
The top panel of Fig. \[fig:r0\] confirms our earlier conclusion that the clustering of early-type galaxies is stronger than that of late-type galaxies. At , early-types typically have a real space clustering amplitude that is $1.5-1.7$ times greater than that of late-types.
Discussion and conclusions {#sec:conc}
==========================
We have used the 2dFGRS to study the dependence of clustering on spectral type for samples spanning a factor of twenty in galaxy luminosity. The only previous attempt at a bivariate luminosity-morphology/spectral type analysis of galaxy clustering was performed by Loveday (1995) using the Stromlo-APM redshift survey. They were able to probe only a relatively narrow range in luminosity around $L^{\star}$, which is more readily apparent if one considers the median magnitude of each of their magnitude bins (see Fig. 3b of Norberg 2001). The scatter between spectral and morphological types illustrated in Fig. \[fig:jon.eta\] precludes a more detailed comparison of our results with those of earlier studies, based on morphological classifications.
In Norberg (2001), we used the 2dFGRS to make a precise measurement of the dependence of galaxy clustering on luminosity. The clustering amplitude was found to scale linearly with luminosity. One of the aims of the present paper is to identify the phenomena that drive this relation. In particular, there are two distinct hypotheses that we wish to test. The first is that there is a general trend for clustering strength to increase with luminosity, regardless of the spectral type of the galaxy. The second is that different types of galaxies have different clustering strengths, which may vary relatively little with luminosity, but a variation of the mix of galaxy types with luminosity results in a dependence of the clustering strength on luminosity.
Madgwick (2002) estimated the luminosity function of 2dFGRS galaxies for different spectral classes, and found that, in going from early-type to late-types, the slope of the faint end of the luminosity function becomes steeper while the characteristic luminosity becomes fainter. Another representation of the variation of the luminosity function with spectral class is shown in Fig. \[fig:ndens\], where we plot the fraction of early and late type galaxies in absolute magnitude bins. The plotted fractions are derived from the volume-limited samples listed in Tables 1, 2 and 3. The mix of spectral types changes dramatically with luminosity; faint samples are dominated by late-types, whereas early-types are the most common galaxies in bright samples. Similar trends were found for galaxies labelled by morphological type in the SSRS2 survey by Marzke (1998).
We find that the change in the mix of spectral types with luminosity is not the main cause for the increase in the clustering strength of the full sample with luminosity. To support this assertion, we plot in Fig. \[fig:bl\] the variation of clustering strength with luminosity normalized, for [*each*]{} spectral class, to the clustering strength of a fiducial sample of galaxies, [*i.e.*]{} the sample which covers the magnitude range $-19.5\,\ge\,M_{b_{\rm J}}-5\log_{10}\,h\,\ge\,-20.5$. For a galaxy sample with best fitting correlation function parameters $r^{i}_{0}$ and $\gamma^{i}$, we define the relative bias with respect to the sample of the same type by $$\frac{b^{i}}{b^{\star}}\Big|_{\eta}(r) =
\sqrt{\frac{ (r^{i}_{0})^{\gamma_{i}}}{r^{\gamma}_{0}}
r^{\gamma-\gamma_{i}}}\Big|_{\eta},
\label{eq:bias}$$ where $r_{0}$ and $\gamma$ are the best fitting power-law parameters for the fiducial sample. In Fig. \[fig:bl\], we plot the relative bias evaluated at a fixed scale, $r = 4.89$, which is the correlation length of the reference sample for all -classified galaxies. A scale dependence in Eq. \[eq:bias\] arises if the slopes of the real space correlation functions are different for the galaxy samples being compared. In practice, the term $r^{\gamma-\gamma_{i}}$ is close to unity for the samples considered. The dashed line shows a fit to the bias relation defined by the open symbols. The solid line shows the effective bias relation obtained by , which is defined in a slightly different way to the effective bias computed here. From Fig. \[fig:bl\], we see that the trend of increasing clustering strength with luminosity in both spectral classes is very similar for galaxies brighter than $L>0.5L^{\star}$. At the brightest luminosity, corresponding to $\sim\,4 L^{\star}$, the clustering amplitude is a factor of $2-2.5$ times greater than at $L^{\star}$. This increase is much larger than the offset in the relative bias factors of early and late types at any given luminosity. We conclude that the change in correlation length with absolute magnitude found by Norberg (2001) is primarily a luminosity effect rather than a reflection of the change in the mix of spectral types with luminosity.
Benson (2001) showed that a dependence of clustering strength on luminosity is expected in hierarchical clustering cold dark matter universes because of the preferential formation of the brightest galaxies in the most massive, strongly clustered dark halos. The close connection between the spectral characteristics of galaxies and their clustering properties discussed in this paper provides further evidence that the galaxy type is also related to the mass of the halo in which galaxy forms.
Acknowledgments {#acknowledgments .unnumbered}
===============
The 2dFGRS is being carried out using the 2 degree field facility on the 3.9m AngloAustralian Telescope (AAT). We thank all those involved in the smooth running and continued success of the 2dF and the AAT. We thank the referee, Dr. J. Loveday, for producing a speedy and helpful report. PN is supported by the Swiss National Science Foundation and an ORS award, and CMB acknowledges the receipt of a Royal Society University Research Fellowship. This work was supported in part by a PPARC rolling grant at Durham.
Baugh, C.M., 1996, MNRAS, 280, 267.
Benson, A.J., Baugh, C.M., Cole, S., Frenk, C.S., Lacey, C.G., 2000a, MNRAS, 316, 107.
Benson, A.J., Cole, S., Frenk, C.S., Baugh, C.M., Lacey, C.G., 2000b, MNRAS, 311, 793.
Benson, A.J., Frenk, C.S., Baugh, C.M., Cole, S., Lacey, C.G., 2001, MNRAS, 327, 1041. Berlind, A.A., Weinberg, D.H., 2002, ApJ submitted (astro-ph/0109001) Bland-Hawthorn, J., (the 2dFGRS Team), 2002, in preparation.
Blanton, M., 2000, ApJ, 544, 63.
Bromley, B.C., Press, W.H., Lin, H., Kirshner, R.P., 1998, ApJ, 505, 25.
Cole, S., Fisher, K.B., Weinberg, D.H., 1994, MNRAS, 267, 785.
Cole, S., Hatton, S., Weinberg, D.H., Frenk, C.S., 1998, MNRAS, 300, 945.
Colless, M., (the 2dFGRS Team), 2001, MNRAS, 328, 1039. Davis, M., Geller, M.J., 1976, ApJ, 208, 13.
Davis, M., Peebles, P.J.E., 1983, ApJ, 267, 465.
Evrard, A., (the Virgo Consortium), 2002, ApJ, submitted, astro-ph/0110246.
Folkes, S., (the 2dFGRS Team), 1999, MNRAS, 308, 459.
Guzzo, L., Strauss, M.A., Fisher, K.B., Giovanelli, R., Haynes, M.P., 1997, ApJ, 489, 37.
Hamilton, A.J.S., 1993, ApJ, 417, 19.
Hawkins, E., (the 2dFGRS Team), 2002, in preparation.
Hermit, S., Santiago, B.X., Lahav, O., Strauss, M.A., Davis, M., Dressler, A., Huchra, J.P., 1996, MNRAS, 283, 709.
Iovino, A., Giovanelli, R., Haynes, M., Chincarini, G., Guzzo, L., 1993, MNRAS, 265, 21.
Jenkins, A., Frenk, C.S., Pearce, F.R., Thomas, P.A., Colberg, J.M., White, S.D.M., Couchman, H.M.P., Peacock, J.A., Efstathiou, G., Nelson, A.H., 1998, ApJ, 499, 20.
Jenkins, A., Frenk, C.S., White, S.D.M., Colberg, J., Cole, S., Evrard, A.E., Couchman, H.M.P., Yoshida, N., 2001, MNRAS, 321, 372.
Kaiser, N., 1987, MNRAS, 227, 1.
Kauffmann, G., Nusser, A., Steinmetz, M., 1997, MNRAS, 286, 795.
Kauffmann, G., Colberg, J.M., Diaferio, A., White, S.D.M., 1999, MNRAS, 303, 188.
Kennicutt, R.C., 1992, ApJ, 388, 310.
Kochanek C.S., Pahre M.A., Falco E.E., 2002, ApJ submitted, astro-ph/0011458
Landy, S.D., Szalay, A.S., 1993, ApJ, 412, 64.
Lahav, O., Saslaw, W.C., 1992, ApJ, 396, 430.
Loveday, J., 1996, MNRAS, 278, 1025
Loveday, J., Maddox, S.J., Efstathiou, G., Peterson, B.A., 1995, ApJ, 442, 457.
Loveday, J., Tresse, L., Maddox, S.J., 1999, MNRAS, 310, 281.
Maddox, S.J., Efstathiou, G., Sutherland, W.J., Loveday, J., 1990a, MNRAS, 243, 692
Maddox, S.J., Efstathiou, G., Sutherland, W.J., Loveday, J., 1990b, MNRAS, 246, 433
Maddox, S.J., Efstathiou, G., Sutherland, W.J., Loveday, J., 1996, MNRAS, 283, 1227
Maddox, S.J., (the 2dFGRS Team), 2002, in preparation.
Madgwick, D.S., (the 2dFGRS Team), 2002, submitted to MNRAS, astro-ph/0107197.
Marzke, R.O., da Costa, L.N., Pellegrini, P.S., Willmer, C.N.A., Geller, M.J., 1998, ApJ, 503, 617.
Norberg, P., (the 2dFGRS Team), 2001, MNRAS, 328, 67. Norberg, P., (the 2dFGRS Team), 2002, MNRAS, submitted, astro-ph/0111011.
Padilla, N.D., Merchan, M.E., Valotto, C.A., Lambas, D.G., Maia, M.A.G., 2001, ApJ, 554, 873.
Peacock, J.A., Smith, R.E., 2000, MNRAS, 318, 1144
Peacock, J.A., (the 2dFGRS team), 2001, Nature, 410, 169.
Santiago, B.X., Strauss, M.A., 1992, ApJ, 387, 9.
Shectman, S.A., Landy, S.D., Oemler, A., Tucker, D.L., Lin, H., Kirshner, R.P., Schechter, P.L., 1996, ApJ, 470, 172.
Schlegel, D.J., Finkbeiner, D.P., Davis, M., 1998, ApJ, 500, 525.
Seljak, U., 2000, MNRAS, 318, 203.
Somerville, R.S., Lemson, G., Sigad, Y., Dekel, A., Kauffmann, G., White, S.D.M., 2001, MNRAS, 320, 289.
Tegmark, M., Bromley, B.C., 1999, ApJ, 518, L69.
Willmer, C.N.A., Da Costa, L.N., Pellegrini, P.S., 1998, AJ, 115, 869.
Zehavi, I., (the SDSS Collaboration), 2002, submitted to ApJ, astro-ph/0106476.
------------------------------ ------------------------------ --------------- --------------- --------------- ----------------- ----------------- -------------
Mag. range Median magnitude $N_{\rm gal}$ $z_{\rm min}$ $z_{\rm max}$ $r_{0}$ $\gamma$ $A(\gamma)$
$M_{b_{\rm J}}-5\log_{10} h$ $M_{b_{\rm J}}-5\log_{10} h$ ()
$-17.5 \quad -18.5$ $-17.98$ 8510 0.0164 0.0724 $5.19 \pm 0.95$ $1.68 \pm 0.12$ 4.14
$-18.0 \quad -19.0$ $-18.46$ 13795 0.0204 0.0886 $4.36 \pm 0.89$ $1.83 \pm 0.10$ 3.58
$-18.5 \quad -19.5$ $-18.93$ 19207 0.0255 0.1077 $4.65 \pm 0.61$ $1.80 \pm 0.08$ 3.68
$-19.0 \quad -20.0$ $-19.40$ 24675 0.0317 0.1302 $4.93 \pm 0.50$ $1.79 \pm 0.10$ 3.71
$-19.5 \quad -20.5$ $-19.85$ 22555 0.0394 0.1500 $4.89 \pm 0.48$ $1.79 \pm 0.05$ 3.71
$-20.0 \quad -21.0$ $-20.30$ 10399 0.0487 0.1500 $5.37 \pm 0.61$ $1.78 \pm 0.11$ 3.75
$-20.5 \quad -21.5$ $-20.74$ 3423 0.0602 0.1500 $6.57 \pm 0.83$ $1.83 \pm 0.23$ 3.58
$-21.0 \quad -22.0$ $-21.19$ 751 0.0739 0.1500 $8.47 \pm 0.97$ $1.80 \pm 0.29$ 3.68
------------------------------ ------------------------------ --------------- --------------- --------------- ----------------- ----------------- -------------
\[tab:eta0\]
------------------------------ ------------------------------ --------------- --------------- --------------- ----------------- ----------------- -------------
Mag. range Median magnitude $N_{\rm gal}$ $z_{\rm min}$ $z_{\rm max}$ $r_{0}$ $\gamma$ $A(\gamma)$
$M_{b_{\rm J}}-5\log_{10} h$ $M_{b_{\rm J}}-5\log_{10} h$ ()
$-17.5 \quad -18.5$ $-18.05$ 1909 0.0163 0.0707 $8.33 \pm 1.82$ $1.87 \pm 0.23$ 3.46
$-18.0 \quad -19.0$ $-18.53$ 3717 0.0203 0.0861 $6.28 \pm 1.46$ $1.98 \pm 0.11$ 3.19
$-18.5 \quad -19.5$ $-18.98$ 6405 0.0253 0.1041 $5.92 \pm 1.00$ $1.83 \pm 0.10$ 3.58
$-19.0 \quad -20.0$ $-19.44$ 10135 0.0314 0.1249 $5.71 \pm 0.57$ $1.87 \pm 0.09$ 3.46
$-19.5 \quad -20.5$ $-19.89$ 11346 0.0388 0.1486 $5.66 \pm 0.56$ $1.87 \pm 0.09$ 3.46
$-20.0 \quad -21.0$ $-20.33$ 6434 0.0480 0.1500 $6.10 \pm 0.72$ $1.80 \pm 0.12$ 3.68
$-20.5 \quad -21.5$ $-20.77$ 2587 0.0590 0.1500 $7.60 \pm 1.02$ $1.87 \pm 0.26$ 3.46
$-21.0 \quad -22.0$ $-21.21$ 686 0.0722 0.1500 $9.74 \pm 1.16$ $1.95 \pm 0.37$ 3.26
------------------------------ ------------------------------ --------------- --------------- --------------- ----------------- ----------------- -------------
\[tab:eta1\]
------------------------------ ------------------------------ --------------- --------------- --------------- ----------------- ----------------- ------------- --
Mag. range Median magnitude $N_{\rm gal}$ $z_{\rm min}$ $z_{\rm max}$ $r_{0}$ $\gamma$ $A(\gamma)$
$M_{b_{\rm J}}-5\log_{10} h$ $M_{b_{\rm J}}-5\log_{10} h$ ()
$-17.5 \quad -18.5$ $-17.96$ 6674 0.0164 0.0734 $4.27 \pm 0.81$ $1.65 \pm 0.12$ 4.29
$-18.0 \quad -19.0$ $-18.44$ 9992 0.0205 0.0901 $3.71 \pm 0.77$ $1.76 \pm 0.11$ 3.82
$-18.5 \quad -19.5$ $-18.90$ 12619 0.0256 0.1099 $4.17 \pm 0.64$ $1.79 \pm 0.10$ 3.71
$-19.0 \quad -20.0$ $-19.37$ 14420 0.0319 0.1333 $4.45 \pm 0.47$ $1.76 \pm 0.09$ 3.82
$-19.5 \quad -20.5$ $-19.82$ 11122 0.0397 0.1500 $4.59 \pm 0.44$ $1.76 \pm 0.07$ 3.82
$-20.0 \quad -21.0$ $-20.26$ 4300 0.0492 0.1500 $5.52 \pm 0.88$ $1.87 \pm 0.13$ 3.46
$-20.5 \quad -21.5$ $-20.71$ 1118 0.0608 0.1500 $6.33 \pm 1.01$ $2.01 \pm 0.29$ 3.12
$-21.0 \quad -22.0$ $-21.17$ 198 0.0749 0.1500 $-$ $-$ $-$
------------------------------ ------------------------------ --------------- --------------- --------------- ----------------- ----------------- ------------- --
\[tab:eta6\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Orthogonal polynomials with respect to a weight function defined on a wedge in the plane are studied. A basis of orthogonal polynomials is explicitly constructed for two large class of weight functions and the convergence of Fourier orthogonal expansions is studied. These are used to establish analogous results for orthogonal polynomials on the boundary of the square. As an application, we study the statistics of the associated determinantal point process and use the basis to calculate Stieltjes transforms.'
address:
- |
Department of Mathematics\
Imperial College\
London\
United Kingdom
- |
Department of Mathematics\
University of Oregon\
Eugene, Oregon 97403-1222.
author:
- Sheehan Olver
- Yuan Xu
title: Orthogonal structure on a wedge and on the boundary of a square
---
[^1] [^2]
Introduction
============
Let $\Omega$ be a wedge on the plane that consists of two line segments sharing a common endpoint. For a positive measure $d\mu$ defined on $\Omega$, we study orthogonal polynomials of two variables with respect to the bilinear form $$\la f, g\ra = \int_{\Omega} f(x,y) g(x,y) d\mu.$$ We also study orthogonal polynomials on the boundary of a parallelogram. Without loss of generality we can assume that our wedge is of the form $$\label{eq:1.1}
\Omega = \{(x_1,1): x_1 \in [0,1]\} \cup \{(1,x_2): x_2 \in [0,1]\}$$ and consider the bilinear form defined by $$\label{eq:1.2}
\la f, g\ra = \int_0^1 f(x,1)g(x,1) w_1(x) dx + \int_0^1 f(1,y)g(1,y) w_2(y) dy.$$
Since $\Omega$ is a subset of the zero set of a quadratic polynomial $l_1(x,y) l_2(x,y)$, where $l_1$ and $l_2$ are linear polynomials, the structure of orthogonal polynomials on $\Omega$ is very different from that of ordinary orthogonal polynomials in two variables [@DX] but closer to that of spherical harmonics. The latter are defined as homogeneous polynomials that satisfy the Laplace equation $\Delta Y = 0$ and are orthogonal on the unit circle, which is the zero set of the quadratic polynomial $x^2 + y^2-1$. The space of spherical polynomials of degree $n$ has dimension 2 for each $n \ge 1$ and, furthermore, one basis of spherical harmonics when restricted on the unit circle are $\cos n \theta$ and $\sin n \theta$, in polar coordinates $(r,\t)$, and the Fourier orthogonal expansions in spherical harmonics coincide with the classical Fourier series.
In [§\[Section:Wedge\]]{}, we consider orthogonal polynomials on a wedge. The space of orthogonal polynomials of degree $n$ has dimension 2 for each $n \ge 1$, just like that of spherical harmonics, and they satisfy the equation $\partial_1 \partial_2 Y = 0$. The main results are
- An explicit expression in terms of univariate orthogonal polynomials when $w_1(x) = w_2(x) =w(x)$ where $w$ is any weight function on $[0,1]$ (Theorem \[thm:ipd-w-op\]),
- Sufficient conditions for pointwise and uniform convergence (Theorem \[thm:pointwise\]), as well as normwise convergence (Corollary \[cor:normwiseconv\]),
- Explicit expression in terms of Jacobi polynomials when $w_1(x) = w_{\a,\g}(x)$ and $w_2(x) = w_{\b,\g}(x)$ (Theorem \[thm:OP-ipd-abg\]),
- Sufficient conditions for normwise convergence (Theorem \[thm:3.4\]).
In [§\[Section:BoundarySquare\]]{} we study orthogonal polynomials on the boundary of a parallelogram, which we can assume as the square $[-1,1]^2$ without loss of generality. For a family of generalized Jacobi weight functions that are symmetric in both $x$ and $y$, we are able to deduce an orthogonal basis in terms of four families of orthogonal bases on the wedge in Theorem \[thm:boundaryOP\]. Furthermore, the convergence of the Fourier orthogonal expansions can also be deduced in this fashion, as shown in Theorem \[thm:squareconv\].
In [§\[Section:Square\]]{} we use orthogonal polynomials on the boundary of the square to construct an orthogonal basis for the weight function $w(\max\{|x|,|y|\})$ on the square $[-1,1]^2$. This mirrors the way in which spherical harmonics can be used to construct a basis of orthogonal polynomials for the weight function $w(\sqrt{x^2+y^2})$ on the unit disk. However, unlike the unit disk, the orthogonal basis we constructed are no longer polynomials in $x,y$ but are polynomials of $x, y$ and $s=\max\{|x|,|y|\}$.
The study is motivated by applications. In particular, we wish to investigate how the applications of univariate orthogonal polynomials can be translated to multivariate orthogonal polynomials on curves. As a motivating example, univariate orthogonal polynomials give rise to a determinantal point process that is linked to the eigenvalues of unitary ensembles of random matrix theory. In [§\[sec:detpointprocess\]]{}, we investigate the statistics of the determinantal point process generated from orthogonal polynomials on the wedge, and find experimentally that they have the same local behavior as a Coulomb gas away from the corners/edges.
In Appendix \[sec:JacobiOperators\], we give the Jacobi operators associated with a special case of weights on the wedge, whose entries are rational. Finally, in Appendix \[sec:Stieltjes\] we show that the Stieltjes transform of our family of orthogonal polynomials satisfies a recurrence that can be built out of the Jacobi operators of the orthogonal polynomials, which can in turn be used to compute Stieltjes transforms numerically. This is a preliminary step towards using these polynomials for solving singular integral equations.
Orthogonal polynomials on a wedge {#Section:Wedge}
=================================
Let $\CP_n^2$ denote the space of homogeneous polynomials of degree $n$ in two variables; that is, $\CP_n^2 =\mathrm{span}\, \{x^{n-k}y^k: 0 \le k \le n\}$. Let $\Pi_n^2$ denote the space of polynomials of degree at most $n$ in two variables.
Orthogonal structure on a wedge
-------------------------------
Given three non-collinear points, we can define a wedge by fixing one point and joining it to other points by line segments. We are interested in orthogonal polynomials on the wedge. Since the three points are non-collinear, each wedge can be mapped to $$\Omega = \{(x_1,1): x_1 \in [0,1]\} \cup \{(1,x_2): x_2 \in [0,1]\}$$ by an affine transform. Since the polynomial structure and the orthogonality are preserved under the affine transform, we can work with the wedge $\Omega$ without loss of generality. Henceforth we work only on $\Omega$.
Let $w_1$ and $w_2$ be two nonnegative weight functions defined on $[0,1]$. We consider the bilinear form define on $\Omega$ by $$\label{eq:ipd-w}
\la f,g\ra_{w_1,w_2} := \int_0^1 f(x,1) g(x,1)w_1(x) dx + \int_0^1 f(1,y) g(1,y)w_2(y) dy.$$ Let $I$ be the polynomial ideal of $\RR[x,y]$ generated by $(1-x)(1-y)$. If $f \in I$, then $\la f, g\ra_{w_1,w_2} =0$ for all $g$. The bilinear form defines an inner product on $\Pi_n^2$, modulo $I$, or equivalently, on the quotient space $\RR[x,y]/I$.
Let $\CH_n^2(w_1,w_2)$ be the space of orthogonal polynomials of degree $n$ in $\RR[x,y]/I$. Then $$\dim \CH_0^2(w_1,w_2)=1 \quad\hbox{and} \quad \dim \CH_n^2(w_1,w_2) =2, \quad n \ge 1.$$ Furthermore, we can choose polynomials in $\CH_n^2(w_1,w_2)$ to satisfy $\partial_x \partial_y p = 0$.
Since $(1-x)(1-y) \CP_{n-2}$ is a subset of $I$, the dimension of $\dim \CH_n^2(w_1,w_2) \le 2$. Applying the Gram–Schmidt process on $\{1, x^k, y^k, k \ge 1\}$ shows that there are two orthogonal polynomials of degree exactly $n$. Both these polynomials can be written in the form of $p(x) + q(y)$, since we can use $xy \equiv x+y-1$ mod $I$ to remove all mixed terms. Evidently such polynomials satisfy $\partial_x \partial_y (p(x) + q(y))=0$.
In the next two subsections, we shall construct an orthogonal basis of $\CH_n^2(w_1,w_2)$ for certain $w_1$ and $w_2$ and study the convergence of its Fourier orthogonal expansions. We will make use of results on orthogonal polynomials of one variable, which we briefly record here.
For $w$ defined on $[0,1]$, we let $p_n(w)$ denote an orthogonal polynomial of degree $n$ with respect to $w$, and let $h_n(w)$ denote the norm square of $p_n(w)$, $$h_n(w) := \int_0^1 |p_n(w;x)|^2 w(x) dx.$$ Let $L^2(w)$ denote the $L^2$ space with respect to $w$ on $[0,1]$. The Fourier orthogonal expansion of $f \in L^2(w)$ is defined by $$f = \sum_{n=1}^\infty {\widehat}f_n(w) p_n(w) \quad \hbox{with} \quad {\widehat}f_n(w) = \frac1{h_n(w)} \int_0^1 f(y) p_n(w;y) w(y)dy.$$ The Parseval identity implies that $$\|f\|_{L^2(w,[0,1])}^2 = \sum_{n=0}^\infty \left|{\widehat}f_n(w) \right|^2 h_n(w).$$ The $n$-th partial sum of the Fourier orthogonal expansion with respect to $w$ can be written as an integral $$\label{eq:partial-sum}
s_n(w;f)(x) := \sum_{k=1}^n {\widehat}f_k(w) p_k(w;x) = \int_{-1}^1 f(y) k_n(w;x,y)w(y)dy,$$ where $k_n(w)$ denotes the reproducing kernel defined by $$\label{eq:reprod-kernel}
k_n(w;x,y) = \sum_{k=0}^n \frac{p_k(w;x) p_k(w;y)}{h_k(w)}.$$
Orthogonal structure for $w_1=w_2$ on a wedge
---------------------------------------------
In the case of $w_1 = w_2 = w$, we denote the inner product by $\la \cdot,\cdot\ra_w$ and the space of orthogonal polynomials by $\CH_n^2(w)$. In this case, an orthogonal basis for $\CH_n^2(w)$ can be constructed explicitly.
\[thm:ipd-w-op\] Let $w$ be a weight function on $[0,1]$ and let $\phi w(x): = (1-x)^2 w(x)$. Define $$\begin{aligned}
\label{eq:PnQ}
\begin{split}
P_n(x,y) & = p_n(w;x)+ p_n(w;y) - p_n(w;1), \quad n= 0,1,2,\ldots,\\
Q_n(x,y) & = (1-x) p_{n-1}(\phi w; x) - (1-y) p_{n-1}(\phi w; y), \quad n=1,2, \ldots.
\end{split}\end{aligned}$$ Then $\{P_n, Q_n\}$ are two polynomials in $\CH_n^2(w)$ and they are mutually orthogonal. Furthermore, $$\label{eq:PnQnorm}
\la P_n, P_n\ra_w = 2 h_n(w) \quad\hbox{and}\quad \la Q_n, Q_n\ra_w = 2 h_{n-1}(\phi w). \quad$$
Since $P_n(x,1) = P_n(1,x)$ and $Q_n(x,1) = - Q_n(1,x)$, it follows that $$\la P_n, Q_m \ra_{w} = \int_0^1 P_n(x,1) Q_m(x,1) w(x)dx + \int_0^1 P_n(1,x) Q_m(1,x) w(x)dx =0$$ for $n \ge 0$ and $m \ge 1$. Furthermore, $$\la P_n, P_m\ra_{w} = 2 \int_0^1 p_n(w;x) p_m(w; x) w(x) dx = 2 h_n(w)\delta_{n,m}$$ by the orthogonality of $p_n(w)$. Similarly, $$\la Q_n, Q_m\ra_{w} = 2 \int_0^1 p_{n-1}(\phi w; x)p_{m-1}(\phi w; x)(1-x)^2 w(x) dx = 2 h_{n-1}(\phi w) \delta_{n,m}.$$ The proof is completed.
Let $L^2(\Omega,w)$ be the space of Lebesgue measurable functions with finite $$\|f\|_{L^2(\Omega,w)}: = \la f,f\ra_w^{\f12} = \left( \|f(\cdot, 1)\|_{L^2(w,[0,1])}^2 + \|f(1,\cdot)\|_{L^2(w,[0,1])}^2\right)^{\f12}$$ norms. For $f \in L^2(\Omega, w)$, its Fourier expansion is given by $$f = {\widehat}f_0 + \sum_{n=1}^\infty \left[ {\widehat}f_{P_n}P_n + {\widehat}f_{Q_n}Q_n \right],$$ where $P_n$ and $Q_n$ are defined in Theorem \[thm:ipd-w-op\] and $${\widehat}f_0:= \frac{ \la f,1\ra_w}{\la 1,1\ra_w}, \qquad
{\widehat}f_{P_n}:= \frac{\la f,P_n \ra_w}{\la P_n,P_n \ra_w}, \qquad {\widehat}f_{Q_n}:= \frac{\la f,Q_n \ra_w}{\la Q_n,Q_n \ra_w}.$$ The partial sum operator $S_n f$ is defined by $$S_n f := {\widehat}f_0 + \sum_{k=1}^n \left[ {\widehat}f_{P_k}P_k + {\widehat}f_{Q_k}Q_k \right],$$ which can be written in terms of an integral in terms of the reproducing kernel $K_n(\cdot,\cdot)$, $$S_n f(x_1,x_2) = \la f, K_n((x_1,x_2), \cdot) \ra_w,$$ where $$K_n((x_1,x_2), (y_1,y_2)):= \frac{1}{\la 1,1\ra_w} + \sum_{k=1}^n \left[ \frac{P_k(x_1,x_2)P_k(y_1,y_2)}{\la P_k,P_k \ra_w}
+ \frac{Q_k(x_1,x_2)Q_k(y_1,y_2)}{\la Q_k,Q_k \ra_w} \right].$$ We show that this kernel can be expressed, when restricted on $\Omega$, in terms of the reproducing kernel $k_n(w;\cdot,\cdot)$ defined at .
\[prop:repkernelsym\] The reproducing kernel $K_n(\cdot,\cdot)$ for $\la \cdot,\cdot\ra_w$ satisfies $$\begin{aligned}
K_n((x,1),(y,1)) &\, = K_n((1,x),(1,y)) \\
&\, = \frac12 k_n(w;x,y) + \frac12 (1-x)(1-y)k_{n-1}(\phi w; x,y), \notag \\
K_n((x,1),(1,y)) & \, = K_n((1,x),(y,1)) \\
& \, = \frac12 k_n(w;x,y) - \frac12 (1-x)(1-y)k_{n-1}(\phi w; x,y). \notag\end{aligned}$$
By and , $$\begin{aligned}
K_n((x,1),(y,1)) & \, = \frac{1} {2h_0(w)} + \sum_{k=1}^n \frac{p_k(w; x) p_k(w;y)}{2 h_k(w)} \\
& \qquad\qquad
+ \sum_{k=1}^n \frac{(1-x)p_{k-1}(\phi w;x) (1-y) p_{k-1}(\phi w;y)}{ 2 h_{k-1}(\phi w)} \\
&\, = \frac12 k_n(w;x,y) + \frac12 (1-x)(1-y)k_{n-1}(\phi w; x,y).\end{aligned}$$ The other case is established similar, using $Q_k(1,y) = - (1-y) p_{k-1}(\phi w; y)$.
It is well-known that the kernel $k_n(w; \cdot,\cdot)$ satisfies the Christoffel–Darboux formula, which plays an important role for the study of Fourier orthogonal expansion. Our formula allows us to write down an analogue of Christoffel–Darboux formula for $K_n(\cdot,\cdot)$, but we can derive convergence directly.
\[thm:pointwise\] Let $f$ be a function defined on $\Omega$. Define $$f_e(x) := \frac12(f(x,1) + f(1,x)) \quad \hbox{and} \quad f_o(x) := \frac12\frac{f(x,1)-f(1,x)}{1-x}.$$ Then $$\begin{aligned}
S_n f(x_1,1) & \,= s_n(w; f_e, x_1) + (1-x_1) s_{n-1} (\phi w; f_o, x_1), \label{eq:Sn=1} \\
S_n f(1,x_2) & \,= s_n(w; f_e, x_2) - (1-x_2) s_{n-1} (\phi w; f_o, x_2). \label{eq:Sn=2} \end{aligned}$$ In particular, if $s_n(w;f_e,x) \to f_e(x)$ and $s_n(\phi w;f_o,x) \to f_o(x)$, pointwise or in the uniform norm as $n\to \infty$, then $S_n f(x)$ converges to $f(x)$ likewise.
By our definition, $$\begin{aligned}
S_n f(x_1,1) = &\, \int_0^1 f(y,1) K_n((x_1,1),(y,1)) w(y) dy + \int_0^1 f(1,y) K_n((x_1,1),(1,y)) w(y) dy \\
= &\, \frac12 \int_0^1 f(y,1) \left[ k_n(w; x_1,y)+ (1-x_1)(1-y) k_{n-1}(\phi w; x_1,y) \right] w(y) dy \\
& \,+ \frac12 \int_0^1 f(1,y) \left[ k_n(w; x_1,y)- (1-x_1)(1-y) k_{n-1}(\phi w; x_1,y) \right] w(y) dy \\
= &\, s_n(w; f_e, x_1) + (1-x_1) s_{n-1} (\phi w;f_o,x_1). \end{aligned}$$ Similarly, $$\begin{aligned}
S_n f(1,x_2) = &\, \int_0^1 f(y,1) K_n((1,x_2),(y,1)) w(y) dy + \int_0^1 f(1,y) K_n((1,x_2),(1,y)) w(y) dy \\
= &\, \frac12 \int_0^1 f(y,1) \left[ k_n(w; x_2,y) - (1-x_2)(1-y) k_{n-1}(\phi w; x_2,y) \right] w(y) dy \\
& \,+ \frac12 \int_0^1 f(1,y) \left[ k_n(w; x_2,y) + (1-x_2)(1-y) k_{n-1}(\phi w; x_2,y) \right] w(y) dy \\
= &\, s_n(w; f_e, x_2) - (1-x_2) s_{n-1} (\phi w;f_o,x_2). \end{aligned}$$ Moreover, since $f_e(x) + (1-x) f_o(x) = f(x,1)$ and $f_e(x) - (1-x) f_o(x) = f(1,x)$, it follows that $$\begin{aligned}
S_n f (x_1,1) - f(x_1) = &\, s_n(w; f_e, x_1) - f_e(x_1)+ (1-x_1) \left(s_{n-1} (\phi w;f_o, x_1) - f_o(x_1) \right),\\
S_n f (1,x_2) - f(x_2) = &\, s_n(w; f_e, x_2) - f_e(x_2)- (1-x_2) \left(s_{n-1} (\phi w;f_o, x_2) - f_o(x_2) \right)\end{aligned}$$ from which we see that the convergence of $s_n(w;f_e)$ and $s_n(\phi w; f_o)$ imply the convergence of $S_n f$.
Since $f\in L^2(\Omega,w)$, it is evident that $f_e \in L^2(w)$. Moreover, $f_o \in L^2(\phi w)$ since $$\int_0^1 |f_o(x)|^2 \phi w(x) dx = \int_0^1 |f(x,1) - f(1,x)|^2 w(x) dx \le 2 \|f\|_{L^2(\Omega, w)}^2.$$ In particular, $s_n (w, f_e)$ and $s_n(\phi w; f_o)$ converge to $f_e$ and $f_o$ in $L^2(w)$ and in $L^2(\phi w)$, respectively.
\[cor:normwiseconv\] If $f\in L^2(\Omega, w)$, then $$\|f - S_n(f) \|_{L^2(\Omega,w)}^2 = 2 \left( \|s_{n} (w;f_e) - f_e\|_{L^2(w)}^2
+ \|s_{n-1} (\phi w;f_o) - f_o\|_{L^2(\phi w)}^2 \right).$$
By the displayed formulas at the end of the proof of the last theorem and $$\int_0^1 |(1-x) g(x)|^2 w(x) dx = \int_0^1 |g(x)|^2 \phi w(x) dx = \|g\|_{L^2(\phi w)}^2,$$ it is easy to see that $$\begin{aligned}
\|S_n f - f\|_{L^2(\Omega, w)}^2 = \, & \|s_n(w; f_e) - f_e+
(1- \{\cdot\}) \left(s_{n-1} (\phi w;f_o) - f_o \right)\|_{L^2(w)}^2 \\
& + \|s_n(w; f_e) - f_e - (1- \{\cdot\}) \left(s_{n-1} (\phi w;f_o) - f_o \right)\|_{L^2(w)}^2 \\
= \, & 2 \left(\|s_{n} (w;f_e) - f_e\|_{L^2(w)}^2
+ \|s_{n-1} (\phi w;f_o) - f_o\|_{L^2(\phi w)}^2 \right),\end{aligned}$$ where we have used the identity $(a+b)^2 + (a-b)^2 = 2 (a^2+b^2)$.
Orthogonal structure on a wedge with Jacobi weight functions
------------------------------------------------------------
For $\a,\g > -1$, let $w_{\a,\g}$ be the Jacobi weight function defined by $$w_{\a,\g}(x):= x^\a (1-x)^\g, \qquad x \in [0,1].$$ We consider the inner product $\la \cdot,\cdot\ra_{w_1,w_2}$ defined in with $w_1(x) = w_{\a,\g}(x)$ and $w_2(x) = w_{\b,\g}(x)$. More specifically, for $\a,\b,\g > -1$ and $\sigma > 0$, we define $$\begin{aligned}
\la f,g\ra_{\a,\b,\g}:= c_{\a,\g} \int_0^1 f(x,1) g(x,1) w_{\a,\g}(x) dx
+\sigma c_{\b,\g} \int_0^1 f(1,y) g(1,y) w_{\b,\g}(y) dy, \end{aligned}$$ where $$c_{\a,\g} := \Big(\int_0^1w_{\a,\g}(x) dx \Big)^{-1} = \frac{\Gamma(\g+\a+2)}{\Gamma(\g+1)\Gamma(\a+1)}.$$
### Orthogonal structure
We need to construct an explicit basis of $\CH_n^2(w_{\a,\g},w_{\b,\g})$. The case $\a = \b$ can be regarded as a special case of Theorem \[thm:ipd-w-op\]. The case $\a \ne \b$ is much more complicated, for which we need several properties of the Jacobi polynomials.
Let $P_n^{(\a,\b)}$ denote the usual Jacobi polynomial of degree $n$ defined on $[-1,1]$. Then $P_n^{(\g,\a)}(2x-1)$ is an orthogonal polynomial with respect to $w_{\a,\g}$ on $[0,1]$. Moreover, $$\begin{aligned}
\label{Jacobi-norm}
h_n^{\a,\g} := &\, c_{\a,\g} \int_0^1 \left[P_n^{(\g,\a)}(2x-1)\right]^2 w_{\a,\g}(x) dx \\
= & \, \frac{(\g+1)_n(\a+1)_n(n+\g+\a+1)}{ n! (\g+\a+2)_n(2n+\g+\a+1)}\notag\end{aligned}$$ by [@Szego (4.3.3)]. Furthermore, $P_n^{(\a,\b)}(1) = \binom{n+\a}{n}$ and, in particular, $P_n^{(0,\b)}(1) =1$. Our construction relies on the following lemma.
For $m > n \ge 0$, $$\begin{aligned}
I_{m,n}^{\a,\g}:= & c_{\a,\g} \int_{0}^1 P_n^{(\g,\a)}(2x-1) P_{m-1}^{(\g+2,\a)}(2x-1) (1-x)^{\g+1} x^\a dx \\
= & \begin{cases} 0, & n > m, \\
\displaystyle{ \frac{- m (\g+1)_m(\a+1)_m}{m!(2m+\g+\a+1) (\g+\a+2)_m},} & n = m, \\
\\
\displaystyle{\frac{(\g+1)(\a+1)_{m-1}(\g+1)_n}{(\g+\a+2)_{m}n!}}, & n < m.
\end{cases}\end{aligned}$$
Since $P_n^{(\g,\a)}(2x-1)$ is an orthogonal polynomial of degree $n$ with respect to $(1-x)^\g x^{\a}$ on $[0,1]$, $I_{m,n}^{\g,\a} = 0$ for $n > m$ holds trivially. For $m \ge n$, we need two identities of Jacobi polynomials. The first one is, see [@Szego (4.5.4)] or [@DLMF (18.9.6)], $$(2m + \g+\a+1) (1-x)P_{m-1}^{(\g+2, \a)}(2x-1) = (m+\g+ 1) P_{m-1}^{(\g+1,\a)}(2x-1)- m P_m^{(\g+1,\a)}(2x-1)$$ and the second one is the expansion, see [@DLMF (18.18.14)], $$P_m^{(\g+1,\a)}(2x-1) = \frac{(\a+1)_m}{(\g+\a+2)_m}\sum_{k=0}^m \frac{(\g+\a+1)_k (2k+\g+\a+1)}{(\a+1)_k (\g+\a+1)}
P_k^{(\g,\a)}(2x-1).$$ Putting them together shows that $$\begin{aligned}
\label{eq:Jacobi-A}
(1-x) P_{m-1}^{(\g+2,\a)}(2x-1) = & \frac{(\g+1)(\a+1)_{m-1}}{(\g+\a+1)_{m+1}} \\
& \times \sum_{k=0}^{m-1} \frac{(\g+\a+1)_k (2k+\g+\a+1)}{(\a+1)_k}P_k^{(\g,\a)}(2x-1) \notag \\
& - \frac{m}{m+\g+\a+1}P_m^{(\g,\a)}(2x-1). \notag \end{aligned}$$ Substituting this expression into $I_{m,n}^{\g,\a}$ and using the orthogonality of the Jacobi polynomials and , we conclude that, for $m-1 \ge n$, $$I_{m,n}^{\a,\g} = \frac{(\g+1)(\a+1)_{m-1}}{(\g+\a+2)_{m}} \frac{ (\g+1)_n }{n!}.$$ Hence, the case $m > n$ follows. The same argument works for the case $n =m$.
What is of interest for us is the fact that the dependence of $I_{m,n}^{\g,\a}$ on $n$ and $\a$ is separated, which is critical to prove that $Q_n$ in the next theorem is orthogonal.
\[thm:OP-ipd-abg\] Let $P_0(x,y) =1$ and, for $n =1,2,\ldots$, define $$\begin{aligned}
P_n(x,y) = & \,P_n^{(\g,\a)}(2x-1) + P_n^{(\g,\b)}(2y-1) - \binom{n+\g}{n}, \label{eq:op-P}\\
Q_n(x,y) = & \, \frac{(\g+\a+2)_{n}}{(\a+1)_{n-1}} (1-x)P_{n-1}^{(\g+2,\a)}(2x-1) \label{eq:op-Q}\\
& \qquad \qquad \qquad - \s^{-1} \frac{(\g+\b +2)_{n}}{(\b+1)_{n-1}} (1-y)P_{n-1}^{(\g+2,\b)}(2 y -1). \notag \end{aligned}$$ Then $\{P_n, Q_n\}$ are two polynomials in $\CH_n^2(w_{\a,\g},w_{\b,\g})$ and $$\label{eq:ipdPQ}
\la P_n, Q_n \ra_{\a,\b,\g} = \frac{(\b-\a) (\g+1)_{n+1}}{(2n+\g+\a+1)(2n+\g+\b+1) (n-1)!}.$$ In particular, the two polynomials are orthogonal to each other if $\b = \a$. Furthermore $$\begin{aligned}
\la P_n, P_n \ra_{\a,\b,\g} = &\, h_n^{\a,\g}+ \sigma h_n^{\b,\g} \\
\la Q_n,Q_n\ra_{\a,\b,\g} = &\,
\frac{(\g+1)_2 (\a+\g+2)_n^2}{(\a+\g+2)_2 (\a+1)_{n-1}^2} h_{n-1}^{\a,\g+2} + \s^{-1}
\frac{(\g+1)_2 (\b+\g+2)_n^2}{(\b+\g+2)_2 (\b+1)_{n-1}^2} h_{n-1}^{\b,\g+2} .\end{aligned}$$
Since $P_n^{(\g,\a)}(1) = P_n^{(\g,\b)}(1) = \binom{n+\g}{n}$, our definition shows that $$\begin{aligned}
\la P_n, Q_m \ra_{\a,\b,\g} = \frac{(\g+\a+2)_{m}}{(\a+1)_{m-1}} I_{m,n}^{\a,\g}
-\frac{(\g+\b+2)_{m}}{(\b+1)_{m-1}} I_{m,n}^{\b,\g}. \end{aligned}$$ By the identity in the previous lemma, if $n > m$, then $ \la P_n, Q_m \ra_{\a,\b,\g} =0$ since both $I_{m,n}^{\a,\g}=0$ and $I_{m,n}^{\b,\g} =0$, whereas if $n < m$, then $$\la P_n, Q_m \ra_{\a,\b,\g} = \frac{(\g+1) (\g+1)_n}{n!} -\frac{(\g+1) (\g+1)_n}{n!} =0.$$ The case $n = m$ follows from a simple calculation. Moreover, for $m \ne n$, $$\begin{aligned}
\la P_n, P_m\ra_{\a,\b,\g} = & \, c_{\a,\g} \int_0^1 P_n^{(\g,\a)}(2x-1) P_m^{(\g,\a)}(2x-1)
(1-x)^\g x^\a dx \\
& + c_{\g,\g} \int_0^1 P_n^{(\b,\g)}(2x-1) P_m^{(\g,\b)}(2x-1)(1-x)^\g x^\b dx = 0 \end{aligned}$$ by the orthogonality of the Jacobi polynomials, and it is equal to $h_n^{\g,\a} + h_n^{\g,\b}$ for $m=n$. Similarly, $$\begin{aligned}
\la Q_n, Q_m & \ra_{\a,\b,\g} = \frac{(\g+\a+2)_{m}}{(\a+1)_{m-1}} c_{\a,\g}
\int_0^1 P_{n-1}^{(\g+2,\a)}(2x-1) P_{m-1}^{(\g+2,\a)}(2x-1)(1-x)^{\g+2} x^\a dx \\
& + \s^{-1} \frac{(\g+\b+2)_{m}}{(\b+1)_{m-1}} c_{\b,\g}\int_0^1 P_{n-1}^{(\g+2,\b)}(2x-1) P_{m-1}^{(\g+2,\b)}(2x-1)
(1-x)^{\g+2} x^\b dx = 0.\end{aligned}$$ To derive the norm of $\la Q_n,Q_n\ra$, we need to use $c_{\g,\a} = (\g+1)_2/(\a+\g+2)_2 c_{\g+2,\a}$. The proof is completed.
\[cor:OP-Rn\] For $n =1,2,\ldots$, define $$\label{eq:op-R}
R_n(x,y) = Q_n(x,y) - \frac{\la P_n,Q_n\ra_{\a,\b,\g}}{h_n^{\g,\a}+ \s h_n^{\g,\b}} P_n(x,y).$$ Then, for $\a \ne \b$, $\{P_n, R_n\}$ are two polynomials in $\CH_n^2(w_{\a,\b,\g})$ and they are mutually orthogonal. Moreover, $$\la R_n, R_n\ra_{\a,\b,\g} = \la Q_n,Q_n\ra_{\a,\b,\g} - \frac{\la P_n,Q_n\ra_{\a,\b,\g} }{\la P_n,P_n\ra_{\a,\b,\g}}.$$
### Fourier orthogonal expansions
Let $L^2(\Omega, w_{\a,\g}, w_{\b,\g})$ be the space of functions defined on $\Omega$ such that $f(1,1)$ is finite and the norm $$\|f\|_{L^2(\Omega, w_{\a,\g}, w_{\b,\g})} = \left(c_{\a,\g} \int_0^1 |f(x,1)|^2 w_{\a,\g}(x) dx
+ \s c_{\b,\g} \int_0^1 |f(1,y)|^2 w_{\b,\g}(y) dy \right)^{\f12}$$ is finite for every $f$ in this space. For $ f\in L^2(\Omega, w_{\a,\g}, w_{\b,\g})$ we consider the Fourier orthogonal expansion with respect to $\la \cdot,\cdot\ra_{\a,\b,\g}$. With respect to the orthogonal basis $\{P_n,R_n\}$ in Theorem \[thm:OP-ipd-abg\] and Corollary \[cor:OP-Rn\], the Fourier orthogonal expansion is defined by $$f = {\widehat}f_0 + \sum_{n=1}^\infty \left[ {\widehat}f_{P_n}P_n + {\widehat}f_{R_n}R_n \right],$$ where $${\widehat}f_0:= \frac{ \la f,1\ra_{\a,\b,\g}}{\la 1,1\ra_{\a,\b,\g}}, \qquad
{\widehat}f_{P_n}:= \frac{\la f,P_n \ra_{\a,\b,\g}}{\la P_n,P_n \ra_{\a,\b,\g}},
\qquad {\widehat}f_{R_n}:= \frac{\la f,R_n \ra_{\a,\b,\g}}{\la R_n,R_n \ra_{\a,\b,\g}}.$$ Its $n$-th partial sum is defined by $$S_n^{\a,\b,\g} f: = {\widehat}f_0 + \sum_{k =1}^n \left[ {\widehat}f_{P_k}P_k + {\widehat}f_{R_k}R_k \right].$$ In this case, we do not have a closed form for the reproducing kernel with respect to $\la \cdot,\cdot\ra_{\a,\b,\g}$. Nevertheless, we can relate the convergence of the Fourier orthogonal expansions to that of the Fourier–Jacobi series. For $w_{\a,\g}$, we denote the partial sum defined in by $s_n^{\a,\g} f$.
For $f$ defined on $\Omega$, we define $f_1(x) = f(x,1)$ and $f_2(x) = f(1,x)$, and $$g_1(x): = \frac{f(x,1)-f(1,1)}{1-x} \quad \hbox{and}\quad g_2(y): = \frac{f(1,y)-f(1,1)}{1-y}.$$ It is easy to see that if $f(\cdot,1) \in L^2(w_{\a,\g},[0,1])$, then $g_1 \in L^2(w_{\a,\g+2}, [0,1])$, and if $f(1,\cdot) \in L^2(w_{\b,\g},[0,1])$, then $g_2 \in L^2(w_{\b,\g+2}, [0,1])$.
\[thm:3.4\] Let $\a, \b, \g > -1$. Then the Fourier orthogonal expansion converges in $f\in L^2(\Omega, w_{\a,\g}, w_{\b,\g})$. Furthermore, for $f(\cdot,1) \in L^2(w_{\a,\g})$ and $f(1,\cdot) \in L^2(w_{\b,\g})$, $$\begin{aligned}
\| f - S_n^{\a,\b,\g} f\|_{\a,\b,\g} \le &\, c \left( \|f_1 - s_n^{\a,\g} f_1 \|_{L^2(w_{\a,\g})}
+ \| f_2 - s_n^{\b,\g} f_2\|_{L^2(w_{\b,\g})} \right) \\
& + c \left( \|g_1 - s_n^{\a,\g+2} g_1 \|_{L^2(w_{\a,\g+2})}
+ \| g_2- s_n^{\b,\g+2} g_2\|_{L^2(w_{\b,\g+2})} \right),\end{aligned}$$ where $c$ is a constant that depends only on $\a,\b,\g$.
Since polynomials are dense on $\Omega$, by the Weierstrass theorem, the orthogonal basis $\{P_n,R_n\}$ is complete, so that the Fourier orthogonal expansion converges in $L^2(\Omega, w_{\a,\g}, w_{\b,\g})$. By the Parseval identity, $$\| f - S_n^{\a,\b,\g} f\|_{\a,\b,\g}^2 = \sum_{k= n+1}^\infty |{\widehat}f_{P_k}|^2 \la P_k, P_k\ra_{\a,\b,\g} +
\sum_{k= n+1}^\infty |{\widehat}f_{R_k}|^2 \la R_k, R_k \ra_{\a,\b,\g}.$$
Throughout this proof we use the convention $A \sim B$ if $c_1 B \le A \le c_2 A$, where $c_1$ and $c_2$ are constants that are independent of varying parameters in $A$ and $B$. By and the fact that $\Gamma(n+\a+1)/n! \sim n^\a$, it is easy to see that $h_n^{\a,\g} \sim n^{-1}$, so that $$\la P_n,P_n \ra_{\a,\b,\g} \sim n^{-1}, \qquad \la Q_n,Q_n \ra_{\a,\b,\g} \sim n^{2\g+3}, \qquad
\la P_n, Q_n\ra_{\a,\b,\g} \sim n^\g,$$ and, consequently, $$\la R_n R_n\ra_{\a,\b,\g} \sim n^{2 \g+3} - n^{2 \g}/ n^{-1} \sim n^{2\g+3}.$$ The Fourier–Jacobi coefficients of $f_1$ and $f_2$ are denoted by ${\widehat}{f_1}_n^{\a,\g}$ and ${\widehat}{f_2}_n^{\b,\g}$, respectively. It follows readily that ${\widehat}f_{P_n} \sim {\widehat}{f_1}_n^{\a,\g}
+ {\widehat}{f_2}_n^{\b,\g}$, consequently, $$\begin{aligned}
\sum_{k=n+1}^\infty |{\widehat}f_{P_k}|^2 \la P_k,P_k\ra_{\a,\b,\g} \le &\, c \sum_{k=n+1}^\infty
\Big( |{\widehat}{f_1}_k^{\a,\g}|^2 h_k^{\a,\g} + |{\widehat}{f_2}_k^{\b,\g}|^2 h_k^{\b,\g}\Big) \\
\le &\, c \left( \|f_1 - s_n^{\a,\g} f_1 \|_{L^2(w_{\a,\g})}
+ \| f_2 - s_n^{\b,\g} f_2\|_{L^2(w_{\b,\g})} \right). \end{aligned}$$ We now consider the estimate for $R_n$ part. By the definition of $R_n$, $$\begin{aligned}
\la f, R_n\ra_{\a,\b,\g} \sim \la f, Q_n \ra_{\a,\b,\g} - n^{\g+1} \la f, P_n \ra_{\a,\b,\g}. \end{aligned}$$ It is easy to see that $$\sum_{k=n+1}^\infty \frac{ | k^{\g+1}\la f, P_k \ra_{\a,\b,\g} |^2}{\la R_k,R_k\ra_{\a,\b,\g}} \sim \sum_{k=n+1}^\infty
k^{-1} | \la f, P_k \ra_{\a,\b,\g} |^2 \sim
\sum_{k=n+1}^\infty k^{-2} |{\widehat}f_{P_k}|^2 \la P_k,P_k\ra_{\a,\b,\g},$$ so that we only have to work with the term $\la f, Q_k \ra_{\a,\b,\g}$. The definition of $Q_k$ shows that $\la 1, Q_k\ra_{\a,\b,\g} = 0$, which leads to the identity $$\begin{aligned}
\la f, Q_k \ra_{\a,\b,\g} = &\, \frac{(\g+\a+2)_k}{(\a+k)_{k-1}} c_{\a,\g}\int_0^1 (f(x,1) - f(1,1)) Q_k(x,1) x^\a (1-x)^\g dx \\
& + \frac{(\g+\b+2)_k}{(\b+n)_{k-1}} c_{\b,\g}\int_0^1 (f(1,y) - f(1,1)) Q_k(1,y) y^\b (1-y)^\g dy\\
= &\,\frac{(\g+\a+2)_k}{(\a+k)_{k-1}} {\widehat}{g_1}_k^{\a,\g+2} h_k^{\a,\g+2} +
\frac{(\g+\b+2)_k}{(\b+n)_{k-1}} {\widehat}{g_2}_k^{\b,\g+2} h_k^{\b,\g+2}. \end{aligned}$$ Consequently, it follows that $$\begin{aligned}
\sum_{k=n+1}^\infty \frac{|\la f, Q_k \ra_{\a,\b,\g} |^2}{\la R_k,R_k\ra_{\a,\b,\g}}
& \le c \sum_{k=n+1}^\infty \left( k |{\widehat}{g_1}_k^{\a,\g+2} h_k^{\a,\g+2}|^2
+ k |{\widehat}{g_2}_k^{\b,\g+2} h_k^{\b,\g+2}|^2 \right) \\
& \le c \sum_{k=n+1}^\infty \left( |{\widehat}{g_1}_k^{\a,\g+2}|^2 h_k^{\a,\g+2}+ |{\widehat}{g_2}_k^{\b,\g+2}|^2 h_k^{\b,\g+2}\right) \\
& = c \left( \|g_1 - s_n^{\a,\g+2} g_1 \|_{L^2(w_{\a,\g+2})} + \| g_2- s_n^{\b,\g+2} g_2\|_{L^2(w_{\b,\g+2})} \right).\end{aligned}$$ The proof is completed.
For this we need the following identity, $$\begin{aligned}
& c_{\a,\g} \int_0^1 g(x) (1-x) P_n^{(\g+2,\a)}(2x-1) w_{\a,\g}(x) dx \\
& \qquad = \sum_{k=0}^{n} a_{k,n} c_{\a,\g} \int_{0}^1 g(x) P_n^{(\g,\a)}(2x-1) w_{\a,\g}(x) dx
= \sum_{k=0}^n b_{k,n} {\widehat}g_k^{\a,\g} \end{aligned}$$ where, by , $a_{k,n} \sim k^{\g+1}/n^{\g+2}$ so that $b_{k,n} \sim k^{\g}/n^{\g+2}$ for $0 \le k \le n-1$, and $a_{n,n} \sim 1$ so that $b_{n,n} \sim n^{-1}$. This shows that $$\begin{aligned}
\la f, Q_n \ra_{\a,\b,\g} = &\, \frac{(\g+\a+2)_{n}}{(\a+1)_{n-1}} \sum_{k=0}^n b_{k,n} {\widehat}{f_1}_k^{\a,\g}+
\frac{(\g+\b+2)_{n}}{(\a+1)_{n-1}} \sum_{k=0}^n b_{k,n} {\widehat}{f_2}_k^{\b,\g} \\
\sim &\, \sum_{k=0}^{n-1} k^{\g} {\widehat}{f_1}_k^{\a,\g} + n^{\g+1} {\widehat}{f_1}_n^{\a,\g} +
\sum_{k=0}^{n-1} k^{\g} {\widehat}{f_2}_k^{\b,\g} + n^{\g+1} {\widehat}{f_2}_n^{\b,\g}.\end{aligned}$$ Evidently, $$\sum_{n=1}^\infty \frac{| n^{\g+1} {\widehat}{f_2}_n^{\a,\g} |^2}{\la R_n,R_n\ra_{\a,\b,\g}} \sim \sum_{n=1}^\infty
n^{-1} |{\widehat}{f_2}_n^{\a,\g} |^2 \sim \sum_{n=1}^\infty|{\widehat}{f_2}_n^{\a,\g} |^2 h_n^{\a,\g}.$$ Moreover, using the classical Hardy inequality, $$\begin{aligned}
\sum_{n=1}^\infty \frac{1}{ \la R_n, R_n \ra_{\a,\b,\g}} \left| \sum_{k=0}^{n-1} k^{\g} {\widehat}{f_1}_k^{\a,\g} \right|^2
\le &\, c \sum_{n=1}^\infty \left| \frac{1}{n} \sum_{k=0}^{n-1} \left(\frac{k}{n}\right)^{\g+\f12} {\widehat}{f_1}_k^{\a,\g}k^{-\f12} \right|^2 \\
\le &\, 4 c \sum_{n=1}^\infty \left| {\widehat}{f_1}_n^{\a,\g}\right |^2 n^{-1}
\le c_1 \sum_{n=1}^\infty \left| {\widehat}{f_1}_n^{\a,\g}\right |^2 h_n^{\a,\g}. \end{aligned}$$ Clearly the same estimates hold if we replace $\a$ by $\b$. Putting these together gives $$\sum_{n=0}^\infty |{\widehat}f_{R_n}|^2 \la R_n,R_n\ra_{\a,\b,\g}=
\sum_{n=1}^\infty \frac{|\la f, Q_n \ra_{\a,\b,\g}|^2}{ \la R_n, R_n \ra_{\a,\b,\g}}
\le c \sum_{n=1}^\infty \left(\left| {\widehat}{f_1}_n^{\a,\g}\right |^2 h_n^{\a,\g}+
\left| {\widehat}{f_1}_n^{\b,\g}\right |^2 h_n^{\b,\g}\right).$$ The proof is completed.
Orthogonal polynomials on the boundary of the square {#Section:BoundarySquare}
====================================================
Using the results in the previous section, we can study orthogonal polynomials on a parallelogram. Since orthogonal structure is preserved under an affine transformation, we can assume without loss of generality that the parallelogram is the square $[-1,1]^2$.
For $\a,\g > -1$, let $\varpi_{\a,\g}$ be the weight function $$\varpi_{\a,\g}(x):= |x|^{2\a+1} (1-x^2)^\g.$$ We consider orthogonal polynomials of two variables on the boundary of $[-1,1]^2$ with respect to the bilinear form $$\begin{aligned}
\label{eq:int-boundary}
\la f, g \ra = & \, c_{\a,\g}\int_{-1}^1 [f(x,-1) g(x,-1) + f(x,1) g(x,1)] \varpi_{\a,\g}(x) dx \\
&+ c_{\b,\g} \int_{-1}^1 [f(-1,y) g(-1,y)+ f(1,y) g(1,y)] \varpi_{\b,\g}(y) dy \notag\end{aligned}$$ for $\a,\b,\g > -1$. Since $(1-x^2)(1-y^2)$ vanishes on the boundary of the square, the bilinear form defines an inner product modulo the ideal generated by this polynomial, or in the space $$\RR[x,y]/I : = \RR[x,y] /\la (1-x^2) (1-y^2) \ra.$$ Let $\CB\CV_n^2$ denote the space of orthogonal polynomials in $\RR[x,y]/I$ with respect to the inner product $\la \cdot,\cdot\ra$.
For $n \ge 0$, the dimension of $\CB\CV_n^2$ is given by $$\dim \CB\CV_n^2 = n+1, \quad n =0,1,2, \quad \hbox{and} \quad \dim \CB\CV_n^2 =4, \quad n \ge 3.$$
For convenience, we define, for $n \ge 3$, $$q_{n,1}(x,y) = (1-x^2) x^{n-2} \quad \hbox{and}\quad q_{n,2}(x,y) = (1-x^2) x^{n-3} y,$$ Furthermore, we define $q_{n,3}(x,y) = q_{n,2}(y,x)$ and $q_{n,4}(x,y) = q_{n,1}(y,x)$. We choose a basis of $\RR[x,y]/I$ as $$\{1,x,y, 1-x^2, x y, 1-y^2\} \cup \{q_{n,1}, q_{n,2},q_{n,3},q_{n,4}, n \ge 3\}.$$ Applying the Gram–Schmidt process on this basis will yield a basis of orthogonal polynomials in $\RR[x,y]/I$. We will determine an explicit basis.
\[prop1\] Let $n \ge 3$. With respect to $\la \cdot,\cdot\ra_{\a,\b}$,
1. $q_{n,1}$ is orthogonal to $q_{n,i}$, $i =2,3,4$ and to $y, xy$ and $1-y^2$;
2. $q_{n,2}$ is orthogonal to $q_{n,i}$, $i =1,3,4$ and to $1,x, 1-x^2, 1-y^2$.
For $i = 3$ and $4$, $q_{n,1} q_{m,i}$ contains a factor $(1-x^2)(1-y^2)$, so that $\la q_{n,1}, q_{m,i} \ra_{\a,\b} =0$ for all $m \ge 3$. The same argument shows that $q_{n,1}$ is orthogonal to $1-y^2$. Since $q_{n,1}$ depends on on $x$ and $q_{n,2}$ contains a single $y$, it is easy to see that $\la q_{n,1}, q_{n,2} \ra_{\a,\b} =0$ by the symmetry. The same argument shows that $q_{n,1}$ ir orthogonal to $y$ and $xy$. This completes the proof of the first item. The proof for the second item is similar.
Recall that the inner product $\la \cdot,\cdot\ra_{\a,\b,\g}$ studied in the previous section contains a fixed parameter $\sigma$. For fixed $\a,\b$ and $\d_1,\d_2 \in \{0,1\}$, we define $p_{m,1}^{\a+\d_1,\b+\d_2,\g}$ and $p_{m,2}^{\a+\d_1,\b+\d_2,\g}$ to be a basis of $\CH_m^2(w_{\a+\d_1,\g},w_{\b+\d_2,\g})$ for a particular choice of $\s$ defined by $$\label{eq:sigma}
\s_{\d_1,\d_2} = \frac{ c_{\b,\g} c_{\a+\d_1,\g}}{c_{\a,\g} c_{\b+\d_2,\g}}.$$ For example, $p_{m,i}^{\a,\b,\g}$ are defined with $\s_{0,0} =1$ and $p_{m,i}^{\a+1,\b,\g}$ are defined with $\s_{1,1} = (\a+\g+2)/(\a+1)$. For each pair of $\a+\d_1, \, \b+\d_2$, we can choose, for example, $p_{m,1}^{\a+\d_1,\b+\d_2,\g} = P_m$ defined in and take $p_{m,2}^{\a+\d_1,\b+\d_2,\g}
= Q_m$ defined in or $p_{m,2}^{\a+1,\b+1,\g} = R_m$ defined in .
\[thm:boundaryOP\] For $n = 0, 1,2,$ a basis for $\CB\CV_n$ is denoted by $Y_{n,i}$ and given by $$\begin{aligned}
& Y_{0,1}(x,y) = 1, \quad Y_{1,1}(x,y) = x \quad Y_{1,2}(x,y) = y, \\
& Y_{2,1}(x,y) = p_{1,1}^{\a,\b,\g}(x^2,y^2),\quad Y_{2,2}(x,y) = xy, \quad Y_{2,3}(x,y) = p_{1,2}^{\a,\b,\g}(x^2,y^2).\end{aligned}$$ For $n \ge 3$, the four polynomials in $\CB\CV_n^2$ that are linearly independent modulo the ideal can be given by $$\begin{aligned}
Y_{2m,1}(x,y) & = p_{m,1}^{\a,\b,\g}(x^2,y^2),\\
Y_{2m,2}(x,y) & = p_{m,2}^{\a,\b,\g}(x^2,y^2),\\
Y_{2m,3}(x,y) & = x y \,p_{m-1,1}^{\a+1,\b+1,\g}(x^2,y^2),\\
Y_{2m,4}(x,y) & = x y \, p_{m-1,2}^{\a+1,\b+1,\g}(x^2,y^2)\end{aligned}$$ for $n =2m \ge 2$, and $$\begin{aligned}
Y_{2m+1,1}(x,y) & = x \,p_{m,1}^{\a+1,\b,\g}(x^2,y^2),\\
Y_{2m+1,2}(x,y) & = x \,p_{m,2}^{\a+1,\b,\g}(x^2,y^2),\\
Y_{2m+1,3}(x,y) & = y \,p_{m,1}^{\a,\b+1,\g}(x^2,y^2), \\
Y_{2m+1,4}(x,y) & = y \,p_{m,2}^{\a,\b+1,\g}(x^2,y^2)\end{aligned}$$ for $n=2m+1 \ge 3$. In particular, these bases satisfy the equation $\partial_x^2 \partial_y^2 u = 0$.
The proof relies on the parity of the integrals. For example, it is easy to see that $\la x f(x^2,y^2), g(x^2,y^2) \ra =0$ and $\la y f(x^2,y^2), g(x^2,y^2) \ra =0$ for any polynomials $f$ and $g$, which implies, in particular, that $\la Y_{2m,i}, Y_{2n+1,j}\ra =0$ for $i,j = 1,2,3,4$. Furthermore, it is easy to see that $\la x y f(x^2,y^2), g(x^2,y^2)\ra = 0$ for any polynomials $f$ and $g$. Hence, $\la Y_{2m,i},Y_{2k,j}\ra =0$ for $i =1,2$ and $j=3,4$. Furthermore, using the relation $$\label{eq:4.3}
\int_{-1}^1 f(x^2) |x|^{2\a+1} (1-x^2)^\g dx = \int_0^1 f(x) x^\a (1-x)^\g dx,$$ it is easy to see that $$\begin{aligned}
\la Y_{2m,i}, Y_{2 k,j} \ra & = \la p_{m,i}^{\a, \b,\g}, p_{k,j}^{\a,\b,\g} \ra_{\a,\b,\g}, \quad i,j = 1, 2 \\
\la Y_{2m,i}, Y_{2 k,j} \ra & = \frac{c_{\a,\g}}{c_{\a+1,\g}} \la p_{m,i}^{\a+1, \b+1,\g},
p_{k,j}^{\a+1, \b+1,\g} \ra_{{\a+1, \b+1,\g}}, \quad i,j = 3,4,\end{aligned}$$ where in the second identity, we have adjusted the normalization constants of integrals from $c_{\a,\g}$ and $c_{\b,\g}$ to $c_{\a+1,\g}$ and $c_{\b+1,\g}$, respectively, and used our choice of $\s_{1,1}$. Hence, with our choice of $\s_{0,0}$ and $\s_{1,1}$, we see that $Y_{2m,i}$ is orthogonal to $Y_{2k,j}$ for $i,j =1,2$ and $i,j = 3,4$, respectively. Similarly, by the same consideration, we obtain that $$\begin{aligned}
\la Y_{2m+1,i}, Y_{2 k+1,j} \ra &\, = \frac{c_{\a,\g}}{c_{\a+1,\g}}
\la p_{m,i}^{\a+1,\b,\g}, p_{k,j}^{\a+1,\b,\g} \ra_{\a+1, \b,\g}, \quad i,j = 1, 2 \\
\la Y_{2m+1,i}, Y_{2 k+1,j} \ra & \, = \la p_{m,i}^{\a,\b+1,\g}, p_{k,j}^{\a,\b+1,\g} \ra_{\a,\b+1,\g}, \quad i,j = 3,4,\end{aligned}$$ which shows, with our choice of $\s_{0,1}$ and $\s_{1,0}$, that $Y_{2m+1,i}$ is orthogonal to $Y_{2k+1,j}$ for $i,j =1,2$ and $i,j = 3,4$, respectively. Finally, since $\partial_x \partial_y p_{n,i}^{\a,\b}(x,y) =0$, we see that $Y_{n,j} = \xi(x,y) u(x) + \eta(x,y) v(x)$, where $\xi$ and $\eta$ are linear polynomial of $x,y$, so that it is evident that $\partial_x^2 \partial_y^2 Y_{n,j}(x,y)=0$.
In our notation, the case $\a = -\frac12$ $\b= - \f12$ and $\g =0$ corresponds to the inner product in which the integrals are unweighted.
Let $L^2([-1,1]^2, \varpi_{\a,\g}, \varpi_{\b,\g})$ be the space of functions defined on the boundary of $[-1,1]^2$ such that $f(\pm 1, \pm 1)$ are finite and the norm $$\begin{aligned}
\|f\|_{L^2(\varpi_{\a,\g}, \varpi_{\b,\g})} = &
\left( c_{\a,\g} \int_{-1}^1 \left( |f(x,1)|^2+|f(x,-1)|^2 \right) \varpi_{\a,\g}(x)dx \right. \\
& \left. + c_{\b,\g} \int_{-1}^1 \left( |f(1,y)|^2+|f(-1,y)|^2 \right) \varpi_{\b,\g}(y)dy \right)^{\f12}.\end{aligned}$$ is finite for every $f$. For $f \in L^2([-1,1]^2, \varpi_{\a,\g}, \varpi_{\b,\g})$, its Fourier orthogonal expansion is defined by $$f = \sum_{n=0}^2 \sum_{i=1}^{n+1} {\widehat}f_{n,i} Y_{n,i}^{\a,\b,\g} + \sum_{n=3}^\infty
\sum_{i=1}^{4} {\widehat}f_{n,i} Y_{n,i}^{\a,\b,\g},
\qquad {\widehat}f_{n,i} = \frac{\la f, Y_{n,i}^{\a,\b,\g}\ra}{\la Y_{n,i}, Y_{n,i}^{\a,\b,\g}\ra}.$$ For $n \ge 2$, let $S_n (f)$ denotes its $n$-th partial sum defined by $$S_n f = \sum_{k=0}^2 \sum_{i=1}^{k+1} {\widehat}f_{k,i} Y_{k,i}^{\a,\b,\g} + \sum_{k=3}^n
\sum_{i=1}^{4} {\widehat}f_{k,i} Y_{k,i}^{\a,\b,\g}.$$
For fixed $\a,\b,\g$, let $\la \cdot,\cdot\ra_{\a+\delta_1,\b+\delta_2,\g}$ be the inner product defined in the previous section with $\s = \s^{\a,\b,\g}$. For $f$ defined on $[-1,1]^2$, we define four functions $$\begin{aligned}
F_{e,e}(x,y) &\, = \tfrac14\left[ f(x,y) + f(-x,y)+ f(x,-y) + f(-x,-y) \right], \\
F_{e,o}(x,y) &\, = \tfrac14\left[ f(x,y) + f(-x,y) - f(x,-y) - f(-x,-y) \right], \\
F_{o,e}(x,y) &\, = \tfrac14\left[ f(x,y) - f(-x,y) + f(x,-y) - f(-x,-y) \right], \\
F_{o,o}(x,y) &\, = \tfrac14\left[ f(x,y) - f(-x,y) - f(x,-y) + f(-x,-y) \right], \end{aligned}$$ where the subindices indicate the parity of the function. For example, $F_{e,o}$ is even in $x$ variable and odd in $y$ variable. By definition, $$f(x,y) = F_{e,e}(x,y) + F_{e,o}(x,y) + F_{o,e}(x,y) +F_{o,o}(x,y).$$ We further define $$\begin{aligned}
G_{0,0}(x,y) &\, = F_{e,e}(x,y), \quad G_{0,1}(x,y) = y^{-1} F_{e,o}(x,y) , \\
G_{1,0}(x,y) & \, = x^{-1} F_{o,e}(x,y), \quad G_{1,1}(x,y) = x^{-1} y^{-1} F_{o,o}(x,y)\end{aligned}$$ and define $\psi: \RR^2 \mapsto \RR^2$ by $\psi: (x,y) \mapsto (\sqrt{x},\sqrt{y})$. Changing variables in integrals as in , we see that if $f\in L^2([-1,1]^2, \varpi_{\a,\g}, \varpi_{\b,\g})$, then $G_{\delta_1,\delta_2}\circ \psi \in L^2(\CB, w_{\a+\delta_1,\g}, w_{\b+\delta_2,\g})$ for $\delta_i \in \{0,1\}$.
\[thm:squareconv\] For $f\in L^2([-1,1]^2, \varpi_{\a,\g}, \varpi_{\b,\g})$, $$\begin{aligned}
S_{2 m} f (x,y) = &\, S_m^{\a,\b,\g} G_{0,0}\circ \psi (x^2,y^2)+ y S_{m-1}^{\a,\b+1,\g} G_{0,1} \circ \psi (x^2,y^2) \\
&\, + x S_{m-1}^{\a+1,\b,\g} G_{1,0} \circ \psi(x^2,y^2)
+ x y S_{m-1 }^{\a+1,\b+1,\g} G_{1,1} \circ \psi(x^2,y^2), \end{aligned}$$ $$\begin{aligned}
S_{2 m+1} f (x,y) = &\, S_m^{\a,\b,\g} G_{0,0}\circ \psi (x^2,y^2)+ y S_{m}^{\a,\b+1,\g} G_{0,1} \circ \psi (x^2,y^2) \\
&\, + x S_{m}^{\a+1,\b,\g} G_{1,0} \circ \psi(x^2,y^2)
+ x y S_{m-1}^{\a+1,\b+1,\g} G_{1,1} \circ \psi(x^2,y^2). \end{aligned}$$ In particular, the norm of $S_n f - f$ is bounded by those of $S_{m}^{\a+\d_1,\b+\d_2,\g} G_{\d_1,\d_2} - G_{\d_1,\d_2}$ as in Theorem \[thm:3.4\].
Using the parity of the function, it is easy to see that $$\frac{\la f, Y_{2m,i}\ra}{\la Y_{2m,i}, Y_{2m,i}\ra} = \frac{\la F_{e,e}, Y_{2m,i} \ra}{\la Y_{2m,i}, Y_{2m,i}\ra} =
\frac{\la G_{0,0} \circ \psi, p_{2m,i}^{\a,\b,\g}\ra_{\a,\b,\g}}{\la p_{2m,i}^{\a,\b,\g}, p_{2m,i}^{\a,\b,\g}\ra_{\a,\b,\g}}, \qquad i = 1,2,$$ where we have used the fact that $F_{e,e}$ is even in both variables and use the change of variables in integrals as in . The similar procedure can be used in the other three cases, as $G_i(x,y)$ is even in both variables, and the result is $$\begin{aligned}
\frac{\la f, Y_{2m,i}\ra}{\la Y_{2m,i}, Y_{2m,i}\ra} &\, = \frac{\la F_{o,o}, Y_{2m,i} \ra}{\la Y_{2m,i}, Y_{2m,i}\ra} =
\frac{\la G_{1,1} \circ \psi, p_{m,i}^{\a+1,\b+1,\g}\ra_{\a+1,\b+1,\g}}{\la p_{2m,i}^{\a+1,\b+1,\g}, p_{2m,i}^{\a+1,\b+1,\g}
\ra_{\a+1,\b+1,\g}}, \quad i = 3,4, \\
\frac{\la f, Y_{2m+1,i}\ra}{\la Y_{2m+1,i}, Y_{2m+1,i}\ra} &\, = \frac{\la F_{e,o}, Y_{2m+1,i} \ra}{\la Y_{2m+1,i}, Y_{2m+1,i}\ra} =
\frac{\la G_{0,1} \circ \psi, p_{m,i}^{\a,\b+1,\g}\ra_{\a,\b+1,\g}}{\la p_{2m,i}^{\a,\b+1,\g}, p_{2m,i}^{\a,\b+1,\g}
\ra_{\a,\b+1,\g}}, \quad i = 1,2, \\
\frac{\la f, Y_{2m+1,i}\ra}{\la Y_{2m+1,i}, Y_{2m+1,i}\ra} &\, = \frac{\la F_{o,e}, Y_{2m+1,i} \ra}{\la Y_{2m+1,i}, Y_{2m+1,i}\ra} =
\frac{\la G_{1,0} \circ \psi, p_{m,i}^{\a+1,\b,\g}\ra_{\a+1,\b,\g}}{\la p_{2m,i}^{\a+1,\b,\g}, p_{2m,i}^{\a+1,\b,\g}
\ra_{\a+1,\b,\g}}, \quad i = 3,4. \end{aligned}$$ Since $S_n^{\a+\d_1, \b+ \d_2, \g} G_{\d_1,\d_2} \circ \psi(x^2, y^2) \to G_{\d_1,\d_2} (x,y)$ and $$f(x,y) = G_{0,0}(x,y) + y G_{0,1}(x,y) + y G_{1,0}(x,y) + xy G_{1,1}(x,y),$$ the last statement is evident.
Orthogonal system on the square {#Section:Square}
===============================
Let $w$ be a nonnegative weight function defined on $[0,1]$. Define $$W(x,y) = w(\max \{|x|,|y|\}), \qquad (x,y) \in [-1,1]^2.$$ We construct a system of orthogonal functions with respect to the inner product $$\la f,\g\ra_W = \int_{-1}^1 \int_{-1}^1 f(x,y) g(x,y) W(x,y) dx dy.$$ by making use of the orthogonal polynomials on the boundary or the square, studied in the previous section. Our starting point is the following integral identity derived from changing variables $(x,y)
\mapsto (s \xi,s\eta)$, $$\label{eq:int-square}
\int_{-1}^1 \int_{-1}^1 f(x,y) w(\max \{|x|,|y|\})dxdy = \int_0^s s \int_\CB f(s \xi, s \eta) d\s(\xi,\eta) w(s) ds,$$ where $\int_\CB d \sigma$ denotes the integral on the boundary of the square, $$\int_\CB f(\xi,\eta)d\s(\xi,\eta) = \int_{-1}^1 \left[f(\xi,1)+ f(\xi,-1)\right] d\xi + \int_{-1}^1 \left[f(1 \eta)+ f(-1, \eta)\right] d\eta.$$
Our orthogonal functions are similar in structure to orthogonal polynomials on the unit disk that are constructed using spherical harmonics. However, these function are polynomials in $(s, \xi,\eta)$ for the $(x,y) = (s \xi, s\eta)
\in [-1,1]^2$, but not polynomials in $(x,y)$.
Let $\CB\CV_n^2$ be the space of orthogonal polynomials on the boundary of $[-1,1]^2$ with respect to the inner product $$\la f,g\ra = \int_\CB f(\xi,\eta) g(\xi,\eta) d\s(\xi,\eta) ,$$ which is the inner product with $\a = -\f12$, $\b = - \f12$ and $\g =0$ studied in the previous section. Let $Y_{n,i}$ be an orthogonal basis for $\CB\CV_n^2$. For $n \le 2$, they are defined by, see Theorem \[thm:boundaryOP\], $$\begin{aligned}
Y_{0,1}(x,y) & =1, \quad Y_{1,1}(x,y) = x, \quad Y_{1,2}(x,y) = y; \\
Y_{2,1}(x,y) & = x^2 - \frac{2}{3}, \quad Y_{2,2}(x,y) = xy,\quad
Y_{2,3}^2(x,y) = y^2 - \frac{2}{3},\end{aligned}$$ whereas for $n \ge 3$, they are constructed in Theorem \[thm:boundaryOP\]. For $n \ge k $, denote by $P_{m, 2n-2k}$ the orthogonal polynomial of degree $m$ with respect to $t^{2n-2k+1} w(t)$ on $[0,1]$ and with $P_{0,2n-2k}(s):=1$. For $n \ge 0$ and $0 \le k \le n$, we define $$Q_{k,i}^n (x,y):= P_{k, 2n-2k}(s) s^{n-k} Y_{n-k,i}\left(\frac{\xi}{s}, \frac{\eta}{s} \right),$$ where $i = 1,\ldots, \min\{n+1,4\}$.
In the coordinates $(x,y) = s(\xi,\eta)$, the system of functions $$\{Q_{k,i}^n: i = 1,\ldots, \min\{n+1,4\}, \,\, 0\le k\le n, \,\, n\ge 0\}$$ is a complete orthogonal basis for $L^2(W; [-1,1]^2)$.
Changing variables $x = s \xi$ and $y= \eta$ shows $$\begin{aligned}
\la Q_{k,i}^n, Q_{l,j}^m \ra_{W} = & \int_0^1 P_{k, 2n-2k}(s)P_{l, 2m-2l}(s) s^{n-k+m-l+1} w(s)ds \\
& \times \int_{\CB} Y_{n- k,i}(\xi,\eta) Y_{m- l,j}(\xi,\eta) d\s(\xi,\eta).\end{aligned}$$ The second integral is zero if $i \ne j$ and $n-k \ne m-l$, whereas the second integral is zero when $n-k = m-l$ and $k \ne l$, so that $\la Q_{k,i}^n, Q_{l,j}^m \ra_{W} =0$ if $ i\ne j$, $k\ne l$ and $n \ne m$. By definition, $s^{n-k} Y_{n-k,i}\left(\frac{\xi}{s}, \frac{\eta}{s} \right)$ is a polynomial of degree $n-k$ in the variable $s$, so that $Q_{k,i}^n$ is a polynomial of degree $n$. To show that the system is complete, we show that if $\la f, Q_{k,i}^n\ra =0$ for all $k,i,n$, then $f(x,y)=0$. Indeed, by the orthogonality of polynomials on the boundary, we see that $$\begin{aligned}
f (x,y) = f(s \xi,s \eta) & = \sum_{k=0}^n s^k \sum_{j=0}^k a_{j,k} \xi^j \eta^{k-j} \\
& = \sum_{k=0}^n s^k \sum_{m=0}^k \sum_{i=1}^{\min\{m+1,4\}} b_{m,i}^k Y_{m,i}(\xi,\eta)\end{aligned}$$ modulo the ideal. Changing order of summation shows that $$\begin{aligned}
f(x,y) = & \sum_{m=0}^n \sum_{i=1}^{\min\{m+1,4\}} \left(\sum_{k=0}^{n-m} b_{m,i}^{k+m} s^k \right) s^m Y_{m,i}(\xi,\eta)\\
& = \sum_{m=0}^n \sum_{i=1}^{\min\{m+1,4\}} \left(\sum_{k=0}^{n-m} c_{m,i,k} P_{k,2m}(x) \right) s^m Y_{m,i}(\xi,\eta).\end{aligned}$$ This completes the proof.
Sampling the associated determinantal point process {#sec:detpointprocess}
===================================================
Associated with an orthonormal basis $q_0(x),\ldots,q_N(x)$ is a determinantal point process, which describes $N$ points $\lambda_1,\ldots,\lambda_N$ distributed according to $$\det \begin{pmatrix} K_N(\lambda_1,\lambda_1) & \cdots & K_N(\lambda_1, \lambda_N) \cr
\vdots & \ddots & \vdots \cr
K_N(\lambda_N,\lambda_1) & \cdots & K_N(\lambda_N,\lambda_N)
\end{pmatrix}$$ where $$K_N(x,y) = \sum_{k=0}^N q_k(x)q_k(y)$$ is the reproducing kernel, see [@RMTHandbookDetProcess] for an overview of determinantal point processes.
In the particular case of univariate orthogonal polynomials with respect to a weight $w(x)$, the associated determinantal process is equivalent to a Coulomb gas—that is, the points are distributed according to $${1 \over Z_N} \prod_{k=1}^N w(x_k) \prod_{k < j} |\lambda_k - \lambda_j|^2$$ where $Z_N$ is the normalization constant—as well as the eigenvalues of unitary ensembles, see for example [@DeiftOrthogonalPolynomials] for the case of an analytic weight on the real line or [@JUEUniversality] for the case of a weight supported on $[-1,1]$ with Jacobi-like singularities.
In the case of our orthogonal polynomials on the wedge, the connection with Coulomb gases and random matrix theory is no longer obvious: the interaction of the points is not Coulomb (that is, it can not be reduced to a Vandermonde determinant squared times a product of weights), nor is there an obvious distribution of random matrices whose eigenvalues are associated with the points[^3]. We note that there are recent universality results due to Kroó and Lubinsky on the asymptotics of Christoffel functions associated with multivariate orthogonal polynomials [@UniversalityBall; @UniversalityMultivariateOPs], but they do not apply in our setting.
![Monte Carlo calculation of the probability that no point satisfying $y = 1$ sampled according to the determinantal point process associated to the wedge orthogonal polynomials with $\alpha = \beta = \gamma = 0$ lies in a neighbourhood of four different points. $N$ is the total number of basis elements and points. We have scaled the statistics so that the variance is one, and have used 10,000 samples. []{data-label="fig:detprocess"}](Figs/wedgedetprocess10.pdf "fig:"){width=".7\textwidth"} ![Monte Carlo calculation of the probability that no point satisfying $y = 1$ sampled according to the determinantal point process associated to the wedge orthogonal polynomials with $\alpha = \beta = \gamma = 0$ lies in a neighbourhood of four different points. $N$ is the total number of basis elements and points. We have scaled the statistics so that the variance is one, and have used 10,000 samples. []{data-label="fig:detprocess"}](Figs/wedgedetprocess20.pdf "fig:"){width=".7\textwidth"} ![Monte Carlo calculation of the probability that no point satisfying $y = 1$ sampled according to the determinantal point process associated to the wedge orthogonal polynomials with $\alpha = \beta = \gamma = 0$ lies in a neighbourhood of four different points. $N$ is the total number of basis elements and points. We have scaled the statistics so that the variance is one, and have used 10,000 samples. []{data-label="fig:detprocess"}](Figs/wedgedetprocess30.pdf "fig:"){width=".7\textwidth"}
![Monte Carlo calculation of the probability that no point satisfying $y = 1$ sampled according to the Coulomb gas on the wedge lies in a neighbourhood of four different points. $N$ is the total number of basis elements and points. We have scaled the statistics so that the variance is one, and have used 10,000 samples. []{data-label="fig:coulomb"}](Figs/coulomb10.pdf "fig:"){width=".7\textwidth"} ![Monte Carlo calculation of the probability that no point satisfying $y = 1$ sampled according to the Coulomb gas on the wedge lies in a neighbourhood of four different points. $N$ is the total number of basis elements and points. We have scaled the statistics so that the variance is one, and have used 10,000 samples. []{data-label="fig:coulomb"}](Figs/coulomb20.pdf "fig:"){width=".7\textwidth"} ![Monte Carlo calculation of the probability that no point satisfying $y = 1$ sampled according to the Coulomb gas on the wedge lies in a neighbourhood of four different points. $N$ is the total number of basis elements and points. We have scaled the statistics so that the variance is one, and have used 10,000 samples. []{data-label="fig:coulomb"}](Figs/coulomb30.pdf "fig:"){width=".7\textwidth"}
Using the algorithm for sampling determinantal point processes associated with univariate orthogonal polynomials [@InvariantEnsembleSampling], which is trivially adapted to the orthogonal polynomials on the wedge, we can sample from this determinantal point process. We use this algorithm to calculate statistics of the points. In Figure \[fig:detprocess\], we use the sampling algorithm in a Monte Carlo simulation to approximate the probability that no eigenvalue is present in a neighbourhood of three points for $\alpha = \beta = \gamma = 0$. That is, we take 10,000 samples of a determinantal point process, and calculate the distance of the nearest point to $z_0$, for $z_0$ equal to $(1,1)$, $(0,1)$, $(0.5,1)$ and $(0.7,1)$. The plots are of a complementary empirical cumulative distribution function of these samples. This gives an estimation of the probability that no eigenvalue is in a neighbourhood of $z_0$. We have scaled the distributions so that the empirical variance is one: this ensures that the distributions tend to a limit as $N$ becomes large, which is the regime where universality is present.
![Comparison of the gap probability of the determinantal point process associated to the wedge orthogonal polynomials and Coulomb gas near $(0.5,1)$ for $N = 101$ points. We have scaled the statistics so that the variance is one, and have used 10,000 samples. []{data-label="fig:opvcoulomb"}](Figs/opvcoulomb.pdf){width=".7\textwidth"}
In Figure \[fig:coulomb\] we plot the same statistics but for samples from the unweighted Coulomb gas on the wedge, which has the distribution $${1 \over Z_N} \prod_{k < j} \|\lambda_k - \lambda_j\|^2$$ for $\lambda_k$ supported on the wedge. As this is a Vandermonde determinant squared, it is also a determinantal point process with the basis arising from orthogonalized complex-valued polynomials $1, (x + i y), (x+ i y)^2, \ldots$ [@NormalMatrixModel]. We approximate this orthogonal basis using the modified Gram–Schmidt algorithm with the wedge inner product calculated via Clenshaw–Curtis quadrature. Again, this fits naturally into the sampling algorithm of [@InvariantEnsembleSampling], hence we can produce samples of this point process. What we observe is that, while our determinantal point process is not a Coulomb gas, it appears to be in the same universality class as the Coulomb gas away from the edge and corner, as the statistics follow the same distribution. This universality class matches that of the Gaussian Unitary Ensemble, as seen in Figure \[fig:opvcoulomb\] where we compare the three for $N = 50$.
Conclusion
==========
We have introduced multivariate orthogonal polynomials on the wedge and boundary of a square for some natural choices of weights. We have also generated a complete orthogonal basis with respect to a suitable weight inside the square. We have looked at determinantal point process statistics and observed a relationship between the resulting statistics and Coulomb gases, suggesting that, away from the corner and edge, they are in the same universality class.
One of the motivations for this work is to solve singular integral equations and evaluate their solutions on contours that have corners, in other words, to generalized the approach of [@RMSSOSIEs]. Preliminary work in this direction is included in Appendix \[sec:Stieltjes\], which shows how the recurrence relationship that our polynomials satisfy can be used to evaluate Stieltjes transforms.
Jacobi operators {#sec:JacobiOperators}
================
By necessity, multivariate orthogonal polynomials have block-tridiagonal Jacobi operators corresponding to multiplication by $x$ and $y$. We include here the recurrences associated with the inner product $ \la f,g\ra_{\a,\a,\g}$ (that is, $\b = \a$) that encode the Jacobi operators as they have a particularly simple form. The following lemma gives a linear combination of our orthogonal polynomials that vanish on $x=1$:
\[prop:vanish\] For $\beta = \alpha$, we have $$(\alpha+\gamma+2) Q_1(x,y)- P_1(x,y) +(1+\alpha) P_0(x,y) = 2 (\alpha+\gamma+2)(1-x)$$ and for $n=1,2,\ldots,$ $$\begin{aligned}
(n + \gamma + \alpha + 2) & Q_{n + 1}(x,y) - (n + 1) P_{n + 1}(x,y) - (n + \gamma) Q_n(x, y) + (n + a + 1) P_n( x, y) \cr
&=
2 (1 - x) (2 n + \gamma + \alpha + 2) P_n^{(\gamma + 1, \alpha)}(2 x - 1)\end{aligned}$$
Assume $(1-x)(1-y) = 0$. Then $$\begin{aligned}
(1-x) P_0(x,y) & = {1 \over 2} Q_1(x,y) - {1 \over 2 (2 + \gamma+\alpha) } P_1(x,y) + {1+\gamma \over 2 (2+\gamma+\alpha)} P_0(x,y) \cr
(1-x)P_1(x,y) &= {\gamma + \alpha + 2 \over 2 (4 + \gamma + \alpha)} Q_2( x, y)
-
{\gamma + \alpha + 2 \over (3 + \gamma + \alpha) (4 + \gamma + \alpha)} P_2( x, y) \cr
&- {1+\alpha \over
4 + \gamma + \alpha} Q_1( x, y)
+ {4 + 3 \alpha + \gamma (3 + \gamma + \alpha)\over2 (2 + \gamma + \alpha) (4 + \gamma + \alpha)}
P_1( x, y) \cr
& -{(1 + \gamma) (1 + \alpha) \over 2 (2 + \gamma + \alpha) (3 + \gamma + \alpha)}
P_0( x, y) \cr
(1-x) Q_1(x,y) = & -{1 \over 2 (4 + \gamma + \alpha)} Q_2( x, y)
+
{1\over (3 + \gamma + \alpha) (4 + \gamma + \alpha)} P_2( x, y) \cr
&+ {(3 + \gamma) \over
2 (4 + \gamma + \alpha)} Q_1( x, y)
- {(2 + \gamma) \over (2 + \gamma + \alpha) (4 + \gamma + \alpha)}
P_1( x, y) \cr
& + {(1 + \gamma) (2 + \gamma) \over 2 (2 + \gamma + \alpha) (3 + \gamma + \alpha)}
P_0( x, y)\end{aligned}$$ and, for $n=2,3,\ldots$, $$\begin{aligned}
(1 - x) P_n(x,y) = & {(1 + \gamma + \alpha + n) (n + \gamma + \alpha + 2) \over
2 (1 + \gamma + \alpha + 2 n) (2 + \gamma + \alpha + 2 n)}
Q_{n+1}(x,y) \cr
&- {(1 + \gamma + \alpha + n) (n + 1)\over
2 (1 + \gamma + \alpha + 2 n) (2 + \gamma + \alpha + 2 n)}
P_{n+1}(x,y) \cr
& - { (\alpha + n) (1 + \gamma + \alpha + n) (1 + \gamma + \alpha + 2 n) \over (1 + \gamma + \alpha +
2 n) (2 + \gamma + \alpha + 2 n) (\gamma + \alpha + 2 n)}
Q_n(x,y) \cr
& + {(1 + \gamma) (\gamma + \alpha) + 2 (1 + \gamma + \alpha) n + 2 n^2 \over
2 (2 + \gamma + \alpha + 2 n) (\gamma + \alpha + 2 n)}
P_n(x,y) \cr
&+ {(n + \alpha) (n + \alpha - 1) \over
2 (1 + \gamma + \alpha + 2 n) (2 n + \gamma + \alpha)}
Q_{n-1}(x,y) \cr
&- {(n + \alpha) (n + \gamma) \over
2 (1 + \gamma + \alpha + 2 n) (2 n + \gamma + \alpha)} P_{n-1}(x,y) \cr
(1 - x) Q_n(x,y) = &- {n (2 + \gamma + \alpha + n) \over 2 (2 + \gamma + \alpha + 2 n) (2 n + \gamma + \alpha + 1)}
Q_{n+1}(x,y) \cr
&+ {n (1 + n) \over 2 (2 + \gamma + \alpha + 2 n) (2 n + \gamma + \alpha + 1)}
P_{n+1}(x,y) \cr
& + {(1 + \gamma) (2 + \gamma + \alpha) + 2 (1 + \gamma + \alpha) n +
2 n^2 \over 2 (\gamma + \alpha + 2 n) (2 + \gamma + \alpha + 2 n)}
Q_n(x,y) \cr
& - {n (1 + \gamma + n) \over (\gamma + \alpha + 2 n) (2 + \gamma + \alpha + 2 n)}
P_n(x,y) \cr
&- {(1 + \gamma + n) (\alpha + n - 1)\over 2 (\gamma + \alpha + 2 n) (2 n + \gamma + \alpha + 1)}
Q_{n-1}(x,y) \cr
& {(1 + \gamma + n) (\gamma + n)\over 2 (\gamma + \alpha + 2 n) (2 n + \gamma + \alpha + 1)} P_{n-1}(x,y)
\end{aligned}$$
The first equation follows from Proposition \[prop:vanish\], since, for $y = 1$, we have (using [@DLMF (18.9.5)] to increment the first parameter) $$\begin{aligned}
(2n+\gamma + \alpha + 1)(1 - x) P_n(x,y) = &(n + \gamma + \alpha + 1) (1 - x) P_n^{(\gamma + 1, \alpha)}( 2 x - 1) \cr
&- (n + \alpha) (1-x)P_{n - 1}^{( \gamma+ 1, \alpha)}( 2 x - 1)
\end{aligned}$$
The second equation also follows from Proposition \[prop:vanish\], since, for $y = 1$, we have (using [@DLMF (18.9.6)] to decrement the first parameter) $$\begin{aligned}
(2n+\gamma + \alpha + 1)(1 - x) Q_n(x,y) = & -n (1 - x) P_n^{(\gamma + 1, \alpha)}( 2 x - 1) \cr
&+ (n + \gamma + 1) (1-x)P_{n - 1}^{( \gamma+ 1, \alpha)}( 2 x - 1).
\end{aligned}$$
The recurrences for multiplication by $1-y$ follow from the symmetries $P_n(x,y) = P_n(y,x)$ and $Q_n(x,y) = -Q_n(y,x)$.
Stieltjes transform of orthogonal polynomials {#sec:Stieltjes}
=============================================
Consider the Stieltjes transform $${\cal S}_\Omega f(z) = \int_\Omega {f(x,y) \over z - (x + i y)} \ds,$$ where $\ds$ is the arc-length differential. Just as in one-dimensions, the Stieltjes transform of weighted multivariate orthogonal polynomials satisfies the same recurrence as the orthogonal polynomials themselves
Suppose ${\mathbb P}_n$ are a family of orthogonal polynomials with respect to $w(x,y)$. Then, for $n = 1, 2, \ldots$, $$z {\cal S}_\Omega[{\mathbb P}_n w](z) = {\cal S}_\Omega[\zeta {\mathbb P}_n w](z)$$ In particular, if ${\mathbb P}_n$ satisfies the recurrence relationships $$\begin{aligned}
x {\mathbb P}_n &= C_{n}^x {\mathbb P}_{n-1} + A_{n}^x {\mathbb P}_n + B_{n}^x {\mathbb P}_{n+1} \\
y {\mathbb P}_n &= C_{n}^y {\mathbb P}_{n-1} + A_{n}^y {\mathbb P}_n + B_{n}^y {\mathbb P}_{n+1}
\end{aligned}$$ then for $A_n^z = A_{n}^x +\I A_{n}^y$, $B_n^z = B_{n}^x +\I B_{n}^y$ and $C_n^z = C_{n}^x +\I C_{n}^y$ we have $$\begin{aligned}
z {\cal S}_\Omega[{\mathbb P}_n w](z)&= C_n^z {\cal S}_\Omega[{\mathbb P}_{n-1} w](z)+A_n^z {\cal S}_\Omega[{\mathbb P}_n w](z) +B_{n}^z {\cal S}_\Omega[{\mathbb P}_{n+1} w](z)
\end{aligned}$$
We will identify ${\mathbb R}^2$ and ${\mathbb C}$ and use the notation $\zeta = x + i y$. Note that $$z \int_\Omega { f(\zeta) \over z-\zeta} \ds =
\int_\Omega {(z - \zeta) f(\zeta) \over z - \zeta} \ds + \int_\Omega {\zeta f(\zeta ) \over z - \zeta} \ds =
\int_\Omega f(\zeta ) \ds+ {\cal S}_\Omega[\zeta f](z)$$ The first integral is zero if $f$ is orthogonal to $1$.
While this holds true for all families of multivariate orthogonal polynomials, in general, satisfying a single recurrence is not sufficient to determine ${\cal S}_\Omega[{\mathbb P}_n w](z)$. However, since our blocks are square, in our case it is:
If $B_n^z = B_{n}^x +\I B_{n}^y$ is invertible, then $$\begin{aligned}
{\cal S}_\Omega[{\mathbb P}_{n+1} w](z)&= z (B_{n}^z)^{-1} {\cal S}_\Omega[{\mathbb P}_n w](z) - (B_{n}^z)^{-1} C_n^z {\cal S}_\Omega[{\mathbb P}_{n-1} w](z)- (B_{n}^z)^{-1} A_n^z {\cal S}_\Omega[{\mathbb P}_n w](z)
\end{aligned}$$
This means that we can calculate the Stieltjes transform in linear time by solving the recurrence equation, using explicit formulas for the $n=0$ and $n=1$ terms. Unfortunately, the results are numerically unstable for both $z$ on and off the contour. Here we sketch an alternative approach built on (F.W.J.) Olver’s and Miller’s algorithm, see [@DLMF Section 3.6] for references in the tridiagonal setting and [@GautschiOPs Section 2.3] for the equivalent application to calculating Stieltjes transforms of univariate orthogonal polynomials.
For $z$ off the contour, we can successfully and stably calculate the Stieltjes transform using a block-wise version of Olver’s algorithm, which is equivalent to solving the $2n +1 \times 2n+1$ block-tridiagonal system $$\begin{aligned}
\vc q_0 &= 1 \cr
C_k^z \vc q_{k-1} + (A_k^z - z I)\vc q_k + B_k^z \vc q_{k+1}& = 0 \qquad \hbox{for} k = 1,2,\ldots,n-1 \cr
\vc q_n &= \begin{pmatrix} 0\cr 0\end{pmatrix}\end{aligned}$$ where $\vc q_0 \in {\mathbb C}^1$ and $\vc q_k \in {\mathbb C}^2$ for $k=1,2,\ldots,n$. Then $${\cal S}_\Omega[{\mathbb P}_k w](z) \approx {\cal S}_\Omega[{\mathbb P}_{0} w](z) \vc q_k$$ Olver’s algorithm consists of performing Gaussian elimination adaptively until a convergence criteria is satisfied.
For $z$ on or near the contour, we no longer see quick decay in the Stieltjes transform (it is no longer a minimal solution to the recurrence), hence $n$ must be prohibitively large. Instead, we adapt Olver’s algorithm in a vein similar to Miller’s algorithm to allow for a non-decaying tail. We do so by calculating two additional solutions $\vc q^1$, and $\vc q^2$ (with the same block-sizes as before) satisfying: $$\begin{aligned}
\vc q_0^j &= 1 \cr
C_k^z \vc q_{k-1}^j + (A_k^z - z I)\vc q_k^j + B_k^z \vc q_{k+1}^j& = 0 \quad \hbox{for}\quad k = 1,2,\ldots,n-1 \cr
\qquad\vc q_{n}^1 = \begin{pmatrix} 1 \cr 0 \end{pmatrix} \qquad &\hbox{and}\qquad \vc q_{n}^2 = \begin{pmatrix} 0 \cr 1 \end{pmatrix}.\end{aligned}$$ These three solutions avoid picking up the exponentially growing solution that forward recurrence does. Thus we can solve a $3 \times 3$ system for constants $a,b$ and $c$ satisfying $$\begin{aligned}
a(z) \vc q_0 + b(z) \vc q_0^1 + c(z) \vc q_0^2 = {\cal S}_\Omega[{\mathbb P}_0 w](z) \\
a(z) \vc q_1 + b(z) \vc q_1^1 + c(z) \vc q_1^2 = {\cal S}_\Omega[{\mathbb P}_1 w](z) \end{aligned}$$ We immediately have that $$\label{eq:olvermiller}
{\cal S}_\Omega[{\mathbb P}_k w](z) = a(z) \vc q_k + b(z) \vc q_k^1 + c(z) \vc q_k^2 \qquad\hbox{for}\qquad k = 0, 1, \ldots, n.$$
\[fig:olvermiller\]
{width=".8\textwidth"}
While this holds true for all $n$, we note that in practice we need to choose $n$ bigger than the number of coefficients in order to observe numerical stability, see Figure \[fig:olvermiller\]. We also find that there are still stability issues near the corner. Resolving these issues is ongoing research.
[99]{}
A. Borodin, Chapter 11: Determinantal Point Processes, in *The Oxford Handbook of Random Matrix Theory*, eds. G. Akemann, J. Baik, P. Di Francesco, Oxford University Press, 2011
L.-L. Chau and O. Zaboronsky, On the structure of correlation functions in the normal matrix model, *Comm. Math. Phys.*, 196(1) (1998), 203–247
P. Deift, [*Orthogonal Polynomials and Random Matrices: a Riemann-Hilbert Approach*]{}, American Mathematical Soc., 1999
C. F. Dunkl and Y. Xu, *Orthogonal Polynomials of Several Variables, 2nd ed.* Encyclopedia of Mathematics and its Applications **155**, Cambridge University Press, 2014.
W. Gautschi, *Orthogonal Polynomials: Computation and Approximation*, Oxford University Press, 2004
A. Kroó and D. Lubinsky, Christoffel Functions and Universality on the Boundary of the Ball, *Acta Math. Hungarica*, 140 (2013), 117–133
A. Kroó and D. Lubinsky, Christoffel Functions and Universality in the Bulk for Multivariate Orthogonal Polynomials, *Can. J. Maths*, 65 (2013), 600–620
A. B. J. Kuijlaars and M. Vanlessen, Universality for eigenvalue correlations from the modified Jacobi unitary ensemble, *Int. Maths Res. Not.* 2002.30 (2002) 1575–1600
F. W. J. Olver, A. B. Olde Daalhuis, D. W. Lozier, B. I. Schneider, R. F. Boisvert, C. W. Clark, B. R. Miller, and B. V. Saunders, eds. NIST Digital Library of Mathematical Functions. http://dlmf.nist.gov/, Release 1.0.16 of 2017-09-18.
S. Olver, N. Raj Rao and T. Trogdon, Sampling unitary ensembles, *Rand. Mat.: Th. Appl.*, 4 (2015) 1550002
R.M. Slevinsky amd S. Olver, A fast and well-conditioned spectral method for singular integral equations, *J. Comp. Phys.*, 332 (2017) 290–315
G. Szegő, *Orthogonal Polynomials*, 4th edition, Amer. Math. Soc., Providence, RI. 1975.
[^1]: The second author was supported in part by NSF Grant DMS-1510296.
[^2]: Communicated by Alan Edelman.
[^3]: If there is such a random matrix distribution, one would expect it to be a pair of commuting random matrices, whose joint eigenvalues give points on the wedge.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The physical mechanisms that induce the transformation of a certain mass of gas in new stars are far from being well understood. Infrared bubbles associated with [H[ii]{}]{} regions have been considered to be good samples for investigating triggered star formation. In this paper we report on the investigation of the dust properties of the infrared bubble N4 around the H[ii]{} region [G11.898+0.747]{}, analyzing its interaction with its surroundings and star formation histories therein, with the aim of determining the possibility of star formation triggered by the expansion of the bubble. Using *Herschel* PACS and SPIRE images with a wide wavelength coverage, we reveal the dust properties over the entire bubble. Meanwhile, we are able to identify six dust clumps surrounding the bubble, with a mean size of 0.50 pc, temperature of about 22 K, mean column density of 1.7 $\times10^{22}$ 2, mean volume density of about 4.4 $\times10^{4}$ cm$^{-3}$, and a mean mass of 320 . In addition, from PAH emission seen at 8 , free-free emission detected at 20 cm and a probability density function in special regions, we could identify clear signatures of the influence of the H[ii]{} region on the surroundings. There are hints of star formation, though further investigation is required to demonstrate that N4 is the triggering source.'
author:
- 'Hong-Li Liu, Jin-Zeng Li,Yuefang Wu, Jing-Hua Yuan,Tie Liu, G. Dubner, S. Paron, M. E. Ortega, Sergio Molinari, Maohai Huang, Annie Zavagno, Manash R. Samal, Ya-Fang Huang, Si-Ju Zhang'
title: Interactions of the Infrared bubble N4 with the surroundings
---
Introduction {#s1}
============
Massive stars (of OB spectral type with masses greater than $\sim10 M_{\odot}$) are thought to form in clusters within molecular cloud complexes. Intense radiative and mechanical outputs from massive stars strongly affect their parent molecular clouds in two opposite ways: by disrupting molecular clouds and by triggering star formation. In effect, on the one hand, the natal molecular clouds can be accelerated beyond their escape velocities by feedback from massive stars through expanding H[ii]{} regions, outflows, stellar winds, and supernova explosions, leading to the disruption of the clouds which can stop the eventual star-forming processes. On the other hand, this feedback is capable of sweeping up the surroundings and accumulating them into condensations gravitationally bound, inducing the formation of new generations of stars.
Over the past several decades, numerous studies have focused on star formation triggered in the environs of H[ii]{} regions [e.g., @deh05; @deh06; @zav07; @deh08; @deh09; @ogu10; @elm11; @ken12; @dal13; @sam14; @liu15; @dal15]. Two different mechanisms have been proposed as models of star formation triggered in the surroundings of H[ii]{} regions: the “collect and collapse" process [C & C, @elm77] and “radiation-driven implosion" [RDI, @ber89; @lef94]. In the C & C process, the ultraviolet radiation from ionizing sources produces an ionization front (IF), creating an expanding H[ii]{} region. The supersonic expansion of the H[ii]{} region drives the shock front (SF) in front of the IF. Finally, a shell of circumstellar gas can be collected between the IF and the SF. In due time, the shell may become denser and collapse to form stars. Several numerical simulations [e.g., @hos05; @hos06; @dal07] have suggested that expanding H[ii]{} regions can trigger star formation through the C & C process only if ambient molecular material is massive enough. This process has been tentatively detected to be at work in several well-known H[ii]{} regions such as Sh 104 [@deh03], RCW 79 [@zav06], and RCW 120 [@deh09]. By contrast, in the RDI model, the IF drives the SF into molecular clouds surrounding the H[ii]{} region, stimulating the collapse of pre-existing subcritical clumps. Recent numerical simulations [e.g., @mia06; @mia09; @bis11] have demonstrated that the RDI model can successfully interpret star formation in bright-rimmed clouds (BRCs). This prediction seems to have been confirmed by observations of BRCs [e.g., @rea09; @liu12] indicating that star formation concentrated along the central axis of the clouds might be the consequence of the RDI process.
@tho12 and @ken12 speculated that around 14%-30% of the massive young stellar objects (MYSOs) in the Milky Way might have been induced by expanding bubbles/H[ii]{} regions on the basis of the association of a large sample of IR bubbles [@chu06; @chu07] with Red MSX MYSOs [@urq08]. However, the exact processes of triggered star formation are far from being understood. It is hoped that detailed studies of the physical interaction of bubbles/H[ii]{} regions together with their environs will lead to a better understanding of star formation associated with bubbles/H[ii]{} regions [@liu15].
Far-infrared observations carried out by the *Herschel Space Observatory* enable us to get a better insight into the physical properties of the bubble / H[ii]{} region, with unprecedented resolutions and sensitivities in that spectral range. Thanks to the widespread wavelength coverage from 70 to 500 , the dust temperature and column density distributions of all of the entire bubbles/H[ii]{} regions can be readily revealed by fitting spectral energy distributions (SEDs) pixel by pixel. Additionally, *Herschel* observations can reveal star formation surrounding the system. The widespread coverage is adequate for better constraining the SED profiles of YSOs at earlier phases and more accurately estimating their physical parameters such as stellar mass and luminosity. Moreover, *Herschel* observations coupled with suitable molecular lines offering kinetic and dynamical information are very useful tools for understanding the interactions of the bubbles/H[ii]{} regions with their surroundings. The present paper presents a comprehensive study of the bubble N4 [@chu06] using a combination of *Herschel* and molecular line observations.
The bubble N4 appears as an almost complete ring in the IR regime (see Figure \[fig-rgb1\]), enclosing the H[ii]{} region G11.898+0.747 [@loc89]. N4 is centered at , with a radius of $\sim 2\arcmin$, corresponding to $1.9\pm0.5$ pc at a distance of $3.2\pm0.9$ kpc [@chu06]. By analyzing the spatial distribution of candidate YSOs around N4, @wat10 found no evidence for triggered star formation. However, they claimed that the presence of triggered star formation can not be ruled out due to the lack of a complete sample of the YSO population as well as maps of the molecular gas and ionized gas in the environs. In contrast, the expanding motion of N4 uncovered by the observations of $J=1-0$ of , 3co, and 18o carried out by @li13 suggests a signature of possible triggered star formation. Therefore, the issue of star formation in N4 merits a more comprehensive investigation.
The purpose of this study is to analyze interactions of the bubble N4 with its surroundings and star formation histories therein, and to explore the possibility of triggered star formation. This paper is organized as follows: the *Herschel* and molecular line observations together with archival IR and radio data are presented in Section 2, the results are described in Section 3, the discussion is arranged in Section 4, and finally, we give a summary in Section 5.
Observations and data reduction {#s2}
===============================
*Herschel* Observations {#s-ob1}
-----------------------
The bubble N4 was observed as part of the [^1] survey. In this survey, the Photodetector Array Camera $\&$ Spectrometer [PACS, @pog10] at 70 and 160 and the Spectral and Photometric Imaging Receiver [SPIRE, @gri10] at 250, 350, and 500 worked simultaneously in the parallel photometric mode. The observations were run in two orthogonal scanning directions at a scan speed of 60 arcsec s$^{-1}$. The measured angular resolutions of these bands are 10$\farcs$7, 13$\farcs$9, 23$\farcs$9, 31$\farcs$3, and 43$\farcs$.8 [@tra11], respectively, corresponding to $0.12\pm0.05$ to $0.49\pm0.21$ pc at the distance to N4. The detailed descriptions of the pre-processing of the data up to usable high-quality images can be found in @tra11.
To search for early stages of star formation in N4, we took 56 point sources from the Curvature Threshold Extractor package (CuTeX) catalog. They are encompassed within the 2.5 times radii of the bubble, of detection in the PACS 70 band. Details about the catalog and the photometric procedures considered in CuTeX can be found in @mol11.
Molecular Line Observations
---------------------------
[lllllll]{} H$_{2}$O $6_{1,6}-5_{2,3}$ & 22.235077 & 120 & 0.46 & 80 & 0.05 & 0.21\
CH$_{3}$OH $7_{0,7}-6_{1,6}$ & 44.069476 & 60.5 & 0.48 & 170 & 0.07 & 0.11\
HCO$^{+}$ (1-0) & 89.188526 & 31.5 & 0.38 & 250 & 0.08 & 0.05\
o-H$_{2}$CO $2_{1,2}-1_{1,1}$ & 140.83952 & 23.4 & 0.33 & 360 & 0.09 & 0.04
Single-dish observations of molecular lines were performed with the Korean VLBI Network (KVN) 21-m radio telescope at Yonsei site, Seoul on 2014 December 28. The front end of the single-dish telescope is equipped with the four receivers working at 22, 43, 86, and 129 GHz simultaneously. A digital filter bank (DFB) with a 64 MHz bandwidth split into 4096 channels was adopted, resulting in velocity resolutions of 0.21 , 0.11 , 0.05 , and 0.04 , respectively. The pointing and tracking accuracies were better than 3$\arcsec$ [@kim11].
Since mapping observations are quite time-consuming, we only chose six dense clumps, located on the border of N4, to make single-point observations. To investigate the dynamical and kinetic properties (see Section \[s-mle\] and \[s-sfc\]) and star formation activity in the clumps, H$_{2}$O (22 GHz), CH$_{3}$OH (44 GHz), HCO$^{+}$ (89 GHz), and o-H$_{2}$CO (141 GHz) were observed simultaneously. The data were calibrated to antenna temperature ($T_{a}^{*}$) using the chopper wheel method. A summary of observations is given in Table \[tbl-1\], including the half-power beam width ($\theta_{\mathrm{HPBW}}$), the beam efficiency ($\eta$), the typical systematic temperature ($T_{\mathrm{sys}}$), the rms noise level ($T_{\mathrm{rms}}$) and the velocity resolution ($\delta V$). The data were analyzed and visualized with the software GILDAS [@gui00].
Archival data {#s-ad}
-------------
To study the physical characteristics of the bubble and its associated YSO candidates, auxiliary data of images and point sources were complemented by the GLIMPSE [@ben03], MIPSGAL [@car09] and *WISE* [@wri10] surveys. The images of the Spitzer Infrared Array Camera (IRAC) at 3.6, 4.5, 5.8, and 8.0 , together with the Multiband Imaging Photometer for *Spitzer* (MIPS) at 24 were retrieved from the *Spitzer* Archive.[^2]. The angular resolutions of the images in the IRAC bands are $< 2\arcsec$ and it is $\sim 6\arcsec$.0 in the MIPS 24 band. In addition, the *WISE* survey has mapped the entire sky in four infrared bands centered at 3.4, 4.6, 12, and 22 , with a resolution of 6$\farcs$1, 6$\farcs$4, 6$\farcs$5, and 12$\farcs$0, respectively [@wri10]. Point sources were obtained from the *AllWISE* Source and MIPSGAL Catalogs.[^3] Both catalogs contain the photometries of point sources cross matched with the Two Micron All Sky [2MASS, @skr06] survey. The 20 cm radio continuum image was used from the Multi-Array Galactic Plane Imaging Survey [MAGPIS, @hel06] archive[^4] was used to analyze the properties of the H[ii]{} region associated with N4, . Its angular resolution is less than 6$\farcs$0 and the sensitivity is better than 0.15 mJy beam$^{-1}$.
![ Composite three-color image of N4. The IRAC 4.5 , 8.0 , and 24 are colored in blue, green, and red, respectively. The image is centered at .[]{data-label="fig-rgb1"}](N4rgb.pdf){width="48.00000%"}
Results {#s3}
=======
Dust Emission {#s-dust}
-------------
### Dust Temperature and Column Density Distributions
*Herschel* observations with the widespread wavelength coverage can reveal the dust properties of the entire bubble. To do this, a SED pixel-by-pixel fitting was performed to obtain the distributions of dust temperature ($T_{\mathrm{dust}}$) and H$_{2}$ column density ($N_{\mathrm{H_{2}}}$). Assuming optically thin dust emission, a graybody function for a single temperature can be expressed as follows: $$\label{eq-1}
I_{\nu} = \kappa_{\nu 0}(\nu/ \nu_{0})^{\beta}\ B_{\nu}(T_{\mathrm{d}})\ \mu \ m_\mathrm{H}\ N(\mathrm{H}_{2}),$$ where $I_{\nu}$ is the surface brightness, $B_{\nu}(T_{\mathrm{d}})$ is the blackbody function for the dust temperature of $T_{\mathrm{d}}$, the mean molecular weight $\mu$ is assumed to be 2.8 [e.g., @kau08; @sad13]; and $m_\mathrm{H}$ is the mass of a hydrogen atom. The dust opacity per unit mass of both dust and gas is defined as $\kappa_{\nu}=\kappa_{\nu 0}(\nu/ \nu_{0})^{\beta}$, where $\kappa_{\nu 0}$ is assumed to be 0.1 cm$^{2}~$g$^{-1}$ at 1 THz [@bec90] under a gas-to-dust mass ratio of 100, and $\beta$ is fixed to 2, which is a statistical value found in a large sample of H[ii]{} regions [@and12].
The *Herschel* data of the four bands at 160, 250, 350, and 500 were used for the SED fitting. Since emission at 70 may trace hotter components such as very small grains (VSGs) and warmer material heated by protostars, it was not considered in the SED fitting for a single temperature. The four images were convolved to the same resolution 43$\farcs$8, and rebinned to the same pixel size 11$\farcs$5. To minimize the contribution from the line of sight, we subtracted the backgrounds for each image using a reference area away from the bubble. The corresponding backgrounds were estimated to be 634 MJy sr$^{-1}$ for 160 , 387 MJy sr$^{-1}$ for 250 , 205 MJy sr$^{-1}$ for 350 and 79 MJy sr$^{-1}$ for 500 . By treating $T_{\mathrm{dust}}$ and $N_{\mathrm{H_{2}}}$ as free parameters, the SED fitting was performed using the IDL program MPFITFUN[^5] [@mar09]. In principle, the 870 image should have been considered in the SED fitting. Although the 870 data constrain the dust temperature, as they are sensitive to colder components than *Herschel* data, they miss the bulk of extended emission, which is the major interest in our current work. Thus, the image at 870 has been excluded in the SED fitting.
The resulting $T_{\mathrm{dust}}$ and $N_{\mathrm{H_{2}}}$ maps are shown in Figure \[fig-NT\], where the black contours stand for the emission at 24 and the purple contour delineates a level of $N_{\mathrm{H_{2}}} = 9\times10^{21}$ 2, sculpting a shell structure. Emission traced by 24 mainly comes from hot dust which can reach very high temperatures after absorbing high-energy photons [e.g., @deh10 references therein]. The hot dust distribution shown in Figure \[fig-NT\](a) indicates a dust heating from the interior to the edge of N4 (also see Figure \[fig-rgb1\]), therefore, possible temperature gradients along this direction would be expected. However, the dust distributions in the interior and on the edge of N4 appear to be smooth. Given the resolution of 6$\arcsec$ at 24 , the smoothness could be interpreted as the temperature gradients on small scales not resolved by the angular resolution of 43$\farcs$8 of the dust temperature map. Additionally, we note that there is a cold region of $\sim 22$ K within the bubble, which may be not real due to the low flux found there [@and12].
On the large scale, the temperature distribution nearly matches with 24 emission (see Figure \[fig-NT\](a)). That is, the temperatures of cold dust in the bubble direction are hotter than those away the bubble, suggesting that the H[ii]{} region enclosed by N4 has been heating its surroundings. On the small scale, there are three regions that are especially warmer than other places in the $T_{\mathrm{dust}}$ map. The first region, located near the bubble center, it can be attributed to the heating of the central ionizing stars (see Section \[s-HII\]). The second warm region, with which some YSO candidates are associated (see Section \[s-yso\]), is to the north of N4. Hence, this warm region could be a result of the heating from embedded protostars therein. However, the third warm region in the northern border of N4 has no bright 24 counterpart or potential associated YSO candidate. Figure \[fig-NT\](b) shows that the column density of this region is lower than that of its neighbors. This difference suggests that this warm region may be heated premarily by the ionizing photons leaking from the H[ii]{} region.
For the H$_{2}$ column density distribution in Figure \[fig-NT\](b), a shell morphology is inscribed as seen in IR bands. In the shell, there is an anti-correlation between $N_{\mathrm{H_{2}}}$ and $T_{\mathrm{dust}}$. This may be attributed to a lower penetration of the external heating from the H[ii]{} region into dense regions. If the shell is bounded by a level of $9\times10^{21}$ cm$^{-2}$ as shown in Figure \[fig-NT\](b), which covers a region of $\sim 1.5$ radius of the bubble, the mean $N_{\mathrm{H_{2}}}$ column density and total mass of the shell can be estimated to be $1.1\times10^{22}$ cm$^{-2}$ and $5.2\times10^{3}$ , respectively. The total mass of the neutral material of the entire bubble is estimated to be $\sim~5.5\times10^{3}$ .
{width="80.00000%"}
. \[fig-NT\]
### Dense Clumps {#s-DC}
Six dense clumps were identified within the shell structure in the $\mathrm{H}_{2}$ column density image. These sources were decomposed visually from the image by ellipses bounded by the level of $9\times10^{21}$ cm$^{-2}$ enclosing the majority of fluxes. The six clumps have different physical sizes and peak column densities ($N_{\mathrm{H_{2},peak}}$). Their average temperatures ($<T_{dust}>$) were obtained from the dust temperature image. We estimated the total mass (gas+dust) and the volume density for each clump.
[llllllllll]{}\[tbh!\]
A & 18:08:53.14 & -18:18:25.5 & 24.3 & 44.6 & 0.4 $\pm$ 0.1 & 22.3 $\pm$ 0.2 & 1.0 $\pm$ 0.1 & 4.0 $\pm$ 1.4 & 0.8 $\pm$ 0.5\
B & 18:09:00.66 & -18:16:30.0 & 36.7 & 51.9 & 0.6 $\pm$ 0.2 & 22.1 $\pm$ 0.5 & 1.9 $\pm$ 0.3 & 5.0 $\pm$ 2.1 & 3.5 $\pm$ 2.5\
C & 18:08:55.59 & -18:14:44.5 & 34.9 & 49.9 & 0.6 $\pm$ 0.2 & 23.1 $\pm$ 0.9 & 2.8 $\pm$ 0.5 & 7.8 $\pm$ 3.6 & 4.5 $\pm$ 3.4\
D & 18:08:48.57 & -18:14:23.1 & 34.1 & 40.8 & 0.5 $\pm$ 0.1 & 22.3 $\pm$ 0.7 & 1.6 $\pm$ 0.2 & 5.2 $\pm$ 2.1 & 1.9 $\pm$ 1.3\
E & 18:08:45.39 & -18:15:34.6 & 35.8 & 41.0 & 0.5 $\pm$ 0.2 & 21.7 $\pm$ 0.7 & 1.5 $\pm$ 0.1 & 4.7 $\pm$ 1.8 & 1.9 $\pm$ 1.3\
F & 18:08:45.18 & -18:17:40.7 & 37.4 & 35.2 & 0.5 $\pm$ 0.1 & 23.6 $\pm$ 0.7 & 1.3 $\pm$ 0.1 & 4.4 $\pm$ 1.6 & 1.4 $\pm$ 0.9
Table \[tbl-2\] gives a summary of the aforementioned parameters, where the main source of error is the uncertainty of the distance to the bubble. The average size, dust temperature, column density, number density, and mass of the six clumps are $\sim0.5$ pc, $\sim22$ K, $\sim1.7\times10^{22}$ cm$^{-2}$, $\sim4.4\times10^{4}$ cm$^{-3}$, and $\sim3.2\times10^{2}$ , respectively. The average density $> 10^{4}$ cm$^{-3}$ is consistent with the detections of dense molecular lines within these clumps (see Section \[s-mle\]).
Molecular Line Emission {#s-mle}
-----------------------
[lllllllll]{}\[tbh!\]
A & 23.61 $\pm$ 0.06 & 2.62 $\pm$ 0.11 & 1.30 $\pm$ 0.12 & 3.62 $\pm$ 0.16 & 23.53 $\pm$ 0.25 & 2.30 $\pm$ 0.52 & 0.40 $\pm$ 0.11 & 0.99 $\pm$ 0.19\
B & 23.58 $\pm$ 0.06 & 2.03 $\pm$ 0.08 & 1.84 $\pm$ 0.12 & 4.01 $\pm$ 0.15 & 23.88 $\pm$ 0.08 & 1.91 $\pm$ 0.19 & 0.97 $\pm$ 0.11 & 1.98 $\pm$ 0.17\
C-I & 23.56 $\pm$ 0.21 & 2.54 $\pm$ 0.21 & 0.99 $\pm$ 0.11 & 3.84 $\pm$ 0.31 & 24.51 $\pm$ 0.38 & 4.32 $\pm$ 0.70 & 0.53 $\pm$ 0.12 & 2.42 $\pm$ 0.41\
C-II & 26.68 $\pm$ 0.06 & 1.79 $\pm$ 0.21 & 2.91 $\pm$ 0.11 & 4.74 $\pm$ 0.26 & 26.87 $\pm$ 0.04 & 1.16 $\pm$ 0.11 & 1.78 $\pm$ 0.12 & 2.19 $\pm$ 0.29\
D & 25.00 $\pm$ 0.06 & 3.06 $\pm$ 0.10 & 1.47 $\pm$ 0.11 & 4.77 $\pm$ 0.17 & 25.11 $\pm$ 0.12 & 2.00 $\pm$ 0.35 & 0.67 $\pm$ 0.11 & 1.43 $\pm$ 0.19\
E & 25.28 $\pm$ 0.06 & 2.46 $\pm$ 0.06 & 2.33 $\pm$ 0.11 & 6.12 $\pm$ 0.17 & 25.47 $\pm$ 0.09 & 2.23 $\pm$ 0.28 & 0.96 $\pm$ 0.12 & 2.26 $\pm$ 0.21\
F & 24.86 $\pm$ 0.09 & 4.02 $\pm$ 0.24 & 0.85 $\pm$ 0.12 & 3.64 $\pm$ 0.23 & 25.36 $\pm$ 0.20 & 2.91 $\pm$ 0.51 & 0.49 $\pm$ 0.11 & 1.53 $\pm$ 0.22
![HCO$^{+}$ (1-0) and o-H$_{2}$CO $2_{1,2}-1_{1,1}$ spectra of the six clumps. The black dashed lines indicate a systematic velocity of 24.7 for the bubble. The systematic velocity is determined from the average spectra of the HCO$^{+}$ and H$_{2}$CO lines of the six clumps, which is coincident with the results of , 3co, and 18o (1-0) observations [@li13]. $T_{\mathrm{mb}}$ represents the main-beam temperature. The name of each clump is noted in the top left of each panel.[]{data-label="fig-coreSp"}](coreSp.pdf){width="48.00000%"}
Figure \[fig-coreSp\] presents HCO$^{+}$ (1-0) and o-H$_{2}$CO $2_{1,2}-1_{1,1}$ spectra of the six clumps. Since the molecules H$_{2}$O and CH$_{3}$OH were not detected, their spectra are not be displayed. Emission of HCO$^{+}$ (1-0) and o-H$_{2}$CO $2_{1,2}-1_{1,1}$ is strong enough to be detected in all clumps. Their detections indicate high densities for the six clumps, since the critical densities of both lines are $4.5\times10^{4}$ , and $1.3\times10^{5}$ for a kinetic temperature of 20 K [@shi15]. Using the software GILDAS, we retrieved the observed parameters from the spectra of the six clumps, including the peak velocity ($V_{LSR}$), line width (FWHM) and the main-beam temperature ($T_{mb}$), and the velocity-integrated intensity ($\int T_{mb} \ \mathrm{d}V$) of the different spectra (see Table \[tbl-3\]).
It is interesting to analyze the centroid velocity variances of the six clumps. The spectra of HCO$^{+}$ and o-H$_{2}$CO obviously show the velocities blueshifted to a systematic velocity of 24.7 in the southeastern (SE) part of the bubble (i.e., Clumps A and B) and the redshifted velocities in the northwestern (NW) part (i.e., Clumps D and E). The two remaining clumps show no evident velocity shifts. Observations of *J=1-0* of , 3co, and 18o also have already revealed the same velocity difference between the SE and the NW [see Figure 3 of @li13]. Due to the lack of blueshifted profiles in the front of the bubble and redshifted ones in the back cap of the bubble in the CO observations, @li13 suggested that the bubble N4 may be expanding along the SE-NW direction with an inclination relative to the sky plane. This scenario is compatible with @bea10 interpretations of molecular gas studies around a number of IR bubbles, where they conclude that the SFs driven by massive stars tend to create rings. Therefore, the shell of dust and gas in the bubble N4 is presumably assembled by the expanding H[ii]{} region.
Assuming that the observed o-H$_{2}$CO line is optically thin and in local thermodynamic equilibrium (LTE), we estimate the column density of this molecular species toward each dense clump presented in Section \[s-DC\] using [@goldsmith99]: $$N = \frac{8 \pi k \nu^{2}}{h c^{3} A_{ul}} \frac{Q_{rot}}{g_{u}}~{\rm exp}(\frac{E_{u}}{T_{rot}}) \int{T_{mb}~{\rm dv}},
\label{eqNh2co}$$ where $A_{ul} = 5.4 \times 10^{-5}$ s$^{-1}$, $E_{u} = 21.9$ K, and $g_{u} = 15$. Assuming that the dust and the gas are coupled at the same temperature, in LTE conditions we can approximate $T_{dust}=T_{kin}=T_{rot}$. Thus, for $T_{rot}$ we used the $T_{dust}$ value obtained for each clump (see Section \[s-DC\]). Given a temperature of $\sim22$ K averaged over the six clumps, the o-H$_{2}$CO partition function $Q_{rot}$ was estimated to be $\sim50$ by the extrapolation to the relation between $Q_{rot}$ and $T_{rot}$ from the CDMS Catalog.[^6] The obtained column density values for each clump are presented in Table \[colabund\], which also includes the o-H$_{2}$CO abundances ($X =$ N(o-H$_{2}$CO)/N(H$_{2}$)). The abundances were estimated using the H$_{2}$ column densities presented in Section \[s-DC\]. We note that o-H$_{2}$CO emission is generally optically thick, therefore, the assumption of optically thin emission of o-H$_{2}$CO results in underestimated column density and abundance.
The obtained o-H$_{2}$CO column densities and abundances are quite similar to those obtained toward other photo-dominated regions (PDRs), such as the Horsehead PDR [@guzman11], and the Orion Bar [@vanderwiel; @vanderwielE]. Considering that the dust grains are cold ($T_{dust} \sim 22$ K) in the dense clumps around the N4 bubble and given the abundances presented in Table \[colabund\], we suggest that in this region the formaldehyde may be formed mainly in the gas phase with a probable contribution of photo-desorption of the grain mantles as was found in the Horsehead PDR [@guzman11].
Observations of other o-H$_{2}$CO and p-H$_{2}$CO lines toward this region would be very useful to give an important observational probe of the gas phase and grain surface chemistry toward PDRs.
[lcc]{} Clump & N(o-H$_{2}$CO) & $X$\
& ($\times10^{12}$ cm$^{-2}$) & (10$^{-10}$)\
A & 6.26 & 6.2\
B & 12.66 & 6.6\
C & 28.20 & 9.9\
D & 9.05 & 5.6\
E & 14.61 & 9.7\
F & 9.13 & 7.0\
YSOs Associated with N4 {#s-yso}
-----------------------
To understand more about star formation histories associated with the bubble, candidate YSOs associated with N4 have been initially identified using the online SED fitting tool[^7] of @rob06 [@rob07] and then further demonstrated and classified with color-color diagrams. The SED fitting tool invokes a grid of 20,000 two-dimensional Monte Carlo radiation transfer models [e.g., @whi04]. With 10 viewing angles for each model, there are actually 200,000 YSO SED models. This tool works as a linear regression method to fit these models to the multi-wavelength photometry measurements of a given source. The goodness/badness of each fit could be quantified by a specified value of the best $\chi^{2}$ ($\chi^{2}_{best}$), whereby YSO candidates could be robustly distinguished from reddened photospheres of main-sequence and giant stars since YSOs require a thermal emission component to reproduce the shapes of their mid-IR excesses. In addition, this method can make use of any available data to constrain the shape of SED profiles as well as possible (see Appendix \[s-app\]). However, we have to acknowledge that the disk and stellar parameters inferred from the SED models can be very uncertain [@off12 and T.P. Robitaille 2015, private communication]. Therefore, if these uncertain parameters are used to classify the identified YSO candidates with the scheme of YSO classification of @rob06, the classification results must be unreliable. To avoid this uncertainty, the color selection schemes of @gut09 and @koe12 have been adopted to group the YSO candidates identified by the SED fitting into three “Classes": Class I, Class II, and Transition Disk (TD) YSOs. Class I objects are deeply embedded protostars with a dominant infalling envelope, Class II objects are surrounded by a substantial accreting disk [@lad87], and TD objects are more evolved protostars where the inner parts of the disks have been cleared by photoevaporation of the central stars or by planet-forming processes [@gut09; @yua14; @par15]. We believe that using the infrared color schemes is a better way to categorize the YSO candidates because the infrared color schemes are a powerful proxy for measuring the excess emission [e.g., @all04; @gut08; @gut09; @koe12]. Apart from this, the candidate YSOs, which have been singled out by the SED fitting, can be further confirmed by such color schemes.
Sixty potential YSOs within 5$\arcmin$ from the bubble center have been identified and classified. Of these, there are 8 Class I, 12 Class II, and 40 TD objects. The classification results are listed in Table \[tbl-5\] and the detailed processes can be found in the Appendix \[s-app\]. We note that the resulting 60 YSO candidates are incomplete but robust (see the Appendix \[s-app\]). Despite this incompleteness, the resulting population with robust identifications is enough for simply learning about the distributions of the YSOs associated with N4.
Figure \[fig-N4rgbIII\] shows the spatial distribution of the 60 YSO candidates. One can see that the majority ($\sim82\%$) of YSOs are spatially correlated with the PDRs as traced by 8 emission (see Section \[s-HII\]), which indicates the good spatial association of these potential YSOs with the bubble N4 and the strong impacts that the enclosed H[ii]{} region is having on the surrounding star formation. In addition, there is an overdensity of the number of YSO candidates residing on the verge of the bubble. This is also demonstrated by the statistics of all candidates as a function of the normalized bubble radius (see Figure \[fig-ysoStat\]). The statistics suggest that a large number ($\sim40\%$) of YSO candidates is located within $1.0-1.5$ times the radius of the bubble. Such spatial distribution of YSOs in N4 agrees well with the statistical results based on the study of the association of a large sample of YSOs with IR bubbles [@ken12; @tho12].
[lllllll]{}\[tbh!\]
Y1 & J180831.89-181639.1 & 272.133 & -18.278 & 7 & 13.4 & II\
Y2 & J180832.49-181616.6 & 272.135 & -18.271 & 7 & 35.7 & TD\
Y3 & J180834.88-181751.0 & 272.145 & -18.298 & 7 & 18.3 & TD\
Y4 & J180835.68-181814.8 & 272.149 & -18.304 & 7 & 11.0 & TD\
Y5 & J180839.36-181813.8 & 272.164 & -18.304 & 7 & 20.9 & TD\
Y6 & J180839.71-181841.5 & 272.165 & -18.312 & 7 & 47.7 & TD\
Y7 & J180840.11-181902.3 & 272.167 & -18.317 & 7 & 21.6 & TD\
Y8 & J180841.62-181600.0 & 272.173 & -18.267 & 7 & 12.9 & TD\
Y9 & J180842.08-181457.5 & 272.175 & -18.249 & 7 & 8.2 & TD\
Y10 & J180842.19-181329.1 & 272.176 & -18.225 & 7 & 34.6 & TD\
Y11 & J180842.69-181930.1 & 272.178 & -18.325 & 7 & 34.0 & TD\
Y12 & J180843.58-181441.8 & 272.182 & -18.245 & 7 & 38.3 & TD\
Y13 & J180844.11-181456.4 & 272.184 & -18.249 & 7 & 30.4 & TD\
Y14 & J180844.15-181401.9 & 272.184 & -18.234 & 7 & 43.4 & TD\
Y15 & J180844.97-181457.3 & 272.187 & -18.249 & 7 & 40.2 & TD\
Y16 & J180845.97-181612.4 & 272.192 & -18.270 & 8 & 32.2 & II\
Y17 & J180846.14-181844.0 & 272.192 & -18.312 & 7 & 32.0 & TD\
Y18 & J180847.47-181416.2 & 272.198 & -18.238 & 7 & 22.3 & TD\
Y19 & J180847.65-181450.0 & 272.199 & -18.247 & 7 & 32.9 & TD\
Y20 & J180847.67-181441.7 & 272.199 & -18.245 & 7 & 12.6 & TD\
Y21 & J180848.33-181357.2 & 272.201 & -18.233 & 7 & 33.9 & TD\
Y22 & J180849.08-181441.8 & 272.205 & -18.245 & 4 & 0.2 & I\
Y23 & J180849.78-181340.5 & 272.207 & -18.228 & 7 & 6.5 & TD\
Y24 & J180850.19-181816.2 & 272.209 & -18.305 & 7 & 27.3 & II\
Y25 & J180850.83-181836.6 & 272.212 & -18.310 & 7 & 14.0 & II\
Y26 & J180851.55-181831.9 & 272.215 & -18.309 & 7 & 7.6 & TD\
Y27 & J180852.42-181744.2 & 272.218 & -18.296 & 7 & 22.5 & II\
Y28 & J180852.59-181356.6 & 272.219 & -18.232 & 7 & 7.4 & II\
Y29 & J180852.97-181336.2 & 272.221 & -18.227 & 7 & 43.8 & TD\
Y30 & J180853.04-181742.2 & 272.221 & -18.295 & 8 & 5.5 & I\
Y31 & J180853.35-181130.9 & 272.222 & -18.192 & 7 & 41.9 & TD\
Y32 & J180853.65-181208.4 & 272.224 & -18.202 & 7 & 29.8 & TD\
Y33 & J180853.94-181422.4 & 272.225 & -18.240 & 5 & 0.3 & I\
Y34 & J180854.47-181406.5 & 272.227 & -18.235 & 4 & 0.1 & I\
Y35 & J180854.69-181230.5 & 272.228 & -18.208 & 7 & 39.4 & TD\
Y36 & J180854.87-181409.5 & 272.229 & -18.236 & 4 & 0.1 & I\
Y37 & J180855.80-181136.3 & 272.233 & -18.193 & 7 & 16.7 & TD\
Y38 & J180856.26-181505.7 & 272.234 & -18.252 & 7 & 26.5 & II\
Y39 & J180856.64-181420.4 & 272.236 & -18.239 & 7 & 12.4 & II\
Y40 & J180857.00-181354.7 & 272.238 & -18.232 & 4 & 0.1 & II\
Y41 & J180857.48-181853.3 & 272.240 & -18.315 & 7 & 25.0 & TD\
Y42 & J180858.07-181506.0 & 272.242 & -18.252 & 7 & 12.8 & TD\
Y43 & J180858.12-181807.2 & 272.242 & -18.302 & 7 & 26.4 & TD\
Y44 & J180858.75-181806.5 & 272.245 & -18.302 & 7 & 22.4 & TD\
Y45 & J180858.77-181629.7 & 272.245 & -18.275 & 11 & 25.1 & I\
Y46 & J180859.32-181316.1 & 272.247 & -18.221 & 7 & 9.8 & II\
Y47 & J180859.44-181332.8 & 272.248 & -18.226 & 7 & 0.9 & I\
Y48 & J180859.47-181855.8 & 272.248 & -18.316 & 7 & 14.1 & TD\
Y49 & J180900.26-181347.5 & 272.251 & -18.230 & 7 & 11.3 & TD\
Y50 & J180900.28-181921.8 & 272.251 & -18.323 & 7 & 31.4 & TD\
Y51 & J180900.48-181903.2 & 272.252 & -18.318 & 7 & 28.2 & TD\
Y52 & J180902.92-181316.2 & 272.262 & -18.221 & 7 & 41.9 & TD\
Y53 & J180904.56-181638.3 & 272.269 & -18.277 & 7 & 6.6 & TD\
Y54 & J180904.95-181806.2 & 272.271 & -18.302 & 7 & 23.3 & TD\
Y55 & J180905.60-181621.4 & 272.273 & -18.273 & 8 & 2.1 & II\
Y56 & J180906.19-181639.1 & 272.276 & -18.278 & 7 & 0.9 & TD\
Y57 & J180906.60-181847.9 & 272.278 & -18.313 & 7 & 19.7 & TD\
Y58 & J180908.60-181706.1 & 272.286 & -18.285 & 7 & 39.2 & TD\
Y59 & MG011.8545+00.7327 & 272.216 & -18.313 & 10 & 8.7 & I\
Y60 & MG011.9455+00.7481 & 272.248 & -18.226 & 9 & 6.6 & II
Properties of the Ionized Region {#s-HII}
--------------------------------
![(a) Image at 8 (grayscale) overlaid with free-free continuum emission at 20 cm (contours). The yellow contours start from $5 \sigma$ (i.e., $1 \sigma=0.15$ mJy Beam$^{-1}$) with a step of $8 \sigma$. Class I, Class II, and transition disk YSO candidates are symbolized by square, triangle, and circle symbols, respectively. The dashed rectangle is a selected region for a close-up view of 4.5 emission. (b) Image at 4.5 (colorscale) overlaid with dust emission at 250 (contours). The white contours start from 18 Jy Beam$^{-1}$ with a step of 5.5 Jy Beam$^{-1}$. The ellipses depict the dust clumps extracted in Section \[s-DC\], and the cross symbols represent their peak positions. A scale bar of 1 pc is shown on the bottom left.[]{data-label="fig-N4rgbIII"}](imIRAC4YSO.pdf "fig:"){width="48.00000%"} ![(a) Image at 8 (grayscale) overlaid with free-free continuum emission at 20 cm (contours). The yellow contours start from $5 \sigma$ (i.e., $1 \sigma=0.15$ mJy Beam$^{-1}$) with a step of $8 \sigma$. Class I, Class II, and transition disk YSO candidates are symbolized by square, triangle, and circle symbols, respectively. The dashed rectangle is a selected region for a close-up view of 4.5 emission. (b) Image at 4.5 (colorscale) overlaid with dust emission at 250 (contours). The white contours start from 18 Jy Beam$^{-1}$ with a step of 5.5 Jy Beam$^{-1}$. The ellipses depict the dust clumps extracted in Section \[s-DC\], and the cross symbols represent their peak positions. A scale bar of 1 pc is shown on the bottom left.[]{data-label="fig-N4rgbIII"}](imIRAC2.pdf "fig:"){width="48.00000%"}
![Plot of the number counts of the 60 YSO candidates as a function of angular distance from the center of the bubble N4. The distance is normalized by the bubble radius. Error bars are determined via Poisson statistics.[]{data-label="fig-ysoStat"}](ysoDistHisto.pdf){width="45.00000%"}
Figure \[fig-N4rgbIII\] (a) presents the 8 image overlaid with radiation at 20 cm. The 8 emission predominantly comes from two prominent polycyclic aromatic hydrocarbons (PAHs) at 7.7 and 8.6 . They are good tracers of PDRs delineating the IF created by massive stars. The 20 cm radiation represents free-free continuum emission from ionized gas. As shown in Figure \[fig-N4rgbIII\] (a), most of the ionized gas is surrounded by the ring of PDRs, showing the strong effects of the H[ii]{} region on its surroundings. Additionally, the ring of PDRs is also spatially well correlated with the 250 emission (see Figure \[fig-N4rgbIII\] (b)), which is sensible to cold dust in thermal equilibrium. As such, the cold dust wrapping ionized gas suggests a positive association of the bubble N4 with the H[ii]{} region, enhancing their intensive mutual interaction.
Several properties of the H[ii]{} region can be derived from 20 cm free-free continuum emission. @chu78 [@win83; @qui06] demonstrated the existence of a relationship between the H[ii]{} region electron temperature, $T_{e}$, and the Galactocentric distance, $R_{\mathrm{Gal}}$. A detailed analysis of a sample of 76 H[ii]{} regions with the highest quality data suggested that the electron temperature along the Galactic disk decreases approximately as a function of $R_{\mathrm{Gal}}$ [@qui06]: $$\label{eq-4}
T_{e} = (5780\pm350) + (287\pm46)\ R_{\mathrm{Gal}}.$$ With an estimated Galactocentric distance of 5.5 kpc to N4, it yields an electron temperature of $7400\pm600$ K. If the H[ii]{} region reaches the equilibrium at this temperature, the electron density, $n_{\mathrm{e}}$, the mass of ionized gas, $M_{\mathrm{ion}}$, and the Lyman photons per second, $N_{\mathrm{Lym}}$, from the exciting star(s), can be estimated following @kur94: $$\label{eq-5}
n_{\mathrm{e}} = 2.878\times 10^{4}\ [(\frac{\theta}{\mathrm{arcsec}})^{-3}\ (\frac{D}{\mathrm{kpc}})^{-1} \ (\frac{\nu}{\mathrm{GHz}})^{0.1}\ (\frac{T_{\mathrm{e}}}{\mathrm{K}})^{0.35}\ (\frac{S_{\nu}}{\mathrm{Jy}})]^{0.5}\ \mathrm{cm}^{-3},$$
$$\label{eq-6}
M_{\mathrm{ion}} = \frac{4}{3}\ \pi \ r^{3}_{\mathrm{H{\,\sc{\mathrm{ii}}}}}\ n_{e}\ \ m_\mathrm{\mathrm{p}},$$
$$\label{eq-7}
N_{\mathrm{Ly}} = 7.588\times 10^{48}\ (\frac{T_{\mathrm{e}}}{\mathrm{K}})^{-0.5}\ (\frac{\nu}{\mathrm{GHz}})^{0.1}\ (\frac{S_{\nu}}{\mathrm{Jy}})\ (\frac{D}{\mathrm{kpc}})^{2} \ \mathrm{ph}\ \mathrm{s}^{-1},$$
where $S_{\nu}$ is the integrated flux density at the special frequency $\nu$ over the angular size $\theta$; $r_{\mathrm{H{\,\sc{\mathrm{ii}}}}}$ is the radius of the H[ii]{} region, and $m_\mathrm{p}$ is the proton mass. If the $\theta$ and $r_{\mathrm{H{\,\sc{\mathrm{ii}}}}}$ are considered to be approximately equal to those of the bubble, a total of $32.8\pm0.01$ Jy at 20 cm measured by integrating over the 5 $\sigma$ contour (see Figure \[fig-N4rgbIII\]) yields an estimated $n_{\mathrm{e}}$, $M_{\mathrm{ion}}$, and $N_{\mathrm{Lym}}$ of $75\pm15$ 2, $50\pm38$ , and $(1.5\pm1.1)\times10^{48}$ $\mathrm{ph}\ \mathrm{s}^{-1}$, respectively. The errors of these parameters result mainly from the uncertainty in the distance of N4. According to @mar05, the estimated $N_{\mathrm{Lym}}$ (Log($N_{\mathrm{Lym}}$) $\simeq48.18$) indicates that a main O8.5V-O9V star was responsible for the ionization of N4. These are lower limits given that any ionizing photons absorbed by dust or moving away from the bubble were not accounted for.
The central exciting stars of N4 may be located near the very bright 24 emission inside the bubble. Based on the statistical analysis for the seven well-defined bubbles including N4, @deh10 argued that the hot dust grains seen at 24 inside the ionized region can be heated by absorption of the Lyman continuum photons which are more numerous near exciting stars. This indicates that the ionizing stars of N4 may reside near the very bright 24 emission inside the bubble (see Figures \[fig-rgb1\] and \[fig-NT\](a)). This explanation can be further supported by the positive association of intense free-free continuum radiation from ionized gas with the bright 24 emission (see Figures \[fig-NT\](a) and \[fig-N4rgbIII\](a)). Observations of spectroscopy of ionizing stars [@mar10] are expected to reveal their natures.
Discussion {#s4}
==========
Star Formation in Clumps {#s-sfc}
------------------------
The good agreement of the PDRs with cold dust emission and the associated YSO candidates therein are good indicators of the strong impact of the H[ii]{} region on its surroundings and on star formation processes. In what follows, we investigate the six identified dust clumps in a context of star formation within them.
In the case of the Aquila Rift complex, a column density threshold of $N_{\mathrm{H_{2}}} > 7\times10^{21}$ cm$^{-2}$ was estimated for the formation of prestellar cores [@and11]. For the six clumps in N4, the mean column density of $1.4\times10^{22}$ cm$^{-2}$ suggests that they may be capable of forming stars. A mass-size relationship is generally adopted to predict whether the clumps will form low-mass stars or high-mass stars [@kau10]. Based on the statistical analysis of nearby clouds without high-mass star formation (i.e., Pipe Nebula, Taurus, Perseus, and Ophiuchus) and known samples with high-mass star formation such as those of @beu02 [@mue02; @hil05], and @mot07, a limiting mass-size relation of $m(r)\geq 870$ $(r/\mathrm{pc})^{1.33}$ was found to be an approximate threshold for high-mass star formation [@kau10]. Figure \[fig-kpmass\] shows the mass versus size relation for the six clumps in N4. Given uncertainties of the two parameters, the two clumps (Clumps B and C) are above the threshold, indicating that they may be forming massive stars, whereas the other four clumps below the threshold may be inclined to form low-mass stars. Indeed, several YSO candidates are close to all clumps but Clump F and six of these (Y22, Y33, Y34, Y36, Y45 and Y59, see Figure \[fig-N4rgbIII\]) are classified as Class I YSOs, however, there is no YSO candidate located at the peak positions of the six clumps. These facts suggest that the interiors of six clumps may be in the process of forming stars but without active ongoing star-forming activities.
No clear evidence for ongoing star formation has also been demonstrated in the observations of the four molecular lines toward the peak positions of the six clumps. The 22 GHz H$_{2}$O and 44 GHz Class I CH$_{3}$OH masers are known to occur in low- and/or high-mass star formation regions [e.g., @for99; @kur04]. These two types of masers are generally thought to be associated with molecular outflows [e.g., @cod04; @kur04]. Optically thick lines of HCO$^{+}$ and H$_{2}$CO are widely used to search for asymmetric line profiles indicative of infall or outflow motions [e.g., @wu07; @che10 and references therein]. Therefore, these four molecular lines are an important signpost of ongoing star formation. In the bubble N4, it turns out that no detection of the 22 GHz H$_{2}$O and 44 GHz Class I CH$_{3}$OH masers has been obtained toward the peak positions of all clumps, and the asymmetric red profiles (i.e., a double peak spectral profile with red peak higher than blue peak) have been detected only in Clump C by the HCO$^{+}$ and H$_{2}$CO lines (see Figure \[fig-coreSp\]).
As a result, no scenario of ongoing star formation is confirmed in the center of any of the six clumps except for Clump C. On the one hand, this result may be a consequence of the poor angular resolution of the present observations (23-120$\arcsec$, corresponding to 0.26-1.35 pc). On the other hand, it can result from the incompleteness of the current sample of YSO candidates. Therefore, we have correlated the six clumps with the YSOs in the RMS survey database [@lum13], the sources in the Methanol Multi-Beam survey [@gre10], and the compact H[ii]{} region catalog in the CORNISH survey [@pur13]. It turns that there is only one source found in the RMS survey, which corresponds to the YSO candidate Y45 (see Figure \[fig-N4rgbIII\]). In addition, extended green objects (EGOs) identified based on their extended 4.5 emission are thought to be massive YSO outflow candidates [@cyg08]. Figure \[fig-N4rgbIII\](b) shows *Spitzer* 4.5 emission overlaid with cold dust radiation at 250 . One can see that six clumps except for Clumps C and D peak at dark 4.5 emission, and no brighter extended 4.5 structures with respect to their backgrounds are detected for Clumps C and D, indicating that there may be either no active ongoing star formation or deeply embedded protostars at the centers of these clumps.
Due to the aforementioned lack of the H$_{2}$O, Class I CH$_{3}$OH maser detections, and other signs of ongoing star formation, it is difficult to confirm that the asymmetric red profiles are produced by outflow motions. The fact that Clump C is located close to the strong ionization zone implies that it is fully exposed to intense feedback from the main O8.5V-O9V star such as the compression of the ionized region and stellar winds, forming the different two different velocity components and leading to the asymmetric line profiles. Alternatively, this type of line profile may originate from the clump rotation. Hence, it is worthwhile to determine the mechanism responsible for the asymmetric line profile in Clump C using a map observation with higher angular resolution.
![Mass-size relationship of the six clumps. The dashed line represents an empirical relation of $m(r) = 870$ $(r/\mathrm{pc})^{1.33}$, which is expected to be as a threshold for forming high-mass stars [@kau10].[]{data-label="fig-kpmass"}](size_mass.pdf){width="45.00000%"}
Feedback From Massive Stars {#s-feedback}
---------------------------
Massive stars are expected to affect their surroundings. Inside the bubble, the hot dust traced by 24 emission has been presumably heated by absorption of Lyman continuum photons whose free-free radiation is detected through radio continuum emission at 20 cm. At the edge of the bubble, the shell of cold dust emission is associated with the PDR that delimits the H[ii]{} region, indicating the bright illumination of the far UV photons created by massive stars. In the northeast of the bubble, brighter PAH emission is seen extending beyond the IF. This can be attributed to far UV photons leaking from the H[ii]{} region due to small-scale inhomogeneities in the IF and in the surrounding medium [@zav07]. Likewise, the southwestern filament with bright PAH emission can also be attributed to leaking radiation. All of the above observed facts demonstrate the strong influence of the expanding H[ii]{} region on the surroundings. Therefore, the YSO candidates associated with them must suffer the physical action of the H[ii]{} region. Moreover, 8 emission shows concave-convex characteristics, indicating that the IF is locally distorted. The outline of the cold dust emission, and likewise that of the IF, shows that this type of distortion occurs in the adjacent clouds. In this case such distortion could be attributed to compressions of the expanding IF. Such compression can be characterized by the probability density function (PDF) of the column density.
In turbulent simulations [@tre12a; @tre12b], it is argued that when the ionized gas pressure overweighs the ram pressure of the turbulence, the ionization compression could produce PDF forms with two lognormal distributions: $$\label{eq-pdf}
p(\eta)=\frac{p_{0}}{\sqrt{2\pi\sigma_0^2}}\exp\left(\frac{-(\eta-\mu_0)^2}{2\sigma_0^2}\right)+
\frac{p_{1}}{\sqrt{2\pi\sigma_1^2}}\exp\left(\frac{-(\eta-\mu_1)^2}{2\sigma_1^2}\right),$$ where $\eta = \mathrm{ln}(N/\bar{N})$, $p_{i}$, $\mu_i$ and $\sigma_i$ are the integral, mean, and dispersion of each component, respectively. $\bar{N}$ represents the mean column density over a large region. The first lognormal form at low densities is generally thought to be a result of isothermal supersonic turbulence [@vaz94; @vaz08; @fed08; @fed10]. The second lognormal distribution in the PDF at high densities is believed to be caused by the compression from the ionized gas pressure [@tre12a; @tre12b; @tre14]. Currently, this sort of PDF form has been observed in several massive star-forming regions, and interpreted as ionization compressions [e.g., @sch12; @tre14].
Therefore, the exploration of the PDF form in the bubble N4 can help to improve our understanding of the influences of the H[ii]{} region on the surroundings. The column density PDF is investigated over three regions (i.e., Circles 1, 2 and 3, see Figure \[fig-NT\]) covering the ionized gas and the major surroundings. The three regions are concentric with an equal separation of 0.01 degree. Figure \[fig-pdf\] displays the column density PDFs toward the three regions well fitted by the sum of two lognormal distributions. The mean column density $\bar{N}=8.0\times10^{21}$ 2 is averaged over the region of Circle 3. The fitting parameters can be found in Table \[tbl-6\]. As shown in Figure \[fig-pdf\], all the column density PDFs of the three regions show the second lognormal forms. Given the association of the H[ii]{} region with the bubble N4, these forms might be caused by compression of ionized gas. We note that the regions smaller than the Circle 1 do not fit two lognormal distributions well. This may be due to the fact that in this region there are not enough sample points. Furthermore, we observe that the integral of the second lognormal component ($p_{1}$) decreases as the radius of the target region increases. As suggested by @tre14, this trend could be due to the fact that the larger the region is, the less important the amplitude of the compressed lognormal becomes since more unperturbed gas is added to the distribution, indicative of less significance of compression of the ionized gas in larger regions.
[ccccccc]{}\[tbh!\]
1& 0.035 & -0.091 & 0.200 & 0.006 & 0.430 & 0.238\
2& 0.035 & -0.078 & 0.181 & 0.005 & 0.383 & 0.271\
3& 0.040 & -0.081 & 0.200 & 0.003 & 0.488 & 0.221
![Column density PDFs over the target regions (see Figure \[fig-NT\]). $\overline{N_{\mathrm{H_{2}}}}$ represents mean the density. The best lognormal PDF fit (solid curve) is done by the sum of two lognormal distributions (dashed curves, see Eq. \[eq-pdf\]). Error bars are determined via Poisson statistics.[]{data-label="fig-pdf"}](PDF_large008.pdf "fig:"){width="45.00000%"} ![Column density PDFs over the target regions (see Figure \[fig-NT\]). $\overline{N_{\mathrm{H_{2}}}}$ represents mean the density. The best lognormal PDF fit (solid curve) is done by the sum of two lognormal distributions (dashed curves, see Eq. \[eq-pdf\]). Error bars are determined via Poisson statistics.[]{data-label="fig-pdf"}](PDF_large009.pdf "fig:"){width="45.00000%"} ![Column density PDFs over the target regions (see Figure \[fig-NT\]). $\overline{N_{\mathrm{H_{2}}}}$ represents mean the density. The best lognormal PDF fit (solid curve) is done by the sum of two lognormal distributions (dashed curves, see Eq. \[eq-pdf\]). Error bars are determined via Poisson statistics.[]{data-label="fig-pdf"}](PDF_large01.pdf "fig:"){width="45.00000%"}
Collect and Collapse Process {#s-cc}
----------------------------
The association of collected molecular gas with PDR regions and the compression from ionized gas confirm the strong influence of the H[ii]{} region on the adjacent medium. In the context of the C & C process, such an impact may stimulate the formation of a new generation of stars. To test the occurrence of this process, the dynamical time of the H[ii]{} region can be compared with the fragmentation time of the surrounding molecular clouds [@deh03; @deh05; @zav06; @zav07; @zav10].
According to the well-known expansion law [@spi78; @dys80], if an H[ii]{} region evolves in an homogeneous molecular cloud, the dynamical age ($t_\mathrm{dyn}$) of the H[ii]{} region can be expressed as follows: $$\begin{aligned}
\label{eq-9}
\nonumber
R_{\mathrm{St}} &=& (3N_{\mathrm{Ly}}/4\pi n_{\mathrm{H},0}^{2}\alpha_{\mathrm{B}})^{1/3}, \\
t_{\mathrm{dyn}}&=&\frac{4}{7}\ \frac{R_{\mathrm{St}}}{C_{\mathrm{H{\,\sc{\mathrm{ii}}}}}}\ [(\frac{R_{\mathrm{IF}}}{R_{\mathrm{St}}})^{7/4}-1],
\end{aligned}$$ where $R_{\mathrm{St}}$ is the radius of the Str$\mathrm{\ddot{o}}$mgren sphere, $n_{0}$ is the initial particle density of ambient neutral gas, $\alpha_{B}$ = $2.6\times10^{-13}(10^{4} \mathrm{K}/T)^{0.7}$ cm$^{3}$ s$^{-1}$ [@kwa97] is the the coefficient of the radiative recombination, $C_{\mathrm{H{\,\sc{\mathrm{ii}}}}}$ is the isothermal sound speed of ionized gas assumed to be 10 km s$^{-1}$, and $R_{\mathrm{IF}}$ is the radius of the IF. If we assume that $R_{\mathrm{IF}}$ is approximately equal to the radius of N4, the total mass of $\sim555$ of the neutral and ionized gas can yield an initial number density of $n_{\mathrm{H},0}= \sim1.2\times10^{4}$ . Given the estimated $N_{\mathrm{Ly}}$ (see Section \[s-HII\]), $t_\mathrm{dyn}$ of N4 is therefore $\sim1.0$ Myr. This timescale, however, is uncertain since the actual evolution of the H[ii]{} regions is not in a strictly uniform medium. Bearing this in mind, the estimated dynamical age should be considered with this caveat.
On the other hand, following @whi94, gravitational fragmentation of the shell of collected material can be expected when $$\label{eqa11}
t_{\mathrm{frag}}=1.56 \ a_{.2}{}^{7/11} N _{49}{}^{-1/11} n_3{}^{-5/11},$$ where $a_{.2}$ is the sound speed ($a_{s}$) inside the shell in units of 0.2 , $N_{49}$ is the ionizing photon flux ($N_{Ly}$) in units of $10^{49}$ ph s$^{-1}$, and $n_{3}$ is the initial gas atomic number density ($n_{\mathrm{H},0}$) in units of $10^{3}$ cm$^{-3}$. An estimate of $a_{s}=0.3$ at a dust temperature of 22 K gives $t_{\mathrm{frag}}=\sim0.3$ Myr.
In comparison, the fragmentation time is much shorter than the dynamical age. This indicates that the shell of collected gas has had enough time to fragment during the lifetime of N4, which is consistent with the signature of six dust fragments condensed out of the shell. Therefore, the C & C process is presumably at work in the bubble N4. In addition, the spatial distribution of some YSO candidates shows an overdensity of the number of YSOs at the edge of the bubble (see Figure \[fig-N4rgbIII\] and \[fig-ysoStat\]). In combination with such an overdensity and the existence of ionization compression, the aforementioned timescales show that a scenario of triggered star formation in the bubble N4 through the C & C process is possible.
Summary {#s5}
=======
Taking advantage of observations of *Herschel* and of four molecular lines (i.e., H$_{2}$O $6_{1,6}-5_{2,3}$, CH$_{3}$OH $7_{0,7}-6_{1,6}$, HCO$^{+}$ (1-0), and o-H$_{2}$CO $2_{1,2}-1_{1,1}$), together with auxiliary archival data involving four public surveys (i.e., GLIMPSE, MIPSGAL, *WISE*, 2MASS, and MAGPIS), we have investigated the interactions of the bubble N4 with the adjacent medium and explored the possibility of triggered star formation. The main results are summarized below.
1. The distributions of the dust temperature and column density toward N4 show an anti-correlation: the higher the column density, the colder the dust temperature. This is attributed to the penetration degree of the external heating from the associated H[ii]{} region.
2. A shell structure standing out of the column density map harbors six dense dust clumps. These clumps have a mean size of $\sim0.5$ pc, temperature of $\sim22$ K, column density of $\sim1.7\times10^{22}$ 2, volume density of $\sim4.4\times10^{3}$ cm$^{-3}$, and mass of $\sim3.2\times10^{2}$ . Two out of the six may be massive enough to form high-mass stars, while the remaining could form low-mass stars. At the present sensitivity and angular resolution, the observations of the four molecular lines toward the six clumps did not reveal a clear evidence of ongoing star formation.
3. The H[ii]{} region associated with the bubble is likely to be excited by an O8.5V-O9V type star with a dynamical age of $\sim 1.0$ Myr. The velocity difference between the southeastern clumps and the northwestern ones as shown in the spectra of HCO$^{+}$ and o-H$_{2}$CO, suggests that the bubble is expanding.
4. The shell of cold dust emission associated with PDRs, and the compressions of ionized gas characterized by the column density PDF of the target regions demonstrate that the expanding bubble has a strong influence on the its surroundings.
5. In the context of the C & C mechanism, we find that the shell of collected matter has enough time to fragment during the lifetime of N4. From compression of ionized gas, the overdensity of the number of YSO candidates at the edge of the bubble, and the timescales involved, we suggest that triggered star formation might have taken place in the bubble N4 but its definitive demonstration requires more detailed molecular lines observations.
We thank the anonymous referee for the comments that much improved the quality of this paper. This work is supported by the National Natural Science Foundation of China through grants of 11503035, 11573036, 11373009, 11433008, and 11403040; the International S&T Cooperation Program of China through grand of 2010DFA02710; and the Beijing Natural Science Foundation through the grant of 1144015; and the Chinese Academy of Sciences. G.D., S.P., and M.O. acknowledge support from ANPCyT and CONICET (Argentina) grants. SPIRE has been developed by a consortium of institutes led by Cardiff Univ. (UK) with Univ. Lethbridge (Canada); NAOC (China); CEA, LAM (France); IFSI, Univ. Padua (Italy); IAC (Spain); Stockholm Observatory (Sweden); Imperial College London, RAL, UCL-MSSL, UKATC, Univ. Sussex (UK); Caltech, JPL, NHSC, Univ. Colorado (USA). This development has been supported by national funding agencies: CSA (Canada); NAOC (China); CEA, CNES, CNRS (France); ASI (Italy); MCINN (Spain); SNSB (Sweden); STFC (UK); and NASA (USA). PACS has been developed by a consortium of institutes led by MPE (Germany) with UVIE (Austria); KU Leuven, CSL, IMEC (Belgium); CEA, LAM(France); MPIA (Germany); INAFIFSI/ OAA/OAP/OAT, LENS, SISSA (Italy); IAC (Spain). This development has been supported by the funding agencies BMVIT (Austria), ESA-PRODEX (Belgium), CEA/CNES (France), DLR (Germany), ASI/INAF (Italy), and CICYT/MCYT (Spain). We are grateful to the KVN staff. The KVN is a facility operated by the Korea Astronomy and Space Science Institute. We have used the NASA/IPAC Infrared Science Archive to obtain data products from the *Spitzer*-GLIMPSE, *Spitzer*-MIPSGAL, *WISE*, and 2MASS surveys.
Identification of YSO candidates {#s-app}
================================
[ll]{}\[tbh!\]
Selected point sources (ALLWISE and MIPSGAL catalogs) & 191\
Good fittings of YSO SEDs ($\chi^{2}_{best}/N_{data}\leq7$) & 77\
PAH-feature emission & 17\
Final YSO candidadates & 60
To our knowledge, if more data at longer wavelengths are taken into account, the SED fitting will be better constrained. For example, Figure \[fig-preyso\] shows the resulting SED fits with the same parameter inputs except for different data inputs. Obviously, the SED fitting results with the 70 data inputs (Figure \[fig-preyso\] (c)) are best constrained, followed by the cases with the 22 but without 70 data (Figure \[fig-preyso\] (b)) inputs, whereas the cases without the data inputs at both wavelengths (Figure \[fig-preyso\] (a)) are most poorly constrained. Therefore, the *ALLWISE* and MIPSGAL catalogs were taken into account. The former catalog includes photometries at 22 and the latter one includes photometries at 24 . In addition, these two catalogs have been cross matched with the 2MASS and/or GLIMPSE surveys, ensuring more photometries to be used in the SED fitting.
From the two catalogs, 191 point sources were picked out. First, we retrieved a sample of 164 sources detected with a ratio of signal-to-noise (S/N) of $>10$ in the four *WISE* bands and with an extension flag of $<2$ from the *ALLWISE* Source catalog. The former simply ensures sources with detections at more wavelengths and with good enough data qualities and the latter guarantees the sources as a point source. Also, this sample has been complemented by that retrieved from the MIPSGAL catalog. In this catalog, a sample of 32 point sources was chosen with a photometry uncertainty of $<0.2$ in the MIPS 24 and the four IRAC bands. These restrictions are equivalent to those mentioned previously. Except for 5 sources overlapping in these two samples, a total of 191 point sources were obtained. In addition, we tried searching for photometries at 70 for these selected sources. After investigating several published literatures, only one object, J180858.77-181629.7, has been found with 70 photometries [@lum13]. Moreover, the 56 point sources from the CuTeX catalog (see Section \[s-ob1\]) were also inspected by a cross matching with the chosen 191 sources within a search radii of $2\arcsec$ in the software Topcat.[^8] In a combination of visual inspections, only five photometries at 70 from the CuTeX catalog were considered in the following SED fittings. In total, 6 out of 191 point sources have 70 photometries.
In the fittings to the selected 191 sources, the distance was constrained in the range of 2.3-4.1 kpc, and the extinction was in the range of 0 to 16 Mag, which was estimated from the column density map by a relation of $A_{\mathrm{v}} = 5.34 \times 10^{-22}
N(\mathrm{H_2})$ [@deh09]. After the SED fitting to the sources, the YSO candidates were determined by the following criteria. That is, the candidates should be those (1) with $\chi^{2}_{best}/N_{\mathrm{data}}\leq7$ and (2) fitted well only by the YSO models, not by a model of a stellar photosphere with a foreground extinction. The former criterion has been determined by visually inspecting dozens of profiles of resulting SED fittings. We note that the value of 7 is a little greater than those adopted in the literature [e.g., @pov09; @pov13; @kan09]. This could be due to the fact that more data points ($N_{\mathrm{data}}$) have been taken into account in this work. Figure \[fig-preyso\] exemplifies a source satisfied with the above criteria. The candidate J180858.77-181629.7 has a good fit of the YSO models (Figure \[fig-preyso\] (c)) but a poor fit of a stellar photosphere (Figure \[fig-preyso\] (d)), indicating that this object is not a field star. These criteria result in 77 YSO candidates including 74 sources from the *ALLWISE* catalog and the remaining 3 from the MIPSGAL catalog.
The color selection schemes were used to categorize the 77 YSO candidates. To begin, the classification method of @koe12 was adopted for the 74 candidates. According to the criteria in various color spaces, which can kick out possible contaminants such as star-forming galaxies, broad-line AGNs, and knots of shock emission, 16 PAH-contaminated emission sources were removed. The remaining uncontaminated sources are considered to be Class I YSOs if their colors follow (1) $[3.4]-[4.6]>1.0$ and (2) $[4.6]-[12]>2.0$; Class II YSOs if their colors follow (1) $[3.4]-[4.6]-\sigma_{12}>0.25$ and (2) $[4.6]-[12]-\sigma_{23}>1.0$, where $ \sigma_{12}, \sigma_{23}$ are the combined measurement errors of two corresponding *WISE* bands (i.e., \[3.4\] and \[4.6\], \[4.6\] and \[12\], respectively); and TD YSOs if their colors follow (1) $[4.6]-[22]>2.5$ and (2) $[3.4]<14$. In the end, 7 Class I, 11 Class II, and 40 TD YSOs are obtained by the above constraints, which are also shown in Figure \[fig-cc\] (a)-(b). For the three candidates from the MIPSGAL catalog, the color selection method of @gut09 was considered. Similarly, one PAH-feature emission was ruled out. The remaining 2 sources are classified into one Class I YSO since its colors follow (1) $[4.5]-[5.8]>0.7$, and (2) $[3.6]-[4.5]>0.7$; and one Class II YSO since its colors follow (1) $[4.5]-[8.0]-\sigma_{24}>0.5$, (2) $[3.6]-[5.8]-\sigma_{13}>0.35$, (3) $[3.6]-[5.8]+\sigma_{13}\leq \frac{0.14}{0.04} \times (([4.5]-[8.0]-\sigma_{24})-0.5)+0.5$, and (4) $[3.6]-[4.5]-\sigma_{12}>0.35$, where $\sigma_{24}, \sigma_{13}, \sigma_{12}$ are the combined measurement errors of two corresponding IRAC bands (i.e., \[4.5\] and \[8.0\], \[3.6\] and \[5.8\], \[3.6\] and \[4.5\], respectively). The above constraints are displayed in Figure \[fig-cc\](c)-(d). In total, built on the 77 SED-identified YSO candidates, the color schemes confirm 60 out of them as potential YSOs: 8 Class I, 12 Class II and 40 TD objects. The results of the classification can be found in Table \[tbl-5\]. Table \[tbl-9\] summarizes the number of sources in each step for identifying YSO candidates. Their photometries are tabulated in Table \[tbl-10\] and Table \[tbl-11\]. The plots of the SED fitting to the 60 YSO candidates are available in Figures 10 and 11.
The resulting 60 potential YSOs are incomplete toward N4 but robust. The strict criteria on the aforementioned source selections result in the incompleteness. For instance, if the S/Ns of photometries at the selected wavelengths were set to 5, there would be more sources to be in question, resulting in more YSO candidates. However, the lower S/Ns are directly related to the photometry qualities, reducing the robustness of the resulting YSO candidates. For the sake of the robustness, the strict criteria were be carried out. Additionally, the combination of the SED fitting method with the color selection schemes makes the 60 YSO candidates reliable enough.
[llllllll]{}\[tbh!\]
J180831.89-181639.1 & 10.697 $\pm$ 0.042 & 10.311 $\pm$ 0.031 & 7.278 $\pm$ 0.083 & 3.314 $\pm$ 0.055 & 16.589 $\pm$ – & 13.393 $\pm$ 0.022 & 11.821 $\pm$ 0.027\
J180832.49-181616.6 & 10.541 $\pm$ 0.03 & 10.434 $\pm$ 0.032 & 7.963 $\pm$ 0.045 & 3.337 $\pm$ 0.028 & 11.7 $\pm$ 0.026 & 11.194 $\pm$ 0.024 & 10.823 $\pm$ 0.024\
J180834.88-181751.0 & 10.074 $\pm$ 0.031 & 9.927 $\pm$ 0.031 & 6.823 $\pm$ 0.048 & 5.931 $\pm$ 0.067 & 13.385 $\pm$ 0.029 & 11.75 $\pm$ 0.037 & 11.113 $\pm$ 0.034\
J180835.68-181814.8 & 9.551 $\pm$ 0.025 & 9.47 $\pm$ 0.025 & 7.305 $\pm$ 0.049 & 5.895 $\pm$ 0.108 & 13.831 $\pm$ 0.049 & 12.006 $\pm$ – & 10.925 $\pm$ –\
J180839.36-181813.8 & 11.207 $\pm$ 0.056 & 11.146 $\pm$ 0.07 & 7.554 $\pm$ 0.101 & 4.809 $\pm$ 0.087 & 15.853 $\pm$ 0.065 & 13.108 $\pm$ 0.03 & 12.014 $\pm$ 0.021\
J180839.71-181841.5 & 11.005 $\pm$ 0.05 & 10.881 $\pm$ 0.061 & 6.792 $\pm$ 0.046 & 4.154 $\pm$ 0.036 & 16.762 $\pm$ – & 13.47 $\pm$ 0.037 & 11.907 $\pm$ 0.026\
J180840.11-181902.3 & 11.032 $\pm$ 0.042 & 10.849 $\pm$ 0.057 & 8.29 $\pm$ 0.09 & 4.622 $\pm$ 0.043 & 15.16 $\pm$ – & 13.997 $\pm$ 0.091 & 12.005 $\pm$ –\
J180841.62-181600.0 & 10.152 $\pm$ 0.031 & 10.41 $\pm$ 0.029 & 7.238 $\pm$ 0.042 & 4.466 $\pm$ 0.097 & 11.908 $\pm$ 0.023 & 10.873 $\pm$ 0.021 & 10.554 $\pm$ 0.019\
J180842.08-181457.5 & 10.534 $\pm$ 0.037 & 10.949 $\pm$ 0.043 & 7.628 $\pm$ 0.049 & 4.491 $\pm$ 0.04 & 12.144 $\pm$ 0.023 & 11.197 $\pm$ 0.022 & 10.89 $\pm$ 0.026\
J180842.19-181329.1 & 10.587 $\pm$ 0.034 & 10.666 $\pm$ 0.036 & 8.87 $\pm$ 0.062 & 6.489 $\pm$ 0.102 & 14.399 $\pm$ 0.043 & 12.301 $\pm$ 0.042 & 11.352 $\pm$ 0.031\
J180842.69-181930.1 & 10.072 $\pm$ 0.026 & 10.023 $\pm$ 0.03 & 8.238 $\pm$ 0.101 & 5.769 $\pm$ 0.076 & 15.183 $\pm$ 0.058 & 12.526 $\pm$ 0.07 & 11.322 $\pm$ 0.045\
J180843.58-181441.8 & 9.652 $\pm$ 0.035 & 9.506 $\pm$ 0.03 & 5.882 $\pm$ 0.021 & 3.244 $\pm$ 0.03 & 15.211 $\pm$ 0.053 & 12.255 $\pm$ 0.037 & 10.903 $\pm$ 0.024\
J180844.11-181456.4 & 11.345 $\pm$ 0.045 & 11.696 $\pm$ 0.051 & 7.217 $\pm$ 0.028 & 2.897 $\pm$ 0.024 & 16.351 $\pm$ – & 14.165 $\pm$ 0.055 & 12.771 $\pm$ 0.055\
J180844.15-181401.9 & 10.359 $\pm$ 0.03 & 10.617 $\pm$ 0.032 & 6.835 $\pm$ 0.042 & 4.243 $\pm$ 0.041 & 16.05 $\pm$ 0.094 & 13.476 $\pm$ 0.035 & 12.264 $\pm$ 0.031\
J180844.97-181457.3 & 11.433 $\pm$ 0.053 & 11.782 $\pm$ 0.048 & 7.052 $\pm$ 0.036 & 2.66 $\pm$ 0.03 & 16.942 $\pm$ – & 14.029 $\pm$ 0.035 & 12.598 $\pm$ 0.033\
J180845.97-181612.4$^{(a)}$ & 10.931 $\pm$ 0.083 & 10.484 $\pm$ 0.055 & 5.624 $\pm$ 0.053 & 3.33 $\pm$ 0.047 & 16.876 $\pm$ – & 13.625 $\pm$ 0.035 & 12.049 $\pm$ 0.024\
J180846.14-181844.0 & 10.287 $\pm$ 0.055 & 10.479 $\pm$ 0.05 & 5.907 $\pm$ 0.018 & 3.516 $\pm$ 0.03 & 15.098 $\pm$ 0.035 & 12.466 $\pm$ 0.021 & 11.317 $\pm$ 0.023\
J180847.47-181416.2 & 10.423 $\pm$ 0.036 & 10.442 $\pm$ 0.031 & 6.884 $\pm$ 0.031 & 2.591 $\pm$ 0.033 & 16.78 $\pm$ – & 13.575 $\pm$ 0.055 & 12.05 $\pm$ 0.045\
J180847.65-181450.0 & 10.064 $\pm$ 0.053 & 9.849 $\pm$ 0.06 & 5.602 $\pm$ 0.054 & 3.312 $\pm$ 0.084 & 15.599 $\pm$ 0.061 & 12.603 $\pm$ 0.046 & 11.213 $\pm$ 0.039\
J180847.67-181441.7 & 10.371 $\pm$ 0.067 & 10.074 $\pm$ 0.07 & 5.573 $\pm$ 0.063 & 2.92 $\pm$ 0.06 & 16.419 $\pm$ – & 13.655 $\pm$ 0.034 & 11.849 $\pm$ 0.026\
J180848.33-181357.2 & 9.8 $\pm$ 0.042 & 9.635 $\pm$ 0.045 & 6.494 $\pm$ 0.047 & 3.071 $\pm$ 0.042 & 16.221 $\pm$ 0.085 & 12.853 $\pm$ 0.079 & 11.306 $\pm$ 0.058\
J180849.08-181441.8 & 11.845 $\pm$ 0.09 & 10.265 $\pm$ 0.056 & 5.752 $\pm$ 0.036 & 2.808 $\pm$ 0.05 & – & – & –\
J180849.78-181340.5 & 8.674 $\pm$ 0.027 & 8.714 $\pm$ 0.028 & 6.506 $\pm$ 0.024 & 2.82 $\pm$ 0.034 & 10.041 $\pm$ 0.026 & 9.383 $\pm$ 0.032 & 9.162 $\pm$ 0.03\
J180850.19-181816.2 & 9.1 $\pm$ 0.028 & 8.73 $\pm$ 0.026 & 5.591 $\pm$ 0.061 & 3.453 $\pm$ 0.071 & 14.469 $\pm$ 0.037 & 11.537 $\pm$ 0.024 & 10.137 $\pm$ 0.024\
J180850.83-181836.6 & 9.318 $\pm$ 0.03 & 9.01 $\pm$ 0.03 & 5.884 $\pm$ 0.036 & 2.808 $\pm$ 0.034 & 15.385 $\pm$ – & 12.583 $\pm$ 0.022 & 10.647 $\pm$ 0.027\
J180851.55-181831.9 & 8.991 $\pm$ 0.026 & 8.963 $\pm$ 0.026 & 5.555 $\pm$ 0.024 & 2.763 $\pm$ 0.034 & 9.631 $\pm$ 0.024 & 9.25 $\pm$ 0.021 & 9.161 $\pm$ 0.021\
J180852.42-181744.2 & 9.108 $\pm$ 0.05 & 8.729 $\pm$ 0.028 & 5.133 $\pm$ 0.024 & 2.65 $\pm$ 0.079 & 15.479 $\pm$ – & 12.151 $\pm$ 0.032 & 10.418 $\pm$ 0.027\
J180852.59-181356.6 & 10.289 $\pm$ 0.049 & 9.885 $\pm$ 0.04 & 5.064 $\pm$ 0.054 & 1.387 $\pm$ 0.019 & 15.335 $\pm$ 0.06 & 12.364 $\pm$ – & 11.101 $\pm$ –\
J180852.97-181336.2 & 9.418 $\pm$ 0.027 & 9.402 $\pm$ 0.03 & 6.039 $\pm$ 0.042 & 1.711 $\pm$ 0.016 & 14.05 $\pm$ 0.033 & 11.437 $\pm$ 0.037 & 10.218 $\pm$ 0.029\
J180853.04-181742.2$^{(b)}$ & 10.116 $\pm$ 0.054 & 8.856 $\pm$ 0.03 & 4.771 $\pm$ 0.025 & 2.119 $\pm$ 0.044 & 13.289 $\pm$ 0.036 & 12.471 $\pm$ 0.049 & 11.942 $\pm$ 0.051\
J180853.35-181130.9 & 11.43 $\pm$ 0.04 & 11.741 $\pm$ 0.035 & 8.139 $\pm$ 0.085 & 5.616 $\pm$ 0.071 & 16.122 $\pm$ 0.079 & 13.566 $\pm$ 0.037 & 12.434 $\pm$ 0.037\
J180853.65-181208.4 & 11.036 $\pm$ 0.045 & 11.233 $\pm$ 0.046 & 8.071 $\pm$ 0.051 & 5.005 $\pm$ 0.089 & 15.465 $\pm$ 0.064 & 13.051 $\pm$ 0.03 & 12.031 $\pm$ 0.026\
J180853.94-181422.4$^{(c)}$ & 10.641 $\pm$ 0.052 & 9.607 $\pm$ 0.036 & 4.724 $\pm$ 0.047 & 1.55 $\pm$ 0.037 & – & – & –\
J180854.47-181406.5 & 10.503 $\pm$ 0.046 & 9.409 $\pm$ 0.04 & 4.15 $\pm$ 0.022 & 0.32 $\pm$ 0.009 & – & – &\
J180854.69-181230.5 & 10.053 $\pm$ 0.04 & 9.929 $\pm$ 0.035 & 7.437 $\pm$ 0.096 & 5.093 $\pm$ 0.071 & 13.979 $\pm$ 0.024 & 11.683 $\pm$ 0.021 & 10.695 $\pm$ 0.021\
J180854.87-181409.5 & 10.185 $\pm$ 0.041 & 9.126 $\pm$ 0.035 & 4.286 $\pm$ 0.03 & 0.956 $\pm$ 0.03 & – & – & –\
J180855.80-181136.3 & 12.069 $\pm$ 0.064 & 12.526 $\pm$ 0.074 & 7.954 $\pm$ 0.072 & 5.717 $\pm$ 0.093 & 15.822 $\pm$ 0.087 & 13.738 $\pm$ 0.074 & 12.892 $\pm$ 0.045\
J180856.26-181505.7 & 10.283 $\pm$ 0.058 & 9.398 $\pm$ 0.043 & 5.627 $\pm$ 0.075 & 0.797 $\pm$ 0.062 & 15.32 $\pm$ – & 13.878 $\pm$ 0.062 & 12.337 $\pm$ 0.035\
J180856.64-181420.4 & 9.631 $\pm$ 0.041 & 9.154 $\pm$ 0.047 & 4.51 $\pm$ 0.021 & 1.295 $\pm$ 0.022 & 16.241 $\pm$ 0.084 & 12.831 $\pm$ 0.028 & 11.059 $\pm$ 0.024\
J180857.00-181354.7 & 11.014 $\pm$ 0.066 & 10.209 $\pm$ 0.047 & 5.523 $\pm$ 0.019 & 2.025 $\pm$ 0.022 & – & – &\
J180857.48-181853.3 & 10.859 $\pm$ 0.057 & 10.794 $\pm$ 0.066 & 7.455 $\pm$ 0.029 & 4.348 $\pm$ 0.038 & 15.821 $\pm$ 0.075 & 13.12 $\pm$ 0.068 & 11.663 $\pm$ 0.039\
J180858.07-181506.0 & 8.04 $\pm$ 0.026 & 8.086 $\pm$ 0.029 & 4.4 $\pm$ 0.016 & 1.647 $\pm$ 0.032 & 9.067 $\pm$ 0.034 & 8.456 $\pm$ 0.053 & 8.27 $\pm$ 0.02\
J180858.12-181807.2 & 10.55 $\pm$ 0.044 & 10.425 $\pm$ 0.049 & 5.956 $\pm$ 0.033 & 2.655 $\pm$ 0.033 & 14.648 $\pm$ – & 12.495 $\pm$ – & 12.088 $\pm$ 0.072\
J180858.75-181806.5 & 10.519 $\pm$ 0.042 & 10.435 $\pm$ 0.054 & 5.78 $\pm$ 0.017 & 3.046 $\pm$ 0.037 & 15.81 $\pm$ 0.065 & 12.795 $\pm$ 0.022 & 11.453 $\pm$ 0.021\
J180858.77-181629.7$^{(e)}$ & 8.532 $\pm$ 0.024 & 6.901 $\pm$ 0.022 & 2.865 $\pm$ 0.018 & 0.219 $\pm$ 0.026 & 16.035 $\pm$ – & 13.621 $\pm$ 0.045 & 11.328 $\pm$ 0.026\
J180859.32-181316.1 & 11.16 $\pm$ 0.047 & 10.745 $\pm$ 0.04 & 6.736 $\pm$ 0.091 & 4.302 $\pm$ 0.074 & 15.747 $\pm$ 0.078 & 13.754 $\pm$ 0.033 & 12.725 $\pm$ 0.033\
J180859.44-181332.8 & 8.857 $\pm$ 0.03 & 7.786 $\pm$ 0.023 & 4.949 $\pm$ 0.021 & 2.629 $\pm$ 0.027 & 14.689 $\pm$ 0.039 & 12.413 $\pm$ 0.047 & 10.703 $\pm$ 0.027\
J180859.47-181855.8 & 11.603 $\pm$ 0.047 & 11.933 $\pm$ 0.051 & 7.688 $\pm$ 0.044 & 4.464 $\pm$ 0.034 & 14.72 $\pm$ 0.049 & 13.208 $\pm$ 0.045 & 12.366 $\pm$ 0.035\
J180900.26-181347.5 & 9.892 $\pm$ 0.037 & 9.958 $\pm$ 0.038 & 6.826 $\pm$ 0.055 & 3.903 $\pm$ 0.037 & 11.435 $\pm$ 0.023 & 10.455 $\pm$ 0.022 & 10.073 $\pm$ 0.019\
J180900.28-181921.8 & 11.454 $\pm$ 0.039 & 11.642 $\pm$ 0.045 & 8.351 $\pm$ 0.081 & 5.308 $\pm$ 0.089 & 15.867 $\pm$ 0.08 & 13.298 $\pm$ 0.035 & 12.136 $\pm$ 0.027\
J180900.48-181903.2 & 11.585 $\pm$ 0.053 & 11.687 $\pm$ 0.073 & 7.693 $\pm$ 0.068 & 5.275 $\pm$ 0.094 & 16.271 $\pm$ 0.104 & 14.157 $\pm$ 0.085 & 13.071 $\pm$ 0.055\
J180902.92-181316.2 & 9.691 $\pm$ 0.028 & 9.827 $\pm$ 0.034 & 8.412 $\pm$ 0.045 & 5.596 $\pm$ 0.078 & 11.372 $\pm$ 0.022 & 10.364 $\pm$ 0.022 & 9.922 $\pm$ 0.019\
J180904.56-181638.3 & 9.987 $\pm$ 0.032 & 9.923 $\pm$ 0.034 & 5.869 $\pm$ 0.015 & 3.132 $\pm$ 0.028 & 14.576 $\pm$ 0.032 & 12.207 $\pm$ 0.023 & 11.139 $\pm$ 0.021\
J180904.95-181806.2 & 10.578 $\pm$ 0.035 & 10.54 $\pm$ 0.041 & 8.39 $\pm$ 0.094 & 5.864 $\pm$ 0.095 & 15.07 $\pm$ 0.083 & 12.745 $\pm$ 0.065 & 11.608 $\pm$ 0.05\
J180905.60-181621.4$^{(d)}$ & 11.084 $\pm$ 0.066 & 10.292 $\pm$ 0.039 & 5.544 $\pm$ 0.015 & 2.607 $\pm$ 0.024 & 12.639 $\pm$ 0.023 & 11.881 $\pm$ 0.022 & 11.549 $\pm$ 0.019\
J180906.19-181639.1 & 10.814 $\pm$ 0.033 & 10.514 $\pm$ 0.039 & 6.314 $\pm$ 0.017 & 3.348 $\pm$ 0.032 & 15.044 $\pm$ 0.045 & 12.744 $\pm$ 0.038 & 11.67 $\pm$ 0.029\
J180906.60-181847.9 & 11.425 $\pm$ 0.057 & 11.481 $\pm$ 0.077 & 8.474 $\pm$ 0.059 & 6.018 $\pm$ 0.065 & 14.951 $\pm$ 0.052 & 12.994 $\pm$ 0.033 & 12.082 $\pm$ 0.03\
J180908.60-181706.1 & 11.316 $\pm$ 0.05 & 11.686 $\pm$ 0.073 & 8.709 $\pm$ 0.055 & 5.842 $\pm$ 0.054 & 15.388 $\pm$ 0.104 & 13.396 $\pm$ 0.145 & 12.225 $\pm$ 0.068
[lllllllll]{}\[tbh!\]
MG011.8545+00.7327$^{(a)}$ & 2.04 $\pm$ 0.02 & 14.739 $\pm$ 0.037 & 13.029 $\pm$ 0.046 & 11.361 $\pm$ 0.034 & 8.786 $\pm$ 0.031 & 7.866 $\pm$ 0.039 & 6.767 $\pm$ 0.033 & 5.163 $\pm$ 0.024\
MG011.9455+00.7481 & 2.78 $\pm$ 0.02 & 14.689 $\pm$ 0.039 & 12.413 $\pm$ 0.047 & 10.703 $\pm$ 0.027 & 8.567 $\pm$ 0.031 & 7.882 $\pm$ 0.041 & 7.093 $\pm$ 0.032 & 6.18 $\pm$ 0.025
![ Example of the fitting of models to an object. The filled circles and triangle symbolize the input fluxes and the upper limit, respectively. The black line shows the best fit. The gray lines show a set of fits that satisfy $\chi^{2}-\chi_{best}^{2}< 2 \times N_{data}$. The dashed line means the stellar photosphere corresponding to the central source of the best fitting model as it would look in the absence of circumstellar dust but with interstellar extinction. The panels (a)-(c) show the resulting SED fittings without the input fluxes at 24 and 70 , with the input flux at 24 , and with the input fluxes at 24 and 70 , respectively. The panel (d) is to check whether the source cannot be fit simply by a star with interstellar extinction.[]{data-label="fig-preyso"}](J18085877-1816297_a.pdf "fig:"){width="24.00000%"} ![ Example of the fitting of models to an object. The filled circles and triangle symbolize the input fluxes and the upper limit, respectively. The black line shows the best fit. The gray lines show a set of fits that satisfy $\chi^{2}-\chi_{best}^{2}< 2 \times N_{data}$. The dashed line means the stellar photosphere corresponding to the central source of the best fitting model as it would look in the absence of circumstellar dust but with interstellar extinction. The panels (a)-(c) show the resulting SED fittings without the input fluxes at 24 and 70 , with the input flux at 24 , and with the input fluxes at 24 and 70 , respectively. The panel (d) is to check whether the source cannot be fit simply by a star with interstellar extinction.[]{data-label="fig-preyso"}](J18085877-1816297_b.pdf "fig:"){width="24.00000%"} ![ Example of the fitting of models to an object. The filled circles and triangle symbolize the input fluxes and the upper limit, respectively. The black line shows the best fit. The gray lines show a set of fits that satisfy $\chi^{2}-\chi_{best}^{2}< 2 \times N_{data}$. The dashed line means the stellar photosphere corresponding to the central source of the best fitting model as it would look in the absence of circumstellar dust but with interstellar extinction. The panels (a)-(c) show the resulting SED fittings without the input fluxes at 24 and 70 , with the input flux at 24 , and with the input fluxes at 24 and 70 , respectively. The panel (d) is to check whether the source cannot be fit simply by a star with interstellar extinction.[]{data-label="fig-preyso"}](J18085877-1816297_c.pdf "fig:"){width="24.00000%"} ![ Example of the fitting of models to an object. The filled circles and triangle symbolize the input fluxes and the upper limit, respectively. The black line shows the best fit. The gray lines show a set of fits that satisfy $\chi^{2}-\chi_{best}^{2}< 2 \times N_{data}$. The dashed line means the stellar photosphere corresponding to the central source of the best fitting model as it would look in the absence of circumstellar dust but with interstellar extinction. The panels (a)-(c) show the resulting SED fittings without the input fluxes at 24 and 70 , with the input flux at 24 , and with the input fluxes at 24 and 70 , respectively. The panel (d) is to check whether the source cannot be fit simply by a star with interstellar extinction.[]{data-label="fig-preyso"}](J18085877-1816297_d.pdf "fig:"){width="24.00000%"}
![Color-color diagrams for (a) *WISE* bands 1, 2, and 3; (b) WISE bands 1, 2, and 4; (c) IRAC bands 1, 2, and 3; (d) IRAC bands 1, 2, 3, and 4. Red triangles symbolize Class I YSOs, blue pluses Class II YSOs, and cyan diamond transition disk YSOs. The arrows show an extinction vector of $A_{\mathrm{K}}$ = 5 mag following the extinction laws of @koe14 [@fla07]. The combined measurement errors are shown on the bottom right. []{data-label="fig-cc"}](wise_cc.pdf "fig:"){height="4cm" width="8.75cm"} ![Color-color diagrams for (a) *WISE* bands 1, 2, and 3; (b) WISE bands 1, 2, and 4; (c) IRAC bands 1, 2, and 3; (d) IRAC bands 1, 2, 3, and 4. Red triangles symbolize Class I YSOs, blue pluses Class II YSOs, and cyan diamond transition disk YSOs. The arrows show an extinction vector of $A_{\mathrm{K}}$ = 5 mag following the extinction laws of @koe14 [@fla07]. The combined measurement errors are shown on the bottom right. []{data-label="fig-cc"}](irac_cc.pdf "fig:"){height="3.8cm" width="8.75cm"}
Allen, L. E., Calvet, N., D’Alessio, P., et al. 2004, , 154, 363
Anderson, L. D., Zavagno, A., Deharveng, L., et al. 2012, , 542, A10
Andr[é]{}, P., Men’shchikov, A., K[ö]{}nyves, V., & Arzoumanian, D. 2011, Computational Star Formation, 270, 255
Beaumont, C. N., & Williams, J. P. 2010, , 709, 791
Beckwith, S. V. W., Sargent, A. I., Chini, R. S., & Guesten, R. 1990, , 99, 924
Benjamin, R. A., Churchwell, E., Babler, B. L., et al. 2003, , 115, 953
Bertoldi, F. 1989, , 346, 735
Beuther, H., Walsh, A., Schilke, P., et al. 2002, , 390, 289
Bisbas, T. G., W[ü]{}nsch, R., Whitworth, A. P., Hubber, D. A., & Walch, S. 2011, , 736, 142
Carey, S. J., Noriega-Crespo, A., Mizuno, D. R., et al. 2009, , 121, 76
Chen, X., Shen, Z.-Q., Li, J.-J., Xu, Y., & He, J.-H. 2010, , 710, 150
Churchwell, E., Povich, M. S., Allen, D., et al. 2006, , 649, 759
Churchwell, E., Smith, L. F., Mathis, J., Mezger, P. G., & Huchtmeier, W. 1978, , 70, 719
Churchwell, E., Watson, D. F., Povich, M. S., et al. 2007, , 670, 428
Codella, C., Lorenzani, A., Gallego, A. T., Cesaroni, R., & Moscadelli, L. 2004, , 417, 615
Cyganowski, C. J., Whitney, B. A., Holden, E., et al. 2008, , 136, 2391
Dale, J. E., Bonnell, I. A., & Whitworth, A. P. 2007, , 375, 1291
Dale, J. E., Ercolano, B., & Bonnell, I. A. 2013, , 431, 1062
Dale, J. E., Haworth, T. J., & Bressert, E. 2015, arXiv:1502.05865
Deharveng, L., Lefloch, B., Kurtz, S., et al. 2008, , 482, 585
Deharveng, L., Lefloch, B., Massi, F., et al. 2006, , 458, 191
Deharveng, L., Lefloch, B., Zavagno, A., et al. 2003, , 408, L25
Deharveng, L., Schuller, F., Anderson, L. D., et al. 2010, , 523, A6
Deharveng, L., Zavagno, A., & Caplan, J. 2005, , 433, 565
Deharveng, L., Zavagno, A., Schuller, F., et al. 2009, , 496, 177
Dyson, J. E., & Williams, D. A. 1980, New York, Halsted Press, 1980. 204 p.,
Elmegreen, B. G. 2011, EAS Publications Series, 51, 45
Elmegreen, B. G., & Lada, C. J. 1977, , 214, 725
Federrath, C., Klessen, R. S., & Schmidt, W. 2008, , 688, L79
Federrath, C., Roman-Duval, J., Klessen, R. S., Schmidt, W., & Mac Low, M.-M. 2010, , 512, A81
Flaherty, K. M., Pipher, J. L., Megeath, S. T., et al. 2007, , 663, 1069
Forster, J. R., & Caswell, J. L. 1999, , 137, 43
Goldsmith, P. F., & Langer, W. D. 1999, , 517, 209
Green, J. A., Caswell, J. L., Fuller, G. A., et al. 2010, , 409, 913
Griffin, M. J., Abergel, A., Abreu, A., et al. 2010, , 518, L3
Guilloteau, S., & Lucas, R. 2000, Imaging at Radio through Submillimeter Wavelengths, 217, 299
Gutermuth, R. A., Megeath, S. T., Myers, P. C., et al. 2009, , 184, 18
Gutermuth, R. A., Myers, P. C., Megeath, S. T., et al. 2008, , 674, 336
Guzm[á]{}n, V., Pety, J., Goicoechea, J. R., Gerin, M., & Roueff, E. 2011, , 534, A49
Helfand, D. J., Becker, R. H., White, R. L., Fallon, A., & Tuttle, S. 2006, , 131, 2525
Hill, T., Burton, M. G., Minier, V., et al. 2005, , 363, 405
Hosokawa, T., & Inutsuka, S.-i. 2006, , 646, 240
Hosokawa, T., & Inutsuka, S.-i. 2005, , 623, 917
Kang, M., Bieging, J. H., Povich, M. S., & Lee, Y. 2009, , 706, 83
Kauffmann, J., Bertoldi, F., Bourke, T. L., Evans, N. J., II, & Lee, C. W. 2008, , 487, 993
Kauffmann, J., & Pillai, T. 2010, , 723, L7
Kendrew, S., Simpson, R., Bressert, E., et al. 2012, , 755, 71
Kim, K.-T., Byun, D.-Y., Je, D.-H., et al. 2011, Journal of Korean Astronomical Society, 44, 81
Koenig, X. P., Leisawitz, D. T., Benford, D. J., et al. 2012, , 744, 130
Koenig, X. P., & Leisawitz, D. T. 2014, , 791, 131
Kurtz, S., Churchwell, E., & Wood, D. O. S. 1994, , 91, 659
Kurtz, S., Hofner, P., & [Á]{}lvarez, C. V. 2004, , 155, 149
Kwan, J. 1997, , 489, 284
Lada, C. J. 1987, Star Forming Regions, 115, 1
Lefloch, B., & Lazareff, B. 1994, , 289, 559
Li, J.-Y., Jiang, Z.-B., Liu, Y., & Wang, Y. 2013, Research in Astronomy and Astrophysics, 13, 921
Liu, H.-L., Wu, Y., Li, J., et al. 2015, , 798, 30
Liu, T., Wu, Y., Zhang, H., & Qin, S.-L. 2012, , 751, 68
Lockman, F. J. 1989, , 71, 469
Lumsden, S. L., Hoare, M. G., Urquhart, J. S., et al. 2013, , 208, 11
Mueller, K. E., Shirley, Y. L., Evans, N. J., II, & Jacobson, H. R. 2002, , 143, 469
Markwardt, C. B. 2009, Astronomical Data Analysis Software and Systems XVIII, 411, 251
Martins, F., Schaerer, D., & Hillier, D. J. 2005, , 436, 1049
Martins, F., Pomar[è]{}s, M., Deharveng, L., Zavagno, A., & Bouret, J. C. 2010, , 510, A32
Miao, J., White, G. J., Nelson, R., Thompson, M., & Morgan, L. 2006, , 369, 143
Miao, J., White, G. J., Thompson, M. A., & Nelson, R. P. 2009, , 692, 382
Molinari, S., Schisano, E., Faustini, F., et al. 2011, , 530, A133
Motte, F., Bontemps, S., Schilke, P., et al. 2007, , 476, 1243
Ogura, K. 2010, Astronomical Society of India Conference Series, 1, 19
Offner, S. S. R., Robitaille, T. P., Hansen, C. E., McKee, C. F., & Klein, R. I. 2012, , 753, 98
Paron, S., Ortega, M. E., Dubner, G., et al. 2015, , 149, 193
Poglitsch, A., Waelkens, C., Geis, N., et al. 2010, , 518, L2
Povich, M. S., Churchwell, E., Bieging, J. H., et al. 2009, , 696, 1278
Povich, M. S., Kuhn, M. A., Getman, K. V., et al. 2013, , 209, 31
Purcell, C. R., Hoare, M. G., Cotton, W. D., et al. 2013, , 205, 1
Quireza, C., Rood, R. T., Bania, T. M., Balser, D. S., & Maciel, W. J. 2006, , 653, 1226
Reach, W. T., Faied, D., Rho, J., et al. 2009, , 690, 683
Robitaille, T. P., Whitney, B. A., Indebetouw, R., Wood, K., & Denzmore, P. 2006, , 167, 256
Robitaille, T. P., Whitney, B. A., Indebetouw, R., & Wood, K. 2007, , 169, 328
Sadavoy, S. I., Di Francesco, J., Johnstone, D., et al. 2013, , 767, 126
Samal, M. R., Zavagno, A., Deharveng, L., et al. 2014, , 566, A122
Schneider, N., Csengeri, T., Hennemann, M., et al. 2012, , 540, L11
Shirley, Y. L. 2015, , 127, 299
Skrutskie, M. F., Cutri, R. M., Stiening, R., et al. 2006, , 131, 1163
Spitzer, L. 1978, New York Wiley-Interscience, 1978. 333 p.,
Thompson, M. A., Urquhart, J. S., Moore, T. J. T., & Morgan, L. K. 2012, , 421, 408
Traficante, A., Calzoletti, L., Veneziani, M., et al. 2011, , 416, 2932
Tremblin, P., Audit, E., Minier, V., & Schneider, N. 2012a, , 538, A31
Tremblin, P., Audit, E., Minier, V., Schmidt, W., & Schneider, N. 2012b, , 546, A33
Tremblin, P., Schneider, N., Minier, V., et al. 2014, , 564, A106
Urquhart, J. S., Hoare, M. G., Lumsden, S. L., Oudmaijer, R. D., & Moore, T. J. T. 2008, Massive Star Formation: Observations Confront Theory, 387, 381
van der Wiel, M. H. D., van der Tak, F. F. S., Ossenkopf, V., et al. 2009, , 498, 161
van der Wiel, M. H. D., van der Tak, F. F. S., Ossenkopf, V., et al. 2010, , 510, C1
Vazquez-Semadeni, E.1994, , 423, 681
V[á]{}zquez-Semadeni, E., Gonz[á]{}lez, R. F., Ballesteros-Paredes, J., Gazol, A., & Kim, J. 2008, , 390, 769
Watson, C., Hanspal, U., & Mengistu, A. 2010, , 716, 1478
Whitney, B. A., Indebetouw, R., Bjorkman, J. E., & Wood, K. 2004, , 617, 1177
Whitworth, A. P., Bhattal, A. S., Chapman, S. J., Disney, M. J., & Turner, J. A. 1994, , 268, 291
Wink, J. E., Wilson, T. L., & Bieging, J. H. 1983, , 127, 211
Wright, E. L., Eisenhardt, P. R. M., Mainzer, A. K., et al. 2010, , 140, 1868
Wu, Y., Henkel, C., Xue, R., Guan, X., & Miller, M. 2007, , 669, L37
Yuan, J.-H., Wu, Y., Li, J. Z., & Liu, H. 2014, , 797, 40
Zavagno, A., Anderson, L. D., Russeil, D., et al. 2010, , 518, L101
Zavagno, A., Deharveng, L., Comer[ó]{}n, F., et al. 2006, , 446, 171
Zavagno, A., Pomar[è]{}s, M., Deharveng, L., et al. 2007, , 472, 835
{width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"}
{width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"} {width="24.00000%"}
[^1]: , the *Herschel* infrared Galactic Plane Survey, is an Open Time Key project on the *Hershcel Space Observatory* (HSO) aiming to map the entire Galactic Plane in five infrared bands. This survey covers a $|b|<1^{\circ}$ wide strip of the Milky Way Galactic plane in the longitude range $-60^{\circ} \leq l \leq 60^{\circ}$
[^2]: <http://sha.ipac.caltech.edu/applications/Spitzer/SHA/>
[^3]: <http://irsa.ipac.caltech.edu/frontpage/>
[^4]: <http://third.ucllnl.org/gps/>
[^5]: <http://www.physics.wisc.edu/~craigm/idl/fitting.html>
[^6]: [www.astro.uni-koeln.de/cdms/catalog](www.astro.uni-koeln.de/cdms/catalog)
[^7]: <http://caravan.astro.wisc.edu/protostars/>
[^8]: <http://www.star.bris.ac.uk/~mbt/topcat/>
| {
"pile_set_name": "ArXiv"
} |
**Reply to Comment by J.A.Garcia, arXiv:0705.0143 (to appear in PRD)**
**Subir Ghosh**\
Physics and Applied Mathematics Unit,\
Indian Statistical Institute,\
203 B. T. Road, Calcutta 700108, India.
\
We observe that there is no clash between the works [@gar] and [@g1]. In the Comment [@gar] the author has shown that one can construct a Lagrangian model of a point particle with a Magueijo-Smolin (MS) form of dispersion relation in a canonical phase space provided one modifies the Lorentz generator to $$J^{\mu\nu}_{DSR}=(x^\mu -\frac{(xp)}{l}\eta ^\mu )p^{\nu}-(x^\nu -\frac{(xp)}{l}\eta ^\nu )p^{\mu}.
\label{1}$$ On the other hand in [@g1] I have shown that one can keep the Lorentz generator $$J^{\mu\nu}=x^\mu p^{\nu}-x^\nu p^{\mu}
\label{2}$$ unchanged provided a non-canonical symplectic structure is used.
In my opinion the above two formalisms are complimentary and there is no reason to treat the former [@gar] as an improvement, but for a bias of the author of [@gar] against the introduction of a non-canonical phase space.
Furthermore, it is crucial to keep in mind that from the point of view of DSR, the Lagrangian in (4) [@gar] with the chosen form of $J^{\mu\nu}_{DSR}$ is fundamental and the coordinate $x^\mu$, (with its non-canonical behavior under Lorentz transformation), is the [*[physical]{}*]{} coordinate and $p^{\mu}$ is the [*[physical]{}*]{} momentum. According to DSR, results obtained in $x^\mu ,p^\mu $ variables (and [*[not]{}*]{} in $X^\mu ,P^\mu $) should be compared with experiments. Thus (4) in [@gar] should be considered as the starting point and (1) in [@gar] is obtained in a particular parameterization. This does not mean that the DSR model is trivially related to normal particle model. This is because in order to get the correct behavior of a DSR particle one has to convert the normal particle equations (in $X^\mu ,P^\mu $) to equations involving physical DSR coordinates $x^\mu ,p^\mu $ using $$X^\mu =(F^{-1})^{\mu}_{\nu}x^\nu$$ as given in [@gar]. This will lead to new $\kappa$-DSR physics since coordinates and momenta get mixed up under Lorentz transformation. In this way one can exploit the canonical framework to generate DSR behavior. This sort of approach is discussed extensivly in [@g2] in a related context where it is also shown that dynamical inputs are required in order to extrapolate kinematical equations in canonical framework to equations in DSR framework.
[99]{} J.A.Garcia, arXiv:0705.0143 (to appear in PRD). S.Ghosh, Phys.Rev. D74, 084019 (hep-th/0608206). S.Ghosh and P.Pal, hep-th/0702159 (to appear in PRD).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The regularity of an edge ideal of a finite simple graph $G$ is at least the induced matching number of $G$ and is at most the minimum matching number of $G$. If $G$ possesses a dominating induced matching, i.e., an induced matching which forms a maximal matching, then the induced matching number of $G$ is equal to the minimum matching number of $G$. In the present paper, from viewpoints of both combinatorics and commutative algebra, finite simple graphs with dominating induced matchings will be mainly studied.'
address:
- 'Department of Pure and Applied Mathematics, Graduate School of Information Science and Technology, Osaka University, Toyonaka, Osaka 560-0043, Japan'
- ' Department of Mathematics, Kyoto Sangyo University, Motoyama, Kamigamo, Kita-ku, Kyoto 603-8555, Japan'
- 'Department of Mathematics, Faculty of Science, Shizuoka University, 836 Ohya, Suruga-ku, Shizuoka 422-8529, Japan'
- 'Department of Pure and Applied Mathematics, Graduate School of Information Science and Technology, Osaka University, Toyonaka, Osaka 560-0043, Japan'
author:
- Takayuki Hibi
- Akihiro Higashitani
- Kyouko Kimura
- Akiyoshi Tsuchiya
title: Dominating induced matchings of finite graphs and regularity of edge ideals
---
Introduction {#introduction .unnumbered}
============
The regularity of an edge ideal of a finite simple graph has been studied by many articles including [@BC1205], [@BC1302], [@DHS], [@HaVThypergraph], [@Katzman], [@KAM], [@Kummini], [@MMCRTY], [@Nevo], [@VanTuyl], [@Woodroofe-regularity] and [@Zheng]. Recall that a finite graph is [*simple*]{} if it possesses no loop and no multiple edge.
Let $G$ be a finite simple graph on the vertex set $[n] = \{ 1, \ldots, n \}$ with the edge set $E(G)$ and $S = K[x_{1}, \ldots, x_{n}]$ the polynomial ring in $n$ variables over a field $K$ with standard grading. The [*edge ideal*]{} of $G$ is the ideal $I(G) \subset S$ which is generated by those squarefree quadratic monomials $x_{i}x_{j}$ with $\{ i, j \} \in E(G)$. Following the previous paper [@HHKO], we continue our research on the relation between the regularity $\operatorname{reg}(S/I(G))$ of the quotient ring $S/I(G)$ and the matching number, the minimum matching number together with the induced matching number of $G$.
A [*matching*]{} of $G$ is a subset $\mathcal{M} \subset E(G)$ such that, for $e$ and $e'$ belonging to $\mathcal{M}$ with $e \neq e'$, one has $e \cap e'
= \emptyset$. A maximal matching of $G$ is a matching $\mathcal{M}$ of $G$ for which $\mathcal{M} \cup \{ e \}$ cannot be a matching of $G$ for all $e \in E(G) \setminus \mathcal{M}$. An [*induced matching*]{} is a matching $\mathcal{M}$ of $G$ such that, for $e$ and $e'$ belonging to $\mathcal{M}$ with $e \neq e'$, there is no edge $f \in E(G)$ with $e \cap f \neq \emptyset$ and $e' \cap f \neq \emptyset$. The [*matching number*]{} of $G$, denoted by $\operatorname{match}(G)$, is the maximum cardinality of the matchings of $G$ and the [*minimum matching number*]{} of $G$, denoted by $\operatorname{min-match}(G)$, is the minimum cardinality of the maximal matchings of $G$. Furthermore, the [*induced matching number*]{} of $G$, denoted by $\operatorname{ind-match}(G)$, is the maximum cardinality of the induced matching of $G$.
The basic inequalities, due to [@Katzman] and [@Woodroofe-regularity], among the above three invariants together with $\operatorname{reg}(S/I(G))$ are $$\operatorname{ind-match}(G) \leq \operatorname{reg}S/I(G) \leq \operatorname{min-match}(G) \leq \operatorname{match}(G).$$ In addition, one can easily prove the inequality $$\operatorname{match}(G) \leq 2 \operatorname{min-match}(G),$$ see Proposition \[claim:m<2m\]. Naturally, one question arises: Given integers $p, c, q, r$ satisfying $$0 < p \leq c \leq q \leq r \leq 2q,$$ we can ask if there exists a finite simple graph $G$ for which $$\operatorname{ind-match}(G) = p, \, \,
\operatorname{reg}S/I(G) = c, \, \,
\operatorname{min-match}(G) = q, \, \,
\operatorname{match}(G) = r.$$ In Section $1$, this question and its related problems will be studied.
Cameron and Walker [@CW] succeeded in characterizing a finite simple graph $G$ with $\operatorname{ind-match}(G) = \operatorname{match}(G)$. For example, if $G$ is a star or a star triangle, then one has $\operatorname{ind-match}(G) = \operatorname{match}(G)$. We say that a finite connected simple graph $G$ is a [*Cameron–Walker graph*]{} if $\operatorname{ind-match}(G) = \operatorname{match}(G)$ and if $G$ is neither a star nor a star triangle. Thus in particular for a Cameron–Walker graph $G$, one has $$\operatorname{ind-match}(G) = \operatorname{reg}S/I(G) = \operatorname{min-match}(G) = \operatorname{match}(G).$$ From a viewpoint of commutative algebra, the study on Cameron–Walker graphs is done in [@HHKO]. In Section $2$, we treat some classes of finite simple graphs which contain Cameron–Walker graphs as a subclass and investigate these combinatorial properties.
A [*dominating induced matching*]{} of $G$ is an induced matching which also forms a maximal matching of $G$. Every Cameron–Walker graph possesses a dominating induced matching. Clearly a finite simple graph $G$ with a dominating induced matching satisfies the equalities $$\operatorname{ind-match}(G) = \operatorname{reg}S/I(G) = \operatorname{min-match}(G).$$ However, there is a finite simple graph $G$ which possesses no dominating induced matching, but satisfies the equality $\operatorname{ind-match}(G) = \operatorname{min-match}(G)$. A characterization of finite simple graphs possessing dominating induced matchings is easy, see Proposition \[ind-match=max-match\].
Our first work is to find a characterization of finite simple graphs $G$ satisfying $\operatorname{ind-match}(G) = \operatorname{min-match}(G)$ (Theorem \[ind-match=min-match\]).
Recall that a [*vertex cover*]{} of a finite simple graph $G$ on $[n]$ is a subset $C \subset [n]$ for which $C \cap e \neq \emptyset$ for all $e \in E(G)$. A [*minimal vertex cover*]{} of $G$ is a vertex cover $C$ of $G$ for which no proper subset of $C$ can be a vertex cover of $G$. A finite simple graph $G$ is called [*unmixed*]{} if all minimal vertex covers have the same cardinality. Our second work is to characterize unmixed graphs with dominating induced matchings (Theorem \[claim:unmixedDIM\]).
Finally, in Section $3$, the algebraic study of finite simple graphs with dominating induced matchings will be discussed. In [@HHKO] it is shown that every Cameron–Walker graph is vertex decomposable, hence sequentially Cohen–Macaulay. However, there is a finite simple graph $G$ with a dominating induced matching such that $G$ is not sequentially Cohen–Macaulay. We cannot escape from the temptation to find a characterization of vertex decomposable graphs with dominating induced matchings. However, to find a complete characterization seems to be rather difficult. We try to find a class ${\mathcal A}$ of vertex decomposable graphs with dominating induced matchings such that ${\mathcal A}$ contains all Cameron–Walker graphs. In addition, various examples will be supplied.
Matching number, induced matching number, and regularity
========================================================
Let $G$ be a finite simple graph. A [*matching*]{} of $G$ is a subset $\mathcal{M} \subset E(G)$ such that $e \cap e' = \emptyset$ for all $e, e' \in \mathcal{M}$ with $e \neq e'$. We denote by $\operatorname{match}(G)$ (resp. $\operatorname{min-match}(G)$), the maximum (resp. minimum) cardinality among maximal matchings of $G$. Two edges $e, e' \in E(G)$ are said to be [*$3$-disjoint*]{} if $e \cap e' = \emptyset$ and there is no edge $f \in E(G)$ with $e \cap f \neq \emptyset$ and $e' \cap f \neq \emptyset$. An [*induced matching*]{} is a set of edges which are pairwise $3$-disjoint. We denote by $\operatorname{ind-match}(G)$, the maximum cardinality among induced matchings of $G$. By Katzman [@Katzman] and Woodroofe [@Woodroofe-regularity], we have $$\label{eq:all-ineq}
\operatorname{ind-match}(G) \leq \operatorname{reg}S/I(G) \leq \operatorname{min-match}(G) \leq \operatorname{match}(G).$$ In this section, we investigate the problem to construct a finite simple connected graph with given values of these $4$ invariants.
We first note the relation between $\operatorname{match}(G)$ and $\operatorname{min-match}(G)$.
\[claim:m<2m\] Let $G$ be a finite simple graph. Then $\operatorname{match}(G) \leq 2 \operatorname{min-match}(G)$.
Let $\{u_i, v_i \}$, $i = 1, 2, \ldots, q$ be edges of $G$ which form a maximal matching with $q = \operatorname{min-match}(G)$. Let $e$ be an edge in $G$. Then $e$ contains at least one vertex of $2q$ vertices $u_1, v_1, \ldots, u_q, v_q$. Therefore there is no matching which consists of $2q+1$ edges.
Then the following problem naturally occurs:
Let $p, c, q, r$ be integers satisfying $$0 < p \leq c \leq q \leq r \leq 2q.$$ Construct a finite simple connected graph $G$ satisfying $$\operatorname{ind-match}(G) = p, \quad
\operatorname{reg}S/I(G) = c, \quad
\operatorname{min-match}(G) = q, \quad
\operatorname{match}(G) = r.
$$
When we ignore the condition for the regularity, we can do. The following result might be known, however we give a proof of it for the sake of completeness.
\[ind-min-match\] For arbitrary integers $p, q, r$ with $0 < p \leq q \leq r \leq 2q$, there exists a finite simple connected graph $G$ which satisfies $$\operatorname{ind-match}(G) = p, \qquad \operatorname{min-match}(G) = q, \qquad \operatorname{match}(G) = r.$$
Let $a, b, m, n$ be non-negative integers with $m \leq n$ and $1 \leq n$. Let us consider the following graph $G_{a,b,m,n}$:
(400,270)(10,-70) (50,170)[$G_{a,b,m,n}$:]{} (200,100)[(100,50)]{} (170,125) (190,125) (205,125) (210,125) (215,125) (205,150) (210,150) (215,150) (230,125) (200,75) (170,175) (190,175) (230,175) (150,110) (150,95) (150,90) (152,85) (170,75) (250,110) (250,95) (250,90) (248,85) (230,75) (50,25) (100,25) (175,25) (225,25) (300,25) (350,25) (50,-5) (50,-35) (30,5) (30,-25) (100,-5) (100,-35) (80,5) (80,-25) (175,-5) (175,-35) (155,5) (155,-25) (225,-5) (225,-35) (300,-5) (300,-35) (350,-5) (350,-35) (120,5) (125,5) (130,5) (255,5) (260,5) (265,5) (170,125)[(0,1)[50]{}]{} (190,125)[(0,1)[50]{}]{} (230,125)[(0,1)[50]{}]{} (200,75)[(-3,-1)[150]{}]{} (200,75)[(-2,-1)[100]{}]{} (200,75)[(-1,-2)[25]{}]{} (200,75)[(1,-2)[25]{}]{} (200,75)[(2,-1)[100]{}]{} (200,75)[(3,-1)[150]{}]{} (50,25)[(0,-1)[60]{}]{} (50,25)[(-1,-1)[20]{}]{} (50,-5)[(-1,-1)[20]{}]{} (100,25)[(0,-1)[60]{}]{} (100,25)[(-1,-1)[20]{}]{} (100,-5)[(-1,-1)[20]{}]{} (175,25)[(0,-1)[60]{}]{} (175,25)[(-1,-1)[20]{}]{} (175,-5)[(-1,-1)[20]{}]{} (225,25)[(0,-1)[60]{}]{} (300,25)[(0,-1)[60]{}]{} (350,25)[(0,-1)[60]{}]{} (25,-40)[$\underbrace{\phantom{aaaaaaaaaaaaaaaaaaaaaaaaaaaaaa}}$]{} (100,-55)[$a$]{} (215,-40)[$\underbrace{\phantom{aaaaaaaaaaaaaaaaaaaaaaaaaaa}}$]{} (280,-55)[$b$]{} (165,177)[$\overbrace{\phantom{aaaaaaaaaaaaa}}$]{} (195,190)[$2m$]{} (190,100)[$K_{2n}$]{}
Then $G_{a,b,m,n}$ satisfies $$\begin{aligned}
\operatorname{ind-match}(G_{a,b,m,n}) &= a+b+1, \\
\operatorname{min-match}(G_{a,b,m,n}) &= a+b+n, \\
\operatorname{match}(G_{a,b,m,n}) &= 2a+b+(n-m) + 2m = 2a + b + n + m.
\end{aligned}$$ Therefore we obtain a desired graph $G$ if we can choose $a, b, n, m$ satisfying $$\begin{aligned}
a+b+1 &= p, \\
a+b+n &= q, \\
2a+b+n+m &=r.
\end{aligned}$$
Indeed, we can choose such $a,b,m,n$. First, by $a+b=p-1$, we take $n = q-(a+b) = q-p+1 > 0$. Then $$r = 2a+b+n+m = a+ (p-1) + (q-p+1) + m,$$ and we have $a+m = r-q$. Note that $a+b=p-1$ and $a+m=r-q$.
*Case 1*: $r-q \leq q-p+1$. We can take $a=0$, $b=p-1$, and $m=r-q$. (Then $m \leq n$.)
*Case 2*: $r-q > q-p+1$. We set $m=q-p+1$ and $a=r-2q+p-1 (>0)$. Then $m=n$ and $b = (p-1) - (r-2q+p-1) = 2q-r \geq 0$. The last inequality follows from the condition $r \leq 2q$.
Also, the difference between the regularity and the induced matching number as well as the difference between the minimum matching number and the regularity can be arbitrary large.
Recall that the regularity of a (standard graded) $S$-module $M$ is defined by $$\operatorname{reg}(M) = \max \{ j-i \; : \; \beta_{ij} (M) \neq 0 \},$$ where $\beta_{ij} (M) := \dim_K [\operatorname{Tor}_i^S (K, M)]_j$, the $ij$th Betti number of $M$.
\[CompleteBipartite5cycle4cycle\] For arbitrary non-negative integers $a, b$, there exists a finite simple connected graph $G$ satisfying $$\label{eq:distance-reg}
\begin{aligned}
\operatorname{ind-match}(G) &= \operatorname{reg}(S/I(G)) - a, \\
\operatorname{min-match}(G) &= \operatorname{reg}(S/I(G)) + b.
\end{aligned}$$
Although for a given integer $c$, there exists a simple connected graph $G$ with $\operatorname{reg}S/I(G) = c$ (for example, the cycle of length $3c-1$ is such a graph), we do not know whether there exists a finite simple connected graph $G$ satisfying $\operatorname{reg}S/I(G) = c$ together with (\[eq:distance-reg\]) for given integers $a,b,c$.
Let $G$ be a finite simple graph on $V$. When we identify the vertices of $G$ with the variables of the underlying polynomial ring $S$ of the edge ideal $I(G)$, we denote $S=K[V]$. Theorem \[CompleteBipartite5cycle4cycle\] immediately follows by the following lemma.
\[CompleteBipartite5cycle4cycle-invariants\] Let $a, b$ be non-negative integers. Let $G_{a,b}$ be the graph consisting of a complete bipartite graph $K_{1, a+b+1}$ with the bipartition $\{ x \} \sqcup \{ y_1, \ldots, y_{a+b+1} \}$ and $5$-cycles attaching to each $y_1, \ldots, y_{a+1}$ and $4$-cycles attaching to each $y_{a+2}, \ldots, y_{a+b+1}$. We denote by $V_{a,b}$, the vertex set of $G_{a,b}$. Then $$\begin{aligned}
\operatorname{ind-match}(G_{a,b}) &= a + b +2, \\
\operatorname{min-match}(G_{a,b}) &= 2a + 2b + 2, \\
\operatorname{match}(G_{a,b}) &= 2a + 2b + 3, \\
\operatorname{reg}(K[V_{a,b}]/I(G_{a,b})) &= 2a + b + 2.
\end{aligned}$$
In order to prove Lemma \[CompleteBipartite5cycle4cycle-invariants\], we use the following two results.
\[Woodroofe:regularity\] If a graph $G$ has an induced subgraph $H$ which consists of disjoint union of $m$ edges and cycles $C_{3 i_1 + 2}, \ldots, C_{3 i_n + 2}$, then $\operatorname{reg}S/I(G) \geq m + n + \sum_{j=1}^n i_j$.
\[KM:regularity\] Let $I_1, \ldots, I_s$ be squarefree monomial ideals of $S$. Then $$\operatorname{reg}S/(I_1 + \cdots + I_s) \leq \sum_{j=1}^s \operatorname{reg}S/I_j.$$
We first compute $\operatorname{ind-match}(G_{a,b})$. Note that we cannot choose $2$ edges which are $3$-disjoint in $G_{a,b}$ from each $4$-cycle or each $5$-cycle. The same is true for $K_{1, a+b+1}$. Therefore $\operatorname{ind-match}(G_{a,b}) \leq a+b+2$. Indeed, there exist $a+b+2$ edges of $G_{a,b}$ which form an induced matching of $G_{a,b}$: we choose an edge which does not contain $y_i$ from each $4$-cycle, the edge which is $3$-disjoint with $\{ x, y_i \}$ from each $5$-cycle, and $\{ x, y_1 \}$.
We next compute $\operatorname{reg}K[V_{a,b}]/I(G_{a,b})$. Take an edge from each $4$-cycle. Then the graph which consists of these $b$ edges and $a + 1$ copies of $5$-cycles is an induced subgraph of $G_{a,b}$. Therefore by Lemma \[Woodroofe:regularity\], we have $$\operatorname{reg}(K[V_{a,b}]/I(G_{a,b})) \geq b + 2(a+1).$$
In order to prove the opposite inequality, we define subgraphs $G_1, \ldots, G_{a+b+1}$ of $G_{a,b}$. For $1 \leq i \leq a+1$, let $G_i$ be the subgraph of $G_{a,b}$ consisting of the $5$-cycle containing $y_i$ and $\{ x, y_i \}$. Also for $a+2 \leq i \leq a+b+1$, let $G_i$ be the subgraph of $G_{a,b}$ consisting of the $4$-cycle containing $y_i$ and $\{ x, y_i \}$. Then $E(G_{a,b}) = \bigcup_{i=1}^{a+b+1} E(G_i)$. Since $$\operatorname{reg}(K[V_{a,b}]/I(G_i)K[V_{a,b}]) = \left\{
\begin{alignedat}{3}
&2, &\quad &1 \leq i \leq a+1, \\
&1, &\quad &a+2 \leq i \leq a+b+1,
\end{alignedat}
\right.$$ we have $\operatorname{reg}(K[V_{a,b}]/I(G_{a,b})) \leq 2(a+1)+b$ by Lemma \[KM:regularity\].
Therefore $\operatorname{reg}(K[V_{a,b}]/I(G_{a,b})) = b + 2(a+1)$ holds.
Finally we compute $\operatorname{min-match}(G_{a,b})$ and $\operatorname{match}(G_{a,b})$. Note that $$\begin{aligned}
&\operatorname{min-match}(C_4) = \operatorname{match}(C_4) = 2, \\
&\operatorname{min-match}(C_5) = \operatorname{match}(C_5) = 2, \\
&\operatorname{min-match}(K_{1, a+b+1}) = \operatorname{match}(K_{1, a+b+1}) = 1.
\end{aligned}$$ Let $\mathcal{M}$ be a maximal matching of $G$. If $\{ x, y_i \} \notin \mathcal{M}$ for $i = 1, \ldots, a+b+1$, then we have $\# \mathcal{M} = 2(a+1) + 2b$. If $\{ x, y_i \} \in \mathcal{M}$ for some $1 \leq i \leq a+1$, then the cardinality of $\mathcal{M}$ is either $1 + 2(a+1) + 2b$ or $1 + 1 + 2a + 2b = 2(a+1) + 2b$. If $\{ x, y_i \} \in \mathcal{M}$ for some $a+2 \leq i \leq a+b+1$, then we have $\# \mathcal{M} = 2(a+1) + 2b$. Therefore we have the desired assertions.
Now we return to the first inequalities (\[eq:all-ineq\]). There are the following $8$ cases:
1. $\operatorname{ind-match}(G) = \operatorname{reg}S/I(G) = \operatorname{min-match}(G) = \operatorname{match}(G)$.
2. $\operatorname{ind-match}(G) = \operatorname{reg}S/I(G) = \operatorname{min-match}(G) < \operatorname{match}(G)$.
3. $\operatorname{ind-match}(G) = \operatorname{reg}S/I(G) < \operatorname{min-match}(G) = \operatorname{match}(G)$.
4. $\operatorname{ind-match}(G) = \operatorname{reg}S/I(G) < \operatorname{min-match}(G) < \operatorname{match}(G)$.
5. $\operatorname{ind-match}(G) < \operatorname{reg}S/I(G) = \operatorname{min-match}(G) = \operatorname{match}(G)$.
6. $\operatorname{ind-match}(G) < \operatorname{reg}S/I(G) = \operatorname{min-match}(G) < \operatorname{match}(G)$.
7. $\operatorname{ind-match}(G) < \operatorname{reg}S/I(G) < \operatorname{min-match}(G) = \operatorname{match}(G)$.
8. $\operatorname{ind-match}(G) < \operatorname{reg}S/I(G) < \operatorname{min-match}(G) < \operatorname{match}(G)$.
For each case, is there a finite simple connected graph $G$ satisfying the inequalities? The following theorem is an answer to the question.
\[graph-ineq\] There exists a finite simple connected graph $G$ satisfying the inequalities. In particular, we can construct an infinite family of finite simple connected graphs satisfying each inequalities except for the case (v).
A graph $G$ is called chordal if any cycle in $G$ of length more than $3$ has a chord. Hà and Van Tuyl [@HaVThypergraph] proved that $\operatorname{reg}S/I(G) = \operatorname{ind-match}(G)$ holds for a chordal graph $G$.
*Case (i):* The Cameron–Walker graphs [@HHKO] are just such graphs $G$.
*Case (ii):* The path graph $P_{6n}$ with $6n$ vertices $(n \geq 1)$ satisfies the inequalities. Indeed, $\operatorname{ind-match}(P_{6n}) = \operatorname{min-match}(P_{6n}) =2n$ and $\operatorname{match}(P_{6n}) = 3n$. (Note that $P_{6n}$ has a dominating induced matching; see Section \[sec:DIM\].)
*Case (iii):* The complete graph $K_n$ with $n$ vertices $(n \geq 4)$ satisfies the inequalities. Indeed, $\operatorname{ind-match}(K_n) = 1$ and $\operatorname{min-match}(K_n) = \operatorname{match}(K_n) = \lfloor n/2 \rfloor \geq 2$. Also, since $K_n$ is a chordal graph, it follows that $\operatorname{ind-match}(K_n) = \operatorname{reg}(S/I(K_n))$ by [@HaVThypergraph].
*Case (iv):* The fully whiskered graph $W(K_n)$ of the complete graph $K_n$ $(n \geq 3)$ satisfies the inequalities. Here $W(K_n)$ is defined as follows: let $x_1, \ldots, x_n$ be vertices of $K_n$ and let $y_1, \ldots, y_n$ be new vertices. Then $W(K_n)$ is the graph on $\{ x_1, \ldots, x_n,\, y_1, \ldots, y_n \}$ whose edge set is $E(K_n) \cup \{ \{ x_i, y_i \} : i=1, 2, \ldots, n \}$. Note that $W(K_n)$ is also a chordal graph. Thus $\operatorname{ind-match}(G) = \operatorname{reg}S/I(G)$ holds. Since $\operatorname{ind-match}(G) = 1$, $\operatorname{min-match}(G) = \lceil n/2 \rceil$ and $\operatorname{match}(G) = n$, $G$ satisfies the desired inequalities.
*Case (v):* The $5$-cycle $C_5$ satisfies the desired inequalities. Indeed, $\operatorname{ind-match}(C_5) = 1$ and $\operatorname{reg}S/I(C_5) = \operatorname{min-match}(C_5) = \operatorname{match}(C_5) = 2$.
*Cases (vi) and (viii)*: Consider the graph $G_{a,b}$ on Lemma \[CompleteBipartite5cycle4cycle-invariants\]. If $a \neq 0$ and $b=0$, then the graph satisfies the inequalities (vi). If $a \neq 0$ and $b \neq 0$, then the graph satisfies the inequalities (viii).
*Cases (vii):* For an integer $k \geq 2$, let $H_k$ be the graph on $$V_k := \{ u,v, \; x_1, x_2, \ldots, x_k, \; y_1, y_2, \ldots, y_k, \;
z_{11}, z_{12}, z_{21}, z_{22}, \ldots, z_{k1}, z_{k2} \}$$ with the following edges: $$\label{eq:label-edgeH_k}
\begin{alignedat}{3}
\mathfrak{e}_i &:= \{ u, x_i \}, &\
\mathfrak{f}_i &:= \{ v, y_i \}, \\
\mathfrak{g}_{i1} &:= \{ x_i, z_{i1} \}, &\
\mathfrak{g}_{i2} &:= \{ x_i, z_{i2} \}, \\
\mathfrak{h}_{i1} &:= \{ y_i, z_{i1} \}, &\
\mathfrak{h}_{i2} &:= \{ y_i, z_{i2} \},
\end{alignedat}
\quad i= 1, 2, \ldots, k;
$$ see Figure \[fig:H\_k\].
Then we see from Lemma \[H\_k\] below that $H_k$ satisfies the inequalities (vii).
(200,230)(-20,-105) (-20,90)[$H_k$:]{} (-15,0) (15,0) (45,0) (75,0) (105,0) (110,0) (115,0) (145,0) (175,0) (0,30) (60,30) (160,30) (0,-30) (60,-30) (160,-30) (80,90) (80,-90) (-28,5)[$z_{11}$]{} (14,5)[$z_{12}$]{} (32,5)[$z_{21}$]{} (74,5)[$z_{22}$]{} (132,5)[$z_{k1}$]{} (174,5)[$z_{k2}$]{} (-15,30)[$x_{1}$]{} (45,30)[$x_{2}$]{} (165,30)[$x_{k}$]{} (-15,-35)[$y_{1}$]{} (45,-35)[$y_{2}$]{} (165,-35)[$y_{k}$]{} (68,90)[$u$]{} (68,-95)[$v$]{} (-15,0)[(1,2)[15]{}]{} (45,0)[(1,2)[15]{}]{} (145,0)[(1,2)[15]{}]{} (-15,0)[(1,-2)[15]{}]{} (45,0)[(1,-2)[15]{}]{} (145,0)[(1,-2)[15]{}]{} (15,0)[(-1,2)[15]{}]{} (75,0)[(-1,2)[15]{}]{} (175,0)[(-1,2)[15]{}]{} (15,0)[(-1,-2)[15]{}]{} (75,0)[(-1,-2)[15]{}]{} (175,0)[(-1,-2)[15]{}]{} (80,90)[(-4,-3)[80]{}]{} (80,90)[(-1,-3)[20]{}]{} (80,90)[(4,-3)[80]{}]{} (80,-90)[(-4,3)[80]{}]{} (80,-90)[(-1,3)[20]{}]{} (80,-90)[(4,3)[80]{}]{}
\[H\_k\] The graph $H_k$ ($k \geq 2$) in the proof of Theorem \[graph-ineq\] satisfies $$\begin{aligned}
\operatorname{ind-match}(H_k) &= k, \\
\operatorname{min-match}(H_k) &= \operatorname{match}(H_k) = 2k, \\
\operatorname{reg}(K[V_k]/I(H_k)) &= k+1.
\end{aligned}$$
In order to prove Lemma \[H\_k\], we use a Lyubeznik resolution ([@Lyubeznik]), which is a subcomplex of the Taylor resolution.
Let $I$ be a monomial ideal of $S$ and $m_1, \ldots, m_{\mu}$ the minimal monomial generators of $I$. The free basis $e_{i_1 \cdots i_s}$ of the Taylor resolution is said to be $L$-admissible if $\operatorname{lcm}(m_{i_t}, \ldots, m_{i_s})$ is not divisible by $m_q$ for all $1 \leq t < s$ and for all $q < i_t$. We will denote an $L$-admissible symbol $e_{i_1 \cdots i_s}$ by $[m_{i_1}, \ldots, m_{i_s}]$. The degree of an $L$-admissible symbol $[m_{i_1}, \ldots, m_{i_s}]$ is defined by the degree of $\operatorname{lcm}(m_{i_1}, \ldots, m_{i_s})$. (Recall that we consider the standard grading on the polynomial ring $S$.) An $L$-admissible symbol $[m_{i_1}, \ldots, m_{i_s}]$ is said to be maximal if there is no $L$-admissible symbol $[m_{j_1}, \ldots, m_{j_t}]$ such that $\{ i_1, \ldots, i_s \} \subsetneq \{ j_1, \ldots, j_t \}$. A Lyubeznik resolution $(\mathcal{L}_{\bullet}, d_{\bullet})$ of $I$ (with respect to the above order of the minimal monomial generators) is the subcomplex of the Taylor resolution generated by all $L$-admissible symbols, which is also a free resolution of $S/I$.
We first compute $\operatorname{ind-match}(H_k)$. Let $\mathcal{M}$ be a maximal induced matching of $H_k$, i.e., $\mathcal{M}$ is an induced matching of $H_k$ and there is no induced matching which properly contains $\mathcal{M}$. Suppose that $\{ u, x_1 \} \in \mathcal{M}$. If $\{ v, y_j \} \in \mathcal{M}$ ($j = 1, 2, \ldots, k$), then each of the rest edges is not $3$-disjoint with at least one of $\{ u, x_1 \}, \{ v, y_j \}$. Hence $\# \mathcal{M} = 2$. If $\{ v, y_j \} \notin \mathcal{M}$ for all $j$, then $\mathcal{M}$ contains exactly one of $\{ y_{\ell}, z_{\ell 1} \}, \{ y_{\ell}, z_{\ell 2} \}$ for each $\ell = 2, 3, \ldots, k$. Therefore $\# \mathcal{M} = 1 + (k-1) = k$. When $\{ u, x_j \}, \{ v, y_j \} \notin \mathcal{M}$ for all $j$, exactly one of $\{ x_j, z_{j1} \}, \{ x_j, z_{j2} \},
\{ y_j, z_{j1} \}, \{ y_j, z_{j2} \}$ belongs to $\mathcal{M}$. Thus $\# \mathcal{M} = k$. Therefore we have $\operatorname{ind-match}(H_k) = k$.
Let $\mathcal{M}'$ be a maximal matching of $H_k$. We show that $\# \mathcal{M}' = 2k$. Suppose that $\{ u, x_1 \} \in \mathcal{M}'$. If $\{ v, y_1 \} \in \mathcal{M}'$, then $\mathcal{M}' \setminus \{ \{ u, x_1 \}, \{ v, y_1 \} \}$ is a maximal matching of $k-1$ copies of the $4$-cycle whose cardinality is $2(k-1)$. Hence $\# \mathcal{M}' = 2k$. If $\{ v, y_j \} \in \mathcal{M}'$ for $j \neq 1$, then $\mathcal{M}' \setminus \{ \{ u, x_1 \}, \{ v, y_j \} \}$ is a maximal matching of the disjoint union of $2$ copies of $P_3$ (the path graph with $3$ vertices) and $k-2$ copies of the $4$-cycle. Hence it follows that $\# \mathcal{M}' = 2+ 2 + 2(k-2) = 2k$. If $\{ v, y_j \} \notin \mathcal{M}'$ for all $j$, then $\mathcal{M}' \setminus \{ \{ u, x_1 \} \}$ is a maximal matching of the disjoint union of $P_3$ and $k-1$ copies of the $4$-cycle. Hence $\# \mathcal{M}' = 1 + 1 + 2(k-1) = 2k$. When $\{ u, x_j \}, \{ v, y_j \} \notin \mathcal{M}$ for all $j$, $\mathcal{M}'$ is a maximal matching of $k$ copies of the $4$-cycle. Therefore $\# \mathcal{M}' = 2k$.
Finally, we compute $\operatorname{reg}(K[V_k]/I(H_k))$. We consider the following decomposition of $H_k$: (a) the $4$-cycle with vertices $x_j, z_{j1}, y_j, z_{j2}$ and the edge $\{ x_j, u \}$ ($j=1, 2, \ldots, k$); (b) the star graph on $\{ v, y_1, \ldots, y_k \}$. The edge ideal of each decomposed graph is of regularity $1$. Thus we have $\operatorname{reg}(K[V_k]/I(H_k)) \leq k+1$ by Lemma \[KM:regularity\].
For the opposite inequality $\operatorname{reg}(K[V_k]/I(H_k)) \geq k+1$, we consider the induced subgraph of $H_k$ on $$W_k = \{ u, v, \; x_1, x_2, \ldots, x_k, \; y_1, y_2, \ldots, y_k, \;
z_{11}, z_{21}, \ldots, z_{k1} \}.$$ We use the labeling of edges of $H_k$ as in (\[eq:label-edgeH\_k\]). For the sake of simplicity, we use $z_i$ (resp. $\mathfrak{g}_{i}$, $\mathfrak{h}_{i}$) instead of $z_{i1}$ (resp. $\mathfrak{g}_{i1}$, $\mathfrak{h}_{i1}$) for $i = 1, 2, \ldots, k$ and denote this graph by $H_k'$; see Figure \[fig:H\_k’\].
(200,230)(-20,-105) (-20,90)[$H_k'$:]{} (0,0) (60,0) (105,0) (110,0) (115,0) (160,0) (0,30) (60,30) (160,30) (0,-30) (60,-30) (160,-30) (80,90) (80,-90) (-15,-3)[$z_{1}$]{} (45,-3)[$z_{2}$]{} (165,-3)[$z_{k}$]{} (-15,30)[$x_{1}$]{} (45,30)[$x_{2}$]{} (165,30)[$x_{k}$]{} (-15,-35)[$y_{1}$]{} (45,-35)[$y_{2}$]{} (165,-35)[$y_{k}$]{} (68,90)[$u$]{} (68,-95)[$v$]{} (0,0)[(0,1)[30]{}]{} (60,0)[(0,1)[30]{}]{} (160,0)[(0,1)[30]{}]{} (0,0)[(0,-1)[30]{}]{} (60,0)[(0,-1)[30]{}]{} (160,0)[(0,-1)[30]{}]{} (80,90)[(-4,-3)[80]{}]{} (80,90)[(-1,-3)[20]{}]{} (80,90)[(4,-3)[80]{}]{} (80,-90)[(-4,3)[80]{}]{} (80,-90)[(-1,3)[20]{}]{} (80,-90)[(4,3)[80]{}]{} (27,60)[$\mathfrak{e}_1$]{} (60,60)[$\mathfrak{e}_2$]{} (120,60)[$\mathfrak{e}_k$]{} (35,-70)[$\mathfrak{f}_1$]{} (65,-70)[$\mathfrak{f}_2$]{} (120,-70)[$\mathfrak{f}_k$]{} (-12,12)[$\mathfrak{g}_1$]{} (48,12)[$\mathfrak{g}_2$]{} (162,12)[$\mathfrak{g}_k$]{} (-12,-19)[$\mathfrak{h}_1$]{} (48,-19)[$\mathfrak{h}_2$]{} (162,-19)[$\mathfrak{h}_k$]{}
By Hochster’s formula for Betti numbers (see also [@Kimura Lemma 3.1]), it is enough to prove that $\beta_{2k+1, 3k+2} (K[W_k]/I(H_k')) \neq 0$. In order to prove this, we use a Lyubeznik resolution. We identify edges of $H_k'$ and minimal monomial generators of $I(H_k')$.
When $k=2$, $H_2'$ is the $8$-cycle. (Hence we know that $\operatorname{reg}K[W_2]/I(H_2') = 3$.) Let us consider the Lyubeznik resolution of $I(H_2')$ with respect to the following order of edges of $H_2'$ (which corresponds to the order of minimal monomial generators of $I(H_2')$): $$\label{eq:orderH_2'}
\mathfrak{e}_1, \mathfrak{h}_1, \mathfrak{f}_k, \mathfrak{g}_k,
\mathfrak{e}_k, \mathfrak{g}_1, \mathfrak{f}_1, \mathfrak{h}_k.
$$ We denote the resolution by $(\mathcal{L}_{\bullet}^{(2)}, d_{\bullet}^{(2)})$. Then the maximal $L$-admissible symbols are $$[\mathfrak{e}_1, \mathfrak{h}_1, \mathfrak{f}_k, \mathfrak{g}_k,
\mathfrak{e}_k, \mathfrak{f}_1],
[\mathfrak{e}_1, \mathfrak{h}_1, \mathfrak{f}_k, \mathfrak{g}_k,
\mathfrak{g}_1, \mathfrak{h}_k].$$ Put $$\xi^{(2)} := [\mathfrak{e}_1, \mathfrak{h}_1, \mathfrak{f}_k, \mathfrak{g}_k,
\mathfrak{e}_k]
- [\mathfrak{e}_1, \mathfrak{h}_1, \mathfrak{f}_k, \mathfrak{g}_k,
\mathfrak{g}_1] \in \mathcal{L}_5^{(2)}.$$ Then it is easy to see that $1 \otimes \xi^{(2)}
\in \operatorname{Ker}(1 \otimes d_{5}^{(2)}) \setminus \operatorname{Im}(1 \otimes d_6^{(2)})$. Also $\deg \xi^{(2)} = 8$. Therefore we have $\beta_{5,8} (K[V_2]/I(H_2')) \neq 0$.
Next assume that $k \geq 3$. For each $i = 2, \ldots, k-1$, consider the induced subgraph of $H_k'$ on $\{ u, x_i, z_i, y_i, v \}$. We denote it by $L_i^{(k)}$. Also let $C_8^{(k)}$ be the induced subgraph of $H_k'$ on $\{ u, x_1, z_1, y_1, v, y_k, z_k, x_k \}$, which is the $8$-cycle. Note that $E(H_k') = \bigcup_{i=2}^{k-1} E(L_i^{(k)}) \cup E(C_8^{(k)})$.
Let us consider the Lyubeznik resolution of $I(L_i^{(k)})$ with respect to the following order of edges of $L_i^{(k)}$: $\mathfrak{g}_i, \mathfrak{h}_i, \mathfrak{f}_i, \mathfrak{e}_i$. We denote the resolution by $(\mathcal{L}_{\bullet}^{(k,i)}, d_{\bullet}^{(k,i)})$. Then the maximal $L$-admissible symbols are $[\mathfrak{g}_i, \mathfrak{f}_i, \mathfrak{e}_i],
[\mathfrak{g}_i, \mathfrak{h}_i, \mathfrak{f}_i]$ and it is easy to see that $1 \otimes [\mathfrak{g}_i, \mathfrak{h}_i]
\in \operatorname{Ker}(1 \otimes d_2^{(k,i)}) \setminus \operatorname{Im}(1 \otimes d_3^{(k,i)})$. Also $\deg [\mathfrak{g}_i, \mathfrak{h}_i] = 3$.
We also consider the Lyubeznik resolution of $I(C_8^{(k)})$ with respect to the ordering as in (\[eq:orderH\_2’\]). Then the same argument with $H_2'$ is valid.
Now let us consider the Lyubeznik resolution of $I(H_k')$ with respect to the following order of edges of $H_k'$: $$\label{eq:orderH_k'}
\begin{aligned}
&\mathfrak{g}_2, \mathfrak{h}_2, \mathfrak{f}_2, \mathfrak{e}_2, \\
&\qquad \quad \ldots, \\
&\qquad \mathfrak{g}_{k-1}, \mathfrak{h}_{k-1}, \mathfrak{f}_{k-1},
\mathfrak{e}_{k-1}, \\
&\qquad \qquad
\mathfrak{e}_1, \mathfrak{h}_1, \mathfrak{f}_k, \mathfrak{g}_k,
\mathfrak{e}_k, \mathfrak{g}_1, \mathfrak{f}_1, \mathfrak{h}_k.
\end{aligned}
$$ We denote the resolution by $(\mathcal{L}_{\bullet}^{(k)}, d_{\bullet}^{(k)})$. Note that for $i=1, \ldots, k-2$, the $i$th row of (\[eq:orderH\_k’\]) corresponds to $L_{i+1}^{(k)}$ and the last row of (\[eq:orderH\_k’\]) corresponds to $C_8^{(k)}$. These graphs are only connected by the vertices $u,v$. By the definition of the ordering of the minimal monomial generators of $I(H_k')$, it is easy to see that the $L$-admissible symbols of $(\mathcal{L}_{\bullet}^{(k)}, d_{\bullet}^{(k)})$ are obtained by each $L$-admissible symbols for $L_2^{(k)}, \ldots, L_{k-1}^{(k)}$ and $C_8^{(k)}$. A similar claim is true when we consider the maximal $L$-admissible symbols. Put $$\begin{aligned}
\xi^{(k)}
:= &[\mathfrak{g}_2, \mathfrak{h}_2,
\ldots, \mathfrak{g}_{k-1}, \mathfrak{h}_{k-1}, \,
\mathfrak{e}_1, \mathfrak{h}_1, \mathfrak{f}_k, \mathfrak{g}_k,
\mathfrak{e}_k] \\
- &[\mathfrak{g}_2, \mathfrak{h}_2,
\ldots, \mathfrak{g}_{k-1}, \mathfrak{h}_{k-1}, \,
\mathfrak{e}_1, \mathfrak{h}_1, \mathfrak{f}_k, \mathfrak{g}_k,
\mathfrak{g}_1].
\end{aligned}$$ Then $\xi^{(k)} \in (\mathcal{L}_{2(k-2)+5}^{(k)})_{3(k-2)+8}$. Also $1 \otimes \xi^{(k)}
\in \operatorname{Ker}(1 \otimes d_{2(k-2)+5}^{(k)}) \setminus
\operatorname{Im}(1 \otimes d_{2(k-2)+6}^{(k)})$ follows. Therefore we have $\beta_{2k+1,3k+2} (K[V_k]/I(H_k')) \neq 0$ as desired.
Can we construct an infinite family of finite simple connected graphs $G$ satisfying (v)?
A finite simple connected graph $G$ satisfying the inequalities (v) might be rare. Actually, when the number of vertices of $G$ is at most $7$, there is no such a graph $G$ with (v) except for $C_5$.
\[7vertex\] Let $G$ be a finite simple connected graph with at most $7$ vertices. Then $\operatorname{match}(G) = \operatorname{reg}S/I(G) > \operatorname{ind-match}(G)$ if and only if $G$ is a $5$-cycle.
After submitting the paper, Biyikoğlu and Civan [@BC1503 Theorem 3.17] proved that there is no finite simple connected graph $G$ satisfying (v) except for $C_5$.
A graph with a dominating induced matching {#sec:DIM}
==========================================
In [@HHKO], the authors studied the Cameron–Walker graphs. In this section, we treat some classes of graphs which contain Cameron–Walker graphs as a subclass and investigate these combinatorial properties.
We first recall some definitions on graphs.
Let $G$ be a finite simple graph on the vertex set $V$. Let $W$ be a subset of $V$. We denote by $G_W$ the induced subgraph of $G$ on $W$: the vertex set of $G_W$ is $W$ and the edge set of $G_W$ consists of all edges of $G$ which are contained in $W$. We write $G \setminus W$ instead of $G_{V \setminus W}$. In particular, when $W = \{ x \}$, consisting of $1$ vertex, we write $G \setminus x$ instead of $G \setminus \{ x \}$. For a vertex $x \in V$, we denote by $N_G (x)$ the set of neighbours of $x$. Also we set $N_G [x] := N_G (x) \cup \{ x \}$. The degree of $x$ is defined by $\deg_G (x) := \# N_G (x)$. For a subset $W \subset V$, we set $N_G (W) = \bigcup_{x \in W} N_G (x)$ and $N_G [W] = \bigcup_{x \in W} N_G [x]$. We sometimes omit the lower subscript $G$ on these notation if there is no fear of confusion.
A subset $W \subset V$ is called *independent* if no two vertices of $W$ are adjacent in $G$. An independent set $W$ is said to be maximal if there is no independent set of $G$ which properly contains $W$. Also a subset $C \subset V$ is called a *vertex cover* of $G$ if all edges of $G$ meet with $C$. A vertex cover $C$ is said to be minimal if there is no vertex cover of $G$ which is properly contained in $C$. Note that $C$ is a minimal vertex cover of $G$ if and only if $V \setminus C$ is a maximal independent set of $G$. A graph is said to be unmixed if all minimal vertex covers (equivalently, all maximal independent sets) of $G$ have the same cardinality. When $G$ is unmixed, the edge ideal $I(G)$ is height unmixed.
An edge of $G$ is called a *leaf edge* if it contains a degree $1$ vertex. Also a triangle of $G$ is called a *pendant triangle* if its two vertices are of degree $2$ and the rest vertex is of degree more than $2$.
A Cameron–Walker graph $G$ satisfies the equalities (i) in the previous section: $$\label{eq:CWgraph}
\operatorname{ind-match}(G) = \operatorname{reg}S/I(G) = \operatorname{min-match}(G) = \operatorname{match}(G).$$ Recall that a Cameron–Walker graph $G$ consists of a connected bipartite graph with the vertex partition $X \sqcup Y$ such that there is at least one leaf edge attached to each vertex $x_i \in X$ and that there may be possibly some pendant triangles attached to a vertex $y_j \in Y$. Choose one leaf edge which contains $x_i$ for each $x_i \in X$. Then these edges and the edges consisting of two degree $2$ vertices of all pendant triangles form an induced matching of $G$. It also forms a maximal matching of $G$. Thus for a Cameron–Walker graph $G$, there exists an induced matching of $G$ which is also a maximal matching of $G$. Such a matching is called a *dominating induced matching* or an *efficient edge domination set*. There exists a graph which does not have a dominating induced matching. For example, let $G_0$ be the graph on the vertex set $\{ 1,2, \ldots, 6 \}$ with edges $\{ 1,2 \}, \{ 2,3 \}, \{ 3,4 \}, \{ 4,5 \}, \{ 3,6 \}$. Then it is easy to see that an induced matching consisting of one edge is not a maximal matching of $G_0$. Other induced matching of $G_0$ is only $\{ \{ 1,2 \}, \{ 4,5 \} \}$, which is also not a maximal matching; $\{ \{ 1,2 \}, \{ 4,5 \}, \{ 3,6 \} \}$ is a matching of $G_0$. On graph theory, it has been studied the problem of determining whether a given finite simple graph has a dominating induced matching. This problem is known to be NP-complete in general; see e.g., [@CMMR; @LMS].
A graph with a dominating induced matching is characterized as follows. It is easy to check but we give a proof of this for the completeness.
\[ind-match=max-match\] Let $G$ be a finite simple graph on $V$. Then $G$ has a dominating induced matching if and only if there is an independent set $W$ such that $G \setminus W$ is a disjoint union of edges. When this is the case, the set of edges of $G \setminus W$ forms a dominating induced matching of $G$.
Let $\mathcal{M} = \{ e_1, e_2, \ldots, e_s \}$ be a dominating induced matching of $G$. Let $W$ be the set of vertices which do not appear in each $e_i$. Then $W$ is an independent set of $G$ since $\mathcal{M}$ is a maximal matching.
For example, the path graph $P_{6n}$ with $6n$ vertices and the $6n$-cycle $C_{6n}$ have dominating induced matchings; for instance, take $W = \{ 3, 6, \ldots, 6n \}$.
Let $G$ be a finite simple graph. Since $\operatorname{ind-match}(G) \leq \operatorname{min-match}(G)$ holds in general, if $G$ has a dominating induced matching, then $\operatorname{ind-match}(G) = \operatorname{min-match}(G)$ holds.
\[rmk:ind-match<min-match\] For a finite simple graph $G$, the inequality $\operatorname{ind-match}(G) \leq \operatorname{min-match}(G)$ follows via $\operatorname{reg}S/I(G)$; see (\[eq:all-ineq\]).
We prove this inequality by pure combinatorics. Let $\mathcal{M} = \{ e_1, \ldots, e_s \}$ be an induced matching of $G$ and $\mathcal{M}' = \{ e_1', \ldots, e_t' \}$ a maximal matching of $G$. Then for each $e_k \in \mathcal{M}$, there exists $e_{i_k}' \in \mathcal{M}'$ with $e_{i_k}' \cap e_{k} \neq \emptyset$ because of the maximality of $\mathcal{M}'$ . Since $\mathcal{M}$ is an induced matching, it follows that $i_k \neq i_j$ if $k \neq j$. Therefore we have $\operatorname{ind-match}(G) \leq \operatorname{min-match}(G)$.
We next characterize a finite simple graph $G$ with $\operatorname{ind-match}(G) = \operatorname{min-match}(G)$.
\[ind-match=min-match\] Let $G$ be a finite simple graph on $V$. Then $G$ satisfies $\operatorname{ind-match}(G) = \operatorname{min-match}(G)$ if and only if the vertex set $V$ can be partitioned as $$ V = \{ v_{i1}, v_{i2} \; : \; i = 1, 2, \ldots, \alpha + \beta \}
\sqcup \{ z_1, \ldots, z_{\alpha} \}
\sqcup \{ w_1, \ldots, w_{\gamma} \},
$$ where $\alpha, \beta, \gamma$ are non-negative integers, so that the edge set of $G$ is of the following form $$ \{ e_i \; : \; i = 1, 2, \ldots, \alpha + \beta \} \\
\cup \{ e_i' \; : \; i = 1, 2, \ldots, \alpha \} \cup E',
$$ where we set $e_i = \{ v_{i1}, v_{i2} \}$ ($i = 1, 2, \ldots, \alpha + \beta$) and $e_i' = \{ v_{i1}, z_i \}$ ($i = 1, 2, \ldots, \alpha$), and an edge in $E'$ is one of the following:
1. an edge containing $z_i$ ($i = 1, 2, \ldots, \alpha$);
2. an edge consisting of an end vertex of $e_i$ ($i = \alpha + 1, \alpha + 2, \ldots, \alpha + \beta$) and $w_j$ ($j = 1, 2, \ldots, \gamma$);
3. an edge consisting of $v_{i1}$ ($i = 1, 2, \ldots, \alpha$) and $w_j$ ($j = 1, 2, \ldots, \gamma$).
[**(“If”)**]{} It is easy to see that $e_1, \ldots, e_{\alpha + \beta}$ form an induced matching of $G$. Hence we have $\alpha + \beta \leq \operatorname{ind-match}(G)$. On the other hand, $e_1', \ldots, e_{\alpha}', e_{\alpha + 1}, \ldots, e_{\alpha + \beta}$ form a maximal matching of $G$ since the rest vertices are $v_{12}, \ldots, v_{\alpha 2}, w_1, \ldots, w_{\gamma}$. Hence we have $\operatorname{min-match}(G) \leq \alpha + \beta$. By combining these inequalities, we have $\operatorname{ind-match}(G) = \operatorname{min-match}(G) = \alpha + \beta$.
[**(“Only If”)**]{} Put $s = \operatorname{ind-match}(G) = \operatorname{min-match}(G)$. Let $\mathcal{M} = \{ e_1, \ldots, e_s \}$ be an induced matching of $G$ with cardinality $s$ and $\mathcal{M}' = \{ e_1', \ldots, e_s' \}$ a maximal matching of $G$ with cardinality $s$. As noted in Remark \[rmk:ind-match<min-match\], for each $e_k \in \mathcal{M}$, there exists an edge $e_{i_k}' \in \mathcal{M}'$ with $e_{i_k}' \cap e_k \neq \emptyset$, and $i_k \neq i_j$ if $k \neq j$. Also since $\# \mathcal{M} = s = \operatorname{ind-match}(G)$, $i_k$ is uniquely determined by $k$. Therefore we may assume that $e_k' \neq e_k$ for $k = 1, \ldots, \alpha$ ($0 \leq \alpha \leq s$) and $e_k' = e_k$ for $k = \alpha + 1, \ldots, s$. Set $z_k = e_k' \setminus e_k$ ($k = 1, \ldots, \alpha$) and $W = V \setminus (\bigcup_{k=1}^s e_k \cup \{ z_1, \ldots, z_{\alpha} \})$. Then it is easy to see that $$V = \left( \bigcup_{k=1}^s e_k \right)
\sqcup \{ z_1, \ldots, z_{\alpha} \} \sqcup W$$ is a desired partition.
As mentioned before Remark \[rmk:ind-match<min-match\], if $G$ has a dominating induced matching, then $\operatorname{ind-match}(G) = \operatorname{min-match}(G)$ holds. But the converse is false; the graph $G_0$ (see the beginning of this section) does not have a dominating induced matching, but $\operatorname{ind-match}(G_0) = \operatorname{min-match}(G_0) = 2$; for example, $\{ \{ 2, 3 \}, \{ 4, 5 \} \}$ is a maximal matching with cardinality $2$.
Now we return to the graph with a dominating induced matching. We consider the problem which graph with a dominating induced matching is unmixed. Let $G$ be a finite simple graph on the vertex set $V$ with a dominating induced matching. Then $V$ can be decomposed as $W \sqcup M$ where $W$ is an independent set of $G$ and $G_M$ consists of $m := \operatorname{min-match}(G)$ disconnected edges $\{ x_{j1}, x_{j2} \}$, $j=1, \ldots, m$. Set $$W_0 := \{ w \in W \; : \; \text{$w$ is not an isolated vertex of $G$} \}.$$ Also set $$\begin{aligned}
m_1 &:= \# \{ j \; : \;
\text{$\deg_G x_{j1} = 1$ or $\deg_G x_{j2} = 1$} \}, \\
m_2 &:= \# \{ j \; : \;
\text{$\deg_G x_{j1} \geq 2$ and $\deg_G x_{j2} \geq 2$} \}.
\end{aligned}$$ Note that $m_1 + m_2 = m$. For a subset $U \subset V$, we denote by $IN_G (U, W)$ (or $IN (U, W)$ if there is no fear of confusion), the set of isolated vertices of $G \setminus N_G [U]$ which are contained in $W_0$.
Let $M_2$ be a subset of $M$ satisfying the following $2$ conditions:
1. $\# (M_2 \cap \{ x_{j1}, x_{j2} \}) \leq 1$ for all $j = 1, \ldots, m$;
2. $\deg_G x \geq 2$ for all $x \in M_2$,
Then $G' := G \setminus N_G [M_2]$ has a dominating induced matching if $G'$ is not an edgeless graph. Indeed $V' := V(G') = V \setminus N_G [M_2]$, the vertex set of $G'$, can be decomposed as $W' \sqcup M'$, where $W' = W \setminus N_G [M_2]$ and $M' = M \setminus N_G [M_2]$. We use notation $m_1', m_2', W_0'$ for $G'$ with respect to this decomposition of $V'$ with a similar meaning to $G$. If $G'$ is an edgeless graph, then we set $m_1' = m_2' = 0$ and $W_0' = \emptyset$.
\[claim:unmixedDIM\] We use the same notation as above. Let $G$ be a finite simple graph on $V$ with a dominating induced matching. Then $G$ is unmixed if and only if for some (and then all) decomposition $V = W \sqcup M$, the following condition $(\flat)$ is satisfied:
Condition $(\flat)$
: For each subset $M_2 \subset M$ with the properties $(\ast 1)$, $(\ast 2)$, the following $2$ conditions are satisfied: $$\begin{aligned}
(\flat 1) &\ m_2 - m_2' = \# N_G (M_2) - \# M_2. \\
(\flat 2) &\
\# W_0 \leq 2 m_2 - \# N_G (M_2)
+ \# M_2 + \# IN_G (M_2, W).
\end{aligned}$$
\[rmk:GV\] We use the same notation as in Theorem \[claim:unmixedDIM\].
1. The empty set $M_2 = \emptyset$ is regarded as satisfying $(\ast 1)$ and $(\ast 2)$. Then $(\flat 1)$ is satisfied as both-hand sides are $0$. Also $(\flat 2)$ must be $\# W_0 \leq 2 m_2$. Indeed if $G$ is unmixed, this inequality holds; see Lemma \[claim:GV\] below.
2. The left-hand side of $(\flat 1)$ is equal to the cardinality of the following set: $$\mathcal{I}_{G, M_2} := \left\{ j \; : \;
\begin{aligned}
&\deg_G x_{j1} \geq 2, \ \deg_G x_{j2} \geq 2,
\ \text{and one of the following is satisfied:} \\
&(i) \ M_2 \cap \{ x_{j1}, x_{j2} \} \neq \emptyset; \\
&(ii) \ N_G (x_{j1}) \setminus \{ x_{j2} \} \subset N_G (M_2); \\
&(iii) \ N_G (x_{j2}) \setminus \{ x_{j1} \} \subset N_G (M_2)
\end{aligned}
\right\}.$$ Indeed, $m_2 - m_2'$ counts the number of edges $\{ x_{j1}, x_{j2} \}$ with $\deg_G x_{j1} \geq 2$ and $\deg_G x_{j2} \geq 2$ such that both $x_{j1}$ and $x_{j2}$ are not vertices of $G'$ or one of $\deg_{G'} x_{j1} = 1$ and $\deg_{G'} x_{j2} = 1$ holds.
We first prove the following lemma.
\[claim:GV\] We use the same notation as above. Let $G$ be a finite simple graph on $V$ with a dominating induced matching: $V = W \sqcup M$. If $G$ is unmixed, then $\# W_0 \leq 2 m_2$.
Let us consider the following subset of $V$: $$\label{eq:MinimalVertexCover}
C_0 := \left( \bigcup_{\genfrac{}{}{0pt}{}{j}{\deg_G x_{j1} \geq 2, \, \deg_G x_{j2} \geq 2}}
\{ x_{j1}, x_{j2} \} \right)
\cup \left( \bigcup_{\genfrac{}{}{0pt}{}{j}{\deg_G x_{j k} \geq \deg_G x_{j \ell} = 1}}
\{ x_{j k} \} \right).$$ Then $C_0$ is a minimal vertex cover of $G$. Since $G$ is unmixed, we have $\operatorname{height}I(G) = m_1 + 2 m_2$. By Gitler and Valencia [@GV Corollary 3.4], we have $2 \operatorname{height}I(G) \geq \# W_0 + \# M$. Note that $\# M = 2 (m_1 + m_2)$. Hence $$\# W_0 \leq 2 \operatorname{height}I(G) - \# M
= 2 (m_1 + 2 m_2) - 2 (m_1 + m_2)
= 2 m_2.$$
Now we prove Theorem \[claim:unmixedDIM\].
We first assume that $G$ is unmixed. Let $W \sqcup M$ be a decomposition of $V$ where $W$ is an independent set of $G$ and $G_M$ consists of disconnected edges $\{ x_{j1}, x_{j2} \}$, $j=1, \ldots, m$. Since $C_0$ in (\[eq:MinimalVertexCover\]) is a minimal vertex cover of $G$, the cardinality of any minimal vertex cover of $G$ is $m_1 + 2 m_2$. Let $M_2$ be a subset of $M$ satisfying the conditions $(\ast 1)$ and $(\ast 2)$. Then $G' = G \setminus N_G [M_2]$ is an edgeless graph or a graph with dominating induced matching as noted before Theorem \[claim:unmixedDIM\]. Considering the minimal vertex cover of $G$ which is disjoint with $M_2$, we have $m_1 + 2 m_2 = \# N_G (M_2) + m_1' + 2 m_ 2'$. (If $G'$ is an edgeless graph, then we consider both $m_1'$ and $m_2'$ as $0$.) Also focusing on the number of edges of $G_M$, we have $m_1 + m_2 = \# M_2 + m_1' + m_2'$. Then we have $$\label{eq:m2-m2'}
m_2 - m_2' = \# N_G (M_2) - \# M_2.$$ Hence $(\flat 1)$ holds. Also note that when $G$ is unmixed, $G'$ is also unmixed since the union of a minimal vertex cover of $G'$ and $N_G (M_2)$ is a minimal vertex cover of $G$. Then by Lemma \[claim:GV\], we have $\# W_0' \leq 2 m_2'$. Since $$\# W_0' = \# W_0 - (\# N_G (M_2) - \# M_2) - \# IN_G (M_2, W)$$ and (\[eq:m2-m2’\]), we have $$\begin{aligned}
\# W_0 &= \# W_0' + (\# N_G (M_2) - \# M_2) + \# IN_G (M_2, W) \\
&\leq 2 m_2' + (\# N_G (M_2) - \# M_2) + \# IN_G (M_2, W) \\
&= 2 (m_2 - \# N_G (M_2) + \# M_2)
+ (\# N_G (M_2) - \# M_2) + \# IN_G (M_2, W) \\
&= 2 m_2 - \# N_G (M_2) + \# M_2 + \# IN_G (M_2, W).
\end{aligned}$$ Thus $(\flat 2)$ also holds.
We next assume that the decomposition $V=W \sqcup M$ satisfies the condition $(\flat)$. As noted in Remark \[rmk:GV\], the inequality $\# W_0 \leq 2 m_2$ is satisfied. We use induction on $m$.
When $m = 1$, there are 2 cases: $(m_1, m_2) = (1,0), (0,1)$.
If $(m_1, m_2) = (1,0)$, then $\# W_0 \leq 2 m_2 = 0$. Therefore it follows that $G$ is a graph consisting of a single edge with isolated vertices and thus $G$ is unmixed.
If $(m_1, m_2) = (0,1)$, then $\# W_0 \leq 2 m_2 = 2$. Also, since $m_2 = 1 > 0$, we have $\# W_0 > 0$. Hence $\# W_0 = 1, 2$. We first assume that $\# W_0 = 1$. Since $\deg_G x_{11}, \deg_G x_{12} \geq 2$, it follows that $G$ is a triangle with isolated vertices. Thus it is unmixed. We next assume that $\# W_0 = 2$. Put $W_0 = \{ w_1, w_2 \}$. Take $M_2 = \{ x_{12} \}$. Then $M_2$ satisfies the conditions $(\ast 1)$ and $(\ast 2)$. By $(\flat 2)$, we have $$\begin{aligned}
2 = \# W_0 &\leq 2 m_2 - \# N_G (M_2) + \# M_2 + \# IN_G (M_2, W) \\
&= 2 - \# N_G (x_{12}) + 1 + \# IN_G (\{ x_{12} \}, W).
\end{aligned}$$ Hence $\deg_G x_{12} = \# N_G (x_{12}) \leq \# IN_G (\{ x_{12} \}, W) + 1$. Note that $\deg_G x_{12} = 2,3$. If $\deg_G x_{12} = 3$, then $\{ x_{12}, w_1 \}, \{ x_{12}, w_2 \} \in E(G)$ and $IN_G (\{ x_{12}\}, W) = \emptyset$. This contradicts to $\deg_G x_{12} \leq \# IN_G (\{ x_{12} \}, W) +1$. Hence $\deg_G x_{12} = 2$. The same is true for $x_{11}$. Therefore we conclude that the edge set of $G_{W_0 \cup M}$ is, by renumbering the vertices, $\{ \{ x_{11}, w_1 \}, \{ x_{12}, w_2 \}, \{ x_{11}, x_{12} \} \}$, and thus $G$ is unmixed.
We next assume that $m \geq 2$. Since $C_0$ is a minimal vertex cover of $G$ with cardinality $m_1 + 2 m_2$, it is sufficient to prove that the cardinality of any minimal vertex cover of $G$ is $m_1 + 2 m_2$. Let $C$ be a minimal vertex cover of $G$ which is not of the form $C_0$. Then there exists a vertex in $M$, say $x_{m 2}$, with $x_{m 2} \notin C$ such that $\deg_G x_{m 2} \geq 2$. Then $N_G (x_{m 2}) \subset C$. As noted in Lemma \[claim:min-ver-cov\] below, we have that $C \setminus N_G (x_{m 2})$ is a minimal vertex cover of $G \setminus N_G [x_{m 2}]$.
Put $G'' := G \setminus N_G [x_{m 2}]$. Then $G''$ is also a graph with a dominating induced matching. Let $W'' \sqcup M''$ be the decomposition of the vertex set $V'' = V(G'')$ induced by the decomposition $V=W \sqcup M$. Note that $m_1'' + m_2'' = m'' = m - 1$. Then it is sufficient to prove that $V'' = W'' \sqcup M''$ also satisfies the condition $(\flat)$ for $G''$. Indeed, when this is the case, it follows that $G''$ is unmixed by inductive hypothesis. Therefore $$\# (C \setminus N_G (x_{m 2})) = m_1'' + 2 m_2''.$$ Consider the condition $(\flat)$ with $\{ x_{m 2} \}$. By $(\flat 1)$, we have $$m_2 - m_2'' = \# N_G (x_{m 2}) - 1.$$ Hence $$\begin{aligned}
\# C &= \# (C \setminus N_G (x_{m 2})) + \# N_G (x_{m 2}) \\
&= (m_1'' + 2 m_2'') + (m_2 - m_2'' + 1) \\
&= m_1'' + m_2'' + 1 + m_2 \\
&= m + m_2 \\
&= m_1 + 2 m_2,
\end{aligned}$$ as required.
Now we prove that $V'' = W'' \sqcup M''$ also satisfies the condition $(\flat)$ for $G''$.
Let $M_2''$ be a subset of $M''$ satisfying $(\ast 1)$ and $(\ast 2)$ for $G''$. We need to prove that $(\flat 1)$ and $(\flat 2)$ are satisfied. In order to prove $(\flat 1)$, we use the description of the left-hand side of $(\flat 1)$ as in Remark \[rmk:GV\] (2). Put $M_2 = M_2'' \cup \{ x_{m 2} \}$. Note that $M_2$ and $\{ x_{m2} \}$ satisfy $(\ast 1)$ and $(\ast 2)$ for $G$. Also note that the right-hand side of $(\flat 1)$ for $(G, M_2)$ is $$\label{eq:neighbour}
\# N_G (M_2) - \# M_2
= (\# N_{G''} (M_2'') - \# M_2'') + (\# N_G (x_{m 2}) - 1)$$ because $\# M_2 = \# M_2'' + 1$ and $\# N_G (M_2) = \# N_{G''} (M_2'') + \# N_G (x_{m 2})$.
Now, let $j$ be an index with $\deg_G x_{j1} \geq 2$ and $\deg_G x_{j2} \geq 2$. Recall that the left-hand side of $(\flat 1)$ for $(G, M_2)$ is the number of $j$ for which one of the condition (i), (ii), (iii) inside $\mathcal{I}_{G, M_2}$ is satisfied. We compare the satisfaction of the condition for the pair $(G, M_2)$ with that for the pair $(G'', M_2'')$. If $j \neq m$, and $\deg_{G''} x_{j1} \geq 2$ and $\deg_{G''} x_{j2} \geq 2$, then the satisfaction of each of the conditions (i), (ii), (iii) inside $\mathcal{I}_{G, M_2}$ and $\mathcal{I}_{G'', M_2''}$ is equivalent. If $j=m$, then note that $x_{m 2} \in \{ x_{m2} \} \subset M_2$, that is $\{ x_{m 2} \} \cap \{ x_{m1}, x_{m2} \} \neq \emptyset$ as well as $M_2 \cap \{ x_{m1}, x_{m2} \} \neq \emptyset$, which corresponds to the condition (i) inside $\mathcal{I}_{G, \{ x_{m2} \}}$, $\mathcal{I}_{G, M_2}$, respectively. If $j \neq m$, and $\deg_{G''} x_{j1} \leq 1$ or $\deg_{G''} x_{j2} \leq 1$, then one of the following is satisfied: $$\begin{aligned}
N_G (x_{j1}) \setminus \{ x_{j2} \}
&\subset N_G (x_{m 2}) \subset N_G (M_2), \\
N_G (x_{j2}) \setminus \{ x_{j1} \}
&\subset N_G (x_{m 2}) \subset N_G (M_2).
\end{aligned}$$ These correspond to the conditions (ii), (iii) inside $\mathcal{I}_{G, \{ x_{m2} \} }$, $\mathcal{I}_{G, M_2}$. Note that when $j \neq m$, and $\deg_{G''} x_{j1} \geq 2$ and $\deg_{G''} x_{j2} \geq 2$, the cases (i), (ii), (iii) inside $\mathcal{I}_{G, \{ x_{m2} \} }$ do not occur. Combining these with Remark \[rmk:GV\] (2), we have that the lefthand-side of ($\flat 1$) for $M_2$ with respect to $G$ is equal to the sum of the lefthand-side of ($\flat 1$) for $\{ x_{m 2} \}$ with respect to $G$ and the lefthand-side of ($\flat 1$) for $M_2''$ with respect to $G''$. Hence by assumption for $G$, we have that the lefthand-side of ($\flat 1$) for $M_2''$ with respect to $G''$ is equal to $$(\# N_G (M_2) - \# M_2)
- (\# N_G (\{ x_{m 2} \}) - \# \{ x_{m 2} \})$$ By (\[eq:neighbour\]), it is equal to $$\# N_{G''} (M_2'') - \# M_2'',$$ as desired.
Finally we prove the inequality ($\flat 2$) for $M_2''$ with respect to $G''$. Let $W_0''$ be the set of vertices in $V'' \cap W'' = V'' \cap W$ which are not isolated in $G''$. Then $$\# W_0'' = \# W_0 - (\# N_G (x_{m 2}) - 1) - \# IN_G (\{ x_{m 2} \}, W).$$ Also $$\begin{aligned}
\# N_{G''} (M_2'') &= \# N_G (M_2) - \# N_G (x_{m2}), \\
\# IN_{G''} (M_2'', W'')
&= \# IN_G (M_2, W) - \# IN_G (\{ x_{m 2} \}, W).
\end{aligned}$$ Furthermore, it follows from the assumption $(\flat 1)$ for $\{ x_{m 2} \}$ with respect to $G$ that $m_2'' = m_2 - \# N_G (x_{m 2}) + 1$. Then $$\begin{aligned}
&2 m_2'' - \# N_{G''} (M_2'') + \# M_2'' + \# IN_{G''} (M_2'', W'')
- \# W_0'' \\
&= 2 (m_2 - \# N_G (x_{m 2}) + 1)
- (\# N_G (M_2) - \# N_G (x_{m 2}))
+ (\# M_2 - 1) \\
&+ (\# IN_G (M_2, W) - \# IN_G (\{ x_{m 2} \}, W))
- (\# W_0 - (\# N_G (x_{m 2}) - 1) - \# IN_G (\{ x_{m 2} \}, W)) \\
&= 2 m_2 - \# N_G (M_2) + \# M_2 + \# IN_G (M_2, W) - \# W_0 \geq 0
\end{aligned}$$ by the assumption $(\flat 2)$ for $M_2$ with respect to $G$. Hence $(\flat 2)$ for $M_2''$ with respect to $G''$ is also satisfied as desired.
\[claim:min-ver-cov\] $C \setminus N_G (x_{m 2})$ is a minimal vertex cover of $G \setminus N_G [x_{m 2}]$.
We first prove that $C \setminus N_G (x_{m 2})$ is a vertex cover of $G \setminus N_G [x_{m 2}]$. Let $e$ be an edge of $G \setminus N_G [x_{m 2}]$. Then $e \cap N_G [x_{m 2}] = \emptyset$. Also, since $C$ is a vertex cover of $G$, it follows that $e \cap C \neq \emptyset$. Combining these facts we have $e \cap (C \setminus N_G (x_{m 2})) \neq \emptyset$.
We next prove the minimality of $C \setminus N_G (x_{m 2})$. Assume that $C' \subsetneq C \setminus N_G (x_{m 2})$ is a vertex cover of $G \setminus N_G [x_{m 2}]$. Then $C' \cup N_G (x_{m 2}) \subsetneq C$. We derive a contradiction by proving that $C'' := C' \cup N_G (x_{m 2})$ is a vertex cover of $G$.
Let $e$ be an edge of $G$. If $e \cap N_G (x_{m 2}) \neq \emptyset$, then $e \cap C'' \neq \emptyset$. If $e \cap N_G (x_{m 2}) = \emptyset$, then $x_{m2} \notin e$ and $e$ is an edge of $G \setminus N_G [x_{m 2}]$. Since $C'$ is a vertex cover of $G \setminus N_G [x_{m 2}]$, it follows that $e \cap C' \neq \emptyset$. Therefore $e \cap C'' \neq \emptyset$.
A graph $G$ is called *forest* if $G$ has no cycle. The chordalness of a graph with a dominating induced matching is characterized as follows:
\[chordal-vertex-edge-decomp\] Let $G$ be a finite simple graph on $V$ with a dominating induced matching. Let $$\mathcal{M} = \{ \{ x_{j1}, x_{j2} \} \; : \;
i=1, 2, \ldots, m \}$$ be a matching of $G$ so that $$W = V \setminus \bigcup_{j=1}^m \{ x_{j1}, x_{j2} \}$$ is an independent set of $G$.
Let $\widetilde{G}$ be the graph obtained by identifying $x_{j1}$ and $x_{j2}$ for $j=1, 2, \ldots, m$. That is $\widetilde{G}$ is a graph on the vertex set $\widetilde{V} := W \cup \{ x_{1}, \ldots, x_m \}$ with the edge set $$E(\widetilde{G}) = \{ \{ w, x_j \} \; : \;
\text{$\{ w, x_{j1} \} \in E(G)$ or $\{ w, x_{j2} \} \in E(G)$} \}.$$
Then $G$ is chordal if and only if $\widetilde{G}$ is a forest.
We first prove that if $G$ is not chordal, then $\widetilde{G}$ is not a forest, in other words, $\widetilde{G}$ has a cycle. Assume that $G$ has a chordless cycle $C$ of length $\ell$ with $\ell > 3$. Let $\widetilde{C}$ be the subgraph of $\widetilde{G}$ obtained from $C$ by the same operation as we obtain $\widetilde{G}$ from $G$. If there is no $j$ such that both of $x_{j1}, x_{j2}$ are vertices of $C$, then $\widetilde{C}$ is also a cycle. Hence $\widetilde{G}$ has a cycle. If both of $x_{j1}, x_{j2}$ are vertices of $C$, then these must be adjacent in $C$ because $C$ is a chordless cycle. Since $\ell > 3$, the other adjacent vertices $y_{i_1}$, $y_{i_2}$ of $x_{j1}, x_{j2}$ are different. Note that $y_{i_1}, y_{i_2} \in W$. Then $y_{i_1}, y_{i_2}, x_{j}$ are vertices of $\widetilde{C}$. It then follows that $\widetilde{C}$ is a cycle of $\widetilde{G}$.
Next suppose that $G$ is chordal. Assume that $\widetilde{G}$ has a cycle. Let $\widetilde{C}$ be a minimal cycle of $\widetilde{G}$ and let $\ell$ be the length of $\widetilde{C}$. Since $\widetilde{G}$ is a bipartite graph, $\ell$ must be even and thus $\ell \geq 4$. Let $C$ be a cycle of $G$ corresponding to $\widetilde{C}$ with the minimum length. Then the length of $C$ is greater than or equal to $\ell \geq 4$. Since $G$ is chordal, $C$ must have a chord $e$. We may assume that $e = \{ w, x_{j1} \}$ where $w \in W$. Since $C$ is a cycle, there are two paths from $w$ to $x_{j1}$; we take with the shorter length; let $y_0 = w, y_1, y_2, \ldots, y_k = x_{j1}$ be a sequence of vertices of such path in $C$ where $k \geq 2$ and $\{ y_i, y_{i+1} \} \in E(C)$ for $i = 0, 1, \ldots, k-1$. If $k > 2$, then $\{ w, x_{j1} \}$ must be a chord of $\widetilde{C}$, a contradiction. If $k=2$, then $y_1 = x_{j2}$ and $C \setminus \{ y_1 \}$ is also a cycle corresponding to $\widetilde{C}$. This contradicts to the minimality of $C$.
Some algebraic properties
=========================
In this section, we investigate algebraic properties of the edge ideal of a graph with a dominating induced matching.
In [@HHKO], it is proved that a Cameron–Walker graph is vertex decomposable, in particular, it is sequentially Cohen–Macaulay. But there is a graph with a dominating induced matching which is not sequentially Cohen–Macaulay; the $6$-cycle is such an example; see [@FVT Proposition 4.1].
We obtain some class of vertex decomposable graphs among graphs with a dominating induced matching.
Recall that a graph $G$ on $V$ is called vertex decomposable (see [@Woodroofe-vertex-decomp Lemma 4]) if $G$ is an edgeless graph or there exists $v \in V$ with the following $2$ properties:
1. $G \setminus v$ and $G \setminus N[v]$ are vertex decomposable;
2. no independent set in $G \setminus N[v]$ is a maximal independent set in $G \setminus v$.
We call $v \in V$ a *shedding vertex* of $G$ if $v$ satisfies (VD2). Note that for a vertex $v \in V$, if there exists $w \in V$ such that $N[w] \subset N[v]$, then $v$ is a shedding vertex ([@Woodroofe-vertex-decomp Lemma 6]).
\[DIM-VD\] Let $G$ be a finite simple graph on $V$ with a dominating induced matching. Assume that there exists a decomposition $V = W \sqcup M$ satisfying the following property, where $W = \{ y_1, \ldots, y_r \}$ is an independent set and $G_M$ consists of $m$ disconnected edges $\{ x_{j1}, x_{j2} \}$, $j = 1, \ldots, m$. Moreover assume that for each $j=1, \ldots, m$, one of the following is satisfied:
1. $\deg_G x_{j1} = 1$ or $\deg_G x_{j2} = 1$;
2. $\deg_G x_{j1} = \deg_G x_{j2} = 2$ and there is $y_{i_j} \in W$ such that $x_{j1}, x_{j2} \in N_G(y_{i_j})$;
3. $\deg_G x_{jk} = 3$ and $\deg_G x_{jl} = 2$ where $\{ k, l \} = \{ 1, 2 \}$, and there is $y_{i_j} \in W$ such that $N_G (y_{i_j}) = \{ x_{j1}, x_{j2} \}$;
4. $\deg_G x_{j1} = \deg_G x_{j2} = 3$ and there are distinct three vertices $y_{i_{j1}}, y_{i_{j2}}, y_{i_{j3}} \in W$ such that $\{ x_{j1}, y_{i_{j1}} \}, \{ x_{j2}, y_{i_{j2}} \} \in E(G)$, $N_G (y_{i_{j3}}) = \{ x_{j1}, x_{j2} \}$, and there is a pendant triangle attached to at least one of $y_{i_{j1}}, y_{i_{j2}}$.
Then $G$ is vertex decomposable.
\[rmk:DIM-VD-pendant-trianlgle\] The each condition (ii), (iii), (iv) of Theorem \[DIM-VD\] is concerned with the existence of a pendant triangle.
Indeed, the condition (ii) means that $G$ has a pendant triangle attached to $y_{i_j}$; the condition (iii) means that $G$ has a pendant triangle attached to $x_{jk}$; the condition (iv) is explicit.
In order to prove Theorem \[DIM-VD\], we use the following lemma.
\[pendant-triangle-shedding\] Let $G$ be a finite simple graph on $V$. Suppose that $G$ has a pendant triangle attached to $v \in V$. Then $v$ is a shedding vertex.
Let $v_1, v_2$ be the two degree $2$ vertices of a pendant triangle attached to $v$. Then $N[v_1] = \{ v, v_1, v_2 \} \subset N[v]$ holds. Hence $v$ is a shedding vertex; see before Theorem \[DIM-VD\].
Now we prove Theorem \[DIM-VD\].
We use induction on $r = \# W$. If $r=1$, then $G$ is chordal and thus $G$ is vertex decomposable by Woodroofe [@Woodroofe-vertex-decomp Corollary 7].
Suppose that $r \geq 2$. If the cases (iii) and (iv) do not occur, then $G$ is a Cameron–Walker graph and thus, $G$ is vertex decomposable by [@HHKO Theorem 3.1].
If there is an edge $\{ x_{j1}, x_{j2} \}$ with the condition (iii), say, $\deg x_{j1} = 3$ and $\deg x_{j2} = 2$, then $x_{j1}$ is a shedding vertex because of Remark \[rmk:DIM-VD-pendant-trianlgle\] and Lemma \[pendant-triangle-shedding\]. Therefore we only need to prove that both $G \setminus x_{j1}$ and $G \setminus N[x_{j1}]$ are vertex decomposable. Indeed $G \setminus x_{j1}$ is the disjoint union of single edge $\{ x_{j2}, y_{i_j} \}$ and $G' := G \setminus \{ x_{j1}, x_{j2}, y_{i_j} \}$. Since the vertex set of $G'$ can be decomposed as $W' \sqcup M'$ where $W' = W \setminus \{ y_{i_j} \}$ and $M' = M \setminus \{ x_{j1}, x_{j2} \}$, $G'$ has a dominating induced matching. Also $G'$ satisfies the assumption of the theorem with this decomposition of the vertex set since $N_G (y_{i_j}) = \{ x_{j1}, x_{j2} \}$. Hence we conclude that $G'$, and thus $G \setminus x_{j1}$ is vertex decomposable by inductive hypothesis. Also the vertex set of $G \setminus N[x_{j1}]$ can be decomposed as $W'' \sqcup M''$ where $W'' = W \setminus N(x_{j1})$ and $M'' = M \setminus \{ x_{j1}, x_{j2} \}$. Thus $G \setminus N[x_{j1}]$ has a dominating induced matching. Since $\deg_G x_{j' k} \geq \deg_{G \setminus N[x_{j1}]} x_{j'k}$, we can easily see that this decomposition satisfies the assumption of the theorem. Hence by inductive hypothesis, we conclude that $G \setminus N[x_{j1}]$ is also vertex decomposable.
Suppose that there is an edge $\{ x_{j1}, x_{j2} \}$ with the condition (iv). We may assume that $G$ has a pendant triangle attached to $y_{i_{j1}}$. Then $y_{i_{j1}}$ is a shedding vertex by Lemma \[pendant-triangle-shedding\]. Hence it is enough to prove that both $G \setminus y_{i_{j1}}$ and $G \setminus N [y_{i_{j1}}]$ are vertex decomposable. We first consider $G \setminus y_{i_{j1}}$. Since the vertex set of this graph can be decomposed as $(W \setminus \{ y_{i_{j1}} \}) \sqcup M$, this graph has a dominating induced matching. We check that each $j' = 1, \ldots, m$ satisfies one of the conditions (i), (ii), (iii), (iv) with respect to $G \setminus y_{i_{j1}}$. If $j'$ satisfies the condition (i) (resp. (ii)) with respect to $G$, then $j'$ satisfies the condition (i) (resp. (ii) or (i)) with respect to $G \setminus y_{i_{j1}}$. Assume that $j'$ satisfies the condition (iii) (resp. (iv)) with respect to $G$. Since $\deg_G y_{i_{j1}} \geq 3$ and $\deg_G y_{i_{j'}} = 2$ (resp. $\deg_G y_{i_{j' 3}} = 2$), the vertex $y_{i_{j1}}$ is different from $y_{i_{j'}}$ (resp. $y_{i_{j' 3}}$). Hence $j'$ satisfies the condition (iii) or (ii) (resp. (iv) or (iii)). Therefore this decomposition satisfies the assumption of the theorem. Hence by inductive hypothesis, we conclude that $G \setminus y_{i_{j1}}$ is vertex decomposable. We next consider $G \setminus N[y_{i_{j1}}]$. In this case, the vertex set of $G \setminus N[y_{i_{j1}}]$ is decomposed as $W' \sqcup M'$ where $$\begin{aligned}
W' &= W \setminus \left( \{ y_{i_{j1}} \} \cup
\bigcup_{\genfrac{}{}{0pt}{}{\text{$j'$ satisfying (iii)}}{\{ x_{j' 1}, x_{j' 2} \} \cap N_G (y_{i_{j1}}) \neq \emptyset}} \{ y_{i_{j'}} \}
\cup
\bigcup_{\genfrac{}{}{0pt}{}{\text{$j'$ satisfying (iv)}}{ \# (\{ x_{j' 1}, x_{j' 2} \} \cap N_G (y_{i_{j1}})) = 1}} \{ y_{i_{j' 3}} \} \right), \\
M' &= (M \setminus N_G (y_{i_{j1}})) \cup
\bigcup_{\genfrac{}{}{0pt}{}{\text{$j'$ satisfying (iii)}}{\{ x_{j' 1}, x_{j' 2} \} \cap N_G (y_{i_{j1}}) \neq \emptyset}} \{ y_{i_{j'}} \}
\cup
\bigcup_{\genfrac{}{}{0pt}{}{\text{$j'$ satisfying (iv)}}{ \# (\{ x_{j' 1}, x_{j' 2} \} \cap N_G (y_{i_{j1}})) = 1}} \{ y_{i_{j' 3}}\}.
\end{aligned}$$ Then we can easily see that $G \setminus N_G [y_{i_{j1}}]$ has a dominating induced matching. For example, let $j'$ be an index satisfying (iii) and $\{ x_{j' 1}, x_{j' 2} \} \cap N_G (y_{i_{j1}}) \neq \emptyset$, say $x_{j' 1} \in N_G (y_{i_{j1}})$. Then $x_{j' 1} \notin M'$ but $y_{i_{j'}} \in M$ and $\{ x_{j' 2}, y_{i_{j'}} \}$ is an edge of $G \setminus N[y_{i_{j1}}]$. Note that in this case, $\deg_{G \setminus N [y_{i_{j1}}]} y_{i_{j'}} = 1$. Then it is also easy to see that the assumption of the theorem is satisfied with this decomposition. Therefore $G \setminus N[y_{i_{j1}}]$ is vertex decomposable by inductive hypothesis.
Although we provide the characterization for a graph with a dominating induced matching to be unmixed in Theorem \[claim:unmixedDIM\], we can obtain a clearer characterization for the unmixedness of the class of graphs in Theorem \[DIM-VD\]. It is sufficient to consider a connected graph which is not a single edge.
\[DIM-VD-unmixed\] Let $G$ be a finite simple connected graph on $V$ with a dominating induced matching. Assume that there exists a decomposition $V = W \sqcup M$ satisfying the assumption of Theorem \[DIM-VD\] (where $W \neq \emptyset$). We use the same notation as in Theorem \[DIM-VD\] and before Theorem \[claim:unmixedDIM\]. Then $G$ is Cohen–Macaulay if and only if $\# W = m_2$ and for all $y_i \in W$, there is just one edge $\{ x_{j_i 1}, x_{j_i 2} \}$ of $G$ such that both $\{ x_{j_i 1}, y_i \}$ and $\{ x_{j_i 2}, y_i \}$ are edges of $G$.
We first note that $G$ is Cohen–Macaulay if and only if $G$ is unmixed because $G$ is vertex decomposable by Theorem \[DIM-VD\].
[**(“Only If”)**]{} Let $M_2$ be the union of the following subsets $V_1, \ldots, V_4$ of $M$: $V_1$ is the set of the vertices $x_{jk} \in V$ where $\{ x_{j1}, x_{j2} \}$ is an edge of type (i) of Theorem \[DIM-VD\] with $\deg x_{jk} \geq \deg x_{jl} = 1$ ($\{ k,l \} = \{ 1,2 \}$); $V_2$ is the set of the vertices $x_{j1}$ where $\{ x_{j1}, x_{j2} \}$ is an edge of type (ii) of Theorem \[DIM-VD\]; $V_3$ is the set of the vertices $x_{jk}$ where $\{ x_{j1}, x_{j2} \}$ is an edge of type (iii) of Theorem \[DIM-VD\] with $\deg x_{jk} = 3$ and $\deg x_{jl} = 2$ ($\{ k,l \} = \{ 1,2 \}$); $V_4$ is the set of the vertices $x_{jk}$ where $\{ x_{j1}, x_{j2} \}$ is an edge of type (iv) of Theorem \[DIM-VD\] and the numbers of pendant triangles attached to each $y_{i_j k}$ is less than or equal to that of to $y_{i_j l}$ with the notation in Theorem \[DIM-VD\] (iv) ($\{ k,l \} = \{ 1,2 \}$). Clearly, $M_2$ satisfies the condition $(\ast 1)$. The condition $(\ast 2)$ is also satisfied because $G$ is connected. Note that $\# M_2 = m$, in particular $m_2' = 0$. Since $G$ is unmixed, by $(\flat 1)$ of Theorem \[claim:unmixedDIM\], it follows that $$m_2 = \# N_G (M_2) - \# M_2.$$ Also $W \subset N_G (M_2)$ holds. Actually, take $y_i \in W$. Since $G$ is connected, there exists an edge $\{ x_{j1}, x_{j2} \}$ such that $\{ x_{jk}, y_i \}$ is an edge of $G$. If $\{ x_{j1}, x_{j2} \}$ is of type (i) or (ii) of Theorem \[DIM-VD\], then it is easy to see that $y_i \in N_G (M_2)$. If $\{ x_{j1}, x_{j2} \}$ is of type (iii) of Theorem \[DIM-VD\] and $x_{jk} \notin M_2$, then $x_{j \ell} \in M_2$ and $\deg_G x_{jk} = 2$. It then follows that $y_i \in N_G (x_{j \ell}) \subset N_G (M_2)$. If $\{ x_{j1}, x_{j2} \}$ is of type (iv) of Theorem \[DIM-VD\] and $x_{jk} \notin M_2$, then there is a pendant triangle attached to $y_i$. Let $x_{j' 1}, x_{j' 2}$ be the two vertices of the pendant triangle of degree $2$. Since $\{ x_{j' 1}, x_{j' 2} \}$ is of type (ii) of Theorem \[DIM-VD\], it follows that $y_{i} \in N_G (M_2)$.
The inclusion $W \subset N_G (M_2)$ implies that $N_G (M_2) = V(G) \setminus M_2$. Therefore $$\begin{aligned}
m_2 &= \#N_G (M_2) - \# M_2 \\
&= (\# V(G)- \# M_2) - \# M_2 \\
&= \# V(G) - 2 \# M_2 \\
&= (2m + \# W) - 2 m \\
&= \# W.
\end{aligned}$$
Suppose that there exists $y_i \in W$ such that $\{ x_{j 1}, y_i \}, \{ x_{j 2}, y_i \},
\{ x_{j' 1}, y_i \}, \{ x_{j' 2}, y_i \} \in E(G)$ for $j \neq j'$. It then follows that both $\{ x_{j1}, x_{j2} \}$ and $\{ x_{j' 1}, x_{j' 2} \}$ are of type (ii) of Theorem \[DIM-VD\]. In particular, both $\{ y_i, x_{j1}, x_{j2} \}$ and $\{ y_i, x_{j' 1}, x_{j' 2} \}$ form pendant triangles attached to $y_i$. Assume that there are $\alpha \geq 2$ pendant triangles attached to $y_i$; set the two degree $2$ vertices of each pendant triangle as $\{ x_{j_k 1}, x_{j_k 2} \}$, $k=1, 2, \ldots, \alpha$. Put $M_2'' = \{ x_{j_1 1}, x_{j_2 1}, \ldots, x_{j_\alpha 1} \}$. Then $M_2''$ satisfies the condition $(\ast 1)$ and $(\ast 2)$. Note that $N_G (M_2'') = \{ y_i, x_{j_1 2}, x_{j_2 2}, \ldots, x_{j_\alpha 2} \}$ and $m_2 - m_2'' = \alpha$. Therefore by $(\flat 1)$ of Theorem \[claim:unmixedDIM\], we have $$\alpha = m_2 - m_2'' = \# N_G (M_2'') - \# M_2''
= (\alpha + 1) - \alpha = 1,$$ this contradict to $\alpha \geq 2$. Since $m_2 = \# W$, the assertion follows.
[**(“If”)**]{} Let $X$ be a maximal independent set of $G$. In order to prove that $G$ is unmixed, it is sufficient to show that $\# X = m$. Set $$\begin{aligned}
E_1 &:= \{ \{ x_{j1}, x_{j2} \} \in E(G) \; : \;
\text{$\deg_G x_{j1} = 1$ or $\deg_G x_{j2} = 1$} \}, \\
E_2 &:= \{ \{ x_{j1}, x_{j2} \} \in E(G) \; : \;
\text{$\deg_G x_{j1} \geq 2$ and $\deg_G x_{j2} \geq 2$} \}
\end{aligned}$$ and $V_k := \bigcup_{e \in E_k} e$ for $k=1,2$. Since $V = V_1 \sqcup V_2 \sqcup W$, $$\# X = \# (X \cap V_1) + \# (X \cap (V_2 \cup W)).$$
Take $\{ x_{j1}, x_{j2} \} \in E_1$. Assume that $\deg_G x_{j1} = 1$. Then $\deg_G x_{j2} \geq 2$ because $G$ is connected. If $x_{j2} \notin X$, then $x_{j1} \in X$ because the maximality of $X$. This implies $\# (X \cap \{ x_{j1}, x_{j2} \}) = 1$ and $\# (X \cap V_1) = m_1$.
By assumption, $V_2 \sqcup W$ can be decomposed as $\bigcup_{i=1}^r \{ y_i, x_{j_i 1}, x_{j_i 2} \}$. We claim that $X \cap \{ y_i, x_{j_i 1}, x_{j_i 2} \} = 1$ for each $i$.
Since $\{ y_i, x_{j_i 1} \}, \{ y_i, x_{j_i 2} \},
\{ x_{j_i 1}, x_{j_i 2} \} \in E(G)$, it follows that $\# (X \cap \{ y_i, x_{j_i 1}, x_{j_i 2} \}) \leq 1$. Assume that $y_i \notin X$. If $\deg y_i \geq 3$, then $\deg_G x_{j_i 1} = \deg_G x_{j_i 2} = 2$ and it follows from the maximality of $X$ that exactly one of $x_{j_i 1}, x_{j_i 2}$ belongs to $X$. When $\deg y_i = 2$, if neither $x_{j_i 1}$ nor $x_{j_i 2}$ belong to $X$, then $X \cup \{ y_i \}$ is also an independent set. This contradicts to the maximality of $X$. Thus (exactly) one of $x_{j_i 1}, x_{j_i 2}$ belongs to $X$. Hence $\# (X \cap (V_2 \cup W)) = m_2 = \# W$.
Therefore $$\# X = \# (X \cap V_1) + \# (X \cap (V_2 \cup W))
= m_1 + m_2 = m,$$ as desired.
We show an example satisfying the assumption of Theorem \[DIM-VD\].
The graph $G$ in Figure \[fig:CMDIM\] has a dominating induced matching, which is not a Cameron–Walker graph. It also satisfies the assumption of Theorem \[DIM-VD\] with the displayed decomposition of the vertex set.
Then $m_2 = \# W = 3$. We can also easily see that this graph satisfies the assumption for the vertex in $W$ of Theorem \[DIM-VD-unmixed\]. Hence $G$ is Cohen–Macaulay by Theorem \[DIM-VD-unmixed\].
![A Cohen–Macaulay graph with a dominating induced matching[]{data-label="fig:CMDIM"}](C_M_dim.eps){width="120mm"}
We close the paper by giving some more examples of a graph with a dominating induced matching which does not satisfy the assumption on Theorem \[DIM-VD\].
We first show some Cohen–Macaulay graphs with a dominating induced matching.
1. The path graph $P_4$ with $4$ vertices is a Cohen–Macaulay graph. Also it has a dominating induced matching. Indeed, set $V(P_4) = \{ 1,2,3,4 \}$ and $E(P_4) = \{ \{ 1,2 \}, \{ 2,3 \}, \{ 3,4 \} \}$. Then we see that $P_4$ has a dominating induced matching with the decomposition $V(P_4) = W \sqcup M$ where $W = \{ 1,4 \}$ and $M = \{ 2,3 \}$.
2. The graph $G_1$ on the vertex set $\{ 1,2,3,4,5,6 \}$ whose edge set is $$E(G_1)
= \{ \{ 1,3 \}, \{ 1,4 \}, \{ 1,5 \}, \{ 2,4 \}, \{ 2,5 \}, \{ 2,6 \},
\{ 3,4 \}, \{ 5,6 \} \}$$ is a graph with a dominating induced matching. This is Cohen–Macaulay, in particular, unmixed.
(100,60) (0,45)[$G_1$:]{} (20,20) (40,20) (60,20) (80,20) (40,40) (60,40) (15,10)[$3$]{} (40,10)[$4$]{} (60,10)[$5$]{} (80,10)[$6$]{} (35,45)[$1$]{} (60,45)[$2$]{} (40,40)[(-1,-1)[20]{}]{} (40,40)[(0,-1)[20]{}]{} (40,40)[(1,-1)[20]{}]{} (60,40)[(-1,-1)[20]{}]{} (60,40)[(0,-1)[20]{}]{} (60,40)[(1,-1)[20]{}]{} (40,20)[(-1,0)[20]{}]{} (60,20)[(1,0)[20]{}]{}
We next show an unmixed graph with a dominating induced matching but not Cohen–Macaulay.
The graph $G_2$ on the vertex set $\{ 1,2,3,4,5,6 \}$ whose edge set is $$\{ \{ 1,3 \}, \{ 1,4 \}, \{ 1,5 \}, \{ 1,6 \}, \{ 2,3 \}, \{ 2,4 \},
\{ 2,5 \}, \{ 2,6 \}, \{ 3,4 \}, \{ 5,6 \} \}$$ is a graph with a dominating induced matching. This is unmixed but not Cohen–Macaulay.
(100,60) (0,45)[$G_2$:]{} (20,20) (40,20) (60,20) (80,20) (40,40) (60,40) (15,10)[$3$]{} (40,10)[$4$]{} (60,10)[$5$]{} (80,10)[$6$]{} (35,45)[$1$]{} (60,45)[$2$]{} (40,40)[(-1,-1)[20]{}]{} (40,40)[(0,-1)[20]{}]{} (40,40)[(1,-1)[20]{}]{} (60,40)[(-1,-1)[20]{}]{} (60,40)[(0,-1)[20]{}]{} (60,40)[(1,-1)[20]{}]{} (40,20)[(-1,0)[20]{}]{} (60,20)[(1,0)[20]{}]{} (40,40)[(2,-1)[40]{}]{} (60,40)[(-2,-1)[40]{}]{}
Finally, we show a sequentially Cohen–Macaulay graph with a dominating induced matching but not unmixed.
The graph $G_3$ on the vertex set $\{ 1,2,3,4,5,6 \}$ whose edge set is $$\{ \{ 1,3 \}, \{ 3,4 \}, \{ 2,4 \}, \{ 2,5 \}, \{ 2,6 \}, \{ 5,6 \} \}$$ is a graph with a dominating induced matching.
(100,60) (0,45)[$G_3$:]{} (20,20) (40,20) (60,20) (80,20) (40,40) (60,40) (15,10)[$3$]{} (40,10)[$4$]{} (60,10)[$5$]{} (80,10)[$6$]{} (35,45)[$1$]{} (60,45)[$2$]{} (40,40)[(-1,-1)[20]{}]{} (60,40)[(-1,-1)[20]{}]{} (60,40)[(0,-1)[20]{}]{} (60,40)[(1,-1)[20]{}]{} (40,20)[(-1,0)[20]{}]{} (60,20)[(1,0)[20]{}]{}
This is not unmixed since both $\{ 3,4,5,6 \}$ and $\{ 2,3,6 \}$ are minimal vertex covers of $G_3$. Also, since $G_3$ is chordal, it is sequentially Cohen–Macaulay by Francisco and Van Tuyl [@FVT].
The third author is partially supported by JSPS Grant-in-Aid for Young Scientists (B) 24740008.
We thank anonymous referees for reading the manuscript carefully.
[30]{} T. Biyikoğlu and Y. Civan, Vertex decomposable graphs, codismantlability, Cohen–Macaulayness and Castelnuovo-Mumford regularity, *Electron. J. Combin.* [**21**]{} (2014), 17pp. T. Biyikoğlu and Y. Civan, Bounding Castelnuovo-Mumford regularity of graphs via Lozin’s transformation, preprint, arXiv:1302.3064v1.
T. Biyikoğlu and Y. Civan, Castelnuovo-Mumford regularity of graphs, preprint, arXiv:1503.06018.
K. Cameron and T. Walker, The graphs with maximum induced matching and maximum matching the same size, *Discrete Math.* [**299**]{} (2005), 49–55.
D. M. Cardoso, E. A. Martins, L. Media, and O. Rojo, Spectral results for the dominating induced matching problem, preprint, arXiv:1311.2748v1.
H. Dao, C. Huneke and J. Schweig, Bounds on the regularity and projective dimension of ideals associated to graphs, *J. Algebraic Combin.* [**38**]{} (2013), 37–55.
C. A. Francisco and A. Van Tuyl, Sequentially Cohen–Macaulay edge ideals, *Proc. Amer. Math. Soc. * **135** (2007), 2327–2337.
I. Gitler and C. E. Valencia, Bounds for invariants of edge-rings, *Comm. Algebra* **33** (2005), 1603–1616.
H. T. Hà and A. Van Tuyl, Monomial ideals, edge ideals of hypergraphs, and their graded Betti numbers, *J. Algebraic Combin. * **27** (2008), 215–245
T. Hibi, A. Higashitani, K. Kimura, and A. B. O’Keefe, Algebraic study on Cameron–Walker graphs, *J. Algebra* [**422**]{} (2015), 257–269. G. Kalai and R. Meshulam, Intersections of Leray complexes and regularity of monomial ideals, *J. Combin. Theory Ser. A* **113** (2006), 1586–1592.
M. Katzman, Characteristic-independence of Betti numbers of graph ideals, *J. Combin. Theory, Ser. A* **113** (2006), 435–454.
F. Khosh-Ahang and S. Moradi, Regularity and projective dimension of edge ideal of $C_5$-free vertex decomposable graphs, *Proc. Amer. Math. Soc.* [**142**]{} (2014), 1567–1576.
K. Kimura, Non-vanishingness of Betti numbers of edge ideals, in: Harmony of Gröbner bases and the modern industrial society, World Scientific, 2012, pp. 153–168.
M. Kummini, Regularity, depth and arithmetic rank of bipartite edge ideals, *J. Algebraic Combin.* [**30**]{} (2009), 429–445.
M. C. Lin, M. J. Mizrahi, J. L. Szwarcfiter, Exact algorithms for dominating induced matchings, preprint, arXiv:1301.7602v2.
G. Lyubeznik, A new explicit finite free resolution of ideals generated by monomials in an $R$-sequence, *J. Pure Appl. Algebra* [**51**]{} (1988), 193–195.
M. Mahmoudi, A. Mousivand, M. Crupi, G. Rinaldo, N. Terai, and S. Yassemi, Vertex decomposability and regularity of very well-covered graphs, *J. Pure Appl. Algebra* [**215**]{} (2011), 2473–2480.
E. Nevo, Regularity of edge ideals of $C_4$-free graphs via the topology of the lcm-lattice, *J. Combin. Theory, Ser. A* [**118**]{} (2011), 491–501.
A. Van Tuyl, Sequentially Cohen–Macaulay bipartite graphs: vertex decomposability and regularity, *Arch. Math. (Basel)* [**93**]{} (2009), 451–459.
R. Woodroofe, Vertex decomposable graphs and obstructions to shellability, *Proc. Amer. Math. Soc. * **137** (2009), 3235–3246.
R. Woodroofe, Matchings, coverings, and Castelnuovo-Mumford regularity, *J. Commut. Algebra* [**6**]{} (2014), 287–304. X. Zheng, *Resolutions of facet ideals*, Comm. Algebra **32** (2004), 2301–2324.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We systematically study the large-$q_T$ (or small-$b$) matching of transverse momentum dependent (TMD) distributions to the twist-2 integrated parton distributions. Performing operator product expansion for a generic TMD operator at the next-to-leading order (NLO) we found the complete set of TMD distributions that match twist-2. These are unpolarized, helicity, transversity, pretzelosity and linearly polarized gluon distributions. The NLO matching coefficients for these distributions are presented. The pretzelosity matching coefficient is zero at the presented order, however, it is evident that it is non-zero in the following orders. This result offers a natural explanation of the small value of pretzelosity found in phenomenological fits. We also demonstrate that the cancellation of rapidity divergences by the leading order soft factor imposes the necessary requirement on the Lorentz structure of TMD operators, which is supported only by the TMD distributions of leading dynamical twist. Additionally, this requirement puts restrictions on the $\gamma_5$-definition in the dimensional regularization.'
address:
- |
Departamento de Física Teórica II, Universidad Complutense de Madrid (UCM),\
28040 Madrid, Spain
- |
Institut für Theoretische Physik, Universität Regensburg,\
D-93040 Regensburg, Germany
author:
- 'Daniel Gutiérrez-Reyes'
- Ignazio Scimemi
- 'Alexey A. Vladimirov'
bibliography:
- 'Refbib.bib'
title: 'Twist-2 matching of transverse momentum dependent distributions'
---
Introduction
============
The transverse momentum dependent (TMD) factorization theorems for semi-inclusive deep inelastic scattering (SIDIS) and Drell-Yan type processes formulated in [@GarciaEchevarria:2011rb; @Collins:2011zzd; @Chiu:2012ir; @Echevarria:2014rua] allow a consistent treatment of rapidity divergences in the definition of spin (in)dependent TMD distributions. They also provide a self-contained definition of TMD operators which can be considered individually by standard methods of quantum field theory without referring to a scattering process. In particular, the large-$q_T$ (or small-$b$) matching of TMD distributions on the corresponding integrated functions can be evaluated. Such consideration is practically very important because the resulting matching coefficients serve as an initial input to many models and phenomenological ansatzes for TMD distributions. The unpolarized TMD distribution is the most studied case and it has been treated using different regularization schemes at the next-to-leading order (NLO) [@GarciaEchevarria:2011rb; @Becher:2010tm; @Collins:2011zzd; @Echevarria:2014rua; @Aybat:2011zv; @Cherednikov:2008ua; @Cherednikov:2007tw] and the next-to-next-to-leading order (NNLO) [@Echevarria:2015usa; @Echevarria:2016scs; @Gehrmann:2014yya; @Gehrmann:2012ze]. For polarized distributions such a program has been performed only for helicity, transversity and linearly polarized distributions at NLO [@Bacchetta:2013pqa; @Echevarria:2015uaa]. However, these works miss a systematic discussion on the relevant renormalization schemes, which are fundamental to establish their calculation and to provide a spring to higher order analysis. By this article we open a series of articles devoted to the study of the small-$b$ matching of polarized TMD distributions. The primary goal of this letter is to provide a dedicated and consistent study of the leading twist (twist-2) matching of the TMD operators.
The quark and gluon components of the generic TMD operators are $$\begin{aligned}
\label{def:TMD_OP_Q}
\Phi_{ij}(x,{\bm{b}})&=\int \frac{d\lambda}{2\pi}e^{-ixp^+\lambda}\bar q_i{\left(}\lambda n+{\bm{b}}{\right)}\mathcal{W}(\lambda,{\bm{b}})q_j{\left(}0{\right)},
\\ \label{def:TMD_OP_G}
\Phi_{\mu\nu}(x,{\bm{b}})&=\frac{1}{xp^+}\int \frac{d\lambda}{2\pi}e^{-ixp^+\lambda}
F_{+\mu}{\left(}\lambda n+{\bm{b}}{\right)}\mathcal{W}(\lambda,{\bm{b}})F_{+\nu}{\left(}0{\right)},\end{aligned}$$ where $n$ is the lightlike vector and we use the standard notation for the lightcone components of vector $v^\mu=n^\mu v^-+\bar n^\mu v^++g_T^{\mu\nu}v_\nu$ (with $n^2=\bar n^2=0$, $n\cdot\bar n=1$, and $g_T^{\mu\nu}=g_{\mu\nu}-n^\mu \bar n^\nu-\bar n^\mu n^\nu$). The operator $\mathcal{W}$ is $$\begin{aligned}
\mathcal{W}(\lambda,{\bm{b}})=\tilde W_{n}^T{\left(}\lambda n+{\bm{b}}{\right)}\sum_X |X\rangle\langle X|\tilde W_{n}^{T\dagger}{\left(}0{\right)},\end{aligned}$$ with Wilson lines $W$ taken in the appropriate representation of gauge group . The staple contour of the gauge link results in the rapidity divergences, the unique feature of TMD operators. The rapidity divergences are removed by the proper rapidity renormalization factor $R$, which is built from the TMD soft factor, $$\begin{aligned}
\label{eq:softf}
S({\bm{b}}) &=
\frac{{{\text{Tr}}}_\text{color}}{N_c}
\langle 0|{\left[}S_n^{T\dagger} \tilde S_{\bar n}^T {\right]}({\bm{b}})
{\left[}\tilde S^{T\dagger}_{\bar n} S_n^T{\right]}(0)|0\rangle,\end{aligned}$$ where $S_n$ and ${\tilde S}_{\bar n}$ stand for soft Wilson lines along $n$ and $\bar n$ (for the precise definition of $W^T$ and $\tilde S^T$ see e.g.[@Echevarria:2016scs]). The structure of factor $R$ follows from the TMD factorization theorem [@GarciaEchevarria:2011rb; @Collins:2011zzd; @Echevarria:2014rua; @Echevarria:2016scs; @Chiu:2012ir] and depends on the rapidity regularization scheme. However, the expressions for rapidity-divergence-free quantities, such as evolution kernels and matching coefficients are independent on the scheme. In the following we use the $\delta$-regularization scheme formulated in [@Echevarria:2015byo; @Echevarria:2016scs]. This scheme uses the infinitesimal parameter $\delta$ as a regulator for rapidity divergences in combination with the usual dimensional regularization (with $d=4-2\epsilon$, $\epsilon>0$) for ultraviolet and collinear divergences. Such combination appears to be very visual and practically convenient. The central statement of the TMD factorization theorem is the complete elimination of rapidity divergences by the rapidity renormalization factor $R$. In the $\delta$-regularization scheme where $R=1/\sqrt{S({\bm{b}})}$, the rapidity divergences take the form of $\ln\delta$ and do not mix with divergences in $\epsilon$, which yield the exact cancellation of $\ln\delta$ at finite $\epsilon$. This non-trivial demand is necessary for a consistent higher-then-NLO evaluation and requires the matching of regularizations for different field modes (see [@Echevarria:2016scs]). It also results into the correspondence between TMD processes and the jet production [@Vladimirov:2016dll].
The hadron matrix elements of the TMD operators with open vector and spinor indices (\[def:TMD\_OP\_Q\],\[def:TMD\_OP\_G\]) are to be decomposed over all possible Lorentz variants, which define TMD parton distribution functions (TMDPDFs). In the literature, this decomposition has been made in the momentum space (for spin-1/2 hadrons it can be found in [@Goeke:2005hb; @Bacchetta:2008xw] (for quark operators) and in [@Mulders:2000sh] (for gluon operators)). However, it is convenient to consider TMD distributions in the impact parameter space, where it is naturally defined. The correspondence between decomposition in momentum and impact parameter spaces can be found in e.g.[@Boer:2011xd; @Echevarria:2015uaa]. In this work we need only a part of the complete decomposition, $$\begin{aligned}
\label{TMD_Q_dec}
&\Phi_{q{\leftarrow}h,ij}(x,{\bm{b}})=\langle h|\Phi_{ij}(x,{\bm{b}})|h\rangle =\frac{1}{2}\Big(
f_1\gamma^-_{ij} +g_{1L} S_L(\gamma_5\gamma^-)_{ij}
\\{\nonumber}&+(S_T^\mu
i\gamma_5\sigma^{+\mu})_{ij}h_1+(i\gamma_5\sigma^{+\mu})_{ij}{\left(}\frac{g_T^{\mu\nu}}{2}+\frac{ b^\mu b^\nu}{ {\bm{b}}^2}{\right)}\frac{S_{T}^\nu }{2}h_{1T}^{\perp}+...\Big),
\\\label{TMD_G_dec}
&\Phi_{g{\leftarrow}h,\mu\nu}(x,{\bm{b}})=\langle h|\Phi_{\mu\nu}(x,{\bm{b}})|h\rangle
\\{\nonumber}&
= \frac{1}{2}\Big(-g_T^{\mu\nu}f_1^g -
i\epsilon_T^{\mu\nu}S_Lg_{1L}^g +2h_1^{\perp g} {\left(}\frac{g_T^{\mu\nu}}{2}+\frac{ b^\mu b^\nu}{{\bm{b}}^2}{\right)}+...\Big),\end{aligned}$$ where the vector $b^\mu$ is a 4-dimensional vector of the impact parameter ($b^+=b^-=0$ and $ -b^2\equiv {\bm{b}}^2>0$), and $S_{T,L}$ are components of the hadron spin vector defined in Eq.(\[def:spin\]). On the r.h.s. of Eqs. (\[TMD\_Q\_dec\],\[TMD\_G\_dec\]) and in the rest of the letter we omit arguments of TMD distributions $(x,{\bm{b}})$, unless they are necessary. Note that in Eq. (\[TMD\_Q\_dec\]) we use the normalization for the distribution $h_{1T}^\perp$ different from the traditional one [@Goeke:2005hb]. The traditional definition can be recovered substituting $h_{1T}^{\perp}\rightarrow h_{1T}^{\perp} {\bm{b}}^2 M^2$, with $M$ being the mass of hadron. In the section \[sec:trans\_and\_pretz\] we argue that such normalization is natural.
In Eqs.(\[TMD\_Q\_dec\],\[TMD\_G\_dec\]) we write only the TMD distributions that match the twist-2 integrated distributions. The dots include the TMD distributions that match the twist-3 and higher parton distribution functions (PDF). The reported distributions are usually addressed as helicity ($g_{1L}$ and $g_{1L}^g$), transversity ($h_1$), pretzelosity ($h_{1T}^\perp$) and linearly polarized gluon ($h_1^{\perp g}$) distributions. The small-$b$ matching of these distributions has been performed separately for quarks [@Bacchetta:2013pqa] and gluons [@Echevarria:2015uaa] in different renormalization schemes. Furthermore, the pretzelosity distribution has been overlooked by these groups. In this letter, we present a uniform and consistent NLO matching of these TMD distributions.
Small-$b$ operator product expansion {#sec:OPE}
====================================
The small-$b$ operator product expansion (OPE) is the relation between TMD operators and lightcone operators. Its leading order can be written as $$\begin{aligned}
\label{OPE_Q}
&\Phi_{ij}(x,{\bm{b}})=
\\{\nonumber}&\quad
\Big[{\left(}C_{q{\leftarrow}q}({\bm{b}}){\right)}_{ij}^{ab}\otimes \phi_{ab}\Big](x)+\Big[{\left(}C_{q{\leftarrow}g}({\bm{b}}){\right)}_{ij}^{\alpha\beta}\otimes
\phi_{\alpha\beta}\Big](x)+...,
\\\label{OPE_G}
&\Phi_{\mu\nu}(x,{\bm{b}})=
\\{\nonumber}&\quad\Big[{\left(}C_{g{\leftarrow}q}({\bm{b}}){\right)}_{\mu\nu}^{ab}\otimes \phi_{ab}\Big](x)+\Big[{\left(}C_{g{\leftarrow}g}({\bm{b}}){\right)}_{\mu\nu}^{\alpha\beta}\otimes
\phi_{\alpha\beta}\Big](x)+...,\end{aligned}$$ where symbol $\otimes$ denotes the Mellin convolution in the variable $x$. The functions $C({\bm{b}})$ are dimensionless, i.e. they depend on ${\bm{b}}$ only logarithmically. The dots represent the power suppressed contributions, which presently have been studied only for the unpolarized case (see discussion in [@Scimemi:2016ffw]). At this order of OPE, the functions $\phi(x)$ are the formal limit of the TMD operators $\Phi(x,{\bm{0}})$. The hadronic matrix elements of $\phi$ are the PDFs $$\begin{aligned}
\label{PDF_Q_dec}
\phi_{q{\leftarrow}h,ij}(x)&=&\langle h|\phi_{ij}(x)|h\rangle \\{\nonumber}&=&\frac{1}{2}\Big(
f_q(x)\gamma^-_{ij} +\Delta f_q(x) S_L(\gamma_5\gamma^-)_{ij}
\\{\nonumber}&&+(S_T^\mu
i\gamma_5\sigma^{+\mu})_{ij}\delta f_q(x)\Big)+\mathcal{O}{\left(}\frac{M}{p^+}{\right)},
\\\label{PDF_G_dec}
\phi_{g{\leftarrow}h,\mu\nu}(x)&=&\langle h|\phi_{\mu\nu}(x)|h\rangle
\\{\nonumber}&=& \frac{1}{2}{\left(}-g_T^{\mu\nu}f_g(x) -
i\epsilon_T^{\mu\nu}S_L\Delta f_g{\right)}+\mathcal{O}{\left(}\frac{M}{p^+}{\right)},\end{aligned}$$ where $M$ is the mass of hadron, $S_{L}$ and $S_T$ are the components of the hadron spin vector $$\begin{aligned}
\label{def:spin}
S^\mu=S_L{\left(}\frac{p^+}{M}\bar n^\mu-\frac{M}{2p^+}n^\mu{\right)}+S_T^\mu ,\end{aligned}$$ and $\epsilon_T^{\mu\nu}=\epsilon^{+-\mu\nu}=n_\alpha \bar n_\beta \epsilon^{\alpha\beta\mu\nu}$. For future convenience we introduce the universal notation $$\begin{aligned}
\Phi_q^{[\Gamma]}=\frac{{\text{Tr}}(\Gamma \Phi)}{2},\qquad \Phi_g^{[\Gamma]}=\Gamma^{\mu\nu}\Phi_{\mu\nu}.\end{aligned}$$ Both sides of Eq. (\[OPE\_Q\],\[OPE\_G\]) should be supplemented by the ultraviolet renormalization constants. Additionally, the TMD operator on the l.h.s. is to be multiplied by the rapidity renormalization factor $R$. The renormalized TMD operator has the form $$\begin{aligned}
\label{TMD_ren}
\Phi^{\text{ren}}\!(x,{\bm{b}};\mu,\zeta)
=Z(\mu,\zeta|\epsilon)R({\bm{b}},\mu,\zeta|\epsilon,\delta)\Phi(x,{\bm{b}}|\epsilon,\delta),\end{aligned}$$ where we explicitly show the dependence on regularization parameters on the r.h.s.. The dependence on $\epsilon$ and $\delta$ cancels in the product. The renormalization factors are independent on the Lorentz structure but dependent on parton flavor. The explicit expressions for these factors up to NNLO can be found in [@Echevarria:2016scs; @Echevarria:2015byo].
The cancellation of rapidity divergences for the spin-dependent distributions is a non-trivial statement. Let us consider the small-$b$ OPE for a generic TMD quark operator. At one loop we find $$\begin{aligned}
\label{gen_OPE}
&\Phi_q^{[\Gamma]}=\Gamma^{ab}\phi_{ab}+a_sC_F\pmb B^{\epsilon}\Gamma(-\epsilon)\Big[
\\{\nonumber}&
-(\gamma^+\gamma^-\Gamma+\Gamma\gamma^-\gamma^+)^{ab}
+\bar x{\left(}\frac{g_T^{\alpha\beta}}{2}-\frac{ b^\alpha b^\beta}{4\pmb B}\epsilon{\right)}(\gamma^\mu \gamma_\alpha \Gamma \gamma_\beta \gamma_\mu)^{ab}
\\&{\nonumber}+{\left(}\frac{1}{(1-x)_+}-\ln{\left(}\frac{\delta}{p^+}{\right)}{\right)}{\left(}\gamma^+\gamma^-\Gamma+\Gamma\gamma^-\gamma^+ +
\frac{i\epsilon \gamma^+\!\!\!\not b\Gamma}{2\pmb B}+\frac{i\epsilon \Gamma\!\!\!\not b\gamma^+}{2\pmb B}{\right)}^{ab}
\\&{\nonumber}-\frac{i\pi}{2}{\left(}\gamma^+\gamma^-\Gamma-\Gamma\gamma^-\gamma^+ +
\frac{i\epsilon \gamma^+\!\!\!\not b\Gamma}{2\pmb B}-\frac{i\epsilon \Gamma\!\!\!\not b\gamma^+}{2\pmb B}{\right)}^{ab}
\Big]\otimes \phi_{ab}+\mathcal{O}(a_s^2),\end{aligned}$$ where $\pmb B= {\bm{b}}^2/4>0$, $a_s=g^2/(4\pi)^{d/2}$, and we use the standard PDF notation, $[f(x)]_+=f(x)-\delta(\bar x)\int dy f(y)$ and $\bar x=1-x$. In this expression, we omit the gluon operator contribution for simplicity. The complex term in the last line of Eq. (\[gen\_OPE\]) is the artifact of $\delta$-regularization. The logarithm of $\delta$ represents the rapidity divergence which is to be eliminated by the factor $R$ which at this perturbative order reads $$\begin{aligned}
R&=1+
2a_sC_F\pmb B^\epsilon\Gamma(-\epsilon)
{\nonumber}\\
&\times \Big(
\mathbf{L}_{\sqrt{\zeta}}+2\ln{\left(}\frac{\delta}{p^+}{\right)}-\psi(-\epsilon)-\gamma_E\Big)+\mathcal{O}(a_s^2),\end{aligned}$$ where $\mathbf{L}_X=\ln{\left(}\pmb B X^2 e^{2\gamma_E}{\right)}$. The rapidity divergence cancels in the product $R\Phi$ if and only if $$\begin{aligned}
\label{gG=0}
\gamma^+\Gamma=\Gamma \gamma^+=0\ ,\end{aligned}$$ yielding $$\begin{aligned}
\label{eq:OPE_qq}
&&R\Phi_q^{[\Gamma]}=\Gamma^{ab}\phi_{ab}+a_sC_F\pmb B^{\epsilon}\Gamma(-\epsilon)
\\{\nonumber}&&\times \Big[
{\left(}-4+\frac{4}{(1-x)_+}+2\delta(\bar x)(\mathbf{L}_{\sqrt{\zeta}}-\psi(-\epsilon)-\gamma_E){\right)}\Gamma^{ab}
\\{\nonumber}&&
+\bar x{\left(}\frac{g_T^{\alpha\beta}}{2}-\frac{ b^\alpha b^\beta}{4\pmb B}\epsilon{\right)}(\gamma^\mu \gamma_\alpha \Gamma \gamma_\beta \gamma_\mu)^{ab}
\Big]\otimes \phi_{ab}+\mathcal{O}(a_s^2)\ .\end{aligned}$$ The cancellation of rapidity divergences is the fundamental pre-requisite to obtain the matching coefficients of the renormalized operator $\Phi$ and $\phi$.
The conditions analogue to Eq. (\[gG=0\]) for the gluon operator are $$\begin{aligned}
\label{Lambda+=0}
\Gamma^{+\mu}=\Gamma^{-\mu}=\Gamma^{\mu+}=\Gamma^{\mu-}=0.\end{aligned}$$ They follow from OPE for a generic gluon TMD operator $\Phi^{\mu\nu}$ similar to Eq. (\[gen\_OPE\]), which we do not present here, since it is rather lengthy and not instructive. The conditions in Eq. (\[gG=0\],\[Lambda+=0\]) are satisfied only for the following Lorentz structures $$\begin{aligned}
\Gamma^q=\{\gamma^+,\gamma^+\gamma^5,\sigma^{+\mu}\},&&
\Gamma^g=\{g_T^{\mu\nu},\epsilon_T^{\mu\nu}, b^\mu b^\nu/{\bm{b}}^2\},\end{aligned}$$ which exactly correspond to the Lorentz structures for the so called leading dynamical twist TMD distributions. In this way, the relations Eq. (\[gG=0\],\[Lambda+=0\]) provide a definition of the leading dynamical twist for TMD operators that can be used with no reference to a particular cross-section. On the other hand, our consideration shows that TMD operators of non-leading dynamical twist have rapidity singularities that are not canceled by the soft factor in Eq. (\[eq:softf\]). While we have no knowledge of a calculation of the correction to the leading order of TMD factorization, our finding demonstrates that it has a different structure of rapidity divergences (which can spoil the factorization). The relation in Eq. (\[gG=0\]) will be used in the next section to fix the definition of $\gamma_5$ in the dimensional regularization.
In order to calculate the matching coefficients, we consider the quark and gluon matrix elements with the momentum of parton set to $p^\mu=p^+\bar n^\mu$. This choice of kinematic is allowed for consideration of twist-2 contribution only (which is the case of this article). Then, the calculations are greatly simplified. In particular, the perturbative corrections to the parton matrix element of $\phi$’s are zero, due to the absence of a scale in the dimensional regularization. Therefore, such matrix elements are equal to their renormalization constant, i.e. has not finite in $\epsilon$-terms. In practice, it implies that the matching coefficient is the $\epsilon$-finite part of the parton matrix element of the renormalized TMD operator (\[TMD\_ren\]). The evaluation of OPE for a general Lorentz structure (as in Eq. (\[eq:OPE\_qq\])) is not very representative because one needs only the components associated with the TMDPDFs. Therefore, we project out the required components and present the expressions for each particular distribution.
Helicity distribution {#sec:helicity}
=====================
In the case of helicity distributions the Lorentz structures for quark and gluon operators are $$\begin{aligned}
\Gamma=\gamma^+\gamma^5 ,\quad \Gamma^{\mu\nu}=i\epsilon_T^{\mu\nu}.\end{aligned}$$ The corresponding orthogonal projectors are $$\begin{aligned}
\label{hel:proj}
\overline{\Gamma}=\mathcal{N}_{\text{sch.}}\frac{\gamma^-\gamma^5}{2} ,\quad \overline{\Gamma}^{\mu\nu}=i\mathcal{N}_{\text{sch.}}\frac{\epsilon_{T}^{\mu\nu}}{2},\end{aligned}$$ where the factor $\mathcal{N}_{\text{sch.}}$ depends on the definition of $\gamma_5$ matrix in dimensional regularization. Historically the most popular schemes (for QCD calculations) are ’t Hooft-Veltman-Breitenlohner-Maison (HVBM) [@tHooft:1972tcz; @Breitenlohner:1977hr], and Larin scheme [@Larin:1993tq; @Larin:1991tj]. In both schemes the combination $\gamma^+\gamma^5$ can be presented as $$\begin{aligned}
\gamma^+\gamma^5=\frac{i}{3!}\epsilon^{+\nu\alpha\beta}\gamma_{\nu}\gamma_\alpha\gamma_\beta,\end{aligned}$$ where $\epsilon^{\mu\nu\alpha\beta}$ is the antisymmetric Levi-Civita tensor. The difference between schemes is hidden in the definition of Levi-Civita tensor. In HVBM the $\epsilon^{\mu\nu\alpha\beta}$ is defined only for 4-dimensional set of indices. I.e. $\epsilon^{\mu\nu\alpha\beta}=1$ if $\{\mu\nu\alpha\beta\}$ is even permutation of $\{0,1,2,3\}$, $\epsilon^{\mu\nu\alpha\beta}=-1$ if the permutation is odd, and $\epsilon^{\mu\nu\alpha\beta}=0$ for any another case. In Larin scheme the $\epsilon$-tensor is non-zero for all set of $d$-dimensional indices. The value of individual components are undefined, however, the product of two $\epsilon$-tensors is defined, $\epsilon^{\mu_1\nu_1\alpha_1\beta_1}\epsilon^{\mu_2\nu_2\alpha_2\beta_2}=-g^{\mu_1\mu_2}g^{\nu_1\nu_2}g^{\alpha_1\alpha_2}g^{\beta_1\beta_2}+g^{\mu_1\nu_2}g^{\nu_1\mu_2}g^{\alpha_1\alpha_2}g^{\beta_1\beta_2}-..$, where the dots mean all $4!$ permutations of indices with alternating signs.
The drawback of both schemes is the violation of Adler-Bardeen theorem for the non-renormalization of the axial anomaly. This must be fixed by an extra finite renormalization constant $Z_{qq}^5$, derived from an external condition, see detailed discussion in [@Larin:1993tq; @Matiounine:1998re; @Ravindran:2003gi]. The NNLO calculation of polarized deep-inelastic-scattering and Drell-Yan process in refs.[@Matiounine:1998re; @Ravindran:2003gi] made in (HVBM) have shown that the finite renormalization is required only for the the quark-to-quark part (both singlet and non-singlet cases). The same finite renormalization constant can be used for Larin scheme up to $\epsilon$-singular terms at NNLO [@Moch:2014sna]. However, it seems that for higher order terms (in $\epsilon$ or in the coupling constant) the constant should be modified [@Moch:2014sna].
Needless to say, that Larin scheme is far more convenient then HVBM, because it does not violate Lorentz invariance. However, Larin scheme, as it is originally formulated and used in the modern applications [@Moch:2014sna], is inapplicable for TMD calculations. The point is that it does violate the definition of the leading dynamical twist Eq. (\[gG=0\]). Indeed, in the Larin scheme we have $$\begin{aligned}
\gamma^+\Gamma=\gamma^+{\left(}\gamma^+\gamma^5{\right)}_{\text{Larin}}=\frac{i}{3!}\epsilon^{+\nu\alpha\beta}\gamma^+\gamma_{\nu}\gamma_{\alpha}\gamma_{\beta}\neq 0,\end{aligned}$$ because there is a contribution when all indices $\{\nu\alpha\beta\}$ are transverse. Note, that in HVBM scheme there is not such problem, since in the 4-dimensional $\epsilon^{+\nu\alpha\beta}$, one of the indices is necessarily “$-$”. To ensure the existence of Eq. (\[gG=0\]) we perform a light modification of Larin scheme, and call it *Larin$^+$* scheme. We define $$\begin{aligned}
\label{Larin+}
(\gamma^+\gamma^5)_{\text{Larin}^+}=\frac{i\epsilon^{+-\alpha\beta}}{2!}\gamma^+\gamma_\alpha\gamma_\beta=\frac{i\epsilon_T^{\alpha\beta}}{2!}\gamma^+\gamma_\alpha\gamma_\beta.\end{aligned}$$ The $\epsilon_T$-tensor is $d$-dimensional, and for calculations it should be supplemented by the relation $$\begin{aligned}
\epsilon_T^{\alpha_1\beta_1}\epsilon_T^{\alpha_2\beta_2}=-g_T^{\alpha_1\alpha_2}g_T^{\beta_1\beta_2}+g_T^{\alpha_1\beta_2}g_T^{\beta_1\alpha_2}.\end{aligned}$$ In the case the $\epsilon$-tensor is 4-dimensional, the definition Eq. (\[Larin+\]) coincides with HVBM. The normalization factors presented in the Eq. (\[hel:proj\]) are $$\begin{aligned}
\mathcal{N}_{\text{sch.}}&=&\left\{\begin{array}{cc}
{\displaystyle}{1}& \text{HVBM},
\\
{\displaystyle}{(1-\epsilon)^{-1}(1-2\epsilon)^{-1}} & \text{Larin}^+.
\end{array}\right.\end{aligned}$$
The NLO calculation is straightforward and parallel to unpolarized calculation, which is presented in details in [@Echevarria:2016scs]. We write the matching onto integrated distribution as $$\begin{aligned}
R\Phi_{q}^{[\gamma^+\gamma_5]}&=\Delta C_{q\leftarrow q}\otimes \phi_{q}^{[\gamma^+\gamma_5]}+
\Delta C_{q\leftarrow g}\otimes \phi_{g}^{[\epsilon_T^{}]}{\nonumber}\\
R\Phi_{g}^{[\epsilon_T^{}]}&=\Delta C_{g\leftarrow q}\otimes \phi_{q}^{[\gamma^+\gamma_5]}+
\Delta C_{g\leftarrow g}\otimes \phi_{g}^{[\epsilon_T^{}]}\\end{aligned}$$ where $$\begin{aligned}
{\nonumber}\Delta C_{q{\leftarrow}q}&=\delta(\bar x)+a_s C_F
\Bigg\{2\pmb B^{\epsilon}\Gamma(-\epsilon)\Big[\frac{2}{(1-x)_+}-2
\\&{\nonumber}+\bar x(1+\epsilon)\mathcal{H}_{\text{sch.}}+\delta(\bar x){\left(}\mathbf{L}_{\sqrt{\zeta}}-\psi(-\epsilon)-\gamma_E{\right)}\Big]\Bigg\}_{\epsilon\text{-finite}},
\\
{\nonumber}\Delta C_{q{\leftarrow}g}&=a_s C_F
\Bigg\{2\pmb B^{\epsilon}\Gamma(-\epsilon)\Big[x-\bar x \mathcal{H}_{\text{sch.}}\Big]\Bigg\}_{\epsilon\text{-finite}},\end{aligned}$$ $$\begin{aligned}
{\nonumber}\Delta C_{g{\leftarrow}q}&=a_s C_F
\Bigg\{2\pmb B^{\epsilon}\Gamma(-\epsilon)\Big[1+\bar x \mathcal{H}_{\text{sch.}}\Big]\Bigg\}_{\epsilon\text{-finite}},
\\\label{hel:gg}
\Delta C_{g{\leftarrow}g}&=\delta(\bar x)+a_s C_A
\Bigg\{2\pmb B^{\epsilon}\Gamma(-\epsilon)\frac{1}{x}\Big[\frac{2}{(1-x)_+}-2
\\{\nonumber}&
-2x^2+2x\bar x\mathcal{H}_{\text{sch.}}+
\delta(\bar x){\left(}\mathbf{L}_{\sqrt{\zeta}}-\psi(-\epsilon)-\gamma_E{\right)}\Big]\Bigg\}_{\epsilon\text{-finite}},\end{aligned}$$ where the subscript “$\epsilon$-finite” implies the removal of $\epsilon$-singular terms, as discussed in the end of sec.\[sec:OPE\]. The coefficient $\mathcal{H}_{\text{sch.}}$ accumulates the difference between schemes, $$\begin{aligned}
\mathcal{H}_{\text{sch.}}&=&\left\{\begin{array}{cc}
{\displaystyle}{1+2\epsilon}& \text{HVBM},
\\
{\displaystyle}{\frac{1+\epsilon}{1-\epsilon}} & \text{Larin}^+.
\end{array}\right.\end{aligned}$$ One can see that the expressions within HVBM and Larin$^+$ schemes coincide up to $\epsilon$-suppressed parts.
In the regime of large-$q_T$, the TMD factorization reproduces the collinear factorization. Therefore, it is natural to normalize the helicity TMDPDF such that at large-$q_T$ it reproduces the cross-section for polarized Drell-Yan, which in turn is normalized onto cross-section of unpolarized Drell-Yan process [@Ravindran:2003gi]. The TMD equivalent of this statement is the requirement of equality between helicity and unpolarized matching coefficients $$\begin{aligned}
\Big[Z_{qq}^5({\bm{b}})\otimes \Delta C_{q{\leftarrow}q}({\bm{b}})\Big](x)=C_{q{\leftarrow}q}(x,{\bm{b}}).\end{aligned}$$ The constant $Z_{qq}^5$ is universal, in the sense that it is independent on the rapidity regularization scheme. We find the following finite renormalization constant for the TMD matching $$\begin{aligned}
\label{Zqq5}
Z_{qq}^5=\delta(\bar x )+2 a_sC_F\pmb B^\epsilon \Gamma(-\epsilon){\left(}1-\epsilon-(1+\epsilon)\mathcal{H}_{\text{sch.}}{\right)}\bar x.\end{aligned}$$ Note, that HVBM version of $Z_{qq}^5$ coincides with the NLO part of the one presented in [@Ravindran:2003gi] up to logarithmic terms (which are dependent on the kinematics of process) .
Concluding the section we present the expressions for the helicity TMD distribution in the regime of small-$b$ $$\begin{aligned}
{\nonumber}g_{1L}(x,{\bm{b}})&=[\Delta C_{q{\leftarrow}q}({\bm{b}})\otimes \Delta f_q](x)+[\Delta C_{q{\leftarrow}g}({\bm{b}})\otimes \Delta f_g](x)+\mathcal{O}({\bm{b}}^2),
\\
g^g_{1L}(x,{\bm{b}})&=[\Delta C_{g{\leftarrow}q}({\bm{b}})\otimes \Delta f_q](x)+[\Delta C_{g{\leftarrow}g}({\bm{b}})\otimes \Delta f_g](x)+\mathcal{O}({\bm{b}}^2),\end{aligned}$$ where the matching coefficients are taken in the limit $\epsilon\rightarrow 0$, $$\begin{aligned}
\Delta C_{q{\leftarrow}q}&\equiv C_{q{\leftarrow}q}=\delta(\bar x)+a_s C_F\Big(
-2\mathbf{L}_{\mu}\Delta p_{qq}+2\bar x{\nonumber}\\&{\nonumber}+\delta(\bar x){\left(}-\mathbf{L}_{\mu}^2+2\mathbf{L}_{\mu}\mathbf{l}_\zeta-\zeta_2{\right)}\Big)+\mathcal{O}(a_s^2),
\\
{\nonumber}\Delta C_{q{\leftarrow}g}&=a_s T_F{\left(}-2\mathbf{L}_{\mu} \Delta p_{qg}+4\bar x{\right)}+\mathcal{O}(a_s^2),
\\
{\nonumber}\Delta C_{g{\leftarrow}q}&=a_s C_F{\left(}-2\mathbf{L}_{\mu} \Delta p_{gq}-4\bar x{\right)}+\mathcal{O}(a_s^2),
\\
{\nonumber}\Delta C_{g {\leftarrow}g }&=\delta(\bar x)+a_s C_A\Big(
-2\mathbf{L}_{\mu}\Delta p_{gg}-8\bar x
\\
& +\delta(\bar x){\left(}-\mathbf{L}_{\mu}^2+2\mathbf{L}_{\mu}\mathbf{l}_\zeta-\zeta_2{\right)}\Big)+\mathcal{O}(a_s^2),\end{aligned}$$ with $\mathbf{l}_\zeta=\ln\mu^2/\zeta$. The functions $\Delta p$ are the combination of helicity evolution kernel (which can be found e.g. in [@Moch:2014sna]) and the TMD anomalous dimension. They are $$\begin{aligned}
\label{app:hel_qq}
\Delta p_{qq}(x)&=\frac{2}{(1-x)_+}-1-x
,{\nonumber}\\
\Delta p_{qg}(x)&=2x-1,{\nonumber}\quad
\Delta p_{gq}(x)=2-x,
\\ \Delta p_{gg}(x)&=\frac{2}{(1-x)_+}+2-4x.\end{aligned}$$ The coefficients $\Delta C_{q{\leftarrow}q}$ and $\Delta C_{q{\leftarrow}g}$ have been evaluated in [@Bacchetta:2013pqa]. Our expressions agree with ones presented in [@Bacchetta:2013pqa] apart of $\zeta_2$ term in $\Delta C_{q{\leftarrow}q}$. This disagreement is the result of different renormalization schemes. We use the conventional $\overline{\text{MS}}$ scheme with $e^{\epsilon\gamma_E}$ factor, while $\text{MS}$-scheme of [@Bacchetta:2013pqa] is defined with $\Gamma^{-1}(1+\epsilon)$ factor. The coefficients $\Delta C_{g{\leftarrow}q}$ and $\Delta C_{g{\leftarrow}g}$ have been evaluated in [@Echevarria:2015uaa]. Our expressions agree with expressions presented in the erratum of Ref.[@Echevarria:2015uaa].
Transversity and pretzelosity distributions {#sec:trans_and_pretz}
===========================================
The spinor structure for the transversity TMD operator is usually addressed as $\Gamma=i\gamma_5\sigma^{+\mu}=\epsilon_T^{\mu\nu} \sigma^{+\nu}/2,$ where to obtain the last equality we used that index $\mu$ is transverse. This definition is scheme dependent just as the helicity case. However, since there is no mixture with the gluons at the leading twist, the common practice is to eliminate the $\gamma^5$ or $\epsilon_T$ from the definition of operator. Thus we consider $\Gamma=\sigma^{+\mu}$. The small-$b$ expansion takes the form $$\begin{aligned}
\label{trans->trans}
R\Phi_{q}^{[\sigma^{+\mu}]}&=\Big\{\delta(\bar x)g_T^{\mu\nu}+2a_s C_F\pmb B^\epsilon\Gamma(-\epsilon)\Big[
\\&{\nonumber}g_T^{\mu\nu}{\left(}\frac{2}{(1-x)_+}-2+\delta(\bar x)(\mathbf{L}_{\sqrt{\zeta}}-\psi(-\epsilon)-\gamma_E){\right)}\\&{\nonumber}\qquad\qquad\qquad\qquad-
2\epsilon^2\bar x \frac{ b^\mu b^\nu}{{\bm{b}}^2}\Big]\Big\}\otimes \phi^{[\sigma^{+\nu}]}_{q}.\end{aligned}$$ Comparing this expression with the parameterization Eq. (\[TMD\_Q\_dec\]) we observe that both the transversity distribution and the pretzelosity distributions have the leading twist-2 matching on the integrated transversity PDF.
The transversity and pretzelosity distribution matching coefficients, respectively $ \delta C_{q{\leftarrow}q}$ and $\delta^\perp C_{q{\leftarrow}q}$, are defined as $$\begin{aligned}
\label{trans_OPE}
R\Phi_{q}^{[\sigma^{+\mu}]}&= g^{\mu\nu}_T \delta C_{q{\leftarrow}q}\otimes \phi_{q}^{[\sigma^{+\nu}]}
\\{\nonumber}&
+{\left(}\frac{b^\mu b^\nu}{{\bm{b}}^2}+\frac{g_T^{\mu\nu}}{2(1-\epsilon)}{\right)}\delta^\perp C_{q{\leftarrow}q}\otimes \phi_{q}^{[\sigma^{+\nu}]},\end{aligned}$$ where the factor $(1-\epsilon)$ in the pretzelosity vector structure is necessary to support its tracelessness in dimensional regularization.
Comparing expressions (\[trans\_OPE\]) with (\[trans->trans\]) we obtain $$\begin{aligned}
\delta C_{q{\leftarrow}q}&=\delta(\bar x)+a_s C_F
\Bigg\{2\pmb B^{\epsilon}\Gamma(-\epsilon)\Big[\frac{2}{(1-x)_+}-2
\\{\nonumber}&+\bar x \frac{\epsilon^2}{1-\epsilon}+\delta(\bar x){\left(}\mathbf{L}_{\sqrt{\zeta}}-\psi(-\epsilon)-\gamma_E{\right)}\Big]\Bigg\}_{\epsilon\text{-finite}}.\end{aligned}$$ It results to the following small-$b$ expression for the transversity TMD PDF $$\begin{aligned}
h_{1}(x,{\bm{b}})&=\Big[\delta C_{q{\leftarrow}q}({\bm{b}})\otimes \delta f_q\Big](x)+\mathcal{O}({\bm{b}}^2),\end{aligned}$$ with the matching coefficient $$\begin{aligned}
\delta C_{q{\leftarrow}q}&=\delta(\bar x)+a_s C_F\Big(
-2\mathbf{L}_{\mu}\delta p_{qq}
\\{\nonumber}&
+\delta(\bar x){\left(}-\mathbf{L}_{\mu}^2+2\mathbf{L}_{\mu}\mathbf{l}_\zeta-\zeta_2{\right)}\Big)+\mathcal{O}(a_s^2).\end{aligned}$$ The $\delta p_{qq}$ is the combination of the transversity evolution kernel (see e.g.[@Vogelsang:1997ak; @Mikhailov:2008my]) and TMD anomalous dimension. It is $$\begin{aligned}
\label{app:trans}
\delta p_{qq}(x)=\frac{2}{(1-x)_+}-2.\end{aligned}$$ This expression coincides with the one calculated in [@Bacchetta:2013pqa] up to $\zeta_2$ term (which is absent in [@Bacchetta:2013pqa] due to the usage of a different form of $\overline{\text{MS}}$-scheme).
The matching coefficient of the pretzelosity distribution at finite $\epsilon$ is $$\begin{aligned}
\delta^\perp C_{q{\leftarrow}q}=-4a_s C_F\pmb B^\epsilon \Gamma(-\epsilon)\bar x \epsilon^2.\end{aligned}$$ Here, we can appreciate the consistent and natural counting of the normalization of the pretzelosity provided by Eq. (\[TMD\_Q\_dec\]). We also observe that at this order of perturbation theory the matching coefficient is proportional to $\epsilon$, i.e. zero. Nonetheless, the $\epsilon$-suppressed part will reveal at NNLO, and provide a non-zero contribution. Therefore, we conclude $$\begin{aligned}
\label{pretz}{\nonumber}h_{1T}^\perp(x,{\bm{b}})&=\Big[\delta^\perp C_{q{\leftarrow}q}({\bm{b}})\otimes \delta f_q\Big](x)+\mathcal{O}({\bm{b}}^2)=
\\
& \Big[ {\left(}0+\mathcal{O}(a_s^2){\right)}\otimes \delta f_q\Big](x)+\mathcal{O}({\bm{b}}^2).\end{aligned}$$ This result coincides with the estimation made in [@Bacchetta:2008xw]. According to Eq. (\[pretz\]), the pretzelosity distribution is suppressed numerically. This observation is indeed supported by the measurements of $\sin(3\phi_h-\phi_S)$-asymmetries by HERMES and COMPASS, see e.g.[@Lefky:2014eia] and references within. We also mention that it is not possible to obtain the small-$b$ matching at the helicity distribution. The helicity distribution as a part of pretzelosity distribution is suggested by various model calculations (see [@Avakian:2008dz] and references within).
Linearly polarized gluon
========================
The linearly polarized gluon distribution at small-$b$ matches the unpolarized gluon distribution. The matching of the gluon TMD operator to the unpolarized distribution has the form $$\begin{aligned}
R\Phi_{g}^{\mu\nu}={\left(}\frac{b^\mu b^\nu}{{\bm{b}}^2}+\frac{g_T^{\mu\nu}}{2(1-\epsilon)}{\right)}\Big(
&\delta^L C_{g{\leftarrow}g}\otimes \phi_g^{[g_T]}
\\{\nonumber}&+\delta^L C_{g{\leftarrow}q}\otimes \phi_q^{[\gamma^+]}\Big)+...~,\end{aligned}$$ where dots represent terms proportional to $g_T^{\mu\nu}$ and $\epsilon_T^{\mu\nu}$, i.e. the parts which contribute to the matching of unpolarized and helicity distributions.
The coefficients $\delta^L C$ are[^1] $$\begin{aligned}
\label{eq:dc1}
\delta^L C_{g{\leftarrow}g}&=&\Big(+4a_s C_A \pmb B^\epsilon \Gamma(-\epsilon)\frac{\bar x}{x}\epsilon\Big)_{\epsilon\text{-finite}}.
{\nonumber}\\
\delta^LC_{g{\leftarrow}q}&=&\Big(+4a_s C_F \pmb B^\epsilon \Gamma(-\epsilon)\frac{\bar x}{x}\epsilon\Big)_{\epsilon\text{-finite}}.\end{aligned}$$ Note that there is not rapidity nor renormalization group evolution, which appears at the next perturbative order.
Finally, we obtain following small-$b$ expression for the linearly polarized gluon TMDPDF $$\begin{aligned}
h_1^{\perp g}(x,{\bm{b}}) &=
\\ {\nonumber}&
[\delta^LC_{g{\leftarrow}q}({\bm{b}})\otimes f_q](x)+[\delta^L C_{g{\leftarrow}g}({\bm{b}})\otimes f_g](x)+\mathcal{O}({\bm{b}}^2),\end{aligned}$$ where $$\begin{aligned}
\label{eq:dc2}
\delta^LC_{g{\leftarrow}g}&=&-4a_s C_A \frac{\bar x}{x}+\mathcal{O}(a_s^2)
,
\\
\label{eq:dc3}
\delta^LC_{g{\leftarrow}q}&=&-4a_s C_F \frac{\bar x}{x}+\mathcal{O}(a_s^2).\end{aligned}$$
Conclusions {#sec:Conclusions}
===========
In this letter, we have provided complete discussion on the matching of transverse momentum dependent (TMD) distributions to the twist-2 integrated distributions in the regime of small-$b$ (or equivalently, large-$q_T$). To perform the matching we have evaluated the operator product expansion (OPE) of a generic TMD operator near the light-cone.
As a practical outcome, we derive the complete set of NLO TMD matching coefficients of the twist-2 parton distributions evaluated uniformly at finite $\epsilon$. The TMD distributions that have non-zero matching are helicity ($g_{1L}$, $g_{1L}^g$), transvesity ($h_1$), pretzelosity ($h_{1T}^\perp$) and linearly polarized gluon ($h_1^{\perp g}$) distributions (we do not include the unpolarized TMD distribution in the consideration because it has been considered in many articles. The evaluation performed using the same regularization as this paper can be found in [@Echevarria:2016scs]). The most part of the coefficient functions have been evaluated separately for quarks and gluons by different groups [@Bacchetta:2013pqa; @Echevarria:2015uaa]. We agree with their evaluations (taking into account that in ref.[@Bacchetta:2013pqa], different renormalization scheme has been used).
The evaluation of OPE for a generic TMD operator reveals the condition which should be satisfied in order match the rapidity divergences of a TMD operator and the leading order TMD soft factor Eq. (\[eq:softf\]). The conditions presented in Eqs. (\[gG=0\],\[Lambda+=0\]) restrict the Lorentz structure of the TMD operators. The TMD distributions whose operator meet these conditions, are known as TMD distributions of leading dynamical twist. In this way, we demonstrate that the next-to-leading-dynamical-twist contributions to the TMD factorization theorem (i.e. the power suppressed contributions to the TMD cross-section) necessarily have a different structure of rapidity divergences.
We also provide discussion on the schemes of $\gamma_5$ and $\epsilon_T$-definition in the dimensional regularization, which has been skipped by the previous authors. We have shown that the definition of $\gamma_5$ suggested by the popular Larin scheme [@Larin:1993tq] does not support the condition of the leading dynamical twist, and thus, it is inapplicable in TMD calculations. We suggest an updated version of Larin scheme (Larin$^+$ scheme Eq. (\[Larin+\])), which supports the condition and has simpler properties than the traditional one. Our calculation has been performed in Larin$^+$ and HVBM [@tHooft:1972tcz; @Breitenlohner:1977hr] schemes. At NLO the difference between schemes arises only in the $\epsilon$-suppressed terms. We argue about the normalization of the distributions and derive the finite renormalization constant (\[Zqq5\]) for TMD helicity distributions in both schemes.
The evaluation of the matching has been performed at finite-$\epsilon$. The $\epsilon$-suppressed terms, although do not contribute directly to NLO, contribute to higher perturbative orders (see e.g. discussion in [@Echevarria:2016scs]). The pretzelosity distribution (considered here for the first time) has $\epsilon$-suppressed matching coefficient, which indicates that it has non-zero matching coefficient to transversity distribution at NNLO Eq. (\[pretz\]). This offers a natural explanation of the smallness of this distribution in phenomenological analyses [@Lefky:2014eia]. The complete $\epsilon$-dependent expressions and the general analyses performed in this work open the path to the NNLO evaluation of polarized TMD distributions.
Acknowledgements {#acknowledgements .unnumbered}
================
We acknowledge some very profitable discussions with V. Braun, M. G. A. Buffing, M. Diehl, M. G. Echevarria, T. Kasemets as well as, S.Moch for discussions on the definition of $\gamma_5$. D.G.R. and I.S. are supported by the Spanish MECD grant FPA2014-53375-C2-2-P and FPA2016-75654-C2-2-P.
[^1]: We thank M. Diehl for pointing out a sign typo in eq. (\[eq:dc1\], \[eq:dc2\], \[eq:dc3\]) in the previous version of the paper.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We report new observations of Zn II and Cr II absorption lines in 10 damped [Lyman $\alpha$]{} systems (DLAs), mostly at redshift $z_{\rm abs} \simgt 2.5$. By combining these results with those from our earlier survey (Pettini et al. 1994) and other recent data, we construct a sample of 34 measurements (or upper limits) of the Zn abundance relative to hydrogen \[Zn/H\]; the sample includes more than one third of the total number of DLAs known.
The plot of the abundance of Zn as a function of redshift reinforces the two main findings of our previous study. (1) Damped [Lyman $\alpha$]{} systems are mostly metal-poor, [*at all redshifts sampled*]{}; the column density weighted mean for the whole data set is \[Zn/H\] $= -1.13 \pm 0.38$ (on a logarithmic scale), or approximately 1/13 of solar. (2) There is a large spread, by up to two orders of magnitude, in the metallicities we measure at essentially the same redshifts. We propose that damped [Lyman $\alpha$]{} systems are drawn from a varied population of galaxies of different morphological types and at different stages of chemical evolution, supporting the idea of a protracted epoch of galaxy formation.
At redshifts $z \simgt 2$ the typical metallicity of the damped [Lyman $\alpha$]{} systems is in agreement with expectations based on the consumption of H I gas implied by the recent measurements of $\Omega_{\rm DLA}$ by Storrie-Lombardi et al. (1996a), and with the metal ejection rates in the universe at these epochs deduced by Madau (1996) from the ultraviolet luminosities of high redshift galaxies revealed by deep imaging surveys. There are indications in our data for an increase in the mean metallicity of the damped [Lyman $\alpha$]{} systems from $z > 3$ to $\approx 2$, consistent with the rise in the comoving star formation rate indicated by the relative numbers of $U$ and $B$ drop-outs in the [*Hubble Deep Field*]{}. Although such comparisons are still tentative, it appears that these different avenues for exploring the early evolution of galaxies give a broadly consistent picture.
At redshifts $z < 1.5$ DLAs evidently do not exhibit the higher abundances expected from a simple closed-box model of global chemical evolution, although the number of measurements is still very small. We speculate that this may be due to an increasing contribution of low surface brightness galaxies to the cross-section for damped [Lyman $\alpha$]{} absorption and to the increasing dust bias with decreasing redshift proposed by Fall and collaborators. However, more DLAs at intermediate redshifts need to be identified before the importance of these effects can be assessed quantitatively.
The present sample is sufficiently large for a first attempt at constructing the metallicity distribution of damped [Lyman $\alpha$]{} systems and comparing it with those of different stellar populations of the Milky Way. The DLA abundance histogram is both broader and peaks at lower metallicities that those of either thin or thick disk stars. At the time when our Galaxy’s metal enrichment was at levels typical of DLAs, its kinematics were closer to those of the halo and bulge than a rotationally supported disk. This finding is at odds with the proposal that most DLAs are large disks with rotation velocities in excess of 200 [km s$^{-1}$]{}, based on the asymmetric profiles of absorption lines recorded at high spectral resolution. Observations of the familiar optical emission lines from H II regions, which are within reach of near-infrared spectrographs on 8-10 m telescopes, may help resolve this discrepancy.\
author:
- Max Pettini
- 'Linda J. Smith'
- 'David L. King'
- 'Richard W. Hunstead'
title: |
THE METALLICITY OF HIGH REDSHIFT GALAXIES:\
THE ABUNDANCE OF ZINC IN 34 DAMPED LYMAN $\alpha$ SYSTEMS FROM $z$ = 0.7 TO 3.4
---
5.0ex
INTRODUCTION
============
In the last twelve months there has been a dramatic increase in our ability to identify normal galaxies at $z \simeq 3$, study their stellar populations, and measure the rates of star formation and metal production in the universe over most of the Hubble time (Steidel et al. 1996; Madau et al. 1996). The most prominent features in the spectra of field galaxies at high-redshift (as is the case in the ultraviolet spectra of nearby star-forming galaxies) are strong interstellar lines which are similar, both qualitatively (in the range of ionization stages seen) and quantitatively (in the strengths of the absorption), to those in damped [Lyman $\alpha$]{} systems; this similarity is consistent with the view that this class of QSO absorbers traces the material available for star formation at $z \simgt 2$ (e.g. Wolfe 1995). The connection between normal galaxies and damped [Lyman $\alpha$]{} systems (DLAs) is a particularly important one to make and clarifying several aspects of this connection remains a priority. The reason is simple: QSOs with known DLAs are typically more than 5 magnitudes brighter than a $L^{\ast}$ galaxy at the same redshift. Consequently, we will inevitably continue to rely mostly on QSO absorption line spectroscopy for the study of physical conditions in the early stages of galaxy formation.
Since 1990 (Pettini, Boksenberg, & Hunstead 1990) we have been conducting a survey of metallicity and dust in DLAs taking advantage of the diagnostic value of weak transitions of Zn II and Cr II. As explained in that paper (see also the critical reappraisal of the technique in Pettini et al. 1997), \[Zn/H\][^1] is a straightforward measure of the degree of metal enrichment analogous to the stellar \[Fe/H\], while \[Cr/Zn\] reflects the extent to which grain constituents are removed from the gas phase and thereby gives an indication of the dust-to-metals ratio. The major results of the survey were reported in Pettini et al. (1994). From the analysis of Zn and Cr abundances in 17 DLAs, mostly at $z \simeq
2$, we concluded that the typical metallicity of the universe at a look-back time of $\sim 13$ Gyr ($H_0 = 50$ [km s$^{-1}$]{} Mpc$^{-1}$; $q_0 = 0.01$) was $Z_{\rm DLA} = 1/10 Z_{\sun}$. We further found that there is a considerable range—by up to two orders of magnitude—in the degree of metal enrichment reached by different damped [Lyman $\alpha$]{} galaxies at essentially the same epoch, and that even at these early stages of galaxy formation dust appears to be an important component of the interstellar medium, leading to the selective depletion of refractory elements from the gas.
A natural next step is to extend the Zn and Cr abundance measurements over a wider range of redshifts than that considered by Pettini et al. (1994) with the ultimate aim of identifying the emergence of heavy elements and dust in galaxies and following their build-up with time. To this end we have continued our survey since 1994; the full sample now consists of 34 DLAs, more than one third of the total number known (Wolfe et al. 1995). In this paper we present the new data and consider the conclusions that can drawn from the whole set of measurements of \[Zn/H\]; preliminary reports have appeared in conference proceedings (e.g. Pettini et al. 1995a; Smith et al. 1996). Our findings on the abundance of dust from consideration of the \[Cr/Zn\] ratio in the same sample have been reported separately (Pettini et al. 1997). Recently, Lu et al. (1996) have addressed similar questions from measurements of \[Fe/H\] in 20 DLAs using high-resolution echelle spectra acquired with the Keck telescope. These authors reach conclusions which are in agreement with those presented here regarding the emergence of heavy elements at high redshifts, although the analysis of \[Fe/H\] is complicated by the fact that this ratio, unlike \[Zn/H\], depends on [*both*]{} the metallicity and dust content of the interstellar medium.
Before proceeding it is useful to point out that in cases where DLAs from the present sample have been reobserved with HIRES on Keck (Wolfe et al. 1994; Wolfe 1995; Prochaska & Wolfe 1997a), \[Zn/H\] has been found to be in good agreement with the values measured in our survey, which is based on 4-m telescope data (see §2 below). While the exceptional quality of the Keck observations has made possible several new aspects of this work, including the study of the relative abundances of a wide range of elements and the analysis of the kinematics of the absorbing gas, the basic survey of metallicity in DLAs can be carried out satisfactorily with 4-m class telescopes. The main reason for this is the optically thin nature of the Zn II and Cr II lines in most DLAs proposed by Pettini et al. (1990) and confirmed by subsequent Keck spectra.\
OBSERVATIONS AND DATA REDUCTION
===============================
The new data reported in this paper consist of observations of 10 DLAs in 9 QSOs obtained between March 1994 and February 1996 (an additional candidate DLA from the low dispersion survey by Storrie-Lombardi et al. (1996b)—at [$z_{\rm abs}$]{}=3.259 in the [$z_{\rm em}$]{}=4.147 BAL QSO 1144$-$073—was shown not to be a damped system by our higher resolution observations of the [Lyman $\alpha$]{} absorption line). In Table 1 we have collected relevant information for the 10 DLAs; the references listed in column (4) are the papers where the damped nature of the absorber was first identified. The absorption redshifts measured from associated metal lines in our blue and red spectra are listed in column (5); with 6 new DLAs at $z_{\rm abs} > 2.5$, we have tripled the number of absorbers in this redshift regime compared with our earlier sample.
The observations, reduction of the spectra and derivation of Zn and Cr abundances followed the procedures described in Pettini et al. (1994) and the interested reader is referred to that paper for a detailed treatment. Briefly, the observations were carried out mostly with the double-beam cassegrain spectrograph of the William Herschel telescope on La Palma, Canary Islands; additional red spectra were secured with the cassegrain spectrograph of the Anglo-Australian telescope at Siding Spring Observatory, Australia. At $z_{\rm abs} > 2.5$ the Zn II $\lambda\lambda 2025.483, 2062.005$ and Cr II $\lambda\lambda 2055.596, 2061.575, 2065.501$ multiplets are redshifted longwards of 7175 Å, where the quantum efficiency of CCDs falls with increasing wavelength. Using EEV and Tektronix CCDs we generally found it necessary to integrate for longer than $\sim 20\,000$ s (column 8 of Table 1) in order to achieve S/N between 9 and 46 (column 9). With a spectral resolution of 0.75–1.1 Å FWHM (column 7), the corresponding $3\sigma$ detection limits for the rest frame equivalent widths of unresolved Zn II and Cr II absorption lines range from $W_0$($3\sigma$) = 66 to 14 mÅ (column 10). The final “depth” of the survey—that is the lowest metallicity that can be measured—depends on the combination of $W_0$($3\sigma$) and the neutral hydrogen column density $N$(H$^0$). Since the values of $N$(H$^0$) in the new DLAs observed span one order of magnitude (see §3 below), it is the sight-lines with the largest column densities of gas which provide the most stringent limits on metal abundances. Accordingly, we have tended to select DLAs for the present survey primarily on the basis of the value of $N$(H$^0$).
In Figure 1 we have reproduced portions of the QSO spectra encompassing the regions where the Zn II and Cr II lines are expected in the 10 DLAs in Table 1. As can be seen from the figure, the absorption lines sought are detected in approximately half of the cases. Table 2 lists redshifts and rest-frame equivalent widths for the detections; in the other cases the $3\sigma$ limits given in column (10) of Table 1 apply.
With the double-beam spectrograph on the WHT we were able to record portions of the blue spectrum of each QSO, centred on the damped [Lyman $\alpha$]{} line, simultaneously with the red arm observations aimed at the Zn II and Cr II lines. The blue detector was either the Image Photon Counting System or a thinned Tektronix CCD; exposure times were the same as those given in column (8) of Table 1. A 600 grooves/mm grating was used to record a 800 Å wide interval of the spectrum with a resolution of $\sim 1.5$ Å FWHM. This configuration was chosen in preference to the higher resolving power achievable with a 1200 grooves/mm grating because a good definition of the QSO continuum is a key factor in determining the accuracy with which $N$(H$^0$) can be deduced from the profile of the damping wings of the [Lyman $\alpha$]{} absorption line. Normalised portions of the blue spectra are shown in Figure 2 together with our fits to the damped [Lyman $\alpha$]{} lines. The theoretical damping profiles are centred at the redshifts of the O I $\lambda 1302.1685$ lines which are encompassed by our blue data.\
ZINC AND CHROMIUM ABUNDANCES
============================
The main results of our survey are collected in Table 3 which includes the 10 DLAs in Table 1 and 7 additional systems for which data have been published since our earlier study (Pettini et al. 1994). Values of the neutral hydrogen column density $N$(H$^0$) are listed in column (3) of Table 4; the typical accuracy of these measurements, including the uncertainty in the placement of the continuum, is $\pm 20$%. $N$(H$^0$) is likely to account for most of the neutral gas in each DLA given the low molecular fractions which apply to these absorbers at high redshifts (Levshakov et al. 1992; Ge & Bechtold et al. 1997; Ćirković et al. 1997).
Columns (3) and (6) of Table 3 give the column densities of Zn$^+$ and Cr$^+$ respectively, deduced from the measured equivalent widths (or upper limits) assuming no line saturation. That this is generally the case is indicated by: (1) the weakness of the absorption lines; (2) the equivalent width ratios of lines within each multiplet which, when measurable, are usually close to the ratios of the corresponding $f$-values (Bergeson & Lawler 1993); and (3) the resolved absorption profiles recorded with HIRES on Keck for many DLA systems, including some in common with the present survey (Lu et al. 1996; Prochaska & Wolfe 1997a). There are of course exceptions, such as the $z_{\rm abs} = 2.5842$ system in Q1209$+$093—see the discussion at §3.12 below. The important point, however, is that it is usually possible with the signal-to-noise ratio and resolution of our data to assess the degree of saturation of the Zn II and Cr II lines.
Column (4) lists the ratios $N$(Zn$^+$)/$N$(H$^0$) derived by dividing the entries in column (3) by those in column (3) of Table 4; comparison with the solar abundance of Zn, log (Zn/H)$_{\sun} = -7.35$ (Anders & Grevesse 1989), then leads to [*underabundances*]{} of Zn by the factors given in column (5). The corresponding values for Cr (log (Cr/H)$_{\sun} = -6.32$) are given in column (8) and column (9) lists the ratio $N$(Cr$^+$)/$N$(Zn$^+$) in cases where it could be determined.
In taking the ratios $N$(Zn$^+$)/$N$(H$^0$) and $N$(Cr$^+$)/$N$(H$^0$) as measures of (Zn/H) and (Cr/H), we implicitly assume that there is little contribution to the observed Zn II and Cr II absorption from ionised gas (which would not produce [Lyman $\alpha$]{} absorption). This is likely to be the case given the large column densities of H I, and indeed there are no indications to the contrary in our data. In particular, we found no significant differences in redshift between the Zn II and Cr II lines, when detected, and O I $\lambda 1302.1685$ which arises only in H I regions. Should this assumption be shown to be incorrect, however, the values of \[Zn/H\] and \[Cr/H\] deduced here would then be upper limits to the true abundances.
We now comment briefly on each DLA in Table 3.\
Q0000$-$263; $z_{\rm abs} = 3.3901$
-----------------------------------
Our observations of this DLA, the highest redshift absorber in the survey, have been described in Pettini et al. (1995a). While Zn II $\lambda 2025.483$ remains undetected, despite the sensitive limit reached in a total exposure time of 58200 s, we do record weak Cr II absorptions at the $4\sigma$ ($\lambda 2055.596$) and $3\sigma$ ($\lambda 2061.575$) significance levels. Cr II $\lambda 2055.596$ is expected to be stronger than Zn II $\lambda 2025.483$ if the fraction of Cr locked up in dust grains is less than about 50%. With $N$(H$^0$) = ($2.5\pm0.5$)$ \times 10^{21}$ cm$^{-2}$ (Savaglio, D’Odorico, & Moller 1994), this is one of the highest column density systems in our sample. We conclude that the abundance of Zn is less than 1/80 of the solar value; this estimate is $\sim 5$ times more sensitive than the previous limit (Savaglio et al. 1994). The abundance of Cr, \[Cr/H\] $\simeq -2.2 \pm 0.1$, is similar to those of other elements measured by Molaro et al. (1996) and Lu et al. (1996), making this DLA one of the most metal-poor in our sample.\
Q0056$+$014; $z_{\rm abs} = 2.7771$
-----------------------------------
This QSO is from the Large Bright Quasar Survey by Chaffee et al. (1991). We deduce $N$(H$^0$) = ($1.3\pm0.2$)$ \times 10^{21}$ cm$^{-2}$ from fitting the core of the damped [Lyman $\alpha$]{} line, in reasonable agreement with the value log $N$(H I) = 21.0 reported by Wolfe et al. (1995).
As can be seen from Figure 1, the Zn II and Cr II absorption lines are broad and shallow in this DLA, spanning $\approx 200$ km s$^{-1}$. The stronger member of the Zn II doublet, $\lambda 2025.483$, falls within the atmospheric A band. Plotting the four absorption lines labelled in Figure 1 on the same velocity scale suggests that most of feature “1” is [*not*]{} due to Zn II$~\lambda 2025.483$, but rather to poorly corrected telluric absorption. From the equivalent widths of Cr II $\lambda 2055.596$ and $\lambda 2065.501$ (features 2 and 4 in Figure 1), which are consistent with the optically thin ratio of 2:1, we deduce a weighted mean $N$(Cr$^+$) = ($2.8\pm0.4$)$ \times 10^{13}$ cm$^{-2}$. This column density of Cr$^+$ produces an equivalent width $W_0 = $($82 \pm 12$) mÅ for Cr II $\lambda 2061.575$; since we measure $W_0 = $($117 \pm 16$) mÅfor feature 3, which is a blend of Cr II$\lambda 2061.575$ and Zn II$\lambda 2062.005$, we conclude that $W_0 = $($35 \pm 20$) mÅ for the latter. This in turn corresponds to $N$(Zn$^+$) = ($3.5\pm2$)$ \times 10^{12}$ cm$^{-2}$. Thus both Zn and Cr appear to be $\approx 20$ times less abundant than in the Sun.
Our red spectrum also shows several Fe II lines from an absorption system at $z_{\rm abs} = 2.3044$, including: Fe II $\lambda 2344.214$ (visible in Figure 1 at $\lambda_{\rm obs} = 7748.46$Å) with $W_0 = $($470 \pm
12$) mÅ; Fe II $\lambda 2367.5905$ with $W_0 = $($64 \pm 6$) mÅ; Fe II $\lambda 2374.4612$ with $W_0 = $($220 \pm 14$) mÅ; and Fe II $\lambda 2382.765$ with $W_0 = $($640 \pm 12$) mÅ.\
Q0201$+$365; $z_{\rm abs} = 2.462$
----------------------------------
Keck observations of this DLA have been published recently by Prochaska & Wolfe (1996) who deduced relatively high abundances of Zn and Cr, respectively $\sim 1/2$ and $\sim 1/8$ of solar. Evidently, even at redshifts as high as 2.5 some galaxies had already undergone significant chemical evolution and enriched their interstellar media in heavy elements to levels comparable with that of the Milky Way today.\
Q0302$-$223; $z_{\rm abs} = 1.0093$
-----------------------------------
Lanzetta, Wolfe, & Turnshek (1995) proposed this as a candidate DLA system on the basis of low-resolution [*IUE*]{} data; a subsequent UV spectrum secured with the Faint Object Spectrograph on the [*Hubble Space Telescope*]{} confirmed that $N$(H$^0$) = ($2.15 \pm 0.35$)$ \times 10^{20}$ cm$^{-2}$ (Pettini & Bowen 1997). Recent WHT observations of Zn II and Cr II lines by Pettini & Bowen (1997) have shown the abundances to be 1/3 and 1/8 of solar respectively. After subtraction of the QSO radial profile from [*HST*]{} WFPC2 images of the field, Le Brun et al. (1997) identified two galaxies which may be producing the absorption; at $z = 1.009$ they would have luminosities $L \approx 0.2 L^{\ast}$ and $\approx L^{\ast}$ and distances of 12 and $27~h_{50}^{-1}$ kpc respectively from the QSO sight-line.\
Q0454$+$039; $z_{\rm abs} = 0.8596$
-----------------------------------
The abundances of Zn and Cr reported by Steidel et al. (1995a) correspond to \[Zn/H\] = $-0.83 \pm 0.08$ and \[Cr/H\] = $-1.01 \pm 0.05$ if the experimentally measured $f$-values of the Zn II and Cr II multiplets (Bergeson & Lawler 1993) are adopted for consistency with the rest of the present study. Deep images of the QSO field both from the ground (Steidel et al. 1995a) and with [*HST*]{} (Le Brun et al. 1997) suggest that the absorber is a compact galaxy with $L \approx 0.25 L^{\ast}$ ($q_0 = 0.05$) at a projected distance of $8~h_{50}^{-1}$ kpc from the QSO.\
Q0836$+$113; $z_{\rm abs} = 2.4651$
-----------------------------------
This is the faintest QSO in our survey (Hunstead, Pettini, & Fletcher 1990) and the S/N of the red spectrum remains modest despite the considerable investment in exposure time (Table 1). Combined with the relatively low H I column density of ($3.8\pm0.4$)$ \times 10^{20}$ cm$^{-2}$, the $3 \sigma$ upper limits to the Zn II and Cr II lines place limits on the abundances of Zn and Cr which are less stringent than in most other DLAs considered: \[Zn/H\] $\leq -0.8$ and \[Cr/H\] $\leq -1.2$.
The blue spectrum shown in Figure 2 was recorded with the IPCS on the WHT in March 1994. Note that, of all the damped [Lyman $\alpha$]{} lines reproduced in Figure 2, this is the only instance where there appears to be weak emission in the core of the absorption line. The line flux, ($2 \pm 0.7$)$ \times 10^{-17}$ erg s$^{-1}$ cm$^{-2}$, agrees within the errors with the value of ($2.9 \pm 0.7$)$ \times 10^{-17}$ erg s$^{-1}$ cm$^{-2}$ reported by Hunstead et al. (1990) from independent data obtained in April 1987 with a different IPCS detector on the AAT. The two sets of observations were obtained with the same slit width (1.2 arcsec) and at the same position angle on the sky (150 degrees).\
Q0841$+$129; $z_{\rm abs} = 2.3745, 2.4764$
-------------------------------------------
The spectrum of this bright ($V \simeq 17$), high redshift ($z \simeq 2.5$, estimated from the onset of the [Lyman $\alpha$]{} forest) BL Lac object discovered by C. Hazard (private communication) shows [*two*]{} DLAs (see Figure 2), making it a highly suitable target for follow-up high resolution observations.
As can be seen from Figure 1, in the lower redshift system we detect features 2 and 3; the strength of the latter indicates a significant contribution from Zn II $\lambda 2062.005$ to the blend. Following a procedure similar to that described for Q0056$+$014 at §3.1 above, we deduce $N$(Cr$^+$) = ($9.5\pm2$)$ \times 10^{12}$ cm$^{-2}$ from the equivalent widths of Cr II $\lambda 2055.596$ and $\lambda 2065.501$. This in turn leads us to estimate that approximately half of the equivalent width of feature 3 is due to Zn II $\lambda 2062.005$ with $W_0 = $($24 \pm 9$) mÅ. Together with the $3 \sigma$ upper limit $W_0$(2025)$ \leq 26$ mÅ for the stronger member of the doublet, this then implies $N$(Zn$^+$) = ($1.8 \pm 0.5$)$ \times 10^{12}$ cm$^{-2}$.
Thus we find that Zn and Cr at $z_{\rm abs} = 2.3745$ are underabundant by factors of 23 and 45 respectively, relative to solar values. Similar, or lower, abundances apply to the $z_{\rm abs} = 2.4764$ DLA, given the lack of detectable Zn II and Cr II lines (see Table 3).\
Q0913$+$072; $z_{\rm abs} = 2.6183$
-----------------------------------
The signal-to-noise ratios of our spectra of this bright QSO are among the highest in the survey—see Table 1 and Figures 1 and 2. The column density of neutral hydrogen is however comparatively low, $N$(H$^0$) = ($2.3 \pm 0.4$)$ \times 10^{20}$ cm$^{-2}$. The lack of Zn II and Cr II absorption even at S/N = 46 implies underabundances by factors of more than 14 and 32 respectively.\
Q0935$+$417; $z_{\rm abs} = 1.3726$
-----------------------------------
Lanzetta et al. (1995) estimated $N$(H$^0$)$ \simeq 2 \times 10^{20}$ cm$^{-2}$ for this candidate DLA from low resolution [*IUE*]{} data; a subsequent [*HST*]{} FOS spectrum confirmed that $N$(H$^0$) = ($2.5 \pm 0.5$)$ \times 10^{20}$ cm$^{-2}$ (Lanzetta & Meyer 1996, private communication). With this value of the hydrogen column density, the observations by Meyer, Lanzetta, & Wolfe (1995) imply \[Zn/H\] = $-0.80$ and \[Cr/H\] = $-0.90$.\
Q1104$-$180; $z_{\rm abs} = 1.6616$
-----------------------------------
Smette et al. (1995) identified this DLA in the spectrum of the brighter ($B = 16.7$) component of this gravitationally lensed QSO pair. From AAT observations obtained with an instrumental setup similar to that used in our survey, these authors estimated $N$(H$^0$)$ = 6 \times 10^{20}$ cm$^{-2}$. They also reported detections of Zn II and Cr II absorption lines with equivalent widths $W_0$(2025.483)$ = {\rm (}75 \pm 20$) mÅand $W_0$(2055.596)$ = {\rm (}57 \pm 20$) mÅ respectively. If the lines are unsaturated \[Zn/H\] = $-0.80$ and \[Cr/H\] = $-1.30$.\
Q1151$+$068; $z_{\rm abs} = 1.7736$
-----------------------------------
Even though the damped [Lyman $\alpha$]{} line falls in the crowded near-UV spectrum of this $z_{\rm em} = 2.762$ QSO (see Figure 2), our estimate $N$(H$^0$) = ($2.0\pm 0.5$)$ \times 10^{21}$ cm$^{-2}$ is in very good agreement with log $N$(H I)$ = 21.3$ published by Turnshek et al. (1989). The ratios of equivalent widths within the Zn II and Cr II multiplets strongly suggest that the lines are optically thin; Zn and Cr are both underabundant by a factor $\approx 40$.
Our red spectrum, which covers the region $5500-5900$ Å, shows three C IV $\lambda\lambda1548,1550$ doublets at $z_{\rm abs} = 2.5629$, 2.7069 and 2.7551 respectively.\
Q1209$+$093; $z_{\rm abs} = 2.5843$
-----------------------------------
This is another high column density DLA; we measure $N$(H$^0$) = ($2.0\pm 0.5$)$ \times 10^{21}$ cm$^{-2}$ which compares well with log $N$(H I)$ = 21.4$ reported by Lu et al. (1993). The Zn II and Cr II lines are the strongest encountered in the entire survey of 34 DLAs (see Table 2). Fitting the absorption profiles requires $b = 50$ km s$^{-1}$ (as usual, $b = \sqrt{2}\sigma$ where $\sigma$ is the one-dimensional velocity dispersion along the line of sight), indicating that several velocity components most likely contribute to the absorption. Similarly, Lu et al. found $b = 122$ km s$^{-1}$ from fitting a single curve-of-growth to the strongest UV absorption lines. Some of the components may well be saturated; so we quote our best estimates of $N$(Zn$^+$) and $N$(Cr$^+$) as lower limits. We conclude that Zn is [*more*]{} abundant than 1/9 solar and Cr more abundant than 1/27 solar. Higher resolution observations are required to establish how close to these limits the true values are.\
Q1328$+$307; $z_{\rm abs} = 0.6922$
-----------------------------------
We have included here the measurements of Zn and Cr abundances in the spectrum of 3C 286 reported by Meyer & York (1992), after appropriate rescaling with the $f$-values of Bergeson & Lawler (1993). Although the measurement was discussed in Pettini et al. (1994), this intermediate redshift DLA was not explicitly included in that sample which consisted exclusively of DLAs at $z_{\rm abs} > 1.7$. CCD images of the QSO field obtained with ground-based telescopes (Steidel et al. 1994) and with [*HST*]{} (Le Brun et al. 1997) show a large ($\approx 10-30~h_{50}^{-1}$ kpc), low surface brightness galaxy which has been proposed as the absorber.\
Q1337$+$113; $z_{\rm abs} = 2.7957$
-----------------------------------
Our measured $N$(H$^0$) of ($8 \pm 2$)$ \times 10^{20}$ cm$^{-2}$ agrees very well with log $N$(H I)$ = 20.9$ reported by Turnshek et al. (1989). When we observed this QSO, in March 1994, we found it to be considerably fainter than the magnitude $V \simeq 18.2$ estimated by Hazard et al. (1986) from POSS plates. Although the S/N achieved is the lowest in the survey (see Table 1 and Figure 1), it is still sufficient to establish that the abundances of Zn and Cr are less than 1/10 and 1/23 of solar respectively.\
Q1946$+$769; $z_{\rm abs} = 2.8443$
-----------------------------------
This $z_{\rm em} = 3.051$ QSO, intrinsically one of the most luminous known, is sufficiently bright to have been studied extensively at echelle resolutions and high S/N with 4-m telescopes (Fan & Tytler 1994; Lu et al. 1995; Tripp, Lu, & Savage 1996). However, the hydrogen column density in the $z_{\rm abs} = 2.8443$ DLA is relatively low, $N$(H$^0$) = ($2 \pm 0.5$)$ \times 10^{20}$ cm$^{-2}$ (Lu et al.). Consequently, the upper limits \[Zn/H\] $\leq -0.82$ and \[Cr/H\] $\leq -1.00$ deduced by these authors are rather uninformative given that the true metallicity is $\sim30$ times lower (\[Fe/H\]$ = -2.44 \pm 0.13$).\
Q2239$-$386; $z_{\rm abs} = 3.2810$
-----------------------------------
This QSO is faint and the absorber is at high redshift; the combination of these two factors resulted in the longest integration time in the survey (see Table 1). Adopting $N$(H$^0$)$ = 5.8 \times 10^{20}$ cm$^{-2}$ measured by Lu & Wolfe (1994), we deduce Zn and Cr underabundances by factors of more than 11 and 13 respectively.
The Cr measurement is based on the weakest member of the triplet, Cr II$~\lambda 2065.501$; $~\lambda 2061.575$ is affected by a strong sky emission line and $~\lambda 2055.596$, which at $z_{\rm abs} = 3.2810$ is redshifted to $\lambda_{\rm obs} = 8802.82$ Å, falls very close to Mn II $\lambda 2606.462$ at $z_{\rm abs} = 2.3777$, the redshift of a second DLA along this line of sight (Lu & Wolfe 1994). Based on the strengths of the other two members of the Mn II triplet, $\lambda 2576.877$ at $\lambda_{\rm obs} = 8703.55$ Åand $\lambda 2594.499$ at $\lambda_{\rm obs} = 8763.97$Å, the feature labelled 2 in the last panel of Figure 1 is mostly Mn II$~\lambda 2606.462$. The two strong absorption lines also visible in this figure are Fe II $~\lambda\lambda 2586.6500,2600.1729$ at $z_{\rm abs} = 2.3777$.\
DISCUSSION
==========
Our total sample, which consists of measurements (or upper limits) of \[Zn/H\] in 34 DLAs over the redshift range $z_{\rm abs} = 0.6922 - 3.3901$, is constructed by combining data for the 17 DLAs in Table 3 with those for the 15 DLAs in Table 3 of Pettini et al. (1994) and with the further addition of two DLAs in Q0528$-$250 (Meyer et al. 1989) which were included in the sample considered by Pettini et al. (1994) but not listed in their Table 3. All the points in Figure 3 are based on the $f$-values of the Zn II doublet by Bergeson & Lawler (1993) and the meteoritic solar abundance of Zn from the compilation by Anders & Grevesse (1989)[^2].
We now consider what implications can be drawn from this extensive survey on the chemical evolution of the neutral content of the universe and on the relationship of damped [Lyman $\alpha$]{} systems to present-day spiral galaxies.\
Chemical Evolution of Damped [Lyman $\alpha$]{} Systems
-------------------------------------------------------
Figure 3 shows the abundance of Zn as a function of redshift. The enlarged sample confirms the two main conclusions reached by Pettini et al. (1994):
\(1) Damped [Lyman $\alpha$]{} systems, [*at all redshifts probed*]{}, are generally metal-poor and presumably arise in galaxies at early stages of chemical evolution.
\(2) There appears to be a large range in the values of metallicity reached by different galaxies at the same redshift, pointing to a protracted ‘epoch of galaxy formation’ and to the fact that chemical enrichment probably proceeded at different rates in different DLA galaxies.
While we find gas with near-solar metallicities at redshifts as high as $z \simeq 2.5$, there are also examples of galaxies with abundances less than 1/10 solar at a time when the disk of the Milky Way differed little from its present-day composition. At redshifts $z \simeq 2 - 2.5$ the full range of metal abundances spans about two orders of magnitude. Although for metallicities $Z_{\rm DLA} \simlt 1/50 Z_{\sun}$ the Zn II lines become vanishingly small and only upper limits to the abundance of Zn can be deduced, we do know from echelle spectroscopy of more abundant astrophysical elements that values of $Z_{\rm DLA} \simlt 1/100 Z_{\sun}$ are not uncommon at $z_{\rm abs} = 2 - 3$ (see Figure 1 of Pettini et al. 1995a).
These two results are considered quantitatively in Table 5 where in the last column we list, for various subsets of our sample, the column density-weighted mean abundance of Zn $${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]} =
{\rm log}\langle{\rm(Zn/H)}_{\rm DLA}\rangle-{\rm log~(Zn/H)}_{\odot},
\label{}$$ where $$\langle{\rm (Zn/H)}_{\rm DLA}\rangle =
\frac{\sum\limits_{i=1}^{n} N{\rm(Zn}^+{\rm)}_i}
{\sum\limits_{i=1}^{n} N{\rm(H}^0{\rm)}_i} ,
\label{}$$
and $\sigma^{\prime}$, the standard deviation from the column density weighted mean, defined as
$$(n - 1)~ \sigma^{\prime 2} =
\sum\limits_{i=1}^{n}
\lgroup {\rm[Zn/H]}_i -
{\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]}\rgroup ^2
\label{}$$
The summations in equations (2) and (3) are over the $n$ DLA systems considered in each subsample.
Under the working assumption that DLAs account for most of the material available for star formation at high redshift, the quantity ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]}$ is a measure of the degree of metal enrichment reached by the universe at a given epoch. This is a general statement which follows from the column density distribution of [Lyman $\alpha$]{} systems (Lanzetta et al. 1995) and which holds irrespectively of the precise nature of the damped absorbers (disks, spheroids, gas clouds yet to collapse into galaxies, etc.), provided that there are no significant biases in the samples of DLAs from which our observations are drawn (Fall 1996).
The values of ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]}$ in Table 5 are strictly upper limits (with the exception of subsample number 1), since the averages include systems for which only upper limits to the abundance of Zn are available. However, we expect the corrections to be small because the systems where the Zn II doublet is below our detection limits are preferentially those with the lowest values of hydrogen column density $N$(H$^0$). Specifically, the fractions of $\sum\limits_{i=1}^{n} N{\rm(H}^0{\rm)}_i$ contributed by DLAs with undetected Zn II lines are 28% for the full sample, and 16%, 16%, and 37% respectively for subsamples 2, 3 and 4. To show that including the upper limits as detections has only a modest effect on the mean values of metallicity deduced, we have recalculated ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]}$ for the full sample twice, substituting $2\sigma$ and $1\sigma$ limits respectively in place of the $3\sigma$ limits used in Table 5 (it could indeed be argued that $3\sigma$ limits for the entire ensemble on Zn non-detections is an overly conservative approach). In this case, ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]}$ decreases from $-1.13 \pm 0.38$ (the value listed in Table 5) to $-1.16 \pm 0.40$ and $-1.20 \pm 0.48$ respectively. On the other hand, [*all*]{} three measurements in subsample 5 ($z_{\rm abs} = 3.0 - 3.5$) are upper limits and accordingly we quote the value of ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]}$ in this redshift interval as an upper limit.
For the full sample of 34 DLAs in the range $z_{\rm abs} = 0.7 - 3.4$ we find ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]} = -1.13 \pm 0.38$. This is the same value as obtained by Pettini et al. (1994) when account is taken of the different $f$-values and solar abundance scale used in our earlier study. For comparison, \[Zn/H\]$_{\rm gas} = -0.19$ along unreddened sight-lines in the solar vicinity (Roth & Blades 1995; Sembach et al. 1995—both analyses used the same $f$-values and solar scale as here). If the interstellar medium (gas+dust) near the Sun has the same composition as the Sun, this would imply that approximately 35% of Zn is in solid form. On the other hand, Pettini et al. (1997) found that for the present sample of DLAs the typical dust-to-metals ratio is approximately half that of the Galactic ISM. If we assume, therefore, that on average 83% of Zn in DLAs is in the gas phase, we obtain ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]}
= -1.13 \pm 0.38 + {\rm log (}1/0.83{\rm )} =
-1.004 \pm 0.38$, and conclude that the column density weighted abundance of Zn in DLAs is 1/11 of that of the Milky Way ISM today.
One of the motivations of the present work was to determine the redshift evolution of the metallicity of DLAs and thereby trace the increase of heavy elements in the universe from the epoch of galaxy formation to the present time. From Figure 4, where our measures of ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]}$ from Table 5 are plotted versus redshift, it can be seen that any such evolution is only mild in the present sample. Between $z = 3$ and 1.5, to which 80% of the sample refers, there appears to be little change from the typical ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]} = -1.13$. This is less surprising, however, when one considers that this redshift interval spans a period of only $\approx 3$ Gyr from 14.3 to 11.4 Gyr ago ($H_0 = 50$ [km s$^{-1}$]{} Mpc$^{-1}$; $q_0 = 0.01$) and that at these epochs evidently there was a large spread in the chemical enrichment of different DLA galaxies.
On the other hand, the upper limit ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]} \leq -1.39$ at $z > 3$ is lower than the means in the other redshift bins, providing tentative evidence for a rapid build-up of elements with time at this epoch. This suggestion is strengthened by the data of Lu et al. (1996) who found that \[Fe/H\] $-2$ in three additional DLAs at $z > 3$. (The correction to \[Fe/H\] for the fraction of Fe in solid form is likely to be small—probably less than a factor of two—at such low metallicities; Pettini et al. 1997). The lowest metallicities measured in DLAs, $Z_{\rm DLA} \simeq -2.5$, are comparable to those thought to apply to the ionised intergalactic medium producing the [Lyman $\alpha$]{} forest at redshifts $z = 2 - 3.5$ (Hellsten et al. 1997), although the large ionization corrections involved make estimates of $Z_{\rm IGM}$ considerably more uncertain than $Z_{\rm DLA}$. It is tempting, therefore, to interpret the rapid increase in metal abundances at $z < 3$ as an indication of the onset of star-formation in galaxies and to speculate that $Z \simeq -2.5$ may be an approximate ‘base’ level of metallicity on which galactic chemical evolution subsequently builds.
The recently realised ability to image high redshift galaxies directly in their ultraviolet stellar continua has led to the first attempts to sketch the global history of star formation over $\sim 80$% of the age of the universe (Madau et al. 1996 and references therein). Determinations of the volume-averaged star formation rate (SFR) from the so-called $B$ and $U$ drop-outs (galaxies with the Lyman limit in the $B$ and $U$ bands respectively) in ground-based surveys (Steidel, Pettini, & Hamilton 1995b) and in the [*Hubble Deep Field*]{} (Madau 1996) do indeed suggest an increase in the SFR from $z \simeq 4$ to $z \simeq 2.75$. As discussed by Madau et al., it is possible to convert the integrated UV luminosity density into a metal ejection rate $\dot{\rho_Z}$ per comoving volume at redshift $z$. Since the massive stars which are the main contributors to the far-UV continuum are also the major producers of heavy elements (at least those released into the ISM by Type II supernovae), the conversion does not depend sensitively on the shape of the IMF in these primordial galaxies. Rather, the principal sources of uncertainty arise from the cosmology assumed and from the amount of dust extinction suffered by the UV continuum.
Bearing in mind these uncertainties, it is of great interest to compare the values of $Z_{\rm DLA}$ deduced here with the metallicities which may be expected on the basis of Madau’s metal ejection rate. Integrating $\dot{\rho_Z}$ in Figure 3 of Madau (1996) from $z = 5.5$ to the present time yields a total density of metals $\rho_Z{\rm (}z = 0{\rm )} \approx 6.2 \times 10^6 M_{\sun}$ Mpc$^{-3}$. This corresponds to an approximately solar metallicity if the present day density of baryons in galaxies is $\rho_{\ast}{\rm (}z = 0{\rm )} \approx {\rm (}2.7 \pm 0.4{\rm )} \times 10^8
M_{\sun}$ Mpc$^{-3}$, or $\Omega_{\ast} \approx 4 \times 10^{-3}$ ($H_0 = 50$ km s$^{-1}$ Mpc$^{-1}$; Madau et al. 1996). In a closed box model, assuming that $\Omega_{\ast}(z = 0) \approx \Omega_{\rm DLA}(z = 4)$ (Storrie-Lombardi, McMahon, & Irwin 1996a) we can take $$\frac{Z(z)}{Z(0)} \simeq \frac{Z(z)}{Z_{\sun}} =
\frac{\displaystyle \int_{5.5}^{z} \dot{\rho_Z} dz}
{\displaystyle \int_{5.5}^{0} \dot{\rho_Z} dz}
\label{}$$ provided the gas consumption into stars from $z = 5.5$ to $z$ is low and $\Omega_{\rm gas}{\rm (}z {\rm)} \gg \Omega_{\ast}{\rm (}z {\rm)}$. The redshift evolution of $\Omega_{\rm DLA}$ (Storrie-Lombardi et al. 1996a) suggests that this may well be the case up to $z \simeq 2$ (as we proposed in Pettini et al. 1994).
The broken line in Figure 4 shows the increase of ${Z(z)}/{Z_{\sun}}$ with decreasing redshift calculated from equation (4) and Madau’s (1996) estimates of $\dot{\rho_Z}$. Evidently, there is rough agreement between the predicted and observed values of $Z_{\rm DLA}$. Given the current uncertainties, we consider it premature to read too much into this comparison. For example, Madau’s $\dot{\rho_Z}$ refers primarily to oxygen and the $\alpha$-elements which presumably are more abundant than zinc and iron by a factor of $2-3$ at these low metallicities (Edvardsson et al. 1993; Carney 1996). On the other hand, the broken line in Figure 4 may well underestimate the metal production rate by similar factors if star-forming galaxies at high redshift are reddened by small amounts of dust, corresponding to ${\rm E(B-V)} \approx 0.1$, as suggested by the observed slopes of the UV continua (Steidel et al. 1996).
Nevertheless, taken at face value, Figure 4 does seem to indicate that in the DLAs we see roughly the same level of metal enrichment as expected from direct observations of star-forming galaxies at these redshifts. More complex galactic chemical evolution models which use as a starting point the gas consumption indicated by the redshift evolution of $\Omega_{\rm DLA}$ (Pei & Fall 1995; Fall 1996) also reproduce the degree of metal enrichment of DLAs and the comoving rate of star formation at $z \simgt 2$. Thus it appears that, to a first approximation at least, these three independent avenues to exploring the epoch of galaxy formation—the consumption of neutral gas, the metal abundance of the absorbers, and the UV luminosity of high-redshift galaxies—give a broadly consistent picture of the early evolution of galaxies.\
### Abundances at $z < 1.5$
The situation is less clear at lower redshifts. Only four measurements make up subsample 1 in Table 5, even though this bin spans a larger interval of time than all the other subsets put together—$\approx 5$ Gyr from 11.4 to 6.3 Gyr ago (again for $H_0 = 50$ [km s$^{-1}$]{} Mpc$^{-1}$; $q_0 = 0.01$). Evidently ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]} = -0.98 \pm 0.33$ is below an extrapolation of Madau’s curve in Figure 4. However, it is difficult to assess how firm this conclusion is, given that 65% of $\sum\limits_{i=1}^{n} N{\rm(H}^0{\rm)}_i$ for subsample 1 is due to the $z_{\rm abs} = 0.6922$ absorber in 3C 286 which appears to be a large, low surface brightness galaxy (see §3.13 above). Possibly such galaxies, whose low metallicities at the present time are thought to be the result of low star formation efficiencies (McGaugh 1994; Padoan, Jimenez, & Antonuccio-Delogu 1997), come to dominate the cross-section for DLA absorption at $z \simlt 1$, if by this epoch most high surface brightness galaxies have already processed a significant fraction of their gas into stars. Furthermore, the build-up of dust which goes hand-in-hand with the production of metals is likely to introduce an increasing bias (with decreasing redshift) against galaxies in advanced stages of chemical evolution, since existing samples of damped [Lyman $\alpha$]{} systems are mostly drawn from magnitude limited optical QSO surveys (Fall & Pei 1993; Pei & Fall 1995).
At $z_{\rm abs} < 1.5$ imaging of DLA absorbers is within current observational capabilities. Although positive identifications based on spectroscopic redshifts have not yet been achieved, the candidates which have been proposed suggest a very diverse population of galaxies. While in some cases the absorbers could be on evolutionary paths similar to that of the Milky Way, the $z_{\rm abs} = 1.0093$ DLA in Q0302$-$223 being a good example (Pettini & Bowen 1997), there are also several instances where galaxies of low luminosity ($L_B \simlt 0.1 L^{\ast}$) or of low surface brightness are indicated (Steidel et al. 1994, 1995, 1997; Le Brun et al. 1997).
Thus both effects considered above—a shift of the DLA population away from ‘normal’ $L^{\ast}$ galaxies and an increasing dust bias—may contribute to the finding that $Z_{\rm DLA}$ does not increase significantly at $z_{\rm abs} < 1.5$ in Figures 3 and 4, contrary to simple expectations in a closed-box model of chemical evolution. However, it will not really be possible to proceed further, and quantify the relative importance of these two effects, without a larger sample of DLAs at intermediate redshifts. Identifying such a sample remains an urgent priority.\
Comparison with Stellar Populations of the Milky Way
----------------------------------------------------
Damped [Lyman $\alpha$]{} systems are commonly thought of as the high redshift counterparts of present-day galactic disks, although we and others (Pettini et al. 1990; York 1988) have often made the point that high column densities of neutral gas are not the prerogative of disk galaxies alone. The sample of \[Zn/H\] measurements now available is sufficiently large to allow a comparison to be made of the distribution of metallicities in DLAs with those of different stellar populations in the Milky Way. In the solar cylinder, stars in the halo, thick disk, and thin disk have distinct dynamical and abundance properties, although the distributions overlap in either parameter taken separately. It is the [*combination*]{} of chemical abundance and kinematic data that studies of Galactic evolution have focussed on; here we attempt to use this information to throw light on the nature of DLA galaxies.
Our measurements of \[Zn/H\] from Figure 3 have been plotted in the top panel of Figure 5 after converting redshift to look-back time in a cosmology compatible with stellar ages. The lower panel in Figure 5 shows the age-metallicity relationship for disk stars determined in the landmark study by Edvardsson et al. (1993). This sample includes stars with the kinematics of both thin and thick disk, defined in terms of the mean velocity perpendicular to the Galactic plane: $\langle |W| \rangle = 19$ [km s$^{-1}$]{} and $\approx 42$ [km s$^{-1}$]{} for thin and thick disk stars respectively (Freeman 1991). In constructing their sample, Edvardsson et al. aimed to include approximately equal numbers of stars in each metallicity bin above $Z = 0.1 Z_{\sun}$; consequently, metal-poor stars are relatively over-represented in Figure 5.
Stellar ages are notoriously uncertain, as is the mapping of redshift to look-back time. However, even allowing for an arbitrary sliding of the points in Figure 5 along the $x$-axes, the metallicity measurements in DLAs evidently do not match the chemical evolution of the Milky Way disk. The [*typical*]{} value $Z_{\rm DLA} = -1.13$ is lower than that of even the most metal-deficient stars in the Edvardsson et al. survey, and at all ages the spread of chemical abundances in the disk is smaller than that of the DLA sample.
This point is reinforced by Figure 6, where we compare the metallicity distribution of DLAs with those of stars in the thin disk, thick disk and halo populations; in Figure 7 we show the comparison with the metallicity histogram for globular clusters. Values for disk stars are from the work by Wyse & Gilmore (1995). These authors combined spectroscopic determinations of \[Fe/H\] for a sample of F and G stars located $1-5$ kpc from the plane with data for samples near the Sun, paying particular attention to including only stars with potential main-sequence lifetimes greater than 12 Gyr. That is, their combined sample should be complete, in the sense of not missing disk stars which have by now evolved away from the main sequence, and the resulting metallicity distributions presumably provide an integrated record of the chemical evolution of the disk. The thin disk distribution in Figure 6 includes the low metallicity tail discussed by Wyse & Gilmore (1995); similarly, the thick disk histogram is consistent with the metal-weak tail shown in Figure 22 of Beers & Sommer-Larsen (1995). The halo sample is from the survey of high proper-motion stars in the solar neighbourhood by Laird et al. (1988), while the histogram in Figure 7 is based on the distribution of \[Fe/H\] in 40 globular clusters plotted in Figure 16 of Carney et al. (1996).
The comparison between the metallicity distribution of DLAs and those of stellar populations in the Galaxy is complicated by the fact that about half of the values which make up the bins with \[Zn/H\]$_{\rm DLA} \leq -0.8$ in Figures 6 and 7 correspond to upper limits of \[Zn/H\] in our survey. Were we to exclude the upper limits from the sample, the resulting distribution would be skewed to higher metallicities. This is also the case if they are included in the sample as detections, as we have done; therefore [*the true distribution of $Z_{\rm DLA}$ is both broader and shifted towards lower metallicities (by undetermined amounts) than the histogram reproduced in Figures 6 and 7*]{}.
Bearing this in mind, the middle and bottom panels of Figure 6 show that the metallicity distribution of DLA galaxies is different from those of long-lived stars in the Galactic disk. Although there is some overlap with the thick disk histogram, the bulk of stars in the disk of the Milky Way apparently formed from gas which was significantly more metal-rich than that giving rise to damped [Lyman $\alpha$]{} systems. The narrow distributions for disk stars in Figure 6 reflect the finding by Edvardsson et al. (1993) that the average metallicity has increased very little over the lifetime of the disk; the scatter at any age in the bottom panel of Figure 5 is nearly as large as the difference in mean metallicity over the entire time span considered. This is also the case for the old open clusters of the Milky Way disk (Friel 1995).
The width of the $Z_{\rm DLA}$ distribution is comparable to those of halo stars and globular clusters, but it peaks at a higher metallicity. This is probably a real effect, rather than being due to the inclusion of upper limits in our sample (as discussed above), since the column density weighted mean metallicity is ${\rm [} \langle{\rm Zn/H}_{\rm DLA}\rangle {\rm ]} = -1.13$. We consider it unlikely that the offset between the observed and true peaks of the $Z_{\rm DLA}$ distribution is as large as required to bring the histograms in the top panel of Figure 6 and in Figure 7 into agreement. Rather we favour the interpretation that, as a whole, the population of DLA galaxies is genuinely more metal enriched than the stellar components of the Galactic halo.
The comparisons discussed above lead to two possible conclusions concerning the nature of damped [Lyman $\alpha$]{} galaxies. The most straightforward interpretation is that a wide range of galaxy morphological types, at different stages of chemical evolution, make up the the DLA population. Available imaging data at $z \simlt 1$ are certainly consistent with this view. A more intriguing possibility is that DLA systems at high redshift arise primarily in the spheroidal component of the present-day galaxy population, by analogy with the interpretation of the $U$ drop-out galaxies put forward by Steidel et al. (1995, 1996). In the Milky Way, the halo and inner bulge may well be related, with the halo having lost $\approx 90$% of its mass to the bulge (e.g. Wyse, Gilmore, & Franx 1997); in this picture the halo-bulge system is an evolutionary sequence parallel to that of the thick disk-thin disk. One could speculate, then, that the distribution of $Z_{\rm DLA}$, with its peak at a higher metallicity than halo stars and globular clusters, reflects different stages in the transition from metal-poor halo to a predominantly metal-rich bulge (Ibata & Gilmore 1995).\
### Divergent Clues from the Absorption Line Profiles?
The message conveyed by Figure 6 contrasts with the interpretation by Wolfe and collaborators of the complex absorption line profiles, often extending over more than 100 [km s$^{-1}$]{}, revealed by high resolution spectroscopy of DLAs (Wolfe 1995; Prochaska & Wolfe 1997b). These authors have argued that in many cases the different components which make up the absorption lines are not distributed at random in velocity; rather, there appears to be a more regular trend of decreasing optical depth with increasing velocity difference from the wavelength where the absorption is strongest. This ‘edge-leading asymmetry’ is the pattern which would be produced by a rotating thick disk, intersected at some distance from the centre, if the average density of gas falls off with distance from the centre and from midplane. Prochaska & Wolfe show that the frequency with which such absorption profiles are encountered is consistent with expectations for randomly oriented disks; this leads them to conclude that most, if not all, DLAs arise in large ($R > 10$ kpc) disks with high rotation velocities ($v_{\rm rot} \simgt 200$ [km s$^{-1}$]{}). Such structures, if common at $z \simgt 2$, are very difficult to explain in currently favoured models of galaxy formation (e.g. Baugh et al. 1997).
The Milky Way is the only galaxy for which we have a record of both chemical abundances and kinematics over its past history. Based on this body of data, the metallicities we measure in the damped [Lyman $\alpha$]{} systems appear incompatible with the rotating disk interpretation put forward by Prochaska & Wolfe. This can be appreciated by considering compilations of metallicities and velocities now available for large samples of stars, such as that published recently by Carney et al. (1996). From their Figures 1 and 3 it can be seen that, of the stars with metal abundances similar to those of DLAs, approximately half have [*retrograde*]{} motions; at a metallicity $Z = -1.1$ the mean velocity relative to the disk rotation is $\langle {\rm V} \rangle \simeq -150$ [km s$^{-1}$]{}. This point is best illustrated by Figure 5 of Carney et al. which shows the metallicity histograms in various intervals of ${\rm V}$; our distribution of $Z_{\rm DLA}$ corresponds to values of ${\rm V}$ in the range $\approx -100$ to $\approx -200$ [km s$^{-1}$]{}. Evidently, when our Galaxy had an average metallicity of $\simlt 1/10$ of solar, it did not exhibit the kinematics of a disk rotating at $\sim 200$ [km s$^{-1}$]{}.
Reconciling these contrasting clues to the nature of damped [Lyman $\alpha$]{} galaxies is an important task for the future. Here we put forward three possible ways out of the current impasse:
1\. Our Galaxy is atypical, and the physical processes which gave rise to its stellar populations were not shared by most other galaxies at high redshifts. Although this possibility cannot be discounted, it is not a very constructive hypothesis to take refuge in, as it will be difficult to test it observationally—at least in the near future.
2\. The absorption profiles are being overinterpreted. A possible concern here is that material whose motion is due not to rotation but to energetic events, such as supernova shocks, may contribute to the ultraviolet absorption lines, since these transitions are sensitive to even relatively small column densities of gas. The ‘edge-leading asymmetry’ interpretation was first proposed by Lanzetta & Bowen (1992) in their analysis of 13 Mg II absorption components spread over 250 [km s$^{-1}$]{} in the $z_{\rm abs} = 0.39498$ DLA in Q1229$-$021. However, it is far from clear that this is really a massive disk; from their analysis of [*HST*]{} images of the field, Le Brun et al. (1997) propose that the absorber is instead a faint ($L_B < 0.1 L^{\ast}$) low surface brightness galaxy. Furthermore, strong Mg II absorption spanning $\approx 300$ [km s$^{-1}$]{} can also be produced by galaxies which are nearly face-on, such as M61 (Bowen, Blades, & Pettini 1996). All these factors cast some doubts on a detailed correspondence between the profiles of ultraviolet absorption lines and the large-scale kinematics of the intervening galaxies.
3\. A third option, and one which we have already proposed, is that DLA galaxies comprise a mix of different morphological types. Thus, it is conceivable that some do exhibit the kinematics of rapidly rotating disks, while others may be spheroids or irregular star-forming galaxies with less ordered velocity fields. This is a hypothesis which [*can*]{} be tested. As more cases become available where both kinematics and chemical abundances are measured in the same DLA, it will be of great interest to examine whether there is any correlation between these two parameters, as found in the stellar populations of the Milky Way.\
SUMMARY AND SUGGESTIONS FOR FUTURE WORK
=======================================
We have assembled the largest sample of damped [Lyman $\alpha$]{} systems for which metallicities have been measured free from the complications introduced by dust depletions. The expanded data set reinforces the two main conclusions reached in our earlier study (Pettini et al. 1994): (1) DLAs are generally metal-poor, [*at all redshifts sampled*]{}; and (2) there is a large spread in abundances at all epochs. We interpret these findings as evidence for a protracted epoch of galaxy formation, and propose that galaxies of different morphological types and at different stages of chemical evolution make up the DLA population.
The metallicity distribution of DLAs is broader and peaks at lower metallicities than those of either the thin or thick disk of our Galaxy. Thus, the chemical abundance data presented here do not support the interpretation of the absorption line profiles in terms of thick disks with rotation velocities $v_{\rm rot} \simgt 200$ [km s$^{-1}$]{} most recently discussed by Prochaska & Wolfe (1997b). This apparent discrepancy may be resolved by further work on both the kinematics and the abundances. With the near-infrared spectrographs now being built for 8-10 m telescopes it will be possible to detect the familiar optical emission lines from star-forming regions in the absorbing galaxies. The widths of these features are likely to be more representative of the global kinematics than the ultraviolet absorption lines which can be so easily affected by local phenomena such as interstellar shocks. On the abundance front, the ratios of chemical elements manufactured in different nucleosynthetic processes have been used to good effect in unravelling the history of star formation in our Galaxy; the same techniques are now beginning to be applied to high redshift DLAs (Pettini et al. 1995b; Lu et al. 1996).
The column density weighted mean metallicity of DLAs at $z \simgt 2$ is in agreement with expectations based on the metal ejection rate deduced by Madau (1996) from the integrated ultraviolet luminosity of star forming galaxies at these redshifts. Our data, when combined with the \[Fe/H\] measurements by Lu et al. (1996), appear to reflect the rapid increase in the comoving star-formation rate between $z \approx 4$ and $\approx 2$ indicated by the relative numbers of $B$ and $U$ drop-outs in the [*Hubble Deep Field*]{}. While these comparisons are of necessity still very approximate, the implication seems to be that observations of DLAs provide a reasonably accurate census of metal enrichment at these epochs. It is encouraging that three independent methods which have been applied to the quest for the epoch of galaxy formation—the global star formation rate deduced from the ultraviolet luminosity of high-redshift galaxies, the rate of consumption of neutral gas implied by the redshift evolution of $\Omega_{\rm DLA}$, and the metallicity of DLAs—apparently give a broadly consistent picture of the universe at $z \simgt 2$.
This is not the case at $z \simlt 1.5$, where $Z_{\rm DLA}$ apparently does not rise as expected from simple models of cosmic chemical evolution. There are a number of plausible explanations for this, including the effects of dust, as discussed extensively by Fall and collaborators, and an increasing contribution of low surface brightness galaxies to the cross-section for DLA absorption. The major obstacle to progress in this area is still the paucity of DLAs with measured element abundances at intermediate redshifts. And yet it is essential to follow the evolution of the DLA population to the present time in order to be confident of our interpretation of the high redshift data. New DLAs at $z \simlt 1$ are still being identified and the sample is slowly growing. With STIS on the [*HST*]{} measurements of \[Zn/H\] can be extended to redshifts lower than the limit $z_{\rm abs} \simeq 0.65$ of ground-based observations. In the next few years the 2dF and Sloan sky surveys (Taylor 1995; Gunn & Weinberg 1995) are expected to increase the number of known DLAs by one order of magnitude. With 8-10 m telescopes it will then be possible to repeat surveys such as this one towards substantially fainter, and potentially more reddened, QSOs. Such programmes should lead to a better assessment of the significance of dust bias in current DLA samples. Finally, with large telescopes we will soon be able to measure element abundances from the optical emission lines of galaxies at redshifts $z \approx 0.1-0.5$. Such data will complement in a very important way the information provided by galaxies selected from their absorption cross-section.\
We are grateful to the UK and Australian Time Assignment committees for generous allocations of telescope time on the WHT and the AAT, and to the technical staff at both Observatories for excellent support with the observations. The WHT Service Observations scheme helped bring this demanding observing programme to completion. We should like to express our sincere thanks to: C. Hazard and P. Hewett for supplying us with positions and finding charts of QSOs with candidate damped [Lyman $\alpha$]{} systems; J. Lewis for assistance with the reduction of some of the spectra; J. Laird and R. Wyse for providing the stellar data used in Figures 6 and 7; C. Jenkins and K. Lipman for help with some statistical aspects of the analysis; M. Fall, G. Gilmore, P. Madau, and S. Ryan for illuminating discussions on several issues relating to stellar populations and galactic chemical evolution; and C. Steidel, J. Prochaska, and D. York for useful comments on an earlier version of the paper. R.W.H. acknowledges financial assistance from the Australian Research Council.
Anders, E., & Grevesse, N. 1989, Geochim. Cosmochim. Acta, 53, 197 Baugh, C.M., Cole, S., Frenk, C.S., & Lacey, C.G. 1997, ApJ, submitted (astro-ph/9703111) Beers, T.C., & Sommer-Larsen, J. 1995, ApJS, 96, 175 Bergeson, S.D., & Lawler, J.E. 1993, ApJ, 408, 382 Bowen, D.V., Blades, J.C., & Pettini, M. 1996, ApJ, 472, L77 Carney, B.W. 1996, PASP, 108, 900 Carney, B.W., Laird, J.B., Latham, D.W., & Aguilar, L.A. 1996, AJ, 112, 668 Chaffee, F.H., Foltz, C.B., Hewett, P.C., Francis, P.A., Weymann, R.J., Morris, S.L., Anderson, S.F., & MacAlpine, G.M. 1991, AJ, 102, 461 Ćirković, M.M., Lanzetta, K.M., Baldwin, J., Williger, G., Carswell, R.F., Potekhin, A.Y., & Varshalovich, D.A. 1997, ApJ, submitted Edvardsson, B., Andersen, J., Gustafsson, B., Lambert, D.L., Nissen, P.E., & Tomkin, J. 1993, A&A, 275, 101 Fall, S.M. 1996, in HST and the High Redshift Universe, ed. N. Tanvir, A. Aragon-Salamanca, & J.V. Wall (Singapore: World Scientific), in press. Fall, S.M., & Pei, Y.C. 1993, ApJ, 402, 479 Fan, X.M., & Tytler, D. 1994, ApJS, 94, 17 Freeman, K.C. 1991, in Dynamics of Disc Galaxies, ed. B. Sundelius (G$\ddot{o}$teborg University, G$\ddot{o}$teborg), 15 Friel, E.D. 1995, ARAA, 33, 381 Ge, J., & Bechtold, J. 1997, ApJ, in press (astro-ph/9701041) Gunn, J.E., & Weinberg, D.H. 1995, in Wide Field Spectroscopy and the Distant Universe, ed. S.J. Maddox & A. Aragon-Salamanca, (Singapore: World Scientific), 3 Hazard, C. 1994, private communication Hazard, C., McMahon, R.G., & Morton, D.C. 1987, MNRAS, 229, 371 Hazard, C., Morton, D.C., McMahon, R.G., Sargent, W.L.W., & Terlevich, R. 1986, MNRAS, 223, 87 Hellsten, U., Dav$\acute{e}$, R., Hernquist, L., Weinberg, D.H., & Katz, N. 1997, ApJ, in press (astro-ph/9701043) Hunstead, R.W., Pettini, M., & Fletcher, A.B. 1990, ApJ, 365, 23 Ibata, R.A., & Gilmore, G. 1995, MNRAS, 275, 605 Laird, J.B., Rupen, M.P., Carney, B.W., & Latham, D.W. 1988, AJ, 96, 1908 Lanzetta, K.M., & Bowen, D.V. 1992, ApJ, 391, 48 Lanzetta, K.M., Wolfe, A.M., & Turnshek, D.A. 1995, ApJ, 440, 435 Le Brun, V. Bergeron, J., Boisse, P., & Deharveng, J.M. 1997, A&A, in press Levshakov, S.A., Chaffee, F.H., Foltz, C.B., & Black, J.H. 1992, A&A, 262, 385 Lu, L., Sargent, W.L.W., Barlow, T.A., Churchill, C.W., & Vogt, S.S. 1996, ApJS, 107, 475 Lu, L., Savage, B.D., Tripp, T.M., & Meyer, D.M. 1995, ApJ, 447, 597 Lu, L., & Wolfe, A.M. 1994, AJ, 108, 44 Lu, L., Wolfe, A.M., Turnshek, D.A., & Lanzetta, K.M. 1993, ApJS, 84, 1 Madau, P., 1996, in Star Formation Near and Far, Proc. 7th Annual Astrophysics Conference in Maryland, ed. S.S. Holt & G.L. Mundy (AIP: New York), in press (astro-ph/9612157) Madau, P., Ferguson, H.C., Dickinson, M., Giavalisco, M., Steidel, C.C., & Fruchter, A. 1996, MNRAS, 283, 1388 McGaugh, S.S. 1994, ApJ, 426, 135 Meyer, D.M., Lanzetta, K.M., & Wolfe, A.M. 1995, ApJ, 451, L13 Meyer, D.M., Welty, D.E., & York, D.G. 1989, ApJ, 343, L37 Meyer, D.M., & York, D.G. 1992, ApJ, 399, L121 Molaro, P., D’Odorico, S., Fontana, A., Savaglio, S., & Vladilo, G. 1996, A&A, 308, 1 Padoan, P., Jimenez, R., & Antonuccio-Delogu, V. 1997, ApJ, in press Pei, Y.C., & Fall, S.M. 1995, ApJ, 454, 69 Pettini, M., Boksenberg, A., & Hunstead, R.W. 1990, ApJ, 348, 48 Pettini, M., & Bowen, D.V. 1997, A&A, submitted Pettini, M., King, D.L., Smith, L.J., & Hunstead, R.W. 1995a, in QSO Absorption Lines, ed. G. Meylan (Berlin: Springer-Verlag), 71 Pettini, M., King, D.L., Smith, L.J., & Hunstead, R.W. 1997, ApJ, in press (April 1, 1997 issue) Pettini, M., Lipman, K., & Hunstead, R.W. 1995b, ApJ, 451, 100 Pettini, M., Smith, L.J., Hunstead, R.W., & King, D.L. 1994, ApJ, 426, 79 Prochaska, J.X., & Wolfe, A.M. 1996, ApJ, 470, 403 Prochaska, J.X., & Wolfe, A.M. 1997a, ApJ, 474, 140 Prochaska, J.X., & Wolfe, A.M. 1997b, in preparation Roth, K.C., & Blades, J.C. 1995, ApJ, 445, L95 Sargent, W.L.W., Boksenberg, A., & Steidel, C.C. 1988, ApJS, 68, 539 Savaglio, S., D’Odorico, S., & Moller, P. 1994, A&A, 281, 331 Sembach, K.R., Steidel, C.C., Macke, R.J., & Meyer, D.M. 1995, ApJ, 445, L27 Smette, A., Robertson, J.G., Shaver, P.A., Reimers, D., Wisotzki, L., & Kohler, T. 1995, A&A Supp, 113, 199 Smith, H.E., Cohen, R.D., & Bradley S.E. 1986, ApJ, 310, 583 Smith, L.J., Pettini, M., King, D.L., & Hunstead, R.W. 1996, in From Stars to Galaxies—the Impact of Stellar Physics on Galaxy Evolution, ed. C. Leitherer, U. Fritze-von Alvensleben & J. Huchra, Astr. Soc. Pacific Conf. Ser., 98, 559 Steidel, C.C., Bowen, D.V., Blades, J.C., & Dickinson, M. 1995a, ApJ, 440, L45 Steidel, C.C., Dickinson, M., Meyer, D.M., Adelberger, K.L., & Sembach, K.R. 1997, ApJ, in press Steidel, C.C., Giavalisco, M., Pettini, M., Dickinson, M., & Adelberger, K.L. 1996, ApJ, 462, L17 Steidel, C.C., Pettini, M., & Hamilton. D. 1995b, AJ, 110, 2519 Steidel, C.C., Pettini, M., Dickinson, M., & Persson, S.E. 1994, AJ, 108, 2046 Storrie-Lombardi, L.J., McMahon, R.G., & Irwin, M.J. 1996a, MNRAS, 283, L79 Storrie-Lombardi, L.J., McMahon, R.G., Irwin, M.J., & Hazard, C. 1996b, ApJ, 468, 121 Taylor, K. 1995, in Wide Field Spectroscopy and the Distant Universe, ed. S.J. Maddox & A. Aragon-Salamanca, (Singapore: World Scientific), 15 Tripp, T.M., Lu, L., & Savage, B.D. 1996, ApJS, 102, 239 Turnshek, D.A., Wolfe, A.M., Lanzetta, K.M., Briggs, F.H., Cohen, R.D., Foltz, C.B., Smith, H.E., & Wilkes, B.J. 1989, ApJ, 344, 567 Wolfe, A.M. 1995, in QSO Absorption Lines, ed. G. Meylan (Berlin: Springer-Verlag), p.13 Wolfe, A.M., Fan, X.-M., Tytler, D., Vogt, S.S., Keane, M.J., & Lanzetta, K.M. 1994, ApJ, 435, L101 Wolfe, A.M., Lanzetta, K.M., Foltz, C.B., & Chaffee, F.H. 1995, ApJ, 454, 698 Wyse, R.F.G., & Gilmore, G. 1995, AJ, 110, 2771 Wyse, R.F.G., Gilmore, G., & Franx, M. 1997, ARAA, 35, in press (astro-ph/9701223) York, D.G. 1988, in QSO Absorption Lines: Probing the Universe, ed. J.C. Blades, D.A. Turnshek, & C.A. Norman (Cambridge: Cambridge Univ. Press), 227
[lcccccc]{} Full Sample & 0.6922 $-$ 3.3901 & 7.8 $-$ 14.7 & 34 & 19 & 15 & $-1.13 \pm 0.38$ & & & & & Subsample 1 & 0.50 $-$ 1.49 & 6.3 $-$ 11.4 & 4 & 4 & 0 & $-0.98 \pm 0.33$ Subsample 2 & 1.50 $-$ 1.99 & 11.4 $-$ 12.7 & 8 & 6 & 2 & $-0.96 \pm 0.44$ Subsample 3 & 2.00 $-$ 2.49 & 12.7 $-$ 13.6 &12 & 6 & 6 & $-1.23 \pm 0.38$ Subsample 4 & 2.50 $-$ 2.99 & 13.6 $-$ 14.3 & 7 & 3 & 4 & $-1.11 \pm 0.27$ Subsample 5 & 3.00 $-$ 3.49 & 14.3 $-$ 14.8 & 3 & 0 & 3 & $\leq -1.39$
[^1]: We use the conventional notation where \[X/Y\] = log (X/Y)$-$log (X/Y)$_{\sun}$
[^2]: This set of atomic parameters and solar abundance introduces a correction of $-0.148$ to the values of \[Zn/H\] published in Pettini et al. (1994)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This proceeding briefly summarizes our recent investigations on the correlations of flow harmonics in 2.76A TeV Pb–Pb collisions with viscous hydrodynamics [VISH2+1]{}. We calculated both the symmetric cumulants $SC^{v}(m, n)$ and the normalized symmetric cumulants $NSC^{v}(m, n)$, and found $v_{2}$ and $v_{4}$, $v_{2}$ and $v_{5}$, $v_{3}$ and $v_{5}$ are correlated, $v_{2}$ and $v_{3}$, $v_{3}$ and $v_{4}$ are anti-correlated. We also found $NSC^{v}(3, 2)$ are insensitive to the QGP viscosity, which are mainly determined by the initial conditions.'
address:
- '$^1$School of Science, Huzhou University, Huzhou 313000, China'
- '$^{2}$Department of Physics and State Key Laboratory of Nuclear Physics and Technology, Peking University, Beijing 100871, China'
- '$^{3}$Collaborative Innovation Center of Quantum Matter, Beijing 100871, China'
- '$^{4}$Niels Bohr Institute, University of Copenhagen, Blegdamsvej 17, 2100 Copenhagen, Denmark'
- '$^{5}$Center for High Energy Physics, Peking University, Beijing 100871, China'
author:
- 'Xiangrong Zhu$^{1,2,3}$, You Zhou$^{4}$, Haojie Xu$^{2,3}$, Huichao Song$^{2,3,5}$'
title: 'Investigating the correlations of flow harmonics in 2.76A TeV Pb–Pb collisions'
---
Introduction\[sec:intro\]
=========================
The ultra-relativistic heavy-ion collision programs at RHIC and LHC have been utilized to the produce extreme conditions to create and study the strongly interacting Quark-Gluon Plasma (QGP), a deconfined state of quarks and gluons. One of the observables to probe the properties of the hot QCD matter is the azimuthal anisotropy in the momentum distribution of the produced particles. The anisotropic flow coefficient $V_{n}$ is generally defined through a Fourier decomposition of the emitted particle distribution as a function of the azimuthal angle $\varphi$, $P(\varphi) = \frac{1}{2\pi} \sum_{n=-\infty}^{+\infty} {\overrightarrow{V_{n}} \, e^{-in\varphi} }$ where $\overrightarrow{V_{n}} =v_{n}\,e^{in\Psi_{n}}$. The $v_{n}$ is the $n$-th order anisotropic flow harmonics and $\Psi_{n}$ is the symmetry plane angle. Recently, the correlations between different order $\overrightarrow{V_{m}}$ and $\overrightarrow{V_{n}}$ have been investigated both theoretically and experimentally, which not only focus on the correlations of the orientations of different flow-vector $\Psi_{n}$ [@Aad:2014fla; @Qiu:2012uy; @Teaney:2012gu; @Jia:2012ju; @Niemi:2015qia] but also on the correlations of the magnitudes of different flow-vector $v_{n}$ [@ALICE:2016kpq; @Aad:2015lwa; @Niemi:2012aj; @Giacalone:2016afq; @Qian:2016pau; @Zhu:2016puf].
In this proceeding, we will briefly review our recent investigations on the correlations of flow harmonics in 2.76A TeV Pb–Pb collisions using the event-by-event viscous hydrodynamics [VISH2+1]{} with different initial conditions and the QGP shear viscosity [@Zhu:2016puf].
Setup of the calculation\[sec:setup\]
=====================================
The [VISH2+1]{} is a (2+1)-d viscous hydrodynamic model to describe the fluid expansion of the QGP with longitudinal boost-invariance [@Song:2007fn; @Shen:2014vra]. In the following calculations, we use an equation of state (EoS) [s95p-PCE]{} [@Huovinen:2009yb], which matches the partially chemical equilibrium hadron resonance gas at low temperature and the lattice QCD data at high temperature. Three different initial conditions, [MC-Glauber]{}, [MC-KLN]{} [@Drescher:2006pi; @Hirano:2009ah], and [AMPT]{} [@Xu:2016hmp], are used in our calculations to study the influence of initial conditions on the correlations of flow harmonics. To explore the sensitivity of the QGP shear viscosity, we choose two values of the specific shear viscosity $\eta/s$ for each initial condition. More specifically, $\eta/s=$ 0.08 and 0.20, for the [MC-Glauber]{} and [MC-KLN]{} initial conditions, and $\eta/s=$0.08 and 0.16 for the [AMPT]{} initial conditions. The hydrodynamic output is converted to final hadron distributions along the freeze-out surface at the temperature $T_{dec}$ = 120 MeV via the Cooper-Frye prescription [@Shen:2011eg; @Song:2013qma]. The initial time of hydrodynamic evolution $\tau_{0}$ and the normalization factors of initial entropy density profiles have been tuned to fit the 0-5% centrality data of $dN/d\eta$ and $p_{\rm T}$ spectra of $\pi$, $K$, and $p$. The bulk viscosity, net baryon density, and the heat conductivity are set to zero to simplify the calculations.
Results and discussion\[sec:results\]
=====================================
{width="0.495\linewidth" height="5cm"} {width="0.495\linewidth" height="5cm"}
We firstly calculate the symmetric cumulants $SC^{v}(m, n)$ defined as $SC^{v}(m, n)= \left< v_{m}^{2} \, v_{n}^{2} \right> - \left< v_{m}^{2} \right> \left< v_{n}^{2} \right>$. The upper panels in Fig. \[fig:scvmvn1\] (left) show the comparison between our calculations and the ALICE measurements. We find that, for these initial conditions and different values of $\eta/s$, [VISH2+1]{} calculations qualitatively capture the centrality dependence of the flow correlations, but not quantitatively. Specially, even though [VISH2+1]{} with [AMPT]{} initial conditions gives good descriptions for the integrated flow $v_{n}$ ($n\leq 4$) [@Zhu:2016puf], it can only reproduce the typical features of the correlations of flow harmonics. This indicates that the correlations between different flow harmonics are more sensitive to the details of hydrodynamic calculations than the individual $v_{n}$ coefficients alone.
Similar to the ALICE data, our model gives negative $SC^{v}(3, 2)$ and positive $SC^{v}(4, 2)$, which suggests $v_2$ and $v_3$ are anti-correlated, while $v_2$ and $v_4$ are correlated. The results reveal that, for a given event, the case with an elliptic flow $v_{2}$ larger than the averaged $\langle v_{2} \rangle$ enhances the probability of finding a triangular flow $v_{3}$ smaller than the averaged $\langle v_{3} \rangle$ and the probability of finding a quadrangular flow $v_{4}$ larger than the averaged $\langle v_{4} \rangle$. The strengths of $SC^{v}(3, 2)$ and $SC^{v}(4, 2)$ are more suppressed with larger $\eta/s$ for each initial condition, which suggests that both $SC^{v}(3, 2)$ and $SC^{v}(4, 2)$ are strongly influenced by the QGP viscosity. By comparing with the symmetric cumulants of the initial state, $SC^{\varepsilon}(m, n)$, we observe the signs of $SC^{v}(3, 2)$ and $SC^{v}(4, 2)$ are determined by the signs of $SC^{\varepsilon}(3, 2)$ and $SC^{\varepsilon}(4, 3)$, respectively.
Figure \[fig:scvmvn1\] (right) presents our predictions for the centrality dependent $SC^{v}(m, n)$ with $(m, n)=$ (5, 2), (5, 3), and (4, 3), together with their corresponding correlators $SC^{\varepsilon}(m, n)$ from the initial state. We observe that, for each initial condition, the [VISH2+1]{} gives positive values for $SC^{v}(5, 2)$ and $SC^{v}(5, 3)$, and negative values for $SC^{v}(4, 3)$. This reveals $v_{2}$ and $v_{5}$, $v_{3}$ and $v_{5}$ are correlated, while $v_{3}$ and $v_{4}$ are anti-correlated. We also notice that their correlation strengths become weaker with the increase of $\eta/s$. The signs of $SC^{v}(5, 2)$ and $SC^{v}(5, 3)$ are consistent with their initial state correlators $SC^{\varepsilon}(5, 2)$ and $SC^{\varepsilon}(5, 3)$. However, $SC^{v}(4, 3)$ and $SC^{\varepsilon}(4, 3)$ show opposite signs for the [MC-Glauber]{} and [AMPT]{} initial conditions. This can be well understood from the proposed relationship of $v_{4}e^{i4\Phi}=a_{0}\varepsilon_{4}e^{i4\Psi_{4}}+a_{1}(\varepsilon_{2}e^{i2\Psi_{2}})^{2}$ [@Gardim:2011xv; @Teaney:2012ke], where the $\varepsilon_2^2$ term makes the dominant contributions in non-central collisions [@Yan:2015jma]. As a result, the signs of $SC^{v}(4, 3)$ are affected by the correlation between $\varepsilon_{2}$ and $\varepsilon_{3}$ and the correlation between $\varepsilon_{3}$ and $\varepsilon_{4}$ rather than the correlation between $\varepsilon_{3}$ and $\varepsilon_{4}$ alone.
{width="0.8\linewidth" height="8cm"}
Figure \[fig:scRvmvn\] shows the normalized correlator of flow harmonics and initial eccentricity coefficients, which are defined as $NSC^{v}(m, n)=SC^{v}(m, n)/\langle v_{m}^2 \rangle\langle v_{n}^2 \rangle$ and $NSC^{\varepsilon}(m, n)=SC^{\varepsilon}(m, n)/\langle \varepsilon_{m}^2 \rangle\langle \varepsilon_{n}^2 \rangle$, respectively. We find that $NSC^{v}(4, 2)$, $NSC^{v}(5, 2)$, and $NSC^{v}(5, 3)$ are sensitive to both initial conditions and $\eta/s$. Meanwhile, their corresponding $NSC^{\varepsilon}$ correlator are separated for different initial conditions. Compared to the ALICE data [@ALICE:2016kpq], the calculated $NSC^{v}(4, 2)$ are roughly fit the data for [AMPT]{} initial conditions and $\eta/s=0.16$ and for [MC-Glauber]{} initial conditions and $\eta/s=0.2$. This indicates the normalized symmetric cumulants can be used to constrain the QGP viscosity for different initial conditions. The $NSC^{v}(3, 2)$ from different combinations of initial conditions and $\eta/s$ are all roughly fit the ALICE data, which also roughly overlap with each other. Such $\eta/s$ independent character of $NSC^{v}(3, 2)$ can be naturally understood from the widely accepted results $v_2\approx k_{1}\varepsilon_2$ and $v_3\approx k_{2}\varepsilon_3$, where $k_{1}$ and $k_{2}$ are the proportion coefficients. Meanwhile, the $NSC^{\varepsilon}(3, 2)$ from the three initial conditions used in our calculations also almost overlap from central to semi-central collisions. In contract, panels (h) and (j) show that, although the $NSC^{\varepsilon}(4, 3)$ strongly depends on the initial conditions, the $NSC^{v}(4, 3)$ almost overlap, which is insensitive to the initial conditions used in our calculation.
Summary\[sec:summary\]
======================
In summary, we investigated the correlations between flow harmonics in 2.76A TeV Pb–Pb collisions using the event-by-event viscous hydrodynamics [VISH2+1]{} with [MC-Glauber]{}, [MC-KLN]{}, and [AMPT]{} initial conditions. We found the symmetric cumulants $SC^{v}(m, n)$ are sensitive to both initial conditions and the QGP shear viscosity, The normalized symmetric cumulants $NSC^{v}(3, 2)$ are mainly determined by the correlation in the initial state, which are insensitive to the QGP viscocity. In contrast, $NSC^{v}(4, 2)$, $NSC^{v}(5, 2)$, $NSC^{v}(5, 3)$ are sensitive to both initial conditions and $\eta/s$. We found that the correlations of flow harmonics are more sensitive to the details of theoretical calculations than individual flow harmonics, which could be used for further constraint the properties of the QGP.
This work is supported by the NSFC and the MOST under grant Nos.11435001 and 2015CB856900, and partially supported by China Postdoctoral Science Foundation under grant No. 2015M570878 and 2015M580908, by the Danish Council for Independent Research, Natural Sciences, and the Danish National Research Foundation (Danmarks Grundforskningsfond).
References {#references .unnumbered}
==========
[9]{} G. Aad [*et al.*]{} \[ATLAS Collaboration\], Phys. Rev. C [**90**]{}, no. 2, 024905 (2014). Z. Qiu and U. Heinz, Phys. Lett. B [**717**]{}, 261 (2012).
D. Teaney and L. Yan, Nucl. Phys. A [**904-905**]{}, 365c (2013).
J. Jia and D. Teaney, Eur. Phys. J. C [**73**]{}, 2558 (2013).
H. Niemi, K. J. Eskola and R. Paatelainen, Phys. Rev. C [**93**]{}, no. 2, 024907 (2016).
J. Adam [*et al.*]{} \[ALICE Collaboration\], arXiv:1604.07663 \[nucl-ex\];
G. Aad [*et al.*]{} \[ATLAS Collaboration\], Phys. Rev. C [**92**]{}, no. 3, 034903 (2015). H. Niemi, G. S. Denicol, H. Holopainen and P. Huovinen, Phys. Rev. C [**87**]{}, no. 5, 054901 (2013).
G. Giacalone, L. Yan, J. Noronha-Hostler and J. Y. Ollitrault, arXiv:1605.08303 \[nucl-th\]. J. Qian and U. Heinz, arXiv:1607.01732 \[nucl-th\]. X. Zhu, Y. Zhou, H. Xu and H. Song, arXiv:1608.05305 \[nucl-th\]. H. Song and U. Heinz, Phys. Lett. [**B658**]{}, 279 (2008); Phys. Rev. C [**77**]{}, 064901 (2008); Phys. Rev. C [**78**]{}, 024902 (2008); H. Song, Ph.D Thesis, The Ohio State University, August 2009, arXiv:0908.3656 \[nucl-th\]. C. Shen, Z. Qiu, H. Song, J. Bernhard, S. Bass and U. Heinz, Comput. Phys. Commun. [**199**]{}, 61 (2016). P. Huovinen and P. Petreczky, Nucl. Phys. [**A837**]{}, 26 (2010); C. Shen, U. Heinz, P. Huovinen and H. Song, Phys. Rev. C [**82**]{}, 054904 (2010).
A. Adil, H. J. Drescher, A. Dumitru, A. Hayashigaki and Y. Nara, Phys. Rev. C [**74**]{} 044905 (2006); H. J. Drescher and Y. Nara, [*ibid.*]{} [**76**]{} 041903 (2007). T. Hirano and Y. Nara, Phys. Rev. C [**79**]{} 064904 (2009); and Nucl. Phys. [**A830**]{} 191c (2009). H. j. Xu, Z. Li and H. Song, Phys. Rev. C [**93**]{}, no. 6, 064905 (2016). C. Shen, U. Heinz, P. Huovinen and H. Song, Phys. Rev. C [**84**]{}, 044903 (2011). H. Song, S. Bass and U. W. Heinz, Phys. Rev. C [**89**]{}, no. 3, 034919 (2014); X. Zhu, F. Meng, H. Song and Y. X. Liu, Phys. Rev. C [**91**]{}, no. 3, 034904 (2015).
F. G. Gardim, F. Grassi, M. Luzum and J. Y. Ollitrault, Phys. Rev. C [**85**]{}, 024908 (2012).
D. Teaney and L. Yan, Phys. Rev. C [**86**]{}, 044908 (2012). L. Yan and J. Y. Ollitrault, Phys. Lett. B [**744**]{}, 82 (2015).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $W$ be a $q$-dimensional irreducible algebraic subvariety in the affine space ${\bf A}^n_\C$, $P_1,...,P_m$ $m$ elements in $\C[X_1,...,X_n]$, and $V(P)$ the set of common zeros of the $P_j$’s in $\C^n$. Assuming that $|W|$ is not included in $V(P)$, one can attach to $P$ a family of non trivial $W$-restricted residual currents in $''{\cal D}^{0,k}(\C^n)$, $1\leq k\leq \min(m,n)$, with support on $|W|$. These currents (constructed following an analytic approach) inherit most of the properties that are fulfilled in the case $q=n$. When the set $|W| \cap V(P)$ is discrete and $m=q$, we prove that for every point $\alpha \in |W|\cap V(P)$ the $W$-restricted analytic residue of a $(q,0)$-form $R d\zeta_I$, $R\in \C[X_1,...,X_n]$, at the point $\alpha$ is the same as the residue on ${\cal W}$ (completion of $W$ in ${\rm Proj}\,\C[X_0,...,X_n]$) at the point $\alpha$ in the sense of Serre ($q=1$) or Kunz-Lipman ($1<q<n$) of the $q$-differential form $(R/P_1\cdots P_q) d\zeta_I$. We will present a restricted version of some affine version of Jacobi’s residue formula and applications of this formula to higher dimensional analogues of Reiss (or Wood) relations, corresponding to situations where the Zariski closures of $|W|$ and $V(P)$ intersect at infinity in an arbitrary way.'
author:
- |
by\
Carlos. A. Berenstein\
[email protected]\
Institute for Systems Research\
University of Maryland, MD 20742, USA\
Alekos Vidras\
[email protected]\
Department of Mathematics and Statistics\
University of Cyprus, Nicosia 1678, Cyprus\
Alain Yger\
[email protected]\
Department of Mathematics,\
University of Bordeaux, 33405 Talence, France.
title: 'Analytic residues along algebraic cycles [^1]'
---
\[section\] \[section\] \[section\] \[section\] \[section\] \[section\] \[section\] \#1[\#1em]{} \#1[[I-.25em\#1]{}]{} \#1[[I-.23em\#1]{}]{} \#1[width.02em height\#1ex depth0ex]{} \#1[width.02em height1.47ex depth\#1ex]{} §[[S]{}]{} ¶[P]{} ${{\rm (}}
\def$[[)]{}]{}
Introduction
============
Let $\Gamma$ be a complete integral curve embedded as a closed subscheme in ${\rm Proj}\, \C[X_0,...,X_n]$ and $\C(\Gamma)$ its function field. Following the exposition of Hübl and Kunz of the Serre’s approach [@huku:gnus], the residue of a meromorphic $(1,0)$-differential form $\omega \in \Omega^1_{\C( \Gamma )/\C}$ at the point $\alpha \in \Gamma$ is defined as follows : let ${\cal M}_1, \dots ,{\cal M}_d$ be the minimal prime ideals of the completion $\widehat {\cal O}_{\Gamma ,\alpha}$ of the local ring of $\Gamma$ at $\alpha$ and let ${\overline R}_j$, $j=1,...,d$, be the integral closures of the “branches” $R_j=\widehat {\cal O}_{\Gamma ,\alpha} /{\cal M}_j$, $j=1,\dots, d$, of the curve $\Gamma $ at the point $\alpha$. Then ${\overline R}_j$ is isomorphic to some algebra of formal power series $\C[[t_j]]$ and in $\C ((t_j))$ the differential $(1,0)$-form $\omega$ can be written as $$\omega=\sum\limits_{k\geq k_{j}} a_k^j\; t_j^k \,,$$ where $a_k^j\in \C$, $k\geq k_{j}$, are complex numbers which are independent of the parameters $t_j$. Define $$\begin{aligned}
{\rm Res}_{\Gamma,\alpha,{\overline R}_j }\, \omega :=a^{j}_{-1}\,,\quad
{\rm Res}_{\Gamma,{\alpha }}\, \omega:=\sum \limits_{j=1}^d a^{j}_{-1}\, . \end{aligned}$$
It was pointed by G. Biernat in [@bier1:gnus] that, if $f_1,...,f_n$ are $n$ germs of holomorphic functions in $n$ variables (with jacobian determinant $J_f\in {\cal O}_n$) such that $(f_1,...,f_{n-1})$ define a germ of curve $\gamma$ (with branches parametrized respectively by $\varphi_1,...,
\varphi_d$) and ${\rm dim}\, [\gamma\cap \{J_f=0\}]=0$, then, for any $h\in {\cal O}_n$, the Grothendieck residue $${\rm Res}_0\, \Big[{h d\zeta_1\wedge \cdots \wedge d\zeta_n
\over f_1\cdots f_n}\Big]:= {1\over (2i\pi)^n} \int\limits_{{|f_1|=\epsilon_1}
\atop {\vdots \atop {|f_n|=\epsilon_n}}}
{h d\zeta_1 \wedge \cdots \wedge d\zeta_n \over f_1\cdots f_n}$$ (with the orientation for the cycle $\{|f_1|=\epsilon_1\,,\dots\,, |f_n|=\epsilon_n\}$ that ensures the positivity of the differential form $d\, {\rm arg}\, f_1 \wedge \cdots \wedge
d\, {\rm arg}\, f_n$ on it) equals $$\sum\limits_{j=1}^d {\rm Res}_{t=0} \, \Bigg[\Big({f_n' h \over f_n J_f}
\circ \varphi_j\Big) (t)\, dt\Bigg]\,;$$ in particular, if $\omega$ denotes the $(1,0)$-meromorphic differential form $$\omega:= {g d\zeta_\alpha \over f_n} \,,\quad g\in {\cal O}_n\,,\ \alpha \in \{1,...,n\}\,,$$ then $${\rm Res}_0 \, \Big[{ {df_1 \wedge \cdots \wedge df_{n-1}\over f_1\cdots f_{n-1}} \wedge \omega}\Big]$$ equals the sum $$\begin{aligned}
\sum\limits_{j=1}^d \nu_j {\rm Res}_{\gamma_j,0}\, [\omega]\,,\end{aligned}$$ where $\gamma_1,...,\gamma_d$ correspond to the irreducible germs of curves attached to the isolated primes in the decomposition of $(f_1,...,f_{n-1})$, and ${\rm Res}_{\gamma_j,0}\, [\omega]$ is defined on the model of the Kunz-Hübl residue, this notion being transposed from the algebraic context to the analytic one. This suggests a natural relation between the approaches developped by Serre-Hübl-Kunz and the analytic residue approach developped by Coleff-Herrera [@coh:gnus] (which precisely allows the transposition of the definition of the Grothendieck residue in the complete intersection case to the setting of currents). The analytic approach we use to define restricted residual currents on a $q$-dimensional reduced analytic space ${\cal Y}\subset U$, where $U$ is an open subset of $\C^n$, will be described in section $2$ as follows : if $f_1,...,f_m$ are $m$ functions holomorphic in $U$, then the map $$\lambda \mapsto \Phi_{Y,f}(\lambda):=\|f\|^{2 \lambda} \, [Y]\,,$$ where $[Y]$ denotes the integration current on $Y=|{\cal Y}|$, can be meromorphically continued as a $'{\cal D}^{(n-q,n-q)}(U)$-map. Moreover, for any $k\in \{1,...,m\}$, for any ordered subset ${\cal I}\subset \{1,...,m\}$ with cardinal $k\leq \min(q,m)$, the analytic continuation of $$\lambda \mapsto \lambda c_k \Phi_{Y,f} (\lambda-k-1) \wedge \overline \partial
\|f\|^2 \wedge \Big( \sum\limits_{l=1}^k (-1)^{l-1} \overline {f_{i_l}}
\bigwedge\limits_{{j=1}\atop {j\not=l}}^k \overline {df_{i_j}}\Big)\,,$$ where $$c_k:={(-1)^{k(k-1)/2} (k-1)! \over (2i\pi)^k}$$ is holomorphic at the origin. Its value at $0$ defines, up to a multiplicative constant, a residual regular holonomic $(n-q,n-q+k)$-current which is supported by $Y\cap V(f)$ ; regular holonomiticity is here understood in the sense of Björk ([@bjo3:gnus], chapter $9$). Properties of such currents are similar to those introduced as above in the case $q=n$. Proposition 2.1 will summarize the different properties of such restricted residual currents. The main case of interest for us will be the case where $m\leq q$ and ${\rm dim} \,(Y\cap V(f)) \leq q-m$, that is $f_1,...,f_m$ define a complete intersection in ${\cal Y}$. In this case, the restricted residue current corresponding to ${\cal I}:=\{1,...,m\}$ is the Coleff-Herrera current on $Y$ $$\Big(\bigwedge \limits_{j=1}^m \overline \partial {1\over f_j}\Big) \wedge [Y]$$ introduced in [@coh:gnus]. It is not surprizing that residual restricted currents in such a complete intersection setting obey the transformation law for multidimensional residue calculus ([@gh:gnus], chapter 6), which we will prove (and use next) in the case $m=q$. If $f_1=P_1,...,f_q=P_q$ are polynomials and $W$ is an affine $q$-dimensional algebraic subvariety of the affine scheme ${\bf A}^n_\C$ such that ${\rm dim}\, (V(P)\cap |W|)=0$, we will prove in section $3$ that the total sum of restricted residues $$\RR
\left[\matrix{ [W]\wedge Q dX_{i_1}\wedge \cdots \wedge dX_{i_q} \cr
\cr P_1,..., P_q}\right]$$ vanishes as soon as the degree of $Q$ is sufficiently small, under a properness assumption on the restriction of $(P_1,...,P_q)$ to $|W|$. We will thus transpose to the restricted case an Abel-Jacobi formula proved in the case $q=n$ and $W={\bf A}^n_\C$ in [@vy:gnus]. Let again $W$ be a $q$-dimensional irreducible algebraic subvariety in the affine scheme ${\bf A}^n_\C$ and $P_1,...,P_q$, $q$ polynomials in $\C[X_1,...,X_n]$ such that $|W|\cap V(P)$ is a discrete (hence finite) algebraic set in $\C^n$. Let $Q\in \C[X_1,...,X_n]$ and ${\cal I}$ a subset in $\{1,...,n\}$ with cardinal $q$. The meromorphic differential form $$\omega:= {Q d\zeta_{i_1}\wedge \cdots \wedge d\zeta_{i_q} \over P_1\cdots P_q}$$ induces an element in $\Omega^q_{\C({\cal W})/\C}$, where ${\cal W}$ denotes the completion of $W$ in ${\rm Proj}\, \C[X_0,...,X_n]$. We will prove in section 4, thanks to the algebraic residue theorem in [@lip2:gnus] and the properties of restricted residual currents that were pointed out in previous sections, that the residue at a closed point $\alpha$ in $|W|\cap V(P)$ (in the sense of Hübl or Lipman [@lip1:gnus]) of the differential form $\omega$ (viewed as an element in $\Omega^q_{\C({\cal W})/\C}$) equals $${\rm Res}_{W,\alpha}\, [\omega]:=\Big\langle
\Big(\bigwedge \limits_{j=1}^q \overline \partial {1\over P_j}\Big) \wedge [W]\,,\,
\psi Q d\zeta_{i_1}\wedge \cdots \wedge d\zeta_{i_q}\Big\rangle \,,$$ where $\psi$ denotes a test-function with compact support in some arbitrary small neighborhood of $\alpha$, such that $\psi \equiv 1$ near $\alpha$. The result is clear when $\alpha$ is a smooth point of $W$, it will follow from the algebraic residue formula combined with a perturbation argument in the case $\alpha$ is a singular point of $W$. As a consequence of the fact that the analytic and algebraic approaches lead to the same restricted residual objects, we will extend (also in section $4$) with an algebraic formulation to such a restricted context the affine Jacobi’s theorem obtained in the non-restricted case $W={\bf A}^n_\C$ in [@vy:gnus].
Let $W$ be a $q$-dimensional irreducible affine algebraic subvariety in ${\bf A}^n_\C$ $$0<q<n$$ and $P_1,...,P_q$ be $q$ polynomials in $\C[X_1,...,X_n]$ such that there exists strictly positive rational numbers $\delta_1,...,\delta_q$ and two constants $K>0$, $\kappa>0$ such that : $$\begin{aligned}
\zeta \in |W|\,,\ \|\zeta\|\geq K \ \Longrightarrow \ \sum\limits_{j=1}^q {|P_j(\zeta)|\over
\|\zeta\|^{\delta_j}} \geq \kappa\,;\end{aligned}$$ then, for any $Q\in \C[X_1,...,X_n]$ such that ${\rm deg}\, Q <\delta_1+\cdots +\delta_q -q$, for any multi-index $(i_1,...,i_q)$ in $\{1,...,n\}^q$, $$\begin{aligned}
\sum\limits_{\alpha\in |W|\cap V(P)}
{\rm Res}_{W,\alpha}\, \Big[{ Qd\zeta_{i_1}\wedge \cdots \wedge d\zeta_{i_q}
\over P_1\cdots P_q}\Big] =0\,.\end{aligned}$$
We will derive (in sections $4$ and 5) some consequences of this result in the spirit of Cayley-Bacharach’s theorem or Wood’s results [@wood1:gnus]. The key point here (compare to the framework of [@huku:gnus] or [@ku2:gnus]) is that the properness assumption along $|W|$ (1.2) which is satisfied by the polynomial map $P:=(P_1,...,P_q)$ does not imply that the Zariski closures of $|W|$ and $V(P_1,...,P_q)$ in $\P^n(\C)$ have an empty common intersection on the hyperplane at infinity.
Restricted residual currents
============================
We begin this section by recalling some basic facts about currents on analytic manifolds, especially integration currents on analytic sets or Coleff-Herrera currents and their “multiplication” with integration currents ; we will inspire ourselves from [@bjo:gnus], [@by2:gnus], [@by3:gnus] or [@meo:gnus]. We start with basic facts about integration on a $q$-dimensional irreducible analytic subset $Y$ in $U\subset \C^n$ [@le:gnus]. The subset $Y_{\rm reg}$ of regular points of $Y$ is a $q$-dimensional complex manifold. The set of singular points $Y_{\rm sing}$ is an analytic subset of $U$ with complex dimension ${\rm dim}\, Y_{\rm sing}<q$. Therefore for any smooth $(q,q)$ test form $\phi_{(q,q)} \in {\cal D}^{(q,q)}(U)$, one can define the action of the integration current $[Y]$ on $\phi_{(q,q)}$ as $$\begin{aligned}
\langle [Y]\,,\, \phi_{(q,q)}\rangle
&=&\int _{Y}\phi _{(q,q)}(\zeta,\bar \zeta)= \int _{Y_{\rm reg}}\phi_{(q,q)}(\zeta,\bar \zeta)+
\int_{Y_{\rm sing}}\phi _{(q,q)}(\zeta,\bar \zeta)\\
&=& \int_{Y_{\rm reg}}\phi_{(q,q)}(\zeta,\bar \zeta)\, . \end{aligned}$$ For ${\rm Re}\, \lambda >0$ and $f_1,...,f_m$ holomorphic in $U$, one can define the $(q,q)$-current $\|f\|^{2\lambda} \, [Y]$ by $$\langle \|f\|^{2\lambda}\, [Y]\,,\, \phi_{(q,q)}\rangle \, :=
\int_{Y_{\rm reg}} \|f\|^{2\lambda} \, \phi_{(q,q)}\,.$$ It is known ([@bjo:gnus],[@bjo3:gnus]) that this current $[Y]$ is a regular holonomic current, which implies, for each point $z_0$ in $U\cap Y$, the existence of a Bernstein-Sato relation $$\begin{aligned}
{\cal Q}_{z_0} \Big(\lambda, \zeta, \bar \zeta, {\partial \over
\partial \zeta}, {\partial \over \partial \bar \zeta}\Big)\;
\Big[ \|f\|^{2(\lambda+1)} \otimes [Y]\Big] =b_{z_0}(\lambda)\, (\|f\|^{2\lambda} \otimes [Y])\end{aligned}$$ ($b_{z_0}\in \C[X]$) valid in a neighborhood of $z_0$. In fact, this does not follow directly from theorem 3.2.6 in [@bjo:gnus] since $\|f\|^2$ is a real analytic function (and not a holomorphic one). Nevertheless, the existence of Bernstein-Sato relations of the form (2.1) remains valid here since $\|f\|^2$ has the particular form $$\|f(\zeta)\|^2=\sum\limits_{j=1}^m f_j(\zeta) \overline {f_j (\zeta)}$$ and the integration current on $Y=\{g_1=\cdots=g_N=0\}$ admits a Siu decomposition $$[Y] =\sum\limits_{1\leq i_1<\cdots <i_{n-q}\leq N}
\, T_{i_1,...,i_{n-q}} \wedge \bigwedge_{l=1}^{n-q} dg_{i_l}\,,$$ where the $T_{i_1,...,i_{n-q}}$ are $(0,n-q)$ currents which are regular holonomic because of Coleff-Herrera type ([@by3:gnus; @meo:gnus; @bjo:gnus]). One can then proceed in $${\bf U}:= \{(\zeta,\overline \zeta)\,:\, \zeta \in U\}\subset
\C^{2n}$$ with blocks of variables $(\zeta,\overline \zeta)$ and profit from the fact that formally $\partial_\zeta$ and $\partial_{\overline \zeta}$ can be considered as derivations respect to independent sets of variables.
Consider then the function of one complex variable defined by $$\begin{aligned}
\lambda \mapsto \Phi_{Y,f}(\lambda ):=\|f\|^{2\lambda}\,[Y]\, .\end{aligned}$$ This function (which is a $'{\cal D}^{(n-q,n-q)} (U)$-current valued function) is well defined and holomorphic in $\{\lambda\in \C\, ;\, {\rm Re}\, \lambda >0 \}$. Thanks to the Bernstein-Sato relations (2.1), it can be continued to the whole complex plane as a meromorphic function. The poles of this meromorphic extension are among strictly negative rational numbers. Furthermore, there is a true pole at any point $\lambda =-k$, $k\in \N^*$.
In fact, we will need a more precise result, where the construction of the meromorphic continuation of (2.2) is involved. What we need is formulated in the following proposition.
Let $Y$ be an irreducible $q$-dimensional analytic subset of $U \subset \C^n$ and $f_1,...,f_m$ $m$ functions holomorphic in $U$. For any $k\in \{1,...,m\}$ and for any ordered subset ${\cal I}\subset \{1,...,m\}$ with cardinal $k\leq \min (q,m)$, the $'{\cal D}^{(n-q,n-q+k)}$-valued map $$\lambda \mapsto \lambda c_k \|f\|^{2(\lambda-k-1)} \, [Y] \wedge
\overline \partial \|f\|^2 \wedge \Big( \sum\limits_{l=1}^k (-1)^{l-1} \overline
{f_{i_l}} \bigwedge_{{j=1}\atop {j\not=l}}^k \overline {df_{i_j}}\Big)$$ $which is holomorphic in ${\rm Re}\, \lambda > k+1$$ can be continued as a meromorphic map to the whole complex plane, with no pole at $\lambda=0$. Its value at $\lambda=0$ defines a residual $(n-q,n-q+k)$-current which is supported by the analytic set $Y\cap \{f_1=\cdots=f_m=0\}=Y\cap V(f)$ and denoted as $$\begin{aligned}
\varphi \in {\cal D}^{(q,q-k)}
\mapsto
\RR \left[\matrix {[Y]\wedge (\cdot) \cr
f_{i_1},...,f_{i_k} \cr
f_1,....,f_m}\right] \, (\varphi)=
\left[\matrix {[Y]\wedge \varphi \cr
f_{i_1},...,f_{i_k} \cr
f_1,....,f_m}\right]\,. \end{aligned}$$
[**Proof.**]{} Assume that $Y$ is defined (in $U$) by the equations $g_1=\cdots=g_N=0$ and that $\nu$ is the multiplicity of the ideal ${\cal O}_{U,y}$ generated by $g_1,...,g_N$ at a generic point $y\in Y$. Let $d=n-q$. One can conclude from [@meo:gnus] that $[Y]$ coincides with the value at $\mu=0$ of the meromorphic $'{\cal D}^{(d,d)}(U)$-valued map $\Psi_g$ $$\mu
\buildrel {\Psi_g}\over {\mapsto} { \mu (d-1)! \over
(2i\pi)^{d} \, \nu}\,
\|g\|^{2\mu} \,
\overline \partial \log \|g\|^2
\wedge \partial \log \|f\|^2 \wedge
\sum\limits_{{j_1<\cdots <j_{d-1}}\atop
{1\leq j_j\leq N}} \,
\bigwedge_{l=1}^{d-1}
\Big({\overline {\partial g_{j_l}}
\wedge \partial g_{j_l} \over \|g\|^2} \Big)\,.$$ In fact, in the general situation where $(g_1,...,g_N)$ define a $q$-purely dimensional cycle ${\cal Z}$ (non necessarily irreducible) in $U$, the integration current (with multiplicities) on ${\cal Z}$ can be expressed as the value at $\lambda=0$ of some meromorphic $'{\cal D}^{(d,d)}(U)$-valued function which can be made explicit in terms of $g_1,...,g_N$ (see theorem 3.1 in [@by3:gnus] for a proof in the algebraic case). Let ${\cal I}\subset \{1,...,m\}$ with cardinal $k\leq \min (q,m)$ and, for ${\rm Re}\, \lambda > k+1$, $$\Theta_{f,{\cal I}} (\lambda):=
\lambda \|f\|^{2(\lambda-k-1)} \,
\overline \partial \|f\|^2 \wedge \Big( \sum\limits_{l=1}^k (-1)^{l-1} \overline
{f_{i_l}} \bigwedge_{{j=1}\atop {j\not=l}}^k \overline {df_{i_j}}\Big)\,.$$
In order to prove the proposition, we can localise the problem and assume that the origin belongs to $Y \cap V(f)$. As in our previous work (see for example [@by2:gnus], pages 32-33, or [@by3:gnus], page 208) we construct an analytic $n$-dimensional manifold ${\cal X}$, a neighborhood $V$ of $0$ in $U$, a proper map $\pi: {\cal X}\rightarrow V$ which realizes a local isomorphism between $V\setminus \{f_1\cdots f_m\, g_1 \cdots g_N=0\}$ and ${\cal X}\setminus \pi^{-1}(\{f_1\cdots f_m \, g_1\cdots g_N=0\})$, such that in local coordinates on ${\cal X}$ (centered at a point $x$), one has, in the corresponding local chart ${\cal U}_x$ around $x$, $$\begin{aligned}
f_j\circ \pi (t)&=& u_j(t)\, t_1^{\alpha_{j1}}\cdots t_n^{\alpha_{jn}}=
u_j(t)\, t^{\alpha_j},
\ j=1,\dots,m \\
g_k \circ \pi (t) &=& v_{k}(t)\, t_j^{\beta_{k1}} \cdots t_n^{\beta_{kn}}=
v_k(t)\, t^{\beta_k},\ k=1,...,N \end{aligned}$$ where the $u_j$, $j=1,...,m$ and the $v_k$, $k=1,...,N$, are non vanishing holomorphic functions in ${\cal U}_x$, at least one of the monomials $t^{\alpha_j}$, $j=1,...,m$ divides all of them (we will denote this monomial as $t^\alpha$), and at least one of the monomials $t^{\beta_k}$, $k=1,...,N$ divides all of them (we will denote this monomial as $t^\beta$).
When $\varphi$ is a $(q,q-k)$-test form with support in $V$, one has, for ${\rm Re}\, \lambda >>0$, $$\int_{V\,\cap \, Y}
\Theta_{f,{\cal I}} (\lambda) \wedge \varphi = \Bigg[ \int_{V}
\Psi_g(\mu) \wedge
\Theta_{f,{\cal I}} (\lambda)
\wedge \varphi \Bigg]_{\mu=0}$$ (the right hand side being continued as a meromorphic function of $\mu$ which has no pole at $\mu=0$). For $\lambda$ fixed with ${\rm Re}\, \lambda >>0$, one can rewrite for ${\rm Re}\, \mu >>0$ the integral $$\int_{V}
\Psi_g(\mu) \wedge
\Theta_{f,{\cal I}} (\lambda)
\wedge \varphi$$ as a sum of integrals of the form $$\begin{aligned}
\int_{{\cal U}_x} \pi^*[\Psi_g] (\mu) \wedge
\pi^* [\Theta_{f,{\cal I}} (\lambda)] \wedge \rho \pi^*(\varphi)\,,\end{aligned}$$ where $\rho$ is a test-function in ${\cal U}_x$ which corresponds to a partition of unity for $\pi^*({\rm Supp}\, \varphi)$. We know from lemma 1.1 and lemma 1.2 in [@by3:gnus] that $$\Big[\pi^* [\Psi_g(\mu)]\Big]_{\mu=0}= \Big[ \Psi_{g\circ \pi} (\mu)\Big]_{\mu=0}$$ is a positive $\partial$ and $\overline \partial$-closed current $\theta_{{\cal U}_x}$ in ${\cal U}_x$, which implies that, as soon as ${\rm Re}\, \lambda >>0$, $$\Bigg[\int_{{\cal U}_x} \pi^*[\Psi_g] (\mu) \wedge
\pi^* [\Theta_{f,{\cal I}} (\lambda)] \wedge \rho \pi^*(\varphi)\Bigg]_{\mu=0}
= \int_{{\cal U}_x}
\theta_{{\cal U}_x} \wedge \pi^* [\Theta_{f,{\cal I}} (\lambda)] \wedge \rho \pi^*(\varphi)\,.$$ On the other hand, in ${\cal U}_x$ and for ${\rm Re}\, \lambda >>0$, a straightforward computation leads to $$\pi^* \Big[\Theta_{f,{\cal I}}\Big] (\lambda)
=\lambda {a^{2\lambda} |t^{\alpha}|^{2\lambda}
\over t^{k \alpha}} \Big( \vartheta + \varpi \wedge
{\overline{d t^{\alpha}} \over \overline {t^\alpha}}\Big)\,,$$ where $\vartheta$ and $\varpi$ are smooth differential forms in ${\cal U}_x$ (with respective types $(0,k)$ and $(0,k-1)$) and $a$ is a strictly positive real analytic function in ${\cal U}_x$. It follows from Stokes’s theorem that $$\begin{aligned}
\int_{{\cal U}_x} \pi^* [\Theta_{f,{\cal I}}] (\lambda)
\wedge \theta_{{\cal U}_x} \wedge
\rho \pi^*(\varphi) =
\int_{{\cal U}_x}
{|t^{\alpha}|^{2\lambda} \over
t^ {k \alpha}} \, \theta_{{\cal U}_x} \wedge \xi_\varphi(\rho\,;\, t,\lambda)\,,\end{aligned}$$ where $(t,\lambda) \mapsto \xi_{\varphi} (\rho\,;\, t,\lambda)$ is a $(n-q,n-q)$-differential form with smooth coefficients (in $t$) depending holomorphically in $\lambda$.
One can see also that, for ${\rm Re}\, \mu >> 0$, $$\pi^* [\Psi_g] (\mu)=
\mu \, b^{2\mu} \, |t^{\beta}|^{2\mu}
\, \Big( {\overline {d t^\beta} \over \overline {t^\beta}} + \eta_{(0,1)} \Big)
\wedge \Big({{d t^\beta} \over {t^\beta}} + \eta_{(1,0)} \Big) \wedge \upsilon \,,$$ where $b$ is a strictly positive real analytic function in ${\cal U}_x$, $\eta_{(0,1)},\, \eta_{(1,0)},\, \upsilon$ are smooth differential forms in ${\cal U}_x$ with respective types $(0,1)\,,\, (1,0)$ and $(d-1,d-1)$. This implies that, if $t_{i_1},...,t_{i_s}$ are the coordinates that appear in $t^\beta$, $$\theta_{{\cal U}_x} = \sum\limits_{l=1}^s \, [t_{i_l}=0]\, \wedge \omega_{i_l} \,,$$ where $\omega_{i_l}$ is a smooth $(d-1,d-1)$-form in ${\cal U}_x$ and $[t_i=0]$ denotes the integration current (without multiplicities) on $\{t_i=0\}$. Therefore, for ${\rm Re}\, \lambda >> 0$, $$\int_{{\cal U}_x} \pi^* [\Theta_{f,{\cal I}}] (\lambda)
\wedge \theta_{{\cal U}_x} \wedge
\rho \pi^*(\varphi)=
\sum\limits_{{l=1}\atop
{(t_{i_l}, t^\alpha)=1}}^s
\int_{\{t_{i_l}=0\}\, \cap\, {\cal U}_x}
{|t^\alpha|^{2\lambda}
\over t^{k \alpha}} \omega_{i_l} (t) \wedge
\xi_\varphi(\rho\,;\, t,\lambda).$$ Such a function of $\lambda$ can be continued to a meromorphic function in the whole complex plane, with no pole at $\lambda=0$ (using Stokes’s theorem). The assertion of the proposition follows, since for ${\rm Re}\, \lambda >>0$, $$\int_{V}
\Psi_g(\mu) \wedge
\Theta_{f,{\cal I}} (\lambda)
\wedge \varphi$$ is a sum of integrals of the form (2.4). $\diamondsuit $ Keeping the notation from above one has the following corollary.
Under the conditions of proposition $2.1$, the residual current defined by $(2.3)$ has the following properties\
$1)$ For any $h\in H(U)$ such that $$\begin{aligned}
\forall K \subset \subset U\cap Y\,, \exists C_K>0\,,\ |h|\leq C_K \|f\|\ {on}\ K\,, \end{aligned}$$ one has $$\RR \left[\matrix {h^k [Y]\wedge (\cdot ) \cr
f_{i_1},...,f_{i_k} \cr
f_1,....,f_m}\right]\equiv 0$$\
$2)$ If $h\in H(U)$ and $$\begin{aligned}
h(z)=0,\; \forall z\in Y\cap V(f)\,, \end{aligned}$$ then one has $$\RR \left[\matrix {{\overline h }[Y]\wedge (\cdot ) \cr
f_{i_1},...,f_{i_k} \cr
f_1,....,f_m}\right]\equiv 0$$
[**Proof.**]{} Let us now suppose that $h$ satisfies (2.6). If we do not perform integration by parts as in (2.5), we have, for ${\rm Re}\, \lambda >>0$, $$\begin{aligned}
&&\int_{{\cal U}_x} \pi^* [\Theta_{f,{\cal I}}] (\lambda)
\wedge \theta_{{\cal U}_x} \wedge
\rho \pi^*( h^k \varphi) \\
&&=
\lambda
\sum\limits_{{l=1}\atop
{(t_{i_l}, t^\alpha)=1}}^s
\int_{\{t_{i_l}=0\}\, \cap\, {\cal U}_x}
{a^{2\lambda} |t^{\alpha}|^{2\lambda}
\over t^{k \alpha}} \Big( \vartheta + \varpi \wedge
{\overline{d t^{\alpha}} \over \overline {t^\alpha}}\Big)
\wedge \omega_{i_l} (t) \wedge \rho \pi^*(h^k \varphi)\,. \end{aligned}$$ Condition (2.6) implies that there exists some positive constant $\kappa$ such that, for any $l=1,...,s$ with $t_{i_l}$ coprime with $t^\alpha$, $$|\pi^* h(t_1,...,\buildrel {i_l}\over{0},...,t_n)|\leq \kappa |t^\alpha|\,,\quad
t\in {\rm Supp}\, \rho\,,$$ which implies that $t^{k\alpha}$ divides $(\pi^* h^k)_{\{|t_{i_l}=0\}}$ on the support of $\rho$. This implies that for such $h$, $$\Bigg[\int_{{\cal U}_x} \pi^* [\Theta_{f,{\cal I}}] (\lambda)
\wedge \theta_{{\cal U}_x} \wedge
\rho \pi^*( h^k \varphi)\Bigg]_{\lambda=0}=0\,,$$ which gives the first assertion of the corollary since $$\int_{V}
\Psi_g(\mu) \wedge
\Theta_{f,{\cal I}} (\lambda)
\wedge \varphi$$ as a sum of integrals of the form (2.4).\
If $h$ vanishes on $Y\cap V(f)$, then, for any $l=1,...,s$ such that $t_{i_l}$ is coprime with $t^{\alpha}$, any coordinate which divides $t^\alpha$ also divides $(\pi^* h)_{|\{t_{i_l}=0\}}$ on the support of $\rho$. This implies that any expression of the form $$\int_{\{t_{i_l}=0\} \cap {\cal U}_x}
{a^{2\lambda} |t^{\alpha}|^{2\lambda}
\over t^{k \alpha}} \Big( \vartheta + \varpi \wedge
{\overline{d t^{\alpha}} \over \overline {t^\alpha}}\Big)
\wedge \omega_{i_l} (t) \wedge \rho \pi^*(\overline h \varphi)$$ has in fact no antiholomorphic singularity (therefore has a meromorphic extension which is polefree at the origin). It follows that for such $h$, one has again $$\Bigg[\int_{{\cal U}_x} \pi^* [\Theta_{f,{\cal I}}] (\lambda)
\wedge \theta_{{\cal U}_x} \wedge
\rho \pi^*( \overline h \varphi)\Bigg]_{\lambda=0}=0\,,$$ which proves the remaining assertion of the corollary since again $$\int_{V}
\Psi_g(\mu) \wedge
\Theta_{f,{\cal I}} (\lambda)
\wedge \varphi$$ is a sum of integrals of the form (2.4). $\quad \diamondsuit$ When $k=m\leq q$, we will use the simplified notation $$\RR \left[\matrix {[Y]\wedge (\cdot) \cr
f_1,...,f_m }\right] \, (\varphi):=
\RR \left[\matrix {[Y]\wedge (\cdot) \cr
f_1,...,f_m\cr
f_1,....,f_m}\right]\,.$$ The transformation law for residual currents can be transposed to the case of restricted residual currents. Since we deal in this paper with restricted residual currents supported by discrete sets, we state the transformation law in this particular setting. One has the following proposition :
Let $Y$ be an irreducible $q$-dimensional analytic subset of $U \subset \C^n$ and $f_1,...,f_q,\, g_1,...,\,g_q$, $2q$ functions holomorphic in $U$ such that $Y\cap V(f)$ and $Y\cap V(g)$ are discrete analytic sets. Assume that there exist $q^2$ holomorphic functions in $U$, $a_{kl}$, $1\leq k,l\leq q$, such that $$g_k(\zeta)=\sum\limits_{l=1}^q a_{kl} (\zeta) \, f_l(\zeta)\,
,\ k=1,...,q\,, \quad \zeta \in Y$$ Then, one has the following equality between restricted residual currents : $$\begin{aligned}
\RR \left[\matrix {[Y]\wedge (\cdot) \cr
f_1,....,f_q}\right] =
\RR \left[\matrix { \Delta \, [Y]\wedge (\cdot)\cr
g_1,....,g_q}\right]\,, \end{aligned}$$ where $\Delta := \det [a_{kl}]_{1\leq k,l\leq q}$.
[**Proof.**]{} In order to prove this equality, we just need to prove it when $U$ is a neighborhood $V$ of a point $\alpha\in Y\cap (V(f)\cup V(g))$ such that $\alpha$ is the only point of $Y\cap(V(f)\cup V(g))$ which lies in this neighborhhood. Thanks to the first assertion in Corollary 2.1, it is enough to test the two currents involved in (2.8) on test forms in ${\cal D}^{(q,0)}(V)$ whose coefficients are holomorphic in a neighborhood of $\alpha$. Let $\varphi$ be such a test form. Since $$\begin{aligned}
&&\overline \partial
\Bigg[\|f\|^{2(\lambda-q)} \, [Y] \wedge
\Big(\sum\limits_{j=1}^q (-1)^{j-1} \overline f_j \bigwedge\limits_{{l=1}\atop {l\not=j}}^q
\overline {df_j}\Big)\Bigg]\\
&&\qquad=\lambda \|f\|^{2(\lambda-q)} \, [Y]\, \wedge \bigwedge_{j=1}^q
\overline {df_j} \\
&&\qquad= \lambda \|f\|^{2(\lambda-q-1)}\, [Y]
\wedge \overline\partial \|f\|^2
\wedge
\Big(\sum\limits_{j=1}^q (-1)^{j-1} \overline f_j \bigwedge\limits_{{l=1}\atop {l\not=j}}^q
\overline {df_l}\Big)\end{aligned}$$ for ${\rm Re}\, \lambda >>0$ and $$\sum\limits_{j=1}^q s_j (\zeta) f_j(\zeta)=1\,,\quad \forall\,
\zeta \in (V\cap Y_{\rm reg})\setminus \{\alpha\},$$ where $$s_j:= {\overline f_j \over \|f\|^2}\,,\quad j=1,...,q,$$ one has, by Stokes’s theorem, that $$\begin{aligned}
\RR \left[\matrix {[Y]\wedge \varphi \cr
f_1,....,f_q}\right]&=&
(-1)^q \omega_q
\int_{Y_{\rm reg}} {\sum\limits_{j=1}^q (-1)^{j-1} \overline f_j \bigwedge\limits_{{l=1}\atop {l\not=j}}^q
\overline {df_j}\over \|f\|^{2q}} \wedge \overline\partial \varphi \nonumber \\
&=& (-1)^q \omega_q
\int_{Y_{\rm reg}} \Big(\sum\limits_{j=1}^q (-1)^{j-1} s_j\bigwedge\limits_{{l=1}\atop {l\not=j}}^q
{ds_l}\Big)\wedge \overline\partial \varphi \, .\end{aligned}$$ Similarly, if we introduce $$t_j:={\overline g_j \over \|g\|^2}\,,\quad j=1,...,q\,,$$ and $$\widetilde {s_j}:=\sum\limits_{l=1}^q a_{lj} t_l\,,\quad j=1,...,q\,,$$ one has also $$\sum\limits_{j=1}^q \widetilde {s_j} (\zeta) f_j(\zeta)=1\,,\quad \forall\,
\zeta\in (V \cap Y_{\rm reg})\setminus \{\alpha\}\, .$$ Let, for $\xi\in [0,1]$ and $j=1,...,q$, $$s_j^{(\xi)} = (1-\xi)\, s_j + \xi\, \widetilde {s_j}\, .$$ Note that we have $$\sum\limits_{j=1}^q s_j^{(\xi)} (\zeta) f_j(\zeta) =1\,,\quad
\forall \xi \in [0,1]\,,\quad
\forall \zeta
\in (V \cap Y_{\rm reg})\setminus \{\alpha\}\, .$$ Therefore, one has, since $$\bigwedge\limits_{j=1}^q \overline\partial_\zeta s_j^{(\xi)}\equiv 0$$ on $(V \cap Y_{\rm reg})\setminus \{\alpha\}$, $${d\over d\xi} \Bigg[
\int_{W_{\rm reg}} \Big(\sum\limits_{j=1}^q (-1)^{j-1} s_j^{(\xi)}\bigwedge\limits_{{l=1}\atop {l\not=j}}^q
{ds_l^{(\xi)}}\Big)\wedge \overline\partial \varphi \Bigg] \equiv 0$$ on $[0,1]$. It follows from (2.9) that $$\begin{aligned}
\RR \left[\matrix {[Y]\wedge \varphi \cr
f_1,....,f_q}\right] &=&
(-1)^q \omega_q \int_{Y_{\rm reg}} \Big(\sum\limits_{j=1}^q (-1)^{j-1} \widetilde {s_j}\bigwedge\limits_{{l=1}\atop {l\not=j}}^q
{\overline \partial \widetilde {s_l}}\Big)\wedge \overline\partial \varphi \\
&=& (-1)^q \omega_q
\int_{Y_{\rm reg}} \Delta\, {\sum\limits_{j=1}^q (-1)^{j-1} \overline g_j \bigwedge\limits_{{l=1}\atop {l\not=j}}^q
\overline {dg_j}\over \|g\|^{2q}} \wedge \overline\partial \varphi \\
&=& \RR \left[\matrix {\Delta\, [Y]\wedge \varphi \cr
g_1,....,g_q}\right] \, . \end{aligned}$$ this concludes the proof of the proposition. $\quad\diamondsuit$ As a consequence of this result, we will state in the algebraic context the following analog of the global transformation law. We need first some piece of notation. Assume that $W$ is a $q$-dimensional irreducible algebraic subvariety in the affine space ${\bf A}^n_\C$ (the integration current on $|W|$ without multiplicities taken into account being denoted as $[W]$) and that $P_1,...,P_q$ are $q$ elements in $\C[X_1,...,X_n]$ such that $|W|\cap V(P_1,...,P_q)$ is a discrete (hence finite) algebraic subset of $\C^n$. For any $Q\in \C[X_1,...,X_n]$, any ordered subset $\{i_1,...,i_q\}$ of $\{1,...,n\}$, we will denote as $$\RR \left[\matrix{ [W]\wedge Q \bigwedge\limits_{l=1}^q dX_{i_l} \cr \cr
P_1,...,P_q}\right]$$ the result of the action of the $W$-restricted current $$\varphi \mapsto \RR \left[\matrix{ [W]\wedge \varphi \cr
P_1,...,P_q}\right]$$ on the $(q,0)$-test form $Q(\zeta) \psi(\zeta) \bigwedge_{l=1}^q d\zeta_{i_l}$, where $\psi$ is any test-function in ${\cal D} (\C^n)$ which equals $1$ in a neighborhood of $|W|\cap V(P)$. If $$\Gamma = \sum\limits_{j=1}^M \nu_j W_j$$ (where $W_1,...,W_M$ are $M$ irreducible algebraic subsets in $\C^n$ and $\nu_j\in \N^*$, $j=1,...,M$) is an effective $q$-dimensional algebraic cycle in the affine space $\C^n$ and $P_1,...,P_q$ are $q$ polynomials such that $W_j\cap V(P)$ is discrete for any $j=1,...,M$, we will also denote as $$\RR \left[\matrix{ [\Gamma]\wedge Q \bigwedge\limits_{l=1}^q dX_{i_l} \cr\cr
P_1,...,P_q}\right]$$ the weighted sum $$\sum\limits_{j=1}^M \nu_j
\RR \left[\matrix{ [W_j]\wedge Q \bigwedge\limits_{l=1}^q dX_{i_l} \cr \cr
P_1,...,P_q}\right]$$
Let $\Gamma$ be an effective $q$-dimensional algebraic cycle in the affine space $\C^n$ and $P_1,...,P_q, R_1,...,R_q$ be $2q$ polynomials such that ${\rm Supp}\, \Gamma \cap
V(P_1,...,P_q)$ and ${\rm Supp}\, \Gamma \cap V(R_1,...,R_q)$ are discrete (hence finite) algebraic subsets of $\C^n$. Assume that there is a $(q,q)$-matrix of polynomials $[A_{k,l}]_{1\leq k,l\leq q}$ such that $$R_k =\sum\limits_{l=1}^q A_{kl} P_l \quad on\quad {\rm Supp}\,
\Gamma\,\quad k=1,..,q\,.$$ Then, for any $Q\in \C[X_1,...,X_n]$, any ordered subset $\{i_1,...,i_q\}$ of $\{1,...,n\}$, one has $$\begin{aligned}
\RR \left[\matrix{ [\Gamma]\wedge Q \bigwedge\limits_{l=1}^q dX_{i_l} \cr \cr
P_1,...,P_q}\right] =
\RR \left[\matrix{ [\Gamma]\wedge \Delta \, Q \bigwedge\limits_{l=1}^q dX_{i_l} \cr\cr
R_1,...,R_q}\right]\,,\end{aligned}$$ where $\Delta$ denotes the determinant of the matrix $[A_{k,l}]_{1\leq k,l\leq q}$.
Another key point about the restricted residual current in the discrete context is the following annihilating property :
Let $Y$ be an irreducible $q$-dimensional analytic subset of $U \subset \C^n$ and $f_1,...,f_q$ be $q$ functions holomorphic in $U$ such that $Y\cap V(f)$ is a discrete analytic set. Then one has, for $k=1,...,q$, $$\begin{aligned}
{\rm Res}\, \left[\matrix {f_k[Y]\wedge (\cdot) \cr
f_1,....,f_q }\right] = 0\end{aligned}$$
[**Proof.**]{} We give here a self-contained proof of the above proposition. Actually, because of the properties quoted in Corollary 2.1, it is enough to show that if $\alpha\in V(P)\cap Y$ and $\varphi$ is a test-function with support arbitrarly small about $\alpha$ with $\varphi=1$ in some neighborhhood $v_\alpha$ of $\alpha$, then, for any function $h \in C^\infty (U)$ which is holomorphic on $v_\alpha$, for any ordered subset ${\cal I}=\{i_1,...,i_q\}\subset \{1,...,n\}$, one has, for $j=1,...,q$, $${\rm Res}\, \left[\matrix {f_j[Y]\wedge h \varphi \, d\zeta_{\cal I} \cr
f_1,....,f_q}\right]=0\,.$$ One can use Stokes’s formula (as in the proof of proposition 2.2) and write $${\rm Res}\, \left[\matrix {f_k[Y]\wedge h \varphi \, d\zeta_{\cal I} \cr
f_1,....,f_q}\right]=
(-1)^q \omega_q \int_{Y_{\rm reg}} h f_k
\Big( \sum\limits_{j=1}^q (-1)^{j-1} s_j \bigwedge\limits_{{l=1}\atop {l\not=j}}^q ds_l \Big)
\wedge \overline \partial \varphi \wedge d\zeta_{\cal I}
\,,$$ where $s_j:=\overline {f_j}/\|f\|^2$, $j=1,...,q$. One can see at once that $$\begin{aligned}
&&f_k
\Big( \sum\limits_{j=1}^q (-1)^{j-1} s_j \bigwedge\limits_{{l=1}\atop {l\not=j}}^q ds_l \Big)
\wedge d\zeta_{\cal I} \wedge [Y]=
\Big(\bigwedge_{{l=1}\atop {l\not=k}}^q ds_l \Big) \wedge
\overline \partial \varphi \wedge d\zeta_{{\cal I}} \wedge [Y]\\
&&\qquad\qquad= \pm d\Bigg[ s_{k'}
\Big(\bigwedge_{{l=1}\atop {l\not=k,k'}}^q ds_l \Big)
\wedge
\overline \partial \varphi \wedge d\zeta_{{\cal I}} \wedge [Y]\Bigg] \end{aligned}$$ for $k'\not=k$, since $s_1 f_1+\cdots +s_q f_q\equiv 1$ on $Y\cap {\rm Supp}\ \overline\partial\varphi$, which shows that $$\int_{Y_{\rm reg}} h f_k
\Big( \sum\limits_{j=1}^q (-1)^{j-1} s_j \bigwedge\limits_{{l=1}\atop {l\not=j}}^q ds_l \Big)
\wedge \overline \partial \varphi \wedge d\zeta_{\cal I} =0$$ as a consequence of Stokes’s formula on $Y$. $\quad \diamondsuit$ We remark here that there is an alternative proof of the last proposition. In fact, when $m\leq q$ and $f_1,...,f_m$ define a complete intersection on $Y$, one can show that the restricted residual current $${\rm Res}\, \left[\matrix{ [Y] \wedge (\cdot) \cr f_1,...,f_m}\right]$$ coincides with the Coleff-Herrera current $\Big(\bigwedge_{j=1}^m \overline\partial (1/f_j)\Big)
\wedge [Y]$ as it is defined in [@coh:gnus]. The proof of this claim can be carried out as it is done in the non restricted case in [@pty:gnus], section 4. Since the proof of this fact is rather tedious, we will not give it here. A consequence of this result is that, when $f_1,...,f_m$ ($m\leq q$) define a complete intersection on $Y$, one has, for $k=1,...,m$, $${\rm Res}\, \left[\matrix {f_k[Y]\wedge (\cdot) \cr
f_1,....,f_m}\right] = f_k \Big(\bigwedge_{j=1}^m \overline\partial {1\over f_j} \Big)\wedge [Y]=0$$ (see [@coh:gnus]). This implies the proposition when $m=q$.
An Abel-Jacobi formula in the restricted case (analytic approach)
=================================================================
One of the key facts about restricted residual currents (as defined through the analytic approach described in section $2$) is that they satisfy (in the $0$-dimensional complete intersection setting) Abel-Jacobi’s formula, exactly as in the non-restricted case (see [@vy:gnus]). Such a result will be, together with the validity of the transformation law in the restricted context) a crucial fact in order to compare our analytic approach and the algebraic one.
Let $W$ be a $q$-dimensional irreducible affine algebraic subvariety of the affine scheme ${\bf A}^n_\C$ $$0<q<n$$ and $P_1,...,P_q$ be $q$ polynomials in $\C[X_1,...,X_n]$ such that there exists strictly positive rational numbers $\delta_1,...,\delta_q$ and two constants $K>0$, $\kappa>0$ with : $$\begin{aligned}
\zeta \in |W|\,,\ \|\zeta\|\geq K \ \Longrightarrow \ \sum\limits_{j=1}^q {|P_j(\zeta)|\over
\|\zeta\|^{\delta_j}} \geq \kappa\,.\end{aligned}$$ Then, for any $Q\in \C[X_1,...,X_n]$ such that ${\rm deg}\, Q <\delta_1+\cdots +\delta_q -q$, for any multi-index $(i_1,...,i_q)$ in $\{1,...,n\}^q$, $$\begin{aligned}
\RR \left[\matrix{ [W]\wedge Q \bigwedge\limits_{l=1}^q dX_{i_l}\cr \cr
P_1,...,P_q}\right] =0\, .\end{aligned}$$
Before we give the proof of this result, let us state an important corollary :
Let $W$ be a $q$-dimensional irreducible algebraic subvariety in the affine scheme ${\bf A}^n_\C$ and ${\cal W}$ be its completion in ${\rm Proj}\, \C[X_0,...,X_n]$. Let $P_1,...,P_q$ be $q$ elements in $\C[X_1,...,X_n]$, with respective degrees $D_1,...,D_q$, such that $$\begin{aligned}
&&|{\cal W}| \cap \Big\{[\zeta_0\,:\, ...\,:\, \zeta_n]\in \P^n(\C)\,;\,
{}^h P_j (\zeta_0,...,\zeta_n)=0,\; j=1,...,q \Big\} \subset \C^n \,,\nonumber \\
&& \end{aligned}$$ where ${}^h P_j$, $j=1,...,q$, denotes the homogeneization of the polynomial $P_j$. Then, for any $Q\in \C[X_1,...,X_n]$ such that ${\rm deg}\, Q <D_1+\cdots +D_q -q$, for any multi-index $(i_1,...,i_q)$ in $\{1,...,n\}^q$, $$\begin{aligned}
\RR \left[\matrix{ [W]\wedge Q \bigwedge\limits_{l=1}^q dX_{i_l}\cr \cr
P_1,...,P_q}\right] =0\end{aligned}$$
[**Proof of corollary 3.1.**]{} Assume that $$|{\cal W}|=\Big\{[\zeta_0\,:\, ...\,:\, \zeta_n]\in \P^n(\C)\,;\,
{}^h {\cal G}_j (\zeta_0,...,\zeta_n)=0,\; j=1,...,N \Big\}\,,$$ where ${\cal G}_1,...,{\cal G}_N$ are homogeneous polynomials in $\widetilde \zeta= (\zeta_0,...,\zeta_n)$. Condition (3.3) implies that $$|{\cal W}| \cap \Big\{[\zeta_0\,:\, ...\,:\, \zeta_n]\in \P^n(\C)\,;\,
{}^h P_j (\zeta_0,...,\zeta_n)=0,\; j=1,...,q \Big\}$$ is a finite set in $\C^n$ ; this implies (through a compacity argument) that there exists $K, \kappa>0$ such that, for any $(\zeta_0,...,\zeta_n) \in \C^{n+1}\setminus \{(0,...,0)\}$ such that $$\Big(|\zeta_1|^2 +\cdots + |\zeta_n|^2\Big)^{1/2} \geq K |\zeta_0|\,,$$ one has $$\sum\limits_{j=1}^q {|{}^h P_j(\widetilde \zeta)|\over
\|\widetilde \zeta \|^{D_j}}
+ \sum\limits_{l=1}^M {|{\cal G}_l(\widetilde \zeta)|\over
\|\widetilde \zeta\|^{{\rm deg}\, {\cal G}_l} } \geq \kappa \,.$$ Condition (3.1) with $\delta_j=D_j$, $j=1,...,q$, holds if we restrict to the affine space $\C^n$. The statement (3.4) follows then from (3.2). $\quad \diamondsuit$
We remark that a proposition similar to proposition 3.1 was proved in the non restricted case ($W={\bf A}^n_\C$) in [@vy:gnus]. Unfortunately, the proof which is given there (and depends heavily on resolution of singularities on the analytic manifold $\P^n(\C)$) cannot immediately be transposed to the restricted case (since the Zariski closure $|{\cal W}|$ of $|W|$ in $\P^n(\C)$ is not a smooth manifold anymore). Instead, we will follow an alternative approach (applicable also for the case $q=n$), based on an argument in the affine space (and not in its compactification $\P^n(\C)$), which was proposed by Ha[ï]{} Zhang in [@zhang:gnus]. Our task has been to adapt this argument to the restricted case. Note that, if $z=Aw$ is a linear change of variables in $\C^n$, one has, for any element in ${\cal D}^{(q,0)}(\C^n)$ $$\RR
\left[\matrix { [W]\wedge \varphi \cr f_1,...,f_q}\right]=
\RR
\left[ \matrix{
[A^{-1} (W)] \wedge A^* \varphi \cr f_1\circ A,...,f_q\circ A}\right]\, .$$ Therefore, we do not loose generality is we assume that ${\cal I}=\{1,...,q\}$ and that the projection $$\Pi~: (\zeta_1,...,\zeta_n) \mapsto (\zeta_1,...,\zeta_q)$$ is a proper map from $|W|$ to $\C^q$ (coordinates can be choosen in such a way that Noether normalization theorem applies respect to any $(q,n-q)$ splitting $\zeta=(\zeta',\zeta'')$ of the set of variables $(\zeta_1,...,\zeta_n)$, see for example [@fl:gnus; @rud:gnus]). For $\delta _i$, $i=1,\dots , q $ which appear in the statement of Proposition 3.1 we choose a positive integer $N$ large enough so that $$\begin{aligned}
N\prod\limits _{{l=1}\atop {l\not=j}}^q\delta_l >2, \quad
j=1, \dots , q \, . \end{aligned}$$ Then, let $$\delta^{[j]}:=
N\prod\limits _{{l=1}\atop {l\not=j}}^q \delta_l\,,\quad j=1,...,q\,,$$ and $$\delta:= N \delta_1 \cdots \delta_q=\delta_j \delta^{[j]}\,,\quad j=1,...,q\, .$$ Similarly, for the polynomials $P_1, \dots, P_q $, one can define, in the affine open set $\C^n \setminus \{P_1\cdots P_q=0\}$, the $C^\infty$ functions $$\widetilde s_j:= {|P_j|^{\delta^{[j]}}
\over P_j\, \sum\limits_{l=1}^q |P_l|^{\delta^{[l]}}}\,,\quad j=1,...,q\, .$$ These functions $\widetilde s_j$, $j=1,...,q$, extend (provided $N >> 1$) to $C^1$ functions in $\C^n \setminus V(P)$, satisfying $$\sum\limits_{j=1}^q \widetilde s_j (\zeta) P_j(\zeta)=1\,,\quad
\zeta\in \C^n \setminus V(P)\, .$$ Let finally $$u_j:=|P_j|^{\delta^{[j]}/2}\,,\quad j=1,...,q$$ and $$S:=\sum\limits_{j=1}^q u_j^2 =\| u\|^2\,.$$ At this point we return to the\
\
[**Proof of proposition 3.1.**]{} One can suppose without any loss of generality that $\{i_1,...,i_q\}=
\{1,...,q\}$ and that the projection $\Pi$ is a proper map from $|W|$ to $\C^q$. Condition (3.1) implies the existence of a strictly positive constant $\kappa_N$ such that $$\begin{aligned}
S(\zeta) \geq \kappa_N \|\zeta\|^{\delta}\,,\quad
\zeta \in |W|\,,\ \|\zeta\|\geq K\, .\end{aligned}$$ Let $$\theta \in {\cal D}(]-3\kappa_N/4,3\kappa_N/4[$$ such that $\theta\equiv 1$ on $[-\kappa_N/4,\kappa_N/4]$ ; for any $R>0$, let the element $\varphi_R$ in $C^1(\C^n)$ defined as $$\varphi_R~: \zeta \mapsto \theta (S(\zeta) /R^\delta)\,.$$ Since the restriction $S_{|_{|W|}}$ is a proper map (all $\delta_j$’s, $j=1,...,q$, being strictly positive) and $V(P)\cap |W|$ is a discrete (hence finite) algebraic subset of $\C^n$ (this follows also from (3.1)), there exists $R_0$ such that for $R>R_0$, $\varphi_R \equiv 1$ in a neighborhood of $|W|\cap V(P)$. Therefore, if $$s_j:={\overline {P_j} \over \|P\|^2}$$ one has (see for example formula (2.9)) $$\begin{aligned}
&&\RR \left[\matrix{ [W]\wedge Q \bigwedge\limits_{l=1}^q dX_{l}\cr \cr
P_1,...,P_q}\right]\nonumber \\
&&\qquad\qquad\qquad = c_q
\int_{|W|_{\rm reg}} \Big(\sum\limits_{j=1}^q (-1)^{j-1} s_j\bigwedge\limits_{{l=1}\atop {l\not=j}}^q
{ds_l}\Big)\wedge Q\, d\zeta'\, \wedge \overline\partial\varphi_R
\nonumber \\ \end{aligned}$$ for any $R>R_0$, where $d\zeta'=\bigwedge_{l=1}^q d\zeta_l$. It follows from an homotopy argument similar to the one which is developped in the proof of proposition 2.2 that $$\begin{aligned}
&&\RR \left[\matrix{ [W]\wedge Q \bigwedge\limits_{l=1}^q dX_{l}\cr \cr
P_1,...,P_q}\right]\nonumber \\
&&\qquad\qquad\qquad =c_q
\int_{|W|_{\rm reg}} \Big(\sum\limits_{j=1}^q (-1)^{j-1} \widetilde
s_j\bigwedge\limits_{{l=1}\atop {l\not=j}}^q
{d\widetilde
s_l}\Big)\wedge Q\, d\zeta'\, \wedge
\overline\partial\varphi_R
\nonumber \\
&&\qquad\qquad\qquad =c_q
\int_{|W|_{\rm reg}} \Big(\sum\limits_{j=1}^q (-1)^{j-1} \widetilde
s_j\bigwedge\limits_{{l=1}\atop {l\not=j}}^q
{\overline\partial \widetilde
s_l}\Big)\wedge Q \, d\zeta'\, \wedge
\overline\partial\varphi_R\nonumber \\\end{aligned}$$ for any $R>R_0$. Since $ P_j \widetilde s_j= u_j^2/S$, $j=1,...,q$, one can rewrite (3.8) as $$\begin{aligned}
&&\RR \left[\matrix{ [W]\wedge Q \bigwedge\limits_{l=1}^q dX_{l}\cr \cr
P_1,...,P_q}\right]\nonumber \\
&&=c_q\, 2^{q-1}
\int_{|W|_{\rm reg}}
\Big(\prod\limits_{j=1}^q {|P_j|\over P_j}\,
u_j^{1-{2\over \delta^{[j]}}}\Big)
{\sum\limits_{j=1}^q (-1)^{j-1}
u_j\bigwedge\limits_{{l=1}\atop {l\not=j}}^q
{d
u_l} \over \|u\|^{2q}} \wedge Q\, d\zeta'\, \wedge
\overline\partial\varphi_R
\nonumber \\
&&={ (-1)^q \, c_q\, 2^{q-1}\over R^\delta}
\int_{|W|_{\rm reg}}
\Big(\prod\limits_{j=1}^q {|P_j|\over P_j}\,
u_j^{1-{2\over \delta^{[j]}}}\Big)
\, {\bigwedge\limits_{l=1}^q du_l \over \|u\|
^{2(q-1)}}
\wedge \theta'\Bigg({\|u\|^2\over R^\delta}\Bigg) \,
Q\, d\zeta'. \nonumber \\\end{aligned}$$ For any order ${\cal J} \subset \{1,...,q\}$, let $$\omega_{\cal J}=\bigwedge\limits_{j=1}^q d\alpha_{{\cal J},l}$$ where $$\alpha_{{\cal J},l} (\zeta_1,...,\zeta_q):=
\cases {{\rm Re}\, \zeta_j\ {\rm if}\ j\in {\cal J} \cr
{\rm Im}\, \zeta_j\ {\rm if}\, j\not\in {\cal J}}\,;$$ then one can write $$d\zeta'=\bigwedge\limits_{l=1}^q d\zeta_l =\sum\limits_{{\cal J}\subset \{1,...,q\}}
i^{q-\#{\cal J}}\, d\omega_{{\cal J}}\, .$$ In order to prove formula (3.4), it is enough to prove that for any ${\cal J}
\subset \{1,...,q\}$, one has, as soon as ${\rm deg}\, Q <\delta_1+\cdots+\delta_q -q$, $$\begin{aligned}
&&\lim\limits_{R\rightarrow +\infty}
\Bigg[{1\over R^\delta}
\int_{|W|_{\rm reg}}
\Big(\prod\limits_{j=1}^q {|P_j|\over P_j}\,
u_j^{1-{2\over \delta^{[j]}}}\Big)
\, {\bigwedge\limits_{l=1}^q du_l \over \|u\|
^{2(q-1)}}
\wedge \theta'\Bigg({\|u\|^2 \over R^\delta}\Bigg) \,
Q\, \omega_{{\cal J}} (\zeta') \Bigg] =0\, .\nonumber\\
&&\end{aligned}$$ Since the restriction of $P=(P_1,...,P_q)$ to each connected sheet ${\cal F}$ (above the $\zeta'$-space) of the $2q$-dimensional real manifold $|W|_{\rm reg}$ is proper, the map $$F_{\cal J,{\cal F}}~:\
\zeta \in {\cal F}\mapsto (u_1,...,u_q, \alpha_{{\cal J},1},...,
\alpha_{{\cal J},q})$$ is a $\R^{2q}$-valued proper map, with topological degree $d_{{\cal J},{\cal F}}$. Moreover, condition (3.6) implies that, for $R>K$, $${\rm Supp} \Big(\theta(S/R^{\delta})\Big)
\subset \{ \zeta \in \C^n\,:\, \|\zeta\| <R\}\, .$$ Actually, for $\|\zeta\|\geq R>K$, one has $$S(\zeta) \geq \kappa_N \|\zeta\|^{\delta}
\geq \kappa_N R^{\delta} > (3 \kappa_N/4) R^{\delta}\, .$$ For such $R$, one has $$\Bigg\|\prod\limits_{j=1}^q {|P_j|\over P_j} \, Q\, \theta' (S/R^{\delta})\Bigg\|_\infty
\leq C\, R^{\,{\rm deg}\, Q}\,,$$ where $C=C(\theta, Q)$ is a positive constant. It follows then from the properness of all maps $F_{{\cal J},{\cal F}}$ and from the positivity of the differential form $$\Big(\prod\limits_{j=1}^q \,
u_j^{1-{2\over \delta^{[j]}}}\Big)
\bigwedge\limits_{l=1}^q du_l$$ in $]0,\infty[^q$ that $$\begin{aligned}
&&{1\over R^\delta} \Bigg|
\int_{|W|_{\rm reg}}
\Big(\prod\limits_{j=1}^q {|P_j|\over P_j}\,
u_j^{1-{2\over \delta^{[j]}}}\Big)
\, {\bigwedge\limits_{l=1}^q du_l \over \|u\|
^{2(q-1)}}
\wedge \theta'\Bigg({\|u\|^2 \over R^\delta}\Bigg) \,
Q\, \omega_{{\cal J}} (\zeta') \Bigg|\\
&& \quad
\leq
{(\sum\limits_{\cal F} d_{{\cal J},{\cal F}})\, C \, R^{\, {\rm deg}\, Q} \over R^{\delta}}
\Bigg( \int_{{\kappa_N R^{\delta}\over 4}
\leq \|u\|^2 \leq {3 \kappa_N R^{\delta}\over 4}}
\Big( \prod\limits_{j=1}^q u_j^{1-{2\over \delta^{[j]}}}\Big)
{\bigwedge\limits_{l=1}^q du_l
\over \|u\|^{2(q-1)}}
\Bigg) \\
&&\quad\quad \times \Bigg(
\int_{\|t\| < R} dt_1\wedge \cdots \wedge dt_q \Bigg) \\
&&\leq
{(\sum\limits_{\cal F} d_{{\cal J},{\cal F}})\, \, C_N \, R^{\, {\rm deg}\, Q + q} \over R^{q \delta}}
\Bigg( \int_{{\kappa_N R^\delta \over 4}
\leq \|u\|^2 \leq {3 \kappa_N R^\delta \over 4}}
\Big( \prod\limits_{j=1}^q u_j^{1-{2\over \delta^{[j]}}}\Big)
\bigwedge\limits_{l=1}^q du_l \Bigg)\\
&&\leq
{(\sum\limits_{\cal F} d_{{\cal J},{\cal F}})\, \, \widetilde C_{N,\vec \delta}
\, R^{\, {\rm deg}\, Q + q} \over R^{q \delta}} \, R^{
{\delta\over 2} \sum\limits_{j=1}^q \Big(1-{1\over \delta^{[j]}}\Big)+
q{\delta \over 2}}\\
&&\leq
(\sum\limits_{\cal F} d_{{\cal J},{\cal F}})\, \, \widetilde C_{N,\vec \delta}
\, R^{{\rm deg}\, Q +q -\delta_1-\,\cdots\, -\delta_q}= {\bf o} (1)\,,\end{aligned}$$ which proves the conclusion (3.10) we need. The proof of proposition 3.1 is therefore completed. $\quad\diamondsuit$
Analytic versus algebraic approach
==================================
Let ${\cal X}$ be an integral $\C$-variety of dimension $q$ and ${\cal D}_1$,...,${\cal D}_q$ be $q$ Cartier divisors on ${\cal X}$ such that $|{\cal D}_1|\cap \cdots \cap
|{\cal D}_q|$ is finite. If $\omega$ is a meromorphic form in $\Omega^q_{\C({\cal X})/\C}$ which has a simple pole along ${\cal D}_1+\cdots+ {\cal D}_q$, one may define (see [@hu:gnus], page 621) the local residue of $\omega$ at any closed point $\alpha$ in $|{\cal D}_1|\cap \cdots \cap |{\cal D}_q|$. That is, if $$\omega = {\eta \over f_1 \cdots f_q}\,,$$ where $\eta \in \omega^q_{\C({\cal X})/\C,\alpha}$ and $f_j=0$, $j=1,...,q$, is a local equation for ${\cal D}_j$ at $\alpha$ then $${\rm Res}_{{\cal X};{\cal D}_1,...,{\cal D}_q,\alpha} \, (\omega)
={\rm Res}_{\C({\cal X})/\C,\alpha} \, \Bigg( \Bigg[ \matrix{\eta \cr
f_1,...,f_q }\Bigg]\Bigg)\,.$$ When ${\cal X}$ is smooth, this definition agrees with the definition in [@gh:gnus], chapter 5, section 1 (see [@lip1:gnus], Appendix A). Adding the hypothesis that ${\cal X}$ is $\C$-complete, one has (see proposition 12.2, page 108, in [@lip2:gnus]) $$\sum\limits_{\alpha \in |{\cal D}_1|\cap \cdots \cap |{\cal D}_q|}
\, {\rm Res}_{{\cal X};{\cal D}_1,...,{\cal D}_q,\alpha} \, (\omega)=0\,,$$ which is known as residue theorem on ${\cal X}$ (it extends the classical residue theorem on a complete integral curve in its algebraic formulation, see [@se:gnus]). Such a residue theorem holds in our analytic setting (and is essentially a consequence of Stokes’s formula). Namely, if $W$ is an integral algebraic $q$-dimensional subscheme in ${\bf A}^n_\C$ (with completion ${\cal W}$ in ${\rm Proj}\, \C[X_0,...,X_n]$) and $P_1,...,P_q$ are $q$ polynomials in $n$ variables such that $|{\cal W}|\cap \{{}^h P_1=\cdots={}^h P_q=0\}$ is finite and included in $\C^n$, then ($[W]$ being understood as the integration current free of multiplicities), $${\rm Res}\left[\matrix{[W]\wedge Q(X) dX_{i_1}\wedge \cdots \wedge dX_{i_q}
\cr P_1,...,P_q}\right]=0$$ when ${\rm deg}\, Q \leq \sum_{j=1}^q {\rm deg}\, P_j -q -1$ (corollary 3.1) for any ordered subset $\{i_1,...,i_q\}\subset \{1,...,n\}$. On the other hand, the transformation law holds for our analytic restricted residue (see corollary 2.2). Such a transformation law remains valid (in its local formulation) for restricted residue symbols defined through the algebraic approach (see theorem 2.4 in [@huku0:gnus]). Finally, the local residue symbol $${\rm Res}_{{\cal W};{\cal D}_1,...,{\cal D}_q,\alpha} \, (\omega)
={\rm Res}_{\C({\cal X})/\C,\alpha} \, \Bigg( \Bigg[ \matrix{\eta \cr
f_1,...,f_q }\Bigg]\Bigg)\,,$$ where $$\omega = {\eta \over f_1 \cdots f_q}\,,$$ $\eta \in \omega^q_{\C({\cal X})/\C,\alpha}$ and $f_j=0$, $j=1,...,q$, is a local equation for ${\cal D}_j$ at $\alpha$, equals to $0$ as soon as $\eta=f_j \widetilde \eta$ for some $\widetilde \eta \in \omega^q_{\C({\cal X})/\C,\alpha}$ (see also [@huku0:gnus], section 2). The same annihilation property is satisfied by the restricted residual current (proposition 2.3). Our goal in this section is to profit from the fact that both restricted residual objects (defined through the algebraic or analytic approach) satisfy the transformation law, the residue formula, the annihilation property, in order to show that they coincide. Therefore, we are able to give an algebraic formulation of the Proposition 3.1, which is the Theorem 1.1 stated in our introduction. In order to do that, we will need the following technical lemma :
Let $|W|$ be an irreducible $q$-dimensional algebraic set in $\C^n$ and $|{\cal W}|$ its Zariski closure in $\P^n(\C)$. Let also $P_1,...,P_q$ be $q$ polynomials in $\C[X_1,...,X_n]$ such that $V(P)\cap |W|$ is a discrete (hence finite) algebraic subset of $\C^n$, with $0 \in V(P)\cap |W|$. Then, there exists $N_0>0$ such that, for any integer $N\geq N_0$, one can find $q n+1$ complex parameters $u_{jk}$, $ j=1,\dots,q$, $k=1,\dots, n$, $t\in \C^*$, so that, if $$\widetilde P_{j}^{(N,u,t)} (X):= tP_j(X) + \Big(\sum\limits_{k=1}^n u_{jk} X_k\Big)^N \,,\quad j=1,...,q\,,$$ one has :
- any point $\alpha\in |W|\cap V(\widetilde P^{(N,u,t)})$ but $0$ belongs to $|W|_{\rm reg}$ ;
- the set $$|{\cal W}|\cap \{[\zeta_0\,:\,...\,:\, \zeta_n]\in \P^n(\C)\,:\, {}^h P^{(N,u,t)}_j (\zeta_0,...,\zeta_n)=0\,,\ j=1,...,q\}$$ is contained in $\C^n$.
[**Proof.**]{} Since $|W|$ is irreducible and $q$-dimensional, one has ${\rm dim}\, |W|_{\rm sing} <q$ ; one can find an algebraic affine hypersurface $H:=\{\zeta\in \C^n\,;\, H(\zeta)=0\}$ (with Zariski closure $|{\cal H}|$) such that $|W|_{\rm sing} \subset H$ and ${\rm dim}\, (|{\cal W}|\cap |{\cal H}|)<q$. Let $N_0 > \deg P_j$, $j=1,...,q$, and $N\geq N_0$. Assume also that $N\geq \rho_{P,W} (0)$, where $\rho_{P,W}(0)$ is the order of vanishing of $P$ at the origin (along $|W|$).
Let $u=[u_{jk}]$, $j=1,...,q$, $k=1,...,n$, be a $(q,n)$ matrix with generic complex entries, $$M_u:=\{\zeta\in \C^n\,:\, u_{j1}\zeta_1+\cdots + u_{jn} \zeta_n=0\,,\ j=1,...,q\}$$ and $|{\cal M}_u|$ its Zariski closure in $\P^n(\C)$. Since ${\rm dim}\, |{\cal W}|=q$ and ${\rm dim}\, (|{\cal W}|\cap |{\cal H}|)<q$, $|{\cal M}_u| \cap |{\cal W}|\subset \C^n$ and $|{\cal M}_u|\cap |{\cal W}|\cap |{\cal H}|=\{0\}$ for $u$ generic. Therefore, for such a generic choice of $u$ ($u=u^{0}$) (this choice will be refined later), for any $t\in \C^*$, the polynomials $$t P_j(X) + (u^{0}_{j1} X_1+\cdots + u_{jn}^0 X_n)^N \,, \quad j=1,...,q\,,$$ define in $\C^n$ an algebraic set $Z^{(N,u^0,t)}$ whose closure ${\cal Z}^{(N,u^0,t)}$ in $\P^n(\C)$ intersects $|{\cal W}|$ only at points in $\C^n$ (note that $0$ is one of these points). The algebraic set $|W|\cap Z^{(N,u^0,t)}$ can be described as $$|W|\cap Z^{(N,u^0,t)}=\{\zeta^{(N,1)}(u^0,t),...,\zeta^{(N,m)} (u^0,t)\} \cup \{0\}\,,$$ where $m$ is fixed (depending on $N$ and $|{\cal W}|$) and the $t\mapsto \zeta^{(N,j)} (u^0,t)$, $j=1,...,m$, are algebraic $\C^n$-valued functions of $t$ which are not identically $0$ and can be classified in two classes, depending on their behavior when $|t|$ tends to zero. A branch $t\mapsto \zeta^{(N,j)} (u^0,t)$ will be in the first class if $\zeta^{(N,j)} (u^0,t)$ tends to zero when $|t|$ tends to $0$. It will be in the second class if $\zeta^{(N,j)} (u^0,t)$ tends to a point in $|W|\cap M_{u^0}$ which is distinct from $0$ when $|t|$ goes to $0$. It follows then from $M_{u^0}\cap |W|\cap H=\{0\}$ that none of the functions $$t\mapsto H(\zeta^{(N,j)} (u^0,t))$$ where $t\mapsto \zeta^{(N,j)} (u^0,t)$ belongs to the second category, can be identically equal to $0$. The behavior of branches of the first category can now be studied when $|t|$ goes to infinity. The assumption on $N$ ensures us that such branches either approach points in $(|W|\cap V(P)) \setminus \{0\}$, either satisfy $$\lim\limits_{|t|\rightarrow \infty} |\zeta^{(N,j)} (u^0,t)|=+\infty$$ in the second alternative. The hypothesis on $u^0$ implies that the function $t\mapsto H(\zeta^{(N,j)} (u^0,t))$ is not identically $0$ if we are in the second alternative. If $u^0$ is conveniently choosen (in terms of the Taylor developments at the first order for $P_1,...,P_q$ at the points in $(|W|\cap V(P)) \setminus \{0\}$, the assertion $t\mapsto H(\zeta^{(N,j)} (u^0,t))\not\equiv 0$ also holds for branches concerned by the first alternative. Finally, for any branch $t\mapsto \zeta^{(N,j)} (u^0,t)$, one has $H(\zeta^{(N,j)} (u^0,t)\not\equiv 0$. Therefore, once $u^0$ has been conveniently chosen, one can pick up $t\not=0$ such that the map $\widetilde P^{(N,u^0,t)}$ satisfies the assertions of the lemma. $\quad \diamondsuit$
We can now relate the analytic and algebraic approaches for restricted residual symbols.
Let ${\cal W}$ be a complete integral $\C$-variety of dimension $q$, embedded in the projective scheme ${\rm Proj}\, \C [X_0,...,X_n]$, $\alpha$ be a closed point in $|{\cal W}|$ such that $\alpha\in \C^n$ and ${\cal D}_1$,...,${\cal D}_q$ be $q$ Cartier divisors on ${\cal W}$ so that the intersection $|{\cal D}_1|\cap \cdots \cap
|{\cal D}_q|$ define a zero-dimensional scheme on ${\cal W}$ in a neighborhood of $\alpha$. If $$\omega = {\eta \over P_1 \cdots P_q}\,,$$ where $\eta = Q\, dX_{i_1}\wedge \cdots \wedge dX_{i_q}$, $Q\in \C[X_1,...,X_n]$ induces an element in $\omega^q_{\C({\cal X})/\C,\alpha}$ and $P_1,...,P_q$ are elements in $\C[X_1,...,X_n]$ such that $P_j$, $j=1,...,q$, is a local equation for ${\cal D}_j$ at $\alpha$, then, for any function $\varphi \in{\cal D}(\C^n)$ with arbitrary small support around $\alpha$ satisfying $\varphi\equiv 1$ in a neighborhood of $\alpha$, one has $$\begin{aligned}
{\rm Res}_{{\cal W};{\cal D}_1,...,{\cal D}_q,\alpha} \, (\omega) =
{\rm Res}\, \left[ \matrix{ [W]\wedge \varphi \, \eta \cr
P_1,..., P_q }\right]\, . \end{aligned}$$
[**Proof.**]{} One can assume for the sake of simplicity that $\alpha=0$. Let ${\cal M}$ be the maximal ideal $(X_1,...,X_n)$ in the local algebra ${\cal O}_{\C[X_1,...,X_n],0}$ and $(I(W))_0$ the localization at $0$ of the radical ideal $$I(W):=\{g\in \C[X_1,...,X_n]\,;\,
g(\zeta)=0 \ \forall \zeta \in |{\cal W}|\cap \C^n\}\, .$$ Choose $p\in \N^*$ such that $${\cal M}^{p} \in \Big([(P_1,...,P_q)_{0}]^2 , I(W)_0 \Big)\, .$$ It follows from the validity of the transformation law and the annihilating property in the algebraic context that, if $$\widetilde P_j(X):=P_j(X)+ \Big( \sum\limits_{k=1}^n u_{jk} X_k \Big)^{p},$$ then one has, for any $\eta= Q\, dX_{i_1}\wedge \cdots \wedge dX_{i_q}$, $Q\in \C[X_1,...,X_n]$, that $$\begin{aligned}
{\rm Res}_{{\cal W};{\cal D}_1,...,{\cal D}_q,0} \, \Big( {\eta \over P_1 \cdots P_q}\Big)
={\rm Res}_{{\cal W};\widetilde {\cal D}_1,...,\widetilde {\cal D}_q,0} \, \Big( {\eta \over \widetilde P_1 \cdots
\widetilde P_q}\Big)\,,\end{aligned}$$ where $\widetilde {\cal D}_j$, $j=1,...,q$, is the Cartier divisor on ${\cal W}$ with local equation $\widetilde P_j$ in a neighborhood of the origin. On the other hand, it follows from Proposition 2.2 and Proposition 2.3 that, for any test-function $\varphi$ with arbitrary small support around the origin, one has also $$\begin{aligned}
{\rm Res}\, \left[ \matrix{ [W]\wedge \varphi \, \eta \cr
P_1,..., P_q }\right] =
{\rm Res}\, \left[ \matrix{ [W]\wedge \varphi\, \eta \cr
\widetilde P_1,..., \widetilde P_q }\right]\, . \end{aligned}$$ If the $u_{jk}$, $j=1,...,q$, $k=1,...,n$ are generic (see for example the construction in the proof of lemma 4.1), the algebraic set $V(\widetilde P)\cap |{\cal W}|\cap \C^n$ is discrete (hence finite). We can then conclude from (4.2) and (4.3) that in order to prove (4.1), it is not restrictive to assume that the algebraic set $V(P)\cap |{\cal W}| \cap \C^n$ is finite, what we will do from now on. The same argument as above shows that, in order to prove (4.1), one can replace $P_j$, $j=1,...,q$, by the polynomial $${1\over t} \widetilde P_j^{(N,u,t)}$$ constructed in lemma 4.1 ($N$ being choosen sufficiently large, certainly such that $N\geq \max \deg P_j$, ${\cal M}^N \subset (I(P)_0,I(W)_0)$) and $\deg Q < q (N-1)$), and this is what we do (preserving the notations $P_j$ and ${\cal D}_j$). As a consequence of the residue formula in the algebraic context (which we recalled at the beginning of this section) and of Corollary 3.1, one has $$\begin{aligned}
&&\sum\limits_{\alpha \in V(P) \cap {\cal W} (\C)}
{\rm Res}_{{\cal W};{\cal D}_1,...,{\cal D}_q,\alpha} \, \Big( {\eta \over P_1 \cdots P_q}\Big)
=\sum\limits_{\alpha \in V(P) \cap {\cal W} (\C)}
{\rm Res}\, \left[ \matrix{ [W]\wedge \varphi \, \eta \cr
P_1,..., P_q }\right]\nonumber \\
&&\end{aligned}$$ whenever $\varphi$ is a test-function in ${\cal D}(\C^n)$ with arbitrary small support around the points $\alpha \in V(P)\cap |{\cal W}|$, such that $\varphi\equiv 1$ in a neighborhood of each of these points ($\varphi_\alpha$ will denote next $\varphi\,\theta_\alpha$, where $\theta_\alpha$ is a test-function with support arbitrary small around $\alpha$ and $\theta_\alpha\equiv 1$ in a neighborhood of $\alpha$). If $\alpha$ is any point in $V(P)\cap |{\cal W}|$ distinct from $0$, ${\cal W}$ is smooth about $\alpha$ (lemma 4.1, first assertion) and we know in this case that $$\begin{aligned}
{\rm Res}_{{\cal W};{\cal D}_1,...,{\cal D}_q,\alpha} \, \Big( {\eta \over P_1 \cdots P_q}\Big)
&=&{\rm Res}\, \left[ \matrix{ [W]\wedge \varphi_\alpha\, \eta \cr
P_1,..., P_q }\right]\,,
\end{aligned}$$ since the construction of our restricted residual currents corresponds to the construction proposed in [@gh:gnus], chap 5, section 1 (this is a consequence of the classical relation between Bochner-Martinelli and Cauchy kernels), which is known to fit with the algebraic approach in the smooth case (as it was recalled at the beginning of this section). Formula (4.1) follows then from (4.4) and from the identifications (4.5). $\quad\diamondsuit$
[**Proof of Theorem 1.1.**]{} We may now transpose to the algebraic context the analytic result stated in proposition 3.1. This gives the statement of the Theorem 1.1 of our introduction, provided we remember that we have $${\rm Res}_{{\cal W};{\cal D}_1,...,{\cal D}_q,\alpha} \, (\omega)
={\rm Res}_{\C({\cal W})/\C,\alpha} \, \Bigg( \Bigg[ \matrix{\eta \cr
f_1,...,f_q }\Bigg]\Bigg)$$ for any point $\alpha$ in $|{\cal W}|\cap |{\cal D}_1|\cap
\cdots \cap |{\cal D}_q| \cap \C^n$ (here we just assume that ${\cal D}_1$,...,${\cal D}_n$ define a $0$-dimensional scheme on $W$, there is no assumption about what happens on $|{\cal W}|\setminus |W|$) and any $\omega$ in $\Omega^q_{\C({\cal W})/\C}$ with simple poles (in $W$) along ${\cal D}_1+\cdots +{\cal D}_q$ ($\eta=f_1\cdots f_q \,\omega$, where $f_j$ denotes a local equation for the Cartier divisor ${\cal D}_j$). Since the reference to the divisors ${\cal D}_1,...,{\cal D}_q$ was implicit in the expression of the element in $\Omega^q_{\C({\cal W})/\C}$, we used the abridged notation ${\rm Res}_{W,\alpha} [\ ]$ instead of ${\rm Res}_{{\cal W};{\cal D}_1,...,{\cal D}_q,\alpha}$ in order to formulate the statement in this theorem. $\diamond $\
As a direct consequence we formulate the restricted version of the Cayley-Bacharach Theorem.
Let $W$ be a $q$-dimensional irreducible affine algebraic subvariety in ${\bf A}^n_\C$ $$0<q<n$$ and $P_1,...,P_q$ be $q$ polynomials in $\C[X_1,...,X_n]$ satisfying the condition $(1.2)$. Assume also that $V(P)$ and $\vert W\vert $ intersect transversally at any of the $k$ points in $V(P)\cap \vert W\vert $. Then any algebraic hypersurface $\{Q=0\}$, $Q\in \C[X_1,...,X_n]$, such that ${\rm deg}\, Q <\delta_1+\cdots +\delta_q -q$, which passes through any $k-1 $ points of the set $V(P)\cap W $ passes through the last one also.
An affine version of Wood’s theorem.
====================================
Let $\gamma_1$, ...,$\gamma_d$ be $d$ pieces of manifold in $\P^n(\C)$ and $|{\cal L}_{0,0}|$ be a line in $\P^{n}(\C)$ which intersects each of the $\gamma_j$ transversally respectively at distinct points $p_{j0}$, $j=1,...,d$. Assume that affine coordinates are such that the support $|{\cal L}_{0,0}|$ is the line $\zeta_1=\cdots=\zeta_{n-1}=0$. Then, for $(\alpha,\beta)\in (\C^{n-1})^2$ close to $(0,0)$, the projective line $$|{\cal L}_{\alpha,\beta}|:=\{[\zeta_0:\,... \, :\zeta_n]\in \P^n(\C)\,;\,
\zeta_k= \alpha_k\, \zeta_n +\beta_k\, \zeta_0\,,\
k=1,...,n-1\}$$ intersects tranversally $\gamma_1$,...,$\gamma_d$ at the respective points $p_1(\alpha,\beta)$,..., $p_d(\alpha,\beta)$ ($p_j(\alpha,\beta)$ being close to $p_{j0}$). In [@wood1:gnus], J. Wood gave a simple criterion for the local germs of manifold $\gamma_1,...,\gamma_d$ to be germs of a global algebraic hypersurface (with degree $d$) $|{\cal H}|$ in $\P^n(\C)$ satisfying the relation such that $$|{\cal H}|\cap |{\cal L}_{\alpha,\beta}| =
\{p_1(\alpha,\beta),...,p_d(\alpha,\beta)\}$$ for $(\alpha,\beta)$ close to $(0,0)$. The (necessary and sufficient) condition he gave can be formulated as follows : $$\begin{aligned}
\sum\limits_{j=1}^d \zeta_n [p_j(\alpha,\beta)]= h_0(\alpha)+\sum\limits_{k=1}^{n-1}
h_k(\alpha)\, \beta_k\,,\end{aligned}$$ where $h_0,...,h_{n-1}$ are germs of holomorphic functions in $\alpha$ at the origin (here $\zeta_n[p]$, where $p$ denotes a point in $\C^n$, means the $n$-th affine coordinate of $p$). Note that the algebraic hypersurface $|{\cal H}|$ (in $\P^n(\C)$) which interpolates $\gamma_1,...,\gamma_d$ is such that its intersection at infinity with any line $|{\cal L}_{\alpha,\beta}|$, with $(\alpha,\beta)$ close to $(0,0)$, is empty. What we would like to state here is an affine analog of this result, ${\bf P}^n(\C)$ being replaced by some irreducible $q$-dimensional affine algebraic subvariety of $\C^n$ ($q=2,...,n$).
Let us first state the following easy consequence of our Theorem 1.1.
Let $W$ be an algebraic irreducible $q$-dimensional subvariety of the affine scheme ${\bf A}^n_\C$ (with $2\leq q\leq n$), $m$ be a positive integer strictly between $0$ and $q$, and $\gamma_1$,...,$\gamma_d$ be $d$ disjoint pieces of $q-m$-dimensional analytic manifold such that $\gamma_j$ lies in $|W|_{\rm reg}$ for $j=1,...,d$. Furthermore, assume that the affine $n+m-q$-dimensional subspace $$L_{0,0}:=\{\zeta \in \C^n\,;\, \zeta_k=0\,,\ k=1,...,q-m\}$$ intersects each $\gamma_j$ transversally respectively at points $p_{j0}$, $j=1,...,d$. Suppose that there are strictly positive rational numbers $\delta_1,...,\delta_m$ and polynomials $P_1,...,P_m$ with $\deg P_j=d_j\geq \delta_j$, $j=1,...,m$, such that
- $|W| \cap V(P)$ is a $q-m$-dimensional variety in $\C^n$ which interpolates the pieces $\gamma_j$ and is such that $|W|\cap V(P)\cap L_{0,0}=\{p_{10},...,p_{d0}\}$ ;
- there exists strictly positive constants $\kappa, K$ such that $$\begin{aligned}
\zeta \in |W|\,,\ \|\zeta\|\geq K \Longrightarrow
\sum\limits_{j=1}^m {|P_j(\zeta)|\over \|\zeta\|^{\delta_j}} +
\sum\limits_{k=1}^{q-m} {|\zeta_k|\over \|\zeta\|} \geq \kappa\, . \end{aligned}$$
Then, for $(\alpha,\beta)$ close to $(0,0)$ in $(\C^{n+m-q})^{q-m} \times \C^{q-m}$, the affine $n+m-q$-dimensional subspace $$L_{\alpha,\beta}:=\Big\{\zeta \in \C^n\,;\, \zeta_k=\sum\limits_{r=1}^{n+m-q}
\, \alpha_{k,r} \zeta_{q-m+r} + \beta_k\,,
\ k=1,...,q-m\Big\}$$ intersects each $\gamma_j$ transversally respectively at the points $p_j(\alpha,\beta)$, $j=1,...,d$ (necesseraly distinct and close to the $p_{j0}$) and one has $$\begin{aligned}
\sum\limits_{j=1}^d \zeta_l [p_j(\alpha,\beta)]=
\sum\limits_{{\underline k\in \N^{q-m}}\atop {|\underline k|
\leq \rho +1}}
h_{\underline k}^{(l)} (\alpha)\, \beta_1^{k_1}\cdots \beta_{q-m}^{k_{q-m}}\,,\
l=q-m+1,...,n\,,\end{aligned}$$ where the $h_{\underline k}^{(l)}$ are germs of holomophic functions in $\alpha$ about the origin and $$\rho:=\sum_{j=1}^m (d_j-\delta_j)$$
[**Proof.**]{} Let, for $k=1,...,q-m$, $$\Lambda_{\alpha,\beta,k}(\zeta):=\zeta_k-\sum\limits_{r=1}^{n+m-q} \alpha_{k,r} \,
\zeta_{q-m+r} -\beta_k\,,\ \zeta \in \C^n \, ;$$ condition $(5.7)$ implies that, when $(\alpha,\beta)$ is sufficiently close to $(0,0)$, one has $$\begin{aligned}
\zeta \in |W|\,,\ \|\zeta\|\geq K \Longrightarrow
\sum\limits_{j=1}^m {|P_j(\zeta)|\over \|\zeta\|^{\delta_j}} +
\sum\limits_{k=1}^{q-m} {|\Lambda_{\alpha,\beta,k} (\zeta)|\over \|\zeta\|}
\geq {\kappa\over 2}\, . \end{aligned}$$ This shows that for $(\alpha,\beta)$ close to $(0,0)$, the only points in $L_{\alpha,\beta} \cap |W| \cap V(P)$ are $d$ points $p_j(\alpha,\beta)$, $j=1,...,d$ which approach the points $p_{10},...,p_{d0}$ (about each of these points, one can use the implicit function theorem in order to describe the intersection $\gamma_j \cap L_{\alpha,\beta}$). This proves the first assertion of the proposition. It follows from proposition 3.1 that, as soon as the multi-index $\underline k\in \N^{q-m}$ is such that $$\sum\limits_{j=1}^m (d_j-1) +1 < \sum\limits_{j=1}^m \delta_j +
\sum\limits_{l=1}^{q-m} (k_l+1)-q=\sum\limits_{j=1}^m \delta_j+
|\underline k|-m\,,$$ then, for $l=q-m+1,...,n$, for any finite ordered subset $\{i_1,...,i_{q-m}\}
\subset \{1,...,n\}$, $${\rm Res}\, \left[\matrix{ [W] \wedge X_l \, \Big(\bigwedge\limits_{j=1}^m
dP_j\Big) \wedge \Big(\bigwedge\limits_{l=1}^{q-m} dX_{i_l}\Big) \cr \cr
P_1,...,P_m, (\Lambda_{\alpha,\beta,1})^{k_1+1},...,
(\Lambda_{\alpha,\beta,q-m})^{k_{q-m}+1} }\right]=0$$ for $(\alpha,\beta)$ such that (5.9) holds. It is immediate to check (use for example formula $(3.7)$) that for such $(\alpha,\beta)$, one has, for any multi-index $\underline k=(k_1,...,k_{q-m})\in \N^{q-m}$, $$\begin{aligned}
&& {\partial^{|\underline k|} \over
\partial \beta_1^{k_1}\cdots
\partial \beta_{q-m}^{k_{q-m}}}
\, {\rm Res}\, \left[\matrix{ [W] \wedge X_l \, \Big(\bigwedge\limits_{j=1}^m
dP_j\Big) \wedge \Big(\bigwedge\limits_{l=1}^{q-m} dX_{i_l}\Big) \cr \cr
P_1,...,P_m, \Lambda_{\alpha,\beta,1},...,
\Lambda_{\alpha,\beta,q-m}}\right] \nonumber \\
&&
\quad =\pm\, {\rm Res}\, \left[\matrix{ [W] \wedge X_l \, \Big(\bigwedge\limits_{j=1}^m
dP_j\Big) \wedge \Big(\bigwedge\limits_{l=1}^{q-m} dX_{i_l}\Big) \cr \cr
P_1,...,P_m, (\Lambda_{\alpha,\beta,1})^{k_1+1},...,
(\Lambda_{\alpha,\beta,q-m})^{k_{q-m}+1} }\right]\, . \end{aligned}$$ Then it follows from $(5.9)$ that the right-hand side of $(5.10)$ (hence the left-hand side) equals identically $0$ when $$|\underline k| > \sum\limits_{j=1}^m (d_j-\delta_j) +1=
\rho+1\,.$$ This proves that, when $(\alpha,\beta)$ is close to $(0,0)$ and $l=m-q+1,...,n$, $$\sum\limits_{j=1}^d \zeta_l [p_j(\alpha,\beta)]
\equiv {\rm Res}\, \left[\matrix{ [W] \wedge \zeta_l \, \Big(\bigwedge\limits_{j=1}^m
dP_j\Big) \wedge \Big(\bigwedge\limits_{l=1}^{q-m} d\Lambda_{\alpha,\beta,l}
\Big) \cr \cr
P_1,...,P_m, \Lambda_{\alpha,\beta,1},...,
\Lambda_{\alpha,\beta,q-m}}\right]$$ is a polynomial expression in $\beta=(\beta_1,...,\beta_{q-m})$ with total degree at most $\rho+1$ (the coefficients being holomorphic functions in $\alpha$). The second assertion of the proposition is proved. $\quad\diamondsuit$ [**Remark.**]{} Note that we recover here as a particular case the necessity of Wood’s condition in the case $W={\bf A}^n_\C$, $m=1$, $\delta_1=d_1=d$, which means precisely that in this case we also impose the restriction $$\{\widetilde \zeta \in \P^n(\C)\,;\, {}^h P_1(\widetilde \zeta)=0\}\cap |{\cal L}_{0,0}|=
\{p_{10},...,p_{d0}\}\,.$$ Furthermore, one can state the following proposition, which appears as a weak converse of proposition 5.1 in the affine setting.
Let $\gamma_1$,...,$\gamma_d$ be $d$ disjoint pieces of $n-m$-dimensional analytic manifold ($1\leq m <n$) in the affine space $\C^n$. Suppose that for any $(\alpha,\beta)\in (\C^{m})^{n-m} \times \C^{n-m}$ close to $(0,0)$, the affine $m$-dimensional subspace $$L_{\alpha,\beta}:=\Big\{\zeta \in \C^n\,;\, \zeta_k=\sum\limits_{r=1}^{m}
\, \alpha_{k,r} \zeta_{n-m+r} + \beta_k\,,
\ k=1,...,n-m\Big\}$$ intersects transversally $\gamma_1$,...,$\gamma_d$ respectively at points $p_1(\alpha,\beta),...,p_d(\alpha,\beta)$. Assume that there exists $D\in \N$ and analytic functions $h_{\underline k}^{(l)}$, $|\underline k|\leq D+1$, $l=n-m+1,...,n$, in a neighborhood of $0$ in $(\C^{m})^{n-m}$ such that for $(\alpha,\beta)$ close to $(0,0)$ in $(\C^{m})^{n-m} \times \C^{n-m}$, for any $l=n-m+1,...,n$, $$\begin{aligned}
\sum\limits_{j=1}^d \zeta_l [p_j(\alpha,\beta)]=
\sum\limits_{{\underline k\in \N^{n-m}}\atop {|\underline k|
\leq D+1}}
h_{\underline k}^{(l)} (\alpha)\, \beta_1^{k_1}\cdots \beta_{q-m}^{k_{n-m}}\, .\end{aligned}$$ Then, one can find a collection of polynomials $(P_\iota)_{\iota \in {\cal J}}$ with degree at most $d+D$ which define an affine algebraic variety $V(P)$ such that for some convenient constants $\epsilon>0$, $\kappa>0$, $K>0$, one has :
- if $$\Gamma_\epsilon:=
\bigcup\limits_{{(\alpha,\beta) \in (\C^{m})^{n-m}\times \C^{n-m}}\atop
{\max (\|\alpha\|,\|\beta\|) <\epsilon}} L_{\alpha,\beta}\,,$$ then $$\begin{aligned}
\zeta \in \Gamma_\epsilon\,,\ \|\zeta\| \geq K \ \Longrightarrow
\max\limits_{\iota \in {\cal J}} |P_\iota (\zeta)| \geq \kappa \|\zeta\|^d\,; \end{aligned}$$
- for $\max (\|\alpha\|,\|\beta\|) <\epsilon$, one has $$\begin{aligned}
L_{\alpha,\beta} \cap V(P)=\{p_1 (\alpha,\beta),...,p_d(\alpha,\beta)\}\, .\end{aligned}$$
Before the proof of this proposition we remark that our approach is directly inspired from [@wood1:gnus] (page 237, proof of the sufficiency). First we observe, as in Wood’s argument, that conditions (5.11) imply that for any integer $\sigma\in \N^*$, for any $l=n-m+1,...,n$, one has, for $(\alpha,\beta)$ close to $(0,0)$ in $(\C^{m})^{n-m} \times \C^{n-m}$, $$\begin{aligned}
\sum\limits_{j=1}^d (\zeta_l[p_j(\alpha,\beta)])^\sigma
=\sum\limits_{{\underline k\in \N^{n-m}}\atop {|\underline k|
\leq D+\sigma}}
h_{\sigma,\underline k}^{(l)} (\alpha)\, \beta_1^{k_1}\cdots \beta_{n-m}^{k_{n-m}}\,,\end{aligned}$$ where the $h_{\sigma,\underline k}^{(l)}$ are analytic functions in $\alpha$ in a neighborhood of $0$.\
Actually, if $\gamma_j$ is defined semi-locally as the smooth complete intersection $$\gamma_j= \{\zeta\in \C^n\,:\, \Phi_{j,1}(\zeta)=\cdots=\Phi_{j,m}(\zeta)=0\}\,,$$ then for any $j=1,...,d$ and for any $k=1,...,n-m$, $$\begin{aligned}
\zeta_k (p_j(\alpha,\beta))&=&{\rm Res} \left[\matrix{\varphi_j \, \zeta_k\,
d\Phi_j \wedge d\Lambda_{\alpha,\beta}
\cr \Phi_{j,1},...,\Phi_{j,m},\Lambda_{\alpha,\beta,1},..., \Lambda_{\alpha,\beta, n-m}}
\right]\\
&=&
{\rm Res} \left[\matrix{\varphi_j \, \Big(\sum\limits_{r=1}^{m} \alpha_{k,r}
\zeta_{n-m+r}+\beta_k\Big)\,
d\Phi_j \wedge d\Lambda_{\alpha,\beta}
\cr \Phi_{j,1},...,\Phi_{j,m},\Lambda_{\alpha,\beta,1},..., \Lambda_{\alpha,\beta, n-m}}
\right]\,,\end{aligned}$$ where $$\Lambda_{\alpha,\beta,k}:=\zeta_k-\sum\limits_{r=1}^m
\alpha_{k,r}\, \zeta_{n-m+r} -\beta_k\,,\quad k=1,...,n-m\,,$$ $$d\Phi_j:= \bigwedge_{k=1}^{m} d\Phi_{j,k}\quad ,\quad d\Lambda_{\alpha,\beta}:=
\bigwedge_{k=1}^{n-m} d\Lambda_{\alpha,\beta,k}\,,$$ and $\varphi_j$ is a test-function which is supported by an arbitrary small neighborhood of $p_{j0}$ and is such that $\varphi\equiv 1$ near this point. This implies that one can also represent in fact any sum $\sum_j \zeta_l[p_j(\alpha,\beta)]$, $l=1,...,n$, in the form (5.11). Since we also have for any $j=1,...,d$, for any $\sigma \in \N^*$ and for any $k=1,...,n-m$ that $${\rm Res} \left[\matrix{\varphi_j \, \Big(\zeta_k-\sum\limits_{r=1}^{m} \alpha_{k,r}
\zeta_{n-m+r}-\beta_k \Big)^{\sigma}\,
d\Phi_j \wedge d\Lambda_{\alpha,\beta}
\cr \Phi_{j,1},...,\Phi_{j,m},\Lambda_{\alpha,\beta,1},..., \Lambda_{\alpha,\beta, n-m}}
\right] =0\,,$$ it follows by induction on $\sigma$ that (5.14) holds in a neighborhood of $(0,0)$ for any $l=1,...,n$, in particular for any $l=n-m+1,...,n$ (here we use the fact that one can choose $\alpha$ generic in a neighborhood of the origin in $(\C^m)^{n-m}$). At this point we return to the proof of the Proposition 5.2.\
[**Proof of Proposition 5.2.**]{} For any $l=n-m+1,...,n$, let $A_l$ be the polynomial in variable $X_l$ (with coefficients analytic in $\alpha$ and polynomial in $\beta$) defined as $$\begin{aligned}
A_l(X_l,\alpha\,;\, \beta) &=& \prod\limits_{j=1}^d (X_l-\zeta_l[p_j(\alpha,\beta)]) \nonumber \\
&=& X_l^d - A_{l1}(\alpha,\beta) X_l^{d-1} + \cdots + (-1)^d A_{ld}(\alpha,\beta) \nonumber \\\end{aligned}$$ ($\alpha$ and $\beta$ close to $0$ in their respective spaces). For any such $\alpha$, denote as $P_{l,\alpha}$ the element in $\C[X_1,...,X_n]$ defined as $$P_{l,\alpha} (X)=A_l\Big(X_l,\alpha\,;\, X_{1}-\sum\limits_{
r=1}^{m} \alpha_{1,r} X_{n-m+r}, ..., X_{n-m}-\sum\limits_{r=1}^{m}
\alpha_{n-m,r} X_{n-m+r}\Big)\, .$$ For each $\alpha$ close to $0$, $P_{l,\alpha}$ is a polynomial in variables $(X_1,...,X_n)$ with total degree less than $d+D$. If $\zeta$ is a point in $\gamma_j$, then one can write, for any $\alpha$ close to $0$ in $(\C^{m})^{n-m}$, for any $l=n-m+1,...,n$, $$\zeta_l= \zeta_l [p_j(\alpha,\beta_\zeta)]\,,$$ where $$\beta_{\zeta,k}=\zeta_k -\sum\limits_{r=1}^{m} \alpha_{k,r}\, \zeta_{n-m+r}\,;$$ then, it follows from the definition (5.15) of $A_l$, hence of $P_{l,\alpha}$, that $\zeta \in V(P_{l,\alpha})$ for any $l=n-m+1,...,n$ and for any $\alpha$ close to zero. All pieces of manifold $\gamma_1,...,\gamma_d$ lie in $V(P_{n-m+1,\alpha},...,P_{n,\alpha})$ for any $\alpha$ close to $0$ in $(\C^{m})^{n-m}$. Moreover, if $\epsilon$ is sufficiently small and $K=K_\epsilon>0$ large enough, then for any $\alpha$ such that $\|\alpha\|<\epsilon$, define the strip $$\Gamma_{\alpha,\epsilon}:= \{\zeta \in \C\,:\, \|\zeta\|\geq K\,,\,
\max\limits_{k=1,...,n-m}
\Big| \zeta_k -\sum\limits_{r=1}^{m} \alpha_{k,r}\, \zeta_{n-m+r} \Big|
<\epsilon\}\; .$$ For any $\zeta \in \Gamma_{\alpha,\epsilon} $ one has that $$\begin{aligned}
\max\limits_{l=n-m+1,...,n} |P_{l,\alpha} (\zeta)|=
\max\limits_{l=n-m+1,...,n}
\prod\limits_{j=1}^d \Big|\zeta_l-
\zeta_l[p_j(\alpha, \beta_\zeta)]\Big| \geq \kappa \|\zeta\|^d \,, \nonumber \\\end{aligned}$$ where $\kappa=\kappa_\epsilon$ is a strictly positive constant (independent of $\alpha$). Let now ${\cal F}$ be the finite subset in $L_{0,0}$ defined as $$\zeta \in {\cal F} \Longleftrightarrow
\forall \, l=n-m+1,...,n\,,\
\exists j\in \{1,...,d\}\,,\ \zeta_l=\zeta_l[p_{j0}]$$ and ${\cal F}':={\cal F}\setminus \{p_{10},...,p_{d0}\}$. There exists an affine form $\Lambda$ in $X_{n-m+1},...,X_n$ such that for any $\zeta \in {\cal F}'$, $$\Lambda(\zeta) \not= \Lambda (p_{j0}), \ j=1,...,d \,.$$ We can define $$B(X_{n-m+1},...,X_n\,;\, \beta) = \prod\limits_{j=1}^d
(\Lambda(X_{n-m+1},...,X_n))-\Lambda[p_j(0,\beta)])$$ and $Q(X)=B(X_{n-m+1},...,X_n\,;\, X_1,...,X_{n-m})$. One can check as before that $V(Q)$ contains the pieces $\gamma_1,...,\gamma_q$. Moreover $Q$ does not vanish at any point in ${\cal F}'$, and $\deg Q \leq d+D$. We now define the family $(P_{\iota})_{\iota \in {\cal J}}$ as the collection of all polynomials $P_{l,\alpha}$, $\|\alpha\|<\epsilon$, $l=n-m+1,...,n$, and the polynomial $Q$. It is clear that $(5.12)$ holds (since it already holds for the collection $(P_{l,\alpha})$, $\|\alpha \|<\epsilon$, $l=n-m+1,...,n$). Since the pieces $\gamma_1,...,\gamma_d$ lie in $V(P)$, one has, for any $(\alpha,\beta)$ such that $\max (\|\alpha\|, \|\beta\|)<\epsilon$, $$\{p_1(\alpha,\beta),...,p_d(\alpha,\beta) \}
\subset L_{\alpha,\beta} \cap V(P)\, .$$ Since points in $V(P) \cap \Gamma_\epsilon$ lie in the compact set $\{\zeta\in \C^n\,:\, \|\zeta\| \leq K\}$ (because of $(5.12)$), points in $L_{\alpha,\beta} \cap V(P)$ uniformly approach points in $L_{0,0}\cap V(P)$ when $(\alpha,\beta)$ tends to $(0,0)$. Since $Q\not=0$ on ${\cal F}'$ and $$V(P_{n-m+1,0},...,P_{n,0})\cap L_{0,0}
\subset {\cal F}\,,$$ one has that $V(P)\cap L_{0,0}=\{p_{10},...,p_{d0}\}$. Therefore, if we refine the choice of $\epsilon$, we can assume that (5.13) holds. This concludes the proof of our proposition. $\quad\diamondsuit$ In the particular case $m=1$, one can be more precise and repeat Wood’s argument in order to obtain the :
Let $\gamma_1$,...,$\gamma_d$ be $d$ disjoint pieces of smooth analytic hypersurface in the affine space $\C^n$. Suppose that for any $(\alpha,\beta)\in (\C^{n-1})^2$ close to $(0,0)$, the affine line $$L_{\alpha,\beta}:=\Big\{\zeta \in \C^n\,;\, \zeta_k=\alpha_k \zeta_n + \beta_k\,,
\ k=1,...,n-1\Big\}$$ intersects transversally $\gamma_1$,...,$\gamma_d$ respectively at points $p_1(\alpha,\beta),...,p_d(\alpha,\beta)$. Assume that there exists $D\in \N$ and analytic functions $h_{\underline k}$, $|\underline k|\leq D+1$, in a neighborhood of $0$ in $\C^{n-1}$ such that for $(\alpha,\beta)$ close to $(0,0)$ in $(\C^{n-1})^2$, $$\begin{aligned}
\sum\limits_{j=1}^d \zeta_n [p_j(\alpha,\beta)]=
\sum\limits_{{\underline k\in \N^{n-m}}\atop {|\underline k|
\leq D+1}}
h_{\underline k} (\alpha)\, \beta_1^{k_1}\cdots \beta_{q-m}^{k_{n-m}}\,. \end{aligned}$$ (one from the $h_{\underline k}$ for $|\underline k|=D+1$ being non identically zero). Then, one can find a polynomial $P$ with degree $d+D$ which defines an affine algebraic variety $V(P)$ such that for some convenient constants $\epsilon>0$, $\kappa>0$, $K>0$, one has :
- if $$\Gamma_\epsilon:=
\bigcup\limits_{{(\alpha,\beta) \in (\C^{m})^{n-m}\times \C^{n-m}}\atop
{\max (\|\alpha\|,\|\beta\|) <\epsilon}} L_{\alpha,\beta}\,,$$ then $$\begin{aligned}
\zeta \in \Gamma_\epsilon\,,\ \|\zeta\| \geq K \ \Longrightarrow
|P (\zeta)| \geq \kappa \|\zeta\|^d\,; \end{aligned}$$
- for $\max (\|\alpha\|,\|\beta\|) <\epsilon$, one has $$\begin{aligned}
L_{\alpha,\beta} \cap V(P)=\{p_1 (\alpha,\beta),...,p_d(\alpha,\beta)\}\,.\end{aligned}$$
[**Proof.**]{} We repeat here Wood’s argument (as we did in the proof of proposition 5.2). For $(\alpha,\beta)$ close to $(0,0)$, the points $p_1(\alpha,\beta),...,p_d(\alpha,\beta)$ are the only intersection points (in the affine space $\C^n$) between the affine line $L_{\alpha,\beta}$ and the affine algebraic hypersurface (with exact degree $d+D$ for $\alpha$ generic) $\{P_{n,\alpha}=0\}$. Since these intersection points are simple (the line hits each $\gamma_j$ transversally), the homogeneous polynomial ${}^ h P_{n,\alpha}$ vanishes at the order at most $D$ at the point $p_{\alpha,\infty}$ at infinity on the projective line $|{\cal L}_{\alpha,0}|$. On the other hand, one has $$|P_{n,\alpha} (\zeta)| \geq \kappa \|\zeta\|^d$$ in a tube $\Gamma_{\alpha,\epsilon}$ along the line $L_{\alpha,0}$ (see (5.16) in the proof of proposition 5.2). This implies that the the homogeneous polynomial ${}^h P_{n,\alpha}$ vanishes at the order at least $D$ at the point $p_{\alpha,\infty}$. Therefore, the hypersurface $\{P_{n,\alpha}=0\}$ (with degree $d+D$) contains the germs $\gamma_1,...,\gamma_d$ (as simple germs) and the germ corresponding to the hyperplane ${\cal H}_\infty$ (this germ being counted with multiplicity $D$). This, combined with the fact that the degree of $P_{n,\alpha}$ is exactly $d+D$, implies that all the polynomials $P_{n,\alpha}$ define the same algebraic hypersurface $H$ in $\C^n$. Since they have the same degree, they are equal (up to some constant) to a polynomial $P$. As the auxiliary construction of the polynomial $Q$ is not needed in the hypersurface case, proposition 5.3 follows immediately from proposition 5.2. $\quad\diamondsuit$ [**Remark.**]{} In the particular case $W={\bf A}^n_\C$, $m=1$, proposition 5.3 appears as the reciprocal assertion to proposition 5.1. The difficulty in the more general case $W={\bf A}^n_\C$, $m>1$, is to be able to interpolate the germs $\gamma_1,...,\gamma_d$ by an algebraic complete intersection $V(P_1,...,P_m)$. It does not seem possible when $m>1$ even if conditions (5.11) are satisfied with $D=0$ (which would mean that the projective variety $\{{}^h P_1=...={}^h P_m=0\}$ corresponding to the complete intersection $V(P)$ that interpolates the pieces $\gamma_j$ does not hit $|{\cal H}_\infty| \cap |{\cal L}_{0,0}|$). We do not have the answer to that question yet. Nevertheless, proposition 5.2 can be seen as an attempt to settle a converse to proposition 5.1 in general.
[99]{} M.F. Atiyah, Resolution of singularities and division of distributions, Comm. Pure. Appl. Math. 23, 1970, 145-150. G. Biernat, [La représentation paramétrique d’un résidu multidimensionel]{}, Rev. Roumaine Math. Pures and Appl. 36, 1991, no. 5-6, 207-211. J.E. Björk, [Residue currents and ${\cal D}$-modules on complex manifolds]{}, preprint, 1996. J.E. Björk, [*Analytic ${\cal D}$-modules and applications*]{}, Kluwer Science Publications, 1993. C. A. Berenstein, A. Yger, [A formula of Jacobi and its consequences]{}, Annales de l’École Normale Supérieure, 24, 1991, 73-83. C. A. Berenstein, A. Yger, Green currents and analytic continuation, Journal d’analyse mathématique, 75, 1990, 1-50. C. A. Berenstein, A. Yger, [Analytic residue theory in non complete intersection case]{}. J. reine angew. Math. 527, 2000, 203-235. J. Chadzynski, T. Krasinski, [A set on which the Lojasiewicz exponent at infinity is attained]{}, Ann. Pol. Nath. LXVII. 2, 1997, p. 191-197. N. Coleff, M. Herrera, [*Les courants résidus associés à une forme méromorphe* ]{}, Lect. Notes in Math. 633, Springer-Verlag, New York, 1978. (N.S.) 33, 1996, no. 3, 295-324. D. Eisenbud, M. Green, J. Harris, [Cayley-Bacharach theorems and conjectures]{}, Bull. Amer. Math. Soc. (N.S.) 33, 1996, no. 3, 295-324. E. Fortuna, S. Lojasiewicz [Sur l’algébricité des ensembles analytiques complexes]{}, J. reine angew. Math. 329, 1981, 215-220. P. Griffiths, J. Harris, [*Principles of algebraic geometry*]{}, Wiley Interscience, New York, 1978. R. Hübl, E. Kunz, Integration of differential forms on schemes, J. reine. angew. Math. 410, 1990, 53-83. R. Hübl, [Residues of regular and meromorphic differential forms]{}, Math. Ann. 300, 1994, 605-628. R. Hübl, E. Kunz, [On the intersection of algebraic curves and hypersurfaces]{}, Math. Z., 1998, 263-278. C.G.J. Jacobi, De relationibus, quae locum habere denent inter puncta intersectionis duarum curvarum vel trium superficierum algebraicarum dati ordinis, simul cum enodatione paradoxi algebraici, Gesammelte Werke, Band III, 329-354. E. Kunz, [*Kähler differentials* ]{}, Advanced topics in Mathematics, Braunschweig, Wiesbaden 1986. E. Kunz, [Über den n-dimensionalen Residuensatz]{}, Jahresber. DMV 94, 1992, 170-188. P. Lelong, [*Fonctions plurisousharmoniques et formes différentielles positives*]{}, Gordon and Beach, Paris-London-New York, 1968. J. Lipman, [*Residues and Traces of differential forms via Hochschild homology*]{}, Contemporary Math. 61, 1987, Amer. Math. Society. J. Lipman, [*Dualizing sheaves, differentials and residues on algebraic varieties*]{}, Astérisque 117, 1984. M. Méo, Résidus dans le cas non nécessairement intersection complète, C. R. Acad. Sci. Paris, t. 333, Série I, 2001, 33-38. M. Passare, A. Tsikh, A. Yger, [Residue currents of the Bochner-Martinelli type]{}, Publicacions Matemàtiques, 44, 2000, 85-117. W. Rudin, A geometric criterion for algebraic varieties, J. Math. Mech. 17, 1968, 671-683. J.P. Serre, [*Groupes algébriques et corps des classes*]{}, Hermann, Paris, 1969. M. Spivak, [*Introduction to differential Geometry*]{}, vol 1. Publish or Perish Press, 1980 W. Stoll, [The growth area of a transcendental analytic set I, II]{}, Math. Annal. 156, 1964, 144-170. A. Vidras, A. Yger, [On some generalizations of Jacobi’s Residue Formula]{}, Annales Scien. de l’École Normale Supérieure, 34, 2001, 131-157. A. Yger, [Lectures at Croce di Magara, June 1998]{}, informal notes. A. Yger, [Résidus, courants résiduels et courants de Green]{}, in [*Géométrie complexe*]{}, F. Norguet, S. Ofman, J. J. Szczeciniarz ed., Actualités Scientifiques et Industrielles 1438, Hermann, 1996. J. A. Wood, [A simple criterion for local hypersurfaces to be algebraic]{}, Duke Mathematical Journal, 51, 1, 1984, 235-237. H. Zhang, Formules de Jacobi et méthodes analytiques. preprint, 2000.
[^1]: AMS classification number: 32A27, (32A25, 32C30)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In Delay Tolerant Networks (DTNs), two-hop routing compromises energy versus delay more conveniently than epidemic routing. Literature provides comprehensive results on optimal routing policies for mobile nodes with homogeneous mobility, often neglecting signaling costs. Routing policies are customarily computed by means of fluid approximation techniques, which assure solutions to be optimal only when the number of nodes is infinite, while they provide a coarse approximation otherwise. This work addresses heterogeneous mobility patterns and multiple wireless transmission technologies; moreover, we explicitly consider the beaconing/signaling costs to support routing and the possibility for nodes to discard packets after a local time. We theoretically characterize the optimal policies by deriving their formal properties. Such analysis is leveraged to define two algorithmic approaches which allow to trade off optimality with computational efficiency. Theoretical bounds on the approximation guarantees of the proposed algorithms are derived. We then experimentally evaluated them in realistic scenarios of multi-class DTNs.'
author:
-
bibliography:
- 'citations.bib'
title: 'Beaconing-Aware Optimal Policies for Two-Hop Routing in Multi-Class Delay Tolerant Networks'
---
Delay Tolerant Networks, 2-Hop Optimal Routing, Approximation Algorithms.
Introduction
============
Delay–tolerant networks (DTNs) are sparse and/or highly mobile wireless ad hoc networks which assure no continuous connectivity. Examples of such networks are those operating in mobile or hash terrestrial environments, or interplanetary networks. Disruption may occur because of the limits of wireless radio range, sparsity of mobile nodes, energy resources, attacks, and noise. One central problem in DTNs is the routing of packets from a source towards the desired destinations. When no *a priori* information is available over the mobility pattern of the nodes, a common technique for overcoming lack of connectivity is to disseminate multiple copies of the packet in the network: this enhances the probability that at least one of them will reach the destination node within a given temporal deadline. This is referred to as *epidemic–style* forwarding, because, alike the spread of infectious diseases, each time a packet–carrying node encounters a new node not having a copy thereof, the carrier may infect this new node by passing on a packet copy; newly infected nodes, in turn, may behave similarly. The destination receives the packet when it meets an infected node. However, epidemic routing is very energy consuming and a convenient compromise of energy versus delay compared to epidemic routing is provided by two-hop routing [@DBLP:journals/pe/GroeneveltNK05] where the infection is limited at the contacts between the source node and intermediary nodes, that is, the source node passes on the packets to be transmitted to all the mobile node she encounters (provided that these last ones do not already have a copy of the packet in their local buffer), and the “infected” mobile node can deliver the packets they are carrying only to the final destination.
We target here the design of optimal two-hop routing policies for DTNs as in [@DBLP:journals/ton/AltmanABP13; @DBLP:journals/tac/AltmanBP11; @DBLP:conf/wiopt/PellegriniAB10]. In two-hop routing, a routing policy controls the decisions taken by the source node to forward/not to forward a given packet to a given mobile node she is encountering at a given time instant. In this work, we further consider network scenarios where mobile nodes are categorized in district multiple classes which may capture different mobility patterns and different available communication technologies onboard and we study the “shape” of optimal routing policies under heterogeneous mix of mobile nodes classes.
In the seminal work [@DBLP:conf/infocom/AltmanNPM09], the authors study optimal static and dynamic control (proved to be threshold based) policies for two-hop DTN when mobile nodes are homogeneous. In this case, optimal policies can be found in closed form. Furthermore, the authors show that when the parameters are unknown it is still possible to obtain a policy that converges to the optimal one by using some adaptive auto–tuning mechanism. Extensions of such adaptive mechanism are proposed in [@DBLP:journals/comcom/GuerrieriCPMM10]. Heterogeneous scenarios are studied in [@DBLP:conf/wiopt/PellegriniAB10] where mobile nodes belong to multiple distinct classes and the routing policy may be class-dependent; the authors resort to fluid approximation to characterize the optimal routing policies; however, even if the proposed approach performs well when the number of mobile nodes goes to infinity, fluid approximation may lead to coarse solutions for finite numbers of mobile nodes (e.g., in other fields in [@DBLP:journals/ior/PerryW11; @DBLP:journals/orl/Randhawa13], fluid approximation is shown to perform well when the number of users is over 200). Optimal control techniques are further proposed under the assumption that the number of copies of the packet is monotonically increasing in time. Along the same lines, in [@DBLP:conf/wd/ChahinAPA11], the authors provide the closed form structure of the controller and provide stochastic algorithms for the learning of the parameters. Timers have been proposed to be associated with messages when stored at mobile nodes, so that after a given threshold the message is discarded. Their optimization under fluid approximation is discussed in [@DBLP:journals/tac/AltmanBP11]. In [@DBLP:journals/tsmc/TembineAAH10] and in [@DBLP:journals/ton/BanerjeeCL10], the authors optimize network performance by designing the mobility of message relays. Evolutionary game theory has been adopted in [@DBLP:journals/cn/AzouziPSK13] to incentivize mobile nodes. Many other related problems have been studied. In [@DBLP:journals/twc/AltmanSP13] the authors deal with the problem of transferring large files from source to destination when all packets are not available at the source prior to the first transmission. In [@DBLP:journals/ton/AltmanP11], a class of replication mechanisms for packets in the network that include coding in order to improve the probability of successful delivery within a given time limit. In [@el2010evolutionary], nodes are considered selfish/non–cooperative and the problem is to incentivize the mobile nodes to forward the packet. The first node that delivers the packet receives a unitary reward. The only cost of the mobile nodes is due to beaconing. The authors study the best strategies by using evolutionary game theory tools. In [@chen2010mobicent], the authors focus on multi–hop routing scenarios in the attempt to design incentives for the mobile nodes to avoid that these nodes behave strategically in terms of edge hiding and edge insertion. The proposed model works only when beaconing costs are not present and each mobile node is subject to the same cost. A similar technological, but not technical, approach is proposed in [@zhu2009smart]. In [@jain2004routing], the authors study different routing strategies when the nodes have a limited buffer and provide an experimental comparison of the strategies. A similar work is presented in [@li2010routing] except that here the authors introduce social information about the nodes. Cooperative aspects are treated in [@ning2011incentive] and [@niyato2010coalition].
Differently than the aforementioned reference literature, (i) we explicitly include in the definition of the optimal routing policy the cost for beaconing messages exchanged by mobile nodes to support packet forwarding; (ii) we allow mobile nodes to discard the packets they are carrying upon expiration of a local temporal deadline (the number of copies of a packet is thus no-longer monotonically as customarily assumed in the reference literature); (iii) we do not resort to fluid approximation, but rather propose optimization algorithms for finite mobile node populations; in details, we introduce an algorithm to find optimal routing policies running in exponential time, as well as approximate polynomial-time algorithms when the number of nodes is finite; in the latter case, formal approximation theoretical bounds are also derived. Finally, we provide a thorough experimental evaluation with realistic settings of the proposed algorithms in terms of approximation ratio, scalability in the number of classes, further evaluating the impact of network parameters onto the optimal routing policies.
Problem statement
=================
Scenario
--------
We consider an environment populated by one [*source*]{} node, one [*sink*]{} node and multiple [*mobile nodes*]{}. *Sink* and *source* node may as well be mobile. A packet is initially held by the source and it must be delivered to the sink no later than time $\tau$. We consider a discrete time representation developing in slots of fixed duration $\Delta$ and we denote the total number of useful time slots as $K = \lfloor{\tau/\Delta}\rfloor$, where the $k$–th time slot corresponds to the time interval $[k \Delta, (k+1)\Delta)$. Mobile nodes are divided according to a set of classes $C$. Each class encodes the features of its nodes, including their [*mobility profile*]{} and [*transmission technology*]{}. In particular, given a class $c \in C$, $N_c$ indicates the total number of mobile nodes belonging to that class; $t_c$ indicates the [*local*]{} time to live, namely the amount of time for which each mobile node of class $c$ will keep a local copy of the packet before discarding it and not accepting it again in the future. The mobility profile of a node of class $c$ is given by the average speed $v_c$. Finally, we describe the transmission technology for class $c$ with the following parameters:
- the communication range for class $c$ is denoted by $R_c$;
- the *beaconing cost* $\beta_c$ captures the energy consumed for the connection control and signaling procedures to support packet transmissions in class $c$,; as an example, the beaconing cost may capture the energy consumed to send/receive beaconing messages to discover nearby mobile nodes;
- the *transmission cost* $\rho_c$ is the energy consumed to transmit a packet to a recipient in class $c$.
Different classes $c$ can have the same technology, e.g., WiFi, sharing beaconing costs. We denote by $\omega\in \Omega$ a technology, and we denote by $C_{\omega}\subseteq C$ the subset of classes adopting technology $\omega$. With overriding of notation, we use indistinguishably $\beta_\omega$ and $\beta_c$ when $c\in C_\omega$. Table \[tab:class\_parameters\] summarizes all the parameters that define a generic class $c \in C$.
$N_c$ number of nodes in class $c$
----------- ------------------------------
$t_c$ packet’s local time to live
$v_c$ average speed
$R_c$ communication range
$\beta_c$ beaconing cost
$\rho_c$ transmission cost
: Set of parameters for a generic class $c \in C$
\[tab:class\_parameters\]
Transmission opportunities between two nodes are given by contacts taking place when each node is within the communication range of the other one. We restrict our setting to a 2–hop routing scheme, where each mobile node can receive the packet only from the source and can forward it only to the sink. Contacts at the source and at the sink are assumed to follow a multi–class Poisson law, where the arrival rate for nodes of class $c$ (either at the source or the sink) is denoted by $\lambda_c$ and computed according to [@DBLP:conf/infocom/AltmanNPM09]: $$\lambda_c=\frac{8wR_cv_c}{\pi L^{2}}$$ where $w$ is a constant set to $1.3693$ and $L$ is the radius of circle in which the nodes move.
When a contact is made between the source and a mobile node that did not receive the packet, the source forwards it according to a [*forwarding policy*]{} $\boldsymbol \mu$ which depends on the current time and the recipient’s class. Given a time slot $k$ and a class $c$, the policy profile at time $k$ is $\boldsymbol \mu(k) = (\mu_{1}(k),\ldots,\mu_{|C|}(k))$ where $\mu_c(k)$ indicates the forward probability at time slot $k$ for class $c$; we also denote with $\mu_c$ the entire policy for such class $c$. In general, when the packet is forwarded, some energy is spent and the packet’s delivery probability is increased. We denote with $F_D(\boldsymbol \mu, K)$ the probability of delivering the message before time $K \Delta$ given policy profile $\boldsymbol \mu$. Obviously, such value is prevented from growing indefinitely by a budget constraint. We call $\Psi$ the upper bound on the total spent energy (including both beaconing and transmissions).
Formal model
------------
We define $X_{c,k}(\boldsymbol \mu)$ as the random variable expressing the number of mobile nodes of class $c$ that have received the packet by time slot $k$, while $Y_{c,k}(\boldsymbol \mu)$ is a random variable expressing the number of mobile nodes of class $c$ that still keep a copy of the packet at time slot $k$. These variables both depend on $\boldsymbol\mu$ and are, in general, different. Indeed, since a mobile node can both receive and discard a packet before time slot $k$, we have that $Y_{c,k}(\boldsymbol \mu) \le X_{c,k}(\boldsymbol \mu)$. Furthermore, we denote by $Q_{c,k,k'}(\mu_c)$ the probability that a mobile node of class $c$ does not receive any packet in time slots $k,\ldots,k'$ as function of $\mu_c$. The expected number of mobile nodes of class $c$ that receive a packet in time slots $0,\ldots,k$ is: $$\mathbb{E}[X_{c,k}(\boldsymbol \mu)] = N_c \cdot (1- Q_{c,0,k}(\mu_c))$$ where $Q_{c,k,k'}(\mu_c) = e^{-\lambda_c \Delta \sum_{i=k}^{k'}\mu_c(i)}$. The expected number of mobile nodes of class $c$ that have the packet at time slot $k$ is: $$\mathbb{E}[Y_{c,k}(\mu_c)] = N_c \cdot (1- Q_{c,\max\{0,k-t_c\},k}(\mu_c))$$ The probability that a packet is delivered by $k\Delta$ to the sink is: $$F_D(\boldsymbol \mu,k) = 1 - \prod_{c\in C}\prod_{h=0}^{k-1} X_{c,h}^*(\lambda_c\Delta,\boldsymbol \mu)\label{eq:OF}$$ where $$X_{c,h}^*(s,\boldsymbol \mu) = \mathbb{E}[e^{-sY_{c,h}(\boldsymbol \mu)}]$$ The budget constraint is formulated as: $$\begin{gathered}
\sum_{c\in C} \rho_c N_c (1- Q_{c,0,K}(\mu_c))+ \\\sum_{\omega\in \Omega}\sum_{k=0}^{K-1}\beta_\omega\cdot \left( 1 - \prod_{c\in C_\omega}\Big(1-\mu_c(k)\Big)\right) \leq \Psi \label{equation:budget}\end{gathered}$$
The left term of the inequality adds up the expected transmission costs with the expected beaconing costs for class $c$, given a policy profile $\boldsymbol\mu$. In particular, transmission costs are obtained by multiplying $\rho_c$ by the expected number of nodes that will receive the packet from slot 0 to slot $K$; on the other side, a beaconing costs of $\beta_c$ is paid for each time slot $h$ with a probability of $\mu_c(h)$, i.e., when a packet transmission could be made.
We aim at finding the optimal policy $\boldsymbol \mu^*$ that maximizes $F_D(\boldsymbol \mu,K)$ without violating budget constraint (\[equation:budget\]).
Problem properties
------------------
We now show some theoretical properties that we shall exploit in addressing the optimization problem introduced above.
\[prop:saturate\_or\_saturate\] Optimal policies either completely consume the budget or prescribe that all the classes transmit for all the slots.\[property:budget\]
It is easy to see that $F_D(\boldsymbol \mu,K)$ is monotonically increasing in $\sum_{h=0}^{K-1} \mu_c(h)$ and that, as a consequence, transmitting for a larger (expected) number of slots cannot result in a lower delivery probability.
Similarly to what proposed in [@DBLP:conf/infocom/AltmanNPM09], we define a threshold–based policy $\mu_c$ as: $$\mu_c(k) =
\begin{cases}
1 & k<h_c \\
\alpha & k=h_c \\
0 & k>h_c
\end{cases}$$ where $\alpha\in [0,1)$. As in the single–class case, optimal policies are threshold based.
\[prope:threshold\_optimal\] Optimal policies are threshold based.
The delivery delay c.d.f. is $F_D(\boldsymbol\mu,K) = 1 - \prod_{c}\Gamma_c(s)$, where $\Gamma_c(s) = \prod_{h=0}^{K-1} X_{c,h}^*(s)$ and $s = \lambda_c\Delta$. Let us denote with $\mu_c$ a *non–threshold* policy for class $c$ and with $\hat{\mu}_c$ a policy obtained by shifting to the left all the non-empty slots of $\mu_c$ and by rounding them so that $\hat{\mu}_c$ matches the definition of threshold policy introduced above. For any $(\mu_c, \hat{\mu}_c)$ obtained in this way we have that $\Gamma_c(s) \ge \hat{\Gamma}_c(s)$ and therefore $$1 - \Gamma_c(s)\cdot \Gamma_{-c}(s) \le 1-\hat{\Gamma}_c(s)\cdot \Gamma_{-c}(s)$$ that is, for any given joint policy $\boldsymbol \mu$, if we substitute the marginal policy of a class $c$ with its threshold version the probability of delivery within $K$ time steps will not decrease.
\[prop:all\_fractional\] Optimal policies can prescribe non–integer thresholds for all the classes.
Consider, for instance, a two–class instance with: $K=20$, $\Delta=100$, $\Psi=0.7$, $N_1=1$, $N_2=2$, $\lambda_1=21\times10^{-5}$, $\lambda_2=20\times10^{-5}, t_1=t_2=K$. We approximated the optimal policy profile by discretizing the values of $h_c$ with a fine grid with step $0.01$. In addition to these points, we considered all the points in which the threshold of one class is integer and the threshold of the other class is calculated in such a way the budget is completely consumed. We evaluated the objective function at all these points and select the maximum. The approximately optimal policy is $h_1=7.87$, $h_2=15.99$. Hence, at the optimum, a fractional part is assigned to both classes.
The optimization problem with objective function Eq. \[eq:OF\] and constraint Eq. \[equation:budget\] is nonlinear and nonconvex.
Nonlinearity is trivial. Nonconvexity is proved by showing the nonconvexity of the feasibility region by computing the Hessian matrix of the budget constraint (\[equation:budget\]) to which we will refer here with $u$ (notice that we restrict our attention on threshold policies from now). Hessian matrix $\mathcal{H}$ is:
The above properties show that the optimization problem is hard. In particular, the adoption of non–convex programming techniques required by the nature of the problem cannot assure to find of global optimal solutions. For these reasons, we focus on the problem of developing approximation algorithms and of studying their theoretical and empirical approximation errors.
Approximation algorithms
========================
In this section, we introduce two approaches to compute the optimal (threshold) policy. We provide the formalization of two approximation algorithms and we discuss their guarantees on the solution quality loss.
Non–polynomial–time approximation scheme {#subsection:gridalgorithm}
----------------------------------------
We start by defining an approximation scheme that does not run in polynomial time, but for which optimality losses can be arbitrarily bounded. We overconstrain the optimization problem, allowing only a single class to have a fractional threshold in its associated policy. This additional constraint is likely to introduce worsenings in the solution quality (see Property \[prop:all\_fractional\]) but it allows us to provide a combinatorial version of the optimization problem. Indeed, once all the classes except one have been assigned integer policies, the potentially fractional policy of the remaining class is univocally determined either by the policy that consumes all the remaining budget or the one that transmits until the last useful time slot (see Property \[prop:saturate\_or\_saturate\]). In addition, we split each slot of length $\Delta$ in sub–slots of length $\epsilon\Delta$ where, for simplicity, $\frac{1}{\epsilon}\in \mathbb{N}$.
This overconstrained problem can be solved optimally by using an enumeration algorithm: we enumerate all the feasible threshold policies and we select the best one (see Property \[prope:threshold\_optimal\]). We report in Algorithm \[alg:gridepsilon\] the necessary steps. At Step 1, the algorithm initializes $F^*$ to be zero. If it is not possible to entirely consume the budget, then the optimal policy profile is to assign $h_c = (K-1)/\epsilon$ to each class $c$ (Step 2–3). Otherwise, the algorithm enumerates all the classes $c$, and for each class $c$ it enumerates all the policy profiles $\boldsymbol\mu = (\mu_{c},\boldsymbol\mu_{-c})$ s.t. $\boldsymbol\mu_{-c}$ is integer and budget $\Psi$ is entirely consumed (Step 5–6). Finally, we keep trace of the best policy found so far.
$F^* \gets 0$
$\boldsymbol \mu^*=\boldsymbol \mu$
$\boldsymbol\mu^* \gets \boldsymbol\mu$
$F^* \gets F_D(\frac{K-1}{\epsilon},\boldsymbol\mu)$
We now describe an efficient method to enumerate all and only the feasible policy profiles $\boldsymbol\mu = (\mu_{c},\boldsymbol\mu_{-c})$ s.t. $\boldsymbol\mu_{-c}$ is integer and budget $\Psi$ is entirely consumed. First, we build a lexicographic order over $C_{-c}= C\setminus c$ and we scan lexicographically the classes in $C_{-c}$. Then, for each $c'\in C_{-c}$ we determine the range of feasible values for $h_{c'}$ on the basis of the policies assigned to the classes $c''\succ c'$ as follows: $$\begin{gathered}
I_{c'}(\boldsymbol\mu_{-c'})=\\\left\{ \max\left\{0, \lceil r_{c'}(\overline{\boldsymbol \mu}_{-c'})\rceil\right\}, \ldots, \min\left\{\frac{K-1}{\epsilon}, \lfloor r_{c'}(\underline{\boldsymbol \mu}_{-c'})\rfloor\right\} \right\}\end{gathered}$$ where $r_{c'}(\boldsymbol\mu_{-c'})$ is computed as follows:
- initially we compute: $$\begin{gathered}
r_{c'}(\boldsymbol\mu_{-c'})=-\frac{\log\left(\frac{N_{c'}+A-\frac{\Psi}{\rho}+\underset{c'':c''\neq c'}{\max}\{\frac{\beta}{\rho} h_{c''}\}}{N_{c'}}\right)}{\lambda_{c'}\Delta\epsilon}\end{gathered}$$ where $$\begin{gathered}
A=\sum\limits_{c'':c''\neq c'}N_{c''}\left(1-e^{-\lambda_{c''}\Delta\epsilon \sum\limits_{k=0}^{\frac{K-1}{\epsilon}}\mu_{c''}(k)}\right)\end{gathered}$$
- if $r_{c'}(\boldsymbol\mu_{-c'})>\underset{c'':c''\neq c'}{\max}\{\frac{\beta\epsilon}{\rho} h_{c''}\}$, then $r_{c'}(\boldsymbol\mu_{-c'})$ is the solution (that can be approximated by means of Netwon algorithm) of the following equation: $$\begin{gathered}
N_{c'}(1-e^{-\lambda_c\epsilon\Delta r_{c'}(\boldsymbol\mu_{-c'})})\\+\sum\limits_{c'':c''\neq c'}N_{c''}\left(1-e^{-\lambda_{c''}\Delta\epsilon \sum\limits_{k=0}^{\frac{K-1}{\epsilon}}\mu_{c''}(k)}\right)\\-\frac{\Psi}{\rho}+\frac{\beta\epsilon}{\rho}r_{c'}(\boldsymbol\mu_{-c'})=0\end{gathered}$$
and where $\overline{\boldsymbol \mu}_{-c'}$ and $\qquad\underline{\boldsymbol \mu}_{-c'}$ are defined in the following way: $$\overline{\boldsymbol \mu}_{-c'}=
\begin{cases}
\overline{\mu}_{c''} = \mu_{c''} & c'' \succ c' \\
\overline{\mu}_{c''} : h_{c''}=\frac{K}{\epsilon}-1 & c' \succ c'' \\
\overline{\mu}_{c} : h_{c}=\frac{K}{\epsilon}-1
\end{cases}$$ $$\qquad\underline{\boldsymbol \mu}_{-c'}=
\begin{cases}
\underline{\mu}_{c''} = \mu_{c''} & c'' \succ c' \\
\underline{\mu}_{c''} : h_{c''}=0 & c' \succ c'' \\
\underline{\mu}_{c} : h_{c}=0
\end{cases}$$
Once the previous steps are done, for every element in $R_{c'}(\boldsymbol\mu_{-c'})$, we assign it to $h_{c'}$ and go to the next class according to the lexicographic order. Finally, once the policies of all the classes $c'\in C_{-c}$ have been assigned, the policy of $c$ is easily given by $h_c=r_c(\boldsymbol \mu_{-c})$.
The above method enumerates all and only the feasible policies consuming exactly the budget in which at most one $h_c$ is fractional.
We need to prove that:
- all the policies except $\mu_c$ are integer,
- the budget is exactly consumed, and
- all and only the feasible policies are enumerated
The first two points are trivial by construction (given that the policy of $c$ is the only potentially non–integer and is computed as the policy that consumes the budget given the policies of all the other classes). To prove the third point, we observe that $I$ is always a well–defined range. Indeed, $\lfloor r_{c'}(\underline{\boldsymbol \mu}_{-c'})\rfloor$ returns the largest $h_{c'}$ that consumes exactly the remaining budget given the budget consumed by all the classes preceding $c'$ in the lexicographic order. Assigning a policy larger than $\min\left\{\frac{K-1}{\epsilon}, \lfloor r_{c'}(\underline{\boldsymbol \mu}_{-c'})\rfloor\right\} $ does not allow one to consume entirely the budget. If the policies assigned to the previous classes are feasible, then $\lfloor r_{c'}(\underline{\boldsymbol \mu}_{-c'})\rfloor$ is always non–negative. As well, $\lfloor r_{c'}(\underline{\boldsymbol \mu}_{-c'})\rfloor$ returns the smallest $h_{c'}$ that consumes exactly the remaining budget given the budget consumed by all the classes preceding $c'$ in the lexicographic order and assuming that the classes that succeed transmit all the slots. Assigning a policy smaller than $\max\left\{0, \lceil r_{c'}(\overline{\boldsymbol \mu}_{-c'})\rceil\right\}$ does no allow one to consume entirely the budget. Even in this case, if the policies assigned to the classes that precede $c$ is feasible, then $\lfloor r_{c'}(\underline{\boldsymbol \mu}_{-c'})\rfloor$ is always smaller than $\frac{K-1}{\epsilon}$. Thus, by construction, for each policy assigned to class $c'$ belonging to $I$, it is always possible to find a feasible policy for the succeeding classes.
The number of policies enumerated by Algorithm \[alg:gridepsilon\] is exponential in the number of classes, being $O((\frac{K-1}{\epsilon})^{|C|-1})$.
We can derive a theoretical lower bound over the quality of the solution found by Algorithm \[alg:gridepsilon\] w.r.t. the optimal solution of the optimization problem.
\[thm:lower\_bound\] Called $\tilde{F}_D$ the value of the solution returned by Alorithm \[alg:gridepsilon\] and $F^*_D$ the value of the optimal solution, we have $\dfrac{\tilde{F}_D}{F^*_D}\geq \dfrac{1-(\frac{1}{2})^{\frac{K-1}{\epsilon}}}{1-(\frac{1}{2})^{|C|\frac{K-1}{\epsilon}}}$.
Call $\boldsymbol \mu^*$ the optimal policy profile and call $\tilde{\boldsymbol \mu}_c$ the policy profile in which $\tilde{h}_{c'}=\lfloor h_{c'}^*\rfloor$ for all $c'\neq c$ and $\tilde{h}_{c}=h_{c}^*$. Obviously, $F^*_D\geq F_D(K,\tilde{\boldsymbol \mu}_c)$. In addition, it is obvious that $F^*_D\geq \tilde{F}_D \geq \max_c\{F_D(K,\tilde{\boldsymbol \mu}_c)\}$. This is because $\tilde{\boldsymbol \mu}_c$ is a feasible policy profile in which at most one policy is fractional that is not assured to consume exactly the budget. We can write a lower bound to $F_D\left(\frac{K}{\epsilon},\tilde{\boldsymbol \mu}_c\right)$ as: $$\begin{gathered}
F_D\left(\frac{K}{\epsilon},\tilde{\boldsymbol \mu}_c\right) = 1 - \prod_{c'\in C}\prod_{k=0}^{\frac{K-1}{\epsilon}} X_{c',k}^*(\lambda_{c'}\epsilon\Delta,\tilde{\boldsymbol \mu}_c) \geq \\1 - \prod_{k=0}^{\frac{K-1}{\epsilon}} X_{c,k}^*(\lambda_{c}\epsilon\Delta,\tilde{\boldsymbol \mu}_c)\end{gathered}$$ By using such lower bound over $F_D\left(\frac{K}{\epsilon},\tilde{\boldsymbol \mu}_c\right)$, we can write: $$\dfrac{\tilde{F}_D}{F^*_D}\geq \max_c\left\{\dfrac{1 - \prod_{k=0}^{\frac{K-1}{\epsilon}} X_{c,k}^*(\lambda_{c}\epsilon\Delta,\boldsymbol \mu^*)}{1 - \prod_{c'\in C}\prod_{k=0}^{\frac{K-1}{\epsilon}} X_{c',k}^*(\lambda_{c'}\epsilon\Delta,\boldsymbol \mu^*)}\right\}$$ since, given $\tilde{\boldsymbol \mu}_c$ and $\boldsymbol \mu^*$, we have $\tilde{h}_c=h^*_c$. Thus, we are interested in: $$\min\max_c\left\{\dfrac{1 - \prod_{k=0}^{\frac{K-1}{\epsilon}} X_{c,k}^*(\lambda_{c}\epsilon\Delta,\boldsymbol \mu^*)}{1 - \prod_{c'\in C}\prod_{k=0}^{\frac{K-1}{\epsilon}} X_{c',k}^*(\lambda_{c'}\epsilon\Delta,\boldsymbol \mu^*)}\right\}$$ where the minimization is over all the parameters. Although the definition of $X^*$ is intricate, a bound can be derived disregarding the exponential nature of all the $X^*$ and considering them as arbitrary values in $[0,1]$. In this case, for reasons of symmetry, the values that minimize the maximum ratio prescribe $X_{c,k}^*=\frac{1}{2}$ for all $c$. This leads to the bound stated in the theorem.
Notice that the theoretical lower bound does not depend on whether the beaconing costs are present. The worst case is when $K=1$ and $|C|\rightarrow \infty$, obtaining a ratio of $1-\frac{1}{2}^\frac{1}{\epsilon}$. However, it can be observed that the worst case ratio goes to one exponentially in $\frac{1}{\epsilon}$. Thus we can obtain a good approximation ratio with a small value of $\frac{1}{\epsilon}$, e.g, the theoretical lower bound over the approximation ratio is about $1-10^{-4}$ when $\frac{1}{\epsilon} =10$. Algorithm \[alg:gridepsilon\] is an approximation scheme (AS), given that the ratio goes to one as $\epsilon$ goes to zero. However, it is not a fully polynomial time AS (FPTAS), its complexity not being polynomial in all the parameters.
Polynomial–time approximation algorithm
---------------------------------------
In this section, we discuss a heuristic approach to approximate the optimal policy in polynomial time. We start by providing Algorithm \[alg:greedy\], a method that greedily maximizes an objective function $G_i$. We shall consider two versions of this function, denoted with $G_1$ and $G_2$, and we will discuss approximation bounds guaranteed by their employment.
$\mu_1(k), \ldots, \mu_{C}(k) \gets 0 \; \forall k$ $k_1, \ldots, k_{C} \gets 0$ $F^* \gets 0$ $\mu_c(h_c + 1) \gets \min\{1,r_c(\boldsymbol\mu)\}$ $\delta_c \gets G_i(\boldsymbol\mu) $ $\mu_c(k_c + 1) \gets 0$ $c^* \gets \displaystyle\arg\max_{c \in C}\{\delta_c\}$ $\mu_{c^*}(k_{c^*} + 1) \gets \min\{1,\hat{\mu}_{c^*}\}$ $k_{c^*} \gets k_{c^*} + 1$ $F^* \gets F^* + \delta_{c^*}$
Algorithm \[alg:greedy\] works on the same discrete-time representation we introduced above, where each time slot has a temporal length of $\epsilon\Delta$. It starts from an initial empty policy and it proceeds considering only integer threshold policies. At each iteration, it appends a locally optimal time slot for a class $c$ meaning that such class will transmit with probability of $1$ for an additional subsequent time slot. Class $c$ is selected as the one that would introduce the largest gain in $G_i$ if a slot is assigned to it. We denote with $k_c$ the integer index for class $c$, referring to the last allocated time slot. Similarly, $\delta_c$ denotes the discrete marginal gain of $G_i$ obtained by allocating a slot to class $c$ in the current policy.
### First version, locally optimizing $F_D$ {#first-version-locally-optimizing-f_d .unnumbered}
In the first version of Algorithm \[alg:greedy\], we consider the maximization of the marginal gain of $F_D$, i.e., the delivery probability. Namely, at step (7) it holds that $G_1(\boldsymbol\mu) = F_D(K,\boldsymbol\mu) - F^*$. In this case, $\delta_c$ represents the benefit, in terms of delivery probability, that an additional time slot for class $c$ would introduce at the current iteration. By exploiting a result presented in [@krause12survey] we are able to provide a bound on the solution quality obtained with this version of the greedy algorithm. The result we shall use can be summarized as follows (see [@krause12survey] for details).
\[thm:submodular\_bound\] Given a ground set $\Omega$, a set function $F : 2^{\Omega} \rightarrow \mathbb{R}$, and a positive integer $W \in \mathbb{N^+}$, let us consider the problem finding $\hat{S}^* = \arg\max_{S \subseteq \Omega, |S| \le W}F(S)$. Then if $F$ is submodular, we have that for every integer $0 \le l \le W$ $$F(S_l) \ge (1 - e^{-l/W}) F(\hat{S}^*)$$ where $S_l \in \Omega$ is the set built after $l$ iterations of the following greedy element–selection rule $$S_i =
\begin{cases}
\emptyset & \text{if}\;$i=0$ \\
S_{i-1} \cup \arg\max_{s \in \Omega}F({S_{i-1} \cup \{s\})} & \text{else}
\end{cases}
\label{eq:greedy-selection-rule}$$
Theorem \[thm:submodular\_bound\] states that greedily maximizing a submodular set function introduces a bounded suboptimality. Eventually, the bound converges to $(1-\frac{1}{e})$ ($\approx 0.63$) when $l = W$, that is, when the maximum number of selections allowed by the cardinality constraint is made.
In order to apply this result to Algorithm \[alg:greedy\], we need to show that the problem of finding an optimal integer policy can be expressed as the maximization of a submodular set function subject to a cardinality constraint. This similarity can be shown by using the following simple formalism. Let us assume that each element in the ground set $e \in \Omega$ is a pair $(c,k)$ where $c \in C$ and $k \in \{1 \ldots \frac{K-1}{\epsilon}\}$. Then, every subset $S \subseteq \Omega$ can be uniquely associated to an integer policy that we denote as $\mu^S$. Intuitively, the correspondence between $S$ and $\mu^S$ is obtained by the following construction rule: $$\mu^S_c(k) =
\begin{cases}
1 & (c,k) \in S \\
0 & \text{else}
\end{cases}$$ Therefore, the objective function for a policy $\mu^S$ can be rewritten as a set function $F(K, S)$.
The second necessary step is to derive a cardinality constraint to define the problem’s feasibility region. In our problem, the feasibility of a policy is determined by the budget limit, namely by Constraint (\[equation:budget\]). For this reason, ideally one would like to find a $W$ such that $|S| > W$ if and only if $\mu^S$ violates Constraint (\[equation:budget\]). However, it can be easily shown that budget feasibility cannot be expressed with a cardinality constraint. The reason is straightforward. The budget of a policy does not solely depend on the number of transmitting slots, but also on how those slots are distributed among the different classes. Nevertheless, a necessary (not sufficient) cardinality upper bound can be determined via the following theorem.
\[thm:cardinality\_constraint\] Any feasible threshold integer policy cannot assign full probability of transmission to more than $W=\min \{ \max_c\{r_c(\boldsymbol\mu^{\emptyset})\}, \frac{K-1}{\epsilon}\}$, where $\boldsymbol\mu^{\emptyset}$ is the empy policy.
Let us assume that $\hat{c}=\arg\max_c\{r_c(\boldsymbol\mu^{\emptyset})\}$. Then, consider a threshold policy $\mu^S$ where $|S| > W$. If $\mu^S$ is budget–feasible then, by definition, the policy obtained in this way should be feasible too: for every $(c,k) \in S$ where $c \neq \hat{c}$ substitute $(c,k)$ with $(\hat{c},h_{\hat{c}} + 1)$. However, by definition of $W$ such a policy cannot be budget–feasible.
Under the above assumption, the optimal integer policy problem can be associated, up to a relaxation of the feasibility constraint, to the maximization of the set function $F_D(K, S)$, subject to $|S| \le W$. In the next step we show the submodularity of $F_D$.
\[prope:submodularity\] The set function $F$ is submodular with respect to $\Omega$.\[prope:submodularity\]
First, let us consider a setting with a single class. From Property \[prope:threshold\_optimal\], we can focus only on threshold policies and rewrite $F$ as a function of $h$, namely the threshold value (this value, in general, can be non–integer). Then it can be easily shown that $F(h)$ is a concave function since the Hessian matrix has strictly negative eigenvalues. Given a function $f: \mathbb{N} \rightarrow \mathbb{R}^+$, then $f(|S|)$ is submodular on the subsets $S$ of an arbitrary set $\Omega$ if and only if $f$ is concave. We can then conclude that $F$ is submodular in the case of a single class. Let us now show submodularity for the case with two classes. Let us denoted with $\Delta F (S|e)$ the marginal gain of $F$ obtained by adding the element $e$ to the set $S$, namely adding a transmitting slot to some class to the policy $\mu^S$. For submodularity to hold, we need to show that for every $S_a$, $S_b$, $e$ such that $S_a \subseteq S_b \subset \Omega$ and $e \in \Omega \setminus S_b$ we have that $\Delta F (S_a|e) \ge \Delta F (S_b|e)$. By definition $e$ adds a slot to a single class, let us assume without loss of generality that this class is $c_1$. Then we have: $$\begin{gathered}
\Delta F (S_a|e) = [1 - (1-(F_{c_1}(S) +\Delta F_{c_1} (S_a|e)))(1-F_{c_2}(S))] \\ - [1 - (1-F_{c_1}(S))(1-F_{c_2}(S))] \\ =(1-F_{c_2}(S_{a}))\Delta F_{c_1} (S_a|e)\end{gathered}$$ and, analogously, $$\Delta F (S_b|e) = (1-F_{c_2}(S_{b}))\Delta F_{c_1} (S_b|e)$$ Since $\Delta F_{c_1} (S_a|e) \ge \Delta F_{c_1} (S_b|e)$ by submodularity of $F_{c_1}$ and $F_{c_2}(S_{b}) \ge F_{c_2}(S_{a})$ by $F_{c_2}$ monotonicity, we have that $F$ is submodular. The same reasoning can be extended to an arbitrary number of classes.
Theorem \[thm:submodular\_bound\] can be applied by showing that Algorithm \[alg:greedy\] corresponds to the greedy element-selection rule reported in (\[eq:greedy-selection-rule\]). It is easy to see that rule (\[eq:greedy-selection-rule\]), when applied to the integer policy problem, proceeds by locally optimal appends in the same way that Algorithm \[alg:greedy\] does. Hence, we are now in the position of state the following theorem:
Let us denote with $S^*$ the policy returned by Algorithm \[alg:gridepsilon\] and with $S^1_l$ is the policy constructed by Algorithm \[alg:greedy\] (version 1) after $l$ iterations. We then have that $F_D(K,S^1_l) \ge (1 - e^{-l/W})F_D(K,S^*)$.
The inequality stated in the theorem follows immediately from the following two properties. First, by applying Theorem \[thm:submodular\_bound\] to Algorithm \[alg:greedy\] (version 1) we have that $F_D(K,S^1_l) \ge (1 - e^{-l/W})F_D(K,\hat{S}^*)$. Second, since $\hat{S}^*$ is the optimal solution of a relaxed version of the integer policy problem, it holds that $F_D(K,S^*) \le F_D(K,\hat{S}^*)$ .
The previous theorem, provides an *online* bound on the solution quality, being it dependent on the number of iterations the algorithm will succeed in performing without violating the budget constraint. An *offline* guarantee can be given by computing the minimum number of slot $s_c$ to be assigned to each class $c$. This number can be computed by setting $\mu_{c'}(i)=1 \; \forall i \; 0 \le i \le K, c' \neq c$ and computing the maximum number of time slots during which $c$ can transmit without saturating the budget.
For any solution $\hat{S}^1$ obtained with Algorithm \[alg:greedy\] (version 1) we have that $F_D(K,\hat{S}^1) \ge (1 - e^{-\sum\limits_{c \in C}{s_c}/W})F_D(K,S^*)$.
### Second version, normalizing $G_1$ with budget costs {#second-version-normalizing-g_1-with-budget-costs .unnumbered}
The second version of our algorithm is an improvement to the previous version that holds when no beaconing costs are considered. Here $G_2$ is obtained by normalizing $G_1$ with the budget cost that an additional time slot will introduce. In other words, $\delta_c$ will represent a ration between benefits and costs. Under the assumption that no beaconing costs are present and that we deal with threshold policies, each transmission has an independent cost and the budget spent by a policy $S$ is given by: $$\psi(S) = \sum\limits_{(c,k) \in S}N_c e^{-\lambda_c \Delta (k-1)}(1 - e^{-\lambda_c \Delta})$$ and, consequently, $$G_2(\boldsymbol\mu) =\frac{G_1(\boldsymbol\mu)}{\psi(\{(c,h_c + 1)\})}$$
If we modify rule (\[eq:greedy-selection-rule\]) by normalizing the objective function by the budget cost for each candidate element, we can again show the equivalence between the new rule and Algorithm \[alg:greedy\] (version 2). As a consequence, we can again resort to a result presented in [@krause12survey] and provide a quality bound on the solution obtained with the combination of the two versions of Algorithm \[alg:greedy\] when beaconing costs are not considered.
If no beaconing costs are present, then it holds that $$\max \{ F_D(K,\hat{S}^1), F_D(K,\hat{S}^2) \} \ge \frac{1}{2} (1 - \frac{1}{e}) \max\limits_{S \subseteq \Omega \\ \psi(S) \le \Psi} F_D(K, S)$$
The proof follows immediately by the consideration made above and a straightforward adaptation of results presented in [@krause12survey].
Experimental evaluation
=======================
In this section, we provide some experimental evaluations of the proposed algorithms. Results are obtained from MATLAB simulations and are aimed at showing the feasibility of our approach and evaluating its performance in terms of solution’s quality. We shall also discuss some qualitative issues observed in the obtained policies.
Experimental setting
--------------------
Each instance of our problem is described by different parameters. In our experiments, we generated istances by considering finite sets of values for each parameter, see Table \[tab:parameters\] for a complete summary. In particular, we devote our attention to three different mobility profiles and to three different transmission technologies for mobile nodes. Mobility profiles are characterized by increasing average speeds. The scenario that we imagine is populated by mobile devices carried by pedestrians, users on bycycles, and users on vehicles, respectively. The transmission technologies we consider provide increasing communication ranges: ZigBee, Bluetooth 4.0. and Wi-Fi Direct. We derive the corresponding values for $\rho$ and $\beta$ by considering the technical specifications of each technology and assuming an application scenario where a single packet has a size of $5$kB and a slot interval $\Delta=10s$. For simplicity, we assign the same number of users to each class.
temporal deadline for delivery ($\tau$)
-----------------------------------------
25, 50, 100, 250
: Parameters used for experiments[]{data-label="tab:parameters"}
radius of the environment ($L$)
---------------------------------
350, 500, 750,1000
: Parameters used for experiments[]{data-label="tab:parameters"}
number of nodes $N_c$
-----------------------
9, 15, 20
: Parameters used for experiments[]{data-label="tab:parameters"}
mobility profiles ($v_c$)
---------------------------
pedestrians (1.5$m/s$)
bicycles (6$m/s$)
vehicles (9$m/s$)
: Parameters used for experiments[]{data-label="tab:parameters"}
transmission technologies
---------------------------
ZigBee ($R=15m$)
Bluetooth 4.0 ($R=50m$)
Wi-Fi Direct ($R=100m$)
: Parameters used for experiments[]{data-label="tab:parameters"}
In the experimental results proposed here, we consider up to 3 classes and a temporal discretization varying according to $\epsilon \in \{1, 1/3, 1/5\}$. The reason behind this choice can be intuitively described by the two graphs of Figure \[fig:bounds\], where we depict the theoretical lower bound from Theorem \[thm:lower\_bound\] with respect to different resolutions and numbers of classes. As it can be seen, a maximum resolution of $\epsilon = 1/5$ represents a reasonable choice to guarantee about $95\%$ of the optimal solution quality without the burden of a prohibitive number of time slots. On the other side, by adopting a maximum number of 3 classes we obtain a case which is fairly close to the worst case (derived for an infinite number of classes) and that is computable by means of our grid algorithm (as discussed in the following—we recall that our grid search requires compute time that is exponential in the number of classes). Finally, we remark that we chose a small number of nodes for simplicity in our experiments, but that the compute time of all our algorithms is linear in the number of nodes.
![Theoretical lower bound over solution’s quality (Theorem \[thm:lower\_bound\]).](img/theoretical_bound_cropped.pdf)
\[fig:bounds\]
Benchmarks
----------
We compare the performance of our algorithm w.r.t. the performance of two heuristic easily–computable algorithms and of an upper bound over the value of the optimal policy.
### Greedy on arrival rate
This algorithm works as follows: it sorts the classes in descending order of $\lambda_c$, then it allocates all the possible budget to the classes from the first one in the order to the last one. For instance, given three classes with $\lambda_1=0.3,\lambda_2=0.2,\lambda_3=0.1$, the algorithm assigns all the possible budget to class $1$ and, if there is a remaining budget, then all the remaining budget is assigned to class $2$ and so on. The rationale is that we expect that the larger the arrival rate the larger the delivery probability. The complexity of this algorithm is obviously easy given that the policy can be found by solving at most $|C|$ equations.
### Class–independent policies
This algorithm searches for the optimal solution of an overconstrained problem in which: the policies related to all the classes are the same, formally $\mu(k)=\mu_c(k)$ for all $c$, and, when the policy is probabilistic, then either the source transmits to all the classes or it does not transmit at all. This last assumption leads to a new formulation of the budget constraint: $$\sum_{c\in C} \rho_c N_c \cdot (1- Q_{c,0,K}(\mu)) + \sum_{\omega\in \Omega}\sum_{k=0}^{K-1}\beta_\omega\cdot \mu(k) \leq \Psi $$ By Property \[property:budget\], the optimal policy is such that the budget $\Psi$ is completely consumed and therefore the above inequality holds with equality. Therefore, the optimization problem reduces to the problem of finding the policy that completely consumes the budget. Formally, interpreting the (class–independent) threshold $h$ as a continuous variable, we can write: $$g(h)= \sum_{c\in C} N_c \cdot e^{-\lambda_c \Delta h} - \sum_{\omega\in \Omega}\left(\sum_{c\in C_\omega} N_c -\frac{\beta_\omega}{\rho_c} \cdot h+ \frac{\Psi}{\rho_c} \right)= 0$$ Function $g$ is a single–variable function strictly monotonically decreasing in $h$ and infinitely differentiable. Such a function admits only one zero, and therefore the above equation admits only one solution. Such a solution can be found (approximately) by using the Newton method, that in this case, due to the property of the function, has a quadratic convergence speed (the number of correct digits roughly at least doubles in every step). Thus, we obtain an approximate solution of high quality within very short time.
### Upper bound over the optimal value
An over bound over the value of the optimal solution can be found by using a variation of the algorithm described in Section \[subsection:gridalgorithm\]. More precisely, we use Algorithm \[alg:gridepsilon\] to enumerate all the policies consuming entirely the budget and we change each policy rounding each $h_c$ to the smallest integer and then adding $1$ for every $c$. Notice that these new policies violate the budget constraint. Among all these policies we find the one maximizing the delivery probability. Its value is an upper bound over the value of the optimal policy. In the graphs we denote this value as $UB$. The proof follows. Call $\boldsymbol \mu^*$ the optimal policy profile with (potentially fractional) thresholds $h_c^*$. Call $\hat{\boldsymbol \mu}$ a generic policy profile obtained as described above. It can be easily observed (it follows from the fact that, fixed the policies of all the classes but one, the policy of the remaining class that consumes entirely the budget is always one) that there alway exists a policy profile $\hat{\boldsymbol \mu}$ such that $\hat{h}_c\geq h^*_c$ for all $c$. Therefore, given that the objective function is strictly monotone in $h_c$, the objective value of $\hat{\boldsymbol \mu}$ is strictly better than the value $\boldsymbol \mu^*$.
Experimental results
--------------------
Figure \[fig:mean\_fub\_tau\] reports how $F_D / UB$ varies as the values of the parameters $\tau,L,N_c$ vary as summarized in Table \[tab:parameters\], $|C|\in\{1,2,3\}$, and $\frac{1}{\epsilon}=5$. For each parameter, we average $F_D / UB$ over the other instances sharing the same value for that parameter. It can be observed that grid search and greedy constructions obtain a remarkable better performance in each case when compared with the benchmarking greedy algorithms based on the arrival rate and the class-independent one. Not exploiting the knowledge about the different classes and solely considering the arrival rate turned out to achieve very similar performances. By increasing the value of $\tau$, it can be seen how this gap with the benchmarks shrinks, suggesting the intuition that when the deadline for packet delivery is large even simplistic policies are able to obtain good delivery probabilities. Another aspect that can be observed is that greedy constructions revealed to be quite effective for the tested cases, since they were able to obtain high performances comparable to the grid search. By increasing the value of $L$, it can be seen how this gap with the benchmarks increases, instead the gap keeps to be approximately constant as $N_c$ and $C$ vary. Interestingly, the approximation ratio of our algorithms is constant (i.e., $>99\%$) w.r.t. all the parameters values.
\[fig:mean\_fub\_tau\]
A more detailed overview on how the performance varies with respect to $\tau$ is shown by the boxplots of Figure \[fig:fub\_tau\_boxplot\]. These graphs show the similarity in performance between the grid search and the greedy constructions algorithms. These last ones obtained worse performances for a limited number of outlier instances. Also it is evident how having finer resolutions remarkably improves the solution’s quality.
\[fig:fub\_tau\_boxplot\]
The above results suggest that greedy constructions seem to be quite effective approaches to approximate the optimal policy requiring, at the same time, much less computational effort than the grid search. In Figure \[fig:time\_greedy\], we show a comparison between computational times obtained with the grid search and the greedy construction algorithms respectively. In particular, we evaluated the algorithms’ scalability when the number of classes grows. To obtain these results we fixed the values of some parameters ($\epsilon = 1/3$, $\tau = 100$, $N_c=10$, $L=500$) and we generated random mobility profiles and transmission technologies by uniformly sampling from the following intervals: $R_c \in [15,50]$, , $v_c \in [1, 15]$ $\rho_i \in [0.05, 0.25]$ $\beta_c \in [3 \times 10^{-7}, 8 \times 10^{-7}]$. It is easy to see how grid search shows an exponential growth in time, while greedy construction proved to be much more efficient even for larger number of classes. Considering a deadline of 1 hour, grid search was not able to compute a solution for more than 4 classes, while greedy construction managed to compute solution up to $800$ classes.
![Time (in seconds) scalability with the number of classes.[]{data-label="fig:time_greedy"}](img/time_greedy_grid.pdf)
Finally, Figure \[fig:policies\] depicts a qualitative evaluation of the policies returned by our algorithms. We consider a reference value for the budget upper bound $\Psi$ and we show how the thresholds of the optimal policy (obtained with grid search) are distributed over the three different technologies. It can be observed how, by increasing the budget, the optimal policy tends to schedule transmissions with all the three technologies. When the budget gets smaller and smaller, then the policy tries to rely more on those technology that have a longer communication range.
![Policy thresholds with different upper bounds on budget.[]{data-label="fig:policies"}](img/policy_psi.pdf)
Conclusions
===========
In this paper we studied two-hop routing for Delay Tolerant Networks when heterogeneous technologies are present taking into account beaconing signal and deadlines after which nodes discard packets. Differently from the literature, whose adopts fluid approximation to find optimal policies—providing exact solution in the limit when the number of nodes is infinite, but a coarse approximation otherwise—, we adopt an operations research approach, formulating the problem as an optimization problem and designing approximation schemes with theoretical bounds. We thoroughly evaluated our algorithms with realistic settings in terms of approximation ratio and compute time as the parameters change. We experimentally showed that for all the generated instances our algorithms have an approximation ratio larger than 99% and that they scale linearly as the values of the parameters increase and therefore they can be applied with extremely large instances.
[Michael Shell]{} Biography text here.
[John Smith]{} Biography text here.
[Jane Smith]{} Biography text here.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Interdependent networks are characterized by two kinds of interactions: The usual connectivity links within each network and the dependency links coupling nodes of different networks. Due to the latter links such networks are known to suffer from cascading failures and catastrophic breakdowns. When modeling these phenomena, usually one assumes that a fraction of nodes gets damaged in one of the networks, which is followed possibly by a cascade of failures. In real life the initiating failures do not occur at once and effort is made replace the ties eliminated due to the failing nodes. Here we study a dynamic extension of the model of interdependent networks and introduce the possibility of link formation with a probability $w$, called healing, to bridge non-functioning nodes and enhance network resilience. A single random node is removed, which may initiate an avalanche. After each removal step healing sets in resulting in a new topology. Then a new node fails and the process continues until the giant component disappears either in a catastrophic breakdown or in a smooth transition. Simulation results are presented for square lattices as starting networks under random attacks of constant intensity. We find that the shift in the position of the breakdown has a power-law scaling as a function of the healing probability with an exponent close to $1$. Below a critical healing probability, catastrophic cascades form and the average degree of surviving nodes decreases monotonically, while above this value there are no macroscopic cascades and the average degree has first an increasing character and decreases only at the very late stage of the process. These findings facilitate to plan intervention in case of crisis situation by describing the efficiency of healing efforts needed to suppress cascading failures.'
author:
- Marcell Stippinger
- János Kertész
title: Enhancing resilience of interdependent networks by healing
---
Introduction {#sec:introduction}
============
Robustness is one of the key issues for network maintenance and design [@Cohen2000; @Callaway2000; @Barabasi2000]. The representation of complex systems has been limited to single networks for a long time [@Newman2010]. In many cases, however, coupling between several networks takes place [@Kivela2013; @non2014]. An important case is that of interdependency [@Buldyrev2010; @Li2012] where there are two kinds of links: connectivity and dependency links. An example of interdependent networks is the ensemble of the Internet and the power supply grid where telecommunication is used to control power plants and electric power is needed to supply communication devices [@Buldyrev2010]. Connectivity links model the relation of the entities within the same sector, spanning in the above example a power supply network and a telecommunication network. Dependency links depict the basic supplies an entity depends on which are supplied by entities in the other network. If a supplier fails its dependent nodes fail as well. The system is viable if a giant component of interconnected units exists in both networks. In the 28 September 2003 blackout in Italy it came to evidence that the interdependency of the two networks makes them more vulnerable than ever thought before [@Buldyrev2010]. Similar relations occur in the economics between banks and firms or funds. Banks are related through interbank loans, firms through supply chains and the interdependence comes from loans and securities. Inappropriate asset proportions can also lead to global avalanches as seen in the subprime mortgage crisis [@May2011].
Interconnecting similar subsystems used to increase capacity was shown beneficial as long as it does not open pathways to cascades [@Brummitt2012]. However, in interdependent networks, the aspect of robustness was considered with the conclusion that broadening the degree distribution of the initial networks enhances vulnerability [@Buldyrev2012]. A cost-intensive intervention to strengthen robustness is to upgrade nodes to be autonomous on some resources [@Schneider2013].
Because failures propagate rapidly in infrastructure networks, they cannot be stopped by installing backup devices during the spreading of the damage. but rather they require already existing systems. After the cascade of failures, damaged devices or elements can be replaced by new, functioning ones *identical* to the originals [@Havlin2012]. In contrast to engineered systems, social or economic networks are highly responsive and may react quickly [@Onnela2010; @Schweitzer2009]. When a failure occurs considerable effort is made to reorganize the network and rearrange the load of failing elements among functioning ones. The role of the failing entities is taken over by *similar* participants. Such processes can be modeled by healing, i.e., substituting some of the failed elements by new ones. The timescale of an economic crisis is wide enough for the network to completely restructure itself [@Schweitzer2009]. So far such mechanisms have only been studied for simple networks [@Wang5; @Kenett2013; @Havlin2012challenges]. Here we extend the original model [@Buldyrev2010] of cascading failures of interdependent networks. After each removal, the healing process attempts to bypass the removed node with a new connectivity link (see Fig. \[fig:cascade\_schema\]). In this paper, we demonstrate how healing acts on interdependent networks.
![\[fig:cascade\_schema\] *a)* Failures, represented by red dots, affect the nodes one by one in a random order. Whenever a node fails, its counterpart, that is, the node in the other network which depends on it, fails as well. In both networks, only the largest connected component (LCC) survives. This constraint can cause further nodes to fail in both networks, which trigger further shrinking of the LCC, and so on, illustrated by the shaded areas. *b)* The neighbors of a failing node try to heal the network, such that two functioning neighbors of a removed node establish a connectivity link with probability $w$.](fig1)
The outline of the paper is as follows. In Sec. \[sec:model\] we define the node failure process in a dynamic way. We introduce initial failures one by one to be able to apply healing at every failure event. Then we relate the original version of cascading failures to our model as a special case and give formulas for comparing the order parameter of the two models. The scaling properties of the healing are explained along with the numeric results in Sec. \[sec:scaling\]. In Sec. \[sec:cascades\] we discuss the properties of the cascades with microscopic insight to the model. Finally we conclude our findings in Sec. \[sec:conclusion\].
The model {#sec:model}
=========
In the standard model of interdependent networks [@Li2012] the computer-generated model-system is built up of two topologically identical networks $A$ and $B$, e.g., square lattices of size $N=L\times L$, where each node has *connectivity* links within the same network. In addition, *dependency* links couple between the networks, which are bidirectional one-to-one relationships connecting randomly selected pairs of nodes from the two networks. If any of the nodes fails its dependent pair fails too. A node in any network can function only if it is connected to the largest connected component of that network the node which it depends on is also functional, otherwise it fails, i.e., it is removed from the network.
The existence of a macroscopic connected component in a single network is treated by percolation theory. In the usual case, for a lattice it describes a second-order phase transition between the phases with and without the existence of a giant component [@Stauffer1994]. Adding interdependency allows cascades of failures to propagate between the two networks. The threshold the network can survive without collapse decreases considerably in this setting [@Li2012].
The collapse due to cascades was shown to be a first order transition if the dependency links have unlimited range while the transition is of second order if the range is less than a critical length $r_c$ [@Li2012; @Danziger2013]. Moreover, the first order transition has a hybrid character with scaling on one of its sides [@Havlin2011; @Baxter2012].
![\[fig:insight\] Part of Network $A$ of a simulated system at $p=0.7$ at *a)* no healing ($w=0.0$) *b)* below the critical healing ($w=0.2$, the average degree stays below $4$) and *c)* slightly above the critical healing ($w=0.4$). *d)* This latter $w=0.4$ system is also represented at $p=0.2$ where one can observe that the nodes get more and more connected and the healing process establishes links between distant nodes.](fig2)
As mentioned in the Introduction we first introduce a dynamic process on the interdependent network model. In the setting of two interdependent networks of general topology this dynamic process consists of the repetition of attacks and relaxations to a rest via cascades. (See Fig. \[fig:cascade\_schema\].) Let us suppose that failures affect the nodes one by one in a random order which defines a timeline. One time step is identified with the external attack of one node. Time is measured by the number of time steps normalized by $N$ for systems of different sizes to be comparable: $$\mathrm{elapsed~time} = 1 - p = \frac{\mathrm{number~of~time~steps}}{N}$$ The externally introduced failure in network $A$ may separate the largest connected component (LCC) into two or more parts where only the largest one survives. All the failed nodes have dependency connections to nodes of the network $B$ causing their failure. Again, the LCC of $B$ may get fragmented and only the largest part survives. This cascading procedure is repeated until no more failures happen. Of course, our model can easily be generalized to any number of interdependent networks and any density of dependency links.
Our aim is to introduce healing into this dynamic model. The procedure is as follows: After an externally introduced failure (which may cut off a part of the LCC) the healing step follows. Two remaining, functioning neighbors of a removed node establish a connectivity link with an independent probability $w$. (See part *b)* in Fig. \[fig:cascade\_schema\].) Then the dependent nodes of the removed nodes are removed from the other network. After the propagation of the failure there, again, two functioning neighbors of a removed node establish a connectivity link probability $w$. Due to the separation of small components, further damages might propagate back and forth within the network, always followed by a healing step. Here, the healing step means that all pairs of neighbors of each failed node is considered as a candidate for a new connectivity link with an independent probability $w$, then, after having selected the candidates, the connectivity links are established simultaneously. The process goes on until no more separation of components occurs. The healing links may change the topology considerably, bridging larger and larger distances as the time goes on (Fig. \[fig:insight\]). Once a critical fraction $(1-p_c)$ of nodes are removed, a catastrophic cascade destroys the remaining system.
The $w=0$ case is simply the dynamic version of the well studied model of Li *et al.*. In [@Li2012] a fraction $(1-q)$ of the original network is destroyed in the first step then the size of the giant component after the relaxation of cascades is traced as a function of $q$. The important difference between this procedure and ours is that in the version of Li *et al.* nodes may be accidentally attacked, which already fail in our step-by-step (dynamic) model. Let $P_\infty$ denote the fraction of remaining nodes as a function of the fraction of attacked nodes $(1-p)$ in the step-by-step model. The number of unattacked but disconnected nodes is $[p-P_\infty(p)]N$. The probability of randomly destroying an already disconnected (but not attacked) node is $(p-P_\infty)/p$, so the implicit relation between the two attacking methods is [^1] $$\label{corresponding_p} 1-p(q) = \int_q^1 1 -
\frac{\widetilde{p}-P_\infty(\widetilde{p})}{\widetilde{p}}
\,\mathrm{d}\widetilde{p}.$$ Due to the small false target ratio in the random attack, the threshold values of the two models are close. The extrapolated threshold value for the infinite system size in case $w=0$ is $p_c = 0.690\pm0.001$, in good agreement with the result of Li *et al.*.
Scaling with the healing probability {#sec:scaling}
====================================
The order parameter $P_\infty$ of our model depends not only on the fraction of attacked nodes but also on the healing parameter. According to one’s intuition, the data show that the critical attack $(1-p_c)$ increases monotonically with $w$.
We executed Monte Carlo simulations of our model with both periodic and open boundary conditions on square lattices starting networks of linear size $L=20$, $40$, $80$, $160$ and $320$ with $960$, $480$, $240$, $120$ and $60$ runs respectively, and we measured that the execution time in our implementation scaled approximately as $N^{2.3}=L^{4.6}$. In the square lattices connectivity links join nodes to their nearest neighbors within the same network. Dependency links were established by first creating the trivial mapping between the topologically identical lattices, then randomly shuffling the end of the links. The $p_c$-s are then obtained averaging over the *vertical* axis: for a given number of surviving nodes $P_\infty(p,w)$, we averaged the proportion of nodes $1-p$ attacked one-by-one. Fig. \[fig:orderparam\] shows the averaged curves for different values of $w$. The shape of the $P_\infty(p,w)$ curves suggests the scaling in the form of anisotropic resizing from the $S(p=1,P_\infty=1)$ point: $$\label{scalingform}
1-P_\infty(1-p,w)=1-a(w)\,P_\infty\left(\frac{1-p}{c(w)}, 0\right)$$ which is asymptotically satisfied in the $w\to 0$ limit.
In the infinite lattice limit, the initial few attacks almost surely occur in different parts of the lattice and do not raise cascades, only the attacked points fail, $P_\infty(p)=p$ if $p \sim 1$. The unit slope at $S$ with respect to $p$ can be expressed by differentiation and yields $a(w)\equiv c(w)$. Let us express the fraction of unattacked nodes relative to the threshold without healing: $\Delta p=p-p_{c0}\leq0$. The change in the threshold value $\Delta p_c(w)=p_c(w)-p_{c0}$ can be identified by the largest $\Delta p$ where $P_\infty$ has an infinite slope (see Fig. \[fig:criticalw\]): $\lim_{\Delta p\to \Delta p_c(w)+0} \frac{\partial}{\partial \Delta p}
P_\infty(1-p_{c0}-\Delta p,w) = \infty$. Substituting it into yields $a(w)=(1-p_{c0}-\Delta p_c(w))/(1-p_{c0})$. The increase in lifetime, $-\Delta p_c(w)$, has a general scaling behavior expressed in $$\label{powerlaw}
-\Delta p(w)= h\,w^\gamma$$ for small $w$-s, in the range $[0.000,0.050]$. For the purpose of precise measurement we created simulation data for all system sizes with step size $0.001$ for $w\in[0.000,0.010]$ additional to that shown in Fig. \[fig:orderparam\]. The measurement is hampered by large fluctuations of the small systems, therefore we extrapolated to infinite system size using standard finite size scaling [@Privman1990]. We used both systems with periodic and open boundary conditions and measured the finite size fluctuations in $p_c$ (approaching from the $p>p_c$ domain in accord with the hybrid character of the transition) which yields slightly different scaling exponents within the error tolerance for the two systems (respectively $\nu_p=1.10\pm0.06$ and $\nu_o=1.20\pm0.06$) from which we deduce $\nu\approx1.15$. The finite size scaling measurements, yielding $p_c = 0.690\pm0.001$, are represented in Fig. \[fig:fss\].
The parameters of Eq. are first fitted for each system size $N=L\times L=20^2, 40^2, 80^2, 160^2 \text{ and } 320^2$, then the infinite size limit is obtained using $1/L$ extrapolation. The systems with periodic and open boundary conditions simulated at different system sizes collapse well yielding $h=0.703\pm0.005$ and $\gamma=1.034\pm0.009$ for the infinite size network.
![\[fig:criticalw\] $p_c$ as a function of $w$ depicts the fraction of unattacked nodes at the transition for the starting $N=320\times320$ system size. $P_\infty(p_c(w)+,w)$ is the giant component size just before the transition as a function of $w$. Its non-zero value shows the jump in the first order transition and its zero value above $w_c = 0.351\pm0.002$ indicates a smooth transition.](fig4)
![\[fig:fss\] The standard deviation of the critical attack fraction $1-p_c$ was used to obtain the length scaling exponents $\nu_p=1.10\pm0.06$ and $\nu_o=1.20\pm0.06$ for periodic (filled) and open (void symbols) boundary conditions as described in Sec. \[sec:scaling\]. Then $p_c$ on the is plotted against $L^{-1/\nu}$ giving good collapse for the infinite system size.](fig5)
![\[fig:degreechange\] *(left and inset)* The average degree on the horizontal axis as a function of the fraction of dead nodes on the vertical axis for the starting $N=320\times320$ system size. The average degree remains constant for $w_c=0.348\pm0.003$. Plotted lines from the right to the left correspond to the range $w=0.340$ to $w=0.360$ respectively with a step size $0.001$, solid lines indicate steps of $0.005$. The shaded areas represent $0.33$ standard deviations in the left part. In the inset, shaded areas are only plotted for solid lines and represent $1.00$ standard deviation. *(right)* The fraction $P_\infty(p)$ of failing nodes as a function of the fraction $p$ of nodes not attacked externally using the same averaging as in Fig. \[fig:orderparam\]. Above $w_c=0.351\pm0.002$ there is no breakdown.](fig6)
Cascades change topology {#sec:cascades}
========================
We call cascades all events involving more nodes than the attacked one and its dependency counterpart. The size (number of nodes involved compared to the starting lattice size) of typical cascades is small up to the point of breakdown.
The healing dynamics changes the network topology and the average degree as well. Fig. \[fig:degreechange\] allows us to describe a transition: below a critical healing threshold $w_c$ we find a sharp breakdown in the number of surviving nodes. The critical healing is defined as the lowest $w$ for which the $P_\infty(p)$ function does not have an infinite slope. In our simulation we observe $w_c = 0.351\pm0.002$. For $w > w_c$ also there is no macroscopic cascade and $P_\infty(p)$ goes smoothly to zero in a second-order transition as $p$ decreases (see also Fig. \[fig:criticalw\]).
The healing performed by the $k$ neighbors introduces $w\binom{k}{2}$ new links on average. A rough mean-field estimate of $w_c$ is the healing probability, which conserves the average degree in the initial settings, leading to $2w_c\binom{k}{2}=k$ (each link joins $2$ nodes). As the square lattice has $k = 4$, the result is $w_c^{\text{mean-field}} = 1/3$. According to the left plot in Fig. \[fig:degreechange\] we find that the average degree $k = 4$ changes least through the simulation for $w_c = 0.348\pm0.004$, which agrees well with the critical healing determined from the $P_\infty$ curves [^2]. The change in the topology along with the trend of the average degree can be observed in Fig. \[fig:insight\]. Below the critical healing $w_c$ the average degree is monotonically decreasing function of $1-P_\infty$ and the connectivity links remain local, conserving the disordered lattice-like topology. Thorough inspection shows that all simulations end with a cascade wiping out all of the remaining network at $p_c(w)$. Above $w_c$ the healing promotes the formation of densely connected regions and connectivity links begin to join distant nodes. We remark that in the terminal stage the defined dynamics removes all nodes and links in both cases. In summary, the difference is that for $w<w_c$ the process terminates with a macroscopic cascade, while for $w>w_c$ there is no macroscopic cascade. In the latter case the average degree increases until it has to decrease due to the small number of remaining nodes.
Conclusions {#sec:conclusion}
===========
We examined the consequences of healing by edge formation in interdependent networks under random attack. We found that the increase in resilience of the network, measured in the number of survived attacks, has power-law scaling with the probability $w$ of healing. By establishing new random links in the neighborhood of the failed nodes, we delayed the collapse of the network through the hindering of cascades. We found that it is possible to completely suppress macroscopic cascading failures for healing probabilities higher than a critical value $w_c$; we demonstrated that this critical healing probability keeps the average degree of the nodes close to the initial value while the network topology changes. By analyzing healing efficiency, these findings can aid in the development of intervention strategies for crisis situations. The presented model contains a number of unrealistic features, like the starting lattice, the unbounded range and the high density of dependency links and the non-locality of the healing links. Further studies should clarify the role of these simplifications.
Acknowledgements {#sec:acknowledgements}
================
This work was partially supported by the European Union and the European Social Fund through project FuturICT.hu (Grant No.: TAMOP-4.2.2.C-11/1/KONV-2012-0013). JK thanks MULTIPLEX, Grant No. 317532. Thanks are due to Éva Rácz for her help at the early stage of this work and to Michael Danziger for a critical reading of the manuscript.
[25]{}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.85.4626) [****, ()](\doibase 10.1103/PhysRevLett.85.5468) [****, ()](\doibase 10.1038/35019019) [**](\doibase
10.1093/acprof:oso/9780199206650.001.0001) (, ) @noop [ ()]{}, in [**](\doibase
10.1007/978-3-319-03518-5_1), , (, ) pp. [****, ()](\doibase
10.1038/nature08932) [****, ()](\doibase
10.1103/PhysRevLett.108.228702) [****, ()](\doibase 10.1038/nature09659) [****, ()](\doibase 10.1073/pnas.1110586109) [****, ()](\doibase
10.1103/PhysRevE.85.066134) [****, ()](\doibase
10.1038/srep01969) [****, ()](\doibase 10.1038/nphys2180) [****, ()](\doibase
10.1126/science.1184819) [****, ()](\doibase
10.1126/science.1173644) [****, ()](\doibase 10.1016/j.physa.2013.01.013) [****, ()](\doibase 10.1038/nphys2819) [****, ()](\doibase 10.1140/epjst/e2012-01695-x) [**](http://www.crcpress.com/product/isbn/9780748402533), ed. (, ) in [**](\doibase 10.1109/SITIS.2013.101) (, ) pp. [****, ()](\doibase
10.1103/PhysRevE.84.066116) [****, ()](\doibase 10.1103/PhysRevLett.109.248701) , ed., [**](http://www.worldscientific.com/worldscibooks/10.1142/1011) (, )
[^1]: The integral can be numerically evaluated more accurately by where is the inverse function of .
[^2]: In triangle lattice we measure $w_c\approx 0.265\pm0.005$, and the heuristic argument gives $w_c=1/5$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce notions of *vector field* and its (discrete time) *flow* on a chain complex. The resulting dynamical systems theory provides a set of tools with a broad range of applicability that allow, among others, to replace in a canonical way a chain complex with a “smaller” one of the same homotopy type. As applications we construct in an explicit, canonical, and symmetry-preserving fashion a minimal free resolution for every toric ring and every monomial ideal. Our constructions work in all characteristics and over any base field. A key subtle new point is that in certain finitely many positive characteristics (which depend on the object that is being resolved) a transcendental extension of the base field is produced before a resolution is obtained, while in all other characteristics the base field is kept unchanged. In the monomial case we show that such a transcendental base field extension *cannot* in general be avoided, and we conjecture that the same holds in the toric case.'
address: |
Department of Mathematics\
University at Albany, SUNY\
Albany, NY 12222
author:
- Alexandre Tchernev
title: |
Dynamical systems on chain complexes\
and canonical minimal resolutions
---
Introduction {#introduction .unnumbered}
============
A standard exercise in differential geometry shows that a vector field on a smooth manifold induces naturally a chain homotopy $V$ on the de Rham complex of the manifold such that $V^2=0$. In his work on discretizing Morse functions and vector fields on manifolds [@Fo1; @Fo2], Forman arrives at the notions of *combinatorial vector field* and its (discrete time) *flow* on the cellular chain complex of a regular CW-complex, and this combinatorial vector field is again a chain homotopy $V$ such that $V^2=0$. Chain homotopies with this property appear prominently also in homological perturbation theory, see [@B; @G; @GL; @L] and references there, and implicitly in other related fields such as algebraic discrete Morse theory [@BW; @JW; @Sk; @Sk2].
In this paper we present (a condensed version of) a general theory of dynamical systems on chain complexes which in some sense mirrors the classical theory of dynamical systems on compact manifolds, and which complements and incorporates as special instances the previously discussed notions. Following Forman’s lead, we call a chain homotopy $V$ on a chain complex ${\mathfrak}F$ a *vector field* on ${\mathfrak}F$ whenever $V^2=0$. In analogy to the manifolds case, such a vector field induces a (discrete time) *flow* $\Phi_V$ on ${\mathfrak}F$, which is an endomorphism of ${\mathfrak}F$ and is chain homotopic to the identity. Iterating this flow produces a well-behaved dynamical system on ${\mathfrak}F$ that preserves the homotopy type. To analyze its asymptotic behaviour, we employ a structure we call a $P$-*grading* or $P$-*stratification* on ${\mathfrak}F$, the presence of which is analogous to having a handlebody decomposition on a manifold, and which is ubiquitous in applications. Such a stratification induces naturally a family of subquotients of ${\mathfrak}F$ that we call the *strata*. We use them to introduce the notion of *Lyapunov structure* for a vector field $V$ on ${\mathfrak}F$, which controls the asymptotic behaviour of the flow $\Phi_V$, and mirrors the way a Lyapunov function controls the asymptotic behaviour of a flow on a compact manifold, see e.g. [@C; @Fr]. A key feature, for which we have not yet discovered a good analogue in the classical dynamical systems on manifolds case, is that a $P$-stratification provides a set of tools for *canonical* creation of vector fields on ${\mathfrak}F$ with desired asymptotic flow behaviour.
While a more detailed analysis of the resulting dynamical systems theory is an ongoing research project, even the basic facts presented here already have a broad range of applications. Leaving uses in topology and in representation theory for later publications, we focus on two long-standing problems from algebraic geometry, commutative algebra and combinatorics — the construction in a canonical, explicit, and *intrinsic* (i.e. symmetry-preserving) manner of minimal free resolutions for toric rings and for monomial ideals.
The monomial ideals case has been a central open problem in commutative algebra since the thesis of Taylor [@T] in 1966, and remains a very active area of research, see e.g. [@MS; @P; @OW] and the references there. The ultimate goal is a construction that is canonical, explicit, preserves the symmetries of the ideal, and works for all ideals and in all characteristics. A solution that works for all ideals in characteristic zero goes back at least to the main result of Yuzvinsky [@Y Theorem 4.3], and is obtained by using in it the canonical splittings given via Moore-Penrose pseudo-inverses, see Remark \[R:pseudo-inverses\]. In the recent paper [@EMO], Eagon, Miller, and Ordog use their amazing combinatorial description of Moore-Penrose pseudo-inverses to give another solution in characteristic zero with a strong combinatorial component, and this solution works also for every ideal in all but finitely many positive characteristics.
To understand why a solution that works in all characteristics has remained elusive so far, we point out a common feature of every construction of a monomial resolution that we are aware of, whether minimal or not: a base field is chosen at the start and is then kept fixed throughout regardless of the properties of the ideal that is being resolved. Perhaps unexpectedly, it turns out that such an approach *cannot* succeed in positive characteristic if one aims for a minimal free resolution that respects the symmetries of the monomial ideal. A key main result of this paper is Theorem \[T:counterexample\] where we provide for each prime $p\ge 2$ a monomial ideal $I(p)$ that does *not* have a symmetry-preserving minimal free resolution for any base field algebraic over $\mathbb F_p$.
The solutions we present in both the monomial and the toric case employ a general strategy common to a number of constructions from representation theory, algebraic geometry, and commutative algebra, among them [@Y], [@EMO], and all constructions that use algebraic Morse theory. In our dynamical systems terminology it is described as follows:
- start from a canonical resolution ${\mathfrak}F$ with a suitable natural $P$-stratification;
- then construct a splitting of each stratum;
- this induces in a natural way a vector field on ${\mathfrak}F$ that has a flow stabilizing after finitely many iterations;
- the stable iterate is then a projection of ${\mathfrak}F$ onto a minimal resolution ${\mathfrak}M$.
Since each stratum is usually canonically obtained from an underlying finite chain complex of finite-dimensional vector spaces over the base field, with each vector space often given a canonical basis, this strategy reduces (completely formally, given our dynamical systems machinery) the minimal resolution problem to the problem of canonically constructing a splitting of such a chain complex of based vector spaces. We feature several ways of producing such a canonical splitting.
Our construction in the monomial ideals case, Theorem \[T:monomial-intrinsic\], obtains the explicit, canonical, and intrinsic minimal free resolution directly from the lcm-lattice [@GPW] of the monomial ideal. It works for all monomial ideals in all characteristics, and keeps the base field unchanged except for certain finitely many positive characteristics (which depend on the ideal), where it produces the desired result only after a transcendental extension of the base field. The starting canonical free resolution here is the *lcm-resolution*, a non-minimal resolution supported on the order complex of the lcm-lattice. On each stratum there is a natural finite family of what we call *matroidal* splittings, induced by the canonical bases of the vector spaces underlying the stratum. When the characteristic does not divide the number of matroidal splittings we obtain a canonical splitting of the stratum from a simple average of all matroidal splittings. When the characteristic divides the number of matroidal splittings, we obtain a canonical splitting of the stratum only after a transcendental extension of the base field; here we form a weighted average of all matroidal splittings with generic weights.
The problem for toric rings, while closely related to the problem for monomial ideals, has been open in all characteristics. Canonical, but non-minimal resolutions were constructed in [@BS] and [@TV1]. Our solution, Theorem \[T:toric-intrinsic\], works in all characteristics and starts with the resolution from [@TV1]. To obtain a canonical splitting of each stratum in characteristic zero we use the Moore-Penrose splitting, and in positive characteristic we choose the weighted average with generic weights of *all* splittings of the canonical finite chain complex that underlies the stratum. Because of this choice, in every positive characteristic we obtain a resolution only after a transcendental base field extension. Of course, see Remark \[R:monomial-remarks\](g), this argument can easily be changed in a manner analogous to the monomial ideals case so that the base field is kept unchanged except in finitely many positive characteristics.
The paper is organized as follows. Sections \[S:vector-fields\]-\[S:asymptotic-properties\] contain a quick introduction to the general theory, with just the bare essentials needed for our applications. In Section \[S:vector-fields\] we introduce the notions of vector field on a chain complex, and the flow of a vector field. We also introduce the notions of $P$-grading/stratification, stratum, Lyapunov structure, and chain-recurrent chain complex. A key point is Proposition \[P:stratification-induced\](c) which shows that, given a vector field on each stratum of a stratified complex ${\mathfrak}F$, one obtains in an explicit canonical way a vector field on ${\mathfrak}F$. In Section \[S:vector-fields-and-splittings\] we review from our new perspective a well-known and well-studied class of vector fields, the class of all splittings. As their flows are already stable, they serve as an important building block to create vector fields with well controlled asymptotic behaviour. We give examples, and elaborate on how to obtain a new splitting from a weighted average of a finite set of splittings. In Section \[S:asymptotic-properties\] we show, see Theorem \[T:asymptotic\], how a Lyapunov structure can be used to control the asymptotic behaviour of the flow of a vector field. The proof we give is an elaborate and explicit generalization of an argument of Yuzvinsky from [@Y Lemma 4.1], but parts of our theorem can be obtained also by using the proof of the basic perturbation lemma from homological perturbation theory [@B; @G].
In Section \[S:toric-rings\] we give our construction, Theorem \[T:toric-intrinsic\], of the explicit, canonical, and intrinsic minimal free resolution of a toric ring. We discuss some of its properties, and compute a simple example.
In Section \[S:monomial-resolutions\] we present, see Theorem \[T:monomial-intrinsic\], our construction of an explicit, canonical, and intrinsic minimal free resolution of a monomial ideal. We state in Problem \[P:mfr\] a much weaker than normally considered version of this construction problem, which reflects the necessity of passing to a base field extension before a solution can be obtained in some positive characteristics. Using our dynamical systems approach we give a solution that has good functorial properties and works in all characteristics. Section \[S:an-example\] is devoted to the computation of a useful example of this monomial construction.
Finally, in Section \[S:monomial-resolutions-in-positive-characteristic\] we show that, see Theorem \[T:counterexample\], in positive characteristic it is not, in general, possible to construct a minimal resolution of a monomial ideal by using only intrinsic properties of the ideal without first taking a transcendental extension of the base field.
The author would like to thank Lucho Avramov, Hank Kurland, and especially Marco Varisco for very useful conversations on various aspects of the theory presented here.
Vector fields {#S:vector-fields}
=============
Throughout this paper rings are associative with unit, modules are left and unitary, and unadorned tensor products are over $\mathbb Z$. While all definitions and results in the first three sections will be formulated in the setting of chain complexes of modules over some base ring, it should be clear to the reader that they are valid also for chain complexes over any co-complete abelian category.
Let ${\mathfrak}F=(F_n,\phi_n)$ be a chain complex of modules over a ring. We will write $B_n({\mathfrak}F)$, $Z_n({\mathfrak}F)$, and $\operatorname{H}_n({\mathfrak}F)$ for the $n$-boundaries, $n$-cycles, and $n$th homology, respectively, of ${\mathfrak}F$. The *homology chain complex of ${\mathfrak}F$* is the chain complex $\operatorname{H}({\mathfrak}F)$ with zero differential and component $\operatorname{H}_n({\mathfrak}F)$ in homological degree $n$.
Recall that if $D$ is a chain homotopy on ${\mathfrak}F$ then the morphism $$\Phi_D=\operatorname{id}_{{\mathfrak}F} - \phi D - D \phi$$ is called the *deformation of ${\mathfrak}F$ along $D$*. Clearly $\Phi_D$ is an endomorphism of ${\mathfrak}F$ and is chain homotopic to the identity.
A *pre-vector field* on ${\mathfrak}F$ is a chain homotopy $D$ on ${\mathfrak}F$ such that $$D\Phi_D=\Phi_D D.$$ A *vector field* on ${\mathfrak}F$ is a chain homotopy $V$ such that $V^2=0$. If this is the case, we call the deformation $\Phi_V$ the *flow* of the vector field $V$.
Recall that a *contraction* of ${\mathfrak}F$ is a chain homotopy $D$ such that $\Phi_D=0$. Thus every contraction is a pre-vector field on ${\mathfrak}F$. Furthermore, a routine computation shows that if $D$ is a contraction, then the homotopy $D\phi D$ is again a contraction and also a vector field on ${\mathfrak}F$.
The following proposition is a straighforward consequence of the definitions.
Let $D$ be a chain homotopy on ${\mathfrak}F$.
\(a) $D$ is a pre-vector field on ${\mathfrak}F$ if and only if $D^2\phi = \phi D^2$. In particular, a vector field on ${\mathfrak}F$ is also a pre-vector field.
\(b) If $D$ is a pre-vector field on ${\mathfrak}F$, then it is also a pre-vector field on both $\operatorname{Ker}\Phi_D$ and $\operatorname{Im}\Phi_D$, and a contraction on $\operatorname{Ker}\Phi_D$. In particular, $\operatorname{Im}\Phi_D$ is chain homotopy equivalent to ${\mathfrak}F$.
\(c) If $D$ is a pre-vector field on ${\mathfrak}F$, then the chain homotopy $V=D\phi D$ is a vector field on ${\mathfrak}F$. Furthermore, $V$ is a vector field on both $\operatorname{Ker}\Phi_D$ and $\operatorname{Im}\Phi_D$, and a contraction on $\operatorname{Ker}\Phi_D$.
Thus, iterating the flow of a vector field on a chain complex produces a well-behaved dynamical system that preserves homotopy type. The structure described below emerges as a key tool in controling the asymptotic behaviour of such a dynamical system. It can be thought of as an analogue in the setting of chain complexes to a handlebody decomposition of a compact smooth manifold.
Let $P$ be a poset. A *$P$-grading* or *$P$-stratification* on the complex ${\mathfrak}F$ is a decomposition $$F_n=\bigoplus_{a\in P} F_n^a$$ for every $n$ such that for each $a\in P$ the collection ${\mathfrak}F(a)=(F_n(a), \phi_n)$ is a subcomplex of ${\mathfrak}F$, where $$F_n(a)=\bigoplus_{x\le a}F_n^x.$$ The summand $F_n^a$ is called the *degree $a$ homogeneous component* of $F_n$. The chain complex $${\overline}{{\mathfrak}F(a)} = {\mathfrak}F(a)\Big/\sum_{b<a}{\mathfrak}F(b)$$ is called the *associated graded strand* at $a$ or the *stratum* at $a$. Observe that we have ${\overline}{{\mathfrak}F(a)}_n = F_n^a$.
$P$-gradings are ubiquitous in algebraic geometry, commutative algebra and algebraic topology. Below we provide several relevant examples. For more details, examples, and applications see e.g. [@CT Section 2] and the references there.
\(a) Let $P=pt$, and set $F_n^{pt}=F_n$. This produces the *trivial* grading/stratification on ${\mathfrak}F$.
\(b) Let $P=\mathbb Z$ and set $F_n^n = F_n$ and $F_n^a = 0$ if $a\ne n$. This produces the $\mathbb Z$-grading/stratification by *homological degree*.
Let $R$ be a $\mathbb Z$-graded ring, and let ${\mathfrak}F$ be a complex of $\mathbb Z$-graded free $R$-modules, with degree-preserving differentials. For each $n$ let $B_n$ be a homogeneous basis of $F_n$, and for each $a$ let $B_n^a$ be the subset of all elements $b\in B_n$ of degree $|b|=a$. Writing $F_n^a$ for the free submodule of $F_n$ with basis the elements from $B_n^a$, we obtain a $\mathbb Z$-stratification of ${\mathfrak}F$. Note that while this stratification depends on the chosen bases $B_n$, the complexes ${\mathfrak}F(a)$ and the strata ${\overline}{{\mathfrak}F(a)}$ do not depend on this choice.
\[Ex:double-complex\] Let $C_{\bullet\bullet}$ be a double complex and let ${\mathfrak}F=Tot^\oplus(C_{\bullet\bullet})$ be the associated total chain complex, see e.g. [@W Section 1.2]. Thus $F_n=\bigoplus_{i+j=n}C_{ij}$ and the two canonical filtrations of ${\mathfrak}F$ induce two canonical $\mathbb Z$-stratifications on ${\mathfrak}F$. For example, setting $F_n^a=C_{a,n-a}$ gives the *stratification by columns* of ${\mathfrak}F$, and the stratum at $a$ is ${\overline}{{\mathfrak}F(a)}=C_{a,\bullet}$ with differential the $a$th vertical differential of $C_{\bullet\bullet}$.
Let $X$ be a regular CW-complex, and let $\Bbbk$ be a commutative ring. Let $C_\bullet(X,\Bbbk)$ be the cellular chain complex of $X$ with coefficients in $\Bbbk$. Recall that the set $A_n$ of the $n$-cells $\sigma$ of $X$ (with closure $\bar\sigma$ and boundary $\dot\sigma$) index a decomposition of $$F_n = \operatorname{H}_n(X^n,X^{n-1},\Bbbk) =
\bigoplus_{\sigma\in A_n}\operatorname{H}_n(\bar\sigma,\dot\sigma, \Bbbk).$$ The face poset $P(X)=\bigcup_{n\ge 0}A_n$ of $X$ has partial order given by $\sigma\le \tau$ if and only if $\sigma\subseteq\bar\tau$. Thus we obtain the standard $P(X)$-stratification of $C_\bullet(X,\Bbbk)$ where $F_n^\sigma = \operatorname{H}_n(\bar\sigma, \dot\sigma ,\Bbbk)$ if $\sigma\in A_n$ and $F_n^\sigma=0$ otherwise.
\[Ex:Taylor-1\] Let $U$ be a finite set, let $\Delta^U$ be the geometric realization of the full simplex on the set of vertices $U$. Thus $\Delta^U$ is a regular CW-complex and the $n$-cells $[\sigma]$ of $\Delta^U$ are indexed by the elements $\sigma$ of the set $A_n$ of $n+1$-subsets of $U$. Let $X$ be a set of variables and let $M(X)=\{f \mid f{\! :}X{\longrightarrow}\mathbb N\}$ be the free abelian monoid on the set $X$ with operation value-wise addition. Let $\Bbbk$ be a commutative ring. We identify each element $f\in M(X)$ with the monomial $\prod_{y\in X}y^{f(y)}$ in the polynomial ring $\Bbbk[X]$ over $\Bbbk$ on the set of variables $X$. For each $v\in U$ let $m_v$ be a monomial in $\Bbbk[X]$ such that $m_u$ does not divide $m_v$ when $u\ne v$. More generally, with every subset $\sigma$ of $U$ we associate the monomial $m_\sigma=\operatorname{lcm}\{m_v\mid v\in\sigma\}$. The set of monomials $L=\{ m_\sigma \mid \sigma\subseteq U\}$, partially ordered by divisibility, is in fact a lattice called the *lcm-lattice* of the ideal $I$ generated in $\Bbbk [X]$ by the set of monomials $\{m_u\mid u\in U\}$, see [@GPW]. Let $d_n^{\sigma\tau}$ be the component of the differential of the cellular chain complex $C_\bullet(\Delta^U, \mathbb Z)$ that sends $\operatorname{H}_n\bigl({\overline}{[\sigma]},\dot{[\sigma]}, \mathbb Z\bigr)$ to $\operatorname{H}_{n-1}\bigl({\overline}{[\tau]},\dot{[\tau]}, \mathbb Z\bigr)$. The *Taylor resolution* [@T] of the ideal $I$ is the chain complex ${\mathfrak}T={\mathfrak}T(I)=(T_n,\delta_n)$ given by $
T_n=
\bigoplus_{\sigma\in A_n}
\operatorname{H}_n\bigl({\overline}{[\sigma]}, \dot{[\sigma]},\mathbb Z\bigr)\otimes\Bbbk[X]
$ with differential $\delta_n$ given by $$\delta_n\Big\vert_{\operatorname{H}_n\bigl({\overline}{[\sigma]}, \dot{[\sigma]},\mathbb Z\bigr)\otimes\Bbbk[X]}
=
\sum_{\tau\in A_{n-1}}d_n^{\sigma\tau}\otimes (m_\sigma/m_\tau).$$ Now by setting $
T_n^a =
\bigoplus_{m_\sigma=a}
\operatorname{H}_n\bigl({\overline}{[\sigma]}, \dot{[\sigma]},\mathbb Z\bigr)\otimes\Bbbk[X]
$ for every $a\in L$ we obtain an $L$-stratification on ${\mathfrak}T$. For $a\in L$ let $\Delta_{\le a}$ (respectively, $\Delta_{<a}$) be the union of all cells $[\sigma]$ of $\Delta^U$ such that $m_\sigma\le a$ (respectively, $m_\sigma<a$). It is straightforward to check $\Delta_{\le a}$ and $\Delta_{<a}$ are regular CW-subcomplexes of $\Delta^U$, and that the stratum ${\overline}{{\mathfrak}T(a)}$ is precisely $${\overline}{{\mathfrak}T(a)} =
C_\bullet(\Delta_{\le a},\Delta_{<a},\mathbb Z)\otimes\Bbbk[X],$$ where $
C_\bullet(\Delta_{\le a},\Delta_{<a},\mathbb Z)=
C_\bullet(\Delta_{\le a},\mathbb Z)/
C_\bullet(\Delta_{<a},\mathbb Z)
$ is the cellular chain complex of the pair $(\Delta_{\le a},\Delta_{<a})$.
Let $P$ be a poset.
\(a) Let $V$ be a vector field on ${\mathfrak}F$. We say that $V$ is *compatible* with a $P$-grading on ${\mathfrak}F$ if for each $a\in P$ and each $n$ the homotopy $V$ maps the degree $a$ homogeneous component $F_n^a$ into the homogeneous component $F_{n+1}^a$.
\(b) A *Lyapunov* (or *chain recurrent*) *structure for $V$ on ${\mathfrak}F$ with values in $P$* is a a $P$-stratification on ${\mathfrak}F$ together with a compatible vector field $V$.
\(c) If given a Lyapunov structure with values in $P$ for a vector field $V$ on ${\mathfrak}F$, then for each $a\in P$ the stratum $${\overline}{{\mathfrak}F(a)}={\mathfrak}F(a)\Big/\sum_{b<a}{\mathfrak}F(b)$$ is also called the *chain recurrent* chain complex of $V$ at $a$.
The following basic properties are immediate from the definitions.
\[P:stratification-induced\] Fix a $P$-stratification on ${\mathfrak}F$.
\(a) A compatible vector field $V$ induces for each $a\in P$ a canonical vector field ${\overline}{V(a)}$ on the stratum ${\overline}{{\mathfrak}F(a)}$.
\(b) Suppose that for each $a\in P$ we are given a vector field $W_a$ on the chain complex ${\overline}{{\mathfrak}F(a)}$ . Via the inclusion ${\overline}{{\mathfrak}F(a)}_n=F_n^a\subseteq F_n$, each $W_a$ induces a vector field (also denoted by $W_a$) on ${\mathfrak}F$ by setting $W_a(F_n^b)=0$ for $b\ne a$.
\(c) Let the vector fields $W_a$ be as in part (b), and let $W=\sum_{a\in P}W_a$. Then $W$ is a vector field on ${\mathfrak}F$ compatible with the given stratification, and such that ${\overline}{W(a)}=W_a$ for every $a\in P$.
For the applications we consider in this paper, one starts with a chain complex ${\mathfrak}F$ that has a natural $P$-stratification, but is “too big”. In that situation one way to find a smaller complex of the same homotopy type is to construct a vector field $W$ on ${\mathfrak}F$ compatible with the stratification and such that the dynamical system produced by iterating the flow $\Phi_W$ stabilizes in any given homological degree after finitely many iterations. Then one can take the stable image as the desired smaller chain complex. Proposition \[P:stratification-induced\](c), while an elementary consequence of the definitions, plays a crucial role in this approach. Note also that, while most of the theory developed so far works also for pre-vector fields, part (c) of this proposition is where we really require vector fields.
\[Ex:Taylor-3\] With notation as in Example \[Ex:Taylor-1\], for each $a\in L$ let $W_a$ be a vector field on $
C_\bullet(\Delta_{\le a},\Delta_{<a},\Bbbk) =
C_\bullet(\Delta_{\le a},\Delta_{<a},\mathbb Z)\otimes\Bbbk
$. Then $W_a\otimes 1$ is a vector field on the chain complex $
{\overline}{{\mathfrak}T(a)} =
C_\bullet(\Delta_{\le a},\Delta_{<a},\Bbbk)\otimes_\Bbbk\Bbbk[X]
$, hence by Proposition \[P:stratification-induced\] we obtain a vector field $W=\sum_{a\in L} W_a\otimes 1$ on the Taylor resolution ${\mathfrak}T$ compatible with the $L$-stratification.
Vector fields and splittings {#S:vector-fields-and-splittings}
============================
Here we briefly review a well-known class of vector fields whose flows are projections, hence already stable. They will be used later as building blocks to construct vector fields with desired asymptotic behaviour.
Recall that the chain complex ${\mathfrak}F$ *splits* or *is splittable* if it is chain homotopy equivalent to its homology complex $\operatorname{H}({\mathfrak}F)$. Clearly this happens if and only if there are morphisms of complexes $f{\! :}{\mathfrak}F{\rightarrow}\operatorname{H}({\mathfrak}F)$ and $g{\! :}\operatorname{H}({\mathfrak}F){\rightarrow}{\mathfrak}F$ and a chain homotopy $D$ on ${\mathfrak}F$ such that $fg=\operatorname{id}_{\operatorname{H}({\mathfrak}F)}$, and $gf=\Phi_D$ (we say that $D$ *realizes* a splitting of ${\mathfrak}F$). It is a standard calculation to verify that this is also equivalent to ${\mathfrak}F$ having a vector field $D$ such that $$\label{E:splitting}
\phi D\phi = \phi.$$ and $$\label{E:partial-splitting}
D\phi D = D.$$ Notice that if a vector field $D$ on ${\mathfrak}F$ satisfies then $\Phi_D$ is an idempotent on ${\mathfrak}F$, hence $\operatorname{Im}\Phi_D$ is homotopic to, and a direct summand of, ${\mathfrak}F$. However, unless an additional condition like holds, $\operatorname{Im}\Phi_D$ will not be isomorphic to $\operatorname{H}({\mathfrak}F)$. This motivates the following terminology.
\[D:splittings\] A *partial splitting* of ${\mathfrak}F$ is a vector field $D$ on ${\mathfrak}F$ that satisfies . A *splitting* of ${\mathfrak}F$ is a partial splitting that satisfies .
Let $D$ be a chain homotopy on ${\mathfrak}F$. Set $N_{n+1}=\operatorname{Im}(\phi D|_{F_n})$, let $M_n=\operatorname{Im}(D\phi|_{F_n})$, and let $C_n=\operatorname{Ker}(D\phi|_{F_n})\cap\operatorname{Ker}(\phi D|_{F_n})$. If $D$ satisfies or then both $D\phi$ and $\phi D$ are idempotents on ${\mathfrak}F$, and $F_n$ decomposes for each $n$ as $$\label{E:near-splitting}
F_n=N_{n+1}\oplus C_n\oplus M_n.$$ Since such a decomposition occurs also in other important cases, we are prompted to make the following more general definition.
We say that a chain homotopy $D$ is a *weak partial splitting* of ${\mathfrak}F$ if for each $n$ the decomposition holds, and $\phi D$ and $D\phi$ induce automorphisms on $N_{n+1}$ and $M_n$, respectively, for each $n$.
\[Ex:adjoint\] Suppose that $\Bbbk$ is a subfield of $\mathbb C$, and that ${\mathfrak}F$ is a chain complex of finite dimensional vector spaces over $\Bbbk$ such that each $F_i$ is given an inner product. For each $i$ let $\phi_i^*$ be the adjoint to $\phi_{i}$. Thus the chain homotopy $\phi^*$ is a vector field on ${\mathfrak}F$, and it is a standard fact from basic linear algebra that $\phi^*$ is a weak partial splitting of ${\mathfrak}F$.
\[P:weak-partial-splitting\] Let $D$ be a weak partial splitting of ${\mathfrak}F$.
\(a) For each $n$ the maps $D$ and $\phi$ induce isomorphisms between $N_n$ and $M_n$.
\(b) Define the chain homotopy $\widehat D$ on ${\mathfrak}F$ by $D(\phi D)^{-1}$ on $N_{n+1}$ and by $0$ on $M_n$ and $C_n$, for each $n$. Then $\widehat D$ is a partial splitting on ${\mathfrak}F$.
\(c) If $D$ satisfies or then $\widehat D = D\phi D(I - D\phi)$.
\(d) If $D$ satisfies then $\widehat D$ is a splitting.
\(e) If $D$ is a partial splitting then $\widehat D = D$.
Note that $D(N_n)=D\phi D(N_n)\subseteq M_n$ hence $D{\! :}N_n {\longrightarrow}M_n$ is a monomorphism. Furthermore, $\phi(M_n)=\phi D\phi(M_n)\subseteq N_n$, hence $\phi{\! :}M_n{\longrightarrow}N_n$ is also a monomorphism. Since $\phi(D(N_n))=N_n$ we see that $\phi{\! :}M_n{\longrightarrow}N_n$ is also epi hence isomorphism. Finally, since $D(\phi(M_n))=M_n$ we see that $D{\! :}N_n{\longrightarrow}M_n$ is also epi hence isomorphism. This takes care of part (a).
\(b) Since $\widehat D(F_n)=D(N_{n+1})=M_{n+1}$ by part (a), it is clear that $\widehat D^2=0$. Furthermore, $\widehat D\phi\widehat D=\widehat D=0$ on $C_n\oplus M_n$ for each $n$. Finally, on $N_{n+1}$ we have $\widehat D\phi\widehat D=\widehat D\phi D(\phi D)^{-1} =
\widehat D$.
\(c) Since $\phi D$ is the identity on $N_{n+1}$, we have that $\widehat D$ is given by $D$ on $N_{n+1}$ and by $0$ on $C_n\oplus M_n$. As $D\phi D(I- D\phi)$ satisfies the same condition, it equals $\widehat D$.
\(d) and (e) are immediate consequences of (b) and (c).
\[R:affine-combination\] Suppose $D_1, \dots, D_n$ are homotopies of ${\mathfrak}F$ that satsify , and $r_1,\dots, r_n$ are central elements of our base ring such that $r_1+\dots +r_n=1$. Then $D=r_1D_1 + \dots + r_nD_n$ satisfies , and therefore $\widehat D$ is a splitting by Proposition \[P:weak-partial-splitting\].
\[R:pseudo-inverses\] Suppose that $\Bbbk$ is a subfield of $\mathbb C$, and that ${\mathfrak}F$ is a chain complex of finite dimensional vector spaces over $\Bbbk$ such that each $F_i$ is given an inner product. For each $i$ let $\phi_i^*$ be the adjoint to $\phi_{i}$. As observed in Example \[Ex:adjoint\], the chain homotopy $\phi^*$ is a weak partial splitting of ${\mathfrak}F$. The partial splitting $\phi^{+}=\widehat{\phi^*}$ in this case is given by taking the well known Moore-Penrose [@M; @P; @Bj] pseudo-inverses $\phi_i^{+}$ of the maps $\phi_{i}$; in particular, $\phi^+$ is a splitting. Indeed, let $f_i(x)=\det(xI - \phi_{i}\phi_i^*)$ be the characteristic polynomial of $\phi_{i}\phi_i^*$. Since $\phi_{i}\phi_i^*$ is self-adjoint, we have $f_i(x)=x^{n_i}g_i(x)$ where $n_i = \dim F_{i-1} - \operatorname{rank}\phi_{i}$, and $g_i(0)\ne 0$. It follows that $$1 - \frac{g_i(x)}{g_i(0)} = x p_i(x)$$ for a canonically and explicitly determined by $\phi_{i}$ and $\phi_i^*$ polynomial $p_i(x)$. It is now standard to check that $p_i(\phi_{i}\phi_i^*)$ is the inverse of $\phi_{i}\phi_i^*$ on $N_{i}=\operatorname{Im}(\phi_{i}\phi_i^*)$, and that $\phi_{i}\phi_i^*p_i(\phi_{i}\phi_i^*)$ is the orthogonal projection of $F_{i-1}$ onto $N_{i}$. Therefore, by Proposition \[P:weak-partial-splitting\], we obtain the explicit formula $$\label{E:pseudo-inverse}
\phi_i^{+} =
\phi_i^* p_i(\phi_{i}\phi_i^*)\phi_{i}\phi_i^* p_i(\phi_{i}\phi_i^*).$$ It is trivial to check from this formula that $\phi_i^{+}$ satisfies the Moore-Penrose conditions [@GVL p. 290] hence equals the Moore-Penrose pseudo-inverse of $\phi_{i}$; in particular, we recovered its standard canonical explicit polynomial expression in terms of $\phi_{i}$ and its adjoint.
\[Ex:pseudo-inverses\] Consider the chain complex of $\mathbb Q$-vector spaces $$0 {\longleftarrow}\mathbb Q
\xleftarrow[\phi_1]{
[
\begin{smallmatrix}
1 & 1 & 1 & 1 & 1 & 1 & 1
\end{smallmatrix}
]
}
\mathbb Q^{7}
\xleftarrow[\phi_2]{
\left[
\begin{smallmatrix}
{\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 \\
-1 & -1 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 \\
{\hphantom{-}}0 & {\hphantom{-}}0 & -1 & -1 & {\hphantom{-}}0 & {\hphantom{-}}0 \\
{\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 & -1 & -1 \\
{\hphantom{-}}1 & {\hphantom{-}}0 & {\hphantom{-}}1 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 \\
{\hphantom{-}}0 & {\hphantom{-}}1 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}1 & {\hphantom{-}}0 \\
{\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}1 & {\hphantom{-}}0 & {\hphantom{-}}1
\end{smallmatrix}
\right]
}
\mathbb Q^{6}
{\longleftarrow}0,$$ where the inner products are the ones making the standard basis on each $\mathbb Q^m$ orthonormal. We have $$\phi_2\phi_2^* =
\left[
\begin{smallmatrix}
0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}0 \\
0 & {\hphantom{-}}2 & {\hphantom{-}}0 & {\hphantom{-}}0 & -1 & -1 & {\hphantom{-}}0 \\
0 & {\hphantom{-}}0 & {\hphantom{-}}2 & {\hphantom{-}}0 & -1 & {\hphantom{-}}0 & -1 \\
0 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}2 & {\hphantom{-}}0 & -1 & -1 \\
0 & -1 & -1 & {\hphantom{-}}0 & {\hphantom{-}}2 & {\hphantom{-}}0 & {\hphantom{-}}0 \\
0 & -1 & {\hphantom{-}}0 & -1 & {\hphantom{-}}0 & {\hphantom{-}}2 & {\hphantom{-}}0 \\
0 & {\hphantom{-}}0 & -1 & -1 & {\hphantom{-}}0 & {\hphantom{-}}0 & {\hphantom{-}}2
\end{smallmatrix}
\right]$$ which has characteristic polynomial $
f_2(x) = \det(xI - \phi_2\phi^*_2) =
x^7 - 12x^6 + 54x^5 - 112x^4 + 105x^3 - 36x^2
$, therefore $$p_2(x) =
\frac{1}{36}(x^4 - 12x^3 + 54x^2 - 112x + 105),$$ and thus we obtain by that $$\phi_2^+ =
\frac{1}{12}
\left[
\begin{smallmatrix}
0& {-5}&{\hphantom{-}}3& {-1}&{\hphantom{-}}5& {-3}&{\hphantom{-}}1\\
0& {-5}& {-1}&{\hphantom{-}}3& {-3}&{\hphantom{-}}5&{\hphantom{-}}1\\
0&{\hphantom{-}}3& {-5}& {-1}&{\hphantom{-}}5&{\hphantom{-}}1& {-3}\\
0& {-1}& {-5}&{\hphantom{-}}3& {-3}&{\hphantom{-}}1&{\hphantom{-}}5\\
0&{\hphantom{-}}3& {-1}& {-5}&{\hphantom{-}}1&{\hphantom{-}}5& {-3}\\
0& {-1}&{\hphantom{-}}3& {-5}&{\hphantom{-}}1& {-3}&{\hphantom{-}}5
\end{smallmatrix}
\right].$$ Doing a similar (but much simpler) computation for $\phi^+_1$ yields that the Moore-Penrose splitting $\phi^+$ has the form $$0{\longrightarrow}\mathbb Q
\xrightarrow[\phi_1^+]{
\left[
\begin{smallmatrix}
1/7 \\
1/7 \\
1/7 \\
1/7 \\
1/7 \\
1/7 \\
1/7
\end{smallmatrix}
\right]
}
\mathbb Q^7
\xrightarrow[\phi_2^+]{
1/12
\left[
\begin{smallmatrix}
0& {-5}&{\hphantom{-}}3& {-1}&{\hphantom{-}}5& {-3}&{\hphantom{-}}1\\
0& {-5}& {-1}&{\hphantom{-}}3& {-3}&{\hphantom{-}}5&{\hphantom{-}}1\\
0&{\hphantom{-}}3& {-5}& {-1}&{\hphantom{-}}5&{\hphantom{-}}1& {-3}\\
0& {-1}& {-5}&{\hphantom{-}}3& {-3}&{\hphantom{-}}1&{\hphantom{-}}5\\
0&{\hphantom{-}}3& {-1}& {-5}&{\hphantom{-}}1&{\hphantom{-}}5& {-3}\\
0& {-1}&{\hphantom{-}}3& {-5}&{\hphantom{-}}1& {-3}&{\hphantom{-}}5
\end{smallmatrix}
\right]
}
\mathbb Q^6 {\longrightarrow}0.$$
\[Ex:Taylor-0\] With notation as in Example \[Ex:Taylor-3\], take $\Bbbk =\mathbb Q$. For each $n$ we have $
C_n(\Delta_{\le a},\Delta_{<a},\mathbb Z) =
\bigoplus_{m_\sigma=a}\operatorname{H}_n\bigl({\overline}{[\sigma]},\dot{[\sigma]},\mathbb Z\bigr)
$. Since each $\operatorname{H}_n\bigl({\overline}{[\sigma]},\dot{[\sigma]},\mathbb Z\bigr)$ is just a copy of $\mathbb Z$ it has a unique up to sign free generator. This yields a natural up to signs basis of $C_n=C_n(\Delta_{\le a},\Delta_{<a},\mathbb Q)$ and therefore induces a natural independent of the signs inner product on $C_n$ in which this basis is orthonormal. Thus, as in Remark \[R:pseudo-inverses\] we obtain an explicit natural splitting $d_a^{+}$ of $C_\bullet(\Delta_{\le a},\Delta_{<a},\mathbb Q)$ by taking the Moore-Penrose pseudo-inverses of its differentials. This induces an explicit natural splitting $\delta_a^{+}=d_a^+\otimes 1$ on the stratum ${\overline}{{\mathfrak}T(a)}$, hence, by Proposition \[P:stratification-induced\], an explicit natural vector field $\delta^{+}=\sum\delta_a^{+}$ on the Taylor resolution ${\mathfrak}T$ compatible with the $L$-stratification.
Asymptotic properties {#S:asymptotic-properties}
=====================
Just as in the classical dynamical systems case, where the asymptotic properties of the flow of a vector field on a smooth compact manifold can be controlled by a Lyapunov function, see e.g. [@C; @Fr], the behaviour in the long run of the flow of a vector field on a chain complex can be controlled by a Lyapunov structure.
To understand this mechanism, recall that if $P$ is a poset and $a\in P$ then the *dimension* of $a$ in $P$ is the supremum $d_P(a)$ of the lengths $n$ of strictly increasing chains $a_0<\dots < a_n=a$ in $P$ terminating at $a$. The *dimension* $\dim P$ of $P$ is the supremum of the dimensions $d_P(a)$ of all the elements $a$ of $P$ . The following has been a main motivating example for us.
\[Ex:Taylor-4\] With notation as in Example \[Ex:Taylor-3\], suppose that $W_a$ is a splitting of ${\overline}{{\mathfrak}T(a)}$ for each $a\in L$. Then, although not explicity included in its statement, the proof of [@Y Lemma 4.1] also shows that $\Phi_W^{k+1}=\Phi_W^k$ whenever $k>\dim L$. Thus, $\dim L$ controls how quickly the flow of $W$ stabilizes. Furthermore, this implies that $\Phi_W^{1+\dim L}$ is a projection of ${\mathfrak}T$ onto a chain homotopic direct summand ${\mathfrak}M_W$ of ${\mathfrak}T$.
Another relevant instance of this mechanism is the proof of the basic perturbation lemma from homological perturbation theory, see e.g. [@GL Basic Perturbation Lemma 2.4.1], its argument going back to [@G; @B; @S]. It can be understood in our language as giving a sufficient condition that certain vector fields have flows that stabilize in any given homological degree after finitely may iterations. As Example \[Ex:double-complex\] suggests, a similar interpretation can be given also to Eagon’s construction of a Wall complex [@E Theorem 1.2]; see also [@EFS Lemma 3.5] for a categorical version of this construction.
The next result, which is the main result in this section, can be thought of as an elaborate and explicit generalization of Yuzvinsky’s argument we referred to in Example \[Ex:Taylor-4\]. It can also be interpreted as a detailed analysis, from our dynamical systems viewpoint, of a special case of the basic perturbation lemma of homological perturbation theory. Either way, this is the key technical fact needed for the applications we consider in this paper.
\[T:asymptotic\] Fix a $P$-grading on ${\mathfrak}F$. For each $c\in P$ let $V_c$ be a vector field on the chain complex ${\overline}{{\mathfrak}F(c)}$, and let $V=\oplus_c V_c$ be the induced vector field on ${\mathfrak}F$. Consider the subposet $P_n=\{c\in P\mid F_n^c\ne 0\}$ of $P$, let $a\in P_n$ be of finite dimension, suppose that for each $c\in P_n$ with $c\le a$ the vector field $V_c$ is a partial splitting on ${\overline}{{\mathfrak}F(c)}$, and let $F_n^c=N^c_{n+1}\oplus C^c_n\oplus M^c_n$ be the corresponding decomposition as in .
Then for any collection of nonnegative integers $\{k_c\mid c\le a\}$ we have $$\sum_{c\le a}\Phi_V^{k_c}(C^c_n)=
\bigoplus_{c\le a}\Phi_V^{k_c}(C^c_n),$$ and each restriction $\Phi_V^{k_c}{\! :}C^c_n {\longrightarrow}\Phi_V^{k_c}(C^c_n)$ is an isomorphism. Furthermore, for each integer $k > d_{P_n}(a)$ we also have:
1. $\Phi_V^k(M_n^a)=0$;
2. $[\Phi_V -\operatorname{id}](C_n^a)\subseteq\bigoplus_{b<a}M_n^b$;
3. $[\Phi_V^{k}-\Phi_V^{k-1}](C_n^a)=0$;
4. $
\Phi_V^k(N_{n+1}^a)\subseteq\bigoplus_{b<a}\Phi_V^{d_{P_n}(b)}(C_n^b);
$
5. $[\Phi_V^{k+1}-\Phi_V^k](F_n^a)=0$;
6. $
\Phi_V^k(F_n^a)\subseteq
\bigoplus_{b\le a}\Phi_V^{d_{P_n}(b)}(C_n^b).
$
7. $[\Phi_V^{k+1}-\Phi_V^k]\bigl({\mathfrak}F(a)_n\bigr)=0$.
8. $
\Phi_V^k\bigl({\mathfrak}F(a)_n\bigr)=
\bigoplus_{b\le a}\Phi_V^{d_{P_n}(b)}(C_n^b).
$
We use induction on $t=d_{P_n}(a)$. When $t=0$ then $a$ is a minimal element of $P_n$, hence ${\mathfrak}F(a)_n={\overline}{{\mathfrak}F(a)}_n$, therefore $V=V_a$ on ${\mathfrak}F(a)_n$ and $\Phi_{V_a}=\Phi_V$ on ${\mathfrak}F(a)_n$; hence all claims of the Theorem are trivially satisfied. This we assume that $t\ge 1$, and that our assertions are true for all elements of $P_n$ of dimension strictly less than $t$.
By the definitions and Proposition \[P:weak-partial-splitting\] we have for every $b\in P_n$ with $b\le a$ that $V_b(C_n^b\oplus M_n^b)=0$, and $V_b(N^b_{n+1})=M^b_{n+1}$, and ${\overline}{\phi(b)}(M^b_{n+1})=N^b_{n+1}$, as well as $$\begin{aligned}
\Phi_{V_b} &=
\begin{cases}
0 &\text{on } N^b_{n+1}\oplus M^b_n; \\
\operatorname{id}&\text{on } C^b_n;
\end{cases}
\\
V_b{\overline}{\phi(b)} &=
\begin{cases}
0 &\text{on } N^b_{n+1}\oplus C^b_n; \\
\operatorname{id}&\text{on } M^b_n;
\end{cases}
\\
{\overline}{\phi(b)}V_b &=
\begin{cases}
0 &\text{on } C^b_n\oplus M^b_n; \\
\operatorname{id}&\text{on } N^b_{n+1}.
\end{cases}
$$ Therefore $\phi(M_n^a)\subseteq N_{n}^a\oplus \bigoplus_{b<a} F_{n-1}^b$, and $$[\operatorname{id}- V\phi](M_n^a)\subseteq\bigoplus_{b<a}M_n^b.$$ Hence $\Phi_V(M_n^a)\subseteq\bigoplus_{b<a}M_n^b$ and (1) follows from our induction hypothesis.
Similarly, since $V({\mathfrak}F(a)_{n-1})= \bigoplus_{b\le a}M_n^b$, we have $V\phi(C^a_n)\subseteq \bigoplus_{b<a}M^b_n$. Also, $\phi V(C^a_n) = 0$ for trivial reasons, and (2) follows. In particular, $\Phi_V(C^a_n)\subseteq C^a_n\oplus \bigoplus_{b<a}M^b_n$ and the $C^a_n$-component of $\Phi_V(x)$ is $x$ for each $x\in C^a_n$. Therefore the same is true for the $C^a_n$-component of $\Phi_V^{k_a}(x)$. It is now immediate that the restriction $\Phi_V^{k_a}{\! :}C^a_n {\longrightarrow}\Phi_V^{k_a}(C^a_n)$ is an isomorphism, and that the sum $\sum_{c\le a}\Phi^{k_c}(C^c_n)$ is a direct sum.
Next we note that (3) is an immediate consequence of (1) and (2), and proceed with the proof of (4). Since $\Phi_{V_a}(N_{n+1}^a)=0$ we get $\Phi_V(N_{n+1}^a)\subseteq \bigoplus_{b<a}F_n^b$, and thus (4) follows by our induction hypothesis using (6). Since $\Phi_{V_a}(F_n^a)=C^a_n$, we see that $\Phi_V(F_n^a)\subseteq C_n^a\oplus\bigoplus_{b<a}F_n^b$, hence (5) and (6) follow from (3) and the induction hypothesis. Now (7) follows from (5) and the induction hypothesis. Finally, (8) is clear from (1), (3), (4), and the induction hypothesis.
\[C:splittings-are-minimal\] Suppose $a$ has finite dimension in both $P_n$ and $P_{n-1}$. Suppose that $V_a$ is a splitting of ${\overline}{{\mathfrak}F(a)}$, and that $V_c$ is a partial splitting of ${\overline}{{\mathfrak}F(c)}$ for each $c\in P_n\cup P_{n-1}$ with $c\le a$.
Then $
\phi\bigl(\Phi^{d_{P_n}(a)}_V(C^a_n)\bigr)\subseteq
\bigoplus_{b<a}\Phi^{d_{P_{n-1}}(b)}_V(C^b_{n-1}).
$
Since ${\overline}{\phi(a)}(C^a_n)=0$ and $\Phi_V(C^a_n)\subseteq C^a_n\oplus\bigoplus_{b<a} M^b_n$ we have $\phi\Phi_V(C^a_n)\subseteq \sum_{b<a}{\mathfrak}F(b)_{n-1}$. As $\phi$ commutes with $\Phi_V$ the desired conclusion is now immediate from Theorem \[T:asymptotic\](8).
Toric rings {#S:toric-rings}
===========
In this section we address the problem of constructing minimal free resolutions of toric rings. These rings arise in geometry as the affine coordinate rings of toric varieties, and in combinatorics and other fields as semigroup rings of affine pointed semigroups (recall that a pointed affine semigroup is one that is isomorphic to a finitely generated subsemigroup of some $\mathbb N^m$). Toric rings are the subject of active current research, see e.g. [@P; @MS; @F] and the references there, and canonical but in general non-minimal resolutions have been constructed in [@BS; @TV1]. For more details on the properties of toric rings and affine pointed semigroups we refer the reader to the excellent expositions in [@MS] and [@P]. We will now describe how to construct their minimal resolutions in a canonical and intrinsic fashion.
Let $Q$ be an affine pointed semigroup. Then $Q$ has a unique minimal generating set $A$. Let $X=\{x_a \mid a\in A\}$ be a set of variables, and let $\Bbbk$ be a field. The map $x_a\mapsto a$ induces a surjective homomorphism of $\Bbbk$-algebras $\Bbbk[X]{\longrightarrow}\Bbbk[Q]$, and a natural $Q$-grading on the polynomial ring $\Bbbk[X]$ over $\Bbbk$ in the set of variables $X$ such that $\deg_Qx_a = a$. This makes the toric ring $\Bbbk[Q]$ into a $Q$-graded $\Bbbk[X]$-module. It is well known that $\Bbbk[Q]$ has a minimal free $Q$-graded resolution over $\Bbbk[X]$. The *Betti degrees* of $Q$ over $\Bbbk$ are the elements of $Q$ that appear as $Q$-degrees of basis elements of the free modules in such a minimal resolution. They can be computed combinatorially directly, without having to compute a minimal resolution first, see e.g. [@MS Theorem 9.2].
The semigroup $Q$ has a natural partial order given by $q\le r$ if and only if $q+s=r$ for some $s\in Q$. Equivalently, $q\le r$ if and only if there is a monomial $m$ in the variables $X$ such that $q+\deg_Qm = r$. This induces a partial order on the set $B=B(Q,\Bbbk)$ of Betti degrees of $Q$ over $\Bbbk$. Let $\mathcal Q$ be the category with objects the elements of $Q$, with morphisms from $q$ to $r$ given by those monomials $m$ in the variables $X$ such that $q +\deg_Qm = r$. Composition of morphisms is given by multiplication of monomials. The full subcategory of $\mathcal Q$ with objects the Betti degrees $B$ of $Q$ over $\Bbbk$ is called the *Betti category* of $Q$ over $\Bbbk$ and denoted by $\mathcal B=\mathcal B(Q,\Bbbk)$.
In [@TV1 Theorem 9.2] a (non-minimal in general) canonical based finite free $Q$-graded resolution ${\mathfrak}F$ of $\Bbbk[Q]$ over $\Bbbk[X]$ was constructed as follows. In homological degree $n$ the module $F_n$ is the free $\Bbbk[X]$-module with homogeneous basis given by all sequences $q_0{\longrightarrow}q_1 {\longrightarrow}\dots {\longrightarrow}q_n$ of composable nonidentity morphisms in the Betti category $\mathcal B(Q,\Bbbk)$, with the $Q$-degree of the sequence being $q_n$. The differential $\phi_n{\! :}F_n{\longrightarrow}F_{n-1}$ is defined as the sum $\phi_n = \sum_{i=0}^n(-1)^i\partial_i$, where the map $\partial_i$ sends the sequence $$\label{E:morphism-sequence}
q_0\xrightarrow{m_1} q_1
\xrightarrow{m_2} \dots \xrightarrow{m_{n-1}}
q_{n-1}\xrightarrow{m_n} q_n$$ to $$\begin{cases}
q_0\rightarrow\dots\rightarrow q_{i-1}\xrightarrow{m_im_{i+1}}
q_{i+1}\rightarrow\dots\rightarrow q_n
&\text{if } 0<i<n;
\\
q_1\rightarrow\dots\rightarrow q_n
&\text{if } i = 0;
\\
m_n ( q_0\rightarrow\dots\rightarrow q_{n-1} )
&\text{if } i=n.
\end{cases}$$ Using the partial order on the set $B$ of Betti degrees we have the natural $B$-stratification of ${\mathfrak}F$ where $F_n^q$ is the free direct summand of $F_n$ with basis the set of all composable sequences of nonidentity morphisms in $\mathcal B$ such that $q_n=q$.
Let us describe the corresponding stratum ${\overline}{{\mathfrak}F(q)}={\mathfrak}F(q)/\sum_{p<q}{\mathfrak}F(p)$. Let $\Delta_{\le q}$ (respectively $\Delta_{<q}$) be the semi-simplicial subset, see e.g. [@W Sections 8.1 and 8.2], of the nerve of $\mathcal B$, consisting of all sequences of composable nonidentity morphisms with $q_n\le q$ (respectively $q_n<q$). It is immediate from the definitions that $${\overline}{{\mathfrak}F(q)} =
\Bbbk[X]\otimes_{\mathbb F} C_{\bullet}(\Delta_{\le q},\Delta_{<q},\mathbb F)$$ where $
C_{\bullet}(\Delta_{\le q},\Delta_{<q},\mathbb F) =
C_{\bullet}(\Delta_{\le q},\mathbb F)/
C_{\bullet}(\Delta_{<q},\mathbb F)
$ is the simplicial chain complex of the pair $(\Delta_{\le q},\Delta_{<q} )$ with coefficients in the prime field $\mathbb F$ of the field $\Bbbk$.
We are now all set to use the machinery we have developed in the previous sections. As a first step we have
\[T:toric-minimal\] Let $\mathbb F$ be the prime field of the field $\Bbbk$. For each $q\in B$ let $W_q$ be a splitting of the chain complex $C_\bullet(\Delta_{\le q},\Delta_{<q},\mathbb F)$, and let $W=\sum 1\otimes W_q$ be the induced vector field on ${\mathfrak}F$ compatible with the $B$-stratification. Let $d=\dim B$.
Then $\Phi^{d+1}_W$ is a projection of ${\mathfrak}F$ onto a free direct summand ${\mathfrak}M_W$. Furthermore, ${\mathfrak}M_W$ is a minimal $Q$-graded free resolution of the toric ring $\Bbbk[Q]$ over $\Bbbk[X]$.
By construction the flow $\Phi_W$ on ${\mathfrak}F$ preserves the $Q$-grading, hence so does its iterate $\Phi^{d+1}_W$. Now $\Phi^{d+1}_W$ is a projection onto its image ${\mathfrak}M_W$ by Theorem \[T:asymptotic\], and, by the same theorem we have in homological degree $n$ that $$({\mathfrak}M_W)_n =
\bigoplus_{q\in B_n}
\Phi_W^{d(q)}(\Bbbk[X]\otimes_{\mathbb F} C^q_n).$$ Thus ${\mathfrak}M_W$ is chain homotopic to ${\mathfrak}F$ and a $Q$-graded projective, hence free, direct summand of ${\mathfrak}F$. In particular ${\mathfrak}M_W$ is a $Q$-graded free resolution of $\Bbbk[Q]$ over $\Bbbk[X]$. Since elements of $\Phi_W^{d(q)}(\Bbbk[X]\otimes_{\mathbb F} C^q_n)$ are homogeneous of degree $q$, it is immediate from Corollary \[C:splittings-are-minimal\] that ${\mathfrak}M_W$ is a minimal resolution as desired.
Next, we indicate how to make a “universal” choice for the splittings $W_q$ in Theorem \[T:toric-minimal\]. In characteristic zero we just take the splittings induced by taking Moore-Penrose pseudo-inverses. In characteristic $p>0$ the situation is more subtle, and we obtain universality only after passing to a certain purely transcendental extension of $\Bbbk$. The main idea is, since in general there seems to be no canonical choice of a splitting over $\Bbbk$, we take the finitely many splittings over $\mathbb F_p$ and average them with generic weights to obtain a canonical splitting over a canonical transcendental extension of $\mathbb F_p$. In spite of its apparent simplicity, this approach provides an esssential step in overcoming a previously unobserved obstruction to such canonical constructions, see Remarks \[R:toric-remarks\].
The following result is an immediate consequence of Proposition \[P:weak-partial-splitting\], Remarks \[R:affine-combination\] and \[R:pseudo-inverses\], and Theorem \[T:toric-minimal\].
\[T:toric-canonical\] Let $\mathbb F$ be the prime subfield of $\Bbbk$.
\(a) Let $\mathbb F=\mathbb Q$. For each $q\in B$, the explicit $\mathbb Q$-basis of the homological degree $n$ component $C_n(\Delta_{\le q},\Delta_{<q},\mathbb Q)$ of $C_\bullet(\Delta_{\le q},\Delta_{<q},\mathbb Q)$, given by the sequences with $q_n=q$, induces a canonical inner product for which this basis is orthonormal. Taking $\delta_q^{+}$ to be the splitting given by the Moore-Penrose pseudo-inverses, see Remark \[R:pseudo-inverses\], of the differentials of $C_\bullet(\Delta_{\le q},\Delta_{<q},\mathbb Q)$ yields a canonical explicit vector field $\phi^{+}=\sum 1\otimes\delta_q^{+}$ on the chain complex ${\mathfrak}F$, hence an explicit canonical minimal resolution ${\mathfrak}M_{\phi^{+}}$ and an explicit canonical projection $\Phi_{\phi^{+}}^{d+1}$ of ${\mathfrak}F$ onto ${\mathfrak}M_{\phi^{+}}$.
\(b) Let $\mathbb F=\mathbb F_p$ where $p\ge 2$ is a prime. For each $q\in B$ let $Spl_Q(q)$ be the (finite) set of all splittings of the chain complex $C_\bullet(\Delta_{\le q},\Delta_{<q},\mathbb F_p)$. Let $Y_q=\{y_\sigma \mid \sigma\in Spl_Q(q)\}$ be a set of indeterminates, and let $Y=\sqcup_{q\in B} Y_q$ be the disjoint union. Let $\mathbb F'=\mathbb F_p(Y)$ (repsectively, $\Bbbk'=\Bbbk(Y)$) be the purely transcendental extension of $\mathbb F_p$ (respectively, $\Bbbk$) on the set of variables $Y$. Then by Remark \[R:affine-combination\] $$W_q =
\frac{\sum_{\sigma\in Spl_Q(q)}\ y_{\sigma}\otimes \sigma}
{\sum_{\sigma\in Spl_Q(q)}\ y_{\sigma}}$$ is a canonical weak partial splitting of $\mathbb F'\otimes C_\bullet(\Delta_{\le q},\Delta_{<q},\mathbb F_p)$, and taking $\widehat{W}_q$ to be the induced by Proposition \[P:weak-partial-splitting\](bd) canonical splitting yields a canonial vector field $W=\sum_{q\in B}1\otimes\widehat{W}_q$ on ${\mathfrak}F'=\Bbbk'\otimes_\Bbbk{\mathfrak}F$ hence a canonical minimal resolution ${\mathfrak}M_W$ of $\Bbbk'[Q]$ over $\Bbbk'[X]$ and a canonical projection $\Phi_W^{d+1}$ of ${\mathfrak}F'$ onto ${\mathfrak}M_W$.
For the remainder of this section we elaborate on how the canonical minimal resolutions constructed in Theorem \[T:toric-canonical\] respect the symmetries of the pointed affine semigroup $Q$. In order to be more specific, we will use the following notation, which will come in handy also in the next section.
Let $\Bbbk$ be a commutative ring, let $S$ be a $\Bbbk$-algebra, and let $M$ be an $S$-module. Given a homomorphism $f{\! :}R{\longrightarrow}S$ of $\Bbbk$-algebras, we write $f_*M$ for the induced $R$-module structure on $M$. Thus the underlying abelian groups of $M$ and $f_*M$ are the same set, and the $R$-module structure on $f_*M$ is given for $r\in R$ and $m\in M$ by $r\cdot m=f(r)m$. Clearly, if $\phi{\! :}M{\longrightarrow}N$ is a homomorphism of $S$-modules then the same map of underlying sets is also a homomorphism $\phi{\! :}f_*M{\longrightarrow}f_*N$ of $R$-modules. Furthermore, when $f$ and $g$ are endomorphisms of $S$ then $(fg)_*M = g_*(f_*M)$, and $g{\! :}S {\longrightarrow}g_* S$ is a homomorphism of $S$-modules.
Now let $G$ be a group of automorphisms of the monoid $Q$. From the combinatorial description of the Betti degrees of $Q$ over $\Bbbk$ it is immediate that the set $B$ is $G$-invariant. The action of $G$ on $Q$ lifts canonically to an action of $G$ as automorphisms on the semigroup algebra $\Bbbk[Q]$, and since $G$ acts as permutations on the minimal generating set $A$ of $Q$, this induces a canonical action of $G$ on the set $X$, and hence on the polynomial ring $\Bbbk[X]$. For $\gamma\in G$ we abuse notation and denote by $\gamma$ also the induced automorphisms on $\Bbbk[Q]$ and $\Bbbk[X]$. We consider the $\Bbbk[X]$-modules $\gamma_*\Bbbk[X]$ and $\gamma_*\Bbbk[Q]$ as $Q$-graded via $\deg_Q x_a=\gamma^{-1}(a)$ and $\deg_Q q = \gamma^{-1}(q)$ for $a\in A$ and $q\in Q$. Then the isomorphisms $\gamma{\! :}\Bbbk[X]{\longrightarrow}\gamma_*\Bbbk[X]$ and $\gamma{\! :}\Bbbk[Q]{\longrightarrow}\gamma_*\Bbbk[Q]$ are isomorphisms of $Q$-graded $\Bbbk[X]$-modules. More generally, for any $Q$-graded $\Bbbk[X]$-module $M$ we consider the $Q$-grading on $\gamma_*M$ where every homogeneous element of degree $q$ in $M$ is homogeneous of degree $\gamma^{-1}(q)$ in $\gamma_*M$.
Next, the action of $G$ on $B$ and on $\Bbbk[X]$ induces an action on the Betti category of $Q$ over $\Bbbk$, and under this action the sequence gets sent by $\gamma\in G$ to $$\gamma(q_0)\xrightarrow{\gamma(m_1)} \gamma(q_1)
\xrightarrow{\gamma(m_2)} \dots \xrightarrow{\gamma(m_{n-1})}
\gamma(q_{n-1})\xrightarrow{\gamma(m_n)}\gamma(q_n).$$ Therefore $\gamma$ lifts to an isomorphism $
\gamma{\! :}C_\bullet(\Delta_{\le q},\Delta_{<q},\mathbb F) {\longrightarrow}C_\bullet(\Delta_{\le\gamma(q)},\Delta_{<\gamma(q)},\mathbb F)
$. In characteristic zero this sends orthonormal bases to orthonormal bases, hence takes $\delta_q^{+}$ to $\delta_{\gamma(q)}^{+}$. In characteristic $p>0$ this produces a bijection of sets of splittings $\gamma{\! :}Spl_Q(q){\longrightarrow}Spl_Q(\gamma(q))$, and therefore a $G$-action on the set $Y$, hence also a field automorphism $\gamma$ of $\Bbbk'$ over $\Bbbk$; which in turn induces a $\Bbbk$-algebra automorphism $\gamma$ of $\Bbbk'[X]$. Since the module $F_n$ (respectively, $\Bbbk'\otimes_\Bbbk F_n$) is free over $\Bbbk[X]$ (respectively, $\Bbbk'[X]$) with basis all sequences , the actions of $G$ on these sequences and on $\Bbbk[X]$ (respectively, $\Bbbk'[X]$) induce an action on $F_n$ (respectively, $F'_n=\Bbbk'\otimes_\Bbbk F_n$) where $\gamma\in G$ sends a linear combination $\sum c_b b$ of sequences $b$ with coefficients $c_b\in\Bbbk[X]$ (respectively, $c_b\in\Bbbk'[X]$) to the linear combination $\sum \gamma(c_b)\gamma(b)$. It is immediate that this produces for each $\gamma\in G$ a $\Bbbk$-automorphism ${\mathfrak}F(\gamma)$ (respectively, ${\mathfrak}F'(\gamma)$) of the resolution ${\mathfrak}F$ (respectively, ${\mathfrak}F'$) such that ${\mathfrak}F(\gamma){\! :}{\mathfrak}F{\longrightarrow}\gamma_*{\mathfrak}F$ (respectively, ${\mathfrak}F'(\gamma){\! :}{\mathfrak}F'{\longrightarrow}\gamma_*{\mathfrak}F'$) is an isomorphism of complexes of free $Q$-graded $\Bbbk[X]$-modules (respectively, $\Bbbk'[X]$-modules). Furthermore, it is straightforward to verify that ${\mathfrak}F(\gamma)$ (respectively, ${\mathfrak}F'(\gamma)$) sends each ${\mathfrak}F(q)$ to ${\mathfrak}F(\gamma(q))$ (respectively, ${\mathfrak}F'(q)$ to ${\mathfrak}F'(\gamma(q))$), and commutes with the vector fields $\phi^{+}$ in characteristic zero (respectively, $W$ in characteristic $p>0$). It follows that ${\mathfrak}F(\gamma)$ (respectively, ${\mathfrak}F'(\gamma)$) commutes with the flow $\Phi_{\phi^{+}}$ (respectively, $\Phi_W$) therefore it induces a $\Bbbk$-automorphism ${\mathfrak}M(\gamma)$ on the minimal resolution ${\mathfrak}M_{\phi^+}$ (respectively, ${\mathfrak}M_W$).
Summarizing the discussion above, we have proved the following result.
\[T:toric-intrinsic\] Let $Q$ be a pointed affine semigroup with group of automorphisms $G$, let $\Bbbk$ be a field, and let the set of variables $X$ be as above.
Then there exist
- a canonical explicitly constructed finitely generated field extension $\Bbbk'$ of $\Bbbk$,
- a canonical explicitly constructed minimal free $Q$-graded resolution ${\mathfrak}M$ of $\Bbbk'[Q]$ over $\Bbbk'[X]$,
- a canonical explicitly constructed action of $G$ as automorphisms of $\Bbbk'$ over $\Bbbk$, and
- a canonical explicitly constructed homomorphism ${\mathfrak}M{\! :}G{\longrightarrow}Aut_\Bbbk({\mathfrak}M)$ of $G$ to the group of automorphisms of the chain complex of $\Bbbk$-vector spaces ${\mathfrak}M$;
such that for each $\gamma\in G$ the map ${\mathfrak}M(\gamma){\! :}{\mathfrak}M {\longrightarrow}\gamma_*{\mathfrak}M$ is an isomorphism of complexes of free $Q$-graded $\Bbbk'[X]$-modules that lifts the isomorphism of $Q$-graded $\Bbbk'[X]$-modules $\gamma{\! :}\Bbbk'[Q]{\longrightarrow}\gamma_*\Bbbk'[Q]$. In characteristic zero one can take $\Bbbk'=\Bbbk$, and in characteristic $p>0$ one can take $\Bbbk'$ to be a purely transcendental extension of $\Bbbk$.
\[R:toric-remarks\] (a) The theorem above shows that our canonical construction of the minimal resolution ${\mathfrak}M$ uses only *intrinsic* properties of $Q$, in the sense that it is canonically invariant under the group of symmetries of $Q$.
\(b) It is natural to ask whether in positive characteristic it is always possible to have an intrinsic construction with $\Bbbk'=\Bbbk$. At this point we do not have an example to the contrary. However, a similar question arises in the closely related case of monomial ideals, where we show in Theorem \[T:counterexample\] that in positive characteristic one must have, in general, $\Bbbk'$ to be at least a transcendental extension of $\Bbbk$. This leads us to conjecture that in the characteristic $p>0$ toric case, in general, for an intrinsic construction of a minimal resolution one needs also $\Bbbk'$ to be at least a transcendental extension of $\Bbbk$. We will observe later in Remark \[R:monomial-remarks\](g) that it is in fact possible to avoid extending the base field in all but finitely many positive characteristics.
\(c) We produced our minimal resolution by starting with the canonical resolution ${\mathfrak}F$ from [@TV1] that uses the normalized bar resolution of the Betti category of $Q$, and then constructing ${\mathfrak}M$ as an explicit direct summand of ${\mathfrak}F$. It should be clear by now to the reader that one could start with any other canonical resolution of $\Bbbk[Q]$ and proceed in an analogous fashion. For example, one could start with the resolution from [@TV1] that comes from the normalized bar resolution of the lub-category of $Q$, or, one could start with the hull resolution of $\Bbbk[Q]$, see [@BS].
\(d) It should be noted that, in characteristic $p>0$, the field extension $\Bbbk'$ that our method needs for the intrinsic construction will depend on the starting canonical resolution ${\mathfrak}F$. Furthermore, in general the number of splittings of a nontrivial chain complex over $\mathbb F_p$ is quite large, which causes the transcendence degree of $\Bbbk'$ over $\Bbbk$ to be quite high. It is therefore an interesting problem, given $Q$, to produce a canonical starting resolution of $\Bbbk[Q]$ that would lead to a “smallest” possible extension $\Bbbk'$ in characteristic $p>0$.
\(e) It is possible, by taking into account orbits of the action of $G$ on the sets of splittings $Spl_Q(q)$, and some other adjustments, to produce versions of Theorem \[T:toric-canonical\] that in positive characteristic yield a field extension $\Bbbk'$ over $\Bbbk$ of a possibly smaller but in general still nonzero transcendence degree . We discuss some of these “other adjustments” in the example below, and incorporate their precise formulation into the statements of our results on monomial ideals in the next section.
\[Ex:toric-minimal\] We will use the following homologically very simple example to illustrate our construction. Let $Q$ be the numeric semigroup generated in $\mathbb N$ by $2$ and $3$. Note that the group of automorphisms of $Q$ is trivial. The corresponding toric ring $\Bbbk[Q]$ has as set of Betti degrees $B=\{0, 6\}$, and the corresponding polynomial ring $R=\Bbbk[x_2, x_3]$ has exactly two monomials of degree $6$, namely $x_2^3$ and $x_3^2$. Thus in the Betti category there are only two sequences of length $1$, namely $$0\xrightarrow{x_2^3} 6, \qquad\text{ and }\qquad
0\xrightarrow{x_3^2} 6.$$ Of course, we also have two trivial sequences of length $0$, namely $0$ and $6$. The resulting resolution ${\mathfrak}F$ has the form $$0{\longleftarrow}R^2
\xleftarrow{
\left[
\begin{smallmatrix}
-x_2^3 & -x_3^2 \\
1 & 1
\end{smallmatrix}
\right]
}
R^2 {\longleftarrow}0.$$ The simplicial chain complexes of the pairs of semi-simplicial sets $(\Delta_{\le 6},\Delta_{<6})$ and $(\Delta_{\le 0},\Delta_{<0})$ have the form (over the prime field $\mathbb F$) $$0{\longleftarrow}\mathbb F
\xleftarrow{
\left[
\begin{smallmatrix}
1 & 1
\end{smallmatrix}
\right]
}
\mathbb F^2 {\longleftarrow}0,
\qquad\text{ and }\qquad
0{\longleftarrow}\mathbb F {\longleftarrow}0,$$ respectively.
\(a) In characteristic zero, using formula gives $$0{\longrightarrow}\mathbb F
\xrightarrow{
\left[
\begin{smallmatrix}
1/2 \\
1/2
\end{smallmatrix}
\right]
}
\mathbb F^2 {\longrightarrow}0,$$ for the Moore-Penrose splitting of $C_\bullet(\Delta_{\le 6},\Delta_{<6}, \mathbb F)$. The resulting vector field $W$ on ${\mathfrak}F$ becomes $$0{\longrightarrow}R^2
\xrightarrow{
\left[
\begin{smallmatrix}
0 & 1/2 \\
0 & 1/2
\end{smallmatrix}
\right]
}
R^2 {\longrightarrow}0,$$ hence the flow $\Phi_W$ is the morphism of chain complexes $$\begin{CD}
0 @<{\hphantom{\left[
\begin{smallmatrix}
1 & (1/2)(x_2^3+x_3^2) \\
0 & 0
\end{smallmatrix}
\right]}}<<
R^2
@<<<
R^2
@<{\hphantom{
\left[
\begin{smallmatrix}
1 & (1/2)(x_2^3+x_3^2) \\
0 & 0
\end{smallmatrix}
\right]
}}
<< 0 \\
@.
@V{
\left[
\begin{smallmatrix}
1 & (1/2)(x_2^3+x_3^2) \\
0 & 0
\end{smallmatrix}
\right]}
VV
@VV{
\left[
\begin{smallmatrix}
1/2 & -1/2 \\
-1/2 & 1/2
\end{smallmatrix}
\right]}
V
@. \\
0
@<{\hphantom{
\left[
\begin{smallmatrix}
1 & (1/2)(x_2^3+x_3^2) \\
0 & 0
\end{smallmatrix}
\right]
}}<<
R^2
@<<<
R^2
@<{\hphantom{
\left[
\begin{smallmatrix}
1 & (1/2)(x_2^3+x_3^2) \\
0 & 0
\end{smallmatrix}
\right]
}}<< 0.
\end{CD}$$ In this case $\Phi_W$ is already a projection and we see that our canonical minimal resolution ${\mathfrak}M=\operatorname{Im}\Phi_W$ is the subcomplex of ${\mathfrak}F$ generated by the sequence $0$ in homological degree $0$, and by the element $(0\xrightarrow{x_2^3} 6) - (0\xrightarrow{x_3^2}6)$ in homological degree $1$. Thus ${\mathfrak}M$ has the form $$0 {\longleftarrow}R \xleftarrow{x_3^2 - x_2^3} R {\longleftarrow}0$$ with respect to this basis.
\(b) In characteristic $p=2$ the complex $C_\bullet(\Delta_{\le 6}, \Delta_{<6}, \mathbb F)$ has exactly $2$ splittings, given by $$0{\longrightarrow}\mathbb F
\xrightarrow[\alpha]{
\left[
\begin{smallmatrix}
1 \\
0
\end{smallmatrix}
\right]
}
\mathbb F^2 {\longrightarrow}0,
\qquad\text{ and }\qquad
0{\longrightarrow}\mathbb F
\xrightarrow[\beta]{
\left[
\begin{smallmatrix}
0 \\
1
\end{smallmatrix}
\right]
}
\mathbb F^2 {\longrightarrow}0.$$ Thus we get $Y_0=\{y_0\}$ and $Y_6=\{y_\alpha, y_\beta\}$. Instead of taking the purely transcendental extension over $\Bbbk$ on the variables $y_0,y_\alpha, y_\beta$, we can “adjust” by taking $\Bbbk'$ to be the fraction field of $\Bbbk[y_0,y_\alpha,y_\beta]/(y_\alpha+y_\beta-1, y_0-1)$, where $\Bbbk[y_0, y_\alpha, y_\beta]$ is the polynomial ring on the variables $y_0,y_\alpha, y_\beta$, and we denote by $y_0,y_\alpha$ and $y_\beta$ also their images in $\Bbbk'$. In particular, $\Bbbk'$ is a purely transcendental extension of $\Bbbk$ of degree $1$. Let $R'=\Bbbk'[X]$ and let ${\mathfrak}F'={\mathfrak}F\otimes_R R'$. Now since $y_\alpha + y_\beta = 1$, we get by Remark \[R:affine-combination\] that $\alpha\otimes y_\alpha + \beta\otimes y_\beta$ is a weak partial splitting on ${\overline}{{\mathfrak}F'(6)} = C_\bullet(\Delta_{\le 6},\Delta_{<6},\mathbb F)\otimes R'$ that satisfies , hence we obtain from Proposition \[P:weak-partial-splitting\](cd) that it is in fact already a splitting. Thus we obtain on ${\mathfrak}F'$ the vector field $W$ $$0{\longrightarrow}R'^2
\xrightarrow{
\left[
\begin{smallmatrix}
0 & y_\alpha \\
0 & y_\beta
\end{smallmatrix}
\right]
}
R'^2 {\longrightarrow}0,$$ hence the flow $\Phi_W$ is the morphism of chain complexes $$\begin{CD}
0 @<{\hphantom{\left[
\begin{smallmatrix}
1 & y_\alpha x_2^3+ y_\beta x_3^2 \\
0 & 0
\end{smallmatrix}
\right]}}<<
{R'}^2
@<<<
{R'}^2
@<{\hphantom{
\left[
\begin{smallmatrix}
1 & y_\alpha x_2^3+ y_\beta x_3^2) \\
0 & 0
\end{smallmatrix}
\right]
}}
<< 0 \\
@.
@V{
\left[
\begin{smallmatrix}
1 & y_\alpha x_2^3+ y_\beta x_3^2 \\
0 & 0
\end{smallmatrix}
\right]}
VV
@VV{
\left[
\begin{smallmatrix}
y_\beta & -y_\alpha \\
-y_\beta & y_\alpha
\end{smallmatrix}
\right]}
V
@. \\
0
@<{\hphantom{
\left[
\begin{smallmatrix}
1 & (1/2)(x_2^3+x_3^2) \\
0 & 0
\end{smallmatrix}
\right]
}}<<
{R'}^2
@<<<
{R'}^2
@<{\hphantom{
\left[
\begin{smallmatrix}
1 & (1/2)(x_2^3+x_3^2) \\
0 & 0
\end{smallmatrix}
\right]
}}<< 0
\end{CD}$$ Just as in part (a), in this case $\Phi_W$ is again already a projection, and we obtain the intrinsic minimal resolution ${\mathfrak}M$ as the subcomplex of ${\mathfrak}F'$ generated over $R'$ again by the trivial sequence $0$ in homological degree $0$, and by $(0\xrightarrow{x_2^3}6) - (0\xrightarrow{x_3^2}6)$ in homological degree $1$. Thus ${\mathfrak}M$ has the form $$0 {\longleftarrow}R' \xleftarrow{x_3^2 - x_2^3} R' {\longleftarrow}0$$ with respect to this basis.
Monomial resolutions {#S:monomial-resolutions}
====================
Let $X$ be a finite set, and let $
M(X) =
\{
f\mid f{\! :}X{\longrightarrow}\mathbb N \text{ is a function of sets}\}
$ be the free abelian monoid with basis the set $X$. Note that $M(X)$ is partially ordered by $f\le g \iff f(x)\le g(x)$ for every $x\in X$. A *monomial ideal* is an ideal $I$ in the monoid $M(X)$. Let $G_I$ be the group of all automorphisms of the monoid $M(X)$ that preserve the ideal $I$. The properties of $I$ invariant under the action of $G_I$ are precisely the *intrinsic* properties of $I$ as an ideal of the monoid $M(X)$.
Given a field $\Bbbk$ we identify the monoid ring $\Bbbk[M(X)]$ with the polynomial ring $\Bbbk[X]$ on the set of variables $X$ via the assignment $f \longmapsto \prod_{x\in X}x^{f(x)}$. Under this identification the monomial ideal $I\subseteq M(X)$ generates an ideal (also denoted by $I$ and called a monomial ideal) in the ring $R=\Bbbk[X]$. The polynomial ring $R$ has a natural $\mathbb Z$-grading, and the monomial ideal $I$ is a homogeneous ideal of $R$.
Since the work of Taylor [@T] it has been a main open problem in commutative algebra to give an explicit canonical description of the minimal $\mathbb Z$-graded free resolution of $I$ over $R$ by using only intrinsic properties of the ideal $I$. As any automorphism in $G_I$ is induced by a permutation on the set $X$, we can identify $G_I$ with the set of those permutations on $X$ that leave the set of monomials in $I$ invariant. Any such permutation $\sigma$ induces an automorphism of the $\mathbb Z$-graded $\Bbbk$-algebra $R$, denoted also by $\sigma$, that leaves the ideal $I$ of $R$ invariant. Therefore, if a canonical construction of a minimal free resolution ${\mathfrak}M$ of $I$ uses only intrinsic properties of $I$ it must be invariant under the action of $G_I$, or, more specifically, there has to be a canonical group homomorphism ${\mathfrak}M{\! :}G_I{\longrightarrow}Aut_\Bbbk({\mathfrak}M)$ from $G_I$ to the group of automorphisms of ${\mathfrak}M$ (as a complex of $\Bbbk$-vector spaces) such that for each $\sigma$ the morphism ${\mathfrak}M(\sigma){\! :}{\mathfrak}M {\longrightarrow}\sigma_*{\mathfrak}M$ is a morphism of chain complexes of $R$-modules that lifts the isomorphism of $R$-modules $\sigma{\! :}I{\longrightarrow}\sigma_*I$.
With this in mind, and considering the results we have already obtained in the previous section on the closely related case of toric rings, we investigate the following (substantially weaker than normally considered) version of the problem of constructing explicitly the minimal free resolution of a monomial ideal. (Recall that for a $\mathbb Z$-graded module $E=\bigoplus_{n\in\mathbb Z} E_n$ over a $\mathbb Z$-graded ring $S$ the *twisted* module $E(k)$ is the $\mathbb Z$-graded $S$-module with homogeneous components $E(k)_n=E_{k+n}$.)
[P:mfr]{} Let $I$ be a monomial ideal in $M(X)$ and let $R=\Bbbk[X]$ be the polynomial ring with the standard $\mathbb Z$-grading, where $\Bbbk$ is a field and $X$ is a finite set of variables. Let ${\mathfrak m}$ be the homogeneous maximal ideal of $R$ generated by the variables. Construct in a canonical explicit way a field extension $\Bbbk'$ of $\Bbbk$, an action of $G_I$ as field automorphisms of $\Bbbk'$ over $\Bbbk$, an acyclic chain complex $${\mathfrak}M \quad = \quad
0 {\longleftarrow}M_0 {\longleftarrow}\dots {\longleftarrow}M_{n-1} \xleftarrow{\mu_n} M_n {\longleftarrow}\dots$$ of finite free $R'=\Bbbk'[X]$-modules with $\mu_n(M_n)\subseteq {\mathfrak m}M_{n-1}$ for each $n\ge 1$, [**and**]{} a group homomorphism $${\mathfrak}M{\! :}G_I {\longrightarrow}Aut_\Bbbk({\mathfrak}M)$$ from $G_I$ to the group of $\Bbbk$-automorphisms of the chain complex of $\Bbbk$-vector spaces ${\mathfrak}M$ such that:
1. there exist $R'$-module isomorphisms $\epsilon{\! :}\operatorname{H}_0{\mathfrak}M {\longrightarrow}IR'$ and $$\theta_n{\! :}M_n{\longrightarrow}\bigoplus_{d\in\mathbb Z} R'(-d)^{\beta_{n,d}}$$ for all $n\ge 0$, such that for $n\ge 1$ each map $$\theta_{n-1}\mu_n\theta_n^{-1}{\! :}\bigoplus_{d\in\mathbb Z} R'(-d)^{\beta_{n,d}} {\longrightarrow}\bigoplus_{d\in\mathbb Z} R'(-d)^{\beta_{n-1,d}}$$ is a morphism of $\mathbb Z$-graded $R'$-modules, and, with respect to the corresponding induced $\mathbb Z$-grading on $\operatorname{H}_0{\mathfrak}M$, so is the map $\epsilon$.
2. there exists an $R'$-module isomorphism $\delta{\! :}\operatorname{H}_0{\mathfrak}M{\longrightarrow}IR'$ such that for each $\sigma\in G_I$ the automorphism $${\mathfrak}M(\sigma){\! :}{\mathfrak}M {\longrightarrow}\sigma_*{\mathfrak}M$$ is a morphism of chain complexes of $R'$-modules, and the induced isomorphism $\bar\sigma{\! :}\operatorname{H}_0{\mathfrak}M {\longrightarrow}\sigma_*\operatorname{H}_0{\mathfrak}M$ satisfies $\sigma\delta = \delta\bar\sigma$.
\[R:weaker-formulation\] Notice that we do [**not**]{} ask for
- a $M(X)$-graded structure on ${\mathfrak}M$, i.e. we do not require a monomial-, fine-, or multi-grading on ${\mathfrak}M$;
- an explicit or canonical construction of the isomorphisms $\epsilon$, and $\theta_n$, i.e. of the $\mathbb Z$-graded structure of each $M_n$; in particular, we do [**not**]{} ask for a canonical or explicit construction of a homogeneous basis of each $M_n$;
- an explicit or canonical construction of the isomorphism $\delta$, i.e. of an identification of the zeroth homology of ${\mathfrak}M$ with the ideal $IR'$;
- the automorphisms ${\mathfrak}M(\sigma)$ to preserve any $\mathbb Z$-graded structure of ${\mathfrak}M$.
In fact we will show later in Theorem \[T:counterexample\] that in characteristic $p>0$ our problem has, in general, no solution with $\Bbbk=\mathbb F_p$ and $\Bbbk'$ an algebraic extension, even in this very weak formulation.
In characteristic zero, a solution goes back at least to the work of Yuzvinsky [@Y]. It is given as a special case of the main result [@Y Theorem 4.1], by taking the splitting $W_a$ of each complex ${\overline}{{\mathfrak}T(a)}$ in Example \[Ex:Taylor-4\] to be $\delta_a^{+}$ as in Example \[Ex:Taylor-0\], the splitting given by taking Moore-Penrose pseudo-inverses. Yuzvinsky’s construction produces the minimal resolution ${\mathfrak}M$ as a subcomplex of the Taylor resolution, and as indicated in Examples \[Ex:Taylor-3\] and \[Ex:Taylor-4\], Yuzvinsky’s proof can be leveraged to also produce an explicit projection of the Taylor resolution onto ${\mathfrak}M$.
In their recent paper [@EMO], Eagon, Miller, and Ordog leverage their ingenious combinatorial description (the “Hedge formula”) of Moore-Penrose pseudo-inverses of differentials in cellular chain complexes, and Eagon’s construction [@E] of a Wall complex, applied to a certain double complex, to give another solution in characteristic zero that has a strong combinatorial component, and works also in all but finitely many positive characteristics.
In the rest of this section we use our dynamical systems theory and present a solution to Problem \[P:mfr\] that works in all characteristics. The construction keeps the base field unchanged in all except certain finitely many positive characteristics (which depend on the ideal), where it produces a transcendental base field extension before obtaining the resolution. Our motivation comes from the influential work [@GPW], where Gasharov, Peeva, and Welker show that the isomorphism class of the lcm-lattice of a monomial ideal determines the structure of the minimal resolution, in the sense that an isomophism between the lcm-lattices of two ideals determines a canonical “relabeling” procedure that transforms a minimal resolution of one ideal into a minimal resolution of the other. It is therefore natural to ask whether it is possible to construct an explicit canonical minimal resolution of a monomial ideal directly from its lcm-lattice in an intrinsic fashion. We will now describe how to accomplish this.
Let $U$ be the set of minimal generators of our monomial ideal $I$ in $M(X)$. Recall from Example \[Ex:Taylor-1\] that the lcm-lattice $L$ of $I$ is the join semi-lattice join-generated in $M(X)$ by the set $U$. Equivalently, representing the elements of $M(X)$ as monomials in the polynomial ring $\Bbbk[X]$ via the identification $f\mapsto m_f=x^f = \prod_{y\in X}y^{f(y)}$, the lcm-lattice $L$ is represented by all monomials that are least common multiples of subsets of the set of monomials $\{m_u\mid u\in U\}$. In this representation the partial order on $L$ is precisely the one given by divisibility of monomials.
Next, set $L'=L\setminus\{1\}$, obtained by removing from $L$ its minimal element $\hat 0 = 1$, and consider the order simplicial complex $\Delta=\Delta(L')$ of the poset $L'$. The $n$-faces of $\Delta$ are given by all chains $$\label{E:n-faces}
a_0< \dots < a_n$$ in $L'$, in particular they have a natural orientation induced by the ordering on their vertices. Thus we obtain canonical bases for the spaces of oriented $n$-chains $C_n(\Delta, \mathbb F)$ and $
C_n(a) := C_n(\Delta_{\le a}, \Delta_{<a}, \mathbb F)=
C_n(\Delta_{\le a},\mathbb F)/C_n(\Delta_{<a},\mathbb F)
$, where $\mathbb F$ is the prime field of $\Bbbk$, and the abstract simplicial complex $\Delta_{\le a}$ (respectively, $\Delta_{<a}$) consists of all faces of $\Delta$ with $a_n\le a$ (resectively, $a_n<a$). In particular, in charateristic zero we obtain a natural inner product on these spaces that makes these bases orthonormal. Choosing a subset $X_i$ of this canonical basis of $C_i(a)=C_i(\Delta_{\le a}, \Delta_{<a},\mathbb F)$ such that it maps under the differential $d_i^a$ of the simplicial chain complex $C_\bullet(\Delta_{\le a},\Delta_{<a},\mathbb F)$ of the pair $(\Delta_{\le a},\Delta_{<a})$ bijectively onto a basis of $\operatorname{Im}d_i^a$ induces a unique map $\tilde d_i^a{\! :}\operatorname{Im}d_i^a{\longrightarrow}C_i(a)$ satisfying $\tilde d_i^a d_i^a\big\vert_{X_i}=\operatorname{id}_{X_i}$. Furthermore, for each canonical basis element $b$ of $C_i(a)$ the unique expression $d_i^a(b)=\sum_{c\in X_i}r_cd_i^a(c)$ produces a canonical $i$-cycle $z(b)=z_{X_i}(b)=b-\sum_{c\in X_i}r_cc$ in $\operatorname{Ker}d_i^a$. It is a routine exercise in linear algebra to show that the set $K_i=\{z(b)\mid b\notin X_i\}$ forms a basis of $\operatorname{Ker}d_i^a$, therefore any subset $Z_i$ of $K_i$ that maps bijectively to a basis of $H_i(a)=\operatorname{H}_i(\Delta_{\le a},\Delta_{<a},\mathbb F)$ under the canonical projection $\pi_i{\! :}\operatorname{Ker}d_i^a {\longrightarrow}H_i(a)$ induces a canonical map $\tilde\pi_i{\! :}H_i(a){\longrightarrow}C_i(a)$ satisfying $\tilde\pi_i \pi_i\big\vert_{Z_i}=\operatorname{id}_{Z_i}$. Therefore choosing such a pair $(X_i,Z_i)$ for each $i\ge 0$ induces canonically a splitting of $C_\bullet(\Delta_{\le a},\Delta_{<a},\mathbb F)$, and we call such a splitting a *matroidal splitting*. Let $MSpl(a)$ be the set of all matroidal splittings of $C_\bullet(\Delta_{\le a}, \Delta_{<a}, \mathbb F)$. Clearly $$\label{E:matroidal-bound}
m(a):= |MSpl(a)| \le
\prod_{i\ge 0}\binom{\dim C_i(a)}{\operatorname{rank}d_i^a}
\binom{\dim C_i(a)-\operatorname{rank}d_i^a}{\dim H_i(a)}.$$ We call $a\in L'$ *critical* over $\Bbbk$ if $char(\Bbbk)$ divides $m(a):= |MSpl(a)|$. We write $L_{crit}'$ for the set of critical elements of $L'$ (over $\Bbbk$). In particular, if $char(\Bbbk)=0$ then $L'_{crit}=\emptyset$, i.e. there are no critical elements. For each $a\in L'_{crit}$ we let $$Y_a=\{y_\sigma \mid \sigma\in MSpl(a)\}$$ be a set of variables. Let $Y=\sqcup_{a\in L'_{crit}}Y_a$ be the disjoint union. Note that $Y$ is finite. We set $$\label{E:field-extension}
\Bbbk' =
\Bbbk(Y)^{\widetilde{}}$$ where $\Bbbk(Y)^{\widetilde{}}$ is the field of fractions of the domain $\Bbbk[Y]^{\widetilde{}}=\Bbbk[Y]/(l_a\mid a\in L'_{crit})$ with $l_a=1-\sum_{\sigma\in MSpl(a)}y_\sigma$ for each $a\in L'_{crit}$. Thus $\Bbbk'$ is a purely transcendental extension of $\Bbbk$, of transcendence degree $|Y|-|L'_{crit}|$, and $\Bbbk'=\Bbbk$ when $char(\Bbbk)=0$.
It is well-known that the simplicial complex $\Delta$ supports a $M(X)$-graded free resolution ${\mathfrak}F=(F_n,\phi_n)$ of $IR'$ over $R'=\Bbbk'[X]$, called the *lcm-resolution* of $IR'$, see [@OW Proposition 3.3.27], described as follows. In homological degree $n\ge 0$ the module $F_n$ is the free $M(X)$-graded $R'$-module with basis the set of all oriented $n$-faces of $\Delta$, each oriented face homogeneous of degree $a_n\in L'$. For $n\ge 1$ the differential $\phi_n{\! :}F_n{\longrightarrow}F_{n-1}$ is given by sending an oriented face $T$ as in to $$\sum_{i=0}^{n-1} (-1)^i T_i
\quad + \quad
(-1)^n (x^{a_n}/x^{a_{n-1}}) T_n ,$$ where for each $0\le i\le n$ the face $T_i$ is the oriented face obtained by omitting $a_i$ from $T$. We have a canonical $L'$-stratification on ${\mathfrak}F$ where for each $a\in L'$ the module $F_n^a$ is the free $R'$-submodule of $F_n$ generated by the oriented faces with $a_n=a$. It is immediate from this description that for the stratum at $a$ we have $${\overline}{{\mathfrak}F(a)}=
C_\bullet(\Delta_{\le a},\Delta_{<a},\mathbb F)\otimes R'.$$
Now all that remains is to specify a splitting $W_a$ on each ${\overline}{{\mathfrak}F(a)}$ and apply Theorem \[T:asymptotic\]. In characteristic zero we can take $W_a=\delta_a^+\otimes 1$, where $\delta_a^+$ is the splitting as in Remark \[R:pseudo-inverses\] given by taking the Moore-Penrose inverses of the differentials of $C_\bullet(\Delta_{\le a},\Delta_{<a},\mathbb F)$ with respect to the natural inner products. Alternatively, in *any* charateristic, we can take $W_a= \widehat{V_a}$, where $$V_a=
\begin{cases}
\sum_{\sigma\in MSpl(a)}\sigma\otimes y_\sigma &\text{ if } a\in L'_{crit}; \\
\frac{1}{m(a)}
\sum_{\sigma\in MSpl_(a)} \sigma\otimes 1 &\text{ otherwise; }
\end{cases}$$ is, by Remark \[R:affine-combination\], a weak partial splitting on $C_\bullet(\Delta_{\le a}, \Delta_{<a},\mathbb F)\otimes R'$ satisfying , and $\widehat{V_a}$ is the induced splitting from Proposition \[P:weak-partial-splitting\].
With this we are all set to use our dynamical systems machinery. By Theorem \[T:asymptotic\] the flow $\Phi_W$ of the vector field $W=\sum_{a\in L'}W_a$ stabilizes after $1 + \dim L'$ iterations, hence $\Phi_W^{1+\dim L'}$ is a projection of ${\mathfrak}F$ onto a projective hence free resolution ${\mathfrak}M$ of $IR'$ over $R'$. Since each $W_a$ preserves $M(X)$-degrees, so do $\Phi_W$ and its iterates, therefore ${\mathfrak}M$ is $M(X)$-graded, and therefore minimal by Corollary \[C:splittings-are-minimal\].
Finally, since under the induced action of $G_I$ an element $\gamma\in G_I$ sends the canonical basis of $C_i(a)$ to the canonical basis of $C_i\bigl(\gamma(a)\bigr)$, we have $G_I$ sending matroidal splittings to matroidal splittings. Therefore constructing the action of $G_I$ on $\Bbbk'$ and the homomorphism of groups ${\mathfrak}M{\! :}G_I {\longrightarrow}Aut_\Bbbk({\mathfrak}M)$ is *mutatis mutandis* that of the corresponding construction for the group $G$ in the proof of Theorem \[T:toric-intrinsic\].
We have thus completed a proof of
\[T:monomial-intrinsic\] Let $I$ be a monomial ideal in the free abelian monoid $M(X)$ on a finite set $X$, with group of automorphisms $G_I$, and let $\Bbbk$ be a field.
Then there exist
- a canonical explicitly constructed finitely generated field extension $\Bbbk'$ of $\Bbbk$,
- a canonical explicitly constructed $M(X)$-homogeneous projection of the lcm-resolution ${\mathfrak}F$ of $I\Bbbk'[X]$ over $\Bbbk'[X]$ onto a direct summand ${\mathfrak}M$ which is a minimal free $M(X)$-graded resolution of $I\Bbbk'[X]$ over $\Bbbk'[X]$,
- a canonical explicitly constructed action of $G_I$ as automorphisms of $\Bbbk'$ over $\Bbbk$, and
- a canonical explicitly constructed homomorphism ${\mathfrak}M{\! :}G_I{\longrightarrow}Aut_\Bbbk({\mathfrak}M)$ from $G_I$ to the group of automorphisms of the chain complex of $\Bbbk$-vector spaces ${\mathfrak}M$;
such that for each $\gamma\in G_I$ the map ${\mathfrak}M(\gamma){\! :}{\mathfrak}M {\longrightarrow}\gamma_*{\mathfrak}M$ is an isomorphism of complexes of free $M(X)$-graded $\Bbbk'[X]$-modules that lifts the isomorphism of $M(X)$-graded $\Bbbk'[X]$-modules $\gamma{\! :}I\Bbbk'[X]{\longrightarrow}\gamma_*(I\Bbbk'[X])$. In characteristic zero one can take $\Bbbk'=\Bbbk$, and in characteristic $p>0$ one can take $\Bbbk'$ to be a purely transcendental extension of $\Bbbk$ of degree $\sum_{a\in L'_{crit}} (m(a)-1)$.
\[C:infinite-primes\] Given any monomial ideal $I$ in $M(X)$, Problem \[P:mfr\] has a solution with $\Bbbk'=\Bbbk=\mathbb F_p$ for every sufficiently big prime $p$.
By we get for each $a\in L'$ $$m(a)\le \prod_{i\ge 0}2^{\dim C_i(a)}2^{\dim C_i(a)}$$ hence for $p\ge \max\{\prod_{i\ge 0}4^{\dim C_i(a)}\mid a\in L'\}$ we get the desired conclusion.
\[R:monomial-remarks\] (a) By now it should be clear to the reader that instead of using the lcm-resolution or the Taylor resolution as a starting point, one could apply our technique starting from any other canonical construction of a based free resolution of a monomial ideal $I$. For example, one could opt for the (in general smaller than the lcm-resolution) resolution supported on the order complex of the Betti poset of $I$, see e.g. [@TV].
\(b) Our construction provides an intrinsic canonical set of $M(X)$-homogeneous generators for the minimal resolution ${\mathfrak}M$, but not a basis. In general it is not possible to produce a canonical and intrinsic (not even up to associates) basis of ${\mathfrak}M$, not even in characteristic zero. This was already pointed out in [@EMO Example 3.8], and that observation continues to hold also under the much weaker requirements of Probem \[P:mfr\]. One can provide a rigorous proof by using an appropriately modified version of the proof of Theorem \[T:counterexample\]. We leave these details to the interested reader.
\(c) Our construction has additional functorial properties not mentioned in Theorem \[T:monomial-intrinsic\]. Given a monomial ideal $J\subset M(X)$ and an element $m\in M(X)$, write $J_{\le m}$ for the monomial ideal generated by the those minimal generators of $J$ that are $\le m$. Denote by $\mathcal I(X)$ the category with objects the monomial ideals in $M(X)$, and with morphisms from $I$ to $J$ those automorphisms $\gamma$ of $M(X)$ such that $\gamma(I)=J_{\le m}$ for some (depending on $\gamma$) monomial $m$. We leave to the reader the routine verification that our construction produces assignments $I \mapsto {\mathfrak}M={\mathfrak}M(I)$ and $\gamma\mapsto {\mathfrak}M(\gamma)$ that make it into a functor ${\mathfrak}M{\! :}\mathcal I(X) {\longrightarrow}\operatorname{{\bf Comp}}$, where $\operatorname{{\bf Comp}}$ is the category of chain complexes of finite free modules as described in [@Tc Section 1].
\(d) Just as in the toric rings case, the transcendence degree of the base field extension that our method requires in positive characteristic depends on the starting non-minimal canonical resolution, see Example \[Ex:mfr-example\](a). For instance, for ideals of low projective dimension but with a sufficiently big number of minimal generators, starting with the lcm-resolution will in general produce a lower transcendence degree than starting with the Taylor resolution. It is an interesting problem whether there exists a canonical non-minimal starting construction that would produce for every ideal a transcendence degree that is lowest among the degrees produced by all other canonical non-minimal starting constructions.
\(e) Given a monomial ideal $I$ and a prime $p\ge 2$, a natural question to ask is what is the smallest transcendence degree required of an extension $\Bbbk'$ of $\Bbbk=\mathbb F_p$ so that Problem \[P:mfr\] has a solution for $I$. Another related question is whether there is in some sense a “minimal” such transcendental extension, perhaps unique up to an isomorphism. At this point it seems that a functorial construction such as ours would not be able to produce that kind of “minimal” extension for every monomial ideal. It is very likely that one would need to sacrifice at least some of the functoriality described in part (c) above, in order to capture the specifics of a given ideal that would lead to such a “minimal” extension.
\(f) Given a monomial ideal $I$, another interesting problem is to describe the set of all primes $p\ge 2$ for which Problem \[P:mfr\] has no solution for $I$ with $\Bbbk'=\Bbbk=\mathbb F_p$.
\(g) By taking averages as in the proof of Theorem \[T:monomial-intrinsic\] of matroidal splittings only (instead of all splittings), one obtains a proof of Theorem \[T:toric-intrinsic\] that keeps the base field unchanged in all but finitely many positive characteristics.
\(h) A matroidal splitting is the same as a splitting induced by what is called a *community* in [@EMO].
An example {#S:an-example}
==========
To illustrate our construction, in this section we compute the following very useful
\[Ex:mfr-example\] Let $\Bbbk$ be a field and let $\mathbb F$ be the prime subfield of $\Bbbk$. Let $X=\{v_0,v_1,v_2,v_3,e_{12},e_{23},e_{31}\}$ and let $I$ be the monomial ideal generated by the set of monomials $$m_0 = v_1v_2v_3e_{12}e_{23}e_{31} ,\quad
m_1 = v_0^2v_2v_3e_{23} ,\quad
m_2 = v_0^2v_1v_3e_{31} ,\quad
m_3 = v_0^2v_1v_2e_{12}.$$ Note this is precisely the ideal $I(3)$ from Definition \[D:the-counterexample\]. The lcm-lattice of $I$ is the set of monomials $\{1, m_0, m_1, m_2, m_3, m_{12}, m_{31}, m_{23}, m\}$, where $$\begin{aligned}
m &= v_0^2v_1v_2v_3e_{12}e_{23}e_{31}, \\
m_{12} &= v_0^2v_1v_2v_3e_{23}e_{31}, \\ m_{23} &= v_0^2v_1v_2v_3e_{31}e_{12}, \\ m_{31} &= v_0^2v_1v_2v_3e_{12}e_{23}. $$ In order to have uniform notation, we also set $e_{ij}=e_{ji}$ and $m_{ij}=m_{ji}$. The simplicial complex $\Delta=\Delta(L')$ has $8$ vertices, $13$ edges, and six $2$-faces. In the canonical bases given by the oriented faces, the complex $C_\bullet(\Delta_{\le m},\Delta_{<m},\mathbb F)$ has the form $$0 {\longleftarrow}\mathbb F
\xleftarrow{
[
\begin{smallmatrix}
1 & 1 & 1 & 1 & 1 & 1 & 1
\end{smallmatrix}
]
}
\mathbb F^{7}
\xleftarrow{
\left[
\begin{smallmatrix}
{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}0 \\
-1 & -1 &{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}0 \\
{\hphantom{-}}0 &{\hphantom{-}}0 & -1 & -1 &{\hphantom{-}}0 &{\hphantom{-}}0 \\
{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}0 & -1 & -1 \\
{\hphantom{-}}1 &{\hphantom{-}}0 &{\hphantom{-}}1 &{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}0 \\
{\hphantom{-}}0 &{\hphantom{-}}1 &{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}1 &{\hphantom{-}}0 \\
{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}0 &{\hphantom{-}}1 &{\hphantom{-}}0 &{\hphantom{-}}1
\end{smallmatrix}
\right]
}
\mathbb F^{6}
{\longleftarrow}0,$$ each of the complexes $C_\bullet(\Delta_{\le m_{ij}},\Delta_{<m_{ij}},\mathbb F)$ for $1\le i<j\le 3$ has the form $$0{\longleftarrow}\mathbb F
\xleftarrow{
[
\begin{smallmatrix}
1 & 1
\end{smallmatrix}
]
}
\mathbb F^2
{\longleftarrow}0,$$ and each of the complexes $C_\bullet(\Delta_{\le m_i},\Delta_{<m_i},\mathbb F)$ for $i=0,\dots,3$ has the form $$0{\longleftarrow}\mathbb F{\longleftarrow}0.$$
\(a) A standard computation shows that in every characteristic we have $72$ matroidal splittings of the complex $C_\bullet(\Delta_{\le m},\Delta_{<m},\mathbb F)$, we have $2$ matroidal splittings for each of the three complexes $C_\bullet(\Delta_{\le m_{ij}}, \Delta_{<m_{ij}},\mathbb F)$, and only one matroidal splitting for each of the four complexes $C_\bullet(\Delta_{\le m_i}, \Delta_{<m_i}, \mathbb F)$. Thus using our methods with the lcm-resolution as a starting point we produce an intrinsic minimal resolution over $\mathbb F_p$ whenever $p\ge 5$. When $p=2$ we produce an intrinsic minimal resolution over a purely transcendental extension of $\mathbb F_2$ of transcendence degree $|Y| - |L'_{crit}| = (72-1) + 3(2-1)=74$, and when $p=3$ we get a minimal resolution over a purely transcendental extension of $\mathbb F_3$ of transcendence degree $72-1=71$. We leave it to the reader ro check that the corresponding computation with the Taylor resolution as starting point produces an intrinsic construction in both $p=2$ and $p=3$ cases after taking a transcendental extension with transcendence degree “only” $17$. On the other hand, when $p\ne 3$ there is an intrinsic construction for the minimal resolution of $I$ without extending the field at all, see Remark \[R:intrinsic-resolution\](a). However, by Theorem \[T:counterexample\], in characteristic $p=3$ it is not possible to avoid taking some kind of transcendental extension. Remark \[R:intrinsic-resolution\](b) gives such an intrinsic construction using an extension of $\mathbb F_3$ of transcendence degree $2$. At this point we do not know if it is possible to have an intrinsic construction where the transcendence degree of the required field extension of $\mathbb F_3$ is exactly $1$.
\(b) In characteristic zero taking Moore-Penrose inverses and using the explicit formula yields the following splittings: $$0{\longrightarrow}\mathbb F
\xrightarrow{
\left[
\begin{smallmatrix}
1/7 \\
1/7 \\
1/7 \\
1/7 \\
1/7 \\
1/7 \\
1/7
\end{smallmatrix}
\right]
}
\mathbb F^7
\xrightarrow{
1/12
\left[
\begin{smallmatrix}
0& {-5}&{\hphantom{-}}3& {-1}&{\hphantom{-}}5& {-3}&{\hphantom{-}}1\\
0& {-5}& {-1}&{\hphantom{-}}3& {-3}&{\hphantom{-}}5&{\hphantom{-}}1\\
0&{\hphantom{-}}3& {-5}& {-1}&{\hphantom{-}}5&{\hphantom{-}}1& {-3}\\
0& {-1}& {-5}&{\hphantom{-}}3& {-3}&{\hphantom{-}}1&{\hphantom{-}}5\\
0&{\hphantom{-}}3& {-1}& {-5}&{\hphantom{-}}1&{\hphantom{-}}5& {-3}\\
0& {-1}&{\hphantom{-}}3& {-5}&{\hphantom{-}}1& {-3}&{\hphantom{-}}5
\end{smallmatrix}
\right]
}
\mathbb F^6 {\longrightarrow}0$$ for $C_\bullet(\Delta_{\le m}, \Delta_{< m}, \mathbb F)$, see Example \[Ex:pseudo-inverses\], and $$0 {\longrightarrow}\mathbb F
\xrightarrow{
\left[
\begin{smallmatrix}
1/2 \\
1/2
\end{smallmatrix}
\right]
}
\mathbb F^2 {\longrightarrow}0$$ for each of the complexes $C_\bullet(\Delta_{\le m_{ij}}, \Delta_{< m_{ij}}, \mathbb F)$ with $1\le i<j\le 3$. The lcm-resolution ${\mathfrak}F$ has the form [$$\begin{aligned}
0 \leftarrow
R^8
\xleftarrow[\phi_1]{
-\left[
\begin{smallmatrix}
\,\, v_0^2& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0& v_1e_{31}& v_1e_{12}& v_1e_{12}e_{31}
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& v_2e_{23}& v_2e_{12}& v_2e_{12}e_{23}
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& v_3e_{23}& v_3e_{31}& v_3e_{31}e_{23}
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0& -1& {\hphantom{-}}0& {\hphantom{-}}0
& -1& {\hphantom{-}}0& {\hphantom{-}}0
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& e_{12}& {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0& {\hphantom{-}}0& -1& {\hphantom{-}}0
& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& -1& {\hphantom{-}}0& {\hphantom{-}}0
& {\hphantom{-}}0& e_{31}& {\hphantom{-}}0 \\ {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0
& {\hphantom{-}}0& -1& {\hphantom{-}}0
& {\hphantom{-}}0& -1& {\hphantom{-}}0
& {\hphantom{-}}0& {\hphantom{-}}0& e_{23} \\ -1& {\hphantom{-}}0& {\hphantom{-}}0& -1
& {\hphantom{-}}0& {\hphantom{-}}0& -1
& {\hphantom{-}}0& {\hphantom{-}}0& -1
& -1& -1& -1 \end{smallmatrix}
\right]
}
&R^{13}
\\[+5pt]
\xleftarrow[\phi_2]{
\left[
\begin{smallmatrix}
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0 \\ e_{12} & {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0 & e_{31}& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0 \\ -1 & -1& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0 & {\hphantom{-}}0& e_{12}&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0&
e_{23} & {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0 & {\hphantom{-}}0& -1&
-1 & {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & e_{31}& {\hphantom{-}}0 \\ {\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& e_{23} \\ {\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & -1& -1 \\ {\hphantom{-}}1 & {\hphantom{-}}0& {\hphantom{-}}1&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0 \\ {\hphantom{-}}0 & {\hphantom{-}}1& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}1& {\hphantom{-}}0 \\ {\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}1 & {\hphantom{-}}0& {\hphantom{-}}1 \end{smallmatrix}
\right]
}
&R^6 \leftarrow 0 \end{aligned}$$ ]{} and the resulting vector field $W=\phi^+$ on ${\mathfrak}F$ is given by $$\begin{aligned}
0 {\longrightarrow}R^8
&\xrightarrow[\phi_1^+]{
\left[
\begin{smallmatrix}
0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0& 1/7 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
1/2 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & 1/2& {\hphantom{-}}0& {\hphantom{-}}0 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0& 1/7 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
1/2 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& 1/2& {\hphantom{-}}0 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0& 1/7 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & 1/2& {\hphantom{-}}0& {\hphantom{-}}0 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& 1/2& {\hphantom{-}}0 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0& 1/7 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0& 1/7 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0& 1/7 \\ 0 & {\hphantom{-}}0& {\hphantom{-}}0& {\hphantom{-}}0&
{\hphantom{-}}0 & {\hphantom{-}}0& {\hphantom{-}}0& 1/7 \end{smallmatrix}
\right]
}
\\[+5pt]
R^{13} &\xrightarrow[\phi_2^+]{
1/12
\left[
\begin{smallmatrix}
0&{\hphantom{-}}0&{\hphantom{-}}0& {-5}&{\hphantom{-}}0&{\hphantom{-}}0&{\hphantom{-}}3
&{\hphantom{-}}0&{\hphantom{-}}0& {-1}&{\hphantom{-}}5& {-3}&{\hphantom{-}}1\\
0&{\hphantom{-}}0&{\hphantom{-}}0& {-5}&{\hphantom{-}}0&{\hphantom{-}}0& {-1}
&{\hphantom{-}}0&{\hphantom{-}}0&{\hphantom{-}}3& {-3}&{\hphantom{-}}5&{\hphantom{-}}1\\
0&{\hphantom{-}}0&{\hphantom{-}}0&{\hphantom{-}}3&{\hphantom{-}}0&{\hphantom{-}}0& {-5}
&{\hphantom{-}}0&{\hphantom{-}}0& {-1}&{\hphantom{-}}5&{\hphantom{-}}1& {-3}\\
0&{\hphantom{-}}0&{\hphantom{-}}0& {-1}&{\hphantom{-}}0&{\hphantom{-}}0& {-5}
&{\hphantom{-}}0&{\hphantom{-}}0&{\hphantom{-}}3& {-3}&{\hphantom{-}}1&{\hphantom{-}}5\\
0&{\hphantom{-}}0&{\hphantom{-}}0&{\hphantom{-}}3&{\hphantom{-}}0&{\hphantom{-}}0& {-1}
&{\hphantom{-}}0&{\hphantom{-}}0& {-5}&{\hphantom{-}}1&{\hphantom{-}}5& {-3}\\
0&{\hphantom{-}}0&{\hphantom{-}}0& {-1}&{\hphantom{-}}0&{\hphantom{-}}0&{\hphantom{-}}3
&{\hphantom{-}}0&{\hphantom{-}}0& {-5}&{\hphantom{-}}1& {-3}&{\hphantom{-}}5
\end{smallmatrix}
\right]
}
R^6 \longrightarrow 0. \end{aligned}$$ Therefore the flow $\Phi=\Phi_W=\Phi_{\phi^{+}}=\operatorname{id}- \phi\phi^+ - \phi^+\phi$ is given by $$\begin{CD}
0 @<<< R^8 @< \phi_1 << R^{13} @< \phi_2 << R^6 @<<< 0 \\
@. @V \Phi_0 VV @V \Phi_1 VV @V \Phi_2 VV @. \\
0 @<<< R^8 @< \phi_1 << R^{13} @< \phi_2 << R^6 @<<< 0,
\end{CD}$$ where $$\Phi_0=
\left[
\begin{smallmatrix}
{\hphantom{-}}1{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&\frac{v_0^2}{7}
\\[+5pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}1{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&\frac{v_1e_{13}}{2}&
\frac{v_1e_{12}}{2}&{\hphantom{-}}0{\hphantom{-}}&\frac{v_1e_{12}e_{31}}{7}
\\[+5pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}1{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&\frac{v_2e_{23}}{2}&{\hphantom{-}}0{\hphantom{-}}&
\frac{v_2e_{12}}{2}&\frac{v_2e_{12}e_{23}}{7}
\\[+5pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}1{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&\frac{v_3e_{13}}{2}&
\frac{v_3e_{13}}{2}&\frac{v_3e_{13}e_{23}}{7}
\\[+5pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&\frac{e_{12}}{7}
\\[+5pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&\frac{e_{13}}{7}
\\[+5pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&\frac{e_{23}}{7}
\\[+5pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}\end{smallmatrix}
\right],$$
$$\Phi_1 =
{
\left[
\begin{smallmatrix}
\frac{6}{7}&0&0&{\frac{-1}{7}}&0&0&{\frac{-1}{7}}&0&0&
{\frac{-1}{7}}&{\frac{-1}{7}}&{\frac{-1}{7}}&{\frac{-1}{7}}
\\[+5pt]
0&\frac{1}{2}&0&\frac{5e_{12}}{12}&{\frac{-1}{2}}&0&
{\frac{-e_{12}}{4}}&0&0&\frac{e_{12}}{12}&\frac{e_{12}}{12}&
\frac{e_{12}}{4}&{\frac{-e_{12}}{12}}
\\[+5pt]
0&0&\frac{1}{2}&\frac{5e_{31}}{12}&0&0&\frac{e_{31}}{12}&
{\frac{-1}{2}}&0&{\frac{-e_{31}}{4}}&\frac{e_{31}}{4}&
\frac{e_{31}}{12}&{\frac{-e_{31}}{12}}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\\[+5pt]
0&{\frac{-1}{2}}&0&{\frac{-e_{12}}{4}}&\frac{1}{2}&0&
\frac{5e_{12}}{12}&0&0&\frac{e_{12}}{12}&\frac{e_{12}}{12}&
{\frac{-e_{12}}{12}}&\frac{e_{12}}{4}
\\[+5pt]
0&0&0&\frac{e_{23}}{12}&0&\frac{1}{2}&\frac{5e_{23}}{12}&0&
{\frac{-1}{2}}&{\frac{-e_{23}}{4}}&\frac{e_{23}}{4}&
{\frac{-e_{23}}{12}}&\frac{e_{23}}{12}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\\[+5pt]
0&0&{\frac{-1}{2}}&{\frac{-e_{31}}{4}}&0&0&\frac{e_{31}}{12}&
\frac{1}{2}&0&\frac{5e_{31}}{12}&{\frac{-e_{31}}{12}}&
\frac{e_{31}}{12}&\frac{e_{31}}{4}
\\[+5pt]
0&0&0&\frac{e_{23}}{12}&0&{\frac{-1}{2}}&{\frac{-e_{23}}{4}}&0&
\frac{1}{2}&\frac{5}{12}\,{x}_{1}&{-\frac{1}{12}\,{x}_{1}}&
\frac{e_{23}}{4}&\frac{e_{23}}{12}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\end{smallmatrix}
\right],
}$$ and $$\Phi_2 = (1/6)
\left[
\begin{smallmatrix}
{\hphantom{-}}1& {-1}& {-1}&{\hphantom{-}}1&{\hphantom{-}}1& {-1}\\
{-1}&{\hphantom{-}}1&{\hphantom{-}}1& {-1}& {-1}&{\hphantom{-}}1\\
{-1}&{\hphantom{-}}1&{\hphantom{-}}1& {-1}& {-1}&{\hphantom{-}}1\\
{\hphantom{-}}1& {-1}& {-1}&{\hphantom{-}}1&{\hphantom{-}}1& {-1}\\
{\hphantom{-}}1& {-1}& {-1}&{\hphantom{-}}1&{\hphantom{-}}1& {-1}\\
{-1}&{\hphantom{-}}1&{\hphantom{-}}1& {-1}& {-1}&{\hphantom{-}}1
\end{smallmatrix}
\right].$$ The posets $L'_n$ indexing the nonzero components of the $L'$-stratification of ${\mathfrak}F$ in homological degree $n$ are $$\begin{aligned}
L'_0 &=L', \\
L'_1 &=\{m_{12}, m_{31}, m_{23}, m\}, \text{ and }\\
L'_2 &=\{m\}.\end{aligned}$$ Thus, according to Theorem \[T:asymptotic\] we must have that $\Phi_0^3$, $\Phi_1^2$, and $\Phi_2$ are projections. In fact even $\Phi_0^2$ is already a projection. Computing these matrices gives $$\Phi_0^2 =
\left[
\begin{smallmatrix}
{\hphantom{-}}1{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&
{\hphantom{-}}0{\hphantom{-}}&\frac{v_0^2}{7}
\\[+5pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}1{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&\frac{v_1e_{31}}{2}&
\frac{v_1e_{12}}{2}&{\hphantom{-}}0{\hphantom{-}}&\frac{2v_1e_{12}e_{31}}{7}
\\[+5pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}1{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&\frac{v_2e_{23}}{2}&
{\hphantom{-}}0{\hphantom{-}}&\frac{v_2e_{12}}{2}&\frac{2v_2e_{12}e_{23}}{7}
\\[+5pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}1{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&\frac{v_3e_{23}}{2}&
\frac{v_3e_{31}}{2}&\frac{2v_3e_{23}e_{31}}{7}
\\[+8pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}\\[+10pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}\\[+10pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}\\[+10pt]
{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}&{\hphantom{-}}0{\hphantom{-}}\end{smallmatrix}
\right],$$ and [ $$\Phi_1^2 =
\left[
{
\begin{smallmatrix}
\frac{6}{7}&0&0&{\frac{-1}{7}}&0&0&{\frac{-1}{7}}
&0&0&{\frac{-1}{7}}&{\frac{-1}{7}}&{\frac{-1}{7}}&{\frac{-1}{7}}
\\[+5pt]
{\frac{-e_{12}}{14}}&\frac{1}{2}&0&\frac{29e_{12}}{84}&
{\frac{-1}{2}}&0&{\frac{-9e_{12}}{28}}&0&0&\frac{e_{12}}{84}&
\frac{e_{12}}{84}&\frac{5e_{12}}{28}&{\frac{-13e_{12}}{84}}
\\[+5pt]
{\frac{-e_{13}}{14}}&0&\frac{1}{2}&\frac{29e_{13}}{84}&0&0&
\frac{e_{13}}{84}&{\frac{-1}{2}}&0&{\frac{-9e_{13}}{28}}&
\frac{5e_{13}}{28}&\frac{e_{13}}{84}&{\frac{-13e_{13}}{84}}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\\[+5pt]
{\frac{-e_{12}}{14}}&{\frac{-1}{2}}&0&{\frac{-9e_{12}}{28}}&
\frac{1}{2}&0&\frac{29e_{12}}{84}&0&0&\frac{e_{12}}{84}&
\frac{e_{12}}{84}&{\frac{-13e_{12}}{84}}&\frac{5e_{12}}{28}
\\[+5pt]
{\frac{-e_{23}}{14}}&0&0&\frac{e_{23}}{84}&0&\frac{1}{2}&
\frac{29e_{23}}{84}&0&{\frac{-1}{2}}&{\frac{-9e_{23}}{28}}&
\frac{5e_{23}}{28}&{\frac{-13e_{23}}{84}}&\frac{e_{23}}{84}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\\[+5pt]
{\frac{-e_{13}}{14}}&0&{\frac{-1}{2}}&{\frac{-9e_{13}}{28}}&0&0&
\frac{e_{13}}{84}&\frac{1}{2}&0&\frac{29e_{13}}{84}&
{\frac{-13e_{13}}{84}}&\frac{e_{13}}{84}&\frac{5e_{13}}{28}
\\[+5pt]
{\frac{-e_{23}}{14}}&0&0&\frac{e_{23}}{84}&0&{\frac{-1}{2}}&
{\frac{-9e_{23}}{28}}&0&\frac{1}{2}&\frac{29e_{23}}{84}&
{\frac{-13e_{23}}{84}}&\frac{5e_{23}}{28}&\frac{e_{23}}{84}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\\[+5pt]
{\frac{-1}{7}}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&0&0&\frac{1}{42}&
\frac{1}{42}&\frac{1}{42}&\frac{1}{42}
\end{smallmatrix}
}
\right].$$ ]{} Thus the minimal resolution ${\mathfrak}M$, being the image of $\Phi^3=\Phi^2$, is provided by our method with a canonical (but non-minimal) intrinsic $M(X)$-homogeneous generating set. While in general one cannot hope for much better, for this particular ideal it is straightforward to see from the form of the matrices above that ${\mathfrak}M$ has as $M(X)$-homogeneous bases $$\bigl\{\quad h_0=\{m_0\},\ h_1=\{m_1\},\ h_2=\{m_2\},\ h_3 =\{m_3\}\quad \bigr\}$$ in homological degree $0$, $$\begin{aligned}
\Bigl\{
\quad
g_{12} = \frac{1}{2}\bigl(\{m_1,m_{12}\} &- \{m_2,m_{12}\}\bigr), \\[+3pt]
g_{31} = \frac{1}{2}\bigl(\{m_3,m_{31}\} &- \{m_1,m_{31}\}\bigr), \\[+3pt]
g_{23} = \frac{1}{2}\bigl(\{m_2,m_{23}\} &- \{m_3,m_{23}\}\bigr), \\[+3pt]
g_0 = \frac{1}{14}\bigl(12\{m_0,m\} &- e_{12}\{m_1,m_{12}\} - e_{12}\{m_2,m_{12}\} \\[+9pt]
&- e_{31}\{m_3,m_{31}\} - e_{31}\{m_1,m_{31}\} \\[+12pt]
&- e_{23}\{m_2,m_{23}\} - e_{23}\{m_3,m_{23}\} \\[+12pt]
&- 2\{m_1,m\} - 2\{m_2,m\} - 2\{m_3,m\} \\[+8pt]
&- 2\{m_{12},m\} - 2\{m_{31},m\} - 2\{m_{23},m\}\bigr)
\quad
\Bigr\}\end{aligned}$$ in homological degree $1$, and $$\begin{gathered}
\Bigl\{\quad
f = \frac{1}{6}\bigl(
\{m_1,m_{12},m\} - \{m_1,m_{31},m\} - \{m_2,m_{12},m\}
+\{m_2,m_{23},m\} \\
+\{m_3,m_{31},m\} - \{m_3,m_{23},m\}\bigr)
\quad\Bigr\}\end{gathered}$$ in homological degree $2$. With respect to these bases ${\mathfrak}M$ takes the form $$0{\longleftarrow}R^4
\xleftarrow{
\left[
\begin{smallmatrix}
\frac{-6}{7}v_0^2 & {\hphantom{-}}0&
{\hphantom{-}}0& {\hphantom{-}}0 \\ \frac{2}{7}v_1e_{12}e_{31}&\frac{-1}{2}v_1e_{31}&
\frac{1}{2}v_1e_{12} & {\hphantom{-}}0 \\ \frac{2}{7}v_2e_{12}e_{23}&\frac{1}{2}v_2e_{23} &
{\hphantom{-}}0&\frac{-1}{2}v_2e_{12} \\ \frac{2}{7}v_3e_{31}e_{23}& {\hphantom{-}}0&
\frac{-1}{2}v_3e_{23} &\frac{1}{2}v_3e_{31} \end{smallmatrix}
\right]
}
R^4 \xleftarrow{
\frac{1}{3}
\left[
\begin{smallmatrix}
0 \\ e_{12} \\ e_{31} \\ e_{23} \end{smallmatrix}
\right]
}
R {\longleftarrow}0,$$ and thus, up to associates, this is exactly the construction from Remark \[R:intrinsic-resolution\](a). Note that the automorphism group of our ideal is $S_3$, where $\tau\in S_3$ acts on $X$ by fixing $v_0$, and sending $v_i$ and $e_{ij}$ to $v_{\tau(i)}$ and $e_{\tau(i)\tau(j)}$ respectively. The basis of ${\mathfrak}M$ above is invariant up to sign under the induced action of $S_3$ on the lcm-resolution ${\mathfrak}F$. Furthermore, this choice of basis is canonical up to sign: in homological degree $0$ this is the only basis-forming subset of the set of canonical generators; in homological degree $1$ each $g_{ij}$ is up to sign the only canonical generator of degree $m_{ij}$, and $g_0$ is the only canonical generator of degree $m$ that is fixed by $S_3$; and in homological degree $2$ up to sign $f$ is the only canonical generator.
Monomial resolutions in positive characteristic {#S:monomial-resolutions-in-positive-characteristic}
===============================================
Throughout this section $p\ge 2$ is a fixed prime.
We show that there is a monomial ideal that does not have a minimal intrinsic resolution over any algebraic extension of $\mathbb F_p$.
\[D:the-counterexample\] We set $$n=n(p)=
\begin{cases}
4 &\text{ if } p=2; \\
p &\text{ otherwise}.
\end{cases}$$ We write $I=I(p)$ for the monomial ideal on the set of $2n+1$ variables $$X=X(p)=\{v_0,\dots,v_n, e_{1,2},e_{2,3},\dots, e_{n-1,n}, e_{n,1}\}$$ generated by the set of monomials $
\{
m_0,m_1,\dots, m_n
\},
$ where $
m=
\prod_{y\in X}y
$ and $$\begin{gathered}
m_0=\frac m {v_0}, \qquad
m_1=\frac {v_0m} {e_{n,1}v_1e_{1,2}}, \qquad
m_n=\frac {v_0m} {e_{n-1,n}v_ne_{n,1}}, \\
\text{ and } \qquad
m_i= \frac {v_0m} {e_{i-1,i}v_ie_{i,i+1}} \qquad \text{ for } i=2,\dots, n-1.\end{gathered}$$ Also, for convenience of notation, for each $e_{i,j}\in X(p)$ we set $e_{j,i}=e_{i,j}$.
The main result in this section is
[T:counterexample]{} Problem \[P:mfr\] does not have a solution for the ideal $I(p)$ with $\Bbbk=\mathbb F_p$ and $\Bbbk'$ an algebraic field extension of $\mathbb F_p$.
While not immediately apparent from the proof presented here, the fundamental fact behind this theorem is that cyclic $p$-groups have nontrivial cohomology over $\mathbb F_p$. Also, most of the work in the proof is to reduce to the case of group action on the minimal resolution that preserves the $\mathbb Z$-grading. For this reason we postpone the proof till the very end of the section, and proceed with all the necessary results for this desired reduction. Our first task is to describe the group of symmetries of $I(p)$, and its minimal free resolution.
\[R:trivial-action\] Let $\Bbbk$ be a field, and let $R=\Bbbk[X]$ be the polynomial ring over $\Bbbk$ in the set of variables $X=X(p)$.
\(a) Inside the symmetric group $S_n$ consider the dihedral subgroup $D_n$ generated by $\sigma=(1,n)(2,n-1)\cdots(\lceil n/2\rceil, \lceil (n+1)/2\rceil)$ and $\rho=(1,\dots,n)$. The action of $S_n$ on the set $\{1,\dots, n \}$ induces an action of $D_n$ on the set $X$ where $\tau\in D_n$ fixes $v_0$, sends $v_i$ to $v_{\tau(i)}$, and sends $e_{i,j}$ to $e_{\tau(i),\tau(j)}$ when $i,j\ge 1$. It is immediate from this formula that $D_n$ preserves the monomial ideal $I(p)$, and it is straightforward to check that there are no other automorphisms of $M(X)$ that preserve $I(p)$.
\(b) Suppose that $char(\Bbbk) = p$, and that we have an action of $D_n$ as ring automorphisms on the polynomial ring $R$ that extends the action of $D_n$ on the set $X$. Since the units of $R$ are the nonzero elements of the field $\Bbbk$ this yields an action of $D_n$ as field automorphisms on $\Bbbk$. If $x\in \Bbbk$ is algebraic over $\mathbb F_p$ and is not fixed by $D_n$, then we look at the splitting field $\widetilde\Bbbk=\mathbb F_p(x)$ of $x$ over $\mathbb F_p$. Consider the normal subgroup $T$ of $D_n$ consisting of the elements of $D_n$ that act trivially on $\widetilde\Bbbk$. It follows that $D_n/T$ is a subgroup of the Galois group of a finite extension of $\mathbb F_p$, hence is cyclic. Therefore our assumptions on $n$ imply that $\rho$ is in $T$, i.e. $\rho$ acts trivially on $x$. Thus if $\Bbbk$ is algebraic over $\mathbb F_p$ we have that $\rho$ acts trivially on $\Bbbk$.
[L:minimal-resolution]{} Let $\Bbbk$ be any field and let $R=\Bbbk[X]$.
\(a) The lcm-lattice $L$ of $I$ consists of the following elements: $$L=\{\
1, \ m_0,\ \dots\ ,\ m_n,\ v_0m/e_{1,2},\ \dots\ ,\ v_0m/e_{n-1,n},\
v_0m/e_{n,1},\ v_0m \ \}.$$
\(b) The minimal free resolution of $R/I$ over $R$ has the form $$0{\longleftarrow}R \overset{\phi_0}{\longleftarrow}H
\overset{\phi_1}{\longleftarrow}G \overset{\phi_2}{\longleftarrow}F {\longleftarrow}0.$$ Also, there are homogeneous bases $\{h_0,h_1,\dots, h_n\}$, $\{g_0, g_{1,2}, \dots, g_{n-1,n}, g_{n,1}\}$, and $\{f\}$ of $H$, $G$, and $F$, respectively, of degrees $$\label{E:basis-degrees}
\begin{split}
|h_0| & = 2n ; \\
|h_1| = \dots = |h_n| & = 2n-1; \\
|g_{1,2}| = \dots = |g_{n-1,n}| = |g_{n,1}| & = 2n+1; \\
|g_0|= |f| & = 2n+2,
\end{split}$$ and such that the maps $\phi_j$ satisfy the following equalities: $$\label{E:differentials}
\begin{split}
\phi_0(h_i) &= m_i \quad\text{for each }i; \\
\phi_1(g_0) &= v_0^2h_0 - e_{n,1}v_1e_{1,2} h_1; \\
\phi_1(g_{1,2}) &= v_2e_{2,3} h_2 - e_{n,1}v_1 h_1; \\
\phi_1(g_{n,1}) &= v_1e_{1,2} h_1 - e_{n-1,n}v_n h_n ; \\
\phi_1(g_{n-1,n}) &= v_ne_{n,1} h_n - e_{n-2,n-1}v_{n-1} h_{n-1}; \\
\phi_1(g_{i,i+1}) &= v_{i+1}e_{i+1,i+2} h_{i+1} - e_{i-1,i}v_i h_i
\quad\text{for}\quad 3\le i\le n-2; \\
\phi_2(f) &= e_{1,2}g_{1,2} + e_{2,3}g_{2,3}+ \dots + e_{n,1}g_{n,1}.
\end{split}$$
\(a) The claim on the lcm-lattice is straightforward to verify.
\(b) This can be proven by taking the equalities and as the definition of ${\mathfrak}F$, and checking directly that this gives a resolution of $R/I$. However, we prefer the following more conceptual proof.
Consider the set of variables $X'=(X\cup\{v_0^2\})\setminus\{v_0\}$, and let $R'=\Bbbk[X']$ be the polynomial ring over $\Bbbk$ on the set of variables $X'$. Thus $R'$ is a subring of $R$, and $R$ is a free, hence flat $R'$-module. Let $\Delta$ be the simplicial complex on the vertex set $\{v_0^2,v_1,\dots, v_n\}$, with oriented facets $\{v_0^2\}$, $e_{1,2}=\{v_1, v_2\}, \dots, e_{n-1,n}=\{v_{n-1}, v_n\}$, and $e_{n,1}=\{v_n, v_1\}$. Thus $\Delta$ is just the disjoint union of a point and an $n$-cycle. Let $I'$ be the nearly-Scarf ideal of $\Delta$ in $R'$, in the sense of Peeva and Velasco [@PV]. Note that $I$ is exactly the ideal generated in $R$ by $I'$. Choose $\{v_0^2\} - \{v_1\}$ as an oriented cycle generating $\operatorname{\widetilde{H}}_0(\Delta,\Bbbk)$, and $\{e_{12}+e_{23}+\dots + e_{n,1}\}$ as an oriented cycle generating $\operatorname{\widetilde{H}}_1(\Delta,\Bbbk)$, and let ${\mathfrak}F'$ be the corresponding minimal free resolution of $R'/I'$ over $R'$, as described in [@PV Theorem 6.1]. From that description it is immediate to check that the complex ${\mathfrak}F := R\otimes_{R'}{\mathfrak}F'$ has homogeneous bases satisfying all desired equalities. As ${\mathfrak}F$ is a resolution of $R/I$ over $R$ by the flatness of $R$, this completes our proof.
The statements and proofs of all remaining results in this section will adhere strictly to the following convention:
If a field $\Bbbk$ is specified, then ${\mathfrak}F$ always stands for the minimal free resolution $$0{\longleftarrow}H \xleftarrow{\phi_1} G \xleftarrow{\phi_2} F {\longleftarrow}0$$ of $I=I(p)$ over $R=\Bbbk[X]$ with bases and maps as described in Lemma \[L:minimal-resolution\]. For uniform notation, we set $g_{ji}=-g_{ij}$. Furthermore, we will always assume we are given an action of $D_n$ as automorphisms of $R$ that extends the action of $D_n$ on $X$. For each $\tau\in D_n$ we will denote by $\tau$ also the induced automorphism of $R$. Clearly, $\tau$ is a symmetry of $I$ such that $\tau(m_i)=m_{\tau(i)}$ for each $i$, and $\tau{\! :}I {\longrightarrow}\tau_*I$ is an isomorphism of $R$-modules.
\[R:intrinsic-resolution\] (a) Supppose $\Bbbk$ is a field of characteristic $\ne p$. In view of Lemma \[L:minimal-resolution\] it is straightforward to check that the complex $$0{\longleftarrow}H \xleftarrow{\tilde\phi_1}
G \xleftarrow{\phi_2} F {\longleftarrow}0$$ is a minimal resolution ${\mathfrak}M$ of $I(p)$, where $\tilde\phi_1$ agrees with $\phi_1$ on all basis elements of $G$ except on $g_0$ where we have $$\begin{aligned}
\tilde\phi_1(g_0) =
v_0^2h_0 - (1/n)\bigl(e_{n,1}v_1e_{1,2} h_1 &+ e_{1,2}v_2e_{2,3} h_2 + \dots \\
&+ e_{n-2,n-1}v_{n-1}e_{n-1,n} h_{n-1} + e_{n-1,n}v_ne_{n,1} h_n\bigr). \end{aligned}$$ Furthermore, the action of $\tau\in D_n$ on $R$ and $I$ lifts naturally to an action ${\mathfrak}M(\tau)$ on ${\mathfrak}M$ that fixes the basis elements $f, g_0$, and $h_0$, and for $1\le i<j\le n$ sends $h_i$ and $g_{ij}$ to $h_{\tau(i)}$ and $g_{\tau(i)\tau(j)}$, respectively. Thus ${\mathfrak}M$ gives a solution to Problem \[P:mfr\] for $I(p)$ that has $\Bbbk'=\Bbbk$ in characteristic $\ne p$.
\(b) Suppose now that $char(\Bbbk)=p$. Let $Y=\{y_1,\dots, y_n\}$ be a set of variables, and let $\Bbbk'$ be the fraction field of the domain $\Bbbk[Y]/J$, where $\Bbbk[Y]$ is the polynomial ring over $\Bbbk$ in the set of variables $Y$, and $J$ is the principal ideal generated by the linear form $y_1+\dots +y_n -1$. We use $y_i$ to denote also the image of $y_i$ in $\Bbbk'$. Each $\tau\in D_n$ acts on $Y$ via $\tau(y_i)= y_{\tau(i)}$. Since this leaves $J$ invariant, it induces an action on $\Bbbk'$ given by the same formula. Thus we get an action on $R'=\Bbbk'[X]$ that leaves $I(p)R'$ invariant. Just as in part (a), for each $\tau\in D_n$ we obtain an action ${\mathfrak}M'(\tau)$ on the $R'$-modules $H'=H\otimes_R R'$, $G'=G\otimes_R R'$ and $F'=F\otimes_R R'$. This gives an action ${\mathfrak}M'(\tau)$ on the chain complex ${\mathfrak}M'$ given by $$0{\longleftarrow}H' \xleftarrow{\phi_1'}
G' \xleftarrow{\phi_2\otimes 1} F' {\longleftarrow}0$$ where $\phi_1'$ agrees with $\phi_1\otimes 1$ on the basis elements $g_{ij}\otimes 1$, and satisfies $$\begin{aligned}
\phi_1'(g_0\otimes 1) =
v_0^2h_0\otimes 1 &- e_{n,1}v_1e_{1,2} h_1\otimes y_1 -
e_{1,2}v_2e_{2,3} h_2\otimes y_2 - \dots \\
&- e_{n-2,n-1}v_{n-1}e_{n-1,n} h_{n-1}\otimes y_{n-1} -
e_{n-1,n}v_ne_{n,1} h_n\otimes y_n . \end{aligned}$$ It is now straightforward to check that ${\mathfrak}M'$ gives a solution to Problem \[P:mfr\] for the ideal $I(p)$ with $\Bbbk'$ a purely transcendental extension of $\Bbbk$ of degree $n-1$.
[L:unique-action-ideal]{} Let $\tau\in D_n$, and suppose $\Bbbk$ is any field. Let $\psi{\! :}I {\longrightarrow}\tau_*I$ be an isomorphism of $R$-modules.
Then $\psi=c\tau$ for some $c\in\Bbbk$. In particular, if $\psi{\! :}I {\longrightarrow}I$ is an isomorphism, then $\psi=c\operatorname{id}$ for some $c\in\Bbbk$.
Note that for $i=2,\dots,n-1$ one has the relation $v_0^2m_0= e_{i-1,i}v_ie_{i,i+1}m_i$ in $I$. Therefore we get $v_0^2\psi(m_0)=e_{\tau(i-1)\tau(i)}v_{\tau(i)}e_{\tau(i)\tau(i+1)}\psi(m_i)$. Smilarly, the relations $v_0^2m_0=e_{n,1}v_1e_{1,2}m_1$ and $v_0^2m_0=e_{n-1,n}v_ne_{n,1}m_n$ yield the relations $$\begin{aligned}
v_0^2\psi(m_0) &=e_{\tau(n)\tau(1)}v_{\tau(1)}e_{\tau(1)\tau(2)}\psi(m_1) \\
v_0^2\psi(m_0) &=e_{\tau(n-1)\tau(n)}v_{\tau(n)}e_{\tau(n)\tau(1)}\psi(m_n).\end{aligned}$$ It follows that $\psi(m_0)=cm_0$ for some $c\in R$. Therefore $$\begin{aligned}
\psi(m_1)&= cv_0^2m_0/(e_{\tau(n)\tau(1)}v_{\tau(1)}e_{\tau(1)\tau(2)})
=cm_{\tau(1)} \\
\psi(m_n)&= cv_0^2m_0/(e_{\tau(n-1)\tau(n)}v_{\tau(n)}e_{\tau(n)\tau(1)})
=cm_{\tau(n)}.\end{aligned}$$ Finally, for $2\le i\le n-1$ we have $$\psi(m_i)=cv_0^2m_0/(e_{\tau(i-1)\tau(i)}v_{\tau(i)}e_{\tau(i)\tau(i+1)})
=cm_{\tau(i)},$$ therefore $\psi=c\tau$. Since $\psi$ is an epimorphism, $c$ is a unit in $R$, hence $c\in\Bbbk$.
\[L:graded-lowest-degrees\] Let $\Bbbk$ be any field, let $\tau$ be any degree preserving ring automorphism of $R$, let $M$ be any finitely generated graded $R$-module, and let $\gamma{\! :}M {\longrightarrow}\tau_* M$ be an isomorphism of $R$-modules such that $\gamma^m =\operatorname{id}$ for some $m\ge 1$. For any $0\ne u\in M$ write $u^*$ for the nonzero homogeneous component of $u$ of lowest degree. Let $\{w_1,\dots, w_k\}$ be a homogeneous generating set of $M$ such that $|\gamma(w_i)^*| = |w_i|$ for each $i$.
Then for each nonzero $u\in M$ and each $s\ge 1$ we have $|\gamma^s(u)^*| = |u^*|$.
Note first that for every nonzero homogeneous $u$ and every $s\ge 1$ we have $|\gamma^s(u)^*|\ge |u|$. Indeed, we can write $u=r_1w_1 + \dots + r_kw_k$ for some homogeneous $r_i\in R$ with $|r_i|+|w_i|=|u|$ for each $i$. Thus $\gamma(u)= \sum \tau(r_i)\gamma(w_i)$ and hence $|\gamma(u)^*| \ge \min\{ |\tau(r_i)| + |\gamma(w_i)^*| \} = |u|$. Now if $|\gamma^s(u)^*|\ge |u|$ for some $s\ge 1$, then writing $\gamma^s(u) = \sum_{i\ge |\gamma^s(u)^*|}u_i$ for homongeneous $u_i$ with $|u_i|=i$, we obtain $|\gamma^{s+1}(u)^*|\ge \min\{|\gamma(u_i)^*|\} \ge \min\{|u_i|\} \ge |u|$, hence the desired inequality follows by induction.
Next, suppose again $u$ is homogeneous. If $|\gamma^s(u)^*| > |u|$ for some $s\ge 1$ then, writing $\gamma^s(u) = \sum_{i\ge |\gamma^s(u)^*|} u_i$ for some homogeneous $u_i$ of $|u_i|=i$, we obtain $\gamma^{s+1}(u) = \sum_{i\ge |\gamma(u)^*|} \gamma(u_i)$ hence $
|\gamma^{s+1}(u)^*| \ge
\min\{|\gamma(u_i)^*|\} \ge
\min\{|u_i|\} = |\gamma^s(u)^*| > |u|.
$ Thus induction gives us that if $|\gamma^t(u)^*|>|u|$ for some $t\ge 1$ then $|\gamma^s(u)^*|> |u|$ for every $s\ge t$. This contradicts the fact that $\gamma^m(u)=u$. It follows that $|\gamma^s(u)^*| = |u|$ for every $s\ge 1$.
Finally, suppose $u\ne 0$ is arbitrary, and write it as a sum of homogeneous components $u=\sum_i u_i$ with $|u_i|=i$. We get $\gamma^s(u) = \sum_{i\ge |u^*|} \gamma^s(u_i)$, and since $|\gamma^s(u_i)^*| = |u_i| = i$ we must have $|\gamma^s(u)^*| = |u^*|$ as desired.
[L:lowest-degrees]{} Let $\Bbbk$ be any field, let $\tau\in D_n$, and let $\psi{\! :}{\mathfrak}F {\longrightarrow}\tau_*{\mathfrak}F$ be a morphism of chain complexes of $R$-modules such that $\psi^m=\operatorname{id}$ for some $m\ge 1$. Then, for each nonzero element $x$ of $H$, $G$, and $F$, we have that $|\psi(x)^*|=|x^*|$.
Let $\pi{\! :}H {\longrightarrow}\operatorname{Coker}\phi_1 = \operatorname{H}_0{\mathfrak}F$ be the canonical projection, and write $\varphi{\! :}\operatorname{H}_0{\mathfrak}F {\longrightarrow}I$ for the isomorphism of $R$-modules induced by $\phi_0$ so that $\phi_0=\varphi\pi$. In particular, we have that $\varphi{\! :}\tau_*\operatorname{H}_0{\mathfrak}F {\longrightarrow}\tau_* I$ is also an isomorphism of $R$-modules. By Lemma \[L:unique-action-ideal\] we have $\varphi{\overline}\psi\varphi^{-1}=c\tau$ for some $c\in\Bbbk$, where ${\overline}{\psi}$ is the isomorphism on homology induced by $\psi$. The result for $x=h_i$ with $0\le i\le n$ is now immediate from the fact that $\ker(\phi_0)=\operatorname{Im}(\phi_1)$ is generated in degrees $\ge 2n+1$. Since the $h_i$s generate $H$, the result for the other nonzero elements $x$ of $H$ follows from Lemma \[L:graded-lowest-degrees\].
Next let $x=g_0$ or $x=g_{i,i+1}$ or $x=g_{n,1}$. Consider the decomposition $\psi(x) = \sum_{i\ge |\psi(x)^*|} u_i$ into homogeneous components with $|u_i|=i$. Then, since $\phi_1$ preserves degrees, $\phi_1\psi(x)=\sum_{i\ge |\psi(x)^*|}\phi_1(u_i)$ is a decomposition of $\psi\phi_1(x)$ into homogeneous components. We already have shown that $|\psi\phi_1(x)^*|=|\phi_1(x)| = |x|\le 2n+2$, and since $\ker(\phi_1)=\operatorname{Im}(\phi_2)$ is generated in degree $2n+2$, we must have $u_i=0$ for $i< |x|$ and $u_{|x|}\ne 0$. Thus $|\psi(x)^*|=|x|$. The result for the other nonzero elements of $G$ is now immediate from Lemma \[L:graded-lowest-degrees\].
Finally, the result for the nonzero elements of $F$ is immediate from the above by the injectivity of $\phi_2$.
\[L:to-homogeneous-action\] Let $\Bbbk$ be any field, let $\tau\in D_n$, and let $\psi{\! :}{\mathfrak}F {\longrightarrow}\tau_*{\mathfrak}F$ be a morphism of chain complexes of $R$-modules such that $\psi^m=\operatorname{id}$ for some $m\ge 1$. Let $\widetilde\psi{\! :}{\mathfrak}F {\longrightarrow}\tau_*{\mathfrak}F$ be the map defined on each basis element $x$ of $H$, $G$, and $F$ by $\widetilde\psi(x) = \psi(x)^*$.
Then $\widetilde\psi$ is an isomorphism of complexes of graded $R$-modules, and $\widetilde\psi^m=\operatorname{id}$.
Let $x$ be a basis element of $H$, $G$, or $F$. Then $\phi(x)$ is a nonzero homogeneous element of ${\mathfrak}F$ of same degree as $x$. We know $\psi[\phi(x)]=\phi[\psi(x)]$, and by Lemma \[L:lowest-degrees\] we also have the decomposition $\psi(x)=\sum_{i\ge |x|}u_i$ into homogenous components. This shows that $\sum_{i\ge |x|}\phi(u_i)$ is a decomposition of $\psi[\phi(x)]$ into homogeneous components. Since we know from Lemma \[L:lowest-degrees\] that $|\psi[\phi(x)]^*|=|\phi(x)|=|x|$, we get $
\widetilde\psi[\phi(x)]=\psi[\phi(x)]^*=\phi(u_{|x|})=
\phi[\psi(x)^*]=\phi[\widetilde\psi(x)]$.
Next, note that by Lemma \[L:graded-lowest-degrees\] and Lemma \[L:lowest-degrees\] we have $|\psi^s(x)^*|=|x^*|$ for every nonzero element $x$ of ${\mathfrak}F$ and every $s\ge 1$. Therefore $\psi^{s+1}(x)^*=\psi^s[\psi(x)^*]^*$ for every $s\ge 1$ and every nonzero $x$. In particular, for every homogeneous basis element $x$ we get $\psi^s(x)^*=\psi^{s-1}[\widetilde\psi(x)]^*$. Iterating this we obtain $
x=\psi^n(x)=\psi^n(x)^*=\psi^{n-1}[\widetilde\psi(x)]^* =
\psi^{n-2}[(\widetilde\psi)^2(x)]^*
= \dots = \psi[\widetilde\psi^{n-1}(x)]^* = \widetilde\psi^n(x)
$.
\[L:homogeneous-action-formulas\] Let $\Bbbk$ be an algebraic extension of $\mathbb F_p$, and let $\psi{\! :}{\mathfrak}F {\longrightarrow}\rho_*{\mathfrak}F$ be a morphism of chain complexes of graded $R$-modules such that $\psi^n=\operatorname{id}$. Then the diagram $$\label{E:unique-action-complex}
\begin{CD}
H @> \psi_0 >> \rho_* H \\
@V \phi_0 VV @VV \phi_0 V \\
I @>> \rho > \rho_* I
\end{CD}$$ is commutative, and we have $$\label{E:action-formulas}
\begin{split}
\psi(h_i) &= h_{\rho(i)} \quad\text{for } 0\le i\le n; \\
\psi(g_{i,j})&=g_{\rho(i),\rho(j)} \\ \psi(g_0) &= g_0 + a_1g_{1,2} + \dots + a_ng_{n,1}; \\
\psi(f)&=f
\end{split}$$ where the $a_i$ are homogeneous elements of $R$ of degree $1$.
Let $\pi$ and $\varphi$ be as in the proof of Lemma \[L:lowest-degrees\]. We obtain therefore an isomorphism of $R$-modules $\widehat\psi=\varphi{\overline}\psi\varphi^{-1}{\! :}I {\longrightarrow}\rho_* I$, where ${\overline}{\psi}$ is the isomorphism on homology induced by $\psi$. Thus $\widehat\psi = c\rho$ for some $c\in\Bbbk$ by Lemma \[L:unique-action-ideal\]. Since by Remark \[R:trivial-action\](b) $\rho$ acts trivially on $\Bbbk$, this yields $
\operatorname{id}= \widehat{\psi}^n = c^n\rho^n = c^n\operatorname{id},
$ hence $c^n=1$ and, as $n$ is a power of $p$, we get $c=1$. It follows that commutes as desired.
Next, since $\psi$ induces an isomorphism on the homogeneous components of $H$ of degrees $2n-1$ and $2n$, and since $\phi_0$ is injective on these components, the desired equalities $\psi(h_i)=h_{\rho(i)}$ for $i=0,\dots,n$ are immediate from the commutativity of . Since $\phi_1\psi(g_{i,j}) = \psi\phi_1(g_{i,j})=\phi_1(g_{\rho(i),\rho(j)})$, the injectivity of $\phi_1$ in degree $2n+1$ yields that $\psi(g_{i,j})=g_{\rho(i),\rho(j)}$ as desired. Now we have $\phi_2\psi(f)=\psi\phi_2(f)=\phi_2(f)$, therefore $\psi(f)=f$ by injectivity of $\phi_2$.
Finally, since $\psi$ preseves degrees, we must have $\psi(g_0)=cg_0 + h'$ where $h'=a_1g_{1,2}+\dots +a_ng_{n,1}$ for some homogeneous $a_i\in R$ of degree $1$ and some constant $c\in \Bbbk$. Since by Remark \[R:trivial-action\](b) $\rho$ acts trivially on $\Bbbk$, it follows that $g_0 = \psi^n(g_0) = c^ng_0 + \sum_{k=1}^nc^{n-k}\psi^{k-1}(h')$. By what we already proved we know that each $\psi^{k-1}(h')$ is again a homogeneous linear combination of the $g_{i,j}$s, and therefore $c^n=1$ hence $c=1$ as $n$ is a power of $p$.
It suffices to show that if $\Bbbk$ is an algebraic field extension of $\mathbb F_p$ then the action on $I$ of the cyclic group $C_n$ generated by $\rho$ in $D_n$ cannot be lifted to an action on ${\mathfrak}F$. Indeed, suppose we have an isomorphism $\psi{\! :}{\mathfrak}F{\longrightarrow}\rho_*{\mathfrak}F$ such that $\psi^n=\operatorname{id}$. Then by Lemma \[L:to-homogeneous-action\] and Lemma \[L:homogeneous-action-formulas\], we may assume that $\psi$ acts via the equations . Since by Remark \[R:trivial-action\](b) $\rho$ acts trivially on $\Bbbk$, going modulo the augmentation ideal $J$ in $R$ generated by the set $\{y-1 \mid y\in X \}$, we obtain an action of the group ring $\Bbbk[C_n]$ on the chain complex of $\Bbbk$-vector spaces $\overline{{\mathfrak}F}= {\mathfrak}F/J{\mathfrak}F$. We see that $\Bbbk[C_n]\bar\phi_1(\bar g_0)$ is $n$-dimensional over $\Bbbk$, hence is a free $\Bbbk[C_n]$-module, therefore $\Bbbk[C_n]\bar g_0$ is also a free $\Bbbk[C_n]$-module, in particular the element $w=(1 + \rho + \dots + \rho^{n-1})\bar g_0$ must be a non-zero $C_n$-invariant element of the $\Bbbk[C_n]$-submodule $G'$ of $\overline G$ generated over $\Bbbk$ by the elements $\bar g_{i,j}$. Since the only invariants of $G'$ are the $\Bbbk$-multiples of $\bar\phi_2(\bar f)$, we see that $w$ is a non-zero element in $\operatorname{Ker}(\phi_1)$, hence $\Bbbk[C_n]\phi_1(\bar g_0)$ cannot be a free $\Bbbk[C_n]$-module, a contradiction.
[99]{}
E. Batzies and V. Welker. Disrete Morse theory for cellular resolutions. , 543:147–168, 2002.
D. Bayer and B. Sturmfels. Cellular resolutions of monomial modules. , 502:123–140, 1998.
A. Bjerhammar. Application of calculus of matrices to method of least squares; with special references to geodetic calculations. , 49, 1951.
R. Brown. The twisted Eilenberg-Zilber theorem. , Edizioni Oderisi, Gubbio, 1965, pp. 33–37.
T. Clark and A. Tchernev. Minimal free resolutions of monomial ideals and of toric rings are supported on posets. , 371(6):3995–4027, 2019.
C. Conley. C.B.M.S. Regional Conf. Series in Math., no. 38, Amer. Math. Soc., Providence, R.I., 1978.
J. Eagon. Partially split double complexes with an associated Wall complex and applications to ideals generated by monomials. , 135:344–362, 1990.
J. Eagon, E. Miller, and E. Ordog. Minimal resolutions of monomial ideals. , 2019, [arXiv:1906.08837.]{}
D. Eisenbud, G. Fløystad, and F.-O. Schreyer. Sheaf cohomology and free resolutions over exterior algebras. , 355(11):4397–4426, 2003.
R. Forman. Morse theory for cell complexes. , 134:90–145, 1998.
R. Forman. Combinatorial vector fields and dynamical systems. , 228:629–681, 1998.
J. Franks. . C.B.M.S. Regional Conf. Series in Math., no. 49, Amer. Math. Soc., Providence, R.I., 1982.
W. Fulton. . Annals of Mathematics Studies, vol. 131, Princeton Univ. Press, 1993.
V. Gasharov, I. Peeva, and V. Welker. The lcm-lattice in monomial resolutions. , 6:521–532, 1999.
G.H. Golub and C.F. Van Loan. . Baltimore: The Johns Hopkins Univ. Press, 2013.
V.K.A.M. Gugenheim. On the chain-complex of a fibration. , 16:398–414, 1972.
V.K.A.M. Gugenheim and L. Lambe. Perturbation theory in differential homological algebra. I. , 33(4):566–582, 1989.
M. J" ollenbeck and V. Welker. Minimal resolutions via algebraic discrete Morse theory. , 197(923), 2009.
L. Lambe. Homological perturbation theory, Hochschild homology, and formal groups. , 134:183–218, 1992.
E. Miller and B. Sturmfels. . Graduate Texts in Mathematics, vol. 227, Springer-Verlag, 2004.
E.H. Moore. On the reciprocal of the general algebraic matrix. , 26(9):394–-95, 1920.
P. Orlik and V. Welker. . Universitext. Springer, Berlin, 2007.
I. Peeva. . Algebra and Applications, vol. 14, Springer-Verlag, 2011.
I. Peeva and M. Velasco. Frames and degenerations of monomial ideals. , 363(4):2029–2046, 2011.
R. Penrose. A generalized inverse for matrices. , 51(3):406–-13, 1955.
W. Shih. Homologie des espaces fibrés. , 13:5–87, 1962.
E. Sk" oldberg. Morse theory from an algebraic viewpoint. , 358(1):115–129, 2006.
E. Sk" oldberg. Algebraic Morse theory and homological perturbation theory. , 26(1):124–129, 2018.
D. Taylor. . PhD thesis, University of Chicago, 1966.
A. Tchernev. Torsion freeness of symmetric powers of ideals. , 359:3357–3367, 2007.
A. Tchernev and M. Varisco. Modules over categories and [B]{}etti posets of monomial ideals. , 143(12):5113–5128, 2015.
A. Tchernev and M. Varisco. Betti categories of graded modules and applications to monomial ideals and toric rings. , 2016, [arXiv:1605.09748.]{}
S. Yuzvinsky. Taylor and minimal resolutions of homogeneous polynomial ideals. , 6:779–793, 1999.
C. A. Weibel. . Cambridge Studies in Advanced Mathematics, vol. 38, Cambridge Univ. Press, Cambridge, 1994.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
This paper considers a general data-fitting problem over a networked system, in which many computing nodes are connected by an undirected graph. This kind of problem can find many real-world applications and has been studied extensively in the literature. However, existing solutions either need a central controller for information sharing or requires slot synchronization among different nodes, which increases the difficulty of practical implementations, especially for a very large and heterogeneous system.
As a contrast, in this paper, we treat the data-fitting problem over the network as a stochastic programming problem with many constraints. By adapting the results in a recent paper [@wang2015random], we design a fully distributed and asynchronized stochastic gradient descent (SGD) algorithm. We show that our algorithm can achieve global optimality and consensus asymptotically by only local computations and communications. Additionally, we provide a sharp lower bound for the convergence speed in the regular graph case. This result fits the intuition and provides guidance to design a ‘good’ network topology to speed up the convergence. Also, the merit of our design is validated by experiments on both synthetic and real-world datasets.
author:
- 'Ying Zhang [^1]'
bibliography:
- 'ref.bib'
- 'ref.bib'
title: Fully Distributed and Asynchronized Stochastic Gradient Descent for Networked Systems
---
[^1]: The author is with Department of Information Engineering, the Chinese University of Hong Kong.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the spin injection efficiency into single and bilayer graphene on the ferrimagnetic insulator Yttrium-Iron-Garnet (YIG) through an exfoliated tunnel barrier of bilayer hexagonal boron nitride (hBN). The contacts of two samples yield a resistance-area product between 5 and 30 k${\Omega}\mu$m$^2$. Depending on an applied DC bias current, the magnitude of the non-local spin signal can be increased or suppressed below the noise level. The spin injection efficiency reaches values from -60% to +25%. The results are confirmed with both spin valve and spin precession measurements. The proximity induced exchange field is found in sample A to be (85 $\pm$ 30) mT and in sample B close to the detection limit. Our results show that the exceptional spin injection properties of bilayer hBN tunnel barriers reported by Gurram et al. are not limited to fully encapsulated graphene systems but are also valid in graphene/YIG devices. This further emphasizes the versatility of bilayer hBN as an efficient and reliable tunnel barrier for graphene spintronics.'
author:
- 'J.C. Leutenantsmeyer'
- 'T. Liu'
- 'M. Gurram'
- 'A.A. Kaverzin'
- 'B.J. van Wees'
bibliography:
- 'references.bib'
date:
-
-
title: Bias dependent spin injection into graphene on YIG through bilayer hBN tunnel barriers
---
Introduction
============
The combination of graphene with other two dimensional layered materials is an elegant way to create atomically thin devices with adjustable properties [@Geim2013; @Han2014; @Roche2015]. The crystalline insulator hexagonal boron nitride is an appealing material for the field of graphene spintronics [@Gurram2018]. Its atomic flatness and sufficiently strong van der Waals interaction with graphene allows the fabrication of heterostructures of 2D materials with minimized contamination, implying good spin transport properties. A long spin diffusion length of 30 $\mu$m has been experimentally achieved in graphene where a bulk flake of hBN was used as protective layer to avoid contamination during the fabrication process [@Drogeler2016]. Therefore, the use of hBN as a pinhole free tunnel barrier is straightforward since these fully encapsulated graphene devices suggest minimized contamination and highly efficient spin transport. Several experimental studies have investigated the spin injection through tunnel barriers of exfoliated hBN [@Yamaguchi2016; @Gurram2016] and large scale hBN grown via chemical vapor deposition [@Fu2014; @Kamalakar2015; @Kamalakar2016; @Gurram2018a]. However, the experimentally demonstrated spin transport lengths are still far below the values suggested by the low intrinsic spin orbit coupling of graphene [@Huertas-Hernando2006].
Having graphene in proximity to magnetic materials is a novel approach to tune the intrinsic properties of graphene. Magnetic graphene is characterized by the induced exchange field [@Leutenantsmeyer2017; @Singh2017; @Wei2016; @Wang2015; @Asshoff2017]. First principle calculations of idealized systems predict an exchange splitting of the graphene spin states to exceed several tens of meV [@Yang2013; @Hallal2017]. However, the experimentally demonstrated exchange fields are still several orders of magnitude below [@Leutenantsmeyer2017; @Singh2017; @Evelt2016].
The realization of graphene devices with a large exchange field requires the tackling of several challenges. The cleanliness of the interface between graphene and YIG is crucial to obtain a strong exchange effect as indicated by the discrepancy between experimentally achieved values and theoretical predictions. Furthermore, the interface and tunnel barrier between the graphene flake and contacts are crucial for the injection of a large spin accumulation and the observation of large spin signals. In our previous works we employed tunnel barriers of oxidized titanium or aluminum to overcome the conductivity mismatch problem [@Schmidt2000; @Rashba2000]. For these types of tunnel barrier the magnitude of the spin signal is limited by pinholes and resulted in a relatively small spin signal of mostly less than 1 ${\Omega}$, which often did not exceed the electrical noise of the measured signals in the sample. In addition, the contamination arising from the PMMA-based fabrication procedure affects the graphene cleanliness negatively. For this study we replace the AlO$_\mathrm{x}$ or TiO$_\mathrm{x}$ tunnel barrier with a bilayer-hBN (bl-hBN) flake, which significantly improves the sample quality and spin signal. Furthermore, we confirm the tunable spin injection reported by Gurram et al. [@Gurram2017] for the graphene/YIG system.
Sample preparation and contact characterization
===============================================
Thin hBN flakes are exfoliated from hBN crystals (HQ Graphene) onto 90 nm SiO$_2$ wafers. The thickness of the flakes is estimated through their optical contrast, which is calibrated by atomic force microscopy. In our microscope (Zeiss Axio Imager.A2m with an EC Epiplan-Neofluar 100x/0.9 objective) bl-hBN corresponds to 2.5% contrast in the green channel. Suitable bl-hBN flakes are picked up by using a dry polycarbonate based transfer method [@Zomer2014] and combined with single- (sample A) or bilayer graphene (sample B) exfoliated from HOPG crystals (ZYB grade, HQ Graphene). The stack is placed on a cleaned 12 $\mu$m YIG grown by liquid phase epitaxy (LPE) on a 600 $\mu$m gadolinium-gallium-garnet substrate (Matesy GmbH). Before the transfer, the YIG substrate for sample A is treated with oxygen plasma to remove organic contaminants and annealed in a 500${^\circ\mathrm{C}}$ furnace in an oxygen atmosphere prior to the transfer of the graphene/bl-hBN stack. The substrate of sample B underwent an additional argon plasma treatment before the annealing step.
The polycarbonate is dissolved in chloroform and the bl-hBN/graphene/YIG stack is cleaned in acetone, isopropanol and sequent annealing for one hour at 350${^\circ\mathrm{C}}$ in an argon-hydrogen atmosphere. Contacts are defined using a standard PMMA-based electron beam lithography process. The electrodes are evaporated at pressures below 10$^{-7}$ mbar and consist of 45 nm cobalt and a 5 nm aluminum capping layer. After the liftoff in warm acetone, the sample (Figs. \[Fig1\]a and \[Fig1\]b) is loaded into a cryostat and kept in vacuum during the characterization. All measurements are carried out at 75 K.
![a) Optical micrograph of the sample A. The outer electrodes (R) are not covered by bl-hBN and used as reference electrodes in both local and non-local measurements. b) Optical micrograph of sample B. c) Schematic measurement of the three-terminal contact resistance. d) All working contacts have a calculated resistance-area product between 5 and 30 k${\Omega}\mu$m$^2$. The full set of IV characteristics is shown in the supplementary information. \[Fig1\]](Fig1.png){width="1\linewidth"}
After loading into the cryostat of the measurement setup, the samples are cooled down to liquid nitrogen temperature and the contacts are characterized in a three-terminal geometry (Fig. \[Fig1\]c) using the outermost contacts as reference electrodes. The resistance-area product is calculated from the current-voltage characteristics and shown for sample A in Fig. \[Fig1\]d. The contacts on sample A and B which employ a bl-hBN tunnel barrier yield a typical resistance-area product between 5 and 30 k${\Omega}\mu$m$^2$, a range comparable to the one reported in [@Gurram2017]. An hBN covered graphene Hall bar sample fabricated in parallel with sample B for comparison yields a carrier density of n = $5\times 10^{12}$ cm$^{-2}$ and a mobility of $\mu = 5400\,\mathrm{cm}^2$/Vs. We found $\mu = 720\,\mathrm{cm}^2$/Vs (estimated via the Shubnikov-de Haas oscillations) in our previous work [@Leutenantsmeyer2017] and conclude that the protective hBN layer significantly improves the graphene charge transport properties on YIG.
Bias-dependent spin injection through bilayer hBN tunnel barriers into single and bilayer graphene on YIG
=========================================================================================================
We now discuss the spin transport in graphene on YIG with a bl-hBN tunnel barrier in a non-local geometry (Fig. \[Fig2\]a). A current of I$_\mathrm{AC}$ = 1 $\mu$A is sourced and modulated with 3.7 Hz between contacts 2 and R2. The ferromagnetic electrode injects a spin current into the graphene underneath contact 2. These spins are diffusing along the graphene channel and are probed by a lock-in as a voltage difference $\mathrm{V}_\mathrm{NL}$ between the detector contact 1 and the reference electrode R1. Using this technique, we can decouple charge and spin transport. The signal can be defined as non-local resistance and calculated via ${\mathrm{R}_\mathrm{NL}}= \mathrm{V}_\mathrm{NL}/\mathrm{I}_\mathrm{AC}$. To characterize the basic spin transport properties of the samples an in-plane magnetic field parallel to the electrodes (B$_\mathrm{app}$) is applied to switch the magnetization of the injector and detector (Fig. \[Fig2\]a). Depending on the relative magnetization alignment of the injector and detector electrodes, the non-local resistance changes between the parallel and the antiparallel resistance states when the contact magnetization switches. This measurement represents a characteristic spin valve behavior (Figs. \[Fig2\]b and \[Fig2\]c) and gives an estimation of the spin relaxation length in the graphene flake (Fig. \[Fig2\]d).
![a) Schematic setup for a non-local spin valve measurement. b) Non-local spin valve measurements of sample A (bl-hBN/graphene). The size of the switch between parallel and antiparallel states of contacts 1 and 2 can be tuned with the applied DC bias and is shown for four different values. c) Sample B (bl-hBN/bl-graphene) shows a comparable dependence on the applied DC bias. Note that the spin signal changes the sign around -92 mV. d) The distance dependent spin valve measurements of sample A allow the estimation of the spin relaxation length from the slope of the linear fit. The same analysis for sample B is discussed in the supplementary material. \[Fig2\]](Fig_2.png){width="1\linewidth"}
To study the effect of the bias on the spin injection, we apply a DC current additionally to the AC current sourced between injector and reference electrode (Fig. \[Fig2\]a). The dielectric strength of hBN is approximately 1.2 V/nm [@Hattori2014]. Therefore, we limit the DC bias current for sample A to 20 $\mu$A, which corresponds to 0.4 – 0.6 V, depending on the IV characteristics of the injector contact. To compare different contacts, we calculate the equivalent voltage ${\mathrm{V_{hBN}}}$ across the hBN tunnel barrier from the applied DC bias current and discuss all results plotted as function of ${\mathrm{V_{hBN}}}$.
Figure \[Fig2\]b contains the spin valve measurements of sample A for four different DC bias currents over distance d = 1.6 $\mu$m. While no spin signal above noise level is visible at -92 mV, a DC bias current of +333 mV results in a clear switching between parallel and antiparallel states with a spin signal of approximately 0.4 ${\Omega}$. Beyond -92 mV, we find an inverted sign of the non-local resistance switching and a spin signal of -0.4 ${\Omega}$ at -155 mV and -0.7 ${\Omega}$ at -257 mV.
Four spin valve measurements of sample B are shown in Fig. \[Fig2\]c. where we find compared to sample A a larger spin signal of up to -2.5 ${\Omega}$ at -356 mV DC bias. The change of the sign of the spin signal occurs in sample B also between -100 mV and 0 mV, a similar range as in the measurements on sample A.
The distance dependence of the spin signal is shown for sample A in Fig. \[Fig2\]d, from which we extract the spin relaxation length $\lambda \sim$ (740 $\pm$ 570) nm. In our previous work we found a comparable value of $\lambda = (490 \pm 40)$ nm for a not hBN protected sample. We conclude that even though the charge transport properties have improved significantly, the spin transport parameters remain similar. The same analysis was applied to sample B, where we found $\lambda \sim$ (2.3 $\pm$ 1) $\mu$m (supplementary material). The bl-hBN tunnel barriers in Fig. \[Fig2\]d show a less clear trend in the distance dependence, resulting in a larger error in $\lambda$. We can attribute this to two origins: an inhomogeneity of the bl-hBN tunnel barriers and an inhomogeneity in the graphene flake. Microscopic cracks in the hBN tunnel barrier could arise during the fabrication and could lead a to a different spin polarization of each contact. This interpretation is also supported by the considerable spread of the resistance-area product of between 5 to 30 k${\Omega}\mu$m$^2$. As a consequence, the values for the spin relaxation length extracted from the distance dependent measurements can only be seen as approximation. However, the consistency with the spin precession measurements as discussed in the following sections confirms the validity of the estimation.
![Non-local spin transport in a) sample A and b) sample B for different DC bias voltages. For comparison the dependence is shown as a function of the bias voltage applied across the hBN barrier. The blue and red curves correspond to the configuration where detector and biased injector contacts are swapped. The spin polarization on the right side of both panels is extracted from the independently measured $\Delta {\mathrm{R}_\mathrm{NL}}$. \[Fig3\]](Fig3.png){width="1\linewidth"}
To extract the DC bias dependence of the spin injection polarization in the cobalt/bl-hBN/graphene/YIG system, we align the magnetization of injector and detector parallel or antiparallel and sweep the DC bias current. $\Delta {\mathrm{R}_\mathrm{NL}}$ = ${\mathrm{R}_\mathrm{NL}}$(P) - ${\mathrm{R}_\mathrm{NL}}$(AP) is calculated and yields the pure spin signal of samples A and B shown in Figs. \[Fig3\]a and \[Fig3\]b. For comparison, both curves are plotted as a function of ${\mathrm{V_{hBN}}}$. While both positive and negative DC biases lead to an enhanced spin injection, a sign change at approximately -80 mV is observed. To extract the bias dependence of the spin injection polarization, we use the unbiased non-local spin signals to calculate the average spin polarization ($\sqrt{{\mathrm{P}_\mathrm{I}}{\mathrm{P}_\mathrm{D}}}$) of injector ${\mathrm{P}_\mathrm{I}}$ and detector ${\mathrm{P}_\mathrm{D}}$. This assumption is justified by the similar shape of the non-local resistances in Figs. \[Fig3\]a and \[Fig3\]b, when injector and detector contacts are swapped. This suggests a similar behavior of both contacts. We can extract a spin polarization via: $$\begin{aligned}
{\mathrm{P}_\mathrm{I}}\cdot {\mathrm{P}_\mathrm{D}}= \frac{\Delta {\mathrm{R}_\mathrm{NL}}\cdot \mathrm{w}}{{\mathrm{R}_\mathrm{sq}}\cdot \lambda} e^{-\mathrm{d}/\lambda} $$ where $\Delta {\mathrm{R}_\mathrm{NL}}$ the spin signal, w the width of the flake, ${\mathrm{R}_\mathrm{sq}}$ the square resistance, $\lambda$ the spin relaxation length and d the injector to detector distance measured from the centers. Under the assumption that ${\mathrm{P}_\mathrm{I}}= {\mathrm{P}_\mathrm{D}}$ we obtain an unbiased spin polarization of 14.65% for sample A and 10.86% for sample B. Because we apply the DC bias only to the injector contact, the spin polarization of the detector remains constant and can be used to extract the dependence of the differential spin injection polarization on the DC bias. We note that the feature of sample A around zero DC bias seems to be a characteristic feature of these particular contacts and does not appear on all contacts on sample A (see supplementary information).
Bias dependent spin precession measurements and estimation of the proximity induced exchange field in bl-hBN/graphene/YIG
=========================================================================================================================
To estimate the strength of the induced exchange field, we apply and rotate a small magnetic field (${\mathrm{B}_\mathrm{app}}$ = 15 mT) in the sample plane (Fig. \[Fig4\]a). The low in-plane coercive field of the YIG films allows us to rotate the YIG magnetization and simultaneously the proximity induced exchange field while leaving the magnetization of the cobalt injector and detector remain unaffected. The resulting modulation of the non-local resistance is a direct consequence of ${\mathrm{B}_\mathrm{app}}+ {\mathrm{B}_\mathrm{exch}}$ and can be only explained by the presence of such [@Leutenantsmeyer2017; @Singh2017].
The analysis of this effect gives us an estimate for the strength of the exchange field and allows us the fitting of the Hanle curves to extract further spin transport parameters. The higher order oscillations that remain in the symmetrized data in Fig. \[Fig4\]b could indicate the presence of local stray fields of the cobalt contacts influencing the local YIG magnetization or an anisotropy arising from the shape of the YIG substrate which might not be fully aligned with the applied magnetic field of 15 mT. Therefore, we apply a smoothing on the data. The resulting curve is shown in red. We estimate the modulation to be (11 $\pm$ 5)% over d = 1.6 $\mu$m, which, given the uncertainty arising from the smoothing process, should be seen as a rather rough approximation. Despite the uncertainty of the exact value of the modulation, the angular dependence indicates the presence of an exchange field in the sample.
![Modulation of spin transport with the exchange field in sample A. a) Schematics of the experiment. ${\mathrm{B}_\mathrm{app}}$ is rotating the YIG magnetization and the exchange field ${\mathrm{B}_\mathrm{exch}}$ in the sample plane while leaving the electrodes and injected spins unaffected. b) The angle dependence of the non-local resistance is measured at T = 10 K and -20 $\mu$A DC bias in parallel and antiparallel alignment. The subtracted spin signal is symmetrized. As a guide to the eye the smoothed data is shown in red, from which we estimate a relative modulation of 11%. c) Fitting of the experimental relative modulation of 11% with our model using $\tau_\mathrm{s}$ and ${\mathrm{B}_\mathrm{exch}}$ as fitting parameters. $\lambda$ = 700 nm and ${\mathrm{B}_\mathrm{app}}$ = 15 mT are fixed parameters. d) Relative modulation of the spin signal calculated from the model using best fit parameters $\tau_\mathrm{s}$ = 14 ps and $\lambda$ = 700 nm, obtained as shown in Fig. \[Fig5\]. ${\mathrm{B}_\mathrm{exch}}$ is varied as indicated, and ${\mathrm{B}_\mathrm{app}}$ = 15 mT. \[Fig4\]](Fig4.png){width="1\linewidth"}
Using the model reported in Leutenantsmeyer et al. [@Leutenantsmeyer2017] we can simulate the modulation of a spin current by exchange field induced precession. To estimate the magnitude of the exchange field leading to 11% modulation, we use $\lambda$ = 700 nm (Fig. \[Fig2\]d) and assume ${\tau_{\mathrm{s}}}$ to be between 5 and 30 ps, a common range for our single layer graphene devices on YIG. To match the experimental modulation, an exchange field between 0 and 250 mT is required (Fig. \[Fig4\]c). To determine the exact value of ${\tau_{\mathrm{s}}}$, we use the parameter pairs of ${\tau_{\mathrm{s}}}$ and ${\mathrm{B}_\mathrm{exch}}$ to fit, as discussed later, the spin precession measurements in Fig. \[Fig5\]a. By comparing both, we find that the both measurement sets can only be fit consistently with ${\tau_{\mathrm{s}}}$ = 14 ps and ${\mathrm{B}_\mathrm{exch}}$ = 85 mT.
Fig. \[Fig4\]d contains the modulation caused by the combination of the applied magnetic field of 15 mT and different values for the exchange field. The expected relative modulation caused by an applied magnetic field of 15 mT with $\lambda$ = 700 nm and $\tau_\mathrm{s}$ = 14 ps does not exceed 0.5%, whereas the observed modulation is clearly larger. To fit the experimentally found modulation of 11%, we have to assume ${\mathrm{B}_\mathrm{exch}}$ = 85 mT. This is a strong indication for the presence of an exchange field in this device. We can conclude that within the uncertainty range of the relative modulation of (11 $\pm$ 5)%, the exchange field in sample A is (85 $\pm$ 35) mT.
![Spin precession measurements in sample A: a) The Hanle spin precession curves from sample A are fit using our exchange model with ${\mathrm{B}_\mathrm{exch}}$ = 85 mT (solid lines) for different DC bias currents. Contact 1 is used as injector, contact 2 as detector (Fig. \[Fig2\]a). We extract b) the calculated spin polarization the injector (${\mathrm{P}_\mathrm{I}}$), c) the spin diffusion coefficient ${\mathrm{D}_{\mathrm{s}}}$ and d) the spin diffusion time ${\tau_{\mathrm{s}}}$. The DC bias dependence ${\mathrm{P}_\mathrm{I}}$ shows a similar dependence as (red line in panel b, Fig. \[Fig3\]d). \[Fig5\]](Fig5.png){width="1\linewidth"}
The Hanle measurements are carried out in parallel and antiparallel alignment of the injector (contact 1) and detector (contact 2), see Fig. \[Fig2\]a for the contact labeling. We extract the spin signal by calculating $[{\mathrm{R}_\mathrm{NL}}(\mathrm{P})-{\mathrm{R}_\mathrm{NL}}(\mathrm{AP})]/2$, shown in Fig. \[Fig5\]a. From the Hanle fit using an exchange field of 85 mT, we extract the polarization of the injector P (Fig. \[Fig5\]b), the spin diffusion coefficient ${\mathrm{D}_{\mathrm{s}}}$ (Fig. \[Fig5\]c) and the spin diffusion time $\tau_\mathrm{s}$ (Fig. \[Fig5\]d). While ${\mathrm{D}_{\mathrm{s}}}$ = (350 $\pm$ 65) cm$^2$/s and $\tau_\mathrm{s}$ = (16 $\pm$ 5) ps remain approximately constant over the applied DC bias range we find a dependence of the injector spin polarization that resembles the DC bias dependence of the injector (Fig. \[Fig3\]a), which implies a consistency in the analysis. Using the spin diffusion coefficient ${\mathrm{D}_{\mathrm{s}}}$ and time $\tau_\mathrm{s}$ extracted from the Hanle measurements, we can calculate the spin relaxation length $\lambda = \sqrt{{\mathrm{D}_{\mathrm{s}}}\tau_\mathrm{s}} = (730 \pm 230)$ nm. When compared to the estimation from the distance dependent spin valve measurements (Fig. \[Fig2\]a) both approaches yield similar values which indicates again the consistency of the analysis.
Note that the rather smooth Hanle curves shown in Fig. \[Fig5\]a could be also fit with a conventional spin precession model that does not include any exchange field. These fittings yield ${\tau_{\mathrm{s}}}$ $\sim$ 25 ps, ${\mathrm{D}_{\mathrm{s}}}$ $\sim$ 800 cm$^2$/s and $\lambda$ $\sim$ 1.4 $\mu$m. Apart from ${\mathrm{D}_{\mathrm{s}}}$ being unrealistically large, the extracted $\lambda$ is two times larger than the result from the independently measured distance dependent spin valves (Fig. \[Fig2\]d) which suggests that the fit of our results with the conventional model is unreliable. Furthermore, if we want to fit the modulation in Fig. \[Fig4\]b with $\lambda$ = 1.4 $\mu$m and ${\tau_{\mathrm{s}}}$ = 25 ps, an exchange field of $\sim$ 60 mT would be required to match the data, even though the Hanle fitting did not include any ${\mathrm{B}_\mathrm{exch}}$. In return, the parameter sets that match 11% modulation do not fit the spin precession measurements unless the values are close to $\lambda$ = 700 nm, ${\tau_{\mathrm{s}}}$ = 14 ps and ${\mathrm{B}_\mathrm{exch}}$ = 85 mT. In conclusion, this analysis underlines the relevance to carry out both, angular modulation of ${\mathrm{R}_\mathrm{NL}}$ and Hanle precession experiments, to characterize the exchange field strength.
Bias dependent spin precession measurements in bl-hBN/bl-graphene/YIG
=====================================================================
In comparison to sample A, sample B is fabricated with a bilayer graphene flake. The extraction of the spin relaxation length via distance dependent spin valve measurements is done in a similar way as for sample A and is shown in the supplementary information in Fig. \[FigS4\]. We extract $\lambda = (2.3 \pm 1)\,\mu$m. The modulation of the non-local resistance by rotating the exchange field in the sample plane is shown in Fig. \[Fig6\]a. The parallel (red) and antiparallel (black) data is measured at 10 K and -366 mV DC bias. The solid line is the smoothed data and used to estimate the relative modulation of the spin signal after subtraction of the parallel and antiparallel data which results in a modulation of 8%.
![a) The non-local resistance can be modulated by 8% by rotating an in-plane magnetic field of 15 mT. The solid lines are smoothed and a guide to the eye. The red line is measured in parallel alignment, the black line in antiparallel configuration. b) Modeling of the 8% modulation with the spin transport parameters of $\lambda$ = 2.3 $\mu$m and $\tau_\mathrm{s}$ = 100 ps. The black curve represents the modulation by the applied magnetic field of 15 mT in the absence of an exchange field, the red curve adds an exchange field of 4 mT. c) The spin relaxation time $\tau_\mathrm{s}$ extracted from the Hanle data in panel d. d) The Hanle spin precession curves of sample B with the fitting curves (lines) for different DC bias currents. The spin relaxation length of $\lambda$ = 2.3 $\mu$m is used as parameters for the fitting. \[Fig6\]](Fig6.png){width="1\linewidth"}
To estimate the exchange field causing this precession, we use $\lambda$ = 2.3 $\mu$m extracted for sample B from the distance dependent measurements and assume $\tau_\mathrm{s}$ = 100 ps, which is later confirmed by the Hanle spin precession measurements. In this particular case, the modulation of the applied magnetic field of 15 mT (black line, Fig. \[Fig6\]d) already induces a modulation close to the experimentally found one. To match the data, a very small exchange field of only 4 mT would be required, leading us to the conclusion that in this device most likely no exchange interaction is present.
Using the Hanle spin precession data, we also extract $\lambda$ = 2.3 $\mu$m with a negligible exchange field. We find consistently over all biases a spin diffusion time of (100 $\pm$ 8) ps and a spin diffusion coefficient of ${\mathrm{D}_{\mathrm{s}}}= \lambda^2/\tau_\mathrm{s}$ = (530 $\pm$ 40) cm$^2$/s, which resembles the values used for the modulation fit and indicates consistency throughout our analysis of the spin transport. The possible absence of the exchange field in sample B stresses the importance of the graphene/YIG interface of these devices. This observation could be also explained with a different proximity effect on each of the two bilayer graphene layers. Nevertheless, sample B shows a similar dependence on the applied DC bias as sample A and shows that the tunable spin injection is also present in the bl-hBN/bl-graphene/YIG system.
Conclusion
==========
We have studied the spin injection through bl-hBN tunnel barriers into single- and bilayer graphene on YIG, showing a more reliable and efficient spin injection compared to TiO$_\mathrm{x}$ tunnel barriers. The bl-hBN tunnel barriers yield a resistance-area product between 5 and 30 k${\Omega}\mu$m$^2$ and the spin injection polarization is found to be tunable through a DC bias current applied to the injector. We observe a sign inversion at approximately -80 mV DC bias applied across the bl-hBN flake. We estimate the proximity induced exchange field through in-plane and out-of-plane spin precession measurements to be around 85 mT in sample A and likely to be absent in sample B. The low magnitude of the exchange field compared to theoretical predictions emphasizes the importance of the graphene/YIG interface on the proximity induced exchange field and confirms our previously reported low exchange strength for graphene/YIG devices. Nevertheless, our results confirm the unique properties of bl-hBN for the reliable spin injection into single and bilayer graphene on YIG and stress the importance of this type of tunnel barrier for future application in graphene spintronics.
Acknowledgements
================
We acknowledge the fruitful discussions with J. Ingla-Aynés, and funding from the European Union’s Horizon 2020 research and innovation program under grant agreement No 696656 and 785219 (‘Graphene Flagship’ core 1 and 2), the Marie Curie initial training network ‘Spinograph’ (grant agreement No 607904) and the Spinoza Prize awarded to B.J. van Wees by the ‘Netherlands Organization for Scientific Research’ (NWO).
**Supplementary Information**
Full set of the hBN tunnel barrier characterization
===================================================
![Full set of the contact characterization of sample A. The inset shows the microscope image with the characterized contacts. All contacts with a bilayer hBN tunnel barrier have a relatively homogeneous resistance-area product. Given the significantly higher resistance of contact 10, we suppose that this contact has a trilayer hBN tunnel barrier. \[FigS1\]](Fig_RA_SampleA_SI.png){width="0.7\linewidth"}
![Extended measurements of the contacts on sample B. The inset shows the microscope image with the characterized contacts. Contact 5 shows a linear metallic behavior, due to the shape of the hBN flake the cobalt is likely in direct contact with the graphene flake. \[FigS2\]](Fig_RA_SampleB_SI.png){width="0.7\linewidth"}
![Extended measurements of DC bias sweeps on sample A. See inset of Fig. \[FigS1\] for the contact numbering. The data is obtained by aligning the injector I and the detector D parallel and antiparallel and subtracting both curves. Since the detection polarization remains constant over the applied bias range, the increase of the non-local resistance corresponds to the increase of the spin injection polarization, which is relatively homogeneous over the contacts. The first two curves are discussed in the main text. \[FigS3\]](FigS3.png){width="0.7\linewidth"}
Estimation of the spin relaxation length in sample B
====================================================
![Distance dependent measurements of the spin valves on sample B. The large difference in the magnitude of the spin signal indicates an inhomogeneous spin polarization of the contacts and could be caused by cracks in the bl-hBN flake. See the inset of Fig. \[FigS2\] for the contact numbering. \[FigS4\]](FigS4.png){width="0.7\linewidth"}
Origin of the background of the Hanle curves in sample B
========================================================
The data shown in the main text in Fig. \[Fig6\] contains only the pure spin signal between injecting and detecting electrode. The spin signal is obtained by aligning the injector and detector parallel and antiparallel and subtracting both curves. The remaining signal is in theory the purely spin dependent signal. Spurious effects that are present in the measured signal are hereby extracted. These effects can be obtained by calculating the background signal by adding the parallel and antiparallel Hanle curves.
In Fig. \[FigS5\]a we show the measured Hanle curves, the extracted spin signal in Fig. \[FigS5\]b and the extracted background signal in Fig. \[FigS5\]c. Both spin and background signal show a dependence on the applied DC bias. The presence of a spin related signal in the background signal is not expected, however, the dependence on the DC bias suggests the opposite case.
![a) The raw data of the Hanle measurements on sample B has a significant background signal that is excluded from b) the spin signal. The dependence of the background signal on the applied DC bias shown in panel c). The background signal is extracted by adding the antiparallel to the parallel Hanle curve. d) To separate the spin and charge dependent contributions to the background signal, we subtract the data measured with the minimized spin signal (0 $\mu$A DC bias) from the individual Hanle background curves and extract the shown background signal. e) The amplitude of the Hanle background signal shows a dependence on the DC bias that roughly resembles the inverted dependence of the injector and detector electrode, which could indicate that the background signal has still a spin related contribution coming from one of the reference contacts. \[FigS5\]](FigS5.png){width="0.7\linewidth"}
To determine the nature of the signal, we normalize the data set to the signal where the spin signal and the spin injection polarization is minimized, which is here the case for a DC bias of 0 $\mu$A (Fig. \[FigS5\]b). This way we can separate the charge and spin dependent signals in the background data that do not depend on the magnetization of the inner detector and injector electrodes. The resulting signal is shown in Fig.S5c. We find a clear dependence on the applied DC bias. We suspect this signal to arise either as contribution from the current reference electrode or as the rotation of the cobalt electrodes at high magnetic fields out of the sample plane.
If we compare the signal amplitude averaged at $\pm$700 mT (\[${\mathrm{R}_\mathrm{NL}}$(+700 mT)+${\mathrm{R}_\mathrm{NL}}$(-700 mT)\]/2), we find a dependence on the DC bias as shown by the red squares in Fig. \[FigS5\]e. This slope approximately resembles that of the DC bias measurements but of opposite sign, which suggests that this signal might be actually spin related. Since the inner injector and detector signals are excluded from this data, we can identify the injector reference contact to be likely the origin. This contact is also biased with the DC current but does not have an hBN tunnel barrier. Therefore, the observation of such large signal is still surprising, especially for of the greater distance of the reference electrode to the detector of 4 $\mu$m instead of 1.9 $\mu$m. At this moment, we are unable to determine the origin of the DC bias dependence of the background signal. Further work is needed for clarification.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The knowledge of quantum phase flow induced under the Weyl’s association rule by the evolution of Heisenberg operators of canonical coordinates and momenta allows to find the evolution of symbols of generic Heisenberg operators. The quantum phase flow curves obey the quantum Hamilton’s equations and play the role of characteristics. At any fixed level of accuracy of semiclassical expansion, quantum characteristics can be constructed by solving a coupled system of first-order ordinary differential equations for quantum trajectories and generalized Jacobi fields. Classical and quantum constraint systems are discussed. The phase-space analytic geometry based on the star-product operation can hardly be visualized. The statement “quantum trajectory belongs to a constraint submanifold” can be changed e.g. to the opposite by a unitary transformation. Some of relations among quantum objects in phase space are, however, left invariant by unitary transformations and support partly geometric relations of belonging and intersection. Quantum phase flow satisfies the star-composition law and preserves hamiltonian and constraint star-functions.'
author:
- 'M. I. Krivoruchenko'
- Amand Faessler
title: |
Weyl’s symbols of Heisenberg operators\
of canonical coordinates and momenta as quantum characteristics
---
Introduction
============
The star-product operation introduced by Groenewold for phase-space functions [@GROE] permits formulation of quantum mechanics in phase space. It uses the Weyl’s association rule [@WEYL1; @WEYL2] to establish one-to-one correspondence between phase-space functions and operators in the Hilbert space. The Wigner function [@WIGNER] appears as the Weyl’s symbol of the density matrix. The skew-symmetric part of the star-product, known as the Moyal bracket [@MOYAL; @BARLE], governs the evolution of symbols of Heisenberg operators. Refined formulation of the Weyl’s association rule is proposed by Stratonovich [@STRA]. The Weyl’s association rule, star-product technique, star-functions, and some applications are reviewed in Refs. [@VOROS; @BAYEN; @CARRU; @BALAZ; @FEDO; @HILL; @KRF; @MIKR].
A one-parameter group of unitary transformations in the Hilbert space $$\mathfrak{U} = \exp(-\frac{i}{\hbar} \mathfrak{H}\tau),
\label{U7}$$ with $\mathfrak{H}$ being Hamiltonian, corresponds to a one-parameter group of canonical transformations in the classical theory [@DIRAC; @SOWUN; @WEYL2], although canonical transformations provide a broader framework [@BLEAF2; @ANDER].
Weyl’s symbols of time dependent Heisenberg operators of canonical coordinates and momenta induce quantum phase flow. Osborn and Molzahn [@OSBOR] construct quantum Hamilton’s equations which determine quantum phase flow and analyze the semiclassical expansion for unconstrained quantum-mechanical systems. An earlier attempt to approach these problems is undertaken in Ref. [@BLEAF1].
The infinitesimal transformations induced by the evolution operator (\[U7\]) in phase space coincide with the infinitesimal canonical transformations induced by the corresponding Hamiltonian function [@DIRAC; @SOWUN; @WEYL2]. The quantum and classical finite transformations are, however, distinct in general, since the star- and dot-products [^1] as multiplication operations of group elements in quantum and classical theories do not coincide. The quantum phase flow curves are distinct from the classical phase-space trajectories. This fact is not well understood (see e.g. Refs. [@BLEAF1; @TCURT]).
Osborn and Molzahn [@OSBOR] made important observation that quantum trajectories in unconstrained systems can be viewed as a “basis” to represent the evolution of quantum observables.
Such a property is usually assigned to characteristics appearing in a standard technique for solving first-order partial differential equations (PDE). The well known example is the classical Liouville equation $$\frac{\partial}{\partial \tau} f(\xi,\tau) = \{f(\xi,\tau),\mathcal{H}(\xi)\}.
\label{class}$$ This equation is solved in terms of characteristic lines which are solutions of classical Hamilton’s equations $$\frac{\partial}{\partial \tau} c^{i}(\xi,\tau) = \{\zeta^{i}, \mathcal{H}(\zeta)\}|_{\zeta = c(\xi,\tau)}
\label{classham3}$$ with initial conditions $c^{i}(\xi,0) = \xi^{i}$. Equations (\[classham3\]) are characteristic equations. They represent a system of first-order ordinary differential equations (ODE) for canonical variables. Physical observables $f(\xi,\tau)$ evolve according to $$f(\xi,\tau) = f(c(\xi ,\tau ),0). \label{charac}$$
It is remarkable that despite quantum Liouville equation is an infinite-order PDE its solutions are expressed in terms of solutions of the quantum Hamilton’s equations which are infinite-order PDE also.
A technical advantage in using the method of characteristics in quantum mechanics stems from the fact that to any fixed order of the semiclassical expansion the quantum Hamilton’s equations can be viewed as a coupled system of first-order ODE for quantum trajectories and generalized Jacobi fields obeying certain initial conditions. The evolution can be considered, respectively, as going along a trajectory in an extended phase space endowed with auxiliary degrees of freedom ascribed to generalized Jacobi fields. The evolution problem can be solved e.g. numerically applying efficient ODE integrators.
Quantum characteristics can be useful, in particular, for solving numerically many-body potential scattering problems by semiclassical expansion of star-functions around their classical values with subsequent integration over the initial-state Wigner function. Among possible applications are transport models in quantum chemistry and heavy-ion collisions [@AICHE; @TUEB1; @TUEB2] where particle trajectories remain striking but an intuitive feature.
A covariant extensions of quantum molecular dynamics (QMD) transport models [@SORGE; @MARIA] is based on the Poincaré invariant constrained Hamiltonian dynamics [@DIRAC2].
We show, in particular, that quantum trajectories exist and make physical sense in the constraint quantum systems also and play an important role similar to that in the quantum unconstrained systems.
The paper is organized as follows: In Sects. II and III, characteristics of unconstraint classical and quantum systems are discussed. Sects. IV and V are devoted to properties of characteristics of constraint classical and quantum systems. Quantum phase flows are analyzed using the star-product technique which we believe to be the most adequate tool for studying the subject.
We give definitions and recall basic features of the method of characteristics in Sect. II.
In Sect. III, fundamental properties of quantum characteristics are derived. The Weyl’s association rule, the star-product technique, and the star-functions are reviewed based on the method proposed by Stratonovich [@STRA]. We show, firstly, that quantum phase flow preserves the Moyal bracket and does not preserve the Poisson bracket in general. Secondly, we show that the star-product is invariant with respect to transformations of the coordinate system, which preserve the Moyal bracket. Thirdly, non-local laws of composition for quantum trajectories and the energy conservation along quantum trajectories are found in Sect. III-D. Applying the invariance of the star-product with respect to change of the coordinate system (\[BRINVA\]) and the energy conservation, we derive new equivalent representations of the quantum Hamilton’s equations Eq.(\[QF2\]) - (\[QF4\]). In Sect. III-E, we derive using the star-product technique the semiclassical reduction of the quantum Hamilton’s equations to a system of first-order ODE involving along with quantum trajectories their partial derivatives with respect to initial canonical variables. Finally, we express the phase-space Green function [@BLEAF1; @MARI] in terms of quantum characteristics and reformulate relation between quantum and classical time-dependent observables [@BRAU] using the method of characteristics.
The possibility of finding quantum trajectories and generalized Jacobi fields by solving a system of ODE gives practical advantages because of the existence of efficient numerical ODE integrators. It would be tempting to extend method of characteristics to constraint systems such as gauge theories, relativistic QMD transport models, etc.
The skew-gradient projection method is found to be useful to formulate classical and quantum constraint dynamics [@NAKA84; @NAKA93; @NAKA01; @KRF; @KRFF; @MIKR]. In Sect. IV, we show that in classical constraint systems characteristic lines exist and the method of characteristics is efficient. The proof we provide does not presuppose that constraint equations can be solved. The phase flow is commutative with the phase flows generated by constraint functions. Characteristic lines, if belong to the constraint submanifold at $\tau = 0$, belong to the constraint submanifold at $\tau > 0$ also.
Sect. V gives description of quantum characteristics in constraint systems. Although the formalism is complete, we encounter unexpected difficulty to formulate simple geometric idea that quantum trajectory belongs to a constraint submanifold. Using tools of the analytic geometry, any idea like that requires the use of composition of functions. In quantum mechanics, one has to use the star-composition. This calls for a modification of usual geometric relations “belong”, “intersect”, and others. In a specific quantum-mechanical sense, the Hamiltonian and constraint functions can be said to remain constant along quantum trajectories, while in the usual geometric sense they obviously don’t. The problem of visualization of relations among quantum objects in phase space is discussed in Sects. III-D and V-B.
Conclusion summarizes results.
Characteristics in classical unconstrained systems
==================================================
The phase space of system with $n$ degrees of freedom is parameterized by $2n$ canonical coordinates and momenta $\xi^{i}=(q^{1},...,q^{n},p_{1},...,p_{n})$ which satisfy the Poisson bracket relations $$\{\xi^{k},\xi^{l} \} = - I^{kl}$$ with $$\left\| I \right\| =\left\|
\begin{array}{ll}
0 & -E_{n} \\
E_{n} & 0
\end{array}
\right\|,$$ where $E_{n}$ is the identity $n\times n$ matrix. The phase space appears as the cotangent bundle $T_{*}\mathbb{R}^{n}$ of $n$-dimensional configuration space $\mathbb{R}^{n}$. The matrix $E_{n}$ imparts to $T_{*}\mathbb{R}^{n} = \mathbb{R}^{2n}$ a skew-symmetric bilinear form. The phase space acquires thereby structure of symplectic space.
In what follows, physical observables are time dependent, whereas density distributions remain constant. Such a picture constitutes the classical analogue of the quantum-mechanical Heisenberg picture.
In the classical unconstrained systems, phase flow: $\xi \to \zeta = c(\xi,\tau)$, is canonical and preserves the Poisson bracket. The classical Hamilton’s equations (\[classham3\]) are first-order ODE. The energy is conserved along classical trajectories $$\mathcal{H}(\xi) = \mathcal{H}(c(\xi,\tau)). \label{clco}$$ The classical Hamilton’s equations (\[classham3\]) can be rewritten as first-order PDE: $$\begin{aligned}
\frac{\partial}{\partial \tau} c^{i}(\xi,\tau) &=& \{c^{i}(\xi,\tau),\mathcal{H}(\xi)\} \label{classham2} \\
&=& \{c^{i}(\xi,\tau), \mathcal{H}(c(\xi,\tau))\}. \label{classham}\end{aligned}$$
The phase-space trajectories can be used to solve the Liouville equation (\[class\]) which is the first-order PDE. Any observable $f(\xi,\tau)$ is expressed in terms of $c(\xi ,\tau )$, as indicated in Eq.(\[charac\]).
Classical trajectories obey the dot-composition law: $$c^{i}(\xi ,\tau_1 + \tau_2 ) = c^{i}(c(\xi ,\tau_1 ),\tau_2).
\label{compcl}$$
Characteristics in quantum unconstrained systems
================================================
The Stratonovich version of the Weyl’s quantization and dequantization [@STRA] is discussed in the next subsection and in more details in Refs. [@BALAZ; @GRAC-1; @GRAC-2; @KRF; @MIKR].
Weyl’s association rule and the star-product
--------------------------------------------
The phase-space variables $\xi^{i}$ correspond to operators $\mathfrak{x}^{k} = (\mathfrak{q}^{1},...,\mathfrak{q}^{n},\mathfrak{p}_{1},...,\mathfrak{p}_{n})$ acting in the Hilbert space, which obey commutation rules $$[ \mathfrak{x}^{k},\mathfrak{x}^{l} ] = -i\hbar I^{kl}.
\label{987897}$$
Operators $\mathfrak{f}$ acting in the Hilbert space admit multiplications by $c$-numbers and summations. The set of all operators constitutes a vector space. The basis of such a space can be labelled by $\xi^{i}$. The Weyl’s basis looks like $$\mathfrak{B}(\xi )= \int \frac{d^{2n}\eta}{(2\pi \hbar )^{n} }
\exp (-\frac{i}{\hbar }\eta _{k}(\xi - \mathfrak{x})^{k}).$$ The association rule for a function $f(\xi )$ and an operator $\mathfrak{f}$ has the form [@STRA] $$f(\xi ) = Tr[\mathfrak{B}(\xi )\mathfrak{f}], \;\;\; \mathfrak{f} = \int \frac{d^{2n}\xi }{(2\pi \hbar )^{n}}f(\xi )\mathfrak{B}(\xi ).
\label{INV}$$ The value of $f(\xi)$ can be treated as the $\xi$-coordinate of $\mathfrak{f}$ in the basis $\mathfrak{B}(\xi )$, while $Tr[\mathfrak{B}(\xi )\mathfrak{f}]$ as the scalar product of $\mathfrak{B}(\xi )$ and $\mathfrak{f}$.
Using Eqs.(\[INV\]) one gets an equivalent association rule $$\mathfrak{f} = f(-i\hbar \frac{\partial}{\partial \eta_k}) \exp(\frac{i}{\hbar}
\eta_k \mathfrak{x}^k)|_{\eta = 0}.
\label{ASSC}$$ The half-Fourier transform, $$f(\xi ) = \int d^{n}\theta \exp \left(-\frac{i}{\hbar}\sum_{a=1}^{n} \theta^{a} p_{a}\right)
<q + \frac{\theta }{2}|\mathfrak{f}|q - \frac{\theta }{2}>,
\label{WEYL2}$$ provides the inverse relation. The Weyl-symmetrized functions of operators of canonical variables have representation [@KAMA91] $$\mathfrak{f} = f(\frac{ \mathfrak{q}_{(1)}^i + \mathfrak{q}_{(3)}^i}{2}, \mathfrak{p}_{(2)}^i),
\label{WEYL3}$$ where the subscripts indicate the order in which the operators act on the right.
Given two functions $f(\xi ) = Tr[\mathfrak{B}(\xi )\mathfrak{f}]$ and $g(\xi ) =Tr[\mathfrak{B}(\xi )\mathfrak{g}]$, one can construct a third function, $$f(\xi )\star g(\xi )=Tr[\mathfrak{B}(\xi )\mathfrak{fg}],$$ called star-product. In terms of the Poisson operator $$\mathcal{P} = - {I}^{kl}\overleftarrow{\frac{\partial }{\partial \xi ^{k}}}\overrightarrow{\frac{\partial }{\partial \xi^{l}}},
\label{POIS}$$ one has $$f(\xi )\star g(\xi )=f(\xi )\exp (\frac{i\hbar }{2}\mathcal{P})g(\xi ).$$ The star-product splits into symmetric and skew-symmetric parts, $$f\star g = f\circ g+\frac{i\hbar}{2} f\wedge g.$$ The skew-symmetric part is known under the name of Moyal bracket.
The Wigner function is the Weyl’s symbol of the density matrix. In the Heisenberg picture, the Wigner function remains constant $W(\xi ,\tau ) =W(\xi ,0 )$, whereas functions representing physical observables evolve with time in agreement with equation $$\begin{aligned}
\frac{\partial }{\partial \tau }f(\xi ,\tau ) = f(\xi ,\tau )\wedge H(\xi ). \label{EVOL} \end{aligned}$$ This equation is the Weyl’s transform of equation of motion for operators in the Heisenberg picture. It is the infinite-order partial differential equation (PDE).
The series expansions of $f(\xi ,\tau )$ over $\tau $ is given by $$f(\xi ,\tau ) = \sum_{s=0}^{\infty }\frac{\tau ^{s}}{s!}\underbrace{(...((}_{s}f(\xi
)\wedge H(\xi ))\wedge H(\xi ))\wedge ...H(\xi )), \label{EXPA}$$ where $f(\xi)=f(\xi ,\tau = 0 )$ is the initial data function.
Using the $\star$-adjoint notations of Ref. [@HAKI], equation (\[EVOL\]) can be represented in the form $$\begin{aligned}
\frac{\partial }{\partial \tau }f(\xi , \tau ) = - Ad_{\star} H[f(\eta ,\tau )](\xi ). \label{HAKI} \end{aligned}$$ Its formal solution, $$\begin{aligned}
f(\xi , \tau ) = \exp( - \tau Ad_{\star} H) f(\xi ,0 ), \label{HAKI2} \end{aligned}$$ is equivalent to Eq.(\[EXPA\]). If the target symbol $f(\xi ,0 )$ is semiclassically admissible, the evolution operator has asymptotic expansion [@OSBOR] $$\begin{aligned}
\exp( - \tau Ad_{\star} H) = \sum_{s=0}^{\infty} \hbar^{2s} \gamma^{(2s)}(\tau). \label{HAKI3} \end{aligned}$$
The power series expansion in $\hbar$ is valid for semiclassically admissible symbols $H$ and $f$. If, however, $f$ is a rapidly oscillating symbol, then (\[HAKI3\]) fails and the solution of the evolution equation becomes of the WKB type whose exponential phase is a symplectic area (see for details Ref. [@OSCO]).
Quantum phase flow preserves the Moyal bracket
----------------------------------------------
Active transformations modify operators $\mathfrak{f}$ and commute with $\mathfrak{B}(\xi)$. Passive transformations change the basis and keep operators fixed. These views are equivalent. We choose the former. Consider transformations depicted by the diagram $$\begin{aligned}
\xi &\stackrel{u}\longrightarrow& \acute{\xi} \nonumber \\
\updownarrow && \updownarrow \nonumber \\
\mathfrak{x} &\stackrel{\mathfrak{U}}\longrightarrow& \acute{\mathfrak{x}} \nonumber\end{aligned}$$ where $\mathfrak{U}$ is given by Eq.(\[U7\]).
The operators of canonical variables are transformed as $
\mathfrak{x}^{i} \rightarrow \acute{\mathfrak{x}}^{i}=\mathfrak{U}^{+}\mathfrak{x}^{i}\mathfrak{U}.
$ The coordinates $\acute{\xi}^{i}$ of new operators $\acute{\mathfrak{x}}^{i}$ in the old basis $\mathfrak{B}(\xi)$ are given by $$\xi^{i} \rightarrow \acute{\xi}^{i} = u^{i}(\xi,\tau) = Tr[\mathfrak{B}(\xi ) \mathfrak{U}^{+} \mathfrak{x}^{i} \mathfrak{U}].
\label{UXIT}$$ Since $\mathfrak{U}$ is the evolution operator, functions $u^{i}(\xi,\tau)$ can be treated as the Weyl’s symbols of operators of canonical coordinates and momenta in the Heisenberg picture. For $\tau = 0$, we have $u^{i}(\xi ,0 ) = \xi^{i}. $
The set of operators of canonical variables is complete in the sense that any operator acting in the Hilbert space can be represented as a function of operators $\mathfrak{x}^{i}$. One can indicate it as follows: $\mathfrak{f} = f(\mathfrak{x})$. The Taylor expansion of $f(\mathfrak{x})$ permits the equivalent formulation of the Weyl’s association rule. Transformations $\mathfrak{f} \rightarrow \acute{\mathfrak{f}}=\mathfrak{U}^{+}\mathfrak{f}\mathfrak{U}$ generate transformations of the associated phase-space functions: $$\begin{aligned}
f(\xi ) \rightarrow \acute{f}(\xi ) &=& f(\xi,\tau) = Tr[\mathfrak{B}(\xi )\mathfrak{U^{+}fU}] \nonumber \\
&=&\sum_{s=0}^{\infty }\frac{1}{s!}\frac{\partial ^{s}f(0)}{\partial \xi^{i_{1}}...\partial \xi ^{i_{s}}}
Tr[\mathfrak{B}(\xi )\mathfrak{U}^{+}\mathfrak{x}^{i_{1}}...\mathfrak{x}^{i_{s}}\mathfrak{U}] \nonumber \\
&=&\sum_{s=0}^{\infty }\frac{1}{s!}\frac{\partial ^{s}f(0)}{\partial \xi
^{i_{1}}...\partial \xi ^{i_{s}}}Tr[\mathfrak{B}(\xi )\acute{\mathfrak{x}}^{i_{1}} ... \acute{\mathfrak{x}}^{i_{s}}] \nonumber \\
&=&\sum_{s=0}^{\infty }\frac{1}{s!}\frac{\partial ^{s}f(0)}{\partial \xi
^{i_{1}}...\partial \xi ^{i_{s}}}u^{i_{1}}(\xi ,\tau)\star ...\star u^{i_{s}}(\xi,\tau) \nonumber \\
&=&\sum_{s=0}^{\infty }\frac{1}{s!}\frac{\partial ^{s}f(0)}{\partial \xi
^{i_{1}}...\partial \xi ^{i_{s}}}u^{i_{1}}(\xi ,\tau)\circ ...\circ u^{i_{s}}(\xi,\tau) \nonumber \\
&\equiv &f(\star u(\xi,\tau)). \label{TF}\end{aligned}$$ Last two lines define the star-composition. The star-function $f(\star u(\xi,\tau))$ is a functional of $u(\xi,\tau)$. The $\circ$-product is not associative in general. However, the indices $i_{s}$ for $s=1,...,2n$ are symmetrized, so the order in which the $\circ$-product is calculated is not important.
The antisymmetrized products $\mathfrak{x}^{[i_{1}}...\mathfrak{x}^{i_{2s}]}$ of even number of operators of canonical variables are $c$-numbers as a consequence of the commutation relations. These products are left invariant by unitary transformations: $\mathfrak{U}^{+}\mathfrak{x}^{[i_{1}}...\mathfrak{x}^{i_{2s}]}\mathfrak{U}=\mathfrak{x}^{[i_{1}}...\mathfrak{x}^{i_{2s}]}$. In phase space, we get $u^{[i_{1}}(\xi,\tau)\star ...\star u^{i_{2s}]}(\xi,\tau) = \xi ^{[i_{1}}\star
...\star \xi ^{i_{2s}]}
$ and, in particular, $$u^{i}(\xi,\tau)\wedge u^{j}(\xi,\tau)=\xi ^{i}\wedge \xi ^{j}=- {I}^{ij}.
\label{AREA}$$
Phase-space transformations induced by $\mathfrak{U}$ preserve the Moyal bracket and do not preserve the Poisson bracket, so the evolution map $\xi \rightarrow \acute{\xi} = u(\xi,\tau)$, is not canonical. Using Eq.(\[TEU\]), one can check e.g. that for $H(\xi) = (\delta_{ij}\xi^{i}\xi^{j})^2$ where $\delta_{ij}$ is the Kronecker symbol functions $u^{i}(\xi,\epsilon)$ do not satisfy the Poisson bracket condition for canonicity to order $O(\epsilon^2 \hbar^2)$.
For real functions $u^{i}(\xi,\tau)$ satisfying Eqs.(\[AREA\]) one may associate Hermitian operators $\acute{\mathfrak{x}}^{i}$ which obey commutation rules for operators of canonical coordinates and momenta. As a result, functions $u^{i}(\xi,\tau)$ appear in the coincidence with a unitary transformation relating $\mathfrak{x}^{i}$ and $\acute{\mathfrak{x}}^{i}$. The conservation of the Moyal bracket for a one-parameter set of continuous phase-space transformations is the necessary and sufficient condition for unitary character of the associated continuous transformations in the Hilbert space.
Change of variables which leaves the star-product invariant
-----------------------------------------------------------
Applying Eq.(\[TF\]) to product $\mathfrak{fg}$ of two operators, we obtain function $f(\zeta) \star g(\zeta)|_{\zeta = \star u(\xi,\tau)}$ associated to operator $\mathfrak{U}^{+}(\mathfrak{fg})\mathfrak{U}$ and function $f(\star u(\xi ,\tau)) \star g(\star u(\xi ,\tau))$ associated to operator $(\mathfrak{U}^{+}\mathfrak{f}\mathfrak{U})(\mathfrak{U}^{+}\mathfrak{g}\mathfrak{U})$. These operators coincide, so do their symbols: $$f(\zeta) \star g(\zeta)|_{\zeta = \star u(\xi,\tau)} = f(\star u(\xi ,\tau)) \star g(\star u(\xi ,\tau)).
\label{BRINVA}$$ The star-product is calculated with respect to $\zeta$ and $\xi$ in the left- and right-hand sides, respectively. Equation (\[BRINVA\]) is valid separately for symmetric and skew-symmetric parts of the star-product of the functions.
The substantial content of Eq.(\[BRINVA\]) is that one can compute the star-product in the initial coordinate system and change variables $\xi \rightarrow \zeta = \star u(\xi,\tau)$, or equivalently, change variables $\xi \rightarrow \zeta = \star u(\xi,\tau)$ and compute the star-product, provided Eq.(\[AREA\]) is fulfilled.
The functions $u^{i}(\xi,\tau )$ define quantum phase flow which represents quantum deformation of classical phase flow.
![Schematic presentation of the star-composition law (\[comp\]). The solid line stands for a quantum trajectory $u^{i}(\xi ,\tau) = u^{i}(\star u(\xi ,s),\tau - s)$ at $0<\tau <t$. The dashed line is assigned to a trajectory $u^{i}(u(\xi ,s),\tau - s)$ which we would have at $s<\tau <t$ for the classical dot-composition law. The distance between the solid and dashed trajectories is of order of $\hbar^2$. []{data-label="fig10"}](dia.eps){width="2.618"}
Composition law for quantum trajectories and energy conservation law
--------------------------------------------------------------------
In the usual geometric sense, quantum characteristics $u(\xi,\tau )$ cannot be considered as trajectories along which physical particles move. The reason lies, in particular, in the star-composition law $$u(\xi ,\tau_1 + \tau_2 ) = u(\star u(\xi ,\tau_1 ),\tau_2)
\label{comp}$$ which is distinct from $u(\xi ,\tau_1 + \tau_2 ) = u(u(\xi ,\tau_1 ),\tau_2)$, see Fig. \[fig10\]. In classical mechanics, the composition law has the form of Eq.(\[compcl\]).
The energy conservation in the course of quantum evolution implies $$H(\xi )=H(\star u(\xi ,\tau )) \label{EC}$$ where $H(\xi )=Tr[\mathfrak{B}(\xi )\mathfrak{H}]$ is Hamiltonian function. $H(\xi )$ is, however, not conserved along quantum trajectories in the usual geometric sense, so $H(\xi ) \neq H(u(\xi ,\tau ))$. In classical mechanics, the conservation law has the form (\[clco\]).
To express the idea that a point particle moves continuously along a phase-space trajectory, one has to use the star-composition (\[comp\]). The dot-composition is not defined in quantum mechanics.
Similarly, $H(u(\xi ,\tau ))$ does not make any quantum-mechanical sense. One has to work with $H(\star u(\xi ,\tau ))$. If so, the only way to express quantitatively the fact of the energy conservation along a phase-space trajectory is to use Eq.(\[EC\]).
The similar problem arises in constraint systems when we want to decide if quantum trajectories belong to a constraint submanifold.
The analytic geometry provides tools to formulate relations among geometric objects. Those relations which are expressed through composition of functions are modified. We discuss if possible to assign a geometric sense to formulas involving the star-composition in Sect. V-B.
Reduction of quantum Hamilton’s equations to a coupled system of ODE for quantum trajectories and generalized Jacobi fields
---------------------------------------------------------------------------------------------------------------------------
Quantum Hamilton’s equations can be obtained applying the Weyl’s transform to evolution equations for Heisenberg operators of canonical coordinates and momenta $$\begin{aligned}
\frac{\partial }{\partial \tau }u^{i}(\xi ,\tau ) &=& u^{i}(\xi ,\tau )\wedge H(\xi ) \label{QF} \\
&=& u^{i}(\xi ,\tau )\wedge H(\star u(\xi ,\tau )) \label{QF2} \\
&=& \zeta ^{i}\wedge H(\zeta )|_{\zeta =\star u(\xi ,\tau )} \label{QF3} \\
&=& \{\zeta ^{i},H(\zeta )\}|_{\zeta =\star u(\xi ,\tau )}. \label{QF4}\end{aligned}$$
To reach the step 2, the energy conservation (\[EC\]) is used. Going from (\[QF2\]) to (\[QF3\]), the change of variables (\[BRINVA\]) is performed. To achieve (\[QF4\]), we exploit $\xi^{i} \wedge f(\xi) = \{\xi^{i},f(\xi)\}$. The time derivative of $u^{i}(\xi ,\tau )$ can be computed classically using the Poisson bracket. The substitution $\zeta =\star u(\xi ,\tau )$ leads, however, to deformation of classical trajectories. Equations (\[QF3\]) and (\[QF4\]) are the quantum analogues of Eq.(\[classham3\]), Eq.(\[QF\]) is the quantum analogue of Eq.(\[classham2\]), and Eq.(\[QF2\]) is the quantum analogue of Eq.(\[classham\]).
As distinct from the de Broglie-Bohm trajectories (see e.g. [@DBB]), $u^{i}(\xi ,\tau )$ are not related to specific states in the Hilbert space.
The functional form of quantum Hamilton’s equations (\[QF\]) is left invariant by the change of variables $\xi \rightarrow \upsilon =\star v_{-}(\xi)$ provided the map $v_{-}$: $\upsilon = v_{-}(\xi)$, preserves the Moyal bracket.
Equations (\[QF\]) are not invariant under canonical transformations. Consider e.g. canonical map: $(q,p) \rightarrow (Q,P)$, with generating function $S_{2}(q,P) = qP + q^3 + qP^2$ such that $p = \partial S_{2}(q,P)/\partial q $ and $Q = \partial S_{2}(q,P)/\partial P$. One can compare $f \circ g$ and $f \wedge g$ in the coordinate systems $(q,p)$ and $(Q,P)$. For functions $f = q$ and $g = p$, one gets, respectively, $f\circ g|_{(q,p)} = qp \neq f\circ g|_{(Q,P)} = qp + 6\hbar^2 Q/(1 + 2P)^5 + O(\hbar^4)$ and $f \wedge g|_{(q,p)} = 1 \neq f \wedge g|_{(Q,P)} = 1 + 24 \hbar^2/(1 + 2P)^6 + O(\hbar^4) $. The symmetric and skew-symmetric parts of the star-product are both not invariant under canonical transformations. Coordinate systems in phase space if related by a canonical transformation provide non-equivalent quantum dynamics. This ambiguity is better known as the operator ordering problem.
The quantum deformation of classical phase flow can be found by expanding $$u^{i}(\xi,\tau) = \sum_{s=0}^{\infty}\hbar^{2s}u^{i}_{s}(\xi,\tau).$$ The right-hand side of Eqs.(\[QF\]) $F^{i}(\zeta) \equiv \{ \zeta^{i},H(\zeta)\}$ is a function of $\zeta = \star u(\xi,\tau)$ (i.e. functional of $u(\xi,\tau)$), so we have to expand $$F^{i}(\star u(\xi,\tau)) = \sum_{s=0}^{\infty}\hbar^{2s}F^{i}_{s}[u(\xi,\tau)]$$ using e.g. the cluster-graph method [@OSBOR; @GRACIA]. Classical trajectories $u^{i}_{0}(\xi,\tau)$ satisfy classical Hamilton’s equations $$\frac{\partial }{\partial \tau } u^{i}_{0} = F^{i}_{0}(u_{0})$$ and initial conditions $u^{i}_{0}(\xi,0) = \xi^{i}$. Given $u^{i}_{0}(\xi,\tau)$, the lowest-order quantum correction $u^{i}_{1}(\xi,\tau)$ can be found by solving first-order ordinary differential equations (ODE) $$\begin{aligned}
\frac{\partial }{\partial \tau }u^{i}_{1}
&=& u^{k}_{1} \frac{\partial F^{i}_{0}(u_{0})}{\partial u_{0}^{k}} \label{QHE} \\
&-& \frac{1}{16}I^{k_{1}l_{1}} I^{k_{2}l_{2}}
J^{i_1}_{0,k_{1}k_{2}}
J^{i_2}_{0,l_{1}l_{2}}
\frac{\partial^2 F^{i}_{0}(u_{0})}{\partial u^{i_1}_{0} \partial u^{i_2}_{0}}\nonumber \\
&-&\frac{1}{24} I^{k_{1}l_{1}} I^{k_{2}l_{2}}
J^{i_1}_{0,k_{1}}
J^{i_2}_{0,k_{2}}
J^{i_3}_{0,l_{1}l_{2}}
\frac{\partial^3 F^{i}_{0}(u_{0})}{\partial u^{i_1}_{0} \partial u^{i_2}_{0} \partial u^{i_3}_{0}} \nonumber\end{aligned}$$ with initial conditions $u^{i}_{1}(\xi,0) = 0$. The functions $J^{i}_{0,k}$ and $J^{i}_{0,kl}$ entering Eq.(\[QHE\]) is a particular case of generalized Jacobi fields $$J^{i}_{r,k_{1}...k_{t}}(\xi, \tau) = \frac{\partial u^{i}_{r}(\xi,\tau)}{\partial \xi^{k_{1}} ... \partial \xi^{k_{t}}}.
\label{GJ}$$ Given $u^{i}_{r}(\xi,\tau)$ and $J^{i}_{r,k_{1}...k_{t}}(\tau,\xi)$ for $0 \leq r \leq s$, the next corrections $u^{i}_{s + 1}(\xi,\tau)$ can be found from first-order ODE involving generalized Jacobi fields (\[GJ\]) with $0 \leq r \leq s$. For a harmonic oscillator, $u^{i}_{s}(\xi,\tau) = 0$ for $s \geq 1$, in which case quantum phase flow is both canonical and unitary.
The generalized Jacobi fields (\[GJ\]) satisfy ODE also. The lowest order equations have the form: $$\begin{aligned}
\frac{\partial }{\partial \tau } J^{i}_{0,k} &=&
\frac{\partial F^{i}_{0}(u_{0})}{\partial u_{0}^{m}} J^{m}_{0,k}, \label{GRADI1} \\
\frac{\partial }{\partial \tau } J^{i}_{0,kl} &=&
\frac{\partial^2 F^{i}_{0}(u_{0})}{\partial u_{0}^{m} \partial u_{0}^{n}} J^{m}_{0,k} J^{n}_{0,l} +
\frac{\partial F^{i}_{0}(u_{0})}{\partial u_{0}^{m}} J^{m}_{0,kl}. \nonumber\end{aligned}$$
The first of these equations describes the evolution of small perturbations along the classical trajectories. Being projected onto a submanifold of constant energy it becomes the Jacobi-Levi-Civita equation [@ARNO]. In stochastic systems, $J^{i}_{0,k}$ grow exponentially with time.
At any fixed level of accuracy of the semiclassical expansion, we have a coupled system of ODE for $u^{i}_{r}(\tau,\xi)$ and $J^{i}_{r,k_{1}...k_{t}}(\tau,\xi)$ subjected to initial conditions $$\begin{aligned}
u^{i}_{0}(\xi,0) = \xi^{i}, ~~~~&~& J^{i}_{0,k}(\xi,0) = \delta^{i}_{k}, \label{J1} \\
u^{i}_{r}(\xi,0) = 0, ~~~~~&~& J^{i}_{r,k_{1}...k_{t}}(\xi,0) = 0, \label{J2}\end{aligned}$$ where $r \ge 1$ and $r \ge 1$ or $t \ge 2$, respectively. The evolution problem can be solved e.g. numerically applying efficient ODE integrators.
A numerical computation of the semiclassical expansion of the quantum phase flow in the elastic scattering of atomic systems is performed in Ref. [@MCQUA].
An alternative approach allowing to reduce the semiclassical quantum dynamics to a closed system of ODE is proposed by Bagrov with co-workers [@BAGR3; @BAGR0; @BAGR1; @BAGR2]. The phase-space trajectories appearing in [@BAGR3; @BAGR0; @BAGR1; @BAGR2] are connected to specific quantum states like in the de Broglie - Bohm theory.
Properties of quantum paths, localization of quantum systems, and a coherent-type representation of the quantum flow are discussed in Ref. [@MIKA04].
The series expansions of $u^{i}(\xi ,\tau )$ and $f(\xi ,\tau ) = f(\star u(\xi ,\tau ))$ over $\tau $ are given by $$\begin{aligned}
&&u^{i}(\xi ,\tau ) = \label{TEU} \\
&&\sum_{s=0}^{\infty }\frac{\tau ^{s}}{s!}\underbrace{(...((}
_{s}\xi^{i} \wedge H(\xi ))\wedge H(\xi ))\wedge...H(\xi )), \nonumber \\
&&f(\star u(\xi ,\tau )) = \label{TEW} \\
&&\sum_{s=0}^{\infty }\frac{\tau ^{s}}{s!}\underbrace{(...((}_{s}f(\xi
)\wedge H(\xi ))\wedge H(\xi ))\wedge \nonumber ...H(\xi )). \nonumber\end{aligned}$$
In general, quantum phase flow is distinct from classical phase flow. This feature holds in integrable systems also, as discussed in Appendix A.
The lowest order operators $\gamma^{(s)}$ entering Eq.(\[HAKI3\]) can be found to be [@OSBOR; @MCQUA; @GRACIA; @KADP] $$\begin{aligned}
&&\gamma^{(0)}(\tau)f(\xi) = f(u_{0}(\xi,\tau)), \label{FINA2} \\
&&\gamma^{(2)}(\tau)f(\xi) = u_{1}^{i}(\xi,\tau) f(u_{0}(\xi,\tau))_{,i} \label{FINA3} \\
&& ~~~~- \frac{1}{16}J^{i}_{0,kl}(\xi,\tau) J^{j,kl}_0(\xi,\tau)f(u_{0}(\xi,\tau))_{,ij} \nonumber \\
&& ~~~~- \frac{1}{24}J^{i}_{0,l}(\xi,\tau)J^{j}_{0,m}(\xi,\tau)J^{k,lm}_{0}(\xi,\tau)f(u_{0}(\xi,\tau))_{,ijk}. \nonumber\end{aligned}$$ Here, the derivatives of $f(u_0(\xi,\tau))$ are calculated with respect to $u^{i}_0$: $$f(u_0(\xi,\tau))_{,i_{1}...i_{s}} = \frac{\partial^{s}f(u_0(\xi,\tau))}{\partial u^{i_{1}}_0...\partial u^{i_{s}}_0}.$$ The Jacobi fields with the upper indices are defined by $$J^{i,k_{1}...k_{t}}_{r}(\xi,\tau) = I^{k_1 j_1} ...I^{k_s j_s}J^{i}_{r,j_{1}...j_{t}}(\xi,\tau).
\label{GJUP}$$
According to Eq.(\[FINA2\]), time dependence of the zero order term $\gamma ^{(0)}(\tau )f(\xi )$ is determined by time dependence of the classical phase-space trajectory and form of the function $f(\xi )$. The similar conclusion holds for $\gamma ^{(2)}(\tau )f(\xi )$: Eq.(\[FINA3\]) tells that time dependence enters through the classical trajectory, the first quantum correction to the classical trajectory, and the classical Jacobi fields. The problem of convergence of a formal power series expansion is always a difficult subject. The convergence rate of the time series depends obviously on the system. Generally, the series expansion in $\tau$ has a finite convergence radius. However, in the itegrable systems one has a truncated series expansion (cf. Eq.(III.27) and Eq.(A.3)).
Green function in phase space and quantum characteristics
---------------------------------------------------------
Using orthogonality condition $$Tr[\mathfrak{B}(\xi)\mathfrak{B}(\zeta)] = (2\pi \hbar)^{n}\delta^{2n}(\xi - \zeta)$$ and Eq.(\[TF\]), we express Green function for the Weyl’s symbols [@BLEAF1; @MARI] in terms of the quantum characteristics: $$\begin{aligned}
D(\xi,\zeta,\tau) &=& Tr[\mathfrak{B}(\xi)\mathfrak{U}^{+} \mathfrak{B}(\zeta)\mathfrak{U}] \nonumber \\
&=& (2\pi \hbar)^{n}\delta^{2n}(\star u(\xi,\tau) - \zeta)
\nonumber \\
&=& (2\pi \hbar)^{n}\delta^{2n}(\xi - \star u(\zeta,-\tau)).
\label{GF}\end{aligned}$$
A compact operator relation between the classical and quantum time-dependent observables is established in Ref. [@BRAU]. Solutions of the quantum and classical Liouville equations, $f(\xi,\tau)$ and $f_{c}(\xi,\tau)$, with initial conditions $f(\xi,0)=f_{c}(\xi,0)$ are related through the product $DD_{c}^{-1}$ where $D_{c}$ is the classical Green function $$D_{c}(\xi,\zeta,\tau) = (2\pi \hbar)^{n}\delta^{2n}(c(\xi,\tau) - \zeta).
\label{2987423897}$$ In terms of the characteristics, we obtain $$f(\xi,\tau) = f_{c}(c(\star u(\xi,\tau),-\tau),\tau).
\label{987987}$$ It is assumed that classical and quantum hamiltonian functions coincide i.e. ${\mathcal H}(\xi) = H(\xi)$.
Given the Green function is known, the quantum trajectories can be found from equation $$u^{i}(\zeta,\tau) = \int \frac{d^{2n}\xi}{(2\pi \hbar)^{n}}\xi^{i}D(\xi,\zeta,-\tau).$$
For $\mathfrak{U} = 1 - \frac{i}{\hbar}\mathfrak{H}\epsilon$ where $\epsilon$ is an infinitesimal parameter, the associated transformations of canonical variables and phase-space functions are given by $
\delta \xi^{i}= \xi^{i} \wedge \epsilon H(\xi ) = \{\xi^{i} ,\epsilon H(\xi )\}$ and $
\delta f(\xi )= f(\xi ) \wedge \epsilon H(\xi )$. The transformations of canonical variables are canonical to order $O(\epsilon)$ only. The infinitesimal transformations of symbols of operators are not canonical. Any function $H(\xi)$ can be used to generate classical phase flow or quantum phase flow, according as the dot-product or the star-product stands for multiplication operation in the set of phase-space functions.
The analogue between unitary and canonical transformations is illustrated by Dirac [@SOWUN] in terms of the generating function $S(q^{\prime},q)$ defined by $\exp(\frac{i}{\hbar}S(q^{\prime},q)) = <q^{\prime}|\mathfrak{U}|q>$. The evolution map $(q,p) \rightarrow (q^{\prime},p^{\prime})$, is canonical for $p = - \partial S(q^{\prime},q)/\partial q$ and $p^{\prime} = \partial S(q^{\prime},q)/\partial q^{\prime}$. The parallelism of the transformations is manifest, but trajectories are complex. The generating function defined by the phase of $<q^{\prime}|\mathfrak{U}|q>$ yields real trajectories. It is not clear, however, if time-dependent symbols of operators are entirely determined by such trajectories.
The Weyl’s symbols of operators of canonical variables $ u^{i}(\xi,\tau)$ are the genuine characteristics in the sense that they allow by equation $f(\xi,\tau) = f(\star u(\xi,\tau),0)$ the entire determination of the evolution of observables. The quantum dynamics is totally contained in $ u^{i}(\xi,\tau)$, whereas the deformation of symbols of the operators calculated at $\star u(\xi,\tau)$ has a kinematical meaning.
Characteristics in classical constraint systems
===============================================
We give first description of second-class constraints systems and of the skew-gradient projection formalism. The details are found elsewhere [@NAKA84; @NAKA93; @NAKA01; @KRF; @KRFF; @MIKR].
Classical constraint systems in phase space
-------------------------------------------
Second-class constraints $\mathcal{G}_{a}(\xi) = 0$ with $a = 1,...,2m$ and $m<n$ have the Poisson bracket relations which form a non-degenerate $2m \times 2m$ matrix $$\det\{\mathcal{G}_{a}(\xi),\mathcal{G}_{b}(\xi)\} \ne 0. \label{NONGEN}$$ If this would not be the case, it could mean that gauge degrees of freedom appear in the system. After imposing gauge-fixing conditions, we could arrive at the inequality (\[NONGEN\]). Alternatively, breaking the condition (\[NONGEN\]) could mean that constraint functions are dependent. After removing redundant constraints, we arrive at the inequality (\[NONGEN\]).
Constraint functions are equivalent if they describe the same constraint submanifold. Within this class one can make transformations without changing dynamics.
For arbitrary point $\xi$ of the constraint submanifold $\Gamma^{*} = \{\xi: \mathcal{G}_{a}(\xi) = 0 \}$, there is a neighbourhood where one may find equivalent constraint functions in terms of which the Poisson bracket relations look like $$\{\mathcal{G}_{a}(\xi),\mathcal{G}_{b}(\xi)\}=\mathcal{I}_{ab} \label{SB}$$ where $$\mathcal{I}_{ab}=\left\|
\begin{array}{ll}
0 & E_{m} \\
-E_{m} & 0
\end{array}
\right\|. \label{SMAT}$$ Here, $E_{m}$ is the identity $m\times m$ matrix, $\mathcal{I}_{ab}\mathcal{I}_{bc}=-\delta _{ac}$. The matrix $\mathcal{I}^{ab} = - \mathcal{I}_{ab}$ is used to lift indices $a,b,\ldots$ up.
The basis (\[SB\]) always exists locally, i.e., in a finite neighbourhood of any point of the constraint submanifold. This is on the line with the Darboux’s theorem (see e.g. [@ARNO]). All symplectic spaces are locally indistinguishable.
![Schematic presentation of skew-gradient projection onto constraint submanifold along commuting phase flows generated by constraint functions.[]{data-label="fig1"}](project.eps){width="5.5"}
Skew-gradient projection formalism
----------------------------------
The concept of the skew-gradient projection $\xi _{s}(\xi )$ of canonical variables $\xi$ onto a constraint submanifold plays important role in the Moyal quantization of constraint systems. Geometrically, the skew-gradient projection acts along phase flows $Id \mathcal{G}^{a}(\xi)$ generated by constraint functions. These flows are commutative in virtue of Eqs.(\[SB\]): Using Eqs.(\[SB\]) and the Jacobi identity, one gets $\{\mathcal{G}^{a},\{\mathcal{G}^{b},f\}\} = \{\mathcal{G}^{b},\{\mathcal{G}^{a},f\}\}$ for any function $f$, so the intersection point with $\Gamma^{*}$ is unique.
To construct the skew-gradient projections, we start from equations $$\{\xi _{s}(\xi ),\mathcal{G}_{a}(\xi )\}=0 \label{CG}$$ which say that point $\xi _{s}(\xi ) \in \Gamma^{*}$ is left invariant by phase flows generated by $\mathcal{G}_{a}(\xi )$. Using the symplectic basis (\[SB\]) for the constraints and expanding $$\xi _{s}(\xi )=\xi +X^{a}\mathcal{G}_{a}+\frac{1}{2}X^{ab}\mathcal{G}_{a}\mathcal{G}_{b}+...$$ in the power series of $\mathcal{G}_{a}$, one gets $$\begin{aligned}
\xi _{s}(\xi )=\sum_{k=0}^{\infty }\frac{1}{k!}\{...\{\{\xi ,\mathcal{G}^{a_{1}}\},\mathcal{G}^{a_{2}}\},...\mathcal{G}^{a_{k}}\} \nonumber \\
\times \mathcal{G}_{a_{1}}\mathcal{G}_{a_{2}}...\mathcal{G}_{a_{k}}.
\label{SGRAD}\end{aligned}$$ Similar projection can be made for function $f(\xi )$: $$\begin{aligned}
f_{s}(\xi )=\sum_{k=0}^{\infty }\frac{1}{k!}\{...\{\{f(\xi ),\mathcal{G}^{a_{1}}\},\mathcal{G}^{a_{2}}\},...\mathcal{G}^{a_{k}}\} \nonumber \\
\times \mathcal{G}_{a_{1}}\mathcal{G}_{a_{2}}...\mathcal{G}_{a_{k}}.
\label{FSG}\end{aligned}$$ One has $$f_{s}(\xi )=f(\xi _{s}(\xi )). \label{FSSF}$$ The projected functions are in involution with the constraint functions: $$\{ f_{s}(\xi ), \mathcal{G}_{a}(\xi) \}=0. \label{involution}$$ Consequently, $f_{s}(\xi )$ does not vary along $Id\mathcal{G}_{a}(\xi)$, since $$\{ f(\xi ), g(\xi) \} \equiv \frac{\partial f(\xi)}{\partial \xi^{i}} (Idg(\xi))^{i}.$$
The skew-gradient projection is depicted schematically in Fig. \[fig1\].
Evolution and skew-gradient projection
--------------------------------------
In the classical second-class constraints systems, one has to start from constructing $\mathcal{H}_{s}(\xi )$ from $\mathcal{H}(\xi )$. The evolution equation for phase-space functions can be converted then to the classical Liouville equation: $$\frac{\partial }{\partial \tau }f(\xi ,\tau )=\{f(\xi ,\tau ),\mathcal{H}_{s}(\xi )\} \label{LIOUCON}$$
Similarly, the canonical variables obey the classical Hamilton’s equations: $$\frac{\partial }{\partial \tau }c^{i}(\xi ,\tau )=\{c^{i}(\xi ,\tau ),\mathcal{H}_{s}(\xi )\} \label{HAMICON}$$ with initial conditions $$c^{i}(\xi ,0)=\xi^{i}. \label{INICON}$$
Equation $$\{\mathcal{G}_{a}(\xi),\mathcal{H}_{s}(\xi )\} = 0
\label{INVO}$$ tells that $\mathcal{G}_{a}(\xi)$ remain constant along $c^{i}(\xi ,\tau )$: $$\mathcal{G}_{a}(\xi) = \mathcal{G}_{a}(c(\xi ,\tau )).
\label{INVO-2}$$
Equations (\[INVO-2\]) show that trajectories do not leave level sets $\{\xi :\mathcal{G}_{a}(\xi )=\mathrm{constant}\}$ and therefore do not leave the constraint submanifold $\Gamma ^{*}=\{\xi :\mathcal{G}_{a}(\xi )=0\}.$
Given $\mathcal{H}_{s}(\xi )$ is constructed, it becomes possible to extend standard theorems of the Hamiltonian formalism to second-class constraints systems without modifications. The novel element is the interplay between the evolution and the skew-gradient projection.
Let the coordinate system $\{\eta^{i}\}$ is obtained from the coordinate system $\{\xi^{i}\}$ by the canonical transformation $\xi \rightarrow \eta = c(\xi ,\tau )$.
![Classical phase flow $c(\xi,\tau)$ is commutative with classical projection $\xi_{s}(\xi)$ onto constraint submanifold $\Gamma^{*}$. []{data-label="fig4"}](cflow.eps){width="7.5"}
Eq.(\[FSG\]) may be applied for $c^{i}(\xi ,\tau )$. Using Eq.(\[INVO-2\]), we replace the arguments of the constraint functions to $c^{i}(\xi ,\tau )$ and replace everywhere $c^{i}(\xi ,\tau )$ with $\eta^{i}$, as long as the Poisson brackets are invariant and the constraint functions are scalars. We arrive at $$\begin{aligned}
c_{s}(\xi ,\tau ) &=&c(\xi _{s}(\xi ),\tau ) \nonumber \\
&=&\xi _{s}(c(\xi ,\tau )). \label{ABBA}\end{aligned}$$ The first line is a consequence of Eq.(\[FSSF\]). The evolution is commutative with the skew-gradient projection. Equation (\[ABBA\]) is illustrated on Fig. \[fig4\].
The Liouville equation can be solved provided phase-space trajectories $c(\xi ,\tau )$ are known. In general, $$f(\xi ,\tau )=f(c(\xi ,\tau ),0). \label{1234}$$ Applying projection (\[FSG\]), one gets $$\begin{aligned}
f_{s}(\xi ,\tau ) &=&f(c(\xi _{s}(\xi ),\tau ),0) \nonumber \\
&=&f(c_{s}(\xi ,\tau ),0). \label{23456}\end{aligned}$$ The first line follows from Eq.(\[FSSF\]). Equation (\[23456\]) shows how to use characteristics in order to solve evolution equations in the classical second-class constraint systems.
The evolution depends on choice of the constraint functions up to a canonical transformation. Suppose we found two sets of the constraint functions $\mathcal{G}_{a}(\xi )$ and $\tilde{\mathcal{G}}_{a}(\xi )$ describing the same constraint submanifold. Each set can be transformed to the standard basis (\[SB\]). Such bases are related by canonical transformations, so one can find a canonical map: $\xi \to \upsilon = v_{-}(\xi)$, such that $\tilde{\mathcal{G}}_{a}(\upsilon) = \mathcal{G}_{a}(v_{-}(\xi))$. The inverse transform is $\upsilon \to \xi = v_{+}(\upsilon)$. The skew-gradient projections $\xi_{s}(\xi)$ and $\upsilon_{s}(\upsilon)$ are related by: $$\upsilon_{s}(v_{-}(\xi)) = v_{-}(\xi_{s}(\xi)).
\label{ohoho}$$ The skew-gradient projection depends on choice of the constraint functions up to a canonical transformation. The same is true for projected Hamiltonian functions: $$\mathcal{H}_{s}(\xi) = \mathcal{H}_{s}^{\prime}(\upsilon)$$ where $\mathcal{H}^{\prime}(\upsilon) = \mathcal{H}(v_{+}(\upsilon))$. Two sets of the constraint functions $\mathcal{G}_{a}(\xi )$ and $\tilde{\mathcal{G}}_{a}(\xi )$ lead to the canonically equivalent Hamiltonian phase flows.
Characteristics in quantum constraint systems
=============================================
The constraint systems represent high interest since all fundamental interactions in the elementary particle physics are based on the principles of gauge invariance. Gauge fixing turns gauge-invariant systems into constraint systems.
In the classical mechanics, the constraint systems can be treated as a limiting case $\lambda \to \infty$ of systems in a potential $V_{\lambda}(q)$ which rapidly increases when the coordinates $q$ go away from the constraint submanifold. In the limit of $\lambda \to \infty$, $V_{\lambda}(q) = 0$ if $q$ belongs to the constraint submanifold and $V_{\lambda}(q) = + \infty$ when $q$ does not belong to the constraint submanifold. The classical systems obtained by imposing the constraints and by the limiting procedure have equivalent dynamic properties [@ARNO].
In the quantum mechanics, this is not the case. The limiting procedure applied to a particular system of Ref. [@COST] to model holonomic constraints, results to the quantum dynamics which depends on the way the constraint submanifold is embedded into the configuration space. From other hand, the quantization of constraint holonomic systems leads to the conclusion that the dynamics is determined by the induced metric tensor only [@KFRF; @KRFF]. The limiting procedure and imposing the constraints are not equivalent schemes of the quantization. In what follows, we discuss the constraint dynamics as it appears in the gauge theories.
The Groenewold-Moyal constraint dynamics has many features in common with the classical constraint dynamics. The projection formalism developed for constraint systems allows, from other hand, to treat unconstrained and constraint systems essentially on the same footing.
Skew-gradient projection in quantum mechanics
---------------------------------------------
We recall that classical Hamiltonian function $\mathcal{H}(\xi)$ and constraint functions $\mathcal{G}_{a}(\xi)$ are distinct in general from their quantum analogues ${H}(\xi)$ and ${G}_{a}(\xi)$. These dissimilarities are connected to ambiguities in quantization of classical systems. It is required only $$\lim_{\hbar \rightarrow 0}{H}(\xi) = \mathcal{H}(\xi), ~~~\lim_{\hbar \rightarrow 0}{G}_{a}(\xi) = \mathcal{G}_{a}(\xi). \nonumber$$ In what follows $$\Gamma^{*} = \{\xi: {G}_{a}(\xi) = 0 \}.
\label{COMA}$$
The quantum constraint functions ${G}_{a}(\xi)$ satisfy $${G}_{a}(\xi )\wedge {G}_{b}(\xi )=\mathcal{I}_{ab}. \label{SBAS}$$
The quantum-mechanical version of the skew-gradient projections is defined with the use of the Moyal bracket $$\xi _{t}(\xi )\wedge {G}_{a}(\xi )=0. \label{CG3}$$
The projected canonical variables have the form $$\begin{aligned}
\xi _{t}(\xi )&=&\sum_{k=0}^{\infty }\frac{1}{k!}(...((\xi \wedge {G}^{a_{1}}) \wedge {G}^{a_{2}})...\wedge {G}^{a_{k}}) \nonumber \\
&&\circ {G}_{a_{1}}\circ {G}_{a_{2}}...\circ {G}_{a_{k}}. \label{SGRAD3}\end{aligned}$$ The quantum analogue of Eq.(\[FSG\]) is $$\begin{aligned}
f_{t}(\xi )&=&\sum_{k=0}^{\infty }\frac{1}{k!}(...((f(\xi )\wedge {G}^{a_{1}})\wedge {G}^{a_{2}})...\wedge {G}^{a_{k}}) \nonumber \\
&&\circ {G}_{a_{1}}\circ {G}_{a_{2}}...\circ {G}_{a_{k}}. \label{SGRAD4}\end{aligned}$$ The function $f_{t}(\xi)$ obeys equation $$f_{t}(\xi ) \wedge {G}_{a}(\xi) = 0.
\label{LAB}$$
The evolution equation which is the analogue of Eq.(\[EVOL\]) takes the form $$\frac{\partial}{\partial t}f(\xi) = f(\xi) \wedge {H}_{t}(\xi) \label{PEV2}$$ where ${H}_{t}(\xi)$ is the Hamiltonian function projected onto the constraint submanifold as prescribed by Eq.(\[SGRAD4\]).
Any function projected quantum-mechanically onto the constraint submanifold can be represented in the form [@MIKR] $$f_{t}(\xi) = \varphi(\star \xi_{t}(\xi)).
\label{GREAT}$$ In the space of projected functions, the set of projected canonical variables $\xi_{t}(\xi)$ is therefore complete.
![Quantum projection $\xi_{t}(\xi)$ defined by Eq.(4.21). The submanifold $\Gamma_{\star} = \{\xi _{t}(\xi ) : \xi \in T_{*}\mathbb{R}^{n}\}$ does not coincide with the constraint submanifold $\Gamma^{*} = \{\xi : {G}_{a}(\xi ) = 0 \}$. The variance is of order $\sim \hbar^2$. The constraint submanifold $\Gamma^{*}$ can be parameterized by classical projection $\Gamma^{*} = \{\xi _{s}(\xi ) : \xi \in T_{*}\mathbb{R}^{n}\}$ constructed with the use of the quantum constraint functions ${G}_{a}(\xi )$. []{data-label="fig5"}](qproj.eps){width="7.5"}
Coordinate star-transformations do not keep geometric relations among quantum objects
-------------------------------------------------------------------------------------
The evolution equation in the quantum constraint systems has the form of Eq.(\[PEV2\]) which is essentially the same as in the quantum unconstrained systems. Replacing $H(\xi )$ by $H_{t}(\xi )$, one can work further with solutions $u(\xi ,\tau )$ of quantum Hamilton’s equations (\[QF\]). It is not required for points $\xi $ to belong to the constraint submanifold, so phase-space trajectories $u(\xi ,\tau )$ occupy the whole phase space.
The quantum phase flow preserves the constraint functions in the following sense: $$G_{a}(\xi )=G_{a}(\star u(\xi ,\tau )). \label{CONCON}$$
The alternative equation $G_{a}(\xi )=G_{a}(u(\xi,\tau ))$ which would carry the conventional geometric meaning uses pre-conditionally the dot-composition law which is not allowed quantum-mechanically. It is obviously violated, so in the usual sense $u(\xi,\tau ) \notin \Gamma^{*}$ for $\tau > 0$ even if $u(\xi,\tau =0) = \xi \in \Gamma^{*}$ (see Fig. \[fig12\]).
Any attempt to decide if $u(\xi,\tau ) \in \Gamma^{*}$ involves the dot-composition e.g. $$u(\xi,\tau )~{\in}~\Gamma^{*} ~ \leftrightarrow ~ \forall a ~ G_{a}(u(\xi,\tau )) = 0.
\label{cbelong}$$ Statements involving the dot-composition are, however, forbidden.
Surprisingly, expressive means of the star-product formalism are not enough to formulate the simple geometric idea that a trajectory belongs to a submanifold.
We wish to find statements admissible quantum-mechanically and from other hand which would support relations of belonging and intersection inherent for geometric objects.
It is tempting to interpret Eqs.(\[CONCON\]) as the evidence that quantum trajectories $u(\xi,\tau )$ do not leave, in a specific quantum-mechanical sense, level sets of constraint functions $\{\xi : G_{a}(\xi ) = \textrm{constant} \}$.
Such a statement has the invariant meaning with respect to unitary transformations: Suppose the map $v_{+}$: $\upsilon \rightarrow \xi = v_{+}(\upsilon)$, corresponds to a unitary transformation in the Hilbert space. The inverse unitary transformation generates the inverse map $v_{-}$: $\xi \rightarrow \upsilon = v_{-}(\xi)$, such that $v_{-}(\star v_{+}(\upsilon)) = \upsilon$ and, by virtue of Eq.(\[BRINVA\]), $v_{+}(\star v_{-}(\xi)) = \xi$. In the coordinate system $\{\upsilon^{i}\}$, the constraint functions become $$G_{a}^{\prime}(\upsilon) = G_{a}(\star v_{+}(\upsilon)).$$ Equation (\[BRINVA\]) allows to change the variables $\xi \to \star v_{+}(\upsilon)$ in Eq.(\[CONCON\]) to give $$G_{a}^{\prime}(\upsilon) = G_{a}^{\prime}(\star u^{\prime}(\upsilon,\tau)) \label{CONCON-2}$$ where $$u^{\prime}(\upsilon,\tau) = v_{-}(\star u( \star v_{+}(\upsilon),\tau))
\label{nonloc}$$ represents the quantum phase flow in the coordinate system $\{\upsilon^{i}\}$. Equations (\[CONCON\]) and (\[CONCON-2\]) are therefore equivalent. They show that “do not leave” represents a predicate invariant under unitary transformations.
The non-local character of relations between the quantum phase flows is displayed in Eq.(\[nonloc\]) explicitly. One can conclude that quantum trajectories do not transform under unitary transformations as geometric objects.
![Constraint submanifolds $\Gamma^{*}$ and $\Gamma^{* \prime}$ (solid lines) and quantum trajectories $u(\xi,\tau)$ and $u^{\prime}(\upsilon,\tau)$ (dashed lines) in unitary equivalent coordinate systems $\{\xi^{i}\}$ and $\{\upsilon^{i}\}$, respectively. As shown, $u(\xi,\tau)$ crosses $\Gamma^{*}$ twice, whereas its image $u^{\prime}(\upsilon,\tau)$ crosses $\Gamma^{* \prime}$ once. Any counting of the intersections rests on an implicit use of the dot-composition, an operation which is forbidden quantum-mechanically. The property of the statements $u(\xi,\tau) \in \Gamma^{*}$ and $u^{\prime}(\upsilon,\tau) \in \Gamma^{* \prime}$ be true or false depends on unitary transformations. From the viewpoints of Eqs.(\[CONCON\]) and (\[CONCON-2\]), $u(\xi,\tau)$ and $u^{\prime}(\upsilon,\tau)$ belong to the level sets of $G_{a}(\xi)$ and $G^{\prime}_{a}(\upsilon)$, respectively. However, from condition $G_{a}(\star u(\xi,\tau)) = 0$ it does not follow that $G_{a}^{\prime}(\star u^{\prime}(\upsilon,\tau)) = 0$ and *vice versa*. Geometric relations among quantum objects, which use the dot-composition, do not have objective meaning. []{data-label="fig12"}](dia-3.eps){width="2.5"}
The coordinate transformation $v_{-}$: $\xi \rightarrow \upsilon = v_{-}(\xi)$ does not superpose $\Gamma ^{*}$ and $$\Gamma ^{*\prime } = \{\upsilon: G_{a}^{\prime}(\upsilon) =0\}.$$ Assuming $\xi \in \Gamma^{*}$, we obtain $G_{a}^{\prime}(v_{-}(\xi)) \neq G_{a}^{\prime}(\star v_{-}(\xi)) =
G_{a}(\xi) = 0$ and therefore $v_{-}(\xi) \notin \Gamma^{*\prime}$ in general. The constraint submanifold does not transform under unitary transformations as a geometric object also.
We see that points of $\Gamma^{*}$ transform differently from $\Gamma^{*}$. They are “not attached to $\Gamma^{*}$”. In new coordinate system, $\Gamma^{*}$ represents a set of new points. To put it precisely, $$\xi \in \Gamma^{*} \nrightarrow \upsilon = v_{-}(\xi) \in \Gamma^{*\prime} = v_{-}(\Gamma^{*}).$$
Unitary transformations affect the visualization of trajectories and submanifolds. The relation “do not leave” supports, however, some features inherent to the usual geometric relations “belong” and “intersect”. One can show e.g. that if quantum trajectories do not leave the level sets of $G_{a}(\xi)$ and each level set of $G_{a}(\xi)$ is a subset of one of the level sets of $F_{a}(\xi)$ then quantum trajectories do not leave the level sets of $F_{a}(\xi)$.
One cannot assign to quantum trajectories definite values of energy and constraint functions. In the coordinate system $\{\xi^{i}\}$ one has $E_{\xi} = H(\star u(\xi,\tau))$, whereas in the coordinate system $\{\upsilon^{i}\}$ one has $E_{\upsilon} = H^{\prime}(\star u^{\prime}(\upsilon,\tau))$ where $H^{\prime}(\upsilon)$ is defined by Eq.(\[hprime\]). The constants $E_{\xi}$ and $E_{\upsilon}$ do not depend on time. However, $E_{\xi} \neq E_{\upsilon}$ in general even if trajectories are related by a unitary transformation. The same conclusion holds for constraint functions, as shown on Fig. \[fig12\].
Finally, the syntax of the star-product formalism is not rich enough to express the simple geometric idea that trajectory belongs to a submanifold.
The star-product geometry admits the statement that quantum trajectories do not leave level sets of the constraint functions. The validity of this statement is not affected by unitary transformations and has the objective meaning. The quantum-mechanical relation “do not leave” is the remnant of usual relations of belonging and intersection inherent to geometric objects. It cannot be completely visualized, however.
Evolution and skew-gradient projection
--------------------------------------
The classical phase flow commutes with the classical skew-gradient projection, as discussed in Sect. IV. We want to clarify if such a property holds for quantum systems.
Given the quantum trajectories $u(\xi ,\tau)$ are constructed, the evolution of arbitrary function can be found with the help of Eq.(\[TF\]) and its projection can be computed using Eq.(\[SGRAD4\]).
The quantum projection applied to arbitrary function cannot be expressed in terms of the same function of the projected arguments Eq.(\[GREAT\]), basically because the classical relation $(fg)_{s} = f_{s}{g}_{s}$ turns to the quantum inequality $(f \star g)_{t} \neq {f}_{t} \star {g}_{t}$. In terms of a function $\varphi(\xi)$ defined for $f(\xi) \equiv f(\xi,0)$ in Eq.(\[GREAT\]), the quantum analogue for Eqs.(\[23456\]) reads $$f_{t}(\xi,\tau) = \varphi(\star u_{t}(\xi,\tau)).
\label{0000}$$ The construction of $\varphi(\xi)$ from $f(\xi)$ is a complicated task, so practical advantages of this equation are not seen immediately.
Equation (\[0000\]) accomplishes solution of the evolution problem for observable $f(\xi,\tau)$ in terms of quantum characteristics.
It remains to prove $$\begin{aligned}
u_{t}(\xi ,\tau ) &=& u(\star \xi _{t}(\xi ),\tau ) \nonumber \\
&=& \xi _{t}(\star u(\xi ,\tau )). \label{QABBA}\end{aligned}$$
The first line is a consequence of the fact that the constraint functions $G_{a}(\xi)$ are Moyal commutative with the projected Hamiltonian function $H_{t}(\xi)$. To arrive at the second line, it is sufficient to use Eq.(\[CONCON\]) to replace arguments of the constraint functions entering the skew-gradient projection.
The quantum phase flow commutes with the quantum projection, as illustrated on Fig. \[fig6\].
The composition law (\[comp\]) for quantum phase flow holds for the constraint systems. It holds for projected quantum trajectories also: $$u_{t}(\xi,\tau_{1} + \tau_{2}) = u_{t}(\star u_{t}(\xi,\tau_{1}),\tau_{2}), \label{concomp}$$ as a consequence of Eqs.(\[QABBA\]).
![Quantum phase flow is commutative with quantum projection operation: $u(\star \xi _{t}(\xi ),\tau ))=\xi _{t}(\star u(\xi ,\tau ))$. The phase-space trajectory $u_{t}(\xi ,\tau )$ does not belong to the submanifold $\Gamma_{\star} = \{\xi _{t}(\xi ) : \xi \in T_{*}\mathbb{R}^{n}\}$ except for $\tau = 0$, so the white planes on Figs. \[fig5\] and \[fig6\] are distinct.[]{data-label="fig6"}](qflow-kopie.eps){width="8.5"}
Conclusions
===========
The method of characteristics for solving evolution equations in classical and quantum, unconstrained and constrained systems has been discussed. The analysis rests on the Groenewold-Moyal star-product technique.
The classical method of characteristics applies to first-order PDE and consists in finding characteristics which are solutions of first-order ODE. For the classical Liouville equation, the corresponding first-order ODE are the Hamilton’s equations and the characteristics of interest are the classical phase-space trajectories.
The quantum Liouville equation is the infinite-order PDE. Nevertheless, it can be solved in terms of quantum characteristics which are solutions of the quantum Hamilton’s equations. These equations represent infinite-order PDE also.
Using the star-product formalism, we showed that to any fixed order in the Planck’s constant, quantum characteristics can be constructed by solving a closed system of ODE for quantum trajectories and generalized Jacobi fields. The quantum evolution becomes local in an extended phase space with new dimensions ascribed to generalized Jacobi fields. This statement holds for constraint systems also.
One-parameter continuous groups of unitary transformations in quantum theory represent the quantum deformation of one-parameter continuous groups of canonical transformations in classical theory. Quantum phase flow, induced by the evolution in the Hilbert space, does not satisfy the condition for canonicity and preserves the Moyal bracket rather than the Poisson bracket. The knowledge of quantum phase flow allows to reconstruct quantum dynamics.
Systems: unconstrained constrained
----------- -------------------------- ------------------------------------
classical $f(c(\xi,\tau),0)$ $f(c_{s}(\xi,\tau),0)$
quantum $f(\star u(\xi,\tau),0)$ $\varphi(\star u_{t}(\xi,\tau),0)$
: Solutions of evolution equations for functions (second column) and projected functions (third column) of classical systems (first row) and quantum systems (second row) in terms of characteristics. $c(\xi,\tau)$ are solutions of classical Hamilton’s equations with hamiltonian function $\mathcal{H}(\xi)$ (second column) and projected hamiltonian function $\mathcal{H}_{s}(\xi)$ (third column). $c_{s}(\xi,\tau)$ are classical projections of $c(\xi,\tau)$. $u(\xi,\tau)$ are solutions of quantum Hamilton’s equations with Hamiltonian function ${H}(\xi)$ (second column) and projected Hamiltonian function ${H}_{t}(\xi)$ (third column). $u_{t}(\xi,\tau)$ are quantum projections of $u(\xi,\tau)$. $\varphi(\xi,0)$ is defined in terms of $f_{t}(\xi,0)$ by Eq.(\[0000\]). Classical and quantum projections are defined by Eqs.(\[FSG\]) and (\[SGRAD4\]), respectively. []{data-label="lab3"}
The results reported in this work are valid for semiclassically admissible functions, i.e. for functions regular in $\hbar$ at $\hbar = 0$. Physical observables are normally associated with classical devices and expressed as classical functions of classical variables. The quantum evolution turns, however, the set of classical functions into the set of semiclassically admissible functions.
The use of the skew-gradient projection formalism allows to treat unconstrained and constraint systems essentially on the same footing. We showed that the skew-gradient projections of solutions of the quantum Hamilton’s equations onto the constraint submanifold comprise the complete information on quantum dynamics of constraint systems.
The formalism we developed applies in particular to the dynamics of gauge-invariant systems which become second class upon gauge fixing. The quantum dynamics of charged particles in external gauge fields on flat and curved manifolds is discussed within the star-product formalism in a gauge-invariant manner in Refs. [@OSKA04; @OSKA05].
The evolution equations for semiclassically admissible functions admit solutions in terms of characteristics in all physical systems, as summarized in Table \[lab3\].
The analytic geometry uses the dot-product and rests on classical ideas how to arrange composition of functions. It is well known that all theorems of geometry can be reformulated using tools of the analytic geometry.
Given the dot-product is replaced with the star-product, we arrive at the star-product geometry with well defined coordinate systems, transformations of the coordinates and equations for functions of the coordinates. However, objects of the star-product geometry, defined algebraically, can hardly be visualized:
We found that quantum trajectories and constraint submanifolds do not transform as geometric objects. The statement “quantum trajectory belongs to a constraint submanifold” can be changed to the opposite by a unitary transformation. The star-composition law (\[comp\]) shows also that the quantum evolution cannot be treated literally as moving along a quantum trajectory.
We attempted to find statements whose validity cannot be reverted by transformations of the coordinate system and which, from other hand, express relations similar to “belong”, “intersect”, etc. A weak but consistent geometric meaning can be attributed to the statement “quantum trajectories do not leave level sets of constraint functions”.
The dot-product composition of linear functions coincides with the star-product composition of linear functions, so under linear transformations straight lines and hyperplanes turn to straight lines and hyperplanes. Relations of the linear algebra, imbedded into the star-product geometry, preserve the consistent geometric meaning.
Finally, this work extended the method of characteristics to quantum unconstrained and constraint systems. From the point of view of applications, it is motivated by the fact of using classical phase-space trajectories in transport models and by the appearance of constraints in relativistic versions of QMD transport models. The method of quantum characteristics represents the promising tool for solving numerically many-body potential scattering problems.
This work is supported by DFG grant No. 436 RUS 113/721/0-2, RFBR grant No. 06-02-04004, and European Graduiertenkolleg GR683.
Quantum phase flow in integrable systems
========================================
A completely integrable classical system admits a canonical transformation which makes the Hamiltonian function depending on half of the canonical variables only (see e.g. [@ARNO]). Such variables if exist can be taken to be canonical momenta which usually referred to as actions. The canonically conjugate coordinates are referred to as angles.
In quantum mechanics, we search for a unitary transformation (or a half-unitary transformation [@BLEAF2]) allowing to express the Hamiltonian as an operator function of operators of canonical momenta. If such a transformation exists, the Hamiltonian commutes with the operators of canonical momenta, so that the canonical momenta are integrals of motion, whereas the operators of canonical coordinates depend linearly on time.
The quantum integrable systems admit an equivalent treatment in the framework of the Groenewold-Moyal dynamics [@KAMA]. Given the Hamiltonian function $H(\xi)$ is known, one has to search for a map $v_{+}$: $$\upsilon \rightarrow \xi = v_{+}(\upsilon),
\label{UNITA}$$ preserving the Moyal bracket, for which the system admits a Hamiltonian function $$H^{\prime}(\upsilon) = H(\star v_{+}(\upsilon))
\label{hprime}$$ depending on actions $\upsilon^{n+1},\ldots,\upsilon^{2n}$, i.e., canonical momenta only. We restrict the discussion by unitary transformations (\[UNITA\]), leaving aside more involved cases described in Refs. [@BLEAF2; @ANDER].
Let $u^{i}(\xi,\tau)$ and $a^{i}(\upsilon,\tau)$ be solutions of Eq.(\[QF\]) with Hamiltonian functions $H(\xi)$ and $H^{\prime}(\upsilon)$, respectively. In the coordinate system $\{\upsilon^{i}\}$, the series expansion (\[TEU\]) is truncated at $s=1$. The quantum Hamilton’s equations give a ’motion by inertia’: $$a^{i}(\upsilon,\tau) = \upsilon^{i} + \{\upsilon^{i},H^{\prime}(\upsilon)\}\tau.
\label{INER}$$ The actions ($i=n+1,...,2n$) remain constant, whereas the angles ($i=1,...,n$) evolve linearly with time. Equations (\[INER\]) can be derived as the Weyl’s transform of the equations of motion for the Heisenberg operators of the canonical coordinates and momenta obtained by a unitary transformation from the initial set of operators the canonical coordinates and momenta.
The Poisson bracket $\{\upsilon^{i},H^{\prime}(\upsilon)\}$ depends for any $i$ on the actions only, so one has $$\begin{aligned}
a^{i}(\upsilon,\tau) \circ a^{j}(\upsilon,\tau) &=& a^{i}(\upsilon,\tau)a^{j}(\upsilon,\tau), \nonumber \\
a^{i}(\upsilon,\tau) \wedge a^{j}(\upsilon,\tau) &=& \{a^{i}(\upsilon,\tau),a^{j}(\upsilon,\tau)\}. \nonumber \end{aligned}$$ The map $a$: $$\upsilon \rightarrow \acute{\upsilon} = a(\upsilon,\tau),$$ showing the evolution in the coordinate system $\{\upsilon^{i}\}$ is both canonical and unitary, as the left-hand side of Eq.(\[INER\]) is a first-order polynomial with respect to the angles.
The actions $\upsilon^{n+1},\ldots,\upsilon^{2n}$ Poisson and Moyal commute with $H^{\prime}(\upsilon)$. Composite functions $v_{-}^{i}(\star u(\star v_{+}(\upsilon),\tau))$, where $v_{-}$ is the inverse unitary map: $$\xi \rightarrow \upsilon = v_{-}(\xi),$$ such that $v_{-}^{i}(\star v_{+}(\upsilon)) = \upsilon^{i}$, obey Eqs.(\[QF\]) and proper initial conditions and coincide with $a^{i}(\upsilon,\tau)$. It can be expressed as follows: $$u^{i}(\xi,\tau) = v_{+}^{i}(\star a(\star v_{-}(\xi),\tau)).
\label{COMBI}$$ The functions $v_{\pm}$ are defined using the star-product and depend on $\hbar$ accordingly.
We thus conclude that quantum phase flow is distinct from classical phase flow for integrable systems also. For a one-dimensional system $H = \frac{1}{2}p^2 + V(q)$, which is a classical integrable sysem for any potential $V(q)$, the first quantum correction to the phase-space trajectories appears to order $O(\hbar^2 \tau^5)$.
In general case, Eq.(\[COMBI\]) shows the connection between quantum phase flows $u^{i}(\xi,\tau)$ and $a^{i}(\upsilon,\tau)$ in two unitary equivalent coordinate systems $\{\xi^{i}\}$ and $\{\upsilon^{i}\}$.
[99]{} H. Groenewold, Physica **12**, 405 (1946). H. Weyl, Z. Phys. **46**, 1 (1927). H. Weyl, *The Theory of Groups and Quantum Mechanics*, Dover Publications, New York Inc. (1931). E. P. Wigner, Phys. Rev. **40**, 749 (1932). J. E. Moyal, Proc. Cambridge Phil. Soc. **45**, 99 (1949). M. S. Bartlett and J. E. Moyal, Proc. Camb. Phil. Soc. [**45**]{}, 545 (1949).
R. L. Stratonovich, Sov. Phys. JETP [**4**]{}, 891 (1957). A. Voros, Ann. Inst. Henri Poincaré [**24**]{}, 31 (1976); [**26**]{}, 343 (1977). F. Bayen, M. Flato, C. Fronsdal, A. Lichnerowicz, D. Sternheimer, Ann. Phys. [**111**]{}, 61 (1978); Ann. Phys. [**111**]{}, 111 (1978). P. Carruthers, F. Zachariasen, Rev. Mod. Phys. **55**, 245 (1983). M. Hillery, R.F. O’Connell, M.O. Scully, E. P. Wigner, Phys. Rep. **106**, 121 (1984). N. L. Balazs and B. K. Jennings, Phys. Rep. **104**, 347 (1989). B. Fedosov, *Deformation Quantization*, Akademie Verlag, Berlin (1996). M. I. Krivoruchenko, A. A. Raduta, A. Faessler, Phys. Rev. [**D73**]{}, 025008 (2006). M. I. Krivoruchenko, Talk given at the XIII Annual Seminar “Nonlinear Phenomena in Complex Systems: Chaos, Fractals, Phase Transitions, Self-organization”, Minsk, Belarus, May 16-19, 2006; e-Print Archive: hep-th/0610074.
P. A. M. Dirac, *The Principles of Quantum Mechanics*, 3-rd ed., Claredon Press, Oxford (1947). P. A. M. Dirac, Phys. Zeits. Sowjetunion **3**, 64 (1933). B. Leaf, J. Math. Phys. **10**, 1971 (1969); **10**, 1980 (1969). A. Anderson, Phys. Lett. **B305**, 67 (1993); Phys. Lett. **B319**, 157 (1993); Ann. Phys. **232**, 292 (1994).
T. A. Osborn, F. H. Molzahn, Ann. Phys. [**241**]{}, 79 (1995). B. Leaf, J. Math. Phys. [**9**]{}, 769 (1968). T. Curtright and C. Zachos, J. Phys. [**A32**]{}, 771 (1999).
J. Aichelin, Phys. Rep. **202**, 233 (1991). S. W. Huang, A. Faessler, Guo-Qiang Li [*et al.*]{}, Prog. Part. Nucl. Phys. **30**, 105 (1993). E. Lehmann, R. K. Puri, A. Faessler [*et al.*]{}, Prog. Part. Nucl. Phys. **30**, 219 (1993).
H. Sorge, H. Stöcker and W. Greiner, Ann. Phys. **192**, 266 (1989). T. Maruyama *et al.*, Nucl. Phys. [**A534**]{}, 720 (1991). P. A. M. Dirac, Rev. Mod. Phys. **21**, 392 (1949); Proc. Roy. Soc. **A246**, 326 (1958).
B. R. McQuarrie, T. A. Osborn, and G. C. Tabisz, Phys. Rev. [**A58**]{}, 2944 (1998). V. G. Bagrov, V. V. Belov and M. F. Kondrat’eva, Teor. Mat. Fiz. [**98**]{}, 48 (1994) \[Theor. Math. Phys. [**98**]{}, 34 (1994)\]. V. G. Bagrov, V. V. Belov, M. F. Kondrat’eva, A. M. Rogova, and A. Yu. Trifonov, J. Moscow Phys. Soc.[**3**]{}, 1 (1993). V. G. Bagrov, V. V. Belov, A. M. Rogova, and A. Yu. Trifonov, Mod. Phys. Lett. [**B7**]{}, 1667 (1993). V. G. Bagrov, V. V. Belov, A. Yu. Trifonov, Ann. Phys. [**246**]{}, 231 (1996). M. Karasev, Russ. J. Mat. Phys. **11**, 157 (2004). M. S. Marinov, Phys. Lett. A153, 5 (1991). G. Braunss and D. Rompf, J. Phys. [**A26**]{}, 4107 (1993).
M. Nakamura and N. Mishima, Nuovo Cim. **79B**, 287 (1984); Prog. Theor. Phys. **81**, 451 (1989). M. Nakamura and H. Minowa, J. Math. Phys. **34**, 50 (1993). M. Nakamura and K. Kojima, Nuovo Cim. **116B**, 287 (2001). M. I. Krivoruchenko, A. Faessler, A. A. Raduta and C. Fuchs, Int. J. Mod. Phys. **A** [22]{}, 787 (2007).
V. I. Arnold, [*Mathematical Methods of Classical Mechanics*]{}, 2-nd ed., Springer-Verlag, New York Inc. 1989. R. C. T. da Costa, Phys. Rev. **A** 23, 1982 (1981). M. I. Krivoruchenko, Amand Faessler, A. A. Raduta, C. Fuchs, Phys. Lett. **B608**, 164 (2005).
J. M. Gracia-Bondia, Contemporary Math. [**134**]{}, 93 (1992). J. F. Carinena, J. Clemente-Gallardo, E. Follana, J. M. Gracia-Bondia, A. Rivero and J. C. Varilly, J. Geom. Phys. [**32**]{}, 79 (1999).
A. Gracia-Saz, arXiv: math.QA/0411163. P. R. Holland, *The Quantum Theory of Motion*, Cambridge Uni. Press, Cambridge (1993).
T. Hakioglu and A. J. Dragt, J. Phys. **A34**, 6603 (2001). T. A. Osborn and M. F. Kondratieva, J. Phys. **A35**, 5279 (2002). M. I. Krivoruchenko, C. Fuchs, Amand Faessler, nucl-th/0605015. M. V. Karasev and V. P Maslov, *Nonlinear Poisson Brackets*, Nauka, Moscow (1991). M. Karasev, Lett. Math. Phys. **59**, 229 (2001). M. V. Karasev and T. A. Osborn, J. Phys. **A** 37, 2345 (2004). M. V. Karasev and T. A. Osborn, J. Phys. **A** 38, 8549 (2005). M. V. Karasev and V. P. Maslov, Rus. Math. Surv. **39**, 5279 (1984).
[^1]: The dot-product is the usual multiplication operation for numbers, variables and functions. It should not be mixed with scalar product of vectors.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Modern machine learning is based on powerful algorithms running on digital computing platforms and there is great interest in accelerating the learning process and making it more energy efficient. In this paper we present a fully autonomous probabilistic circuit for fast and efficient learning that makes no use of digital computing. Specifically we use SPICE simulations to demonstrate a clockless autonomous circuit where the required synaptic weights are read out in the form of analog voltages. Such autonomous circuits could be particularly of interest as standalone learning devices in the context of mobile and edge computing.'
author:
- Jan Kaiser
- Rafatul Faria
- 'Kerem Y. Camsari'
- Supriyo Datta
bibliography:
- 'library.bib'
title: 'Probabilistic Circuits for Autonomous Learning: A simulation study'
---
Introduction
============
Machine learning, inference and many other emerging applications [@schuman_survey_2017] make use of stochastic neural networks comprising (1) a binary stochastic neuron (BSN) [@ackley_learning_1987; @neal_connectionist_1992] and (2) a synapse that constructs the inputs $I_i$ to the $i^{th}$ BSN from the outputs $m_j$ of all other BSNs.
The output $m_i$ of the $i^{th}$ BSN fluctuates between +1 and -1 with a probability controlled by its input $$m_i (t+\tau_{N}) = {\rm{sgn}}\left[ \tanh \left( {I_i(t)}\right) - r\right]
\label{eq:binary_stochastic_neuron}$$ where $r$ represents a random number in the range $\left[-1,+1\right]$, and $\tau_{N}$ is the time it takes for a neuron to provide a stochastic output $m_i$ in accordance with a new input $I_i$.
Usually the synaptic function, $ I_i (\{m\})$ is linear and is defined by a set of weights $W_{ij}$ such that $$I_i (t+\tau_S) = \sum_{j} W_{ij} m_j(t)
\label{eq:synaptic_function}$$ where $\tau_{S}$ is the time it takes to recompute the inputs $\{I\}$ everytime the outputs $\{m\}$ change. Typically Eqs., are implemented in software, often with special accelerators for the synaptic function using GPU/TPUs [@noauthor_google_2016; @schmidhuber_deep_2015].
The time constants $\tau_N$ and $ \tau_S$. are not important when Eqs., are implemented on a digital computer using a clock to ensure that neurons are updated sequentially and the synapse is updated between any two updates. But they play an important role in clockless operation of autonomous hardware that makes use of the natural physics of specific systems to implement Eqs., approximately. For example, Eq. can be implemented using stochastic magnetic tunnel junctions (s-MTJs) [@camsari_implementing_2017; @camsari_stochastic_2017], while resistive or capacitive crossbars can implement Eq. [@hassan_voltage-driven_2019]. It has been shown that such hardware implementations can operate autonomously without clocks, if *the BSN operates slower than the synapse*, that is, if $\tau_N >> \tau_S$ [@sutton_autonomous_2019].
Stochastic neural networks defined by Eqs. , can be used for inference whereby the weights $W_{ij}$ are designed such that the system has a very high probability of visiting configurations defined by $\{m\}= \{v\}_n$, where $\{v\}_n$ represents a specified set of patterns. However, the most challenging and time-consuming part of implementing a neural network is not the inference function, but the learning required to determine the correct weights $W_{ij}$ for a given application. This is commonly done offline using powerful cloud-based processors and there is great interest in accelerating the learning process and making it more energy efficient so that it can become a routine part of mobile and edge computing.
In this paper we present a new approach to the problem of fast and efficient learning that makes no use of digital computing at all. Instead it makes use of the natural physics of a fully autonomous probabilistic circuit composed of standard electronic components like resistors, capacitors and transistors along with stochastic magnetic tunnel junctions (MTJs).
{width="0.75\linewidth"}
We focus on a fully visible Boltzmann machine (FVBM), a form of stochastic recurrent neural network, for which the most common learning algorithm is based on the gradient ascent approach to optimizing the maximum likelihood function [@koller_probabilistic_2009; @carreira-perpinan_contrastive_2005]. We use a slightly simplified version of this approach, whereby the weights are changed incrementally according to $$W_{ij}(t+\Delta t) = W_{ij}(t) + \epsilon [v_i v_j - m_i m_j - \lambda W_{ij}(t)]$$ where $\epsilon$ is the learning parameter, $\lambda$ is the regularization parameter [@ng_feature_2004] and the bipolar training vectors $\{v\}_n$ are cycled sequentially or as an average. The term $v_iv_j$ is the (average) correlation between the $i$th and the $j$th entry of the training vector $\{v\}_n$. The term $m_im_j$ corresponds to the sampled correlation taken from the model’s distribution. The advantage of this network topology is that the learning rule is local and tolerant to stochasticity.
For our autonomous operation we replace this equation with its continuous time version ($\tau_{L}$: learning time constant) $$\frac{dW_{ij}}{dt} = \frac{v_i v_j - m_i m_j - \lambda W_{ij}}{\tau_{L}}
\label{eq:learning_rule}$$ which we translate into an RC circuit by associating $W_{ij}$ with the voltage on a capacitor C driven by a voltage source $(V_{v,ij} - V_{m,ij})$ with a series resistance R (Fig.): $$C \frac{dV_{ij}}{dt} = \frac{V_{v,ij} - V_{m,ij} -V_{ij}}{R}
\label{eq:learning_circuit}$$ with $v_iv_j=V_{v,ij}/(V_\mathrm{DD}/2)$ and $m_im_j=V_{m,ij}/(V_\mathrm{DD}/2)$. From Fig. \[fig:autonomous\_learning\_circuit\] and comparing Eqs. , it is easy to see how the weights and the learning and regularization parameters are mapped into circuit elements: $W_{ij}=A_v V_{ij}/V_0$, $\lambda = V_0/(A_v V_\mathrm{DD}/2)$ and $\tau_{L}=\lambda RC$ where $A_v$ is the voltage gain of OP3 in Fig. \[fig:autonomous\_learning\_circuit\]. For proper operation the learning time scale $\tau_L$ has to be much larger than the neuron time $\tau_N$ to be able to collect enough statistics throughout the learning process.
A key element of this approach is the representation of the weights $W$ with $voltages$ rather than with programmable $resistances$ for which memristors and other technologies are still in development [@li_review_2018]. By contrast the charging of capacitors is a textbook phenomenon, allowing us to design a learning circuit that can be built today with established technology. The idea of using capacitor voltages to represent weights in neural networks has been presented by several authors for different network topologies in analog learning circuits [@kim_analog_2017; @schneider_analog_1993; @card_learning_1994; @sung_perspective:_2018]. The use of capacitors has the advantage of having a high level of linearity and symmetry for the weight updates during the training process [@li_capacitor-based_2018].
In **Section II**, we will describe such a learning circuit that emulates Eqs. -. The training images or patterns $\{v\}_{n}$ are fed in as electrical signals into the input terminals, and the synaptic weights $W _{ij}$ can then be read out in the form of voltages from the output terminals. Alternatively the values can be stored in a non-volatile memory from which they can subsequently be read and used for inference. In **Section III**, we will present SPICE simulations demonstrating the operation of this autonomous learning circuit.
Clockless learning circuit emulating Eqs.(1)-(3)
================================================
The autonomous learning circuit has 3 parts where each part represents one of the three Eqs. -. On the left hand side of Fig. \[fig:autonomous\_learning\_circuit\], the training data is fed into the circuit by supplying a voltage $V_{v,ij}$ which is given by the $i$th entry of the bipolar training vector $v_i$ multiplied by the $j$th entry of the training vector $v_j$ and scaled by the supply voltage $V_{DD}/2$. The training vectors can be fed in sequentially or as an average of all training vectors. The weight voltage $V_{ij}$ across capacitor $C$ follows Eq. where $V_{v,ij}$ is compared to voltage $V_{m,ij}$ which represents correlation of the outputs of BSNs $m_i$ and $m_j$. Voltage $V_{m,ij}$ is computed in the circuit by using an XNOR gate that is connected to the output of BSN $i$ and BSN $j$. The synapse in the center of the circuit connects weight voltages to neurons according to Eq. . Voltage $V_{ij}$ has to be multiplied by 1 or -1 depending on the current value of $m_j$. This is accomplished by using a switch which connects either the positive or the negative node of $V_{ij}$ to the operational amplifiers OP1 and OP2. Here, OP1 accumulates all negative contributions and OP2 all positive contributions of the synaptic function. The differential amplifier OP3 takes the difference between the output voltages of OP2 and OP1 and amplifies the voltage by amplification factor $A_v$. This voltage conversion is used to control the voltage level of $V_{ij}$ in relation to the input voltage of each BSN. The voltage level at the input of the BSN is fixed by the reference voltage of the BSN $V_0$. However, the voltage level of $V_{ij}$ can be adjusted and utilized to adjust the regularization parameter $\lambda$ in the learning rule (Eq. ). The functionality of the BSN is described by Eq. where the dimensionless input $I(t)$ is given by $I_i(t)=V_\mathrm{i,in}(t)/V_0$. This relates the voltage $V_{ij}$ to the dimensionless weight by $W_{ij}=A_v V_{ij}/V_0$. The hardware implementation of the BSN uses a stochastic MTJ in series to a transistor [@camsari_implementing_2017]. Due to thermal fluctuations of the low-barrier magnet (LBM) of the MTJ the output voltage of the MTJ fluctuates randomly but with the right statistics given by Eqn. \[eq:binary\_stochastic\_neuron\]. The time dynamics of the LBM can be obtained by solving the stochastic Landau-Lifshitz-Gilbert (LLG) equation. Due to the fast thermal fluctuations of the LBM in the MTJ, Eq. can be evaluated on a subnanosecond timescale leading to fast generation of samples [@kaiser_ultrafast_2019; @hassan_low-barrier_2019].
Fig. \[fig:autonomous\_learning\_circuit\] just shows the hardware implementation of one weight and one BSN. The size of the whole circuit depends on the size of the training vector $N$. For every entry of the training vector one BSN is needed. The number of weights which is the number capacitive circuits is given by $N(N-1)/2$ where every connection between BSNs is assumed to be reciprocal.
The learning process is captured by Eqs. and . The whole learning process has similarity with the software implementation of persistent contrastive divergence (PCD) [@tieleman_training_2008] since the circuit takes samples from the models distribution ($V_{m,ij}$) and compares it to the target distribution ($V_{v,ij}$) without reinitializing the Markov Chain after a weight update. During the learning process voltage $V_{ij}$ reaches a constant average value where $\frac{dV_{ij}}{dt} \approx 0$. This voltage $V_{ij}=V_{ij,\mathrm{learned}}$ corresponds to the learned weight.
For inference the capacitor $C$ is replaced by a voltage source of voltage $V_{ij,\mathrm{learned}}$. Consequently, the autonomous circuit will compute the desired functionality given by the training vectors. In general, training and inference has to be performed on identical hardware in order to learn around variations. It is important to note that in inference mode this circuit can be used for optimization by performing electrical annealing. This is done by increasing all weights voltages $V_{ij}$ by the same factor over time. In this way the ground state of a Hamiltonian like the Ising Hamiltionian can be found [@sutton_intrinsic_2017; @camsari_scaled_2018].
![**Feeding of training data into the circuit** a) Weight voltage $V_{1,5}$ over time for sequential and average feeding in of the correlation between visible unit $i$ and visible unit $j$ for training a full adder. b) Correlation $v_1v_5$ vs. time $t$. All eight lines of the truth table of a full adder are cycled through where every vector is shown for time $T=1$ ns at a time. c) Enlarged version of subfigure a). For sequential feeding in of data, the voltage change in $v_1v_5$ directly affects $V_{1,5}$.[]{data-label="fig:fig_vivj"}](./fig2.pdf){width="\linewidth"}
$ \ \mathbf{A} \ $ $ \ \mathbf{B} \ $ $\mathbf{C_\mathrm{in}}$ $ \ \mathbf{S} \ $ $\mathbf{C_\mathrm{out}}$ **Dec** $P_\mathrm{Ideal}$
-------------------- -------------------- -------------------------- -------------------- --------------------------- --------- --------------------
-1 -1 -1 -1 -1 0 0.125
-1 -1 1 1 -1 6 0.125
-1 1 -1 1 -1 10 0.125
-1 1 1 -1 1 13 0.125
1 -1 -1 1 -1 18 0.125
1 -1 1 -1 1 21 0.125
1 1 -1 -1 1 25 0.125
1 1 1 1 1 31 0.125
: **Truth Table of Full Adder** Every 0 in the binary representation of the Full Adder is replaced by a -1 in the bipolar representation. “Dec” represents the decimal conversion of each line. $P_\mathrm{Ideal}$ is the ideal probability distribution were every line’s probability is $p=1/8=0.125$.[]{data-label="table: FA"}
SPICE simulations
=================
In this section the autonomous learning circuit in Fig. \[fig:autonomous\_learning\_circuit\] is simulated in SPICE. We show how the proposed circuit can be used for both inference and learning. As examples we demonstrate the learning on a Full Adder (FA) and on 5x3 digits images. For all SPICE simulations the following parameters are used for the stochastic MTJ in the BSN implementation: Saturation magnetization $M_S=1100 \ \mathrm{emu/cc}$, LBM diameter $D=22$ nm, LBM thickness $l=2$ nm, TMR=110%, damping coefficient $\alpha=0.01$, temperature $T=300 \ \mathrm{K}$ and demagnetization field $H_D=4 \pi M_S V$ with $V=(D/2)^2 \pi l$. For the transitors, 14 nm HP-FinFET PTM models were used [@noauthor_predictive_2012] with fin number $fin=1$ for the inverters and $fin=2$ for XNOR-gates. Ideal operational amplifiers and switches are used in the synapse. For this parameter set, the reference voltage of each BSN is given by $V_0= 50 \ \mathrm{mV}$ [@camsari_implementing_2017]. The characteristic time of the BSNs $\tau_N$ is in the order of 100 ps [@hassan_low-barrier_2019] and much larger than the time it takes for the synaptic connections, namely the resistors and operational amplifiers, to propagate BSN outputs to neighboring inputs ($\tau_N \gg \tau_S$).\
Learning addition {#learning-addition .unnumbered}
-----------------
As first training example we use the probability distribution of a full adder (FA). The truth table/probability distribution of a full adder with bipolar variables is shown in table \[table: FA\]. To learn this distribution the correlation terms $v_iv_j$ in the learning rule have to be fed into the voltage node $V_{v,ij}$. The correlation is dependent on what training vector / truth table line is fed in. For the second line for the truth table for example $v_1v_2=-1 \cdot -1 =1$ and $v_1v_3=-1 \cdot 1 =-1$ with A being the first node, B the second node and so on. In Fig. \[fig:fig\_vivj\] b) the correlation $v_1v_5$ is shown. For the sequential case the value of $v_1v_5$ is obtained by circling through all lines of the truth table where each training vector is shown for $1$ ns. $A$ and $C_\mathrm{out}$ in table \[table: FA\] only differ in the fourth and fifth line for which $v_1v_5=-1$. For all other cases $v_1v_5=1$. The average of all lines is shown as red solid line. Fig. \[fig:fig\_vivj\] a) shows how the weight voltage $V_{ij}$ with $i=1$ and $j=5$ for FA learning and the first $1000$ ns of training. The following learning parameters have been used for the FA: $\tau_L=62.5$ ns where $C=1$ nF and $R=5 \ \mathrm{k\Omega}$, $A_v=10$ and $R_F=1 \ \mathrm{M\Omega}$. This choice of learning parameters ensures that $\tau_L \gg \tau_N$. Due to the averaging effect of the RC-circuit both sequential and average feeding of the training vector result in similar learning behavior as long as the RC-constant is much larger than the timescale of sequential feeding. Fig. \[fig:fig\_vivj\] c) shows the enlarged version of Fig. \[fig:fig\_vivj\] a). For the sequential feeding, voltage $V_{1,5}$ changes substantially every time $v_iv_j$ switches to -1.
![**Training of a Full Adder in SPICE** a) Probability distribution of a trained full adder is compared to the ideal distribution with inputs binary inputs $A$, $B$, $C_\mathrm{in}$ and outputs $S$ and $C_\mathrm{out}$. The training is performed for $5000$ ns. Blue bars are the probability distribution extracted from SPICE simulations by creating a histogram of the configurations of [$m$]{} over the last $2500 \ \mathrm{ns}$ of training. Red bars are the probability distribution obtained by using the weight voltages at the end of training. b) Kullback–Leibler divergence between the trained and the target distribution defined as $KL(P_\mathrm{Ideal} || P_\mathrm{Train}(t))=\sum_\mathbf{\mathrm{m}} P_\mathrm{Ideal}(\mathbf{\mathrm{m}}) \log(P_\mathrm{Ideal}(\mathbf{\mathrm{m}})/P_\mathrm{Train}(\mathbf{\mathrm{m}},t))$ . Following parameters have been used in the simulations: $C=1 \ \mathrm{nF}$, $R=5 \ \mathrm{k\Omega}$, $R_F=1 \ \mathrm{M\Omega}$, $A_v=10$, $V_0=50 \ \mathrm{mV}$.[]{data-label="fig:fig_FA"}](./fig3.pdf){width="\linewidth"}
In Fig. \[fig:fig\_FA\] a) the probability distribution of the FA $P_\mathrm{Train}$ is shown after $5000 \ \mathrm{ns}$ of learning and compared to the ideal distribution $P_\mathrm{Ideal}$. The training vector is fed in as an average. Fig. \[fig:fig\_FA\] a) also shows the normalized histogram $P_\mathrm{SPICE}$ of the sampled BSN configurations taken from the last $2500 \ \mathrm{ns}$ of learning. For both $P_\mathrm{Train}$ and $P_\mathrm{SPICE}$ the eight trained configurations of table \[table: FA\] are the dominant peaks. To monitor the training process, the Kullback-Leibner divergence between the trained to the ideal probability distribution $KL(P_\mathrm{Ideal} || P_\mathrm{Train}(t))$ is plotted as a function of training time $t$ in Fig. \[fig:fig\_FA\] b). During training the KL divergence decreases over time until it reaches a constant value at about 0.33. Fig. \[fig:fig\_FA\] shows that the probability distribution of a FA can be learned very fast with the proposed autonomous learning circuit.\
Learning image completion {#learning-image-completion .unnumbered}
-------------------------
As second example the circuit is utilized to train 10 5x3 pixel digit images shown in Fig.\[fig:fig4\] a). The network is trained for 3000 ns and the bipolar training data is fed in as average of the 10 $v_iv_j$ terms for every digit. The same learning parameters as in the previous section are used here. In \[fig:fig4\] b) the KL divergence is shown as a function of time between the trained and the ideal probability distribution where the ideal distribution has 10 peaks with each peak being 10 % for each digit. Most of the learning happens in the first 500 ns of training, however, the KL divergence still reduces slightly during the later parts of learning. After 3000 ns the KL divergence reaches a value of around 2.
For inference we replace the capacitor by a voltage source where every voltage is given by the previously learned voltage $V_{ij}$. The circuit is run for 10 instances where every instance has a unique clamping pattern of 6 pixels representing one of the 10 digits. The clamped inputs are show in Fig. \[fig:fig4\] c). The input of a clamped BSN is set to $\pm V_\mathrm{DD}/2$. Each instance is run for 100 ns and the outputs of the BSNs monitored. The BSNs fluctuate between the configurations given by the learned probability distribution. In Fig. \[fig:fig4\] d) the heat map of the output of the BSNs is shown. For every digit the most likely configuration is given by the trained digit image. To illustrate this point, the amount of BSN fluctuations is reduced by increasing the learned weight voltages by factor of $I_0=2$. This step is equivalent of reducing the temperature of the Boltzmann machine by a factor of 2 [@aarts_simulated_1989]. The circuit is again run in inference mode for 100 ns with the same clamping patterns. In Fig. \[fig:fig4\] e) the heatmap is shown. The circuit locks in into the learned digit configuration. This shows that in inference mode the circuit can be utilized for image completion.
![**Training and Testing of 5x3 Digit images** a) 5x3 digit images from 0 to 9. b) Kullback Leibner divergence during training for 3000 ns using the autonomous circuit. c)-e) Image completion: For inference, 6 unique pixels are clamped for every digit (as shown in c)). Subfigure d) ande) show the heatmap of BSN outputs during inference for running the circuit for 100 ns for d) $I_0=1$ and e) $I_0=2$.[]{data-label="fig:fig4"}](./figdigit.pdf){width="\linewidth"}
Conclusions
===========
We have used full SPICE simulations to demonstrate the feasibility of a clockless autonomous circuit running without any digital component with the learning parameters set by circuit parameters. Due to the fast BSN operation, samples are drawn at subnanosecond speeds leading ultrafast learning, as such the learning speed should be at least multiple orders of magnitudes faster compared to other computing platforms [@adachi_application_2015; @korenkevych_benchmarking_2016; @terenin_gpu-accelerated_2019]. We believe the approach can be extended to other machine learning algorithms to design autonomous circuits. Such standalone learning devices could be particularly of interest in the context of mobile and edge computing.
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was supported in part by ASCENT, one of six centers in JUMP, a Semiconductor Research Corporation (SRC) program sponsored by DARPA. KYC gratefully acknowledges support from Center for Science of Information (CSoI), an NSF Science and Technology Center, under grant CCF-0939370.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The class of $\mu$-compact sets can be considered as a natural extension of the class of compact metrizable subsets of locally convex spaces, to which the particular results well known for compact sets can be generalized. This class contains all compact sets as well as many noncompact sets widely used in applications. In this paper we give a characterization of a convex $\mu$-compact set in terms of properties of functions defined on this set. Namely, we prove that the class of convex $\mu$-compact sets can be characterized by continuity of the operation of convex closure of a function (= the double Fenchel transform) with respect to monotonic pointwise converging sequences of continuous bounded and of lower semicontinuous lower bounded functions.'
author:
- |
M.E. Shirokov[^1]\
\
Steklov Mathematical Institute, Moscow, Russia
title: 'Characterization of convex $\mu$-compact sets'
---
The properties of compact sets in the context of convex analysis have been studied by many authors (see [@1; @4; @5] and the references therein). It is natural to ask about possible generalizations of the results proved for compact convex sets to noncompact sets. In [@2] one such generalization concerning the particular class of sets called $\mu$-compact sets is considered. In [@2; @3] it is shown that for this class of sets, which includes all compact convex sets as well as some noncompact sets widely used in applications, many results of the Choquet theory [@1] and of the Vesterstrom-O’Brien theory [@4; @5] can be proved. In this paper we give a characterization of a convex $\mu$-compact set in terms of properties of functions defined on this set.
In what follows $\mathcal{A}$ is a bounded convex complete separable metrizable subset of some locally convex space.[^2] Let $C(\mathcal{A})$ be the set of all continuous bounded functions on the set $\mathcal{A}$ and $M(\mathcal{A})$ be the set of all Borel probability measures on the set $\mathcal{A}$ endowed with the weak convergence topology [@6 Chapter II, §6]. Let $\mathrm{co}f$ and $\overline{\mathrm{co}}f$ be the convex hull and the convex closure of a function $f$, which are defined respectively as the maximal convex and the maximal convex closed (that is, lower semicontinuous) functions majorized by $f$ [@7].[^3]
With an arbitrary measure $\mu\in M(\mathcal{A})$ we associate its barycenter (average) $\mathbf b\, (\mu)\, \in \, \mathcal{A}$, which is defined by the Pettis integral (see ) $$\label{eq1}
\mathbf b(\mu)\ =\ \int_{\mathcal{A}}x \mu(dx).$$ For arbitrary $x\in\mathcal{A}\,$ let $M_{x}(\mathcal{A})$ be a convex closed subset of the set $M(\mathcal{A})$ consisting of such measures $\mu$ that $\,\mathbf b(\mu) = x$.
The barycenter map $$\label{b-map}
M(\mathcal{A})\ \ni \ \mu\; \mapsto \; \mathbf b(\mu)\ \in
\mathcal{A}$$ is continuous (this can be shown easily by applying Prokhorov’s theorem [@6 Ch.II, Th.6.7]). Hence the image of any compact subset of $M(\mathcal{A})$ under this map is a compact subset of $\mathcal{A}$. The $\mu$-compact sets are defined in [@2; @3] by the converse requirement.
**Definition.** A set $\mathcal{A}$ is called $\,\mu$-[*compact* ]{} if the preimage of any compact subset of $\mathcal{A}$ under barycenter map (\[b-map\]) is a compact subset of $M(\mathcal{A})$.
Any compact set is $\mu$-compact, since compactness of $\mathcal{A}$ implies compactness of $M(\mathcal{A})$ [@6]. The $\mu$-compactness property is studied in detail in [@3], where simple criteria of this property have been established. By using these criteria $\mu$-compactness of the following noncompact sets has been proved:
- the positive parts of the unit balls of the Banach space $\ell_{1}$ and of the Banach space $\mathfrak{T}(\mathcal{H})$ of trace class operators in a separable Hilbert space $\mathcal{H}$;
- the set of positive Borel measures on an *arbitrary* complete separable metric space with the total variation $\leq 1$ endowed with the weak convergence topology;
- the positive parts of the unit balls of the Banach spaces of linear bounded operators in $\ell_{1}$ and in $\mathfrak{T}(\mathcal{H})$ endowed with the strong operator topology.
In particular, this implies $\mu$-compactness of the set of all Borel probability measures on an arbitrary complete separable metric space endowed with the weak convergence topology, of the set of quantum states and of the set of quantum operations endowed with the strong operator topology [@8].
It is essential to note that the $\mu$-compactness property of a convex set is not purely topological but reflects the special relation between the topology and the convex structure of this set [@3].
The following theorem shows that the class of convex $\mu$-compact sets can be characterized by continuity of the operation of convex closure (coinciding with the double Fenchel transform) with respect to monotonic pointwise converging sequences of functions.
**Theorem.** *The following properties are equivalent:*
1. *the set $\mathcal{A}$ is $\mu$-compact;*
2. *for an arbitrary increasing sequence $\{f_{n}\}\subset
C(\mathcal{A})$, converging pointwise to a function $f_{0}\in
C(\mathcal{A})$, the sequence $\{\overline{\mathrm{co}}f_{n}\}$ converges pointwise to the function $\,\overline{\mathrm{co}}f_{0}$;*
3. *for an arbitrary increasing sequence $\{f_{n}\}$ of lower semicontinuous lower bounded functions on $\mathcal{A}$, converging pointwise to a function $f_{0}$, the sequence $\{\overline{\mathrm{co}}f_{n}\}$ converges pointwise to the function $\,\overline{\mathrm{co}}f_{0}$.*
*If these equivalent properties hold then for an arbitrary decreasing sequence $\{f_{n}\}$ of lower semicontinuous bounded functions on the set $\mathcal{A}$, converging pointwise to a lower semicontinuous bounded function $f_{0}$, the sequence $\{\overline{\mathrm{co}}f_{n}\}$ converges pointwise to the function $\overline{\mathrm{co}}f_{0}$.*
**Remark 1.** The functions $\,\overline{\mathrm{co}}f_{n}\,$ in (ii) are not necessarily continuous (only lower semicontinuous). By the generalized Vesterstrom-O’Brien theorem (Theorem 1 in [@3]) these functions are continuous provided the set $\mathcal{A}$ is $\mu$-compact and *stable* (the last property means openness of the convex mixing map $\mathcal{A}\times\mathcal{A}\ni(x,y)\mapsto \frac{1}{2}(x + y) \in
\mathcal{A}$ [@11]).
**Proof.** It is sufficient to show that $\mathrm{(i)}\Rightarrow\mathrm{(iii)}$ and $\mathrm{(ii)}\Rightarrow\mathrm{(i)}$.
$\mathrm{(i)}\Rightarrow\mathrm{(iii)}\,$ Let $\{f_{n}\}$ be an increasing sequence of lower semicontinuous lower bounded functions on the set $\mathcal{A}$, converging pointwise to a function $f_{0}$. We may assume that this sequence consists of nonnegative functions. It suffices to show that the assumption on existence of such $x_{0}\in\mathcal{A}$ that $$\overline{\mathrm{co}}f_{n}(x_{0})\leq\overline{\mathrm{co}}f_{0}(x_{0})-\Delta,
\quad \Delta>0,\quad \forall n,$$ where “$\leq+\infty-\Delta$” means “$\leq\Delta$”, leads to a contradiction.
By Proposition 6 in [@3] we have $$\label{co-rep}
\overline{\mathrm{co}}f_{n}(x_{0})=\inf_{\mu\in
M_{x_{0}}(\mathcal{A})}\mu(f_{n}),\quad n=0,1,2...,\quad
\textup{where}\;\, \mu(f)=\int_{\mathcal{A}} f(y)\,\mu(dy),$$ and this infimum is attained at a particular measure $\mu_{n}$ in $M_{x_{0}}(\mathcal{A})$, t.i. $\overline{\mathrm{co}}f_{n}(x_{0})=\mu_{n}(f_{n})$.
For definiteness suppose $\,\overline{\mathrm{co}}f_{0}(x_{0})\!<\!+\infty$ (the case $\,\overline{\mathrm{co}}f_{0}(x_{0})\!=\!+\infty\,$ is considered similarly). By Fenchel’s theorem (see [@7]) there exists a continuous affine function $\,\alpha\,$ on $\mathcal{A}$ such that $$\label{ineq-p-1}
\alpha(x)\leq f_{0}(x),\quad\forall
x\in\mathcal{A},\quad\textup{è}\quad\overline{\mathrm{co}}f_{0}(x_{0})\leq\alpha(x_{0})+\textstyle{\frac{1}{2}}\Delta.$$ Since the function $\alpha$ is affine, we have $$\label{ineq-p-2}
\begin{array}{c}
\displaystyle
\mu_{n}(\alpha)-\mu_{n}(f_{n})=\alpha(x_{0})-\overline{\mathrm{co}}f_{n}(x_{0})\\\\\displaystyle=
[\alpha(x_{0})-\overline{\mathrm{co}}f_{0}(x_{0})]+[\overline{\mathrm{co}}f_{0}(x_{0})
-\overline{\mathrm{co}}f_{n}(x_{0})]\geq
-\textstyle{\frac{1}{2}}\Delta+\Delta=\frac{1}{2}\Delta.
\end{array}$$ The assumed $\mu$-compactness of the set $\mathcal{A}$ implies relative compactness of the sequence $\{\mu_{n}\}$. By Prokhorov’s theorem this sequence is *tight*, which means that for any $\varepsilon>0$ there exists such compact set $\mathcal{K}_{\varepsilon}\subset\mathcal{A}$ that $\mu(\mathcal{A}\setminus\mathcal{K}_{\varepsilon})<\varepsilon\,$ [@6]. Let $M=\sup_{x\in\mathcal{A}}|\alpha(x)|$ and $\varepsilon_{0}=\frac{1}{4M}\Delta$. By using (\[ineq-p-2\]) we obtain $$\int_{\mathcal{K}_{\varepsilon_{0}}}(\alpha(x)-f_{n}(x))\mu_{n}(dx)\geq
\textstyle{\frac{1}{2}}\Delta-\displaystyle\int_{\mathcal{A}\setminus\mathcal{K}_{\varepsilon_{0}}}(\alpha(x)-f_{n}(x))\mu_{n}(dx)
\geq \textstyle{\frac{1}{4}}\Delta.$$ Hence the set $\mathcal{C}_{n}=\{x\in\mathcal{K}_{\varepsilon_{0}}\,|\,
\alpha(x)\geq f_{n}(x)+\frac{1}{4}\Delta\}$ is not empty for all $n$.
Since the sequence $\{f_{n}\}$ is increasing, the sequence $\{\mathcal{C}_{n}\}$ of *closed* subsets of the *compact* set $\mathcal{K}_{\varepsilon_{0}}$ is monotone: $\mathcal{C}_{n+1}\subseteq\mathcal{C}_{n},\;\forall n$. Hence these exists $x_{*}\in\mathcal{C}_{n}$ for all $n$. This means that $\alpha(x_{*})\geq f_{n}(x_{*})+\frac{1}{4}\Delta\,$ for all $n$ and hence $\alpha(x_{*})> f_{0}(x_{*})$ contradicting (\[ineq-p-1\]).
$\mathrm{(ii)}\Rightarrow\mathrm{(i)}\,$ Suppose the set $\mathcal{A}$ is not $\mu$-compact. Then then there exists a sequence $\{\mu_{k}\}\in M(\mathcal{A})$, which is a not relatively compact and such that the sequence $\{x_{k}=\mathbf{b}(\mu_{k})\}$ converges. By the below Lemma 1 one can consider that this sequence consists of finitely supported measures. By Prokhorov’s theorem the sequence $\{\mu_{k}\}$ is not tight. The below Lemma 2 (with the remark after it) guarantees existence of such $\varepsilon>0$ and $\delta>0$ that for any compact set $\mathcal{K}\subseteq\mathcal{A}$ and any natural $N$ there is such $k>N$ that $\mu_{k}(U_{\delta}(\mathcal{K}))< 1-\varepsilon$, where $U_{\delta}(\mathcal{K})$ is the closed $\delta$-vicinity of the set $\mathcal{K}$ (as a subset of the metric space $\mathcal{A}$). Let $\{\mathcal{K}_{n}\}$ be an increasing sequence of compact subsets of $\mathcal{A}$ such that $\bigcup_{n\in\mathbb{N}}U_{\delta/2}(\mathcal{K}_{n})=\mathcal{A}$. Denote by $d(\cdot\,,\cdot)$ the metric in $\mathcal{A}$. For each natural $n$ consider the continuous bounded function $$\label{f-n}
f_{n}(x)=1-2\,\delta^{-1}\inf_{y\in
U_{\delta/2}(\mathcal{K}_{n})}d(x,y)$$ on the set $\mathcal{A}$ such that $f_{n}(x)=1$, if $x\in
U_{\delta/2}(\mathcal{K}_{n}),$ and $f_{n}(x)<0$, if $x\in
\mathcal{A}\setminus U_{\delta}(\mathcal{K}_{n})$. It is clear that $f_{0}(x)\doteq\lim_{n\rightarrow+\infty}f_{n}(x)\equiv 1$ and hence $\overline{\mathrm{co}}f_{0}(x)\equiv 1$. Let $x_{0}$ be a limit of the sequence $\{x_{k}\}$. To obtain a contradiction it suffices to show that $$\label{ineq}
\overline{\mathrm{co}}f_{n}(x_{0})<1-\varepsilon\quad \forall
n\in\mathbb{N}.$$ By the above-stated property of the sequence $\{\mu_{k}\}$ for each $n$ and any natural $N$ there exists such $k>N$ that $\mu_{k}(U_{\delta}(\mathcal{K}_{n}))< 1-\varepsilon$ and hence $$\overline{\mathrm{co}}f_{n}(x_{k})\leq
\mathrm{co}f_{n}(x_{k})\leq\int_{\mathcal{A}}
f_{n}(x)\mu_{k}(dx)<1-\varepsilon,$$ since $\mu_{k}$ is a measure with finite support. This inequality and lower semicontinuity of the function $\overline{\mathrm{co}}f_{n}$ imply (\[ineq\]).
By using representation (\[co-rep\]) and the monotonic convergence theorem, it is easy to prove the last assertion of the theorem. $\square$
In the above proof the following assertions (well known in the measure theory) are used.
**Lemma 1.** *For an arbitrary sequence $\{\mu_{k}\}\subset M(\mathcal{A})$, which is not relatively compact, there exists a sequence $\{\tilde{\mu}_{k}\}\subset
M(\mathcal{A})$ of finitely supported measures, which is not relatively compact as well, such that $\,\mathbf{b}(\tilde{\mu}_{k})=\mathbf{b}(\mu_{k})$ for all $\,k$.*\
Since the set $M(\mathcal{A})$ can be considered as a complete separable metric space [@6 Chapter II], the above lemma is easily proved by using density of the set of finitely supported measures in $M_{x}(\mathcal{A})$ for each $x\in\mathcal{A}$ [@2 Lemma 1].
**Lemma 2.** *A subset $M_{0}\subseteq M(\mathcal{A})$ is tight if and only if for any $\varepsilon>0$ and $\delta>0$ there exists a compact subset $\,\mathcal{K}(\varepsilon,\delta)\subseteq\mathcal{A}$ such that $$\mu(U_{\delta}(\mathcal{K}(\varepsilon,\delta)))\geq 1-\varepsilon$$ for all $\mu\in M_{0}$, where $U_{\delta}(\mathcal{K}(\varepsilon,\delta))$ is the closed $\delta$-vicinity of the set $\,\mathcal{K}(\varepsilon,\delta)$.*
Since any finite subset of $M(\mathcal{A})$ is tight, the words “*for all $\mu\in M_{0}$*” in the above criterion may be replaced by “*for all $\mu\in M_{0}\setminus M'$, where $M'$ is a finite subset of $M_{0}$*”.
**Proof.** It is easy to see that tightness of the set $M_{0}$ implies validity of the condition in the lemma. Suppose this condition holds. For arbitrary $\varepsilon>0$ and each $n\in\mathbb{N}$ let $\mathcal{K}_{n}=\mathcal{K}\left(\varepsilon
2^{-n}, \varepsilon 2^{-n}\right)$. Then for the compact set $\mathcal{K}=\bigcap_{n\in\mathbb{N}}U_{\varepsilon
2^{-n}}(\mathcal{K}_{n})$ we have $$\mu(\mathcal{A}\setminus\mathcal{K})\leq\sum_{n=1}^{+\infty}\mu(\mathcal{A}\setminus
U_{\varepsilon 2^{-n}}(\mathcal{K}_{n}))
<\sum_{n=1}^{+\infty}\varepsilon 2^{-n}<\varepsilon$$ for all $\mu\in M_{0}$, which means that the set $M_{0}$ is tight. $\square$
It is well known that an arbitrary lower semicontinuous lower bounded function on a metric space can be represented as a pointwise limit of some increasing sequence of continuous bounded functions [@12]. By using the above Theorem and the generalized Vesterstrom-O’Brien theorem mentioned in Remark 1 this observation can be strengthened as follows.
**Corollary.** *An arbitrary lower semicontinuous lower bounded convex (correspondingly, concave) function on a stable $\mu$-compact set $\mathcal{A}$ can be represented as a pointwise limit of some increasing sequence of convex (correspondingly, concave) continuous bounded functions.*
**Remark 2.** In [@3] the weaker version of the $\mu$-compactness property of a set $\mathcal{A}$ defined by the requirement of compactness of the set $M_{x}(\mathcal{A})$ for each $x$ in $\mathcal{A}$ is considered. This property called *pointwise $\mu$-compactness* was used to show that even slight relaxing of the $\mu$-compactness condition in the generalized Vesterstrom-O’Brien theorem leads to breaking its validity. The class of pointwise $\mu$-compact sets is wider than the class of $\mu$-compact sets, in particular, it contains the simplex $\left\{\{x_{i}\}_{i=1}^{+\infty}\,|\,x_{i}\geq0,\,\forall i,\,
\sum_{i=1}^{+\infty}x_{i}\leq 1\right\}\subset\ell_{p}$ for any $p\geq1$, which is $\mu$-compact only for $p=1$ [@3 Proposition 13]. For an arbitrary convex pointwise $\mu$-compact set the assertions of the Krein-Milman theorem and of the Choquet theorem are valid [@3 Proposition 5], but representation (\[co-rep\]) for the convex closure of a lower semicontinuous lower bounded function $f$ does not hold in general [@3 Example 1].
Similar to a convex $\mu$-compact set a convex pointwise $\mu$-compact set can be characterized in terms of properties of functions defined on this set. Namely, one can show that *the following properties are equivalent:*
1. *the set $\mathcal{A}$ is pointwise $\mu$-compact;*
2. *for an arbitrary increasing sequence $\{f_{n}\}\subset
C(\mathcal{A})$, converging pointwise to a function $f_{0}\in
C(\mathcal{A})$, the sequence $\{\mathrm{co}f_{n}\}$ converges pointwise to the function $\,\mathrm{co}f_{0}$.*
Since $$\mathrm{co}f_{n}(x)=\inf_{\mu\in
M^{f}_{x}(\mathcal{A})}\int_{\mathcal{A}} f_{n}(y) \,\mu(dy),\quad
x\in\mathcal{A},\quad n=0,1,2...,$$ where $M^{f}_{x}(\mathcal{A})$ is a subset of $M_{x}(\mathcal{A})$ consisting of finitely supported measures, the implication $\,\mathrm{(i)\Rightarrow(ii)}\,$ in the above assertion is proved by noting that pointwise $\mu$-compactness of $\mathcal{A}$ and Prokhorov’s theorem implies tightness of $\,M^{f}_{x}(\mathcal{A})\,$ and by using Dini’s lemma. The implication $\,\mathrm{(ii)\Rightarrow(i)}\,$ can be established by using the proof of the implication $\,\mathrm{(ii)\Rightarrow(i)}\,$ in the Theorem with $x_{k}=x_{0}$ for all $k$.
The above characterization of a convex pointwise $\mu$-compact set, the Theorem and Corollary 2 in [@3] show that $$\{\textrm{pointwise}\; \mu\textup{-compactness\,of}\;\mathcal{A}
\}\wedge\{\mathrm{co}f=\overline{\mathrm{co}}f\;\forall f\in
C(\mathcal{A})\}=\{\mu\textup{-compactness\,of}\;\mathcal{A}\}.$$
The implication $\,\mathrm{(i)\Rightarrow(iii)}\,$ in the Theorem can be used in study of entropic characteristics of quantum states [@10 §6.2].
[99]{}
E.Alfsen, “Compact convex sets and boundary integrals”, Springer Verlag, 1971.
J.Vesterstrom, “On open maps, compact convex sets and operator algebras”, J. London Math. Soc. V.6, N.2, P.289-297, 1973.
R.O’Brien, “On the openness of the barycentre map”, Math. Ann., V.223, N.3, P.207-212, 1976.
M.E.Shirokov, “On the strong CE-property of convex sets”, Mathematical Notes, V.82, N.3, P.395-409, 2007.
V.Yu.Protasov, M.E.Shirokov, “Generalized compactness in linear spaces and its applications”, Sbornik:Mathematics V.200, N.5, P.697-722, 2009; arXiv:1002.3610 \[math-ph\].
K.Parthasarathy, “Probability measures on metric spaces”, Academic Press, New York and London, 1967.
A.D.Joffe, W.M.Tikhomirov, “Theory of extremum problems”, AP, NY, 1979.
N.N.Vahania, V.I.Tarieladze, “Covariant operators of probability measures in locally convex spaces”, Theory of Probability and its Applications, V.23, N.1, P.1-23, 1978.
C.D.Aliprantis, K.C.Border, “Infinite dimensional analysis”, Springer Verlag, 2006.
A.S.Holevo, “Statistical structure of quantum theory”, Springer Verlag, 2001.
S.Papadopoulou, “On the geometry of stable compact convex sets”, Math.Ann. V.229, P.193-200, 1977.
M.E.Shirokov, “On properties of the space of quantum states and their application to construction of entanglement monotones”, Izvestiya: Mathematics, V.74, N.4, 2010; arXiv:0804.1515 \[math-ph\].
[^1]: e-mail:[email protected]
[^2]: This means that the topology on the set $\mathcal{A}$ is defined by a countable subset of the family of seminorms, generating the topology of the entire locally convex space, and this set is separable and complete in the metric generated by this subset of seminorms.
[^3]: The convex closure of a function is also called the lower (convex) envelope of this function [@1].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Humans have an incredible ability to process and understand information from multiple sources such as images, video, text, and speech. Recent success of deep neural networks has enabled us to develop algorithms which give machines the ability to understand and interpret this information. There is a need to both broaden their applicability and develop methods which correlate visual information along with semantic content. We propose a unified model which jointly trains on images and captions, and learns to generate new captions given either an image or a caption query. We evaluate our model on three different tasks namely cross-modal retrieval, image captioning, and sentence paraphrasing. Our model gains insight into cross-modal vector embeddings, generalizes well on multiple tasks and is competitive to state of the art methods on retrieval.'
address: |
Rochester Institute of Technology\
Rochester NY, USA
bibliography:
- 'egbib.bib'
title: 'Show, Translate and Tell'
---
Introduction
============
Image and text understanding has seen significant progress with the proliferation of convolutional and recurrent neural networks. These neural networks have accomplished outstanding results when applied to individual tasks such as image captioning, cross-modal retrieval and visual-question answering. In each of these tasks, domain transformation is learnt to transfer information between images and text. Off-the-shelf pre-trained networks [@He2015] have been used widely to extract features which represent the objects and their relationships in an image.
Multi-task learning [@devlin2018bert] [@zheng2018same] has recently been applied to natural language processing. By training on multiple tasks jointly, a model learns abstract representations that are task agnostic. This approach can effectively be used as pre-training and is shown to improve many natural language processing tasks [@devlin2018bert]. We investigate this approach and apply it to vision and language tasks namely cross-modal retrieval, image captioning and sentence paraphrasing. Our model learns generalized latent representations of image and text. This approach reduces both inference time and memory requirements when compared to task-specific models.
\[stt\_model\] {width="1.0\linewidth"}
The main contributions of this work include:
- We propose a unified model that jointly trains on images and captions and learns to generate new captions given either an image or text as query.
- We apply this multi-task model to the three different tasks namely cross-modal retrieval, image captioning and sentence paraphrasing.
- We leverage an attention mechanism in this multi-task model and demonstrate improved performance on these tasks.
- We open source our implementation at <https://github.com/peri044/STT>.
Related Work {#sec:literature_review}
============
Recent progress in deep learning has enabled significant advancements in understanding the relationships between visual and language entities. Most of the works focus on extracting advanced deep features and trying to map them to a specific tasks such as image captioning, phrase localization, and cross-modal retrieval.
Aviv et al. [@eisenschtat2017linking] used two way neural networks to optimize euclidean loss between images and text in a common embedding space. Vendrov et al. [@vendrov2015order] proposed to use an order-violation penalty in margin-based ranking loss [@karpathy2015deep] to enforce constraint on the order in which embeddings are learned. In particular, they only use the absolute value of image and text embeddings and use margin-based loss to optimize the model. Faghri et al. [@faghri2017vse++] proved hard negative mining can be useful and showed significant improvements on cross modal retrieval problems. Wehrmann et al. [@brm] proposed to use convolutional text encoders and perform convolutions over characters as opposed to words. They use an embedding matrix for characters and show significant reduction in the number of parameters of the model. You et al. [@cse] proposed to use a local loss along with a global loss to train the image embeddings. Yan et al. [@huang2018learning] proposed to use a multi-label CNN to predict semantic concepts in the image. They use an LSTM network as a sentence encoder to represent sentences and apply margin based ranking loss to bring images and sentences into a common embedding space. Lee et al. [@lee2018stacked] proposed a novel attention mechanism to align image regions with the individual words in a sentence. They compute the attention scores as a similarity metric and optimize the margin based ranking loss.
In this work, we propose a unified model which learns general purpose representations for both images and text which can be applied in the context of multi-task learning.
Show, Translate and Tell
========================
Show, Translate and Tell (STT) is a unified model, which projects images and captions into a common embedding space, and also learns to decode these embeddings into meaningful representations. This approach offers an interesting and insightful way to interpret the semantics of these embeddings.
Method
------
Figure \[fig:stt\_main\_fig\] describes the high-level architecture. During training, images and captions are encoded using Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNNs) respectively. These features are projected into a Common Vector Space (CVS) by using fully connected layers. To align similar concepts of images and text together and map dissimilar concepts far apart in this CVS, we follow [@karpathy2015deep] and enforce a margin based ranking loss to optimize the model parameters. This loss ensures that the model learns the semantic relationships between images and captions. The margin-based ranking loss is given by (1).
$$\begin{array}{r}
L_{rank} = \sum\limits_{m}\sum\limits_{k} max(0, \alpha + s(i,c_k) - s(i,c)) \\ \\+ \sum\limits_{k}\sum\limits_{m} max(0, \alpha + s(c,i_m) - s(c,i))
\end{array}$$
where $s(i,c)$ denotes the cosine similarity between correlated image-caption pair and $s(i, c_k)$ denotes the cosine similarity between uncorrelated image-caption pair. $i$ and $c$ denote the image and caption embeddings obtained from the image and text encoders. This loss forces the similarity of correlated pairs to be higher than uncorrelated pairs by a minimum margin $\alpha$.
**Decoding Image embeddings -** To ensure that the image embeddings are closer to the captions in CVS, we decode the image embeddings using a second RNN (on the right of the Figure 1) into semantically similar captions. We use cross entropy loss at each time step of the RNN, given by (\[ic\_loss\]).
$$\label{ic_loss}
\begin{array}{r}
L_{IC} = -\sum\limits_{t=1}^{N} log P(w_t|I; \theta)
\end{array}$$
where ${P(w_t)}$ is the probability of predicting word ${w_t}$ at timestep $t$, $I$ is the image embeddings from CVS and $\theta$ denotes the parameters of the RNN and the image encoder.
**Decoding Sentence embeddings -** In order to ensure sentence embeddings have semantic meaning, we also decode these embeddings using the same RNN (on the right of the Figure 1), into semantically similar captions. We use cross entropy loss at each timestep of the RNN, given by (\[sp\_loss\]).
$$\label{sp_loss}
\begin{array}{r}
L_{SP} = -\sum\limits_{t=1}^{N} log P(w_t|S; \theta)
\end{array}$$
where ${P(w_t)}$ is the probability of predicting word ${w_t}$ at timestep $t$, $S$ is the input sentence embeddings from CVS and $\theta$ denotes the parameters of the RNN and the sentence encoder. The weights of this RNN are shared during the decoding of image and sentence embeddings.
Combining the above components, the STT model is jointly trained by optimizing the overall objective function in (\[total\_loss\]).
$$\label{total_loss}
\begin{array}{r}
L = \lambda_1L_{rank} + \lambda_2L_{IC} + \lambda_3L_{SP}
\end{array}$$
where $\lambda_1$, $\lambda_2$ and $\lambda_3$ are scalar weights which regulate the importance of individual loss components.
STT model with Attention
------------------------
In order to align fine-grain information in the STT model, we follow [@lee2018stacked] to incorporate attention between image regions and individual words in the sentence. Object proposals are extracted using Faster R-CNN [@girshick2015fast] and top $N$ region proposals are passed through a Resnet-152 CNN for feature extraction. Captions are encoded by an LSTM network and the outputs at individual time steps are collected. The importance of each word over the $N$ regions is calculated by cosine distance and the similarity metric $s(i,c)$ is computed as the aggregate of all the word vectors and regions. During training with attention, we use the average of $N$ region level embeddings as an input to the decoder during image captioning.
Implementation
==============
We use TensorFlow [@abadi2016tensorflow] package throughout our experiments. We resize the images into $256\times256$ and extracted random crops of $224\times224$ for data augmentation. We use Resnet-152 CNN [@he2016deep] as a feature extractor for images and 1-layer LSTM [@hochreiter1997long] network to encode and decode sentences. We use one fully connected layer for image and sentence embeddings. We pre-compute the ResNet-152 [@he2016deep] features and train the image embedding, RNN encoder and decoder modules for 15 epochs. The learning rate is set to 0.0002 and Adam optimizer is used to optimize the parameters of the model. The margin parameter and batch size was set to 0.2 and 128 respectively. For attention models, we follow [@lee2018stacked] and choose $N$ to be 36.
Experiments
===========
We perform extensive experiments to evaluate the proposed model. We choose standard datasets, MSCOCO [@lin2014microsoft] and Flickr30K for training and evaluation. We use 113,287 images from MSCOCO and 29,783 images from Flickr30K for training, 5000 and 1000 images for evaluation. During training, a single sample constitutes one image and two sentence paraphrases. Since each image in [@lin2014microsoft] and [@flickr30k] is annotated with five captions, we form a total of 20 permutations of paraphrases for each image in the dataset. We report recall ($R@K$) as the evaluation metric for retrieval experiments. $R@K$ is the percentage of query samples in which the ground-truth sentences belong to the top $K$ retrieved sentences. For MSCOCO, we report the average of 5-fold cross validation results on 5000 test images. We use BLEU [@papineni2002bleu] and METEOR [@banerjee2005meteor] scores for captioning and paraphrasing experiments.
Cross Modal Retrieval
---------------------
Cross Modal Retrieval is the task of retrieving similar samples given an input query sample. The input query can either be an image or a caption. The images and captions in the test set are passed through image and caption encoders in Figure \[stt\_model\] to extract their corresponding embeddings.
-------------------------------- ---------- ---------- ---------- ----------
Model
R@1 R@10 R@1 R@10
UVS [@uvs] 43.4 85.8 31.0 79.9
Order [@vendrov2015order] 46.7 88.9 37.9 85.9
Aviv [@eisenschtat2017linking] 55.8 - 39.7 -
BRM [@brm] 55.1 93.9 41.2 89.2
CSE [@cse] 56.3 91.5 45.7 90.6
VSE++ [@faghri2017vse++] 58.3 93.3 43.6 87.8
STT **55.1** **92.1** **41.0** **86.0**
STT with att **64.9** **96.8** **49.8** **91.6**
FBB [@engilberge2018finding] 69.8 96.6 55.9 94.0
SCO [@huang2018learning] 69.9 97.5 56.7 94.8
SCAN [@lee2018stacked] 72.7 98.4 55.8 94.8
-------------------------------- ---------- ---------- ---------- ----------
: Results of Cross Modal Retrieval on MSCOCO.
\[result\_mscoco\]
------------------------------ ---------- ---------- ---------- ----------
Model
R@1 R@10 R@1 R@10
UVS [@uvs] 29.8 70.5 22 59.3
BRM [@brm] 36.0 - 31.0 -
VSE++ [@faghri2017vse++] 43.7 82.1 32.3 87.8
CSE [@cse] 44.6 83.8 36.9 79.6
Order [@vendrov2015order] 46.7 88.9 37.9 85.9
STT **38.4** **77.5** **27.1** **68.2**
FBB [@engilberge2018finding] 46.5 82.2 34.9 73.5
SCO [@huang2018learning] 55.5 89.3 41.1 80.1
STT with att **59.2** **91.0** **40.7** **79.0**
SCAN [@lee2018stacked] 67.9 94.4 43.9 82.8
------------------------------ ---------- ---------- ---------- ----------
: Results of Cross Modal Retrieval on Flickr30k.
\[result\_flickr\]
Tables \[result\_mscoco\] and \[result\_flickr\] show cross modal retrieval results of STT model on MSCOCO and Flickr30K respectively. From Table \[result\_mscoco\] and \[result\_flickr\], it is clear that STT model performs well on cross modal retrieval task. STT with attention performs better than all other models on Flickr30K [@flickr30k], except for SCAN [@lee2018stacked]. While SCAN [@lee2018stacked] does great at cross modal retrieval, our model can additionally do captioning and paraphrasing.
Image Captioning
----------------
Images are passed through the encoders and the corresponding embeddings are passed through an RNN decoder in Figure \[fig:stt\_main\_fig\] to generate new captions. Since the model learned to map semantics from different modalities, these embeddings can generate meaningful sentence representations. Table \[im\_capt\] shows the results of image captioning on a test set of 5000 and 1000 images respectively. We use the same test set for retrieval and captioning tasks.
[|r|r|r|r|r|r|]{} Dataset & B@1 & B@2 & B@3 & B@4 & METEOR\
\
& 0.683 & 0.506 & 0.362 & 0.259 & 0.236\
& 0.513 & 0.330 & 0.204 & 0.129 & 0.178\
\
& 0.706 & 0.530 & 0.385 & 0.279 & 0.246\
& 0.611 & 0.427 & 0.293 & 0.203 & 0.193\
Sentence Paraphrasing
---------------------
Similar to captioning, we can use the RNN encoder and decoder to generate paraphrases. The input sentence is passed through an RNN encoder and the generated embeddings are passed through an RNN decoder in Figure \[stt\_model\] to generate a caption with similar meaning. The test set for MSCOCO has 5000 images which results in a total of 25000 ground truth captions. Flickr30k has a total of 5000 ground truth captions (1000 images). To the best of our knowledge, we are the first ones to show sentence paraphrasing in a multi-task setting between vision and language. Table \[sent\_par\] shows the results of sentence paraphrasing on MSCOCO and Flickr30K.
[|r|r|r|r|r|r|]{} Dataset & B@1 & B@2 & B@3 & B@4 & METEOR\
\
& 0.744 & 0.578 & 0.435 & 0.324 & 0.275\
& 0.569 & 0.394 & 0.262 & 0.176 & 0.217\
\
& 0.747 & 0.581 & 0.436 & 0.326 & 0.272\
& 0.673 & 0.493 & 0.353 & 0.252 & 0.221\
\[stt\_model\] ![Sample result of STT model on MSCOCO dataset.[]{data-label="fig:stt_coco"}](mscoco1.png "fig:"){width="1.0\linewidth"}
\[stt\_model\] ![Sample result of STT model with Attention on MSCOCO dataset.[]{data-label="fig:stt_coco_2"}](stt_att_mscoco_2.png "fig:"){width="1.0\linewidth"}
\[stt\_model\] ![Sample result of STT model on Flickr30K dataset.[]{data-label="fig:stt_f30k"}](stt_f30k.png "fig:"){width="1.0\linewidth"}
Discussion
----------
Figures \[fig:stt\_coco\] and \[fig:stt\_coco\_2\] show some sample results of the STT model with attention on MSCOCO. The STT model captures context and relationships between images and captions. All possible permutations of paraphrases were used for training and the model learns to associate image regions to individual words better compared to the model trained without attention. This is evident in Table \[im\_capt\]. The addition of attention improved BLEU [@papineni2002bleu] and METEOR [@banerjee2005meteor] scores for the captioning task. As the number of samples is higher in MSCOCO [@lin2014microsoft], STT performs better compared to Flickr30K [@flickr30k]. Figure \[fig:stt\_f30k\] shows a sample output of STT on Flickr30K [@flickr30k] dataset. Although this particular example is a failed retrieval case, the retrieved captions describe the image well. Flickr30K [@flickr30k] is a challenging dataset due to high correlation between captions of different images. However, our model exhibits good generalization and diversity in such challenging scenarios.
Conclusion
==========
We present a novel multi-task model which learns general purpose embeddings for images and captions. This model captures the semantic relationships between vision and language modalities during training and offers an effective way to interpret the intermediate latent representations. We leverage recent attention mechanisms for further performance boosts. We believe this is the first effort to show sentence paraphrasing in a multi-task setting between vision and language. We evaluate our model on standard benchmark datasets and demonstrate good performance on cross-modal retrieval, image captioning and sentence paraphrasing tasks.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Wim Beenakker\
Theoretical High Energy Physics, IMAPP, Faculty of Science, Mailbox 79, Radboud University Nijmegen, P.O. Box 9010, NL-6500 GL Nijmegen, The Netherlands
- |
Silja Brensing\
DESY, Theory Group, Notkestrasse 85, D-22603 Hamburg, Germany
- |
Michael Krämer, Anna Kulesza\
Institute for Theoretical Particle Physics and Cosmology, RWTH Aachen University\
D-52056 Aachen, Germany
- |
Eric Laenen\
ITFA, University of Amsterdam, Valckenierstraat 65, 1018 XE Amsterdam,\
ITF, Utrecht University, Leuvenlaan 4, 3584 CE Utrecht\
Nikhef Theory Group, Science Park 105, 1098 XG Amsterdam, The Netherlands
- |
Irene Niessen\
Theoretical High Energy Physics, IMAPP, Faculty of Science Mailbox, 79, Radboud University Nijmegen, P.O. Box 9010, NL-6500 GL Nijmegen, The Netherlands
title: 'NNLL resummation for squark-antisquark pair production at the LHC'
---
Introduction {#s:intro}
============
The search for supersymmetry (SUSY) [@Golfand:1971iw; @Wess:1974tw] is a central part of the experimental program at the Large Hadron Collider (LHC). Models of weak-scale SUSY provide a promising solution to the hierarchy problem of the Standard Model (SM) and predict new supersymmetric particles (sparticles) with masses in the TeV range. If they exist, the coloured sparticles, squarks ($\tilde q$) and gluinos ($\tilde g$), would be produced copiously in hadronic collisions and thus offer the strongest sensitivity for supersymmetry searches at the LHC. In the context of the minimal supersymmetric extension of the Standard Model (MSSM) [@Nilles:1983ge; @Haber:1984rc], these particles are produced in pairs due to $R$-parity conservation.
Searches for squarks and gluinos at the proton-proton collider LHC, which has been operating at $\sqrt{S}=7$ TeV in 2010 and 2011, have placed lower limits on squark and gluino masses around 1 TeV [@Aad:2011ib; @Chatrchyan:2011zy]. Once the LHC reaches its design energy of $\sqrt{S}=14$ TeV, SUSY particles with masses up to 3 TeV can be probed [@Aad:2009wy; @Bayatian:2006zz].
Accurate theoretical predictions for inclusive squark and gluino cross sections are needed both to set exclusion limits and, in case SUSY is discovered, to determine SUSY particle masses and properties [@Baer:2007ya; @Dreiner:2010gv]. The inclusion of higher-order SUSY-QCD corrections significantly reduces the renormalization- and factorization-scale dependence of the predictions. In general, the corrections also increase the size of the cross section with respect to the leading-order prediction [@Kane:1982hw; @Dawson:1983fw] if the renormalization and factorization scales are chosen close to the average mass of the produced SUSY particles. Consequently, the SUSY-QCD corrections have a substantial impact on the determination of mass exclusion limits and would lead to a significant reduction of the uncertainties on SUSY mass or parameter values in the case of discovery. The squark-antisquark production processes have been known for quite some time at next-to-leading order (NLO) in SUSY-QCD [@Beenakker:1994an; @Beenakker:1996ch]. Electroweak corrections to the ${\cal O} (\alpha_{\rm s}^2)$ tree-level processes [@Hollik:2008yi] and the electroweak Born production channels of ${\cal O} (\alpha\alpha_{\rm s})$ and ${\cal O} (\alpha^2)$ [@Bornhauser:2007bf] are in general significant for the pair production of SU(2)-doublet squarks $\tilde{q}_L$ and at large invariant masses, but they are moderate for inclusive cross sections.
A significant part of the NLO QCD corrections can be attributed to the threshold region, where the partonic centre-of-mass energy is close to the kinematic production threshold. In this region the NLO corrections are dominated by soft-gluon emission off the coloured particles in the initial and final state and by the Coulomb corrections due to the exchange of gluons between the slowly moving massive sparticles in the final state. The soft-gluon corrections can be taken into account to all orders in perturbation theory by means of threshold resummation techniques [@Sterman:1986aj; @Catani:1989ne]. The Coulomb corrections can be summed to all orders by either using a Sommerfeld factor [@BBPC:BBPC19400460520; @Sakharov:1948yq; @Catani:1996dj; @Fadin:1990wx] or by employing the framework of non-relativistic QCD, where bound-state effects can be included as well [@Fadin:1990wx; @Kiyo:2008bv; @Hagiwara:2008df; @Hagiwara:2009hq; @Kauth:2011bz; @Kauth:2011vg]. In addition, a formalism has been developed in the framework of effective field theories that allows for the combined resummation of soft and Coulomb gluons in the production of coloured sparticles [@Beneke:2009rj; @Beneke:2010da].
Threshold resummation has been performed for all MSSM squark and gluino production processes at next-to-leading-logarithmic (NLL) accuracy [@Kulesza:2008jb; @Kulesza:2009kq; @Beenakker:2009ha; @Beenakker:2010nq; @Beenakker:2011fu]. For squark-antisquark production, in addition to soft-gluon resummation, the Coulomb corrections have been resummed both by using a Sommerfeld factor [@Kulesza:2009kq] and by employing the framework of effective field theories [@Beneke:2010da]. Furthermore, the dominant next-to-next-to-leading order (NNLO) corrections, including those coming from the resummed cross section at next-to-next-to-leading-logarithmic (NNLL) level, have been calculated for squark-antisquark pair-production [@Langenfeld:2009eg].
In this paper we consider threshold resummation at NNLL accuracy for squark-antisquark pair production at the LHC. Compared to the NLL calculation the new ingredients are the one-loop matching coefficients, which contain the NLO cross section near threshold, and the two-loop soft anomalous dimensions. Studies for the pair production of top quarks suggest that the effect of the matching coefficients can be significant [@Bonciani:1998vc] and that NNLL resummation can reduce the scale dependence considerably [@Kidonakis:2001nj; @Langenfeld:2009eg; @Kidonakis:2010dk; @Ahrens:2010zv; @Beneke:2011mq]. We will discuss the impact of the corrections and provide an estimate of the theoretical uncertainty due to scale variation. In addition we will study the impact of the Coulomb gluons on the cross section. We exclude top squarks from the final state in view of potentially large mixing effects and mass splitting in the stop sector [@Ellis:1983ed]. The other squarks are considered as being mass degenerate and all flavours and chiralities are summed over.
The structure of the paper is as follows. In section \[s:resummation\] we discuss the NNLL resummation for squark-antisquark pair-production. In section \[s:Ccoeff\] we present the calculation of the hard matching coefficients required for the NNLL resummation. The numerical results are presented in section \[s:numres\]. We show predictions for the LHC with centre-of-mass energies of $\sqrt{S}=7$ TeV and $\sqrt{S}=14$ TeV. We will conclude in section \[s:conclusion\].
Threshold resummation at NNLL {#s:resummation}
=============================
In this section we briefly review the formalism of threshold resummation for the production of a squark-antisquark pair. The inclusive hadroproduction cross section $\sigma_{h_1h_2\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}}}$ can be written in terms of its partonic version $\sigma_{ij\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}}}$ as $$\begin{gathered}
\label{eq:7}
\sigma_{h_1 h_2 \to {{\tilde{q}}}{{\bar{\tilde{q}}}}}\bigl(\rho, \{m^2\}\bigr)
\;=\; \sum_{i,j} \int d x_1 d x_2\,d\hat{\rho}\;
\delta\left(\hat{\rho} - \frac{\rho}{x_1 x_2}\right)\\
\times\,f_{i/h_{1}}(x_1,\mu^2 )\,f_{j/h_{2}}(x_2,\mu^2 )\,
\sigma_{ij \to {{\tilde{q}}}{{\bar{\tilde{q}}}}}\bigl(\hat{\rho},\{ m^2\},\mu^2\bigr)\,,\end{gathered}$$ where $\{m^2\}$ denotes all masses entering the calculations, $i,j$ are the initial parton flavours, $f_{i/h_1}$ and $f_{j/h_2}$ the parton distribution functions and $\mu$ is the common factorization and renormalization scale. The hadronic threshold for inclusive production of two final-state squarks with mass $m_{{\tilde{q}}}$ corresponds to a hadronic centre-of-mass energy squared that is equal to $S=4m_{{\tilde{q}}}^2$. Thus we define the threshold variable $\rho$, measuring the distance from threshold in terms of energy fraction, as $$\rho \;=\; \frac{4m_{{\tilde{q}}}^2}{S}\,.$$ The partonic equivalent of this threshold variable is defined as $\hat{\rho}=\rho/(x_1x_2)$, where $x_{1,2}$ are the momentum fractions of the partons.
In the threshold region, the dominant contributions to the higher-order QCD corrections due to soft-gluon emission have the general form $$\alpha_{\rm s}^n \log^m\!\beta^2\ \ , \ \ m\leq 2n
\qquad {\rm \ with\ } \qquad
\beta^2 \,\equiv\, 1-\hat{\rho} \,=\, 1 \,-\, \frac{4m_{{\tilde{q}}}^2}{s}\,,
\label{eq:beta}$$ where $s=x_1x_2S$ is the partonic centre-of-mass energy squared. The resummation of the soft-gluon contributions is performed after taking a Mellin transform (indicated by a tilde) of the cross section, $$\begin{aligned}
\label{eq:10}
\tilde\sigma_{h_1 h_2 \to {{\tilde{q}}}{{\bar{\tilde{q}}}}}\bigl(N, \{m^2\}\bigr)
&\equiv \int_0^1 d\rho\;\rho^{N-1}\;
\sigma_{h_1 h_2\to {{\tilde{q}}}{{\bar{\tilde{q}}}}}\bigl(\rho,\{ m^2\}\bigr) \nonumber\\
&= \;\sum_{i,j} \,\tilde f_{i/{h_1}} (N+1,\mu^2)\,
\tilde f_{j/{h_2}} (N+1, \mu^2) \,
\tilde{\sigma}_{ij \to {{\tilde{q}}}{{\bar{\tilde{q}}}}}\bigl(N,\{m^2\},\mu^2\bigr)\,.\end{aligned}$$ The logarithmically enhanced terms are then of the form $\alpha_{\rm
s}^n \log^m N$, $m\leq 2n$, with the threshold limit $\beta\rightarrow 0$ corresponding to $N\rightarrow \infty$. The all-order summation of such logarithmic terms is a consequence of the near-threshold factorization of the cross sections into functions that each capture the contributions of classes of radiation effects: hard, collinear (including soft-collinear), and wide-angle soft radiation [@Sterman:1986aj; @Catani:1989ne; @Bonciani:1998vc; @Contopanagos:1996nh; @Kidonakis:1998bk; @Kidonakis:1998nf]. Near threshold the resummed partonic cross section has the form:
$$\begin{aligned}
\label{eq:12}
\tilde{\sigma}^{\rm (res)} _{ij\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}}}\bigl(N,\{m^2\},&\mu^2\bigr)
=\sum_{I}\,
\tilde\sigma^{(0)}_{ij\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}},I}\bigl(N,\{m^2\},\mu^2\bigr)\,
C_{ij\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}},I}(N,\{m^2\},\mu^2)\nonumber\\
& \times\,\Delta_i (N+1,Q^2,\mu^2)\,\Delta_j (N+1,Q^2,\mu^2)\,
\Delta^{\rm (s)}_{ij\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}},I}\bigl(Q/(N\mu),\mu^2\bigr)\,,\end{aligned}$$
where we have introduced the hard scale $Q^2 = 4m_{{\tilde{q}}}^2$. Before commenting on the different functions in this equation, we recall that soft radiation is coherently sensitive to the colour structure of the hard process from which it is emitted [@Bonciani:1998vc; @Contopanagos:1996nh; @Kidonakis:1998bk; @Kidonakis:1998nf; @Botts:1989kf; @Kidonakis:1997gm]. At threshold, the resulting colour matrices become diagonal to all orders by performing the calculation in an $s$-channel colour basis [@Beneke:2009rj; @Kulesza:2008jb; @Kulesza:2009kq]. The different contributions then correspond to different irreducible representations $I$. For the $q\bar q\to{{\tilde{q}}}{{\bar{\tilde{q}}}}$ process, the $s$-channel basis consists of a singlet $\bf1$ and an octet $\bf8$ representation, while for the $gg\to{{\tilde{q}}}{{\bar{\tilde{q}}}}$ process it contains a singlet $\bf1$, an antisymmetric octet $\bf8_A$ and a symmetric octet $\bf8_S$ representation as presented in e.g. Ref. [@Beneke:2009rj; @Kulesza:2008jb; @Kulesza:2009kq].
In Eq. , $\tilde{\sigma}^{(0)}_{ij \rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}},I}$ are the colour-decomposed leading-order (LO) cross sections in Mellin-moment space. The functions $\Delta_{i}$ and $\Delta_{j}$ sum the effects of the (soft-)collinear radiation from the incoming partons, while the function $\Delta^{\rm (s)}_{ij\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}},I}$ describes the wide-angle soft radiation. Schematically the exponentiation of soft-gluon radiation takes the form [@Sterman:1986aj; @Catani:1989ne] $$\label{eq:3}
\Delta_i\Delta_j\Delta^{\rm(s)}_{ij\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}},I}
\;=\; \exp\Big[L g_1(\alpha_{\rm s}L) + g_2(\alpha_{\rm s}L) + \alpha_{\rm s}g_3(\alpha_{\rm s}L) + \ldots \Big] \,.$$ This exponent captures all dependence on the large logarithm $L=\log N$. Keeping only the $g_1$ term in Eq. constitutes the leading logarithmic (LL) approximation. Including also the $g_2$ term is called the NLL approximation. For the NNLL approximation also the $g_3$ term needs to be taken into account. Explicit expressions for the $g_3$ term and its ingredients are given in Refs. [@Moch:2005ba; @Czakon:2009zw; @Moch:2008qy][^1] and are listed in appendix \[app:g3\].
We also need the matching coefficients $C$, which contain the Mellin moments of the higher-order contributions without the $\log(N)$ terms. To NNLL accuracy, this non-logarithmic part of the higher-order cross section near threshold factorizes into a part that contains the leading Coulomb correction ${\cal C}^{\rm Coul,(1)}$ and a part that contains the NLO hard matching coefficients ${\cal C}^{\rm (1)}$ [@Beneke:2010da]: $$\begin{aligned}
C^{\rm NNLL}=\left(1+\frac{\alpha_{\rm s}}{\pi}\;{\cal C}^{\rm Coul,(1)}(N,\{m^2\},\mu^2)\right)\;\left(1+\frac{\alpha_{\rm s}}{\pi}\;{\cal C}^{\rm (1)}(\{m^2\},\mu^2)\right)\label{eq:matchingcoeff}\end{aligned}$$ The calculation of the NLO hard matching coefficients ${\cal C}^{\rm (1)}$ and the Coulomb contribution ${\cal C}^{\rm Coul,(1)}$ for the squark-antisquark production processes will be discussed in detail in section \[s:Ccoeff\].
Having constructed the NNLL cross-section in Mellin-moment space, the inverse Mellin transform has to be performed in order to recover the hadronic cross section $\sigma_{h_1 h_2 \to {{\tilde{q}}}{{\bar{\tilde{q}}}}}$. In order to retain the information contained in the complete NLO cross sections [@Beenakker:1996ch], the NLO and NNLL results are combined through a matching procedure that avoids double counting of the NLO terms: $$\begin{aligned}
\label{eq:matching}
\sigma^{\rm (NLO+NNLL~matched)}_{h_1 h_2 \to {{\tilde{q}}}{{\bar{\tilde{q}}}}}&\bigl(\rho, \{m^2\},\mu^2\bigr)
=\; \sigma^{\rm (NLO)}_{h_1 h_2 \to {{\tilde{q}}}{{\bar{\tilde{q}}}}}\bigl(\rho, \{m^2\},\mu^2\bigr)
\\[1mm]
& +\, \sum_{i,j}\,\int_\mathrm{CT}\,\frac{dN}{2\pi i}\,\rho^{-N}\,
\tilde f_{i/h_1}(N+1,\mu^2)\,\tilde f_{j/h_{2}}(N+1,\mu^2) \nonumber\\[2mm]
&\times\,
\left[\tilde\sigma^{\rm(res,NNLL)}_{ij\to {{\tilde{q}}}{{\bar{\tilde{q}}}}}\bigl(N,\{m^2\},\mu^2\bigr)
\,-\, \tilde\sigma^{\rm(res,NNLL)}_{ij\to {{\tilde{q}}}{{\bar{\tilde{q}}}}}\bigl(N,\{m^2\},\mu^2\bigr)
{\left.\right|}_{\scriptscriptstyle{\rm (NLO)}}\, \right]. \nonumber\end{aligned}$$ We adopt the “minimal prescription” of Ref. [@Catani:1996yz] for the contour CT of the inverse Mellin transform in Eq. (\[eq:matching\]). In order to use standard parametrizations of parton distribution functions in $x$-space we employ the method introduced in Ref. [@Kulesza:2003wn].
Calculation of the matching coefficients {#s:Ccoeff}
========================================
In this section we will discuss the calculation of the matching coefficients $C$ at one loop. As discussed in equation , the NNLL matching coefficient $C^{\rm NNLL}$ factorizes into the Coulomb contribution and the hard matching coefficient. For NNLL resummation, both terms are needed up to NLO accuracy.
The terms in the NLO cross section which give rise to the Coulomb corrections ${\cal C}^{\rm Coul,(1)}$ in $N$-space do not have the usual phase-space suppression , in view of the Coulombic [$1/\beta$]{} enhancement factor. After performing an expansion of the NLO cross section in $\beta$, the hard matching coefficients ${\cal
C}^{\rm(1)}$ are determined by the terms in the NLO cross section that are proportional to $\beta$, $\beta \log(\beta)$ and $\beta \log^2(\beta)$. Terms that contain higher powers of $\beta$ are suppressed by powers of $1/N$ in Mellin-moment space and do not contribute to the matching coefficient $C$. In contrast to the case of top-pair production in Ref. [@Czakon:2008cx], there is no full analytic result for the real corrections to squark-antisquark production, so we cannot take the explicit threshold limit. For the virtual corrections, which also contain the Coulomb contribution, we will use the full analytic expressions, but for the real corrections we need a different approach.
To obtain the virtual corrections for squark-antisquark production we start from the full analytic calculation as presented in Ref. [@Beenakker:1996ch]. As described in detail in Ref. [@Beenakker:1996ch], the QCD coupling $\alpha_{\rm s}$ and the parton distribution functions at NLO are defined in the $\overline{\rm MS}$ scheme with five active flavours, with a correction for the SUSY breaking in the $\overline{\rm MS}$ scheme. The masses of squarks and gluinos are renormalized in the on-shell scheme, and the top quark and the SUSY particles are decoupled from the running of $\alpha_{\rm s}$.
To obtain the virtual part of the hard matching coefficients, we first need to colour-decompose the result and then expand it in $\beta$. For the first step we only need the colour decomposition of the LO matrix element. Due to the orthogonality of the $s$-channel colour basis, the full matrix element squared is then automatically colour-decomposed: $$|{\cal M}|^2_{{\rm NLO},I}=2\mathrm{Re}({\cal M}_{\rm NLO}{\cal M}^*_{{\rm LO},I}).$$
We are now left with an expression in terms of masses, Mandelstam variables and scalar integrals. Since we need the cross section to ${\cal O}(\beta)$, we have to expand $|{\cal M}|^2$ to zeroth order in $\beta$. For the squark-antisquark production processes the factors that multiply the integrals do not contain negative powers of $\beta$, so we do not have to expand the scalar integrals beyond zeroth order in $\beta$.
The number of integrals that need to be expanded can be reduced. By using the fact that the two outgoing momenta are equal at threshold, we can reduce some of the three- and four-point integrals to two- and three-point integrals respectively. This procedure can be used only for integrals that contain both outgoing momenta. The result of the remaining integrals is explicitly expanded to zeroth order in $\beta$.
Special attention has to be paid to the Coulomb integrals. First, in order to calculate the Coulomb corrections ${\cal C}^{\rm Coul,(1)}$ in $N$-space, we need to know the Coulomb part of the NLO correction in $\beta$-space, corresponding to the leading terms in $\beta$ of the Coulomb integrals. These leading terms are given by [@BBPC:BBPC19400460520; @Sakharov:1948yq; @Catani:1996dj; @Fadin:1990wx]:
$$\begin{aligned}
\sigma_{ij\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}},I}^{\rm Coul,(1)}=-\frac{\alpha_{\rm s}}{\pi}\;\frac{\pi^2}{2\beta}\kappa_{ij\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}},I}\sigma_{ij\rightarrow {{\tilde{q}}}{{\bar{\tilde{q}}}},I}^{\rm(0)}\label{eq:coulomb}\end{aligned}$$
with $\kappa$ colour coefficients that depend on the process and the dimension of the representation. For the $q\bar q$-initiated process they are given by [@Kulesza:2009kq]: $$\kappa_{q\bar q\to{{\tilde{q}}}{{\bar{\tilde{q}}}},\bf1}=-\frac{4}{3} \quad\mathrm{and}\quad \kappa_{q\bar q\to{{\tilde{q}}}{{\bar{\tilde{q}}}},\bf8}=\frac{1}{6}$$ while for the $gg$-initiated process they are: $$\kappa_{gg\to{{\tilde{q}}}{{\bar{\tilde{q}}}},\bf1}=-\frac{4}{3} ,\quad \kappa_{gg\to{{\tilde{q}}}{{\bar{\tilde{q}}}},\bf8_A}=\frac{1}{6}\quad\mathrm{and}\quad \kappa_{gg\to{{\tilde{q}}}{{\bar{\tilde{q}}}},\bf8_S}=\frac{1}{6}\;.$$ The Mellin transform $\tilde\sigma^{\rm Coul,(1)}$ of Eq. is presented in appendix \[app:mellin\]. The function ${\cal C}^{\rm Coul,(1)}$ can be obtained by dividing $\tilde\sigma^{\rm Coul,(1)}$ by the Mellin transform of the LO cross section, which can be found in Ref. [@Kulesza:2009kq].
Secondly, the next term in the $\beta$-expansion of the Coulomb integrals contributes to the hard matching coefficients. Due to their $1/\beta$ behaviour Coulomb integrals cannot be reduced to lower-point integrals, so they need to be expanded explicitly.
To obtain the integrated real corrections at threshold, the key observation is that they are formally phase-space suppressed near threshold unless the integrand compensates this suppression. Therefore we can construct the real corrections at threshold from the singular behaviour of the matrix element squared, which can be obtained using dipole subtraction [@Catani:1996vz; @Catani:2002hc]. We will briefly review the procedure of dipole subtraction and specify how only the singular contributions survive in the threshold limit.
Dipole subtraction makes use of the fact that the cross section can be split into three parts: a part with three-particle kinematics $\sigma^{\{3\}}$, one with two-particle kinematics $\sigma^{\{2\}}$, and a collinear counterterm $\sigma^C$ that is needed for removing the initial-state collinear singularities. These parts are well-defined in $n=4-2\epsilon$ dimensions, but their constituents diverge for $\epsilon\to0$. With the aid of an auxiliary cross section $\sigma^A$, which captures all singular behaviour, all parts are made finite and integrable in four space-time dimensions. This auxiliary cross section is subtracted from the real corrections $\sigma^{\rm R}$ at the integrand level to obtain $\sigma^{\{3\}}$ and added to the virtual corrections $\sigma^{\rm V}$, which defines $\sigma^{\{2\}}$: $$\sigma^{\rm NLO}=\int_{3}\big[{\mathrm{d}}\sigma^{\rm R}-{\mathrm{d}}\sigma^{\rm A}\big]_{\epsilon=0}+\int_{2}\big[{\mathrm{d}}\sigma^{\rm V}+\int_1{\mathrm{d}}\sigma^{\rm A}\big]_{\epsilon=0}+\sigma^{\rm C}\equiv\sigma^{\{3\}}+\sigma^{\{2\}}+\sigma^{\rm C}$$ We will first argue that we can neglect $\sigma^{\{3\}}$. Compared to the case of two-parton kinematics, the phase space of $\sigma^{\{3\}}$ is limited by the energy of the third, radiated massless particle. Near the two-particle threshold, the maximum energy of the radiated particle, and thus the available phase space, equals $E_{\rm max}=\sqrt{s}-2m_{{\tilde{q}}}\propto\beta^2$. Since after subtracting $\sigma^A$ no divergences are left in the integrand of $\sigma^{\{3\}}$, the leading contribution of $\sigma^{\{3\}}$ is at most proportional to $\beta^2$ and can thus be neglected. This leaves us with: $$\sigma^{\rm NLO,thr}=\sigma^{\{2\},\rm thr}+\sigma^{\rm C,\rm thr}=\sigma^{\rm V,thr}+\sigma^{\rm C,thr}+\sigma^{\rm A,thr},$$ so at threshold the real radiation is indeed completely specified by the singular behaviour contained in $\sigma^A$. In Ref. [@Catani:2002hc] the general form of $\sigma^A$ is determined by summing over dipoles that correspond to pairs of ordered partons. These dipoles describe the soft and collinear radiation and reproduce the matrix element squared in the soft and collinear limits. To obtain the cross section, the dipole functions need to be integrated over phase space and in particular over the momentum fraction $x$ that is left after radiation. In the threshold limit the available phase space sets the lower bound of the $x$-integral to $1-\beta^2$, while the upper bound equals 1. Therefore we cannot get a result of ${\cal O}(\beta)$ unless the integrand diverges at $x=1$, which is the case only for soft-gluon radiation. As a result we only need to take into account the dipoles that describe gluon radiation.
Special attention has to be paid to the massive final-state dipole function. In Ref. [@Catani:2002hc] this dipole function has been rewritten in order to simplify the integration. Unfortunately this results in a deformation of the phase space integration which changes exactly the finite terms that we are looking for. Therefore the expression given in Eq. (5.16) of Ref. [@Catani:2002hc] cannot be used for our calculation and we have to use the original dipole function instead. A more detailed argument can be found in appendix \[app:vfactors\]. The divergent part of the dipole function is completely determined by the soft limit. Using the eikonal approximation we obtain the final-state dipole function that correctly reproduces the soft limit at threshold. Its behaviour is given by: $$\left.\langle{\bf V}_{gj,l}\rangle\right|_{\rm div}\propto\sum_{j}\frac{1}{p_g\cdot p_j}\sum_{l\ne j}\left(\frac{p_j\cdot p_l}{p_g\cdot p_j+p_g\cdot p_l}-\frac{1}{2}\frac{m_j^2}{p_g\cdot p_j}\right)T_j\cdot T_l\;,\label{eq:dipole}$$ where $p_g$ is the gluon momentum and the sums run over final-state particles with momenta $p_{j,l}$, masses $m_{j,l}$ and colour charge operators $T_{j,l}$ that are defined in Ref. [@Catani:2002hc]. This expression vanishes at threshold, so the final-state dipoles do not contribute in the threshold limit. For the other dipole functions and the collinear counterterm we can use the equations in [@Catani:2002hc] and take the threshold limit.
After combining the real and the virtual corrections the hard matching coefficients can be obtained by taking the Mellin transform and omitting the Coulomb corrections and the $\log(N)$ terms. The complete expressions for the hard matching coefficients of the squark-antisquark production processes can be found in appendix \[app:Ccoeff\]. Their behaviour for varying gluino mass is shown in Fig. \[fig:Ccoeffplots\]. For the $gg\to\tilde q\bar{\tilde q}$ process the antisymmetric octet $\bf8_A$ contribution to the cross section vanishes because it yields a $p$-wave contribution, which vanishes at threshold.
These matching coefficients have been checked numerically using [PROSPINO]{} [@Beenakker:1996ch] and agree within the numerical accuracy of [PROSPINO]{}. They also agree with the $a_1$ coefficients presented in Ref. [@Langenfeld:2009eg] to the percent level.
Numerical results {#s:numres}
=================
In this section we present numerical results for the NNLL-resummed cross sections with and without the Coulomb contributions. We show the results for squark-antisquark pair-production at the LHC for centre-of-mass energies of 7 TeV and 14 TeV. In order to evaluate hadronic cross sections we use the 2008 NLO MSTW parton distribution functions [@Martin:2009iq] with the corresponding $\alpha_{\rm s}(M_{Z}^2) = 0.120$. We have used a top quark mass of $m_t=172.9$ GeV [@PDG]. The numerical results have been obtained with two independent computer codes.
It should be noted that the Coulomb effects can be screened by the width of the sparticles depending on the specific SUSY scenario. For consistency we will stick to the approach adopted in the NLO calculations, where this screening is not taken into account. In order to study the effects from the hard matching coefficients and the Coulomb corrections separately, we will compare several cross sections with the NLO result and discuss their contribution:
- The NLL matched cross section is based on the calculations presented in [@Beenakker:2009ha; @Kulesza:2008jb; @Kulesza:2009kq] and will be denoted as $\sigma^{\rm NLO+NLL}$.
- The NNLL matched cross section without Coulomb contributions to the resummation $\sigma^{\rm NLO+NNLL\; w/o\; Coulomb}$ contains the soft-gluon resummation to NNLL accuracy matched to the full NLO result. The matching is performed according to Eq. (\[eq:matching\]). The Coulomb correction to the resummation is not included, so ${\cal C}^{\rm Coul,(1)}$ in Eq. is set to zero.
- The NNLL matched cross section $\sigma^{\rm NLO+NNLL}$ does include the Coulomb contribution ${\cal C}^{\rm Coul,(1)}$ from equation . Also in this case Eq. (\[eq:matching\]) has been used to match the cross section to the complete NLO result.
The NLO cross sections are calculated using the publicly available [PROSPINO]{} code [@prospino], based on the calculations presented in Ref. [@Beenakker:1996ch]. The QCD coupling $\alpha_{\rm s}$ and the parton distribution functions at NLO are defined in the $\overline{\rm MS}$ scheme with five active flavours. The masses of squarks and gluinos are renormalized in the on-shell scheme, and the SUSY particles are decoupled from the running of $\alpha_{\rm s}$ and the parton distribution functions. No top-squark final states are considered. We sum over squarks with both chiralities ($\tilde{q}_{L}$ and $\tilde{q}_{R}$), which are taken as mass degenerate. The renormalization and factorization scales $\mu$ are taken to be equal.
We first discuss the scale dependence of the cross sections. Figure \[fig:scale\_dep\] shows the squark-antisquark cross section for $m_{{{\tilde{q}}}}=m_{{{\tilde{g}}}}=1.2$ TeV as a function of the renormalization and factorization scale $\mu$. The value of $\mu$ is varied around the central scale $\mu_0 = m_{{\tilde{q}}}$ from $\mu=\mu_0/5$ up to $\mu=5 \,\mu_0$ and results are shown for the LHC at CM energies of 7 TeV (a) and 14 TeV (b).
For both collider energies we see the usual scale reduction going from LO to NLO. Including the NLL correction and the NNLL contribution without the Coulomb part ${\cal C}^{\rm Coul,(1)}$ improves the behaviour for moderate values of $\mu/\mu_0$, but a fairly strong scale dependence for small values of $\mu/\mu_0$ remains. Upon inclusion of the Coulomb corrections ${\cal C}^{\rm Coul,(1)}$ the scale dependence stabilises over the whole range.
Figure \[fig:scale\_unc\] shows the mass dependence of the scale uncertainty for the NLO, NLO+NLL and NLO+NNLL cross sections at the LHC. The squark and gluino mass have been taken equal and the scale has been varied in the range $m_{{\tilde{q}}}/2\le\mu\le 2m_{{\tilde{q}}}$. As was to be expected from figure \[fig:scale\_dep\], the scale uncertainty reduces as the accuracy of the predictions increases. In the range of squark masses considered here, the NNLL resummation without the Coulomb corrections ${\cal C}^{\rm Coul,(1)}$ already reduces the scale uncertainty to at most 10% for the LHC at a CM energy of 7 TeV and to even lower values for a CM energy of 14 TeV. The inclusion of the Coulomb term ${\cal C}^{\rm Coul,(1)}$ in the resummed NNLL prediction results in a scale uncertainty of only a few percent for both collider energies. The effect of the threshold resummation is more pronounced for a collider energy of 7 TeV, which is due to the fact that the sparticles are produced closer to threshold in that case.
Finally we study the $K$-factors with respect to the NLO cross section: $$K_x=\frac{\sigma^x}{\sigma^{\rm NLO}}~,$$ where $x$ can be NLO+NLL, NLO+NNLL w/o Coulomb or NLO+NNLL. In figure \[fig:K\] we study the mass dependence of the $K$-factor for equal squark and gluino masses.
At the central scale $\mu=m_{{\tilde{q}}}$ the $K$-factor, and thus the theoretical prediction of the cross section, increases as more corrections are included. Also, the effect becomes more pronounced for higher masses. This was to be expected, since in that case the particles are produced closer to threshold. As can be seen in figure \[fig:K\](a), the NNLL resummation without the Coulomb corrections ${\cal C}^{\rm Coul,(1)}$ already results in a 25% increase of the cross section with respect to the NLO cross section for squarks of 2 TeV and a CM energy of 7 TeV. The contribution from the Coulomb term to the resummed NLO+NNLL cross section is larger than the contributions provided by the $g_3$ term in the exponential and the hard matching coefficient ${\cal C}^{\rm (1)}$, yielding a total $K$-factor of 1.45. For the case of a CM energy of 14 TeV, which is shown in figure \[fig:K\](b), the size of the NNLL contributions is smaller, since the sparticles are produced further away from threshold. However, for masses of 3 TeV, the $K$-factor for the NNLL contribution without the Coulomb correction ${\cal C}^{\rm Coul,(1)}$ still yields 1.13, whereas the inclusion of the Coulomb corrections increases this to 1.25. Although the effect from the Coulomb corrections could be somewhat smaller in reality due to the finite lifetime of the squarks, figure \[fig:K\] suggests that the NNLL contribution will remain large.
Figure \[fig:K\] only contains the numbers for equal squark and gluino masses, but the effect of the gluino mass is small, as can be seen in figure \[fig:K-r\]. In figure \[fig:K-r\] the mass ratio $r=m_{{\tilde{g}}}/m_{{\tilde{q}}}$ has been varied. Although some effect can be seen, it is negligible compared to the size of the NNLL corrections. It turns out that this conclusion also holds for a collider energy of 14 TeV. Consequently the NNLL-resummed results are relatively independent of the relation between squark and gluino masses.
The scale dependence of the cross section shows the best stability after including both the hard matching coefficients ${\cal C}^{(1)}$ and the Coulomb contributions ${\cal C}^{\rm Coul,(1)}$. This indicates that all these contributions should be taken into account to achieve the observed cancellation, see also figure 9 in [@Beneke:2010da]. However, the observed reduction in the scale dependence might be modified somewhat by the inclusion of the width of the particles or by matching to the full NNLO result, which is not available. In this context we note that, as a consequence of the NNLL accuracy of resummation, our matched cross section receives additional non-logarithmic NNLO contributions, which would have been consistently treated if matching to NNLO had been possible. A very conservative estimate of the scale uncertainty is provided by the NLO+NNLL w/o Coulomb results, which do not include the Coulomb corrections.\
Conclusions {#s:conclusion}
===========
We have performed the NNLL resummation of threshold corrections for squark-antisquark hadroproduction. In particular, the previously unknown hard matching coefficient ${\cal C}^{\rm(1)}$, needed at this level of accuracy, has been calculated analytically. We have also numerically evaluated the NNLL resummed cross section, matched to the NLO fixed-order expression, for squark-antisquark production at the LHC with CM energies of 7 and 14 TeV. At both collider energies the total cross section increases at the central scale. At 7 TeV collision energy and for a squark mass of 2 TeV, the NLO+NNLL squark-antisquark cross section is larger than the corresponding NLO cross section by as much as 45%. The correction is reduced to 25% if the contributions due to Coulombic interactions are not taken into account. In addition, the scale dependence is reduced significantly, particularly after inclusion of the Coulomb corrections. This information should be used to improve current limits on SUSY masses or, in the case that SUSY is found, to more accurately determine the masses of the sparticles.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work has been supported in part by the Helmholtz Alliance “Physics at the Terascale”, the Foundation for Fundamental Research of Matter (FOM), program 104 “Theoretical Particle Physics in the Era of the LHC", the DFG SFB/TR9 “Computational Particle Physics”, and the European Community’s Marie-Curie Research Training Network under contract MRTN-CT-2006-035505 “Tools and Precision Calculations for Physics Discoveries at Colliders”.
The NNLL functions {#app:g3}
==================
In the following we list the explicit expression for the NNLL function $g^{(3)}_{ij\to {{\tilde{q}}}\bar{{\tilde{q}}},I}$, cf. Refs. [@Moch:2005ba; @Moch:2008qy]. Expressions for the LL and NLL functions can be found in Refs. [@Bonciani:1998vc; @Kulesza:2009kq].
The NNLL functions $g^{(3)}_{q\bar q\to {{\tilde{q}}}\bar{{\tilde{q}}},I}$ and $g^{(3)}_{gg\to {{\tilde{q}}}\bar{{\tilde{q}}},I}$ read $$\begin{aligned}
g^{(3)}_{ij\to {{\tilde{q}}}\bar{{\tilde{q}}},I}\left(\lambda, 4m_{\tilde q}^2,\mu^2 \right) &=&
\frac{A^{(1)}_i b_1^2}{\pi b_0^4}\frac{1}{1-2\lambda}\left[2\lambda^2+2\lambda \log(1-2\lambda)+\frac{1}{2}\log^2(1-2\lambda)\right]\\ {\nonumber}&& +\frac{A^{(1)}_i b_2}{\pi b_0^3}\left[2\lambda+\log(1-2\lambda)+\frac{2 \lambda^2}{1-2\lambda}\right]\\ {\nonumber}&& -\frac{2A^{(1)}_i b_1 \gamma_E}{\pi b_0^2}\frac{\left[2\lambda+\log(1-2\lambda)\right]}{1-2\lambda}+\frac{4 A^{(1)}_i}{\pi} \left(\zeta(2)+\gamma_E^2 \right)\frac{\lambda}{1-2\lambda}\\ {\nonumber}&& -\frac{ A^{(2)}_i b_1}{\pi^2 b_0^3}\frac{1}{1-2\lambda}\left[2\lambda(\lambda+1)+\log(1-2\lambda)\right]\\ {\nonumber}&& +\frac{4 A^{(2)}_i \gamma_E}{\pi^2 b_0} \frac{\lambda}{1-2\lambda}+\frac{2 A^{(3)}_i}{\pi^3 b_0^2} \frac{\lambda^2}{1-2\lambda}-\frac{D^{(2)}_i}{\pi^2 b_0}\frac{\lambda}{1-2\lambda} \\ {\nonumber}&& +\frac{D^{(1)}_{ij\to {{\tilde{q}}}\bar{{\tilde{q}}},I}\, b_1}{2 \pi b_0^2 }\frac{\left[2\lambda+\log(1-2\lambda)\right]}{1-2\lambda}-\frac{2 D^{(1)}_{ij\to {{\tilde{q}}}\bar{{\tilde{q}}},I}\,\gamma_E}{\pi}\frac{\lambda}{1-2\lambda}\\ {\nonumber}&& -\frac{D^{(2)}_{ij\to {{\tilde{q}}}\bar{{\tilde{q}}},I}}{\pi^2 b_0}\frac{\lambda}{1-2\lambda}\\ {\nonumber}&& +\left[\frac{A^{(1)}_i b_1}{\pi b_0^2}\frac{2\lambda+\log(1-2\lambda)}{1-2\lambda}-\frac{4 A^{(1)}_i\gamma_E}{\pi}\frac{\lambda}{1-2\lambda}\right] \log\left(\frac{4m_{\tilde q}^2}{\mu^2}\right)\\ {\nonumber}&& +\frac{A^{(1)}_i}{\pi}\frac{\lambda}{1-2\lambda}\log^2\left(\frac{4m_{\tilde q}^2}{\mu^2}\right)-\frac{2 A^{(2)}_i}{\pi^2 b_0}\frac{\lambda}{1-2\lambda}\log\left(\frac{4m_{\tilde q}^2}{\mu^2}\right)\\ {\nonumber}&& +\frac{D^{(1)}_{ij\to {{\tilde{q}}}\bar{{\tilde{q}}},I}}{\pi}\frac{\lambda}{1-2\lambda}\log\left(\frac{4m_{\tilde q}^2}{\mu^2}\right)\end{aligned}$$ with $\lambda=b_0\alpha_{\rm s}(\mu^2) \log(N) $, $\mu$ the common renormalization and factorization scale, and $\gamma_E$ Euler’s constant. The coefficients of the QCD beta function are denoted by $b_n$ and the first three coefficients are given by [@Tarasov:1980au; @Larin:1993tp] $$\begin{aligned}
b_0&=&\frac{11 C_A-2 n_l}{12 \pi}\,,\\ {\nonumber}b_1&=&\frac{17 C_A^2-5C_A n_l -3C_F n_l}{24 \pi^2}\,,\\ {\nonumber}b_2&=&\frac{1}{(4 \pi)^3}\left[\frac{2857}{54}C_A^3-\frac{1415}{54}C_A^2 n_l -\frac{205}{18} C_A C_F n_l+C_F^2 n_l
+\frac{79}{54}C_A n_l^2 +\frac{11}{9}C_F n_l^2 \right]\,,\\ {\nonumber}\end{aligned}$$ where $n_l$ denotes the number of light quark flavours, $C_A=N_c$ and $C_F=\frac{N_c^2-1}{2N_c}$ with $N_c$ the number of colours. The universal, process independent coefficients up to NNLL accuracy are given by [@Moch:2005ba] $$\begin{aligned}
A^{(1)}_i & = & C_i\,, \\ {\nonumber}A^{(2)}_i & = & \frac{1}{2} C_i \left[ \left( \frac{67}{18}
- \zeta(2) \right) C_A - \frac{5}{9} n_l \right]\,, \\ {\nonumber}A^{(3)}_i & = & \frac{1}{4} C_i \left[ C_A^{2} \left( \frac{245}{24} - \frac{67}{9}\zeta(2) + \frac{11}{6}\zeta(3)
+ \frac{11}{5}\zeta(2)^{2} \right)
+ C_F n_l \left( -\frac{55}{24} + 2\zeta(3)\right) \right. \\ {\nonumber}& & \left. \hspace{2cm}
+\, C_A n_l\, \left( - \frac{209}{108}+ \frac{10}{9}\zeta(2)- \frac{7}{3}\zeta(3) \right) - \frac{n_l^2}{27} \right]\,,\end{aligned}$$ and [@Contopanagos:1996nh; @Vogt:2000ci; @Catani:2003zt] $$\begin{aligned}
D^{(2)}_i&=&C_i\left[C_A \left(-\frac{101}{27}+ \frac{11}{3}\zeta(2)+ \frac{7}{2}\zeta(3)\right)+n_l \left(\frac{14}{27}- \frac{2}{3}\zeta(2) \right) \right]\,,\end{aligned}$$ with the colour factor $C_i=C_F$ for $i=q,\bar{q}$ and $C_i=C_A$ for $i=g$. The process dependent coefficients read [@Bonciani:1998vc] $$\begin{aligned}
{D^{(1)}_{q\bar q(gg)\to {{\tilde{q}}}\bar{{\tilde{q}}},\bf1}}=0\,, \hspace{1cm } {D^{(1)}_{q\bar q(gg)\to {{\tilde{q}}}\bar{{\tilde{q}}},\bf8(8A,8S)}}=-C_A\,,\end{aligned}$$ and [@Czakon:2009zw; @Beneke:2010da] $$\begin{aligned}
{D^{(2)}_{q\bar q(gg)\to {{\tilde{q}}}\bar{{\tilde{q}}},\bf1}}&=&0\,, \\ {\nonumber}{D^{(2)}_{q\bar q(gg)\to {{\tilde{q}}}\bar{{\tilde{q}}},\bf8(8A,8S)}}&=&-\,C_A\left(\frac{1}{2} \left[ \left( \frac{67}{18}
- \zeta(2) \right) C_A - \frac{5}{9} n_l \right]+\frac{C_A}{2}(\zeta(3)-1)+2 \pi b_0 \right)\,.\end{aligned}$$
Mellin transforms of the Coulomb corrections {#app:mellin}
============================================
In this appendix we present the analytical results for the Mellin transforms of the Coulomb corrections in terms of the Euler beta function $\beta$, the digamma function $\Psi$ and the hypergeometric functions $_2F_1$ and $_3F_2$. For the subprocess $q_{i} \bar q_{j} \to {{\tilde{q}}}\bar{{\tilde{q}}}\,$ the expressions for the colour-decomposed Coulomb part in $N$-space are given by $$\begin{aligned}
\tilde\sigma^{\mathrm{Coul},(1)}_{q_{i} \bar q_{j} \to {{\tilde{q}}}\bar{{\tilde{q}}}\,,\bf1}(N)
&=& -\frac{4\alpha_{\rm s}^3\pi^2}{243 m_{{{\tilde{q}}}}^2}
\Bigg[ \frac{4}{N+1}-\frac{4 h}{N+2}\, {_{2}}F_1\left(1,N+2,N+3,h\right) \\ {\nonumber}&& \hspace{1.8cm} +\,4\, I_N(r)+2(r^2-1)\,I_{N+1}(r) \Bigg]\,,\\[3mm]
\tilde\sigma^{\mathrm{Coul},(1)}_{q_{i} \bar q_{j} \to {{\tilde{q}}}\bar{{\tilde{q}}}\,,\bf8}(N)
&=& -\frac{\alpha_{\rm s}^3\pi^2}{2592 m_{{{\tilde{q}}}}^2}\delta_{ij}
\Bigg[ \frac{8 n_l}{N^2+3N+2}+\frac{8}{N+1}+(r^2-1)\frac{4}{N+2} \\ {\nonumber}&& \hspace{2.5cm} +\,4\,r^2\, I_{N+1}(r)+(r^2-1)^2\,I_{N+2}(r) \Bigg]\,- \frac{1}{64}\tilde\sigma^{\mathrm{Coul},(1)}_{q_{i} \bar q_{j} \to {{\tilde{q}}}\bar{{\tilde{q}}}\,,\bf1}(N)\,,\end{aligned}$$ whereas for the subprocess $\,gg \to {{\tilde{q}}}\bar{{\tilde{q}}}\,\,$ they read $$\begin{aligned}
\tilde\sigma^{\mathrm{Coul},(1)}_{gg \to {{\tilde{q}}}\bar{{\tilde{q}}}\,,\bf1}(N)
&=& \frac{\alpha_{\rm s}^3\pi^2 n_l}{288 m_{{{\tilde{q}}}}^2}
\Bigg[ \frac{2}{N+1}+\frac{2}{N+2}\, +\,2\, I_{N+1}(1)-I_{N+2}(1) \Bigg]\,,\\[3mm]
\tilde\sigma^{\mathrm{Coul},(1)}_{gg \to {{\tilde{q}}}\bar{{\tilde{q}}}\,,\bf8A}(N)
&=& -\frac{\alpha_{\rm s}^3\pi^2 n_l}{1536 m_{{{\tilde{q}}}}^2}
\Bigg[ \frac{2}{N+1}+\frac{16}{N+2}\, +\,6\, I_{N+1}(1)+3\,I_{N+2}(1) \Bigg]\,,\\[3mm]
\tilde\sigma^{\mathrm{Coul},(1)}_{gg \to {{\tilde{q}}}\bar{{\tilde{q}}}\,,\bf8S}(N)&=&-\frac{5}{16}\tilde\sigma^{\mathrm{Coul},(1)}_{gg \to {{\tilde{q}}}\bar{{\tilde{q}}}\,,\bf1}(N)\,.\end{aligned}$$ Here $r=m_{\tilde g}/m_{\tilde q}$ and $n_l=5$ denotes the number of light quark flavours. The function $I_{N}(r)$ is given by $$\begin{aligned}
I_{N}(r) &\equiv& \frac{2\sqrt d}{N+1}+2 B_N \log{\left(\frac{r^2+1}{2r}\right)}+B_N\left[\Psi(N+3/2)-\Psi(N+1)\right]\\ {\nonumber}&& +h\,\frac{N+1}{N+3/2}\, B_N\,\, {_3} F_2(1,1,N+2,2,N+5/2,h)\\ {\nonumber}&&-\frac{d}{2(N+3/2)} B_N\,\, {_3}F_2(1,1,3/2,2,N+5/2,d)\\ {\nonumber}&& +\frac{2\, d^{3/2}}{3\left(N+1\right)\left(N+2\right)}\,\, {_3}F_2(1,3/2,2,5/2,N+3,d)\\ {\nonumber}&& -\frac{2^{N+2}}{(N+1)^2}\,\, {_3}F_2(-N,N+1,N+1,N+2,N+2,1/2)\,,\end{aligned}$$ with $$\begin{aligned}
h=-\frac{(r^2-1)^2}{4r^2}, \hspace{1.5cm} \mathrm{and} \hspace{1cm} d=\frac{(r^2-1)^2}{(r^2+1)^2}.\end{aligned}$$ The Euler beta function $\beta (N+1, 1/2)$ is abbreviated by $$\begin{aligned}
B_{N} &\equiv& \beta (N+1, 1/2).\end{aligned}$$
How velocity factors can deform the phase space integration {#app:vfactors}
===========================================================
As mentioned in section \[s:Ccoeff\] the dipole function given in Ref. [@Catani:2002hc] has been modified, which makes it unsuitable for our calculation. We will explicitly show the effect of this modification by considering the change in the second term of Eq. , which is denoted as $I^{\mathrm{coll}}_{gQ,Q}$ in Eq. (5.23) of Ref. [@Catani:2002hc]. In Ref. [@Catani:2002hc] finite pieces of the integrand are taken into account as well, but since we showed in section \[s:Ccoeff\] that the only contribution at threshold comes from the singular part of the integrand, we will omit these terms.
The singular term of the integrand yields a $1/\epsilon$-pole and a finite piece. The pole cancels the pole of the first term of the dipole function , while the finite piece contributes to the hard matching coefficient. In its unmodified form, the finite piece is given by $$\begin{aligned}
\left.I^{\mathrm{coll,unchanged}}_{gQ,Q}\right|_{\mathrm{fin}}&=2\int_0^{y_+}\mathrm{d}y\left[\frac{1}{y}-\frac{\mu^2\sqrt{[2\mu^2+(1-2\mu^2)(1-y)]^2-4\mu^2}}{y(\mu^2+y(1-2\mu^2))\sqrt{1-4\mu^2}}\right]\nonumber\\
&\approx2\int_0^{2(1-2\mu)}\mathrm{d}y\frac{\sqrt{1-4\mu^2}-\sqrt{(1-y/2)^2-4\mu^2}}{y\sqrt{1-4\mu^2}}\,\label{eq:Icoll}\end{aligned}$$ with $$y=\frac{p_g\cdot p_j}{p_g\cdot p_j+p_g\cdot p_l+p_j\cdot p_l},\qquad y_+=\frac{1-2\mu}{1-2\mu^2}\quad\mbox{and}\quad\mu=m/\sqrt s\;.$$ The approximation in the second line of Eq. is suitable near threshold, where $\mu\approx1/2$. Exactly at threshold the finite part equals $4-4\log(2)$ and exactly cancels the contribution from the first term of Eq. .
In Ref. [@Catani:2002hc], the integrand in $I^{\mathrm{coll}}_{gQ,Q}$ has been multiplied by velocity factors in order to simplify the integration: $$\frac{\tilde v_{gQ,Q}}{v_{gQ,Q}}=\frac{(1-y)\sqrt{1-4\mu^2}}{\sqrt{[2\mu^2+(1-2\mu^2)(1-y)]^2-4\mu^2}}\approx\frac{\sqrt{1-4\mu^2}}{\sqrt{(1-y/2)^2-4\mu^2}}\;,$$ where the approximation in the second step holds near threshold. The velocity factor effectively replaces $\sqrt{1-4\mu^2}$ in the denominator of Eq. by $\sqrt{(1-y/2)^2-4\mu^2}$, which amounts to a shift comparable in size to the value of the numerator. In the strict soft limit $y$ vanishes and the velocity factors have no effect. However, we are integrating over gluons that are not soft compared to the energy above threshold $\sqrt{s}-2m$, so we also need the correct behaviour away from the strict soft limit. In fact, if the velocity factors are included the integral $I^{\mathrm{coll}}_{gQ,Q}$ vanishes at threshold, so it does no longer cancel the contribution from the first term of Eq. .
Usually the velocity factors do not pose a problem in calculations using dipole subtraction, since the terms are subtracted from the real part and added to the virtual part. Therefore it does not matter if a dipole function is deformed, as long as the pole is reproduced. Finite contributions can always be moved between $\sigma^{\{2\}}$ and $\sigma^{\{3\}}$. However, we argued that $\sigma^{\{3\}}$ vanishes at threshold due to phase-space suppression, which is not true if the phase space integration is deformed by velocity factors. Therefore we need the unchanged dipole function for this particular calculation.
The hard matching coefficients for squark-antisquark production {#app:Ccoeff}
===============================================================
Here we present the exact expressions for the hard matching coefficients ${\cal C}^{\rm(1)}$ for the squark-antisquark production processes as defined in Eq. . We sum over squarks with both chiralities ($\tilde{q}_{L}$ and $\tilde{q}_{R}$). No top-squark final states are considered and all squarks are considered to be mass-degenerate with mass $m_{{\tilde{q}}}$. Top squarks are taken into account in the loops, where they are taken to be mass-degenerate with the other squarks. The calculation is outlined in section \[s:Ccoeff\] and was done with FORM [@Vermaseren:2000nd]. We first define: $$\beta_{12}(q^2)=\sqrt{1 - \frac{4m_1m_2}{q^2 - (m_1 - m_2)^2}}~,\qquad x_{12}(q^2)=\frac{\beta_{12}(q^2)-1}{\beta_{12}(q^2)+1}$$ and $$m_-^2=m_{\tilde g}^2-m_{\tilde q}^2,\qquad m_+^2=m_{\tilde g}^2+m_{\tilde q}^2,$$ where $m_{{\tilde{g}}}$ is the gluino mass. Denoting the number of light flavours by $n_l=5$, the total number of flavours by $n_f=6$ and the number of colours by $N_c=3$, we also define: $$\begin{aligned}
\gamma_q&=\frac{3}{2}C_F&C_F&=\frac{N_c^2-1}{2N_c}\\
\gamma_g&=\frac{11}{6}C_A-\frac{1}{3}n_l&C_A&=N_c\,.\end{aligned}$$ We denote the factorization scale by $\mu_F$, the renormalization scale by $\mu_R$ and Euler’s constant by $\gamma_E$. For the $q\bar q\to\tilde q\bar{\tilde q}$ process both the singlet $\bf1$ and the octet $\bf8$ representation contribute: $$\begin{aligned}
{\cal C}_{q\bar q\to\tilde q\bar{\tilde q},I}^{(1)}&=\mathrm{Re}\Bigg\{\frac{2C_F}{3}\pi^2+\gamma_g\log\bigg(\frac{\mu_R^2}{m_{\tilde q}^2}\bigg)-\gamma_q\log\bigg(\frac{\mu_F^2}{m_{\tilde q}^2}\bigg)+F_0(m_{\tilde q},m_{\tilde g},m_t)+\frac{19N_c}{24}\\
&\quad+\frac{23}{8N_c}+\frac{1-3N_c^2}{N_c}\log\bigg(\frac{m_+^2}{m_{\tilde q}^2}\bigg)-\frac{2}{N_c}\log\left(2\right)+\bigg(\frac{7N_c}{6}+\frac{2m_{\tilde g}^2}{m_+^2}C_F\bigg)\log\bigg(\frac{m_{\tilde g}^2}{m_{\tilde q}^2}\bigg)\\
&\quad-\frac{m_{\tilde g}^2}{m_{\tilde q}^2}\bigg(\frac{m_-^2}{m_{\tilde q}^2}\log\bigg(\frac{m_-^2}{m_{\tilde g}^2}\bigg)+1\bigg)C_F+\frac{m_{\tilde g}^2}{2m_-^2}\bigg(\frac{m_{\tilde g}^2}{m_-^2}\log\bigg(\frac{m_{\tilde g}^2}{m_{\tilde q}^2}\bigg)-1\bigg)C_F\\
&\quad-\frac{1}{2N_c}\bigg(\frac{m_{\tilde g}^2}{m_{\tilde q}^2}-3\bigg)F_1\left(m_{\tilde q},m_{\tilde g}\right)+\bigg(\frac{m_+^2}{2m_{\tilde q}^2}C_F+\frac{1}{N_c}\bigg)F_2\left(m_{\tilde q},m_{\tilde g}\right)\\
&\quad+2C_F\bigg(\gamma_E^2 - 2\gamma_E\log(2) + \gamma_E\log\bigg(\frac{\mu_F^2}{m_{{\tilde{q}}}^2}\bigg)\bigg)\\
&\quad+\bigg[-\frac{\pi^2}{4}+\log\bigg(\frac{m_+^2}{m_{\tilde q}^2}\bigg)-\log\left(2\right)-\frac{m_{\tilde g}^2}{m_+^2}\log\bigg(\frac{m_{\tilde g}^2}{m_{\tilde q}^2}\bigg)+2+\gamma_E\\
&\quad-\frac{1}{4}\bigg(\frac{m_{\tilde g}^2}{m_{\tilde q}^2}-3\bigg)\left(F_1\left(m_{\tilde q},m_{\tilde g}\right)+F_2\left(m_{\tilde q},m_{\tilde g}\right)\right)\bigg]C_2(I)\Bigg\}.\end{aligned}$$ In this equation the last two lines are proportional to the quadratic Casimir invariant of the representation, which is zero for the singlet and $N_c$ for the octet representation. We have defined the functions: $$\begin{aligned}
F_0(m_{\tilde q},m_{\tilde g},m_t)&=\frac{m_t^2}{2m_{\tilde g}^2}-\bigg(1+\frac{m_{\tilde q}^2}{2m_{\tilde g}^2}\bigg)n_f+\bigg(\frac{m_-^6}{2m_+^2m_{\tilde g}^4}\log\bigg(\frac{m_-^2}{m_{\tilde q}^2}\bigg)+\frac{4m_{\tilde q}^2}{m_+^2}\log\left(2\right)\!\bigg)n_l\\
&\quad+\bigg(\frac{m_t^4}{2m_{\tilde q}^2m_{\tilde g}^2}-\frac{(m_{\tilde q}^2-m_t^2)^2}{4m_{\tilde g}^4}+\frac{m_{\tilde q}^2-m_t^2}{m_{\tilde g}^2}-\frac{1}{12}\bigg)\log\bigg(\frac{m_t^2}{m_{\tilde q}^2}\bigg)\\
&\quad-\frac{m_-^2\big(m_{\tilde g}^2-(m_{\tilde q}-m_t)^2\big)\big(m_{\tilde g}^2-m_{\tilde q}^2+m_t^2\big)}{2m_{\tilde g}^4m_+^2}\beta_{\tilde qt}(m_{\tilde g}^2)\log\left(x_{\tilde qt}(m_{\tilde g}^2)\right)\\
&\quad+\frac{m_t^4-2m_{\tilde q}m_t^3+4m_{\tilde q}^3m_t-4m_{\tilde q}^4}{m_{\tilde q}^2m_+^2}\beta_{\tilde qt}(-m_{\tilde q}^2)\log\left(x_{\tilde qt}(-m_{\tilde q}^2)\right)\\
F_1(m_{\tilde q},m_{\tilde g})&=\mathrm{Li}_2\bigg(\frac{m_-^2}{2m_{\tilde g}^2}\bigg)+\mathrm{Li}_2\bigg(1-\frac{m_-^2}{2m_{\tilde q}^2}\bigg)+\frac{\pi^2}{12}+\log\bigg(\frac{m_-^2}{2m_{\tilde g}^2}\bigg)\log\bigg(\frac{m_+^2}{2m_{\tilde q}^2}\bigg)\\
&\quad+\frac{1}{2}\log^2\!\bigg(\frac{m_{\tilde g}^2}{m_{\tilde q}^2}\bigg)\\
F_2(m_{\tilde q},m_{\tilde g})&=\mathrm{Li}_2\bigg(\frac{m_{\tilde q}^2}{m_{\tilde g}^2}\bigg)-\mathrm{Li}_2\bigg(-\frac{m_{\tilde q}^2}{m_{\tilde g}^2}\bigg)+\log\bigg(\frac{m_+^2}{m_-^2}\bigg)\log\bigg(\frac{m_{\tilde g}^2}{m_{\tilde q}^2}\bigg)\,.\end{aligned}$$ For the $gg\to\tilde q\bar{\tilde q}$ process the antisymmetric octet $\bf8_A$ vanishes because it yields a $p$-wave contribution, which vanishes at threshold. The hard matching coefficients for the singlet $\bf1$ and the symmetric octet $\bf8_S$ do contribute: $$\begin{aligned}
{\cal C}_{gg\to\tilde q\bar{\tilde q},\bf8_A}^{(1)}&=0\\
{\cal C}_{gg\to\tilde q\bar{\tilde q},I}^{(1)}&=\mathrm{Re}\Bigg\{\pi^2\bigg(\frac{5N_c}{12}-\frac{C_F}{4}\bigg)+\gamma_g\log\bigg(\frac{\mu_R^2}{\mu_F^2}\bigg)-\frac{m_{\tilde g}^2N_c}{2m_{\tilde q}^2}\log^2\left(x_{\tilde g\tilde g}(4m_{\tilde q}^2)\right)\\
&\quad+C_F\bigg(\frac{m_+^2m_-^2}{2m_{\tilde q}^4}\log\bigg(\frac{m_+^2}{m_-^2}\bigg)-\frac{m_{\tilde g}^2}{m_{\tilde q}^2}-3\bigg)+\frac{m_+^2N_c}{2m_{\tilde q}^2}\bigg(\mathrm{Li}_2\bigg(\!-\!\frac{m_{\tilde q}^2}{m_{\tilde g}^2}\bigg)-\mathrm{Li}_2\bigg(\frac{m_{\tilde q}^2}{m_{\tilde g}^2}\bigg)\bigg)\\
&\quad+2C_A\bigg(\gamma_E^2 - 2\gamma_E\log(2) + \gamma_E\log\bigg(\frac{\mu_F^2}{m_{{\tilde{q}}}^2}\bigg)\bigg)\\
&\quad+\bigg[\frac{\pi^2}{8}-\frac{1}{2}\mathrm{Li}_2\bigg(\!\!-\!\frac{m_{\tilde q}^2}{m_{\tilde g}^2}\bigg)+\frac{1}{2}\mathrm{Li}_2\bigg(\frac{m_{\tilde q}^2}{m_{\tilde g}^2}\bigg)+\frac{m_{\tilde g}^2}{4m_{\tilde q}^2}\log^2\!\left(x_{\tilde g\tilde g}(4m_{\tilde q}^2)\right)+2+\gamma_E\bigg]C_2(I)\Bigg\}\end{aligned}$$ where in the second equation the representation $I$ can be the $\bf1$ or the $\bf8_S$ and the last line is proportional to the quadratic Casimir invariant of the representation.
[10]{}
Y. Golfand and E. Likhtman, [*[Extension of the Algebra of Poincare Group Generators and Violation of p Invariance]{}*]{}, [*JETP Lett.*]{} [**13**]{} (1971) 323–326. See also <http://www.jetpletters.ac.ru/ps/717/article_11110.shtml> Russian version.
J. Wess and B. Zumino, [*[Supergauge Transformations in Four-Dimensions]{}*]{}, [*Nucl.Phys.*]{} [**B70**]{} (1974) 39–50.
H. P. Nilles, [*[Supersymmetry, Supergravity and Particle Physics]{}*]{}, [ *Phys. Rept.*]{} [**110**]{} (1984) 1–162.
H. E. Haber and G. L. Kane, [*[The Search for Supersymmetry: Probing Physics Beyond the Standard Model]{}*]{}, [*Phys. Rept.*]{} [**117**]{} (1985) 75–263.
, G. Aad [*et. al.*]{}, [*Search for squarks and gluinos using final states with jets and missing transverse momentum with the ATLAS detector in sqrt(s) = 7 TeV proton-proton collisions*]{}, \[[[arXiv:1109.6572]{}](http://xxx.lanl.gov/abs/1109.6572)\].
, S. Chatrchyan [*et. al.*]{}, [*Search for Supersymmetry at the LHC in Events with Jets and Missing Transverse Energy*]{}, \[[[arXiv:1109.2352]{}](http://xxx.lanl.gov/abs/1109.2352)\].
, G. Aad [*et. al.*]{}, [ *[Expected Performance of the ATLAS Experiment - Detector, Trigger and Physics]{}*]{}, tech. rep., 2009.
, G. Bayatian [*et. al.*]{}, [*[CMS technical design report, volume II: Physics performance]{}*]{}, [*J.Phys.G*]{} [**G34**]{} (2007) 995–1579.
H. Baer, V. Barger, G. Shaughnessy, H. Summy, and L.-t. Wang, [*[Precision gluino mass at the LHC in SUSY models with decoupled scalars]{}*]{}, [ *Phys.Rev.*]{} [**D75**]{} (2007) 095010, \[[[hep-ph/0703289]{}](http://xxx.lanl.gov/abs/hep-ph/0703289)\].
H. K. Dreiner, M. Kramer, J. M. Lindert, and B. O’Leary, [*[SUSY parameter determination at the LHC using cross sections and kinematic edges]{}*]{}, [ *JHEP*]{} [**1004**]{} (2010) 109, \[[[ arXiv:1003.2648]{}](http://xxx.lanl.gov/abs/1003.2648)\].
G. L. Kane and J. Leveille, [*[Experimental Constraints on Gluino Masses and Supersymmetric Theories]{}*]{}, [*Phys.Lett.*]{} [**B112**]{} (1982) 227.
S. Dawson, E. Eichten, and C. Quigg, [*[Search for Supersymmetric Particles in Hadron - Hadron Collisions]{}*]{}, [*Phys.Rev.*]{} [**D31**]{} (1985) 1581.
W. Beenakker, R. Hopker, M. Spira, and P. Zerwas, [*[Squark production at the Tevatron]{}*]{}, [*Phys.Rev.Lett.*]{} [**74**]{} (1995) 2905–2908, \[[[hep-ph/9412272]{}](http://xxx.lanl.gov/abs/hep-ph/9412272)\].
W. Beenakker, R. Hopker, M. Spira, and P. M. Zerwas, [*[Squark and gluino production at hadron colliders]{}*]{}, [*Nucl. Phys.*]{} [**B492**]{} (1997) 51–103, \[[[ hep-ph/9610490]{}](http://xxx.lanl.gov/abs/hep-ph/9610490)\].
W. Hollik and E. Mirabella, [*[Squark anti-squark pair production at the LHC: The Electroweak contribution]{}*]{}, [*JHEP*]{} [**0812**]{} (2008) 087, \[[[arXiv:0806.1433]{}](http://xxx.lanl.gov/abs/0806.1433)\].
S. Bornhauser, M. Drees, H. K. Dreiner, and J. S. Kim, [*[Electroweak contributions to squark pair production at the LHC]{}*]{}, [*Phys.Rev.*]{} [ **D76**]{} (2007) 095020, \[[[ arXiv:0709.2544]{}](http://xxx.lanl.gov/abs/0709.2544)\].
G. F. Sterman, [*[Summation of Large Corrections to Short Distance Hadronic Cross-Sections]{}*]{}, [*Nucl.Phys.*]{} [**B281**]{} (1987) 310.
S. Catani and L. Trentadue, [*[Resummation of the QCD Perturbative Series for Hard Processes]{}*]{}, [*Nucl. Phys.*]{} [**B327**]{} (1989) 323.
A. Sommerfeld, [*Atombau und Spektrallinien, II. Band*]{}. Braunschweig: Vieweg, 1939, p. 138.
A. Sakharov, [*[Interaction of an electron and positron in pair production]{}*]{}, [*Zh.Eksp.Teor.Fiz.*]{} [**18**]{} (1948) 631–635.
S. Catani, M. L. Mangano, P. Nason, and L. Trentadue, [*[The Top cross-section in hadronic collisions]{}*]{}, [*Phys.Lett.*]{} [**B378**]{} (1996) 329–336, \[[[ hep-ph/9602208]{}](http://xxx.lanl.gov/abs/hep-ph/9602208)\].
V. S. Fadin, V. A. Khoze, and T. Sjostrand, [*[On the threshold behavior of heavy top production]{}*]{}, [*Z.Phys.*]{} [**C48**]{} (1990) 613–622.
Y. Kiyo, J. H. Kuhn, S. Moch, M. Steinhauser, and P. Uwer, [*[Top-quark pair production near threshold at LHC]{}*]{}, [*Eur.Phys.J.*]{} [**C60**]{} (2009) 375–386, \[[[arXiv:0812.0919]{}](http://xxx.lanl.gov/abs/0812.0919)\].
K. Hagiwara, Y. Sumino, and H. Yokoya, [*[Bound-state Effects on Top Quark Production at Hadron Colliders]{}*]{}, [*Phys.Lett.*]{} [**B666**]{} (2008) 71–76, \[[[arXiv:0804.1014]{}](http://xxx.lanl.gov/abs/0804.1014)\].
K. Hagiwara and H. Yokoya, [*[Bound-state effects on gluino-pair production at hadron colliders]{}*]{}, [*JHEP*]{} [**0910**]{} (2009) 049, \[[[arXiv:0909.3204]{}](http://xxx.lanl.gov/abs/0909.3204)\].
M. R. Kauth, A. Kress, and J. H. Kuhn, [*[Gluino-Squark Production at the LHC: The Threshold]{}*]{}, [[ arXiv:1108.0542]{}](http://xxx.lanl.gov/abs/1108.0542).
M. R. Kauth, J. H. Kuhn, P. Marquard, and M. Steinhauser, [*[Gluino Pair Production at the LHC: The Threshold]{}*]{}, [[arXiv:1108.0361]{}](http://xxx.lanl.gov/abs/1108.0361).
M. Beneke, P. Falgari, and C. Schwinn, [*[Soft radiation in heavy-particle pair production: All-order colour structure and two-loop anomalous dimension]{}*]{}, [*Nucl.Phys.*]{} [**B828**]{} (2010) 69–101, \[[[arXiv:0907.1443]{}](http://xxx.lanl.gov/abs/0907.1443)\].
M. Beneke, P. Falgari, and C. Schwinn, [*[Threshold resummation for pair production of coloured heavy (s)particles at hadron colliders]{}*]{}, [ *Nucl.Phys.*]{} [**B842**]{} (2011) 414–474, \[[[arXiv:1007.5414]{}](http://xxx.lanl.gov/abs/1007.5414)\].
A. Kulesza and L. Motyka, [*[Threshold resummation for squark-antisquark and gluino- pair production at the LHC]{}*]{}, [*Phys. Rev. Lett.*]{} [**102**]{} (2009) 111802, \[[[ arXiv:0807.2405]{}](http://xxx.lanl.gov/abs/0807.2405)\].
A. Kulesza and L. Motyka, [*[Soft gluon resummation for the production of gluino-gluino and squark-antisquark pairs at the LHC]{}*]{}, [*Phys. Rev.*]{} [**D80**]{} (2009) 095004, \[[[ arXiv:0905.4749]{}](http://xxx.lanl.gov/abs/0905.4749)\].
W. Beenakker, S. Brensing, M. Krämer, A. Kulesza, E. Laenen, and I. Niessen, [*[Soft-gluon resummation for squark and gluino hadroproduction]{}*]{}, [ *JHEP*]{} [**12**]{} (2009) 41, \[[[ arXiv:0909.4418]{}](http://xxx.lanl.gov/abs/0909.4418)\].
W. Beenakker, S. Brensing, M. Krämer, A. Kulesza, E. Laenen, and I. Niessen, [*[Supersymmetric top and bottom squark production at hadron colliders]{}*]{}, [*JHEP*]{} [**8**]{} (2010) 1–25, \[[[arXiv:1006.4771]{}](http://xxx.lanl.gov/abs/1006.4771)\].
W. Beenakker, S. Brensing, M. Krämer, A. Kulesza, E. Laenen, L. Motyka, and I. Niessen, [*[Squark and gluino hadroproduction]{}*]{}, [*Int.J.Mod.Phys.*]{} [**A26**]{} (2011) 2637–2664, \[[[ arXiv:1105.1110]{}](http://xxx.lanl.gov/abs/1105.1110)\].
U. Langenfeld and S.-O. Moch, [*[Higher-order soft corrections to squark hadro-production]{}*]{}, [*Phys.Lett.*]{} [**B675**]{} (2009) 210–221, \[[[arXiv:0901.0802]{}](http://xxx.lanl.gov/abs/0901.0802)\].
R. Bonciani, S. Catani, M. L. Mangano, and P. Nason, [*[NLL resummation of the heavy-quark hadroproduction cross- section]{}*]{}, [*Nucl. Phys.*]{} [ **B529**]{} (1998) 424–450, \[[[ hep-ph/9801375]{}](http://xxx.lanl.gov/abs/hep-ph/9801375)\].
N. Kidonakis, E. Laenen, S. Moch, and R. Vogt, [*[Sudakov resummation and finite order expansions of heavy quark hadroproduction cross sections]{}*]{}, [*Phys. Rev.*]{} [**D64**]{} (2001) 114001, \[[[ hep-ph/0105041]{}](http://xxx.lanl.gov/abs/hep-ph/0105041)\].
N. Kidonakis, [*[Next-to-next-to-leading soft-gluon corrections for the top quark cross section and transverse momentum distribution]{}*]{}, [*Phys. Rev.*]{} [**D82** ]{} (2010) 114030, \[[[ arXiv:1009.4935]{}](http://xxx.lanl.gov/abs/1009.4935)\].
V. Ahrens, A. Ferroglia, M. Neubert, B. D. Pecjak, L. L. Yang, [*[Renormalization-Group Improved Predictions for Top-Quark Pair Production at Hadron Colliders]{}*]{}, [*JHEP*]{} [**1009** ]{} (2010) 097, \[[[ arXiv:1003.5827]{}](http://xxx.lanl.gov/abs/1003.5827)\].
M. Beneke, P. Falgari, S. Klein, C. Schwinn, [*[Hadronic top-quark pair production with NNLL threshold resummation]{}*]{}, \[[[ arXiv:1109.1536]{}](http://xxx.lanl.gov/abs/1109.1536)\].
J. R. Ellis and S. Rudaz, [*[Search for Supersymmetry in Toponium Decays]{}*]{}, [*Phys.Lett.*]{} [**B128**]{} (1983) 248.
H. Contopanagos, E. Laenen, and G. Sterman, [*[Sudakov factorization and resummation]{}*]{}, [*Nucl. Phys.*]{} [**B484**]{} (1997) 303–330, \[[[hep-ph/9604313]{}](http://xxx.lanl.gov/abs/hep-ph/9604313)\].
N. Kidonakis, G. Oderda, and G. Sterman, [*[Threshold resummation for dijet cross sections]{}*]{}, [*Nucl. Phys.*]{} [**B525**]{} (1998) 299–332, \[[[hep-ph/9801268]{}](http://xxx.lanl.gov/abs/hep-ph/9801268)\].
N. Kidonakis, G. Oderda, and G. Sterman, [*[Evolution of color exchange in [QCD]{} hard scattering]{}*]{}, [*Nucl. Phys.*]{} [**B531**]{} (1998) 365–402, \[[[hep-ph/9803241]{}](http://xxx.lanl.gov/abs/hep-ph/9803241)\].
J. Botts and G. Sterman, [*[Hard Elastic Scattering in QCD: Leading Behavior]{}*]{}, [*Nucl. Phys.*]{} [**B325**]{} (1989) 62.
N. Kidonakis and G. Sterman, [*[Resummation for QCD hard scattering]{}*]{}, [ *Nucl. Phys.*]{} [**B505**]{} (1997) 321–348, \[[[hep-ph/9705234]{}](http://xxx.lanl.gov/abs/hep-ph/9705234)\].
S. Moch, J. Vermaseren, and A. Vogt, [*[Higher-order corrections in threshold resummation]{}*]{}, [*Nucl.Phys.*]{} [**B726**]{} (2005) 317–335, \[[[hep-ph/0506288]{}](http://xxx.lanl.gov/abs/hep-ph/0506288)\].
M. Czakon, A. Mitov, and G. Sterman, [*[Threshold Resummation for Top-Pair Hadroproduction to Next-to-Next-to-Leading Log]{}*]{}, [*Phys. Rev.*]{} [**D80**]{} (2009) 074017, \[[[ arXiv:0907.1790]{}](http://xxx.lanl.gov/abs/0907.1790)\].
S. Moch and P. Uwer, [*[Theoretical status and prospects for top-quark pair production at hadron colliders]{}*]{}, [*Phys.Rev.*]{} [**D78**]{} (2008) 034003, \[[[arXiv:0804.1476]{}](http://xxx.lanl.gov/abs/0804.1476)\].
S. Catani, M. L. Mangano, P. Nason, and L. Trentadue, [*[The Resummation of soft gluons in hadronic collisions]{}*]{}, [*Nucl.Phys.*]{} [**B478**]{} (1996) 273–310, \[[[ hep-ph/9604351]{}](http://xxx.lanl.gov/abs/hep-ph/9604351)\].
A. Kulesza, G. F. Sterman, and W. Vogelsang, [*[Joint resummation for Higgs production]{}*]{}, [*Phys.Rev.*]{} [**D69**]{} (2004) 014012, \[[[hep-ph/0309264]{}](http://xxx.lanl.gov/abs/hep-ph/0309264)\].
M. Czakon and A. Mitov, [*[On the Soft-Gluon Resummation in Top Quark Pair Production at Hadron Colliders]{}*]{}, [*Phys. Lett.*]{} [**B680**]{} (2009) 154–158, \[[[arXiv:0812.0353]{}](http://xxx.lanl.gov/abs/0812.0353)\].
S. Catani and M. H. Seymour, [*[A general algorithm for calculating jet cross sections in NLO QCD]{}*]{}, [*Nucl. Phys.*]{} [**B485**]{} (1997) 291–419, \[[[hep-ph/9605323]{}](http://xxx.lanl.gov/abs/hep-ph/9605323)\].
S. Catani, S. Dittmaier, M. H. Seymour, and Z. Trocsanyi, [*[The dipole formalism for next-to-leading order QCD calculations with massive partons]{}*]{}, [*Nucl. Phys.*]{} [**B627**]{} (2002) 189–265, \[[[hep-ph/0201036]{}](http://xxx.lanl.gov/abs/hep-ph/0201036)\].
A. D. Martin, W. J. Stirling, R. S. Thorne, and G. Watt, [*[Parton distributions for the LHC]{}*]{}, [*Eur. Phys. J.*]{} [**C63**]{} (2009) 189–285, \[[[arXiv:0901.0002]{}](http://xxx.lanl.gov/abs/0901.0002)\].
K. Nakamura et al. (Particle Data Group), [*J. Phys. G*]{} [**37**]{}, (2010) 075021 and 2011 partial update for the 2012 edition.
see <http://www.thphys.uni-heidelberg.de/~plehn/prospino/> or <http://people.web.psi.ch/spira/prospino/>.
J. Vermaseren, [*[New features of FORM]{}*]{}, [[math-ph/0010025]{}](http://xxx.lanl.gov/abs/math-ph/0010025).
O. V. Tarasov, A. A. Vladimirov, and A. Yu. Zharkov, [*[The Gell-Mann-Low Function of QCD in the Three Loop Approximation]{}*]{}, [*Phys.Lett.*]{} [**B93**]{} (1980) 429–432.
S. A. Larin, J. A. M. Vermaseren, [*[The Three loop QCD Beta function and anomalous dimensions]{}*]{}, [*Phys. Lett.*]{} [**B303** ]{} (1993) 334–336, \[[[hep-ph/9302208]{}](http://xxx.lanl.gov/abs/hep-ph/9302208)\].
A. Vogt, [*[Next-to-next-to-leading logarithmic threshold resummation for deep inelastic scattering and the Drell-Yan process]{}*]{}, [*Phys. Lett.*]{} [**B497** ]{} (2001) 228–234, \[[[hep-ph/0010146]{}](http://xxx.lanl.gov/abs/hep-ph/0010146)\].
S. Catani, D. de Florian, M. Grazzini, P. Nason, [*Soft gluon resummation for Higgs boson production at hadron colliders*]{}, JHEP [**0307** ]{} (2003) 028, .
[^1]: One has to correct for an extra minus sign in front of all $D_{Q\bar Q}$ terms in Eq. (A9) of [@Moch:2008qy].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present integral field spectroscopy of the nebular line emission in a sample of 9 brightest cluster galaxies (BCGs). The sample was chosen to probe both cooling flow and non-cooling flow clusters, as well as a range of cluster X-ray luminosities. The line emission morphology and velocity gradients suggest a great diversity in the properties of the line emitting gas. While some BGCs show evidence for filamentary or patchy emission (Abell 1060, Abell 1668 and MKW3s), others have extended emission (Abell 1204, Abell 2199), while still others have centrally concentrated emission (Abell 2052). We examine diagnostic line ratios to determine the dominant ionization mechanisms in each galaxy. Most of the galaxies show regions with AGN-like spectra, however for two BCGs, Abell 1060 and Abell 1204, the emission line diagnostics suggest regions which can be described by the emission from young stellar populations. The diversity of emission line properties in our sample of BCGs suggests that the emission mechanism is not universal, with different ionization processes dominating different systems. Given this diversity, there is no evidence for a clear distinction of the emission line properties between cooling flow and non-cooling flow BCGs. It is not always cooling flow BCGs which show emission (or young stellar populations), and non-cooling flow BCGs which do not.'
author:
- |
Louise O.V. Edwards[^1]$^{1,2}$, Carmelle Robert$^1$, Mercedes Mollá$^3$, Sean L. McGee$^4$\
$^1$Département de physique, génie physique, et optique, Université Laval, and Centre de recherche en astrophysique du Québec,\
Québec, QC G1K 7P4, Canada\
$^2$Department of Physics and Astronomy, Trent University, Peterborough, ON, Canada, K9J 7B8\
$^3$CIEMAT, Avda. Complutense 22, 28040 Madrid, Spain\
$^4$Department of Physics and Astronomy, University of Waterloo, Waterloo, ON, Canada, N2L 3G1
bibliography:
- 'le1c.bib'
nocite:
- '[@boh01]'
- '[@pet03]'
- '[@cro06]'
- '[@cra05b]'
- '[@con01]'
- '[@wil06]'
- '[@hat07]'
- '[@kau03]'
- '[@ost06]'
- '[@fer08]'
- '[@kew01a]'
- '[@dev91]'
- '[@kew01b]'
- '[@kew02]'
- '[@kaa04]'
title: 'The Diverse Nature of Optical Emission Lines in Brightest Cluster Galaxies: IFU Observations of the Central Kiloparsecs'
---
galaxies: clusters, cooling flows:general – galaxies: evolution – stars: formation – galaxies: stellar content
Introduction
============
The Brightest Cluster Galaxy (BCG) is a giant elliptical, often a cD, and is typically located at the center of the cluster’s gravitational potential. The formation and evolutionary history of these large galaxies has remained an area of active research. Close pairs and multiple nuclei are common in cD galaxies [@lai03], and hence it has been hypothesized that mergers and galactic cannibalism [@ost77] drive their formation.
However, cDs are often found at the peak of the cluster’s X-ray surface brightness, where the hot intracluster gas can potentially cool to form molecular clouds, therefore an additional stellar or gaseous component, arising via the cooling intracluster medium, may be expected. Often BCGs in cooling flow clusters show H$\alpha$ emission [@von06; @cra99], an exciting find as it is a signature of current or recent activity in their cores. Diverse and dramatic morphologies of the H$\alpha$ emission, such as long tails of emitting gas [@fab01], as well as highly concentrated emission [@don00], and even more filamentary structures [@con02; @bla01] have been observed.
@edw07b found that it is in the sample of cooling flow BCGs, compared to the sample of BCGs as a whole, where there exists an increased likelihood of H$\alpha$ emission (found in $\sim$70% of BCGs in cooling flow clusters, but only $\sim$15% of BCGs in the sample as a whole). This implies that the cooling flow status of the cluster is an important factor for activity at the center. Furthermore, it suggests that within cooling flow BCGs, cool molecular clouds, warm ionized hydrogen, and the cooling intracluster medium (ICM) are related. This is consistent with the finding of @edg02, who showed that whenever there are detections of molecular hydrogen in a cooling flow cluster, there are also detections of H$\alpha$ emission. @cra99 obtained optical spectra of 256 X-ray selected BCGs and found that 27% have emission lines and associate most of these with cooling flow clusters.
In cases where the structure of the ionized gas is well studied, complex and irregular morphologies are often spatially correlated with emission features in other wavelength regimes. As @don00 found for Abell 2597 and PKS-0745-19, the ionized gas is filamentary, and similar in extent and structure to the molecular gas emitting in H$_{2}$ 1-0 S(1) line. Other examples of well known H$\alpha$ structures include Abell 1795, which shows a huge tail [@cow83], as well as NGC 4696 in Centaurus (Crawford et al. 2005b) and NGC 1275 in Perseus [@con02; @fab08]. Each of the last three show structures spanning tens of kiloparsecs and harboring the same features as seen for the X-ray emitting gas (Fabian et al. 2001; Crawford et al. 2005a,b). Furthermore, the CO(2-1) and H$\alpha$ emissions are found to be associated in NGC 1275 [@sal06]. @don00 found the H$_{2}$ 1-0 S(1) line emission peak of NGC 1275 to be co-spatial with the peak of H$\alpha$ emission, although confined to the very center of the BCG.
Observations from the [*Chandra*]{} telescope have shown regions of star formation that are associated with bright lumps and filaments of gas whose radiative cooling times are short. In Abell 1795 for example, an excess of blue light is detected from the underlying central dominant galaxy, suggesting a population of hot young stars [@mcn96; @oeg01; @bil08]. Additionally, many rotationally excited transitions of CO been detected, as well as the O VI line, which indicates that gases at a range of temperatures exist (20$\,$K and 100 000$\,$K, respectively). Nevertheless, the amount of H$_{2}$ emission is inconsistent with gas cooling [*directly*]{} from X-ray temperatures into cool clouds. Cooling flow models predict more cooled gas than is observed (Böhringer et al. 2001; Peterson et al. 2003). Possibly, the mass is deposited into molecular clouds which are then reheated by one of several processes - hot stars, shocks, or AGN, for example [@wil02], and only a small fraction of the cooled gas is detected. The ICM and radio sources often appear to be interacting (Croton et al. 2006) and these interactions may form cavities, or bubbles, in the surface brightness of the X-ray gas which move buoyantly through the ICM. In some systems, the bubbles carry enough energy to be able to balance the radiative losses emerging from the center of the clusters in the X-ray band [@piz05; @bir04; @mcn07]. These AGNs are also potential sinks for the cooling gas and can contribute to the ionization of hydrogen gas clouds.
In an attempt to distinguish between these scenarios, this paper examines detailed maps of the line intensity and morphology of the central regions of 9 BCGs. The line emission properties are obtained using the integral field spectrgraphs.
The objective is to identify and characterize the current and recent activity in BCGs. In order to develop non-biased conclusions, it is important to study the high X-ray luminosity systems, those of lower luminosity, and those in cooling flows and non-cooling flows. The star formation rate (SFR) is calculated for young stellar populations and compared to the cooling flow mass deposition rate, which is calculated from the X-ray luminosity within the cooling radius [@fab94]. Careful attention is paid to signatures associated with processes such as the cooling flow phenomenon, AGN activity and galaxy-galaxy interactions.
The paper is organized as follows. In section \[data\], we introduce our galaxy sample selection and outline our observations. In section \[results\] we report our results: the emission line morphology and velocity maps, followed by an investigation into the emission mechanism. In section \[conclchap\], we consider the impact of our results on various galaxy and galaxy group formation hypotheses. Throughout our discussion we will compare our results to those of @wil06 and @hat07 who have recently presented integral field spectroscopy of the central regions of cooling flow BCGs. Unless otherwise stated our analysis assumes the values $\Omega_{\mathrm m}$=0.3 for the matter density parameter, $\Omega_{\Lambda}$=0.7 for the cosmological constant, and $H_0$=70 km$\,$s$^{-1}$$\,$Mpc$^{-1}$ for the Hubble parameter. L$_{X}$ refers to the bolometric X-ray luminosity throughout. We will often refer to the thesis of @edwthesis, within which additional detailed information can be found.
Sample Selection and Data Reduction {#data}
===================================
\[ifuobstab2\] \[ifuobstab2\_oas\]
------------ ---------- ---------- ----- ----------------------- -------------------- -------- -------------------------- -----------------
Cluster BCG z$_{cl}$ D kpc/$^{\prime\prime}$ IFU FOV CF MDR L$_{X}$
Name Name Mpc kpc $\times$ kpc Status M$_{\odot}$$\,$yr$^{-1}$ 10$^{37}$W
Abell 1060 NGC 3311 0.0126 53 0.25 1.3 $\times$ 1.8 No - 0.47
Abell 1204 0.1706 640 2.65 13.5 $\times$ 18.9 Yes 50$^{+40}_{-30}$ 6.77$\pm{1.42}$
Abell 1668 IC 4130 0.0634 256 1.17 5.9 $\times$ 8.2 No - 1.59$\pm{0.24}$
Ophiuchus 0.0280 116 0.55 2.7 $\times$ 3.8 No,Yes - $>$ 4
MKW3s NGC 5920 0.0450 184 0.85 4.3 $\times$ 6.0 Yes 45$^{+10}_{-10}$ 2.68$\pm{0.29}$
Abell 1651 0.0849 337 1.51 7.6 $\times$ 10.6 Yes 231$^{+121}_{-132}$ 8.25
Abell 2052 UGC 9799 0.0345 146 0.68 4.4 $\times$ 6.0 Yes 5$^{+1}_{-1}$ 2.52$\pm{0.20}$
Abell 2199 NGC 2199 0.0310 125 0.59 4.4 $\times$ 6.0 Yes 12$^{+3}_{-3}$ 3.65$\pm{0.15}$
Cygnus-A 0.0561 227 1.04 7.7 $\times$ 10.8 Yes $\sim$250 -
------------ ---------- ---------- ----- ----------------------- -------------------- -------- -------------------------- -----------------
\
Our nine targets are chosen such that line emission in the center of the galaxy is plausible. All BCGs are within 50$\,$kpc of the X-ray center and either previous reports of emission lines exist (most notably from the catalogue of Crawford et al. 1999), or the cluster properties are often associated with emission lines, i.e. a cooling flow (CF) is centered on the BCG and radio emission is detected [@edw07b]. The images and spectra are observed using the integral field spectrographs on the Gemini Telescopes (GMOS IFU) and the William Herschel Telescope (OASIS on the WHT).
The majority of BCGs which have previously been observed with integral field units are in cooling flow clusters (Conselice et al. 2001; Crawford et al. 2005b; Wilman et al. 2006; Hatch et al. 2007). Here, we include non-cooling flow cluster BCGs. The sample is listed in Table \[ifuobstab2\]. The first column lists the name of the cluster, the second the name of the BCG, and the third column lists the cluster redshift (taken from the NASA Extragalactic Database, hereafter NED). The distance and angular scale, assuming h$_{0}$=0.70, are listed in the fourth and fifth columns, respectively. The dimensions of the integral field unit FOV are given in column 6. The cooling flow status (obtained from the literature) is given in column 7 along with the mass deposition rate (MDR) in column 8. The values in these last two columns are accompanied by two important caveats. First, the cooling flow status is not uniformly determined throughout the literature. Often a short cooling time will be used to define a cooling flow, however, a high mass deposition rate, or a central temperature drop will also be used to classify the cluster. Overall these methods generally achieve the same results in terms of classifying the clusters, but can lead to discrepancies for the less powerful systems. Second, the mass deposition rates calculated based on the classical cooling flow paradigm are a factor of 10 to 100 times larger than those computed assuming most of the gas does not cool beyond $\sim$1$\,$keV. The values in the table quoted from ROSAT are the erroneous classic values and should be taken as much higher than the actual mass drop out rates. Note the very large errors on all MDR values. The X-ray luminosity is in column 9 (see the table notes for references). Each BCG and its surrounding local environment is shown in the acquisition images of Figure \[aquiims\]. The spatial extent covered is drawn on the image and spans 5$^{\prime \prime}$$\,$$\times$$\,$7$^{\prime \prime}$ for the GMOS IFU, and 10.3$^{\prime \prime}$$\,$$\times$$\,$7.4$^{\prime \prime}$ for OASIS.
\[ifuobstab\] \[ifuobstab\_oas\]
------------- -------------- --------------- --------------- ----------------- --------- ------------------------- ------------------
Observatory Cluster Name Configuration $\lambda_{0}$ $\Delta\lambda$ IntTime E(B$-$V)$_{Gal}$ $^{a}$ E(B$-$V)$_{int}$
Å Å sec mag mag
GS Abell 1060 R400+r 6300 5710-6830 1640 0.079 0.18 $^{,}$
GN Abell 1204 R400+i 7800 5940-6980 3720 0.017 0
GN Abell 1668 R400+r 6300 5560-6615 3600 0.032 0.3
GS Ophiuchus R400+r 6300/6350 5450-6850 6000 0.591 0.3
GN MKW3s R400+r 6300/6350 5450-6700 5400 0.035 0.3
GS Abell 1651 R400+i 7800/7850 6460-7700 6000 0.027 0.3
WHT Abell 2052 MR661 6610 6130-6740 3600 0.037 0.22
WHT Abell 2052 MR516 5160 4700-5345 1800 0.037 0.22
WHT Abell 2199 MR661 6610 6430-6770 2400 0.012 0.10
WHT Abell 2199 MR516 5160 4700-5345 4800 0.012 0.10
WHT Cygnus-A MR661 6610 5915-6605 3600 0.381 0.6
WHT Cygnus-A MR516 5160 4695-5225 3000 0.381 0.6
------------- -------------- --------------- --------------- ----------------- --------- ------------------------- ------------------
$^{a}$[*Taken from NED. *]{}\
[*References for internal extinction:*]{}
The GMOS IFU observations were completed between February and June, 2006. This was done in queue mode using the two-slit configuration to allow for the largest field of view. The filter and grating pairings were chosen so to observe H$\alpha$ ($\lambda$$_{rest}$6563$\,$Å) at the redshift of the cluster. Three BCGs, and their standard stars, were observed using OASIS on the nights of June 28-29, and July 2, 2005. The 22mm enlarger was used. The MR661 configuration was used to obtain H$\alpha$ and the MR516 configuration was used to obtain H$\beta$ at $\lambda$$_{rest}$4861$\,$Å. Each target was observed close to the zenith in order to reduce the effects of atmospheric absorption. Table \[ifuobstab\] gives the instrument configuration, grating central wavelength, rest wavelength coverage and integration time for each galaxy.
GMOS IFU {#gmos}
--------
Several Gemini specific programs from the package [*gemtools*]{} in the spectral analysis software IRAF were used to reduce and analyze the spectra in the standard fashion (bias subtraction, cosmic ray rejection, flat fielding, wavelength calibration, sky subtraction, and atmospheric extinction-correction). Baseline standard stars were used to perform the absolute flux calibration. The final datacubes of Abell 1651, and most notably Abell 1060, show bright ripples at both sides of the image. These artifacts, were caused by imperfect sky subtraction of the flat field frames during the standard reduction procedures. To make a cleaner continuum image for Abell 1060, the reduction process was slightly modified. Instead of using the pipeline for this case we manually constructed an image of the flat field using only the same wavelength range as the continuum coverage. We then normalized this flat field image and divided it into our continuum image directly. Although this did not result in a perfectly clean frame, it did make for a noticeable improvement: the bright and dark fringes on the right side of the continuum image in Abell 1060 had a flux difference which was 10% of the average continuum level, which decreased to a 2% difference in the reworked image. The variation in the high and low continuum level (discounting fringes) stayed constant at $\sim$35% in each case. These artifacts disappear in the continuum subtracted line images, as both the continuum, and line + continuum images contain the fringes.
Table \[ifuobstab\], column 7, includes a value for the reddening expected from the Milky Way [@shl98]. This galactic extinction was removed using the IRAF task [*deredden*]{}. Subsequently, [*dopcor*]{} and the known cluster redshifts (see Table \[ifuobstab2\]) are applied to deredshift the spectra.
Each hexagonal lenslet is about 0.2$^{\prime \prime}$ and subsampled onto a rectangular grid of 0.1$^{\prime \prime}$ pixels. Each pixel’s spectrum is then median averaged with the value of its 8 closest neighbours in order to increase the S/N and better match the seeing (typically, $\sim$0.8$^{\prime \prime}$ to 1$^{\prime \prime}$). The pixels are slightly subsampled with respect to the seeing, however our results rely on the analysis of even larger regions.
OASIS
-----
The XOasis reduction package was used for the bias subtraction, spectral extraction, flat fielding, wavelength calibration, as well as for a first cosmic ray subtraction and absolute flux calibration. Further cosmic rays were removed by median combination of the frames, except when there were less than three exposures taken. In this instance we used IRAF’s [*imreplace*]{} to create a bad pixel mask and the [*crfix*]{} routine to interpolate across the cosmic ray ruined pixels. To account for accidental pollution created in the housing chamber due to an LED, the sky subtraction itself was preformed using IRAF after the flux calibration was completed in XOasis; the entire sky frame was subtracted from the entire object frame, as the signature of the LED was apparent in both frames (each with the same integration time). Abell 2199 had data taken in the MR516 configuration spread over two nights. After sky subtraction, flux calibration and cosmic ray removal, the datasets were averaged together.
The absolute atmospheric extinction, A(0) for the WHT during the summer months is 0.12, and we used the standard value of E(B-V) of 3.1 for our calculations of the atmospheric extinction. We used the NAOMI AO corrector and the seeing was $\sim$0.7$^{\prime \prime}$, which is greater than the instrumental spatial resolution (0.26$^{\prime \prime}$). Therefore, the surrounding 8 pixels were added to increase the S/N and match the seeing. The corrections of galactic extinction and dereddening were done using the same methods as in the GMOS data.
Line Measurements and Internal Extinction
-----------------------------------------
The spectral line characteristics of flux, central position, and line width (FWHM), are mapped in Section 3. We fit multiple Gaussian profiles to the spectrum of each IFU pixel using the IRAF task [*deblend*]{}. For the nearby H$\alpha$ and \[N \] $\lambda\lambda$ 6548,6584 emission lines we find the best fit for all three lines simultaneously. The task works by taking user specified best guess parameters for the continuum level and location of the line centers. It then varies the parameters of the line model in order to improve the chi square, based on the Levenberg-Marquardt method. Errors quoted on flux levels and central position are also taken from the [*deblend*]{} fit. They are based on the Poisson statistics of the model of the data and determined by fifty monte carlo simulations which are run to create random Gaussian noise. The difference of the fitted flux to the model flux determines the error.
For most galaxies we are able to measure emission lines in H$\alpha$, \[N \] $\lambda\lambda$ 6548,6584 and \[O \] $\lambda$ 5007 and the absorption line of NaD at 5890$\,$Å with pixel-to-pixel S/N$>$10. In some cases we could also detect \[O \] $\lambda\lambda$ 6300,6364, and \[S \] $\lambda\lambda$ 6716,6734 in emission with S/N$>$5. For the GMOS individual spectra, the minimum detectable level, 1$\sigma$ is $\sim$3$\times$10$^{-19}$$\,$erg$\,$s$^{-1}$cm$^{-2}$ (For the OASIS spectra obtained with the MR661 configuration, 1$\sigma$ is $\sim$1$\times$10$^{-18}$$\,$erg$\,$s$^{-1}$cm$^{-2}$). For the OASIS spectra taken with the MR516 configuration, 1$\sigma$ is $\sim$7$\times$10$^{-18}$$\,$erg$\,$s$^{-1}$cm$^{-2}$ and absorption lines of H$\beta$, Fe and Mg$_{b}$ can only reach a 5$\sigma$ detection by adding the spectra over several individual pixels. Unless otherwise stated, the continuum level in the red has been determined by taking a median of a $\sim$100$\,$Å-wide region around 6880$\,$Å if observed using the R400+i configuration, and around 6450$\,$Å if observed in the R400+r configuration (exact continuum regions can be found in Edwards 2007). The rms for each spectrum was calculated in the same spectral windows used to make the continuum image. Average values for the signal-to-noise (S/N) ratio of the continuum near H$\alpha$ are between 10 and 20 (between 8 and 12 for OASIS). For both of the GMOS IFU configurations, R400+r and R400+i, the observed resolution is 3.3$\,$Å, an average of the measured FWHM of observed sky lines at $\sim$6300$\,$Å. The configurations used for OASIS each give an observed FWHM of $\sim$2.5$\,$Å for the standard stars.
In order to calculate qualities such as metallicity and age, and to compare the emission between galaxies, a knowledge of the internal reddening of each galaxy is important. Some elliptical galaxies are known to be dusty [@sad85] and this dust can be uniform, filamentary, or patchy [@lai03]. Our data allow a map of the internal extinction only for Cygnus-A and we show the work for deriving the pixel to pixel extinction for this case in the following paragraph. For all other galaxies, we only have access to an integrated value yielded from slit measurements obtained from the literature. This is by no means ideal, but as only the central few kpc are covered in the FOV, we resign to using this method. These values are listed in column 8 of Table \[ifuobstab\]. In some cases, no known values of the internal extinction for a particular galaxy can be found and so we adopt a value of E(B$-$V)$_{int}$ = 0.3. This value is the average value for X-ray selected BCGs with strong H$\alpha$ emission lines found in @cra99.
For Cygnus-A, we align the images to the peak in the \[O \] $\lambda$ 5007 line image to that in the \[N \] $\lambda$ 6584 line image. These two lines are chosen since they are both strong, and both are high ionization lines. The method is validated as the result is a good match to the continuum peak of both images. The extinction map is subsequently created using the following equation on a pixel by pixel basis:
$$\mathrm{E}(\mathrm{B}-\mathrm{V})_{int}=\frac{2.177}{-0.37~\mathrm{R}}\, \times \, \Big( \mathrm{Log}\Big\{\frac{\mathrm{I}_{o\mathrm{H}\alpha}}{\mathrm{I}_{o\mathrm{H}\beta}}\Big\} - \mathrm{Log}\Big\{\frac{\mathrm{I}_{\mathrm{H}\alpha}}{\mathrm{I}_{\mathrm{H}\beta}}\Big\}\Big). \label{eqnExt}$$
Here, R = 3.1, I$_{o \mathrm{H}\alpha}$/I$_{o \mathrm{H}\beta}$ is 2.85, the theoretical ratio for Case B recombination, and I$_{\mathrm{H}\alpha}$ and I$_{\mathrm{H}\beta}$ are the observed values (Osterbrock & Ferland 2006; Kauffmann et al. 2003). The theoretical I$_{o \mathrm{H}\alpha}$/I$_{o \mathrm{H}\beta}$ ratio of 2.85 may not be the ideal value to use for known Seyfert galaxies, like Cygnus-A, but the actual value is debated. It is often assumed that the H$\alpha$ emission in these systems is enhanced due to collisional processes, and several authors use a value of 3.1 [@gas84; @ost06], although other values have also been determined [@bin90 calculate a value of 3.4]. The extinction map is presented in Figure \[ebvCA\] and shows very high values of E(B$-$V)$_{int}$$\simeq$1.0$\,$-$\,$1.2 at the center where the H$\alpha$ emission is maximal. Only pixel values with S/N$>$5 for both H$\alpha$ and H$\beta$ are plotted. Although the values of extinction determined here may be slightly overestimated due to the choice of intrinsic I$_{o \mathrm{H}\alpha}$/I$_{o \mathrm{H}\beta}$ used, the average value in the NW emission peak from our map is 0.75$\pm$0.2. This is not far from the integrated value of 0.69$\pm$0.04 [@ost06] for Cygnus-A, also calculated using the Balmer decrement but using an intrinsic I$_{o \mathrm{H}\alpha}$/I$_{o \mathrm{H}\beta}$ ratio of 3.08.
Emission Line Morphology, Kinematics and Diagnostics {#results}
====================================================
\[sumtab\]
-------------- ----------- ---------------------- ------------- ---------------------------
Cluster Lines SFR Age$_{old}$ Mass $\rho$$_{old}$
Name $\times$10$^{8}$
M$_{\odot}$yr$^{-1}$ Gyr M$_{\odot}$$\,$kpc$^{-2}$
A1060 SF 0.02$\pm$0.001 10$\pm$6 2$\pm$1
A1204$^{1}$ SF & AGN 7.0$\pm$0.4 - -
A1668 AGN - 7$\pm$3 3$\pm$2
Ophi$_{BCG}$ No$^{2}$ - 10$\pm$5 150$\pm$80
MKW3s AGN$^{3}$ - $>$12 $>$8460
A1651 No - - -
A2052 AGN - $>$4 $>$30
A2199 AGN - $>$ 1 $>$ 3
Cyg-A AGN - - -
-------------- ----------- ---------------------- ------------- ---------------------------
: Summary of IFU Emission Line Properties
\
[*Note 2: There is a clump of gas, Object B, near to the BCG in Ophiuchus showing LINER or AGN-like emission line ratios.*]{}\
[*Note 3: Emission lines in MKW3s are blueshifted with respect to underlying cD population. *]{}\
We present the morphology, kinematic structure, and line diagnostics extracted from the continuum emission and line fluxes. Gaussian profiles are fit on a pixel-to-pixel basis, except for the BCGs in Abell 1668 and MKW3s, where single Gaussian fits are impossible and the flux is measured within a window around the position of the H$\alpha$ line. The emission line maps have all been constructed from continuum-subtracted measurements and no absorption correction is made. The region spectra shown are the integrated values of $\sim$ 30 spectra and are labeled on the the H$\alpha$ images.
In terms of the morphology, we compare the results from the H$\alpha$ and \[N \] $\lambda$ 6584 lines to those in the surrounding continuum for each system, and take note of any particular associated galaxy characteristics such as prominent dust features or nearby neighbours. Maps of \[S \] $\lambda$ 6716 + $\lambda$ 6734 are available in @edwthesis, but not presented here as their morphology mirrors that seen in the \[N \] $\lambda$ 6584 maps. H$\beta$, \[O \] $\lambda$ 5007, and maps of the continuum near H$\beta$ at $\lambda$$_{rest}$4861Å (between 5025 and 5100Å) are presented where available (Abell 2052 and Cygnus-A). Only pixels for which the flux measurement is $>$5$\sigma$ are used in the maps presented.
To examine the kinematics, we show the velocity of the emission lines relative to the cluster radial velocity. When available, the NaD absorption line originating from the underlying galaxy is used for comparison. The velocity and FWHM maps also only show the pixels where the line flux measurement is $>$5$\sigma$.
To diagnose the origin of the emission, we present spectra for representative regions of the BCGs (usually at the location of the H$\alpha$ or continuum peak). Generally, we do not have the H$\beta$ and \[O \] lines required for constructing a BPT diagram [@bpt81] for each galaxy. However, we have plotted values of the \[N \] $\lambda$ 6584/H$\alpha$ line ratio as a function of H$\alpha$ luminosity for central regions in all the BCGs with emission lines. This figure helps separate the ionization mechanism as a value of (\[N \] $\lambda$ 6584/H$\alpha$) $>$ 0.63 is likely to be from an AGN or LINER, rather than star formation [@ost06].
Before discussing the details of each system individually, we offer the reader a quick summary.
- All cluster BCGs show smooth continuum emission, except in the case of Abell 1060. Although the continuum emission is usually smooth, the morphology of the line emission is not uniform throughout the sample. Two BCGs show filamentary emission (in Abell 1668 and MKW3s). There are also two clusters for which the BCG harbours extended emission (Abell 1204 and Abell 2199), as well as two in which the BCG line emission is condensed (Abell 2052 and Cygnus-A). In the cases of Abell 1060 (in the BCG) and Ophiuchus (just outside of the BCG), there are large patches of dust and the line emission follows a similar morphology. Two BCGs show no evidence for line emitting gas (that in Abell 1651 and in Ophiuchus).
- The line kinematics usually vary smoothly in terms of velocity and line width. At times rotation is clearly present (Abell 1060 and Cygnus-A). Also, bulk motions (Abell 1204, Abell 1668, and Ophiuchus) and outflows (MKW3s) are observed. The subtraction of the absorption spectrum for MWK3s allows us to see a clear velocity shift of $\sim$12$\,$Å (550$\,$km$\,$s$^{-1}$) for the emission lines.
- Except for the BCG in Ophiuchus and in Abell 1651, emission lines of H$\alpha$ and \[N \] $\lambda\lambda$ 6548,6584, are prominent (Figure \[specpap\]). For the three galaxies observed with OASIS, there also exist spectra around the H$\beta$ line. The 1$\sigma$ noise level usually dominates any emission or absorption in H$\beta$ for Abell 2052 and Abell 2199. However, the \[O \] $\lambda$ 5007 emission rises above the 5$\sigma$ level in central region (Region 2). Most of the galaxies show AGN-like spectra as summarized in Table \[sumtab\], with very strong forbidden lines with respect to the Balmer emission. Figure \[pplots\] shows this for all galaxies with H$\alpha$ emission. These lines also broaden in some regions of Abell 1204 and Cygnus-A. The BPT diagram of Cygnus-A confirms a Seyfert nucleus in this system. The lines in Object B of Ophiuchus are well described by ionization from a hard source, such as an AGN, but Object B is not the BCG. In the case of Abell 1060 and Abell 1204 for which the emission line ratios show very strong H$\alpha$ emission with respect to the \[N \] $\lambda$ 6584 lines, we may be detecting a young stellar population. We determine the SFR for these last two.
NGC 3311 in Abell 1060
----------------------
[**Morphology**]{} - Figure \[a1060hai\] shows the images of the H$\alpha$ and \[N \] emission lines, as well as the continuum. Note that for this case the \[OI\] emission lines near 6300$\,$Å do not rise above the 1$\sigma$ noise level and therefore we use the median throughout a larger window (6300 and 6500$\,$Å) to build the continuum map.
@vas91 and @lai03 both note a large dust patch that corresponds in position to the obscuring feature going from the SW to the NE seen in the continuum image. In comparing the HST I-band image of the center of the BCG from @lai03, we notice the striking similarity both in extent and morphology between the dust patch they identify, and the regions we identify as strongly emitting in H$\alpha$. The H$\alpha$ contours on the continuum image indicate that much of the H$\alpha$ emission is confined within the dust, and therefore much H$\alpha$ emission could be obscured. Overall, the bright regions in \[N \] follow those seen in the H$\alpha$ emission image. The image of \[S \] $\lambda$ 6716 (not shown) follows the same overall morphology.
[**Kinematics**]{} - Figure \[havel\] shows maps of the relative velocity for the H$\alpha$ and \[N \] $\lambda$ 6584 emitting gas. The appearance of both emission lines is smooth, and we measure a velocity shear of 100$\pm$20$\,$km$\,$s$^{-1}$ across $\sim$1.0$\,$kpc. There is a clumpy distribution in the NaD velocity (clumps of -100 to +100$\,$km$\,$s$^{-1}$), originating from the underlying galaxy, but no shear. The emission line widths range from 130$\,$km$\,$s$^{-1}$ to 200$\,$km$\,$s$^{-1}$, close to the resolution, and have a clumpy distribution (available in Edwards 2007). The velocity shear, low velocity values, and lack of structure in line widths are an indication of rotation of the line emitting gas.
[**Emission Diagnostics**]{} - It is clear from the spectrum presented in Figure \[specpap\] that none of the regions in NGC 3311 show ratios of significant AGN contamination. Figure \[pplots\] shows the line ratios of individual pixels and highlights the area of Region 6, where the H$\alpha$ emission is the strongest, and the line ratios are the lowest, falling well below the AGN/SF cut of 0.63. Based on the line ratios \[N \] $\lambda$ 6584/H$\alpha$ and also (\[S \] $\lambda$ 6716 + $\lambda$ 6731)/H$\alpha$, most regions confidently fall into the composite, or star formation section of the BPT diagram. The (\[S \] $\lambda$ 6716 + $\lambda$ 6731)/H$\alpha$ are not shown here, but all regions have a ratio less than 0.25, much below 0.40, which is typical of an AGN or LINER. The image of the \[N \] $\lambda$ 6584/H$\alpha$ across the center of the galaxy, see @edwthesis, shows the central regions to have the lowest \[N \] $\lambda$ 6584/H$\alpha$ ratios. Thus, it is fairly certain that the ionization is not a result of a hard radiation field supplied by an AGN or LINER.
[**Star Formation Rate and Stellar Populations**]{} - Once the possibility of an AGN has been discarded, a common ionization mechanism explored is the effect of a young population of hot stars, which can produce strong Balmer emission [@cra99; @don00; @wil06]. Another possibility recently calculated by Ferland et al. 2008 is non-radiative heating by cosmic rays which can produce the observed molecular hydrogen lines (at least in the thin filaments of Perseus). @fab08 have shown that star formation within the thin filaments of Perseus may be delayed by magnetic fields. However, at the larger distances of our clusters, it is unlikely that the majority of the flux we observe is from such thin surrounding structures, and it is likely dominated by the bright emission at the center of the BCG. Hence, for this paper, we explore the properties of a young population of stars that could excite the observed H$\alpha$ emission.
A lower limit on the total SFR has been calculated using the method of @ken98. Excluding the dependence on the cosmology chosen, the errors in SFR are most sensitive to the error in H$\alpha$ equivalent width measurement. The intensity of H$\alpha$ emission can be diminished by the presence of H$\alpha$ in absorption. The absorption comes not only from the same ionizing population of young stars, but also from the underlying and massive older stellar population of the galaxy. Although the absorption is below the 1$\sigma$ level in the individual pixel spectra, it can be removed in the integrated spectrum of the totality of H$\alpha$ emitting pixels. We describe this presently.
[*Absorption by the Underlying Population*]{} - The emission appears only in the central 20$^{\prime\prime}$$\times$35$^{\prime\prime}$ (500pc$\times$900pc for Abell 1060) of the IFU field of view, thus, a median of the underlying galaxy spectrum can be isolated by ignoring these central pixels with emission. We assume the older population has the same characteristics as (age and metallicity), and is the dominant luminosity source for the overall luminosity profile of the optical image (that is, over the entire r filter). Therefore, the absorption should scale to this profile. Figure \[1060oldge\] shows the average continuum flux in 5 pixel-wide radial bins where the errorbars show the standard deviation of continuum values in the bin. The region 15-35$\,$pixels ($\sim$400 to 900$\,$pc) from the center is plotted. The flux varies less than 10%, from 3.1 to 3.5$\times$10$^{-18}$$\,$erg$\,$cm$^{-2}$$\,$s$^{-1}$$,$Å$^{-1}$, showing no (or only a weak) trend with radius. Therefore, we simply subtract the integrated absorption spectrum form the integrated emission spectrum (scaled to the same projected physical area). The average of both these spectra are shown in Figure \[1060hatots\].
From Figure \[1060hatots\] we measure the integrated H$\alpha$ absorption to be 3.2$\pm$0.4 $\times$ 10$^{-17}$$\,$erg s$^{-1}$$\,$cm$^{-2}$ per spectrum, on average. With the absorption spectrum in hand, it is straightforward to calculate an age and metallicity of the underlying stellar population by comparing to the results of the population synthesis code of @mol00.
For Abell 1060 the NaD line is available. From the integrated spectrum of non-H$\alpha$ emitting pixels, an equivalent width of 4.5$\pm$0.4$\,$Å, for the absorption is measured. Within our model, this strong absorption can only be fit with a super-solar metallicity (2$\,$Z$_{\odot}$). An age of 1.0$\pm$0.6$\,$$\times$$\,$10$^{10}$$\,$yr provides the best match. The error is based on the subset of models which fit the equivalent width within its measurement error. An estimate of the mass is made by scaling the continuum level (per solar mass) at 6400$\,$Å of the best model to the observed continuum level of the old population. For Abell 1060, 2.0$\pm$0.8 $\times$ 10$^{8}$$\,$M$_{\odot}$$\,$kpc$^{-2}$ are required in order to match the observations. Table \[sumtab\] lists the age and mass of the older stellar population for this, and the other clusters where NaD, Fe I, Mg$_{b}$, or H$\beta$ absorption lines are available (we observe no absorption lines in Abell 1204, Abell 1651, and Cygnus A; details in the derivation of the old stellar populations for all the BCGs can be found in @edwthesis). Lower limits for the old population are quoted for Abell 1668 and Abell 2199 as low level H$\alpha$ is observed in every pixel within the field of view.
[*Star Formation Rate*]{} - We measure a total H$\alpha$ flux of 5.0$\pm$0.3 $\times$ 10$^{-15}$$\,$erg s$^{-1}$$\,$cm$^{-2}$, or an W$_{o}$(H$\alpha$) = $-$106$\,$Å, in the absorption corrected integrated spectrum from all of the H$\alpha$ emitting pixels. This results in a total SFR of 1.4$\pm$0.1 $\times$ 10$^{-2}$$\,$M$_{\odot}$$\,$yr$^{-1}$. The SFR density is 1.2$\pm$0.1 $\times$ 10$^{-7}$$\,$M$_{\odot}$$\,$yr$^{-1}$$\,$pc$^{-2}$. The SFR derived is surely a lower limit as Figure \[a1060cnt\] suggests, and a global measure of the dust absorption in the galaxy has already been accounted for. Assuming that most of the obscured emission is behind this lane, a reasonable estimate on an upper limit to the SFR would be to suppose that all of the obscured emission has the same intensity as in Region 6 which is not behind the dust patch. Therefore, there should be no more than 1.7$\pm$0.1 $\times$ 10$^{-2}$$\,$M$_{\odot}$$\,$yr$^{-1}$ in total, or, 1.5$\pm$0.1 $\times$ 10$^{-7}$$\,$M$_{\odot}$$\,$yr$^{-1}$$\,$pc$^{-2}$. This small amount of star formation found is less than the one derived for cooling flow cluster BCGs. For example, @hic05 found rates of 0.2-219$\,$M$_{\odot}$$\,$yr$^{-1}$ derived from UV excess for cooling flow BCGs, and @edw07b found typical SFRs of 0.3 - 1.6$\,$M$_{\odot}$$\,$yr$^{-1}$ for emitting BCGs. The original MDR of 6$\,$M$_{\odot}$$\,$yr$^{-1}$ found for Abell 1060 would be between 0.06 and 0.6$\,$M$_{\odot}$$\,$yr$^{-1}$ (1 - 2 orders of magnitude below the [*ROSAT*]{} value). However, the [*Chandra*]{} observations [@yam02] show no evidence for any central temperature drop, and so no cooling flow is present in this system. Thus, the fact that we have measured only a very small amount of star formation activity we derive is consistent with this picture.
[*Properties of the Young Population*]{} - We can further quantify the activity by estimating a metallicity, age and mass for the ionizing population. We start by calculating the metallicity of the gas, using abundance ratios as described in @kew02, and then we assume the metallicity of the young stellar population is the same as for the gas. Although this may not be exact, it will help us to restrict the solution for the age of the young stellar population using synthesis models.
It is possible to make a rough estimate of the metallicity in the regions by assuming an average ionization parameter, and comparing the ratios of \[N \] $\lambda$ 6584/(\[S \] $\lambda$ 6716 + $\lambda$ 6731) and \[N \] $\lambda$ 6584/H$\alpha$. Although, as @kew02 strongly caution, both ratios depend strongly on the ionization parameter and this method is not precise. Assuming the ionization parameter is between average values of 5$\,$$\times$$\,$10$^{6}$ and 2$\,$$\times$$\,$10$^{7}$ [@kew02], the metallicities from the \[N \] $\lambda$ 6584/(\[S \] $\lambda$ 6716 + $\lambda$ 6731) ratios are compared against those from \[N \] $\lambda$ 6584/H$\alpha$. The latter generally has two possible values, therefore constraints from the former enable the correct point to be chosen, although large uncertainties exist for the \[S \] $\lambda$ 6731 line measurement. Considering the region within the central 0.60$^{\prime\prime}$ $\times$ 0.70$^{\prime\prime}$ ($\sim$175$\,$pc across), we derive a metallicity of 12+log(O/H) = 9.3$\pm$0.1 (approximately twice solar). This is high, but the mass-metallicity relation of @tre04 shows higher metallicities in more massive galaxies.
The age of the young stellar population is estimated using the evolutionary synthesis code Starburst99 [@lei99]. The H$\alpha$ equivalent width is calculated and matched to results from a run based on an instantaneous burst of star formation using a Salpeter IMF with masses between 1 and 100$\,$M$_{\odot}$. We use a model based on super-solar metallicities (Z=2$\,$Z$_{\odot}$) to best match the metallicity (although a solar metallicity does not significantly affect the age). The best fit model, experimenting with both a mass cut of M$_{up}$ = 100$\,$M$_{\odot}$ and 30$\,$M$_{\odot}$ suggests young ages ($<$10$\,$Myr) as expected for these ongoing bursts in which the youngest stellar populations dominate the spectral energy distribution.
The Starburst99 code also lists the theoretical spectral energy distribution for a 1$\times$10$^{6}$M$_{\odot}$ burst as a function of wavelength. We convert the model continuum luminosity to a flux and scale to the continuum level of the absorption corrected emission spectrum, 6 $\times$ 10$^{-17}$$\,$erg s$^{-1}$$\,$cm$^{-2}$ (the 3$\sigma$ errorbars are $\pm$2 $\times$ 10$^{-17}$$\,$erg s$^{-1}$$\,$cm$^{-2}$). At 10$\,$Myr, this requires that the young population is 3$\pm$1$\times$10$^{6}$M$_{\odot}$. The masses are lower for younger ages and lower theoretical M$_{up}$, for example, at 5$\,$Myr with M$_{up}$=30$_{\odot}$ the mass is 1.0$\pm$0.3$\times$10$^{6}$M$_{\odot}$.
These results show that if the H$\alpha$ emission does indeed result from a young population of stars, it is consistent with a low level of star formation, which in turn is much less than that predicted by direct cooling from the ICM.
The BCG in Abell 1204
---------------------
[**Morphology**]{} - Figure \[a1204reg\], shows the images of the continuum subtracted H$\alpha$, \[N \] $\lambda$ 6584, and the continuum near the H$\alpha$ emission line for the BCG. The line images share the position of the peak intensity with that of the continuum emission (including \[S \] $\lambda$ 6716 which is not shown). They also reveal the existence of a plume of bright H$\alpha$ emission which extends from the central peak to the North, and then towards the East out to the edge of the image. This is in the direction towards several smaller galaxies seen on the acquisition image (Figure \[aquiims\]). There is what appears to be a smaller galaxy in the western corner of the continuum image which has no counter part in the emission line images.
[**Kinematics**]{} - Figure \[havel1204\] shows maps of the relative velocity and FWHM for the H$\alpha$ emitting gas (with the H$\alpha$ emission overlain as contours). Regions of slight blueshifting, as well as redshifting are apparent. The physical scale is from $-$100$\,$km$\,$s$^{-1}$ to $+$150$\,$km$\,$s$^{-1}$, with the most negative values echoing the structure of the H$\alpha$ emission. There is no obvious ordered motion such as the rotation seen in NGC 3311 of Abell 1060. There is however some structure in the relative velocity map as the central emission is clearly blueshifted with respect to the rest of the emission (by up to 200$\,$km$\,$s$^{-1}$). The mean FWHM across the field of view is 130$\,$$\pm$50km$\,$s$^{-1}$, but broadens to 570$\,$$\pm$70$\,$km$\,$s$^{-1}$ at the center, in Region 4. The velocity differences are not as high seen in strong outflows, but this region hosts the highest \[N \] $\lambda$ 6584/H$\alpha$ ratios, suggesting that the emission lines are ionized by an AGN (to be discussed further below). The line widths would be consistent with the central region being closer to the broad line region of the AGN or LINER.
[**Emission Diagnostics**]{} - Comparing Figures \[specpap\] and \[1204lha\] we see that \[N \] $\lambda$ 6584/H$\alpha$ luminosity ratio of star forming Region 1 is lower than line ratio of the AGN affected central Region 4. The line ratios of \[N \] $\lambda$ 6584/H$\alpha$ and also (\[S \] $\lambda$ 6716 + $\lambda$ 6731)/H$\alpha$ suggest that most regions (Regions 1, 2, 7, 8, and 10) are likely a composite of emission from young stellar populations as well as a LINER, while Regions 3, 4, and 5 are from LINER emission only. We cannot distinguish between a Seyfert and a LINER since the ratio of \[O \] $\lambda$ 5007/H$\beta$ was not measured. Integrated along their slit, @cra99 measured a ratio \[O \] $\lambda$ 5007/H$\beta$ $>$ 1 and \[N \] $\lambda$ 6584/H$\alpha$ = 1.4. These ratios will not be uniform throughout the central regions of the BCG. Nevertheless, their integrated value for \[N \] $\lambda$ 6584/H$\alpha$ is between those of our regions. It is therefore useful to keep in mind their result of \[O \] $\lambda$ 5007/H$\beta$ as an average value, possibly closer to a LINER.
[**Star Formation Rate**]{} - It is not possible to calculate a SFR for Regions 3, 4, and 5 for example, as they are clearly dominated by AGN signatures. It is also possible that the SFR calculated in the other regions are affected by the AGN, so in this way the rates presented should be interpreted as upper limits, contaminated by AGN ionization. No strong absorption lines are present, so there is no absorption correction to the integrated H$\alpha$ equivalent width. A total SFR is derived from the combined spectrum of the brightest 255 H$\alpha$ emitting pixels and yields a flux value of 14$\pm$1$\,$$\times$$\,$10$^{-15}$$\,$erg s$^{-1}$$\,$cm$^{-2}$. This yields a SFR of 7.0$\pm$ 0.4$\,$M$_{\odot}$$\,$yr$^{-1}$, or a SFR density of 2.3$\pm$0.1$\,$$\times$$\,$10$^{-8}$$\,$M$_{\odot}$$\,$yr$^{-1}$$\,$pc$^{-2}$.
This upper limit is not inconsistent with the revised value of 50$^{+40}_{-30}$$\,$M$_{\odot}$$\,$yr$^{-1}$ for the MDR assuming the gas does not cool below 0.1keV found by @ode08, as not all the molecular gas will convert to stars, and as the current estimates of MDRs are an order of magnitude [*below*]{} the previously derived rates. Our total SFR is very close to that derived by @ode08 from infrared luminosities inside of a 12.2$^{\prime\prime}$ aperture [@qui08], especially considering our effective aperture is slightly smaller. Our results further suggest that although not the dominant source, part of the luminosity in the O’Dea measures may arise from the central AGN.
We do not attempt to predict a valid age and mass estimate for this source, as contamination from the AGN is important, and the SFR is only an upper limit. The interested reader may refer to Edwards 2007 for an examination into the ages in select less contaminated regions.
IC 4130 in Abell 1668
---------------------
[**Morphology**]{} - There are several pixels which are not well fit by single Gaussian profile fits for this target because of the lower signal, and here, the flux is simply added within a specific window. In this case, Figure \[aha1668\] shows the flux added between 6554 and 6572$\,$Å for H$\alpha$, and between 6575 and 6593$\,$Å for \[N \] $\lambda$ 6584. The figure also shows the continuum near H$\alpha$. Unlike the smooth elliptical distribution of the continuum image, both the H$\alpha$ and \[N \] $\lambda$ 6584 flux images show much more patchy and filamentary structure throughout. The bright regions on the \[N \] $\lambda$ 6584 image correspond to bright regions on the H$\alpha$ image, and the two emission lines share the same overall structure. However, the peak emission is displaced in the images.
[**Kinematics**]{} - A map of the relative H$\alpha$ line velocities is shown in Figure \[havel1668\]. It shows a clear gradient from positive values North of the center ($\gtrsim$ 100$\,$km$\,$s$^{-1}$), to negative values ($\lesssim$ $-$100$\,$km$\,$s$^{-1}$) at the South with the zeropoint near the center of the continuum. The low S/N values at the edge of the figure, as well as a few central patches where emission lines could not be fit as single Gaussians are plotted as having zero relative velocity. Also shown in the figure is the velocity of the NaD lines relative to the average velocity of IC 4130. This map traces the motion of the underlying galaxy, as opposed to the line emitting gas. It shows a smaller variation in magnitude ($\sim\pm$70$\,$km$\,$s$^{-1}$), no smooth gradient, and all of the velocities are negative. This implies that the line emitting gas is not at rest with respect to the underlying galaxy.
[**Emission Diagnostics**]{} - All regions show stronger \[N \] $\lambda$6584 line fluxes than those for H$\alpha$ as can be seen in Figures \[specpap\] and \[pplots\]. Region 8 has the smallest ratio of all the regions, but the value is 1.7$\pm0.2$, including the measurement error from both of the lines. This is still well described by AGN or LINER emission. The facts that the \[N \] $\lambda$6584/H$\alpha$ ratio does vary across the image and with H$\alpha$ luminosity, and that the scaled underlying spectrum has a lower continuum than the emitting spectrum (although some amount of continuum is surely from the nuclear emission) suggest a young population may still exist. However, its lines are completely masked by the presence of the AGN or LINER, and therefore, it is not possible to use the H$\alpha$ emission line to characterize a SFR or an age for the younger stellar populations.
Ophiuchus
---------
[**Morphology**]{} - The continuum subtracted H$\alpha$ flux image for the BCG is shown in Figure \[ahaophi\]. The continuum subtracted \[N \] $\lambda$ 6584 flux image is also shown, but the contours represent the H$\alpha$ flux. There is no line emission at the center of the BCG, rather it is concentrated North of the continuum emission. This concentration of line emission is hereafter referred to as Object B.
There is a drop in the continuum levels of the BCG to the North side, seen in Figure \[aophiconti\]. This is also apparent in Figure \[aquiims\] where the dust is seen as a decrease in flux just to the side of the bright line-emitting Object B. Both these figures show a clear lack of emission at the same projected position as Object B, probably caused by dust. Since the dust feature appears to be very localized relative to the BCG and it is absorbing the galaxy continuum we conclude that the dust patch is in front of the BCG.
[**The Nature of Object B**]{} - By comparing the average redshifted positions of the emission lines in Object B to the average velocity of the BCG NaD absorption line we measure a velocity difference of $+$600$\pm50$$\,$km$\,$s$^{-1}$ between Object B and the BCG, see Figure \[specpap\]. This is well within the typical velocity dispersion of a massive galaxy cluster. There is also a difference of $+$150$\,$km$\,$s$^{-1}$ between the BCG and the cluster velocity (derived from the cluster redshift measurement of z=0.028 from @lah89), though this difference is close to the error of the cluster radial velocity measurement.
Three scenarios can explain the nature of Object B: 1) that it is infalling onto the BCG, 2) that it is a background galaxy, and 3) that it is being ejected from the BCG. The first scenario is supported by the absorption seen in the continuum and acquisition images as well as the relative velocity difference between Object B and the BCG. Scenario 2 is supported by the velocity difference but the dust absorption of the BCG in the region of Object B would have to be a coincidence. If Object B were being ejected away from us, the third scenario could produce the observed velocity difference, but would not explain the dust absorption. Finally, if Object B were being ejected towards us, the dust feature would be explained, but the velocity difference would not.
The hypothesis for which Object B is in front of the BCG and falling onto the BCG or spiraling around it most completely explains our observations. Figure \[havelophi\] shows the H$\alpha$ velocity map measured with respect to the average Object B velocity. This map could indicate some rotation of Object B (where the side near the BCG is rotating in our direction) perhaps perturbed due to some tidal effect with the BCG. In this scenario, Object B could be a small galaxy (or a large molecular cloud) where the interaction with the BCG is compressing the gas resulting in part of the optical and X-ray emission observed.
The integrated spectrum of Object B has \[N \] $\lambda$6584/H$\alpha$ = 3.2$\pm0.3$, a value which is similar to those in its individual regions of (Figure \[1204lha\]), and which indicates an AGN or LINER ionizing source.
Furthermore, \[O \] $\lambda$6300/H$\alpha$ = 0.3$\pm0.05$, and \[S \] $\lambda$6716 $+$ $\lambda$6731/H$\alpha$ = 0.9$\pm0.3$ are consistent with the AGN region of the BPT diagrams, though less clear. The (\[S \] $\lambda$6716 $+$ $\lambda$6731)/H$\alpha$ for instance, varies in the different regions, from 0.72 to 1.26, which is nearer to but on the border of the AGN or LINER side of the BPT diagram.
=2.5in
NGC 5920 in MKW3s
-----------------
[**Morphology**]{} - Figure \[acntmkw\] shows the image of the continuum subtracted H$\alpha$ + \[N \] $\lambda\lambda$ 6548,6584. Deep absorption lines from the underlying galaxy are shifted in velocity with respect to the emission lines. This superposition renders single Gaussian fits to the emission lines impossible. Therefore, the image shown is the addition of all flux between 6530 and 6590$\,$Å. In contrast with smooth spherical continuum emission, the line emission shows a strong elongated feature which crosses the center of the cD galaxy from the NE to the SW.
[**Kinematics**]{} - In Figure \[specpap\], we present the spectrum of the central region (Region 2) for which, in this case, the absorption spectrum is subtracted. Only here, do we subtract the absorption, as in this case as it is deep and shifted in velocity from the emission lines. We measure the absorption by building a spectrum of the underlying galaxy, as in Section 3.1. After the subtraction, the emission lines become easier to distinguish in the integrated spectrum of Region 2. They are clearly blueshifted from the underlying galaxy by $-$560$\pm50$$\,$kms$^{-1}$. This is a known radio galaxy and the velocities are probably indicative of an outflow in our direction.
[**Emission Diagnostics**]{} - Individual pixels are of much too low S/N to plot on Figure \[1204lha\]; we therefore plot the values of the (absorption corrected) integrated regions. The average ratios are also calculated for Regions 1, 3, 4, and 5, there are close to or larger than for Region 2. The values are 1.3, 1.3, 2.6, and 1.7. All emitting regions are thus likely ionized by a hard source, AGN or LINER-like emission.
The BCG in Abell 1651
---------------------
The integrated spectrum shown in Figure \[specpap\] has no emission lines but only H$\alpha$ in absorption. Thus, although this cluster has a cooling core, there are neither signs of gas ionized by an AGN, nor by a population of hot young stars at the center of the BCG.
UGC 9799 in Abell 2052
----------------------
[**Morphology**]{} - Figure \[aha2052\] shows the images of the continuum subtracted H$\alpha$ flux, and the continuum near H$\alpha$ (between 6380 and 6430$\,$Å). These images show smoothly varying emission and share the same peak location. We do not detect any H$\beta$ emission above the 1$\sigma$ level of the noise, the upper limit in the central Region 2 being 4$\times$10$^{-17}$$\,$erg$\,$s$^{-1}$cm$^{-2}$. Lines of \[N \] $\lambda$ 6584, \[S \] $\lambda$ 6716 + $\lambda$ 6734 and \[O \] $\lambda$ 5007, and the continuum around H$\beta$ (which are not shown) display a similar morphology as H$\alpha$.
[**Kinematics**]{} - The relative velocity map of H$\alpha$, Figure \[havel2052\], has a range of $-$250$\,$km$\,$s$^{-1}$ to $+$150$\,$km$\,$s$^{-1}$. The figure shows negative velocities to the South of the center of the emission, and more positive velocities North of the center of the line emitting region. The the line widths (map not shown) vary by 450-700$\,$km$\,$s$^{-1}$ across the galaxy, but show no clear structure. However, compared to cases with clear signs of rotation where the velocity gradient develops smoothly from positive to negative along an axis, in this case, it is less clear as to whether there is rotation or an outflow.
[**Emission Diagnostics**]{} - Figures \[specpap\] and \[1204lha\] show ratios characteristic of ionization due to AGN or LINER activity in most of the pixels. Regions 1, 2, and 6 have \[O \] $\lambda$ 5007 line fluxes of 8.5, 8.2, and 2.4$\times$10$^{-16}$$\,$erg$\,$s$^{-1}$cm$^{-2}$, respectively. There are the only three regions where the line flux is greater than 5$\sigma$ (2.0$\times$10$^{-16}$$\,$erg$\,$s$^{-1}$cm$^{-2}$). Therefore, in these regions, the \[O \] $\lambda$ 5007 emission is necessarily stronger than the H$\beta$ emission and the \[O \] $\lambda$ 5007/H$\beta$ ratio is $\gtrsim$ 6. This helps to constrain the harder emission source as a Seyfert nucleus, over a LINER (\[O \] $\lambda$ 5007/H$\beta$ $>$3.2; also, see Crawford et al. 1999 who obtain an O \] $\lambda$ 5007/H$\beta$ ratio of 9.5). The (\[S \] $\lambda$ 6716 + \[S \] $\lambda$ 6731)/H$\alpha$ line ratio is above 1.10 for all regions and well within the AGN side of the diagram. Therefore, it is not possible to use the H$\alpha$ emission line to search for any young stellar population. However, observing the scatter in the individual pixels of the central Region 2 on Figure \[1204lha\] does suggest that there may be 2 cases: 1) for H$\alpha$ luminosities that increase as the ratio decreases, and hence are affected by star-forming regions, and 2) where the ratios are higher and the H$\alpha$ luminosities are $\sim$1.
NGC 6166 in Abell 2199
----------------------
[**Morphology**]{} - Images of the continuum subtracted H$\alpha$ flux along with the continuum subtracted \[N \] $\lambda$ 6584, and the continuum near H$\alpha$ are shown in Figure \[aha2199\]. The continuum image which includes flux between 6650 and 6700$\,$Å, increases smoothly in brightness towards the center, without any prominent dust features. The peak of the H$\alpha$ emission does coincide with the peak of the continuum flux, however, a second bright peak, not noticeable in any of the other line images is seen to the South. The H$\alpha$ line flux continues to extend towards the East, a morphology that is mirrored in the \[S \] $\lambda$ 6716 + $\lambda$ 6734 emission line image (not shown), although in this case the \[N \] $\lambda$ 6584 line image differs in that it is much more condensed. The elliptical morphology of the continuum in the blue, between 5025 and 5100$\,$Å (not shown) is similar to that seen for the red continuum, though it is less condensed. The individual pixels of the blue configuration show no strong H$\beta$ or \[O \] $\lambda$ 5007 line emission.
[**Kinematics**]{} - The relative velocity map of H$\alpha$ emission is shown in Figure \[havel2199\]. The velocities are calculated with respect to the rest frame of the central galaxy and plotted as such in the figure. However, one can see that clearly the emission lines are not at rest with respect to the central galaxy (z$_{BCG}$=0.03035$\pm$0.00003; de Vaucouleurs et al. 1991), and that the velocity difference within the line emitting gas itself goes from $-$200$\,$km$\,$s$^{-1}$ to $+$200$\,$km$\,$s$^{-1}$. The low level flux of the Eastern side of the image has a complex structure. It is the location of both the Region 6 with the largest velocity, 700$\pm$50$\,$km$\,$s$^{-1}$, and Region 5 with the lowest velocity, 580$\pm$50$\,$km$\,$s$^{-1}$. The pixels which have the most intense line emission are those with the lowest velocities, it could be that they are closer to us and hence subject to less extinction. If this were the case, one would then expect the H$\beta$ line emission to be more smooth, however hardly any emission lines are seen in the image (only Region 8 has detectable H$\beta$ emission). However, the H$\beta$ emission lines are quite possibly effected by absorption which cannot be seen at the S/N level of this data. In this case we would be seeing a curved jet with relative velocity $\sim$500$\,$km$\,$s$^{-1}$. A higher S/N H$\beta$ image would help verify this scenario.
=2.5in
[**Emission Diagnostics**]{} - The spectra in the red and blue are shown in Figure \[specpap\]. The \[N \] $\lambda$ 6584/H$\alpha$ ratios (see Figure \[1204lha\]) for the different regions in this galaxy vary from 1.7 to 10. The \[O \] $\lambda$ 5007/H$\beta$ ratio, measurable only for Region 8, has a value of 1.8$\pm0.5$. This is consistent with the ratios from @cra99 derived from a 6$\,$$^{\prime\prime}$ slit. It is possible that stellar absorption affecting H$\alpha$ and H$\beta$ could be artificially raising the ratios, and hiding the effects of a young population. However, this would not be a significant effect since no absorption is seen in the integrated spectra from the outskirts [@edwthesis].
The only region for which an H$\beta$ emission strength is measurable is Region 8, which shows an H$\alpha$/H$\beta$ ratio of 2.4$\pm0.4$. This is within the theoretical lower limit of 2.8 to within our measurement errors. The ratio is much lower than the integrated value of @cra99, H$\alpha$/H$\beta$ = 9.3. However, as H$\alpha$ is strong in several regions and H$\beta$ is only present in Region 8, the integrated ratio is naturally much larger. Nevertheless H$\beta$ is clearly detected in Region 8 which suggests that the extinction level is lower in this region. It is therefore reasonable to suggest that Region 2 may be closer to the dusty torus of an AGN. It may also be that Region 8 has a population of young stars and that H$\beta$ absorption could be lowering this ratio.
Cygnus-A
--------
[**Morphology**]{} - Figure \[aoIIIbCA\] presents the continuum subtracted H$\alpha$, H$\beta$, \[N \] $\lambda$ 6584, \[O \] $\lambda$ 5007 emission, and the continuum near H$\alpha$ and near H$\beta$. In the line emission images, it is the NW and the central components (Regions 1 and 2) that have the highest flux. In the image of the continuum near H$\alpha$ (between 6380 and 6430$\,$Å), the brightest emission is seen in four blobs: one central, one East of center, one to the SE and another to the NW. The image of the continuum near H$\beta$ (between 4720 and 4830$\,$Å) shows bright emission in the SE and NW blobs, but the central peak is not seen, likely due to intense dust extinction [@jac98].
[**Kinematics**]{} - The relative velocity maps of the H$\alpha$ and \[O \] $\lambda$ 5007 emission lines are shown in Figure \[havelCA\]. The relative velocity map of the \[N \] $\lambda$ 6584 emission is not shown as it harbours very similar morphology and magnitude as to that for the H$\alpha$ relative velocity map. Large scale velocity gradients of $\pm$200$\,$km$\,$s$^{-1}$ are seen for H$\alpha$, and of $\pm$100$\,$km$\,$s$^{-1}$ for \[O \] $\lambda$ 5007, both in the SW and NE direction. On this scale, the velocity gradients in both lines are reminiscent of rotation.
The FWHM of the H$\alpha$ emission line is also shown in Figure \[hafwCA\]. It reveals large widths corresponding to the central emission peak, the location of the central radio point source [@tad03]. There is also an area of large width that extends west of this peak (that is, west of Region 2). This second area extends to Region 8, the direction of the previously detected jet. The large width could be caused by disturbance from the jet.
The morphology and velocity seen in these maps are consistent with the picture given by @jac98. Figure \[cygAgeom\] summarizes their geometry and highlights an opening cone that passes from the NW side (Regions 1 and 2) of the image through the central blob to the SW side of the image. These authors also point out blue condensations (Region 6) and the jet that emanates from the center westwards, passing along south side of the NW blob. Our largest velocity gradients are located perpendicular to the opening cone as well as to the direction of the jet identified by @jac98.
[**Emission Diagnostics**]{} - Figure \[specpap\] shows the red and blue spectra for the central Region 1. As suspected, strong \[O \] $\lambda$ 5007 and \[N \] $\lambda$ 6584 relative to H$\beta$ and H$\alpha$. Figure \[1204lha\] shows that most pixels have \[N \] $\lambda$ 6584/H$\alpha$ ratios higher than $\sim$ 1.4, placing them on the AGN side of the BPT diagram. The highest ratios are in two clumps almost directly South and North of the center of the image, however there appears to be no relationship between the H$\alpha$ luminosity and the value of the line ratio. The lowest line ratios are East of the center, at Region 6. This is the location of bright, blue, condensed clumps of continuum emission (observed using the F622W HST filter) previously discussed in @jac98. The spatial resolution here is not as high as for HST, and neighbouring pixels may be washing out signs of young stars. Indeed, @jac98 are unable to identify any emission of \[O \] $\lambda$ 5007 or H$\beta$ within these blue condensations.
This is the only BCG for which we can constrain the emission mechanism using all four diagnostic lines. The \[O \] $\lambda$ 5007/H$\beta$ map (not shown) is affected by the poorer quality of the H$\beta$ image, but shows no clear structure immediately apparent in the map. However, all of the central pixels show ratios of \[O \] $\lambda$ 5007/H$\beta$ $>$8, disclosing the nature of this AGN to be a Seyfert. The BPT diagram shown in Figure \[bptCA\] classifies virtually every pixel as ionized by AGN emission, and there is a smooth transition from pixels at the Seyfert corner down to the LINER side. This agrees with the well established classification of Cygnus-A being a narrow line radio galaxy. Two additional diagnostics are available for Cygnus-A, the ratio of \[O \] $\lambda$ 6300/\[O \] $\lambda$ 5007 and \[O \] $\lambda$ 5007/H$\alpha$. The former is a good tracer of the ionization parameter (q, as defined in Kewley & Dopita 2002; q=u$\times$c, a measure of the number of ionizing photons per atom at the boundary layer). Both lower and higher ionization states are present. The \[O \] $\lambda$ 5007/H$\alpha$ ratio also shows a region of higher ionization along the edge of emission, North of the NW blob (Region 1), which was noted in the observations of @jac98.
Discussion {#conclchap}
==========
Overview of Results
-------------------
We have investigated the emission line morphologies, dynamics and ionization state in a sample of nine BCGs. Condensed (Abell 2052, Cygnus-A), filamentary and patchy (Abell 1060, Abell 1668, and MKW3s), and extended (Abell 1204, Abell 2199) morphologies are observed and the H$\alpha$ and \[N \] emitting gas usually follows the same morphology. Also of note are two cases, Abell 1060 (a non-cooling flow) and Ophiuchus (at best a low level cooling flow), where the line emitting gas morphology echos that of prominent dust features seen in the optical images. By using BPT diagrams to diagnose the ionization mechanism, we found that in two out of seven emitting BCGs, hot stars are likely perpetrators - NGC 3311 in the non-cooling flow cluster Abell 1060 and the BCG in cooling flow cluster Abell 1204. In these cases, SFRs and ages for the young populations are derived. However, the presence of an optical AGN, in both cooling flow as well as non-cooling flow systems, was strong. Two of the nine BCGs observed had no emission lines. This includes the BCG of Abell 1651 and the BCG Ophiuchus. Unexpectedly in Ophiuchus, it is another source within the cluster (Object B), rather than the BCG that shows line emission. In most cases, the relative velocities are in the neighborhood of $\pm$100 - 200$\,$km$\,$s$^{-1}$, though some higher relative velocities indicative of large inflows (Object B in Ophiuchus) or outflows (MKW3s) are observed.
The properties and main results for each cluster BCG are highlighted in Table \[sumtab\] and discussed briefly below.
- Abell 1060: Previous detections of star formation in this dusty [@lai03] non-cooling flow [@hay06] BCG have been described by @vas91. We characterize several regions of star formation, note that the morphology of the star forming region follows that of the dust, and present a smooth velocity gradient reminiscent of rotation.
- Abell 1204: This cooling flow cluster [@bau05] has the highest redshift of our sample (z = 0.1706) and thus a larger extent of the cD galaxy is viewed. Regions of ionization due to AGN are separated from those indicative of a young stellar population. The latter are found further from the center in a plume that extends towards a chain of smaller cluster galaxies. The relative velocities are modest ($\pm$200$\,$km$\,$s$^{-1}$) and show no signs of rotation. An interaction with the nearby galaxies may be important.
- Abell 1668: This is a non-cooling flow cluster [@sal03], yet strong lines are present. These appear in a filamentary distribution with line ratios signifying ionization from an AGN. The relative velocity of the line emitting gas shows that it is not at rest with the underlying cD.
- Ophiuchus: This cluster was recently observed to have a cool core, even though the X-ray temperature of the cluster itself is very hot at 9-10$\,$keV [@fuj08]. The BCG shows no emission lines. However, emission lines are seen in Object B, at a projected distance of $\sim$2$\,$kpc away from the center of the BCG, which is also at the position of the X-ray centroid. Object B is also cospatial with a dust feature seen on the acquisition image, and the emission lines have a relative velocity of $+$750$\,$km$\,$s$^{-1}$ with respect to the BCG. We attribute this to infall onto the BCG.
- MKW3s: This cooling flow cluster (Kaastra et al. 2004) has previously been observed to have a UV excess and attributed star formation by @mcn89 and by @hic05. We find the emission lines in the GMOS IFU image to be filamentary in morphology, yet well described by ionization from an AGN. The lines are blueshifted by $+$560$\,$km$\,$s$^{-1}$ with respect to the BCG, suggesting an outflow.
- Abell 1651: This cooling flow cluster [@whi00] BCG shows no emission lines.
- Abell 2052: This cooling flow cluster [@bla03; @kaa04] has patchy dust in the center of the BCG [@lai03] and @hic05 and @bla03 have deduced star formation in the BCG from excess UV-IR and U-band continuum emission. We find that the Seyfert signature overwhelms any star formation in the central few arcseconds of the BCG. The morphology of the emission lines here is point-like, in that it shows no asymmetric or extended features. The relative velocities vary on the scale of $\pm$250$\,$km$\,$s$^{-1}$, but are not smooth enough to warrant a classification of rotation.
- Abell 2199: Another cooling flow [@joh02] BCG with previous calculations of star formation based on UV excess [@mcn89]. We find the AGN signature in the emission lines is dominant throughout the extent of the OASIS image. The lines are brightest in the center, but exhibit an extended morphology towards the East. The relative velocities are similar to those seen in Abell 2199, but again the gradient is not regular enough to signify rotation.
- Cygnus-A: This poor cooling flow cluster [@rey96] is a well studied AGN [@jac98; @tad94; @tad03]. We find a morphology consistent with the images of @jac98, line ratios telling of an AGN, and a velocity gradient that suggests rotation. The FWHM of the emission lines traces the direction of the jet, as it becomes wider for the lines in that region.
Certainly gas exists in BCGs and is excited, and the observations above show that many mechanisms are at play. For instance, signs of emission by hot stars are present in both cooling flow and non-cooling flow systems, and AGN-ionized gas is also present in both cooling flow and non-cooling flow systems.
Scenario
--------
@wil06, who conducted a similar study, find little variation of the emission line ratios across the emission nebulae in their high redshift and very luminous sample of 4 cooling flow BCGs, implying a uniform ionization state. They suggest the following single scenario to explain the line emission in all their galaxies: an interaction of smaller cluster galaxies triggers starbursts in cold gas reservoirs, presumably deposited from the cooling flow. This is supported by their observation that different ionization states of the gas vary little spatially, implying a single ionization source for the H$\alpha$ emission. Intriguingly, the current observations for the modest cooling flow case of Abell 1204 support this idea (this is the only cluster in this sample where z $>$ 0.1). In this case, the H$\alpha$, \[N \] and \[S \] emitting gas all share a similar morphology. But, in contradiction to the findings of @wil06, we do find a difference in line rations with respect to position and H$\alpha$ strength with the lower line luminosity pixels showing more star-forming like activity (consistent with results of Hatch et al. 2007). We calculate a total SFR of $\simeq$7$\,$M$_{\odot}$$\,$yr$^{-1}$ for this BCG, which is close to the infrared derived SFR of @ode08 and not inconsistent with the mass deposition rate as not all the molecular gas will convert to stars and as the calculations of MDRs based on the absence of the coolest gas are an order of magnitude [*below*]{} the rates based on the classical cooling flow scenario - 50$\,$M$_{\odot}$$\,$yr$^{-1}$ [@ode08] compared to 675$\,$M$_{\odot}$$\,$yr$^{-1}$ [@whi00]. @wil06 also reasoned that the H$\alpha$ and CO gas are related as they share the same kinematics, and it is this CO gas, which is subsequently disturbed by the passing of nearby neighbors or by an AGN, which will emit the H$\alpha$ line. The observations presented here for Abell 1204 show no hard evidence for an interaction with the nearby neighbors, but, it is tempting to speculate on the likelihood of an interaction with nearby galaxies seen on the acquisition image. It would be interesting to search for molecular gas in this BCG. This is the only cluster in the sample presented in this paper which agrees with the @wil06 hypothesis.
However, the overall “mixed bag" of ionization scenarios seen throughout most of this sample is even more appealing when put into the context of the @hat07 observations. In opposition to the scenario put forth by @wil06, @hat07 find cases where the line emission properties in cooling flow BCGs suggest motions from strong AGN or starburst driven outflows (Abell 2390, Abell 1068), from galaxy passbys (2A 0335+096), and from rotation (Abell 262). They find that the ionization state is not uniform, and do not conclude that one scenario can account for the emission lines seen in cooling flow BCGs, the results presented here are consistent with such conclusions.
One caveat of our study is the small sample size, nonetheless, no difference in the ionization mechanism has been found that clearly separates the cooling flow and non-cooling flow cluster BCGs. Neither has a consistent picture been developed to explain the origin of the line emission throughout the sample. Although, the emission line characteristics are consistent within each BCG and most systems show a hard ionizing source prominent throughout the central few arcseconds, usually well described by LINER emission line ratios.
As found by several other authors [@wil06; @boh02; @don00], this data does not support a simple picture in which X-ray gas cools into molecular clouds subsequently forming stars. This is emphasized by the variation in morphology of the ionized nebula in these systems. The disturbed morphology of the non-AGN ionized gas of Abell 1204 extends towards the direction of several smaller galaxies, suggestive of an interaction with nearby companions. The spectrum of MKW3s is also interesting, with the H$\alpha$ emission line shifted with respected to the underlying spectrum. Although for the majority of cases the line emission is stronger in the central regions of the BCG itself, in Ophiuchus at least, the line emission is localized North of the BCG center, at a distance of $\sim$2$\,$kpc. @don07 find the emission lines in the cooling flow cluster 2A0335+096 are associated with the BCG, as well as a companion. In the case of 2A0335+096, the emission lines are described as dusty, and this is also the case in Ophiuchus and in NGC 3311 of Abell 1060 where the line emission is constrained by what appears on the acquisition images to be strong dust features.
There are of course many possible outcomes for any molecular gas that drops out of the cooling flow onto the BCG. In general, the observations from the dataset presented here, of an overwhelming influence of AGN signatures in most of the sample support the scenario currently put forth for the nature of cooling gas in X-ray clusters where AGN feedback is important. That is, part of the material condensing out of the cooling X-ray gas finds itself in the form of molecular reservoirs at the centers of the BCG, and part rains onto the central black hole. This could trigger an outburst from the AGN which is then hypothesized to reheat the cluster enough to prevent any further cooling. The AGN could simultaneously ionize the molecular gas deposits and influence starbursts. The molecular gas deposits would also be subject to flybys from companions, another mechanism which could trigger a starburst. This complex scenario does explain the observations of line emission resulting from the ionization of hot stars and AGN activity in cooling flows. It could also explain observations of AGN-ionized line emission in non-cooling flow, as the line emission could be triggered after the cooling cluster gas has been reheated by the AGN. The non-cooling flow BCG NGC 3311 of Abell 1060 may also be a part of the nominal $\sim$15% of line emitting BCGs [@edw07b] as it is a non-cooling flow and the emission lines show no evidence for AGN ionization, and thus perhaps the star formation is due to a completely different process. It is however, somewhat more difficult to explain the observations of Ophiuchus, where the line emission is not localized in the BCG. In a future paper, we will combine X-ray and radio observations with our non-detection of line emission from Ophiuchus and Abell 1651 to make a census of the heating and cooling in those systems.
Conclusions
===========
The observations from integral field spectroscopy have revealed the complex nature of the line emission in these galaxies, for which star formation, AGN, or both are important in several systems, but for which no consistency of the emission line characteristics and the cluster properties is seen. The line emission is point-like, filamentary, or extended; it exists in regions that are plagued with large patches of dust, and those that are relatively dust free. The AGN ionizes all of the gas within the image in some cases, whereas regions of star formation are discernible in others. Outflows are seen, as well as gas that has bulk rotations. And the emission could be associated with galaxy interactions in some, but not all cases. Each case with emission lines in a cooling flow shows regions where the emission lines have AGN-like ratios.
The process that is fueling the nominal emission in the non-cooling flows and the other bright galaxies at the center of a cluster is still unknown. However, in cooling flows, the existence of an AGN is correlated with the presence of emission lines. As seen for example in Abell 1204 (this work), 2A0335+096 [@hat07], and Abell 2052 [@bla03], the ionization mechanism can change with the radial distance from the center, showing lines ionized by a hard source at the center, and those ionized by hot stars further out.
This work supports the current ideas on the important role of a massive black hole in these galaxies, as much of the optical line emission is dominated by gas ionized by an AGN. The black hole may act as a sink for the cooling gas [@piz05], may trigger a starburst [@wil06], and its energy output may prevent further cooling of the cluster gas [@bes05; @cro06; @mcn07]. This last role is an important factor in our understanding of the building of large galaxies. The observed galaxy luminosity function undergoes a sharp cut-off at the high end. This cut-off can be explained with the inclusion of X-ray gas and AGN feedback, where the amount of cooling (and hence potential star formation) is regulated by heating from the AGN [@cro06].
Acknowledgments {#acknowledgments .unnumbered}
===============
We are grateful to Jorge Iglesias and Sam Rix, who helped with the OASIS observations at WHT. We thank B. McNamara, H. Martel and L. Drissen for very helpful comments and discussion. We also thank Simon Cantin and Véronique Petit for advice on the OASIS data reduction and plotting tools, and the Gemini Science team for their support in the observing and data reduction for the Gemini data. We would also like to thank the referee for comments which significantly improved the paper. This work was supported by the Natural Sciences and Engineering Research Council of Canada, Université Laval, and le Fonds québécois de la recherche sur la nature et les technologies through research grants to CR and by the Spanish PNAYA projects and AYA2007-67965-C03-03 to MM.
[^1]: Moved to: Infrared Processing and Analysis Center, California Institute of Technology, Pasadena, CA, USA, 91125; [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report on X-ray and gamma-ray observations of the millisecond pulsar (MSP) B1937+21 taken with the *Chandra X-ray Observatory*, *XMM-Newton*, and the *Fermi* Large Area Telescope. The pulsar X-ray emission shows a purely non-thermal spectrum with a hard photon index of $0.9\pm0.1$, and is nearly 100% pulsed. We found no evidence of varying pulse profile with energy as previously claimed. We also analyzed 5.5yr of [*Fermi*]{}survey data and obtained much improved constraints on the pulsar’s timing and spectral properties in gamma-rays. The pulsed spectrum is adequately fitted by a simple power-law with a photon index of $2.38\pm0.07$. Both the gamma-ray and X-ray pulse profiles show similar two-peak structure and generally align with the radio peaks. We found that the aligned profiles and the hard spectrum in X-rays seem to be common properties among MSPs with high magnetic fields at the light cylinder. We discuss a possible physical scenario that could give rise to these features.'
author:
- 'C.-Y. Ng, J. Takata, G. C. K. Leung, K. S. Cheng, and P. Philippopoulos'
title: 'HIGH-ENERGY EMISSION OF THE FIRST MILLISECOND PULSAR'
---
INTRODUCTION
============
Millisecond pulsars (MSPs) are fast-spinning neutron stars with rotation periods ($P$) from a few to tens of milliseconds. They are believed to be old pulsars spun up through accretion process from a companion star. MSPs generally show small spin-down rates () that imply low surface magnetic fields ($B_s\propto\sqrt{P\dot P}$) of the order of $10^8$G, much lower than those of young pulsars. Nonetheless, MSPs exhibit broadband emission as young pulsars do, and are detected across the electromagnetic spectrum from radio to X-ray to gamma-ray bands. Recently there has been significant progress in the high-energy studies of MSPs. The Large Area Telescope (LAT) onboard the *Fermi Gamma-ray Space Telescope* has detected over 60 MSPs in gamma-rays since 2008 [see the LAT second pulsar catalog; @aaa+13]. In addition, the *Chandra X-ray Observatory* and the *XMM-Newton* mission have provided sensitive measurements of the X-ray properties of over a dozen MSPs [see @hnk13 and references therein]. These have significantly expanded the sample and allowed detailed studies of the population.
In this study we focus on the X-ray and gamma-ray emission properties of , which is a representative of an emerging class of MSPs that shows aligned pulse profiles in different energy bands [@gjv+12]. This could possibly indicate a different emission mechanism than that of typical pulsars. [PSR B1937+21]{} (also known as PSR J1939+2134) is the first MSP discovered [@bkh+82] and it remains the second fastest-spinning pulsar known, with $P=1.56$ms and $\dot P=1.05\times10^{-19}$. (Corrections of due to proper motion and differential Galactic rotation are negligible). These timing parameters suggest $B_s=4.1\times10^8$G and spin-down power $\dot E=4\pi^2
I\dot P/P^3=1.1\times10^{36}$[ergs$^{-1}$]{}, where $I=10^{45}$gcm$^2$ is neutron star moment of inertia. This is the second largest among MSPs, only after PSR B1821$-$24.
At high energies, [PSR B1937+21]{} was first detected in X-rays with *ASCA* [@tst+01]. The emission has a non-thermal spectrum and exhibits strong pulsations. The pulse profile shows a narrow peak structure with a pulsed fraction (PF) of 44%. The study also suggests a hint of a second peak in the profile. Subsequent observations with *RXTE* and *BeppoSAX* confirmed the second peak [@chk+03; @ncl+04], and it was claimed that the relative strength of the two peaks and the PF could vary with energy, from PF$=85\%\pm5$% in 1.3–10keV to PF$=54\%\pm7$% in 4–10keV [@ncl+04]. @zav07 reported on a [*Chandra*]{} observation of the source and found an X-ray photon index $\simeq1.2$. Using 1.5yr of [*Fermi*]{} survey data, @gjv+12 detected gamma-ray pulsations from [PSR B1937+21]{}. The pulse profile well aligns with the radio profile, indicating that both emission could originate from the same region in the outer magnetosphere. Spectral analysis using the [*Fermi*]{} data suggests that the phase-averaged pulsar spectrum could be fitted with an exponentially cutoff power-law model [@gjv+12]. However, note that the spectral parameters of this source are not listed in the LAT second pulsar catalog that uses three years of data, due to low detection significance with a test-statistic (TS) value of 10 only [@aaa+13].
[lcccccc]{} [*Chandra*]{}& 2005 Jun 28 & 5516 & ACIS-S & Faint & 3.2 & 49.5\
[*XMM*]{}& 2010 Mar 29 & 0605370101 & MOS1 & Full Frame & 2.6 & 40.4\
& & & MOS2 & Full Frame & 2.6 & 47.5\
& & & PN & Timing & $3\times10^{-5}$ & 40.4
We present a new study of the high-energy emission of [PSR B1937+21]{} using archival X-ray data made with [*Chandra*]{} and [*XMM*]{} and 5.5yr of [*Fermi*]{} survey data. The observations and data reduction are described in Section \[s2\] and the analysis and results are presented in Section \[s3\]. In Section \[s4\], We compare the results with other MSPs and discuss possible physical emission mechanisms. Our findings are summarized in Section \[s5\].
OBSERVATIONS AND DATA REDUCTION {#s2}
===============================
We reprocessed the archival [*Chandra*]{} and [*XMM*]{} data. The former, which have been used in the previous study [@zav07], were taken on 2005 Jun 28 using the ACIS-S detector in the imaging mode and have a time resolution of 3.2s. The [*XMM*]{} observation was taken on 2010 March 29 with the MOS1 and MOS2 cameras in the full frame mode that has a frame time of 2.6s and the PN camera in the timing mode that has a high time resolution of 0.03ms. We performed the [*Chandra*]{} and [*XMM*]{} data reduction using CIAO 4.5 and SAS 12.0, respectively. In the [*XMM*]{} analysis, only `PATTERN` $\leq 12$ events from MOS and `PATTERN` $\leq 4$ events from PN were used to ensure good data quality. After removing periods of high background, we obtained net exposures of 49.5ks, 40.4ks, 47.5ks, and 40.4ks from [*Chandra*]{}, MOS1, MOS2, and PN, respectively. The observation parameters are listed in Table \[table:obs\].
For the gamma-ray analysis, we selected [*Fermi*]{} LAT Pass 7 reprocessed data [P7REP; @bcw+13] taken between 2008 August 4 and 2014 January 17. Class 2 events in the P7REP\_SOURCE\_V15 instrument response function (IRFs) were used throughout this paper, and the data reduction was carried out with the [*Fermi*]{} Science Tools v9r32p5. We restricted the analysis on data with zenith angles less than 100, with spacecraft rocking angle less than 52, and in the 0.1–100GeV energy range.
ANALYSIS AND RESULTS {#s3}
====================
X-Ray Analysis
--------------
[PSR B1937+21]{} is clearly detected in both the [*Chandra*]{} and the [*XMM*]{} observations, and the radial profiles of the images are consistent with a point source, same as what @zav07 found. Using a 3 radius aperture centered on the pulsar, we found $610\pm25$ background-subtracted [*Chandra*]{} counts from the in the 0.5–7keV energy range. Similarly, we extracted $370\pm20$ and $430\pm22$ net counts from MOS1 and MOS2, respectively, using a 20-radius aperture. The PN data have higher background since they were taken in the timing mode. We obtained $1230\pm70$ source counts from a 33-wide rectangular region in 0.5–7keV.
### Timing
We performed timing analysis with only the PN data, as they have a high time resolution. The photon arrival times were first corrected to the Solar system barycenter, then folded using the TEMPO2 photons plug-in[^1] with the ephemeris from radio timing[^2] [@sgc+08]. The resulting X-ray profile in the 0.5–7keV energy range is shown in Figure \[fig:profile\]. We have tried other energy bands between 0.5 and 10keV, and found no energy dependence. The profile exhibits very sharp main pulse and interpulse components $\sim$180 apart. In addition, there is a hint of a third peak in between, albeit it is not statistically significant. A direct comparison with the background level obtained from a nearby region indicates that the emission is nearly 100% pulsed. We also plotted the 1.4GHz radio profile in the figure for comparison. The X-ray pulse and interpulse slightly lag the radio ones.
To quantitatively measure the peak position and width, we followed @aaa+13 to fit the pulse profile using an unbinned maximum likelihood method. We employed a simple model with two asymmetric Lorentzians. The functional form is given by $$g(x)= \frac{2}{\pi}\frac{A_1}{(\sigma_1^-+\sigma_1^+)(1+z_1^2)}
+\frac{2}{\pi}\frac{A_2}{(\sigma_2^-+\sigma_2^+)(1+z_2^2)}~,$$ where $$z_i = \left\{
\begin{array}{lr}
(x-\Phi_i) / \sigma_i^- & \mbox{if } x\leq \Phi_i \\
(x-\Phi_i) / \sigma_i^+ & \mbox{if } x > \Phi_i
\end{array}
\right .~,$$ $A_i$ are the amplitudes, $\Phi_i$ are the peak positions, and $\sigma_i^\pm$ are the width parameters such that the full widths at half-maximum (FWHMs) of the two peaks are given by $\sigma_1^-+\sigma_1^+$ and $\sigma_2^-+\sigma_2^+$. We found that the main pulse and interpulse peak at phase $0.051\pm0.001$ and $0.583^{+0.006}_{-0.005}$, respectively.[^3] The best-fit results with statistical uncertainties at the 1$\sigma$ level are listed in Table \[table:results\] and the model profile is plotted in Figure \[fig:profile\]. The uncertainties are estimated using Monte Carlo simulations. We generated 1000 random realizations of the best-fit model and fitted each one to obtain distributions of the best-fit parameters. The confidence intervals are determined from the most compact regions that contain 68% of the sample. Note that we did not account for any systematic uncertainties, such as the absolute timing accuracy of the instrument. Finally, we searched for bursts that may correspond to the giant pulses and any flux variabilities at longer timescale, but found negative result.
[lccc]{} Position, $\Phi_1$ & $0.054\pm0.001$ & $0.003^{+0.001}_{-0.003}$\
Width parameter, $\sigma_1^-$ & $0.010^{+0.001}_{-0.002}$ & $0.006^{+0.001}_{-0.004}$\
Width parameter, $\sigma_1^+$ & $0.019^{+0.001}_{-0.003}$ & $0.012^{+0.001}_{-0.005}$\
FWHM$_1=\sigma_1^-+\sigma_1^+$ & $0.029^{+0.001}_{-0.003}$ & $0.019^{+0.002}_{-0.006}$\
Radio lag, $\delta_1$ & $0.068\pm0.001$ & $-0.011^{+0.001}_{-0.004}$\
Position, $\Phi_2$ & $0.583\pm0.005$ & $0.540^{+0.006}_{-0.005}$\
Width parameter, $\sigma_2^-$ & $0.012^{+0.003}_{-0.007}$ & $0.015^{+0.001}_{-0.007}$\
Width parameter, $\sigma_2^+$ & $0.020^{+0.001}_{-0.009}$ & $0.023^{+0.001}_{-0.010}$\
FWHM$_2=\sigma_2^-+\sigma_2^+$ & $0.032^{+0.003}_{-0.012}$ & $0.018^{+0.002}_{-0.012}$\
Radio lag, $\delta_2$ & $0.032_{-0.005}^{+0.006}$ & $0.003^{+0.006}_{-0.005}$\
Column density, [$N_{\rm H}$]{} ($10^{22}$[cm$^{-2}$]{}) & $1.2\pm0.2$ &\
Photon index, $\Gamma$ & $0.9\pm0.1$ & $2.38\pm0.07$\
Statistic & $\chi^2_\nu=0.86$ & TS-value=112\
Unabsorbed energy flux ($10^{-12}$ & $0.23_{-0.03}^{+0.04}$ & $16\pm2$\
[ergcm$^{-2}$s$^{-1}$]{}) &&\
Luminosity ($10^{33}$[ergs$^{-1}$]{}) & 0.7 & 49\
Efficiency, $\eta$ & $6\times10^{-4}$ & 0.044
### Spectroscopy
We extracted the pulsar spectrum from the regions stated above. Backgrounds were obtained from nearby regions on the same chip. As the pulsar is 100% pulsed, we extracted the PN spectrum from the on-pulse phase intervals only, between phase 0–0.15 and 0.52–0.65, in order to boost the signal-to-noise ratio. Note that there is a faint source CXO J193939.3+213506 located 13 northeast of the pulsar, which is not resolved by [*XMM*]{}. However, there should be negligible contamination to the pulsar spectrum, since its [*Chandra*]{} count rate in 0.5–7keV is only 2% of that of the pulsar. This is also supported by the $\sim$100% PF of the PN counts.
The spectral fitting was performed in the Sherpa environment. We grouped the ACIS and MOS spectra to at least 20 counts per bin, and the PN spectrum to at least 30 counts per bin. All four spectra (ACIS, MOS1, MOS2, and PN) were fitted jointly in the 0.5–7keV energy range. We employed a simple absorbed power-law model with abundances and the absorption cross sections given by @wam00. This provides a good fit with a reduced $\chi^2$ value of 0.86 over 122 degrees of freedom. The best-fit parameters are listed in Table \[table:results\] and the model is shown in Figure \[fig:spec\]. The fit gives a column density of $N_{\rm H}=(1.2\pm 0.2)\times10^{22}$[cm$^{-2}$]{} and a photon index of $\Gamma=0.9\pm0.1$ (all uncertainties are at the $1\sigma$ confidence level). The absorbed and unabsorbed fluxes in 0.5–7keV are $(1.8\pm0.3)\times10^{-13}$[ergcm$^{-2}$s$^{-1}$]{} and $2.3_{-0.3}^{+0.4}\times
10^{-13}$[ergcm$^{-2}$s$^{-1}$]{}, respectively. For the source distance of 5kpc [@vwc+12], this converts to X-ray luminosity of $L_X=6.8\times10
^{32}$[ergs$^{-1}$]{}, implying an efficiency of $L_X/\dot E=6\times10^{-4}$. To check the cross-calibration, we fit the [*Chandra*]{} and [*XMM*]{} data separately and obtained $N_{\rm H}=1.3_{-0.4}^{+0.5}\times10^{22}$[cm$^{-2}$]{}with $\Gamma_X=0.8\pm0.1$ from the former and $N_{\rm H}=(1.1\pm 0.2)\times10^{22}$[cm$^{-2}$]{}with $\Gamma_X=1.0\pm0.2$ from the latter. These values are fully consistent.
We have also tried a blackbody model, but the fit is slightly worse ($\chi^2_\nu=0.95$) and the temperature seems too high ($kT=1.5$keV) to be physical. Moreover, adding a blackbody component to the power-law model shows no significant improvement to the fit. Nonetheless, we can still post a limit on the thermal emission by adjusting the blackbody parameters until the $\chi^2$ value exceeded a certain level. We obtained a temperature limit of $kT<0.13$keV at the 99% confidence level, for any thermal emission from the polar cap region with a radius $r_{\rm pc}=R_\ast\sqrt{2\pi R_\ast /
cP}=3.7$km, where $R_\ast=10$km is the neutron star radius.
Gamma-Ray Analysis
------------------
### Timing
Using a 1 radius region of interest (ROI) centered on the pulsar position, $3.7\times10^4$ events in the 0.1–100GeV energy range were extracted from the [*Fermi*]{} observation. We applied a barycentric correction to the photon arrival times and folded them using the TEMPO2 fermi plug-in, with the same radio ephemeris as for the X-ray analysis. Pulsations are clearly detected and the $H$-test [@db10] gives a statistic of 78, which corresponds to 7.6$\sigma$ detection. The pulse profile is shown in Figure \[fig:profile\]. We have also tried the 1–100GeV energy range and the profile looks very similar, albeit with a lower $H$-statistic of 19 only.
The gamma-ray profile in the figure shows sharp main pulse and interpulse that resemble the X-ray ones, but the gamma-ray pulses better align with the radio peaks. We fitted the gamma-ray profile with the same algorithm as described in the X-ray analysis above and the best-fit parameters and statistical uncertainties are listed in Table \[table:results\]. Finally, we checked the long term gamma-ray flux evolution of the source, and found no significant variability.
### Spectroscopy
We performed a binned likelihood analysis using a $20\arcdeg\times20\arcdeg$ square ROI. Our model includes all sources in the Second Fermi-LAT Source Catalog [2FGL; @naa+12] within 20 from the pulsar, the Galactic diffuse emission (gll\_iem\_v05.fits), and the extragalactic isotropic emission (iso\_source\_v05.txt). The source spectral models are all adopted from 2FGL. For sources located more than 8 from the pulsar, their spectral parameters are held fixed during the fit. We have tried a power-law model and a power-law with exponential cutoff model (PLEC) for the pulsar spectrum. The latter has a functional form of $$\frac{dN}{dE}\propto E^{-\Gamma}\exp\left(-\frac{E}{E_c}\right)~,$$ where $N$ is the photon flux and $E_c$ is the cutoff energy. We found that both models give low TS values that correspond to a detection significance below 3$\sigma$ level.
[lccccccccc]{} B1821$-$24 &3.1&22&23&7.3&5.1&$1.23\pm0.03$&14&$82.5\pm4$&1\
B1937+21 &1.6&11 &4.1 &10& 5.0 & $0.9\pm0.1$&6.8&$\sim100$ & This work\
B1820$-$30A&5.4&8.3&43&2.5& 7.9&&&&\
J1701$-$3006F&2.3&7.3&7.2& 5.5&6.9&&&&\
J1701$-$3006E&3.2&3.6&10& 2.7&6.9&&&&\
J0218+4232 &2.3&2.4&4.3 &3.1& 2.7 & $1.10\pm0.06$ &3.3&$64\pm6$&2, 3\
B1957+20&1.6&1.1&1.4 &3.0& 2.5 & $\sim2$& $\sim0.5$&& 4\
J1750$-$3703D &5.1&1.4&16& 1.1& 12 & $<50$ &&& 5\
B0021$-$72F&2.6&1.4&4.1&2.1&4.0&&0.055& & 6\
J1740$-$5340A&3.7&1.4&7.9&1.5& 3.4&$1.73\pm0.08$ & 0.22&& 7\
J1701$-$3006D&3.4&1.2&6.6&1.5& 6.9&&&&
To boost the signal, we restricted the analysis on the on-pulse data only, in phase 0.5–0.65 and 0.95–1.1. These intervals are indicated in Figure \[fig:profile\]. We obtained a much higher TS value of 112, i.e., above 10$\sigma$ significance, using the power-law model. The fit gives a photon index of $\Gamma_\gamma=2.38\pm0.07$. The best-fit parameters and the phase-averaged energy flux are listed in Table \[table:results\]. We have also tried the PLEC model and found a sightly better fit (TS = 116) with $\Gamma_\gamma=2.1\pm0.2$ and $E_c=8\pm4$GeV. However, the likelihood ratio test gives $-2\Delta\log(\mbox{likelihood})=\Delta \mbox{TS} <9$. We therefore follow the convention in @aaa+13 to conclude that the improvement is not significantly preferred. The best-fit power-law and PLEC models are plotted in Figure \[fig:fermi\].
DISCUSSION {#s4}
==========
Comparison with Previous Studies
--------------------------------
The [*Chandra*]{} and [*XMM*]{} observations offer better angular resolution and sensitivity than previous X-ray studies with *ASCA*, *RXTE*, and *BeppoSAX*, providing substantial improvements on the spectral and timing measurements. While $\Gamma_X$ we obtained is consistent with the reported values, we found a lower $N_\mathrm{H}$ of $1.2\times10^{22}$[cm$^{-2}$]{}compared to $\sim2\times10^{22}$[cm$^{-2}$]{} given by @tst+01, @chk+03, and @ncl+04. Our result is closer to the total Galactic H[i]{} column density of $1.1\times 10^{22}$[cm$^{-2}$]{} in the direction measured with radio observations [@kbh+05]. We however note that it is not unusual for sources near the Galactic plane to have an X-ray absorption column density larger than the total H[i]{} column density, because X-rays are mostly absorbed by molecular clouds instead of neutral hydrogen atoms [see @hnk13]. The small [$N_{\rm H}$]{} value results in slight smaller unabsorbed flux of $3.0\times
10^{-13}$[ergcm$^{-2}$s$^{-1}$]{} in 2–10keV, as compared to $\sim4\times10^{-13}$[ergcm$^{-2}$s$^{-1}$]{}in @tst+01 and @ncl+04.
For the timing analysis, the [*XMM*]{} PN observation allows us to isolate the pulsar flux from the background emission, revealing for the first time the near 100% PF of the X-ray emission. This is much higher than $\sim$60% found with *ASCA* [@tst+01]. Our results show no evidence of energy-dependent pulse profile in the soft X-ray band, thus, rejecting the claim of a lower ($54\%\pm7$%) PF below 4keV [@ncl+04]. The X-ray peak positions we obtained are mostly consistent with the published values. While a direct comparison with the *RXTE* results may seem to indicate slight offsets ($\Delta \phi_1=0.0028\pm 0.001$, $\Delta \phi_2=0.0078\pm 0.005$), we note that these are smaller than the 40$\mu$s absolute timing accuracy of [*XMM*]{} [@mkc+12], which corresponds to 0.03 in phase units.
[PSR B1937+21]{} is one of the few pulsars with radio giant pulses detected. @chk+03 noticed phase alignment between the radio giant and X-ray pulses, suggesting a possible correlation. Although there is no contemporaneous radio observation accompanied with the [*XMM*]{} timing data, a large number of giant pulses are expected during the 11hr long X-ray exposure, given their observed rate of a few hundreds to a few thousands per hour [e.g., @spb+04; @zps+13]. We found no clustering of X-ray photon arrival times, same as what @chk+03 concluded. Similar negative results have been reported for other pulsars that show giant pulses [e.g., @jrm+04; @bmk+12].
Thanks to more [*Fermi*]{} data and the improved IRF [see @bcw+13], our gamma-ray timing results are fully consistent with, but better constrained than, those from @gjv+12. For the spectral analysis, we found a slightly harder photon index $\Gamma_\gamma=2.38\pm0.07$ compared to $3.02\pm0.18$ [@gjv+12]. On the other hand, our PLEC fit gives a softer $\Gamma_\gamma$ ($2.1\pm0.2$ versus $1.43\pm0.87$), although the two are formally compatible given the large uncertainties of the latter. We also obtained a higher cutoff energy at $E_c=8\pm4$GeV as compared to $E_c=1.2\pm0.7$GeV. However, we note that $E_c$ is not very well determined in both cases.
High-, High-[$B_{\rm lc}$]{} MSPs
---------------------------------
The non-thermal spectrum and large PF of the X-ray emission of [PSR B1937+21]{} indicate a magnetospheric origin. These characteristics seem common among high-MSPs [see @zav07]. We list in Table \[table:highedot\] all known MSPs with $\dot E>10^{35}$[ergs$^{-1}$]{}. Three of them, PSRs B1821$-$24, B1937+21, and J0218+4232, have confirmed X-ray pulsations.[^4] They all exhibit hard power-law spectrum in X-rays with $\Gamma_X\sim 1$, large PF of 64%–100%, general alignment between the X-ray and radio pulses [@khv+02; @jgk+13], and giant radio pulses [@rj01; @wcs84; @kbm+06]. It has been suggested that these properties could be attributed to a strong magnetic fields at the light cylinder. Its strength, [$B_{\rm lc}$]{}, is related to the surface dipole field $B_s$ by $B_{\rm lc}=B_s(R_\ast/R_{\rm lc})^3$, where $R_{\rm lc}=Pc/2\pi$ is the light cylinder radius. All MSPs listed in the table have strong [$B_{\rm lc}$]{} above $10^5$G. In particular, [PSR B1937+21]{} has the highest [$B_{\rm lc}$]{} of $1.0\times 10^6$G, even higher than that of the Crab pulsar.
[lcccccc]{} B1937+21&1.6&110&4.1&9.9&PL &II\
B1821$-$24&3.1&220&23&7.3&PL&II\
J0218+4232&2.3&24&4.3&3.2&PL&II\
J1747$-$4036&1.7&12&1.5& 3.1&&I/II\
B1957+20&1.6&7.6&1.2&2.5&&II\
B1820$-$30A&5.4&83&43& 2.5&&II\
J1902$-$5105&1.7&6.7&1.3 &2.2&&II\
J1810+1744&1.7&4.0&0.9&1.8& &II\
J1125$-$5825&3.1&8.1&4.4& 1.4&&I\
J1446$-$4701&2.2&3.7&1.5&1.3& &I\
J2215+5135&2.6&5.2&2.5&1.3& &I\
J2241$-$5236&2.2&3.3&1.4&1.2& &I\
J1658$-$5324&2.4&3.0&1.7&1.1& &I\
J0034$-$0534&1.9&1.7&0.75&1.0&&II\
J1124$-$3653&2.4&1.6&1.2&0.78& &III\
J0614$-$3329&3.2&2.2&2.4&0.71& &III\
J2043+1711&2.4&1.3&1.0&0.70&&I\
J0102+4839&3.0&1.8&1.9&0.67& &III\
J2214+3000&3.1&1.9&2.2&0.66& &II/III\
J1858$-$2216&2.4&1.1&1.0&0.66& &III\
J0023+0923&3.1&1.5&1.8&0.60& &I\
J2017+0603&2.9&1.3&1.6&0.59& &I\
J1741+1351&3.8&2.2&3.3&0.58&&III\
J0101$-$6422&2.6&1.0&1.1&0.58&&I\
J0613$-$0200&3.1&1.2&1.7&0.53&&I\
J1514$-$4946&3.6&1.6&2.6&0.52& &I\
J0751+1807&3.5&0.72&1.7&0.36&&1\
J0340+4130&3.3&0.65&1.4&0.36& & II/III\
J2047+1053&4.3&1.1&3.0&0.35& &III\
J1600$-$3053&3.6&0.73&1.8&0.35&&I\
J1614$-$2230&3.2&0.38&1.0&0.29&&I\
J1231$-$1411&3.7&0.51&1.6&0.29&&I\
J2051$-$0827&4.5&0.54&2.4&0.24&&I\
J1744$-$1134&4.1&0.41&1.7&0.23&&III\
J1713+0747&4.6&0.34&2.0&0.19&&I\
J2124$-$3358&4.9&0.37&2.4&0.18&BB&III\
J0030+0451&4.9&0.35&2.3&0.18&BB+PL&I\
J2302+4442&5.2&0.38&2.7&0.18& &I/II\
J0437$-$4715&5.8&0.29&2.9&0.14&BB+PL&I\
J0610$-$2100&3.9&0.08&0.69&0.11&&III\
J1024$-$0719&5.2&0.05&0.92&0.06&&I
In Table \[table:fermimsp\] we list all gamma-ray MSPs detected with [*Fermi*]{}LAT in decreasing order of [$B_{\rm lc}$]{}. We classified them according to their phase alignment between the gamma-ray and radio pulses, using a scheme similar to @vjh12: sources with gamma-ray peak lagging, aligned with, and preceding the radio peak are divided into classes I, II, and III, respectively. Class II pulsars are required to have a radio peak aligned with the gamma-ray peak to better than 1/10 of the spin period, and the two peaks should have a similar profile. There are a small number of pulsars, e.g., PSR J2214+3000, that cannot be unambiguously classified. These cases are noted in the table. Finally, we also list in the table the X-ray spectral type of the MSPs, although only a few of them are detected since MSPs are generally faint. It is obvious that most high-[$B_{\rm lc}$]{} MSPs belong to class II and they exhibit non-thermal X-ray emission with hard spectra.
Modeling the Emission of High-[$B_{\rm lc}$]{} MSPs
---------------------------------------------------
@gjv+12 first noted that a group of MSPs show aligned radio, X-ray, and gamma-ray pulse profiles, suggesting the same location for the emission. The authors proposed that the radio emission could be generated in caustics in the outer magnetosphere, same as the gamma-ray emission. As Table \[table:fermimsp\] indicates, the phase alignment is typical among the highest-[$B_{\rm lc}$]{} MSPs. We discuss below possible emission mechanisms in different energy bands, with a focus on their connection with [$B_{\rm lc}$]{}. We qualitatively compare our toy model with the spectral energy distributions (SEDs) of three MSPs — PSRs B1937+21, B1821$-$24, and J0218+4232, which have been detected in both X-rays and gamma-rays.
### Gamma-Ray Emission
We briefly summarize the gamma-ray emission mechanism in the context of the outer gap model [@chr86], in which electrons and positrons are accelerated up to a Lorentz factor of $\gamma\sim 10^7$ near the light cylinder. For a more detailed calculation of the gamma-ray emission process, we refer to @wtc10 [@wtc11] and @tct12. In the outer gap, electrons and positrons are accelerated by the electric field along the magnetic field lines and emit GeV gamma-rays via the curvature radiation process. The magnitude of the electric field is given by $$E_\parallel\sim \frac{f_{\rm gap}B_{\rm lc}R_{\rm lc}}{R_c}~,$$ where $R_c$ is the curvature radius and $f_{\rm gap}$ is the ratio between the gap thickness and the light cylinder radius, typically $f_{\rm gap}\sim0.3$ for MSPs [@tct12].
The Lorentz factor of the accelerated particles can be estimated by the balance between the electric force and the back reaction force of the curvature radiation, $$\begin{aligned}
\gamma_p =& \left(\frac{3R_c^2}{2e}E_\parallel\right)^{1/4}
\sim 5\times 10^6\left(\frac{f_{\rm gap}}{0.3}\right)^{1/2}
\left(\frac{P}{1\,{\rm ms}}\right)^{1/2}\nonumber \\
&\times \left(\frac{B_{\rm lc}}{10^5\,{\rm G}}\right)^{1/4}
\left(\frac{R_{c}}{R_{\rm lc}}\right)^{1/4}~.\end{aligned}$$ The energy of the curvature photons is of the order of GeV, $$\begin{aligned}
E_c=&\frac{3hc\gamma^3}{4\pi R_c}
\sim 0.8\left(\frac{f_{\rm gap}}{0.3}\right)^{1/4}
\left(\frac{P}{1\,{\rm ms}}\right)^{1/2}\nonumber \\
&\times \left(\frac{B_{\rm lc}}{10^5\,{\rm G}}\right)^{3/4}
\left(\frac{R_c}{R_{\rm lc}}\right)^{-1/4}\,{\rm GeV}\end{aligned}$$ and the gamma-ray luminosity from the outer gap is typically $$\begin{aligned}
L_{\gamma}\sim &f_{\rm gap}^3\dot E\sim 1.2\times 10^{32}
\left(\frac{f_{\rm gap}}{0.3}\right)^{3}
\left(\frac{P}{1\,{\rm ms}}\right)^{2}\nonumber \\
&\times \left(\frac{B_{\rm lc}}{10^5\,{\rm G}}\right)^{2}\,{\rm erg\,s^{-1}}~.\end{aligned}$$
### X-Ray Emission
In the standard pulsar theory, non-thermal X-ray emission is attributed to synchrotron radiation from secondary pairs in the outer magnetosphere, which are generated via the pair-creation process between gamma-rays and the background X-rays [e.g., @tc07]. The secondary pairs have an initial Lorentz factor of $\gamma_{\rm max}\sim (1\,{\rm GeV})/(2m_ec^2)\sim 10^3$, then quickly lose their energy via synchrotron radiation and eventually leave the light cylinder with $\gamma_{\rm min}\sim 1/\sin\theta_s\sim 10$, where $\theta_s$ is the pitch angle. During this process, synchrotron cooling gives non-thermal X-rays with $\Gamma_X\sim 1.5$ between $E_{\rm
min}\sim2\times10^{-2}\sin\theta_s (B_{lc}/10^5{\rm G})$keV to $E_{\rm
max}\sim5\sin\theta_s(B_{\rm lc}/10^5\,{\rm G})$keV. This is the general case for young pulsars; they have $\Gamma_X$ observed in the range of 1.5–2 [e.g., @khc+01; @hsg+02].
For high-[$B_{\rm lc}$]{} MSPs, while non-thermal X-rays can be produced near the light cylinder, the very hard photon indices ($\Gamma_X\sim1$) are difficult to explain. A hard spectrum is expected below $E_{\rm min}$, but this is outside the observation bands. @hum05 proposed that the primary particles in MSPs could maintain large momenta and undergo cyclotron resonant absorption of radio emission to produce synchrotron X-rays with a very hard spectrum. However, this process is not efficient between particles with $\gamma\sim10^7$ and radio waves of 0.1–1GHz, unless the $B$-field is of the order of $10^7$–$10^8$G. Such a condition can only occur near the neutron star surface.
One possible scenario to explain the non-thermal X-rays from the outer magnetosphere is inverse-Compton (IC) scattering between the primary particles and radio waves. As we discussed, the radio, X-ray, and gamma-ray emission regions of high-[$B_{\rm lc}$]{} MSPs are likely co-located. Therefore, the radio waves emitted in the outer magnetosphere may possibly irradiate the outer gap region and be up-scattered by the ultra-relativistic particles. We can estimate the energy density of the radio waves in the magnetosphere from the flux density. Assuming a radio spectral index of $\alpha=2$, the typical energy density of $\sim$100MHz radio waves is $U_{\rm ph}\sim10^4$–$10^
5$ergcm$^{-3}$ in high-[$B_{\rm lc}$]{} MSPs. The radiation power of IC scattering from a single particle is then $P_{\rm IC}\sim4\sigma_Tc\gamma^2U_{\rm ph}/3$, where $\sigma_T$ is the Thomson cross section. As a comparison, the power of the curvature radiation is $P_{\rm cur}\sim2e^2c\gamma^4/3R_c^2$. Hence, $$\frac{P_{\rm IC}}{P_{\rm cur}}\sim 0.1 \left(\frac{U_{\rm ph}}{10^4\,{\rm
erg\,cm^{-3}}}
\right)\left(\frac{\gamma}{10^7}\right)^{-2}\left(\frac{R_c}{10^7\,{\rm cm}}
\right)^2~.$$ This suggests that the expected energy flux from the IC process is only slightly smaller than that of the GeV emission and it could be observable in high-[$B_{\rm lc}$]{} MSPs.
To model the IC spectrum, we consider the scattering between outgoing particles in the outer gap and outwardly propagating radio waves. We assume that the radio emission region lies just above the outer gap and approximate the magnetic field lines by concentric circles. The collision angle between the primary particles and radio waves can be crudely estimated by $\sin
\theta_0\sim\sqrt{2f_{\rm gap}}$. For each particle, the IC power per unit energy per unit solid angle is given by $$\frac{dP_{\rm IC}}{d\Omega}\sim {\mathcal D}^2(1-\beta\cos\theta_0)F_{\rm rad}
\frac{d\sigma'}{d\Omega'},$$ where $d\sigma'/d\Omega'$ is the differential Klein-Nishina cross section, $\beta=\sqrt{\gamma^2-1}/\gamma$, $\mathcal{D}=\gamma^{-1}(1-\beta\cos
\theta_1)^{-1}$, $0<\theta_1<1/\gamma$ is the angle between the particle motion and the scattered photon direction, and $F_{\rm rad}$ is the radio spectrum.
The IC spectrum depends sensitively on the radio spectrum. Although the latter is not very clear at low radio frequency $<1$GHz, it is believed that a spectral turnover should exist below 100MHz for MSPs [see @kl01], which is lower than that of young pulsars [$\sim1$GHz; e.g., @kgk07]. For a turnover at 10–100MHz, the corresponding break in the IC spectrum would be at $(1-10)\times (\gamma/10^7)^2$MeV. This is well above the [*Chandra*]{} and [*XMM*]{} energy bands, suggesting that the IC emission could contributed to the observed X-rays. To compare with observations, we assume a broken power-law spectrum in radio with a turnover at 10MHz, i.e. $$F_{\rm rad}(\nu)=A\left \{ \begin{array}{@{\,}ll}
\left(\frac{\nu}{100\,{\rm MHz}}\right)^{\beta_1} & \mathrm{for}~ \nu\ge
10{\rm MHz} \\
\left(\frac{10\,{\rm MHz}}{100\,{\rm MHz}}\right)^{\beta_1}
\left(\frac{\nu}{10\,{\rm MHz}}\right)^{\beta_2} & \mathrm{for}~\nu< 10\,{\rm
MHz}.
\end{array} \right . \label{norm}$$
The spectral index $\beta_1$ is inferred from the observed flux densities at 400MHz and 1.4GHz listed in the ATNF pulsar catalog [@mht+05]. The index below 10MHz, $\beta_2$, is taken to be 0.5, which provides a good match to the X-ray spectra (see below). The normalization $A$ is related to the observed flux density at 400MHz ($F_{400}$) and the source distance $d$ by $$A\sim F_{400}\left(\frac{d}{R_{\rm lc}}\right)^2
\left(\frac{100{\rm MHz}}{400{\rm MHz}}\right)^{\beta_1}~.$$ As the exact geometry of the radio emission region is unknown, we allowed $A$ to vary by a factor of a few to fit the observed SEDs.
We also considered synchrotron X-rays contributed by the secondary pairs. The particles accelerated towards the star would eventually reach the stellar surface and heat up the polar cap region. @tct12 estimated a surface temperature of $\sim 1$MK and a luminosity of $L_r\sim5\times 10^{31}$[ergs$^{-1}$]{}for the thermal emission of PSR B1937+21. These values are below our detection limit of 1.5MK. The thermal X-rays may collide with the GeV gamma-rays to create new pairs, which emit non-thermal X-rays via the synchrotron process. The optical depth of the pair-creation process is $\tau_{X\gamma}\sim L_r\sigma_{X\gamma}R_{\rm lc}/(4\pi R_{\rm lc}^2c kT_r)
\sim 0.02$, where $k_B$ is the Boltzmann constant and $\sigma_{X\gamma}\sim
\sigma_T/3$. The energy distribution of the secondary pairs follows $$\frac{dN_e}{d\gamma_s}(\gamma_s)\sim \frac{m_ec^2}{\dot{E}_{\rm syn}}
\int_{2\gamma_sm_ec^2}^{\infty}Q(E'_{\gamma}) dE'_{\gamma}~,
\label{sdist}$$ where $\gamma_s$ is the Lorentz factor of the secondary pairs, $Q(E_{\gamma})
=F_{\rm cur}(1-e^{-\tau_{X\gamma}})/E_{\gamma}$, $F_{\rm cur}$ is the curvature radiation power per unit energy, and $\dot{E}_{\rm syn}$ is the rate of energy loss from synchrotron radiation. With $\dot{E}_{\rm
syn}=2e^4B^2\sin^2\theta_{s} \gamma_s^2/3m_e^2c^3$, Equation (\[sdist\]) describes a power-law with index $p=2$. The synchrotron spectrum from secondary pairs is $$F_{\rm syn}(E_{\gamma})=\frac{\sqrt{3}e^3B\sin\theta_s}{hm_ec^2}\int
\frac{dN_e}{d\gamma_s}F(x)d\gamma_s~,$$ where $x=E_{\gamma}/E_{\rm syn}$, $E_{\rm syn}=3he\gamma_s^2B\sin\theta_s/4\pi
m_ec$, and $\sin \theta_s\sim \sqrt{2f_{\rm gap}}$.
In Figure \[fig:sed\] we compare our simple model with the observed SEDs of PSRs B1937+21, B1821$-$24, and J0218+4232. It shows that our model provides reasonable fits to the data. The peak flux of the IC and curvature radiation are similar and the IC emission dominates over the synchrotron radiation in the X-ray band above a few keV. Since the IC spectrum is sensitive to the radio spectral index, low-frequency radio measurements below 100MHz in future can offer essential inputs to refine the modeling. For the assumed spectral turnover at 10MHz, the IC emission peaks at $\sim$100keV, which could be detectable with hard X-ray telescopes such as *NuSTAR* or *ASTRO-H*.
### Radio Emission
The radio emission process in pulsars is not clearly understood, nonetheless, it has been suggested that plasma instability could play an important role in the generation of the coherent radio emission [e.g., @uso87]. The characteristic timescale for the instabilities (e.g., two-stream instability) to develop is related to the inverse of the plasma frequency $\omega_p=\sqrt{
4\pi e^2n_e/m_e}$. The electron and positron number density, $n_e$, is proportional to the Goldreich-Julian charge density ($n_{\rm GJ}$) such that $n_e\propto n_{\rm GJ}\sim B_{\rm lc}/(2\pi Pc)$. Therefore, high-[$B_{\rm lc}$]{} MSPs tend to have a higher plasma frequency, and hence a shorter instability timescale, in the outer magnetosphere.
Above the outer gap accelerator, the outgoing and incoming gamma-rays produce electron-positron pairs moving towards the light cylinder and the star, respectively. These flows could develop two-stream instability in the outer magnetosphere. Since there are more pairs created near the inner boundary than near the outer boundary of the outer gap [@tsh04], the outflow is stronger than the inflow at the light cylinder. The Lorentz factor of the two flows are of the order of $\gamma_s\sim 10$, for which the synchrotron cooling timescale is comparable to light-crossing time of the magnetosphere. The instability development timescale can be estimated by $$\tau_i\sim \left(\frac{n_o}{n_i}\right)^{1/3}\gamma_s^{3/2}\omega_{p,o}^{-1}~,$$ where $n_o$ and $n_i$ are the number densities of the outgoing and incoming flows, respectively, and $\omega_{p,o}$ is the plasma frequency of the outgoing flow [@uso87]. With a typical multiplicity of the order of $\kappa\sim 10^3$, we may assume $n_o\sim \kappa n_{\rm GJ}$ and $n_i\sim
n_{\rm GJ}$ in the outer magnetosphere. Hence, the timescale becomes $$\tau_{i}\sim 1.4\times 10^{-8}\left(\frac{\kappa}{10^{3}}\right)^{-1/6}
\left(\frac{P}{1{\,\rm ms}}\right)^{1/2}
\left(\frac{B_{\rm lc}}{10^5{\,\rm G}}\right)^{-1/2}{\,\rm s}~,$$ which is much shorter than the light-crossing time of $\tau_c\sim R_{\rm
lc}/c\sim 1.6\times 10^{-4}(P/1{\,\rm ms})$s. As a result, the instability could develop before the outgoing particles escape the magnetosphere. We speculate that the two-stream instability may generate non-homogeneous and separated plasma clouds, and the scattering or emission process of the plasma cloud eventually produce the observed radio emission in phase with the gamma-ray pulses.
Finally, we note that some high-[$B_{\rm lc}$]{} MSPs, e.g., PSRs B1821$-$24 and J1810+1744, show complex radio profiles with additional components that offset from the gamma-ray peaks. These could be contributed by emission from the polar cap region as in the conventional theory of pulsar radio emission.
CONCLUSIONS {#s5}
===========
We have analyzed X-ray and gamma-ray observations of [PSR B1937+21]{} taken with [*Chandra*]{}, [*XMM*]{} and [*Fermi*]{} LAT. We obtained much improved spectral and timing measurements than previous studies. Our results show that the pulsar X-ray emission is $\sim$100% pulsed and has a purely non-thermal spectrum that can be described by a hard power-law of photon index $\Gamma_X=0.9\pm 0.1$. The X-ray pulse profile consists of two sharp peaks $\sim$180 apart. They generally align with the radio peaks and the phase offsets are less than 7% of the spin period. In gamma-rays, the 5.5yr of [*Fermi*]{} survey data provide a good quality pulse profile in 0.1–100GeV with a significance of over 7$\sigma$. We performed a binned likelihood analysis on the pulsed emission and found that a simple power-law model with $\Gamma_\gamma=2.38\pm0.07$ gives a TS value of 112, corresponding to over 10$\sigma$ significance. Adding an exponential cutoff to the power-law model slightly improves the fit, but the change is not statistically significant.
A comparison of [PSR B1937+21]{} with other MSPs indicates that sources with a strong magnetic field at the light cylinder tend to show a hard, non-thermal X-ray spectrum and good alignment of pulse profiles in different energy bands. The latter suggests that the radio, X-ray, and gamma-ray emission could originate from the same region in the outer magnetosphere. We speculate that radio emission could be generated in the outer gap region when [$B_{\rm lc}$]{} is large, since this could give rise to short instability time scales. We investigate a simple model in which the non-thermal X-rays are contributed by IC scattering between radio waves and primary particles in the outer magnetosphere and by synchrotron radiation from secondary particles. We showed that this toy model is capable to qualitatively reproduce the observed SEDs of the highest-[$B_{\rm lc}$]{}MSPs. Future observations at low radio frequencies and in hard X-rays can help refine the modeling.
We thank the referee for careful reading and useful suggestions and thank Vicky Kaspi and Anne Archibald for useful discussions. JT, GCKL, and KSC are supported by a GRF grant of Hong Kong Government under HKU7009/11P.
[*Facilities:*]{} , ,
natexlab\#1[\#1]{}
, A. A., [Ajello]{}, M., [Allafort]{}, A., [et al.]{} 2013, , 208, 17
, D. C., [Kulkarni]{}, S. R., [Heiles]{}, C., [Davis]{}, M. M., & [Goss]{}, W. M. 1982, , 300, 615
, W., [Huang]{}, H. H., & [Prinz]{}, T. 2010, arXiv:1006.0335
, A. V., [McLaughlin]{}, M. A., [Kondratiev]{}, V. I., & [Ransom]{}, S. M. 2012, , 749, 24
, S., [Grindlay]{}, J. E., [Heinke]{}, C. O., [et al.]{} 2006, , 646, 1104
, S., [van den Berg]{}, M., [Heinke]{}, C. O., [et al.]{} 2010, , 709, 241
, S., [van den Berg]{}, M., [Servillat]{}, M., [et al.]{} 2011, , 730, 81
, J., [Charles]{}, E., & [Wood]{}, . 2013, in Proceedings of the 4th Fermi Symposium, Monterey, California, 2012, eConf C121028, ed. [T. J. Brandt, N. Omodei, & C. Wilson-Hodge]{}, (arXiv:1304.5456)
, K. S., [Ho]{}, C., & [Ruderman]{}, M. 1986, , 300, 500
, G., [Hermsen]{}, W., [Kramer]{}, M., [et al.]{} 2003, , 410, L9
, O. C., & [B[ü]{}sching]{}, I. 2010, , 517, L9
, L., [Johnson]{}, T. J., [Venter]{}, C., [et al.]{} 2012, , 744, 33
, A. K., [Strickman]{}, M. S., [Gwinn]{}, C., [et al.]{} 2002, , 576, 376
, A. K., [Usov]{}, V. V., & [Muslimov]{}, A. G. 2005, , 622, 531
, C., [Ng]{}, C.-Y., & [Kaspi]{}, V. M. 2013, , 768, 64
, R. H. H., [Kong]{}, A. K. H., [Takata]{}, J., [et al.]{} 2012, , 760, 92
, T. J., [Guillemot]{}, L., [Kerr]{}, M., [et al.]{} 2013, , 778, 106
, S., [Romani]{}, R. W., [Marshall]{}, F. E., & [Zhang]{}, W. 2004, , 355, 31
, P. M. W., [Burton]{}, W. B., [Hartmann]{}, D., [et al.]{} 2005, , 440, 775
, J., [Gupta]{}, Y., & [Krzeszowski]{}, K. 2007, , 462, 699
, H. S., [Bailes]{}, M., [Manchester]{}, R. N., [Ord]{}, S. M., & [Jacoby]{}, B. A. 2006, , 640, 941
, L., [Hermsen]{}, W., [Cusumano]{}, G., [et al.]{} 2001, , 378, 918
, L., [Hermsen]{}, W., [Verbunt]{}, F., [et al.]{} 2002, , 577, 917
, A. D., & [Losovsky]{}, B. Y. 2001, , 368, 230
, R. N., [Hobbs]{}, G. B., [Teoh]{}, A., & [Hobbs]{}, M. 2005, , 129, 1993
, A., [Kirsch]{}, M. G. F., [Caballero]{}, I., [et al.]{} 2012, , 545, A126
, L., [Cusumano]{}, G., [L[ö]{}hmer]{}, O., [et al.]{} 2004, , 413, 1065
, P. L., [Abdo]{}, A. A., [Ackermann]{}, M., [et al.]{} 2012, , 199, 31
, R. W., & [Johnston]{}, S. 2001, , 557, L93
, D. A., [Guillemot]{}, L., [Camilo]{}, F., [et al.]{} 2008, , 492, 923
, V. A., [Popov]{}, M. V., [Bartel]{}, N., [et al.]{} 2004, , 616, 439
, M., [Shibata]{}, S., [Torii]{}, K., [et al.]{} 2001, , 554, 316
, J., & [Chang]{}, H.-K. 2007, , 670, 677
, J., [Cheng]{}, K. S., & [Taam]{}, R. E. 2012, , 745, 100
, J., [Shibata]{}, S., & [Hirotani]{}, K. 2004, , 354, 1120
, V. V. 1987, , 320, 333
, C., [Johnson]{}, T. J., & [Harding]{}, A. K. 2012, , 744, 34
, J. P. W., [Weisberg]{}, J. M., [Chael]{}, A. A., [Lee]{}, K. J., & [Lorimer]{}, D. R. 2012, , 755, 39
, Y., [Takata]{}, J., & [Cheng]{}, K. S. 2010, , 720, 178
—. 2011, , 414, 2664
, N. A., [Olive]{}, J.-F., & [Barret]{}, D. 2004, , 417, 181
, J., [Allen]{}, A., & [McCray]{}, R. 2000, , 542, 914
, A., [Cordes]{}, J., & [Stinebring]{}, D. 1984, in Birth and Evolution of Neutron Stars: Issues Raised by Millisecond Pulsars, ed. S. P. [Reynolds]{} & D. R. [Stinebring]{} (Green Bank: NRAO), 63
Zavlin, V. E. 2007, , 308, 297
, V. I., [Popov]{}, M. V., [Soglasnov]{}, V. A., [et al.]{} 2013, , 430, 2815
[^1]: [http://www.physics.mcgill.ca/\$\\sim\$aarchiba/photons\_plug.html](http://www.physics.mcgill.ca/$\sim$aarchiba/photons_plug.html)
[^2]: The timing parameter file is obtained from the LAT second pulsar catalog <http://fermi.gsfc.nasa.gov/ssc/data/access/lat/2nd_PSR_catalog/>.
[^3]: Phase 0 is defined by the maximum of the first Fourier harmonic of the radio profile in the time domain [@gjv+12].
[^4]: @gjv+12 reported evidence of X-ray pulsations from PSR B1957+20, but the overall X-ray emission could be dominated by intra-binary shock emission [@hkt+12].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Bragg coherent x-ray diffractive imaging is a powerful technique for investigating dynamic nanoscale processes in nanoparticles immersed in reactive, realistic environments. Its temporal resolution is limited, however, by the oversampling requirements of 3D phase retrieval. Here we show that incorporating the entire measurement time series, which is typically a continuous physical process, into phase retrieval allows the oversampling requirement at each time step to be reduced leading to a subsequent improvement in the temporal resolution by a factor of 2-20 times. The increased time resolution will allow imaging of faster dynamics and of radiation-dose-sensitive samples. This approach, which we call “chrono CDI," may find use in improving time resolution in other imaging techniques.'
author:
- 'A. Ulvestad'
- 'A. Tripathi'
- 'S. O. Hruszkewycz'
- 'W. Cha'
- 'S. M. Wild'
- 'G. B. Stephenson'
- 'P. H. Fuoss'
title: 'Chrono CDI: Coherent diffractive imaging of time-evolving samples'
---
Introduction {#sec:intro}
============
Understanding nanoscale processes is key to improving the performance of advanced technologies, such as batteries, catalysts, and fuel cells. However, many processes occur inside devices at short length and time scales in reactive environments and represent a significant imaging challenge. Bragg coherent diffractive imaging (BCDI) has emerged as a powerful technique for revealing 3D nanoscale structural information [@Pfeifer2006; @Robinson2009]. With current BCDI methods, 3D image reconstructions of nanoscale crystals have been used to identify and track dislocations [@Ulvestad2015a; @Clark2015a], image cathode lattice strain during battery operation [@Ulvestad2014a; @Ulvestad2015], indicate the presence of surface adsorbates [@Cha2013; @Watari2011b], and reveal twin domains [@Ulvestad2015b; @Huang2015]. The temporal resolution of current BCDI experiments, however, is limited by the oversampling requirements for current phase retrieval algorithms. The insight developed in this work is that, for most physical processes, structural evolution is a continuous process that introduces structural redundancy when measured as a time series. Here, we exploit this redundancy to allow for reduced oversampling (less than the conventionally required factor of 2), thereby improving the measurement rate. In principle, this method can also be used to increase the spatial resolution.
Our new approach, which we call “chrono CDI,” improves the temporal resolution of BCDI by reducing the oversampling requirement along one dimension ($q_z$) at a given time step without significantly compromising image fidelity. To enable this capability, we designed a reconstruction algorithm that simultaneously reconstructs all time states in a series of Bragg rocking curves by utilizing constraints from neighboring time steps. In this work, the initial and final states are assumed to be known in real space. In practice, this situation is achieved provided both states are measured with the required oversampling (OS). The time frame over which the sample is assumed to be static (the measurement time of an individual rocking curve) is reduced in chrono CDI, providing access to faster dynamics in 3D crystals. In addition, this approach can be used to limit the radiation dose in radiation-sensitive samples and/or increase the spatial resolution by allowing for reduced sampling in $q_x$ and $q_y$.
![ \[fig1\] Bragg CDI experiment of a single time-evolving nanocrystal: **(a)** Schematic of a Bragg rocking curve. The rocking curve involves rotating the sample with respect to the incident x-rays $\mathbf{k}_i$. Here, $\mathbf{k}_f$ represents the scattered x-rays, $\mathbf{q}$ the particular scattering vector, and $\mathbf{G}_{hkl}$ the particular Bragg peak. The rocking curve sweeps the 2D area detector through the 3D volume in the $q_z$ direction. Conventionally, all 2D measurements shown by black boxes are required; in order to improve the time resolution, only some 2D measurements (red stars) could be taken. $D_t$ denotes the full 3D diffraction measurement and $\Delta$ the time required to make the full measurement. **(b)** Select time states from the time sequence that was reconstructed from experimental BCDI data measured with greater than the required oversampling. The isosurface is drawn at a constant Bragg electron density and represents the shape of the 85 nm Pd nanocube. The isosurface color is the imaginary part of the image, the phase $\phi$, which is proportional to the $u_{111}$ displacement field.](fig1.pdf)
Figure \[fig1\]a shows a schematic of a Bragg rocking curve. The rocking curve entails rotating the sample with respect to the incident x-ray beam $\mathbf{k}_i$. The sample rotation, labeled schematically by $\theta_{\text{start}}$ and $\theta_{\text{end}}$, displaces the scattering vector $\mathbf{q}=\mathbf{k}_f-\mathbf{k}_i$ from the reciprocal-space lattice point $\mathbf{G}_{hkl}$, the Bragg reflection condition for the HKL lattice planes, so that the 3D intensity distribution (yellow isosurface) can be appropriately sampled and the structure of the nanocrystal (green cube) can be reconstructed. The series of 2D measurements (grey planes in Fig. \[fig1\]a) are stacked to form a 3D dataset $D_t$, where $t$ represents a time index in a series of sequential rocking curve measurements. The total time for the measurement is $\Delta$. Current phase retrieval algorithms require an oversampling of at least 2 of the diffraction pattern in all three dimensions. We refer to an oversampling of 2 as the required oversampling. In addition, the nanocrystal must be approximately static over the measurement time $\Delta$ while $D_t$ is collected, which limits the dynamic timescale that can be observed [@Xiong2014; @Clark2013].
Figure \[fig1\]b shows three select time states from a time series during which a single crystal palladium nanocube (85 nm side length) is exposed to hydrogen gas. Experimental details are given in a recent publication [@Ulvestad2015NC]. The absolute value of the image (shown as an isosurface) corresponds to the Bragg-diffracting electron density [@Ulvestad2015b], while the phase, $\phi$, of the image (color projected onto the isosurface) is proportional to a component of the vector displacement field $\mathbf{u}$ via $\phi=\mathbf{u}\cdot\mathbf{Q}$ [@Pfeifer2006; @Chapman2006a; @Newton2010]. In this case, the Pd (111) Bragg peak was measured and $\phi\sim u_{111}$. In the Pd nanocube, hydrogen intercalation initially causes displacement field changes ($t=42$ minutes) before morphological changes occur ($t=76$ minutes) due to the hydriding phase transformation [@Wicke1996; @Flanagan1991]. The time evolution of the nanocube structure shown in Fig. \[fig1\]b was determined from BCDI experiments performed with an oversampling of 3 in $q_z$ at Sector 34-ID-C of the Advanced Photon Source at Argonne National Laboratory (see Experimental Details in Supplemental Material and Ulvestad, et al.[@Ulvestad2015NC]). Each complete measurement took approximately 2 minutes ($\Delta$ in Fig. \[fig1\]).
Algorithmic Approach {#sec:alg}
====================
To incorporate the redundancy in correlated time series such as those in Fig. \[fig1\], we modify conventional BCDI phase retrieval algorithms. The function minimized by the error reduction phase retrieval algorithms is the modulus error, $\varepsilon_{{\scalebox{0.5}{$\mathcal{M}$}}}^2$, which measures the agreement between the reconstruction’s Fourier moduli and the measured moduli, $$\varepsilon_{{\scalebox{0.5}{$\mathcal{M}$}}}^2(\rho,D)
=
\sum_{{\mathbf{q}}} \Big \vert |\tilde{\rho}| - \sqrt{D} \Big\vert^2,
\label{eq:moderror}$$where $\rho$ is the 3D reconstructed object (in real space), $D$ is the 3D far-field intensity measurement, ${\mathbf{q}}$ is the reciprocal-space coordinate, $\tilde{\rho}={\mathcal{F}}\left[ \rho \right]$, and ${\mathcal{F}}$ is the Fourier transform. Different choices of the function to be minimized lead to different phase retrieval algorithms [@Marchesini2007a]. In chrono CDI, we include a term that depends on reconstructions at other time states, $$\sum_t \bigg[
\varepsilon_{{\scalebox{0.5}{$\mathcal{M}$}}}^2(\rho_t, D_t) +
\sum_{t'\neq t}
w(t,t')
{\mu}(\rho_t,\rho_{t'})
\bigg].
\label{eq:genobjective}$$ In this expression, $t$ indexes the time states, $w(t,t')$ is the weight, $t\neq t'$, and $\mu(\rho_t,\rho_{t'})$ is the miscorrelation term. In this paper, we consider nearest-neighbor correlations in time, a scalar weight parameter $w\geq 0$, and a functional form for the miscorrelation of $$\mu(\rho_t,\rho_{t'})= \sum_{{\mathbf{r}}} \big \vert \rho_t-\rho_{t-1} \big\vert^2
+ \sum_{{\mathbf{r}}} \big \vert \rho_t-\rho_{t+1} \big\vert^2.$$ The $t$th term of the objective in Eq. (\[eq:genobjective\]) then becomes $$ \varepsilon_{{\scalebox{0.5}{$\mathcal{M}$}}}^2(\rho_t, D_t)
+
w \left( \sum_{{\mathbf{r}}} \big \vert \rho_t-\rho_{t-1} \big\vert^2 + \sum_{{\mathbf{r}}} \big
\vert \rho_t-\rho_{t+1} \big\vert^2
\right).
\label{eq:objective}$$ Although other forms are possible, this form has the advantage of being computationally inexpensive. The iterative algorithm is derived by minimizing Eq. (\[eq:objective\]) summed over $t$ (for details, see the Supplemental Material and refs.[@Marchesini2007a; @Tripathi2015]).
Numerical Results {#sec:results}
=================
To evaluate the algorithm’s performance, we carry out iterative phase retrieval on noise-free, simulated data as well as on measured experimental data with different amounts of oversampling.
Simulated Data with Required Oversampling {#sec:simdata}
-----------------------------------------
In this case, the simulated data $D_t^{\rm sim}$ is generated by 3D Fourier transforms of each complex valued Pd nanocube reconstruction in the time series after zero padding to meet the oversampling (OS) requirements of phase retrieval [@Livet2007]. We refer to an oversampling of 2 as the required oversampling (Req. O.S.). In practice, this means the cube size was half of the array size in all three dimensions. The time sequence considered consists of $t=0,12,18,26,34,42,50,58,66$, and $76$ minutes. This time sequence is approximately equally spaced and, as shown in Fig. \[nncorr\], has varying amounts of nearest-neighbor correlation, with an average nearest-neighbor correlation coefficient of 61%. The correlation coefficient $c(t,t') \in [-1,1]$ is defined between two 3D displacement fields at time states $t$ and $t'$ by $$\footnotesize
\frac{ {\sum\limits_{\mathbf{r}} } \big[ u_{111}({\mathbf{r}},t) -
\bar{u}_{111}({\mathbf{r}},t)
\big] \big[ u_{111}({\mathbf{r}},t') -
\bar{u}_{111}({\mathbf{r}},t')
\big]}
{\sqrt{\sum\limits_{\mathbf{r}} \big[ u_{111}({\mathbf{r}},t) - \bar{u}_{111}({\mathbf{r}},t)
\big]^2} \sqrt{\sum\limits_{\mathbf{r}} \big[ u_{111}({\mathbf{r}},t') - \bar{u}_{111}({\mathbf{r}},t')
\big]^2}} \nonumber
$$ where $u_{111}$ is the displacement field projection and $\bar{u}_{111}$ is the average displacement field over the particle.
![Nearest-neighbor correlation coefficient of $u_{111}(\mathbf{r})$ for all pairs in the chosen time sequence. The average nearest-neighbor correlation coefficient is 61%.[]{data-label="nncorr"}](nncorrsequenceplot.pdf)
Figure \[fig2\]a shows the correlation coefficient matrix for the chosen time sequence. Figure \[nncorr\] is a plot of the super-diagonal matrix values. The chosen time sequence is a good balance between having smooth evolution between nearest neighbors and having a large change over the whole time sequence (both the displacement and the amplitude change significantly). The first numerical test of chrono CDI reconstructs the sequence $\rho_t$ from the sequence of $D_t^{\rm sim}$ with the required oversampling.
![**(a)** Time correlation in the \[111\] displacement field projection, $u_{111}({\mathbf{r}})$, during the reconstructed time sequence. **(b)**, Average (over all time states and random starts) modulus error $\varepsilon_{{\scalebox{0.5}{$\mathcal{M}$}}}^2 (\rho_t, D^{\rm sim}_t)$ (black) and average (over all time states and random starts) miscorrelation term ${\mu}(\rho_t,\rho_{t'})$ (blue) normalized by the total intensity in the image as a function of the scalar weight. Error bars represent the standard deviation of the average over all time states obtained from 10 different random starts.[]{data-label="fig2"}](avgobjperfdata.pdf)
The algorithm uses with random initial starts and alternates between the error reduction (ER) and the hybrid input-output (HIO) algorithms using a feedback parameter of $\beta=0.7$ [@Fienup1986]; the support is fixed to the size of the object and is not evolved during the iterative process. At iteration numbers $N=100n$, for $n=1, 2, \ldots, 18$, the algorithm tests whether all reconstructions are correctly oriented with respect to the known initial ($t=0$ minutes) and final ($t=76$ minutes) states by testing whether they are conjugated and reflected (“twin") solutions [@Fienup1986]. Although reconstructing the “twin" image does not affect $\varepsilon_{{\scalebox{0.5}{$\mathcal{M}$}}}^2$, it will negatively impact ${\mu}$, resulting in an artificially high total objective. One constraint used in the present work is that the $\rho_t$ of the initial and final states are known in real and diffraction space. This constraint can be achieved by measuring diffraction datasets at the required oversampling before the experimental dynamics start and after no significant changes are seen in the diffraction data.
Figure \[fig2\]b shows the errors $\varepsilon_{{\scalebox{0.5}{$\mathcal{M}$}}}^2
(\rho_t,
D^{\rm sim}_t)$ and ${\mu}(\rho_t,\rho_{t'})$, averaged over all reconstructed time states and over 10 random starts, as a function of the scalar weight $w$. Both errors are normalized by the total intensity in the image. Figure \[sampmod\] in the Supplemental Material shows the modulus error $\varepsilon_{{\scalebox{0.5}{$\mathcal{M}$}}}^2$ as a function of iteration number for two scalar weight values. The initial and final states ($t=0$ and $t=76$ minutes, respectively) are known in real space. When $w=0$, the modulus error is the lowest, and the miscorrelation term is the largest. These results are expected because the data is noise-free and oversampled at the required oversampling such that a unique solution is fully determined for each $D_t^{\rm sim}$. The weight $w=0$ corresponds to the case when no correlations are taken into account. As $w$ increases, ${\mu}(\rho_t,\rho_{t'})$ decreases, and $\varepsilon_{{\scalebox{0.5}{$\mathcal{M}$}}}^2$ increases for the solution set $\{\rho_t\}$ because their relative contributions to the total objective change. With data sampled at the required oversampling, including information from neighboring time steps in a time series will not improve each individual reconstruction because the complete 3D structural description of the sample at each time is uniquely encoded in the 3D coherent intensity pattern.
Simulated Data with Reduced Oversampling {#sec:undersimdata}
----------------------------------------
We now explore how the additional redundancy from the time series can compensate for reduced oversampling during the rocking curve (e.g. oversampling at less than a factor of 2) at a given time step by reconstructing the time series $D_t^{\rm sim}$ discussed previously but with different degrees of reduced sampling in $q_z$. To start, every third 2D diffraction measurement of the original 84 2D diffraction measurements was selected to form $D_t^{\rm sim}$ for all times except the initial and final time. This leads to data that has 1/3 of the required oversampling. If such a time series were measured experimentally, the measurement time would be reduced by a factor of 3. In assessing algorithm performance, the modulus error was calculated by comparing the far-field exit wave of the reconstructions with the data sampled at the required oversampling. As before, the initial and final states are known, the support is known, and alternating ER/HIO is used as described previously.
![Results for reduced oversampling (OS). **(a)** The average (over all time states and random starts) normalized modulus error as a function of the scalar weight $w$ for 1/3 (blue), 1/20 (black), and 1/84 (red) of the required oversampling. $w=0$ (not shown) produces the same average normalized modulus error values as $w=10^{-4}$. The modulus error reported here is computed with respect to the data with the required oversampling. Error bars represent the standard deviation of the different average (over all time states) values obtained from 10 different random starts. Error bar for $w=0.01$ and 1/20th (black) of the required oversampling is offset for clarity. The black dashed horizontal line shows the normalized average modulus error using the average of the initial and final time states for every time state in the sequence. **(b)** The individual time state errors for 1/3 (blue) required oversampling for a particular random start for $w=0$ (open circle), $w=0.01$ (points), and $w=0.5$ (x).[]{data-label="fig3"}](fig3.pdf)
Figure \[fig3\]a shows the average (over all time states and random starts) modulus error as a function of the scalar weight $w$ for varying amounts of oversampling. The average normalized modulus error does not change from $w=0$ to $w=10^{-4}$. Unlike the results using the required oversampling (shown in Fig. \[fig2\]b) where the lowest modulus error occurs for $w=0$ (no time correlation), in the cases where $q_z$ has been sampled at 1/3 (blue) and 1/20 (black) of the required oversampling, a minimum in the modulus error is observed at a value of $w=0.01$. This modulus error is computed with respect to the datasets that are sampled at the required oversampling, and thus the reconstructions at these minima are the “best" solutions. When only 1/84 (red) of the required oversampling is used (i.e., a single slice from the rocking curve), the modulus error decreases with increasing $w$ and approaches a constant value, which is near the normalized average modulus error when all the reconstructed time states are set to the average of the initial and final state (black dashed horizontal line). In this case, simply using an average of the initial and final states outperforms the reconstruction algorithm, indicating that there are insufficient reciprocal space constraints and that the oversampling in $q_z$ is too low.
On average $w=0.01$ improves the reconstructed time sequence relative to $w=0$, which takes no time correlation into account, for up to 1/20 of the required oversampling. However, it is not clear from the plot of the average whether all time states are being improved equally. Figure \[fig3\]b shows the normalized modulus error at each time state for $w=0, 0.01, 0.5$ at 1/3 of the required oversampling for a particular random start. The states nearest to the known states ($t=0$ and $t=76$ minutes) have lower modulus errors, as expected. By comparing $w=0$ with $w=0.01$, we see that the improvement occurs in all the intermediate states except the state at $t=66$ minutes, which remains essentially unchanged. These results demonstrate that the algorithm improves all reconstructions, even those least correlated with their neighbors (see Fig. \[nncorr\] for a plot of the nearest neighbor correlation). The benefits of chrono CDI are clear when the required oversampling is reduced by up to a factor of approximately 1/20. In these cases, enforcing a degree of nearest-time-step, real-space correlation provides an additional constraint that improves the reconstruction at all intermediate times relative to what can be achieved using conventional phase retrieval.
Experimental Data {#sec:expdata}
-----------------
We now demonstrate chrono CDI on experimental rocking curve data. To simulate varying degrees of reduced oversampling, a subset of the original 2D measurements was selected from the experimental datasets. Figure \[realdifdat\] shows example 2D experimental diffraction measurements from the Pd (111) Bragg rocking curve; see Ulvestad, et al. [@Ulvestad2015NC] for more details. An oversampling of approximately 3 in $q_z$ was used during the original measurement. The reconstruction algorithm is the same as described previously except that the support is not known a priori. Instead, an initial box half the array size in each dimension is used, and the support is updated with the shrinkwrap algorithm [@Marchesini2003; @Chapman2006a] using a Gaussian function with a threshold of $0.01$ and standard deviation of 1.
![Reconstructions of experimental measurement data with reduced oversampling. The real part of the image (shown as an isosurface) corresponds to the reconstructed Bragg electron density, while the complex part of the image (colormap projected onto the isosurface) corresponds to the reconstructed displacement field projection. **(a)** The $t=12$ minutes reconstructions for 1/2 of the required oversampling for $w=0$ and $w=0.01$, and the true solution. **(b)** The same as **(a)** but for the $t=42$ minutes reconstruction.[]{data-label="realdat"}](fig5.pdf)
Figure \[realdat\] shows chrono CDI reconstructions for two representative time states of experimental diffraction data from Pd nanocubes undergoing structural transformations when exposed to hydrogen gas. As before, different amounts of oversampling were investigated. The isosurfaces shown correspond to the reconstructed Bragg electron density, while the color map corresponds to the image phase $\phi$, which is proportional to the $u_{111}$ displacement field. Figure \[realdat\]a shows reconstructions when 1/2 of the required oversampling in $q_z$ is used. Every third slice of the original data (oversampled at a factor of 3 in $q_z$) was used to generate data. This corresponds to an oversampling of 1, which is 1/2 the required oversampling of 2. The reconstruction for $w=0.01$ is much improved compared with $w=0$ and is similar in morphology and lattice displacement to the true solution. Figure \[realdat\]b shows that the same conclusion holds for $t=42$ minutes. See Figure \[cxrealdat\] in the Supplemental Material for central cross-sections that show the amplitude and phase distributions inside the crystal. The average normalized modulus error of the time sequence is improved from $0.2$ to $0.1$ by including nearest-neighbor information (via $w=0.01$). Although the reconstructions do not match exactly, the results convey the same overall physical changes in the crystal.
![Reconstructions of real measurement data with 3/10 of the required oversampling and $w=0.01$. The real part of the image (shown as an isosurface) corresponds to the reconstructed Bragg electron density, while the complex part of the image, $\phi$, is proportional to the displacement field projection. Three states from the reconstructed time sequence are shown. []{data-label="15datremsupp"}](15datremsupp.pdf){width="\linewidth"}
Figure \[15datremsupp\] shows that when 3/10 of the required oversampling is used (corresponding to every 5th slice of the original data), major differences in both the reconstructed Bragg electron density and displacement fields arise as compared with the full-rocking-curve reconstructions. There are also disagreements in the reconstructed phases (proportional to the displacements). We therefore conclude that chrono CDI applied to this particular set of measurement data for the chosen time sequence could have decreased the measurement time by a factor of 2 without losing the essential physics of the transforming crystal. For simulated data it could have reduced the time by up to a factor of 20. The discrepancy is likely due to the finite extent in reciprocal space of the real data, noise in the data, and the unknown support that must be determined via the shrinkwrap algorithm during the reconstruction.
Discussion {#sec:discussion}
==========
In this work, we have shown that our new algorithm improves the time resolution of BCDI by a factor of 2 for experimental data and 20 for simulated data. The algorithm thereby enables BCDI investigations of dynamic structural processes in crystals that were previously out of reach and limiting radiation dose in sensitive samples. The time resolution improvements we demonstrate are achieved by reducing the number of 2D measurements made during a 3D Bragg rocking curve, leading to datasets with less than the required oversampling in $q_z$ at each intermediate time step. The rocking curves across the entire time series are reconstructed simultaneously, enforcing a degree of real-space correlation between solutions at neighboring time steps to account for the reduced oversampling of each individual measurement. The algorithm and its variations should be useful for improving the time resolution of other imaging techniques such as ptychography and tomography[@Mohan2015; @Gibbs2015; @ptycho1; @Hruszkewycz2013; @Tripathi2011] where there is a continuous relationship in real space between nearest neighbor time states.
Acknowledgments {#acknowledgments .unnumbered}
===============
This material was based upon work supported by the U.S. Department of Energy, Office of Science, under Contract No. DE-AC02-06CH11357. This research used resources of the Advanced Photon Source, which is a DOE Office of Science User Facility. Work at the Advanced Photon Source was supported by the Office of Basic Energy Sciences (BES). P. H. F., S. O. H., and G. B. S. were supported by Basic Energy Sciences, Division of Materials Sciences and Engineering. A. T. and S. M. W. were supported by the Office of Advanced Scientific Computing Research and the ROMPR project. A. U. was supported by an Argonne Director’s postdoctoral fellowship.
The authors thank Jesse N. Clark for insightful discussions.
[27]{}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 10.1038/nature04867) [****, ()](\doibase 10.1038/nmat2400) @noop [****, ()]{} [****, ()](\doibase 10.1038/nmat4320) [****, ()](\doibase 10.1021/nl501858u) [****, ()](\doibase 10.1039/C5CP00372E) [****, ()](\doibase 10.1038/nmat3698) [****, ()](\doibase 10.1038/nmat3124) [****, ()](\doibase
10.1021/acs.nanolett.5b01104) [****, ()](\doibase
10.1021/acs.nanolett.5b03568) [****, ()](\doibase
10.1038/srep06765) [****, ()](\doibase
10.1126/science.1236034) [****, ()](\doibase 10.1038/ncomms10092) [****, ()](\doibase 10.1364/JOSAA.23.001179) [****, ()](\doibase
10.1038/nmat2607) [**](\doibase 10.1016/0360-3199(95)00126-3), Vol. () pp. @noop [****, ()]{} [****, ()](\doibase 10.1063/1.2403783) [****, ()](\doibase
10.1016/j.procs.2015.05.205) [****, ()](\doibase 10.1107/S010876730605570X) @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevB.68.140101) [****, ()](\doibase
10.1109/TCI.2015.2431913) [****, ()](\doibase 10.1038/srep11824) [****, ()](\doibase 10.1002/adem.201400443) [****, ()](\doibase
10.1103/PhysRevLett.110.177601) [****, ()](\doibase 10.1073/pnas.1104304108)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The purpose of this work is to survey what is known about the linear independence of spikes and sines. The paper provides new results for the case where the locations of the spikes and the frequencies of the sines are chosen at random. This problem is equivalent to studying the spectral norm of a random submatrix drawn from the discrete Fourier transform matrix. The proof depends on an extrapolation argument of Bourgain and Tzafriri.'
author:
- 'Joel A. Tropp'
bibliography:
- 'spikes-sines.bib'
date: '4 September 2007. Revised 15 April 2008.'
title: On the Linear Independence of Spikes and Sines
---
[^1]
Introduction
============
An investigation central to sparse approximation is whether a given collection of impulses and complex exponentials is linearly independent. This inquiry appears in the early paper of Donoho and Stark on uncertainty principles [@DS89:Uncertainty-Principles], and it has been repeated and amplified in the work of subsequent authors. Indeed, researchers in sparse approximation have developed a much deeper understanding of general dictionaries by probing the structure of the unassuming dictionary that contains only spikes and sines.
The purpose of this work is to survey what is known about the linear independence of spikes and sines and to provide some new results on random subcollections chosen from this dictionary. The method is adapted from a paper of Bourgain–Tzafriri [@BT91:Problem-Kadison-Singer]. The advantage of this approach is that it avoids some of the complicated combinatorial arguments that are used in related works, e.g., [@CRT06:Robust-Uncertainty]. The proof also applies to other types of dictionaries, although we do not pursue this line of inquiry here.
Spikes and Sines
----------------
Let us shift to formal discussion. We work in the inner-product space $\Cspace{n}$, and we use the symbol ${}^\adj$ for the conjugate transpose. Define the Hermitian inner product $\ip{\vct{x}}{\vct{y}} = \vct{y}^\adj \vct{x}$ and the $\ell_2$ vector norm $\norm{\vct{x}} = \absip{ \vct{x} }{\vct{x}}^{1/2}$. We also write $\norm{ \cdot }$ for the spectral norm, i.e., the operator norm for linear maps from $(\Cspace{n}, \ell_2)$ to itself.
We consider two orthonormal bases for $\Cspace{n}$. The standard basis $\{ \onevct_j : j = 1, 2, \dots, n \}$ is given by $$\onevct_j(t) = \begin{cases}
1, & t = j \\
0, & t \neq j
\end{cases}
\qquad\text{for $t = 1, 2, \dots, n$}.$$ We often refer to the elements of the standard basis as or . The Fourier basis $\{ \mathbf{f}_j : j = 1, 2, \dots, n \}$ is given by $$\mathbf{f}_j(t) = \frac{1}{\sqrt{n}} \econst^{2\pi\iunit j t/n}
\qquad\text{for $t = 1, 2, \dots, n$}.$$ We often refer to the elements of the Fourier basis as or .
The (DFT) is the $n \times n$ matrix ${\mathbf{F}}$ whose rows are $\mathbf{f}_1^\adj, \mathbf{f}_2^\adj, \dots, \mathbf{f}_n^\adj$. The matrix ${\mathbf{F}}$ is unitary. In particular, its spectral norm $\norm{ {\mathbf{F}}} = 1$. Moreover, the entries of the DFT matrix are bounded in magnitude by $n^{-1/2}$. Let $T$ and $\Omega$ be subsets of $\{1, 2, \dots, n\}$. We write ${\mathbf{F}}_{\Omega T}$ for the restriction of ${\mathbf{F}}$ to the rows listed in $\Omega$ and the columns listed in $T$. Since ${\mathbf{F}}_{\Omega T}$ is a submatrix of the DFT matrix, its spectral norm does not exceed one.
We use the analysts’ convention that upright letters represent universal constants. We reserve $\cnst{c}$ for small constants and $\cnst{C}$ for large constants. The value of a constant may change at each appearance.
Linear Independence
-------------------
Let $T$ and $\Omega$ be subsets of $\{1, 2, \dots, n\}$. Consider the collection of spikes and sines listed in these sets: $$\coll{X} = \coll{X}(T, \Omega) =
\{ \onevct_j : j \in T \} \cup
\{ \mathbf{f}_j : j \in \Omega \}.$$ Today, we will discuss methods for determining when $\coll{X}$ is linearly independent. Since a linearly independent collection in $\Cspace{n}$ contains at most $n$ vectors, we obtain a simple necessary condition $\abs{T} + \abs{\Omega} \leq n$. Developing sufficient conditions, however, requires more sophistication.
We approach the problem by studying the Gram matrix $\mtx{G} = \mtx{G}(\coll{X})$, whose entries are the inner products between pairs of elements from $\coll{X}$. It is easy to check that the Gram matrix can be expressed as $$\mtx{G} = \begin{bmatrix}
\Id_{\abs{\Omega}} & {\mathbf{F}}_{\Omega T} \\
({\mathbf{F}}_{\Omega T})^\adj & \Id_{\abs{T}}
\end{bmatrix}$$ where $\Id_m$ denotes an $m \times m$ identity matrix and $\abs{\cdot}$ denotes the cardinality of a set.
It is well known that the collection $\coll{X}$ is linearly independent if and only if its Gram matrix is nonsingular. The Gram matrix is nonsingular if and only if its eigenvalues are nonzero. A basic (and easily confirmed) fact of matrix analysis is that the extreme eigenvalues of $\mtx{G}$ are $1 \pm \norm{ {\mathbf{F}}_{\Omega T} }$. Therefore, *the collection $\coll{X}$ is linearly independent if and only if $\norm{ {\mathbf{F}}_{\Omega T} } < 1$*.
One may also attempt to quantify the extent to which collection $\coll{X}$ is linearly independent. To that end, define the $\kappa$ of the Gram matrix, which is the ratio of its largest eigenvalue to its smallest eigenvalue: $$\kappa( \mtx{G} ) = \frac{1 + \norm{ {\mathbf{F}}_{\Omega T} }}
{1 - \norm{ {\mathbf{F}}_{\Omega T} }}.$$ If $\norm{ {\mathbf{F}}_{\Omega T} }$ is bounded away from one, then the condition number is constant. One may interpret this statement as evidence the collection $\coll{X}$ is strongly linearly independent. The reason is that the condition number is the reciprocal of the relative spectral-norm distance between $\mtx{G}$ and the nearest singular matrix [@Dem97:Applied-Numerical p. 33]. As we have mentioned, $\mtx{G}$ is singular if and only if $\coll{X}$ is linearly dependent.
This article focuses on statements about linear independence, rather than conditioning. Nevertheless, many results can be adapted to obtain precise information about the size of $\norm{ {\mathbf{F}}_{\Omega T} }$.
Summary of Results
------------------
The major result of this paper to show that a random collection of spikes and sines is extremely likely to be strongly linearly independent, provided that the total number of spikes and sines does not exceed a constant proportion of the ambient dimension. We also provide a result which shows that the norm of a properly scaled random submatrix of the DFT is at most constant with high probability. For a more detailed statement of these theorems, turn to Section \[sec:both-rand\].
Outline
-------
The next section provides a survey of bounds on the norm of a submatrix of the DFT matrix. It concludes with detailed new results for the case where the submatrix is random. Section \[sec:proofs\] contains a proof of the new results. Numerical experiments are presented in Section \[sec:exper\], and Section \[sec:future\] describes some additional research directions. Appendix \[app:extrap\] contains a proof of the key background result.
History and Results
===================
The strange, eventful history of our problem can be viewed as a sequence of bounds on norm of the matrix ${\mathbf{F}}_{\Omega T}$. Results in the literature can be divided into two classes: the case where the sets $\Omega$ and $T$ are fixed and the case where one of the sets is random. In this work, we investigate what happens when both sets are chosen randomly.
Bounds for fixed sets
---------------------
An early result, due to Donoho and Stark [@DS89:Uncertainty-Principles], asserts that an *arbitrary* collection of spikes and sines is linearly independent, provided that the collection is not too big.
\[thm:up\] Suppose that $\abs{T} \abs{\Omega} < n$. Then $\norm{ {\mathbf{F}}_{\Omega T} } < 1$.
The original argument relies on the fact that ${\mathbf{F}}$ is a Vandermonde matrix. We present a short proof that is completely analytic. A similar argument using an inequality of Schur yields the more general result of Elad and Bruckstein [@EB02:Generalized-Uncertainty Thm. 1].
The entries of the $\abs{\Omega} \times \abs{T}$ matrix ${\mathbf{F}}_{\Omega T}$ are uniformly bounded by $n^{-1/2}$. Since the Frobenius norm dominates the spectral norm, $\normsq{ {\mathbf{F}}_{\Omega T} } \leq \fnormsq{ {\mathbf{F}}_{\Omega T} } \leq \abs{\Omega} \abs{T} / n $. Under the hypothesis of the theorem, this quantity does not exceed one.
Theorem \[thm:up\] has an elegant corollary that follows immediately from the basic inequality for geometric and arithmetic means.
\[cor:additive-up\] Suppose that $\abs{T} + \abs{\Omega} < 2 \sqrt{n}$. Then $\norm{ {\mathbf{F}}_{\Omega T} } < 1$.
The contrapositive of Theorem \[thm:up\] is usually interpreted as an : a vector and its discrete Fourier transform cannot simultaneously be sparse. To express this claim quantitatively, we define the $\ell_0$ “quasinorm” of a vector by $\pnorm{0}{\vct{\alpha}} = \abs{\{ j : \alpha_j \neq 0 \} }$.
Fix a vector $\vct{x} \in \Cspace{n}$. Consider the representations of $\vct{x}$ in the standard basis and the Fourier basis: $$\vct{x} = \sum\nolimits_{j = 1}^n \alpha_j \onevct_j
\qquad\text{and}\qquad
\vct{x} = \sum\nolimits_{j = 1}^n \beta_j \mathbf{f}_j.$$ Then $\pnorm{0}{ \vct{\alpha} } \pnorm{0}{ \vct{\beta} } \geq n$.
The example of the shows that Theorem \[thm:up\] and its corollaries are sharp. Suppose that $n$ is a square, and let $T = \Omega = \{ \sqrt{n}, 2\sqrt{n}, 3\sqrt{n}, \dots, n \}$. On account of the Poisson summation formula, $$\sum\nolimits_{j \in T} \onevct_j
= \sum\nolimits_{j \in \Omega} \mathbf{f}_j.$$ Therefore, the set of vectors $\coll{X}(T, \Omega)$ is linearly dependent and $\abs{T}\abs{\Omega} = n$.
The substance behind this example is that the abelian group $\mathbb{Z} / \mathbb{Z}_{n}$ contains nontrivial subgroups when $n$ is composite. The presence of these subgroups leads to arithmetic cancelations for properly chosen $T$ and $\Omega$. See [@DS89:Uncertainty-Principles] for additional discussion.
One way to eradicate the cancelation phenomenon is to require that $n$ be prime. In this case, the group $\mathbb{Z} / \mathbb{Z}_n$ has no nontrivial subgroup. As a result, much larger collections of spikes and sines are linearly independent. Compare the following result with Corollary \[cor:additive-up\].
\[thm:tao-up\] Suppose that $n$ is prime. If $\abs{T} + \abs{\Omega} \leq n$, then $\norm{ {\mathbf{F}}_{\Omega T} } < 1$.
The proof of Theorem \[thm:tao-up\] is algebraic in nature, and it does not provide information about conditioning. Indeed, one expects that some submatrices have norms very near to one.
When $n$ is composite, subgroups of $\mathbb{Z}/\mathbb{Z}_n$ exist, but they have a very rigid structure. Consequently, one can also avoid cancelations by choosing $T$ and $\Omega$ with care. In particular, one may consider the situation where $T$ is clustered and $\Omega$ is spread out. Donoho and Logan [@DL92:Signal-Recovery] study this case using the , a powerful technique from number theory that can be traced back to the 1930s. See the lecture notes [@Jam06:Notes-Large] for an engaging introduction and references.
Here, we simply restate the (sharp) large sieve inequality [@Jam06:Notes-Large LS1.1] in a manner that exposes its connection with our problem. The of a set is measured as the difference (modulo $n$) between the closest pair of indices. Formally, define $${\rm spread}(\Omega) = \min\{ \abs{ j - k \bmod n } : j, k \in \Omega, j \neq k \}$$ with the convention that the modulus returns values in the symmetric range $\{ - \lceil n/2 \rceil + 1, \dots, \lfloor n/2 \rfloor \}$. Observe that $\abs{\Omega} \cdot {\rm spread}(\Omega) \leq n$.
\[thm:large-sieve\] Suppose that $T$ is a block of adjacent indices: $$\label{eqn:T-block}
T = \{ m + 1, m + 2, \dots, m + \abs{T} \}
\qquad\text{for an integer $m$}.$$ For each set $\Omega$, we have $$\norm{ {\mathbf{F}}_{\Omega T} }^2
\leq \frac{ \abs{T} + n/{\rm spread}(\Omega) - 1 }{n}.$$ In particular, when $T$ has form , the bound $\abs{T} + n/{\rm spread}(\Omega) < n + 1$ implies that $\norm{{\mathbf{F}}_{\Omega T}} < 1$.
Of course, we can reverse the roles of $T$ and $\Omega$ in this theorem on account of duality. The same observation applies to other results where the two sets do not participate in the same way.
The discussion above shows that there are cases where delicately constructed sets $T$ and $\Omega$ lead to linearly dependent collections of spikes and sines. Explicit conditions that rule out the bad examples are unknown, but nevertheless the bad examples turn out to be quite rare. To quantify this intuition, we must introduce probability.
Bounds when one set is random
-----------------------------
In their work [@DS89:Uncertainty-Principles Sec. 7.3], Donoho and Stark discuss numerical experiments designed to study what happens when one of the sets of spikes or sines is drawn at random. They conjecture that the situation is vastly different from the case where the spikes and sines are chosen in an arbitrary fashion. Within the last few years, researchers have made substantial theoretical progress on this question. Indeed, we will see that the linearly dependent collections form a vanishing proportion of all collections, provided that the total number of spikes and sines is slightly smaller than the dimension $n$ of the vector space.
First, we describe a probability model for random sets. Fix a number $m \leq n$, and consider the class $\coll{S}_m$ of index sets that have cardinality $m$: $$\coll{S}_m = \{ S : S \subset \{1, 2, \dots, n\}
\text{ and } \abs{S} = m \}.$$ We may construct a random set $\Omega$ by drawing an element from $\coll{S}_m$ uniformly at random. That is, $$\Prob{ \Omega = S } = \abs{\coll{S}_m}^{-1}
\qquad\text{for each $S \in \coll{S}_m$}.$$ In the sequel, we substitute the symbol $\abs{\Omega}$ for the letter $m$, and we say “$\Omega$ is a random set with cardinality $\abs{\Omega}$” to describe this type of random variable. This phrase should cause no confusion, and it allows us to avoid extra notation for the cardinality.
In the sparse approximation literature, the first rigorous result on random sets is due to Cand[è]{}s and Romberg. They study the case where one of the sets is arbitrary and the other set is chosen at random. Their proof draws heavily on their prior work with Tao [@CRT06:Robust-Uncertainty].
\[thm:qrup\] Fix a number $s \geq 1$. Suppose that $$\label{eqn:qrup-bd}
\abs{T} + \abs{\Omega} \leq
\frac{ \cnst{c} n }{\sqrt{ (s + 1) \log n }}.$$ If $T$ is an arbitrary set with cardinality $\abs{T}$ and $\Omega$ is a random set with cardinality $\abs{\Omega}$, then $$\Prob{ \normsq{ {\mathbf{F}}_{\Omega T} } \geq 0.5 }
\leq \cnst{C} ((s + 1) \log n)^{1/2} n^{-s}.$$ The numerical constant $\cnst{c} \geq 0.2791$, provided that $n \geq 512$.
One should interpret this theorem as follows. Fix a set $T$, and consider all sets $\Omega$ that satisfy . Of these, the proportion that are *not* strongly linearly independent is only about $n^{-s}$. One should be aware that the logarithmic factor in is intrinsic when one of the sets is arbitrary. Indeed, one can construct examples related to the Dirac comb which show that the failure probability is constant unless the logarithmic factor is present. We omit the details.
The proof of Theorem \[thm:qrup\] ultimately involves a variation of the moment method for studying random matrices, which was initiated by Wigner. The key point of the argument is a bound on the expected trace of a high power of the random matrix $\sqrt{n/\abs{\Omega}} \cdot {\mathbf{F}}_{\Omega T}^\adj {\mathbf{F}}_{\Omega T} - \Id_{\abs{T}}$. The calculations involve delicate combinatorial techniques that depend heavily on the structure of the matrix ${\mathbf{F}}$.
This approach can also be used to establish that the smallest singular value of ${\mathbf{F}}_{\Omega T}$ is bounded well away from zero [@CRT06:Robust-Uncertainty Thm. 2.2]. This lower bound is essential in many applications, but we do not need it here. For extensions of these ideas, see also the work of Rauhut [@Rau07:Random-Sampling].
Another result, similar to Theorem \[thm:qrup\], suggests that the arbitrary set and the random set do not contribute equally to the spectral norm. We present one version, whose derivation is adapted from [@Tro07:Conditioning-Random Thm. 10 et seq.].
\[thm:rdm-subdict\] Fix a number $s \geq 1$. Suppose that $$\abs{T} \log n + \abs{\Omega} \leq
\frac{\cnst{c} n}{s}.$$ If $T$ is an arbitrary set of cardinality $\abs{T}$ and $\Omega$ is a random set of cardinality $\abs{\Omega}$, then $$\Prob{ \normsq{ {\mathbf{F}}_{\Omega T} } \geq 0.5 }
\leq n^{-s}.$$
The proof of this theorem uses Rudelson’s selection lemma [@Rud99:Random-Vectors Sec. 2] in an essential way. This lemma in turn hinges on the noncommutative Khintchine inquality [@L-P86:Inegalites-Khintchine; @Buc01:Operator-Khintchine]. For a related application of this approach, see [@CR07:Sparsity-Incoherence].
Theorems \[thm:qrup\] and \[thm:rdm-subdict\] are interesting, but they do not predict that a far more striking phenomenon occurs. A random collection of sines has the following property with high probability. To this collection, one can add an *arbitrary* set of spikes without sacrificing linear independence.
Fix a number $s \geq 1$, and assume $n \geq N(s)$. Except with probability $n^{-s}$, a random set $\Omega$ whose cardinality $\abs{\Omega} \leq n/3$ has the following property. For each set $T$ whose cardinality $$\abs{T} \leq \frac{ \cnst{c} n }{s \log^5 n},$$ it holds that $\normsq{ {\mathbf{F}}_{\Omega T} } \leq 0.5$.
This result follows from the (deep) fact that a random row-submatrix of the DFT matrix satisfies the (RIP) with high probability. More precisely, a random set $\Omega$ with cardinality $\abs{\Omega}$ verifies the following condition, except with probability $n^{-s}$. $$\label{eqn:rip}
\frac{ \abs{\Omega} }{ 2n } \leq \normsq{ {\mathbf{F}}_{\Omega T} }
\leq \frac{ 3 \abs{\Omega} }{ 2n }
\qquad\text{when
$\abs{T} \leq \frac{\cnst{c} \abs{\Omega}}{s \log^5 n}$}.$$ This result is adapted from [@RV06:Sparse-Reconstruction Thm. 2.2 et seq.].
The bound was originally established by Cand[è]{}s and Tao [@CT06:Near-Optimal] for sets $T$ whose cardinality $\abs{T} \leq \cnst{c} \abs{\Omega} / s \log^6 n$. Rudelson and Vershynin developed a simpler proof and reduced the exponent on the logarithm [@RV06:Sparse-Reconstruction]. Experts believe that the correct exponent is just one or two, but this conjecture is presently out of reach.
Let $\cnst{c}$ be the constant in . Abbreviate $m = \cnst{c} \abs{\Omega} / s \log^5 n$, and assume that $m \geq 1$ for now. Draw a random set $\Omega$ with cardinality $\abs{\Omega}$, so relation holds except with probability $n^{-s}$. Select an arbitrary set $T$ whose cardinality $\abs{T} \leq \cnst{c} n / 6 s \log^5 n$. We may assume that $2\abs{T} / m \geq 1$ because $\abs{\Omega} \leq n/3$. Partition $T$ into at most $2\abs{T} / m$ disjoint blocks, each containing no more than $m$ indices: $T = T_1 \cup T_2 \cup \dots \cup T_{2\abs{T} / m}$. Apply to calculate that $$\normsq{ {\mathbf{F}}_{\Omega T} }
\leq \frac{ 2\abs{T}}{m}
\max\nolimits_k \normsq{ {\mathbf{F}}_{\Omega T_k} }
\leq \abs{T} \cdot \frac{ 2s \log^5 n }{ \cnst{c} \abs{\Omega} }
\cdot \frac{ 3 \abs{\Omega} }{ 2n }
\leq \frac{1}{2}.$$ Adjusting constants, we obtain the result when $\abs{\Omega}$ is not too smal.
In case $m < 1$, draw a random set $\Omega$ and then draw additional random coordinates to form a larger set $\Omega'$ for which $\cnst{c} \abs{\Omega'} / s \log^5 n \geq 1$ and $\abs{\Omega'} \leq n/3$. This choice is possible because $n \geq N(s)$. Apply the foregoing argument to $\Omega'$. Since the spectral norm of a submatrix is not larger than the norm of the entire matrix, we have the bound $\normsq{ {\mathbf{F}}_{\Omega T} } \leq \normsq{ {\mathbf{F}}_{\Omega' T} } \leq 0.5$ for each sufficiently small set $T$.
Bounds when both sets are random {#sec:both-rand}
--------------------------------
To move into the regime where the number of spikes and sines is proportional to the dimension $n$, we need to randomize both sets. The major goal of this article is to establish the following theorem.
\[thm:both-rand\] Fix a number $\eps > 0$, and assume that $n \geq N(\eps)$. Suppose that $$\abs{T} + \abs{\Omega} \leq \cnst{c}(\eps) \cdot n.$$ Let $T$ and $\Omega$ be random sets with cardinalities $\abs{T}$ and $\abs{\Omega}$. Then $$\Prob{ \normsq{ {\mathbf{F}}_{\Omega T} } \geq 0.5 }
\leq \exp\bigl\{ -n^{1/2 - \eps} \bigr\}.$$ The constant $\cnst{c}(\eps) \geq \econst^{-\cnst{C}/\eps}$.
Note that the probability bound here is superpolynomial, in contrast with the polynomial bounds of the previous section. The estimate is essentially optimal. Take $\eps > 0$, and suppose it were possible to obtain a bound of the form $$\Prob{ \norm{ {\mathbf{F}}_{\Omega T} } = 1 }
\leq \exp\{ - n^{1/2 + \eps} \}
\qquad\text{where}\qquad
\abs{T} + \abs{\Omega} \leq 2n^{1/2}.$$ According to Stirling’s approximation, there are about $\exp\{ n^{1/2} \log n \}$ ways to select two sets satisfying the cardinality bound. At the same time, the proportion of sets that are linearly dependent is at most $\exp\{-n^{1/2 + \eps} \}$. Multiplying these two quantities, we find that no pair of sets meeting the cardinality bound is linearly dependent. This claim contradicts the fact that the Dirac comb yields a linearly dependent collection of size $2n^{1/2}$.
\[rem:both-rand\] As we will see, Theorem \[thm:both-rand\] holds for every $n \times n$ matrix $\mtx{A}$ with constant spectral norm and uniformly bounded entries: $$\norm{\mtx{A}} \leq 1
\qquad\text{and}\qquad
\abs{a_{\omega t}} \leq n^{-1/2}
\quad\text{for $\omega, t = 1, 2, \dots, n$.}$$ The proof does not rely on any special properties of the discrete Fourier transform.
Random matrix theory
--------------------
Finally, we consider an application of this approach to random matrix theory. Note that each column of ${\mathbf{F}}_{\Omega T}$ has $\ell_2$ norm $\sqrt{\abs{\Omega}/n}$. Therefore, it is appropriate to rescale the matrix by $\sqrt{n/\abs{\Omega}}$ so that its columns have unit norm. Under this scaling, it is possible that the norm of the matrix explodes when $\abs{\Omega}$ is small in comparison with $n$. The content of the next result is that this event is highly unlikely if the submatrix is drawn at random.
\[thm:both-rand-norm\] Fix a number $\delta \in (0, \cnst{c})$. Suppose that $n \geq N(\delta)$ and that $$\abs{T} \leq \abs{\Omega} = \delta n.$$ If $T$ and $\Omega$ are random sets with cardinalities $\abs{T}$ and $\abs{\Omega}$, then $$\Prob{ \sqrt{\frac{n}{\abs{\Omega}}} \norm{ {\mathbf{F}}_{\Omega T} } \geq 9 }
\leq n^{-\cnst{C}}.$$
For $\delta$ in the range $[\cnst{c}, 1]$, it is evident that $$\sqrt{\frac{n}{\abs{\Omega}}} \norm{ {\mathbf{F}}_{\Omega T} }
\leq \cnst{c}^{-1}.$$ Therefore, we obtain a constant bound for the norm of a normalized random submatrix throughout the entire parameter range.
Theorem \[thm:both-rand-norm\] also holds for the class of matrices described in Remark \[rem:both-rand\].
Norms of random submatrices {#sec:proofs}
===========================
In this section, we prove Theorem \[thm:both-rand\] and Theorem \[thm:both-rand-norm\]. First, we describe some problem simplifications. Then we provide a moment estimate for the norm of a very small random submatrix, and we present a device for extrapolating a moment estimate for the norm of a much larger random submatrix. This moment estimate is used to prove a tail bound, which quickly leads to the two major results of the paper.
Reductions {#sec:reductions}
----------
Denote by $\mtx{P}_{\delta}$ a random $n \times n$ diagonal matrix where exactly $m = \lfloor \delta n \rfloor$ entries equal one and the rest equal zero. This matrix can be seen as a projector onto a random set of $m$ coordinates. With this notation, the restriction of a matrix $\mtx{A}$ to $m$ random rows and $m$ random columns can be expressed as $\mtx{P}_{\delta} \mtx{A} \mtx{P}_{\delta}'$, where the two projectors are statistically independent from each other.
\[lem:square-case\] Let $\mtx{A}$ be an $n \times n$ matrix. Suppose that $T$ and $\Omega$ are random sets with cardinalities $\abs{T}$ and $\abs{\Omega}$. If $\delta \geq \max\{\abs{T}, \abs{\Omega}\} / n$, then $$\Prob{ \norm{ \mtx{A}_{\Omega T} } \geq u } \leq
\Prob{ \norm{ \mtx{P}_\delta \mtx{A} \mtx{P}_{\delta}' } \geq u }
\quad\text{for $u \geq 0$}.$$
It suffices to show that the probability is weakly increasing as the cardinality of one set increases. Therefore, we focus on $\Omega$ and remove $T$ from the notation for clarity. Let $\Omega$ be a random subset of cardinality $\abs{\Omega}$. Conditional on $\Omega$, we may draw a uniformly random element $\omega$ from $\Omega^c$, and put $\Omega' = \Omega \cup \{ \omega \}$. This $\Omega'$ is a uniformly random subset with cardinality $\abs{\Omega} + 1$. We have $$\begin{aligned}
\Prob{ \norm{\mtx{A}_{\Omega}} \geq u }
&= \Expect I( \norm{\mtx{A}_{\Omega}} \geq u ) \\
&\leq \Expect I( \norm{\mtx{A}_{\Omega \cup \{\omega\}}} \geq u ) \\
&= \Expect I( \norm{\mtx{A}_{\Omega'}} \geq u ) \\
&= \Prob{ \norm{ \mtx{A}_{\Omega'}} \geq u }\end{aligned}$$ where we have written $I(E)$ for the indicator variable of an event. The inequality follows because the spectral norm is weakly increasing when we pass to a larger matrix, and so we have the inclusion of events $\{ \Omega' : \norm{\mtx{A}_{\Omega}} \geq u \} \subset \{ \Omega' : \norm{\mtx{A}_{\Omega \cup \{\omega\}}} \geq u\}$.
It can be inconvenient to work with projectors of the form $\mtx{P}_\delta$ because their entries are dependent. We would prefer a model where coordinates are selected independently. To that end, denote by $\mtx{R}_{\delta}$ a random $n \times n$ diagonal matrix whose entries are independent 0–1 random variables of mean $\delta$. This matrix can be seen as a projector onto a random set of coordinates with *average* cardinality $\delta n$. The following lemma establishes a relationship between the two types of coordinate projectors. The argument is drawn from [@CR06:Quantitative-Robust Sec. 3].
\[lem:rand-coords\] Fix a number $\delta$ in $[0, 1]$. For every $n \times n$ matrix $\mtx{A}$, $$\Prob{ \norm{ \mtx{P}_{\delta} \mtx{A} }
\geq u }
\leq 2 \Prob{ \norm{ \mtx{R}_{\delta} \mtx{A} }
\geq u }
\quad\text{for $u \geq 0$.}$$ In particular, $$\Prob{ \norm{ \mtx{P}_{\delta} \mtx{A} \mtx{P}_{\delta}' }
\geq u }
\leq 4 \Prob{ \norm{ \mtx{R}_{\delta} \mtx{A} \mtx{R}_{\delta}' }
\geq u }
\quad\text{for $u \geq 0$.}$$
Given a coordinate projector $\mtx{R}$, denote by $\sigma(\mtx{R})$ the set of coordinates onto which it projects. For typographical felicity, we use $\#\sigma(\mtx{R})$ to indicate the cardinality of this set.
First, suppose that $\delta n$ is an integer. For every $u \geq 0$, we may calculate that $$\begin{aligned}
\Prob{ \norm{ \mtx{R}_{\delta} \mtx{A} } \geq u }
&\geq \sum\nolimits_{j = \delta n}^n
\Prob{ \norm{ \mtx{R}_{\delta} \mtx{A} } \geq u
\ | \ \# \sigma( \mtx{R}_{\delta} ) = j }
\cdot \Prob{ \# \sigma(\mtx{R}_{\delta} ) = j } \\
&\geq \Prob{ \norm{ \mtx{R}_\delta \mtx{A} } \geq u
\ | \ \# \sigma( \mtx{R}_\delta ) = \delta n } \cdot
\sum\nolimits_{j = \delta n}^n
\Prob{ \# \sigma(\mtx{R}_\delta ) = j } \\
&\geq \frac{1}{2} \Prob{ \norm{ \mtx{P}_\delta \mtx{A} } \geq u }.\end{aligned}$$ The second inequality holds because the spectral norm of a submatrix is smaller than the spectral norm of the matrix. The third inequality relies on the fact [@JS68:Monotone-Convergence Thm. 3.2] that the medians of the binomial distribution $\textsc{binomial}( \delta, n )$ lie between $\delta n - 1$ and $\delta n$.
In case $\delta n$ is not integral, the monotonicity of the spectral norm yields that $$\Prob{ \norm{ \mtx{R}_{\delta} \mtx{A} } \geq u }
\geq \Prob{ \norm{ \mtx{R}_{\lfloor \delta n \rfloor / n} \mtx{A}}
\geq u }.$$ Since $\mtx{P}_{\lfloor \delta n \rfloor / n} = \mtx{P}_{\delta}$, this point completes the argument.
Small submatrices
-----------------
We focus on matrices with uniformly bounded entries. The first step in the argument is an elementary estimate on the norm of a random submatrix with expected order one. In this regime, the bound on the matrix entries determines the norm of the submatrix; the signs of the entries do not play a role. The proof shows that most of the variation in the norm actually derives from the fluctuation in the order of the submatrix.
\[lem:small-submatrix\] Let $\mtx{A}$ be an $n \times n$ matrix whose entries are bounded in magnitude by $n^{-1/2}$. Abbreviate $\varrho = 1/n$. When $q \geq 2\log n \geq \econst$, $$\left( \Expect \norm{ \mtx{R}_{\varrho} \mtx{A} \mtx{R}_{\varrho}' }^{2q} \right)^{1/2q}
\leq 2q n^{-1/2}.$$
By homogeneity, we may rescale $\mtx{A}$ so that its entries are bounded in magnitude by one. Define the event $\Sigma_{jk}$ where the random submatrix has order $j \times k$. $$\Sigma_{jk} = \{ \#\sigma(\mtx{R}_{\varrho}) = j \text{ and }
\#\sigma(\mtx{R}_{\varrho}') = k \}.$$ On this event, the norm of the submatrix can be bounded as $$\norm{ \mtx{R}_\varrho \mtx{A} \mtx{R}_{\varrho}' }
\leq \fnorm{ \mtx{R}_\varrho \mtx{A} \mtx{R}_{\varrho}' }
\leq \sqrt{jk}.$$ Using elementary inequalities, we may estimate the probability that this event occurs. $$\Probe{\Sigma_{jk}}
= {n \choose j}{n \choose k} \varrho^{j+k} (1-\varrho)^{2n - (j+k)}
\leq \left( \frac{\econst n}{j} \right)^{j}
\left( \frac{\econst n}{k} \right)^{k} n^{-(j+k)}
= (\econst/j)^j \cdot (\econst/k)^k.$$ With this information at hand, the rest of the proof follows from some easy calculations: $$\begin{aligned}
\Expect \norm{ \mtx{R}_{\varrho} \mtx{A} \mtx{R}_{\varrho}' }^{2q}
&= \sum\nolimits_{j,k=1}^n \Expect\left[
\norm{ \mtx{R}_{\varrho} \mtx{A} \mtx{R}_{\varrho}' }^{2q} \ | \
\Sigma_{jk} \right]
\cdot \Probe{ \Sigma_{jk} } \\
&\leq \sum\nolimits_{j,k=1}^n (jk)^q \cdot (\econst/j)^j \cdot (\econst/k)^k \\
&= \left[ \sum\nolimits_{k=1}^n k^q \cdot (\econst/k)^k \right]^2.\end{aligned}$$ A short exercise in differential calculus shows that the maximum term in the sum occurs when $k \log k = q$. Write $k_{\star}$ for the solution to this equation, and note that $k_{\star} \leq q$. Bounding all the terms by the maximum, we find $$\sum\nolimits_{k=1}^n k^q \cdot (\econst/k)^k
\leq n \cdot \exp\{ q \log k_{\star} - k_{\star} \log k_{\star} + k_{\star} \}
\leq n \cdot \exp\{ q \log k_{\star} \}
\leq n \cdot q^q.$$ Combining the last two inequalities, we reach $$\left( \Expect \norm{ \mtx{R}_{\varrho} \mtx{A} \mtx{R}_{\varrho}' }^{2q} \right)^{1/2q}
\leq \left( n^2 \cdot q^{2q} \right)^{1/2q}
= n^{1/q} \cdot q.$$ When $q \geq 2\log n$, the first term is less than two.
This argument delivers a moment estimate that is roughly a factor of $\log q$ smaller than the one stated. This fact can be used to sharpen the major results slightly at a cost we prefer to avoid.
Extrapolation
-------------
The key technique in the proof is an extrapolation of the moments of the norm of a large random submatrix from the moments of a smaller random submatrix. Without additional information, extrapolation must be fruitless because the signs of matrix entries play a critical role in determining the spectral norm. It turns out that we can fold in information about the signs by incorporating a bound on the spectral norm of the matrix. The proof, which we provide in Appendix \[app:extrap\], ultimately depends on the minimax property of the Chebyshev polynomials. The method is essentially the same as the one Bourgain and Tzafriri develop to prove Proposition 2.7 in [@BT91:Problem-Kadison-Singer]. See also [@Tro08:Random-Paving Sec. 7].
\[prop:extrap\] Suppose that $\mtx{A}$ is an $n \times n$ matrix with $\norm{\mtx{A}} \leq 1$. Let $q$ be an integer that satisfies $13 \log n \leq q \leq n/2$. Write $\varrho = 1/n$, and choose $\delta$ in the range $[1/n, 1]$. For each $\lambda \in (0, 1)$, it holds that $$\left( \Expect \norm{ \mtx{R}_\delta \mtx{A}\mtx{R}_\delta' }^{2q} \right)^{1/2q}
\leq 8 \delta^{\lambda} \max\left\{1, n^{\lambda}
\left( \Expect \norm{ \mtx{R}_\varrho \mtx{A}
\mtx{R}_\varrho' }^{2q} \right)^{1/2q} \right\}.$$
Although the statement is a little complicated, we require the full power of this estimate. As usual, the parameter $q$ is the moment that we seek. The proposition extrapolates from a matrix of expected order $1$ up to a matrix of expected order $\delta n$. The parameter $\lambda$ is a tuning knob that controls how much of the estimate is determined by the spectral norm of the full matrix and how much is determined by the norm bound for small submatrices. Indeed, the first member of the maximum reflects the spectral norm bound $\norm{\mtx{A}} \leq 1$.
A tail bound
------------
We are now prepared to develop a tail bound for the random norm $\norm{ \mtx{R}_\delta \mtx{A} \mtx{R}_\delta' }$.
\[lem:tail-bound\] Let $\mtx{A}$ be an $n \times n$ matrix for which $$\norm{ \mtx{A} } \leq 1
\qquad\text{and}\qquad
\abs{ a_{jk} } \leq n^{-1/2}
\quad\text{for $j,k = 1,2, \dots, n$}.$$ Choose $\delta$ from $[1/n, 1]$ and an integer $q$ that satisfies $13 \log n \leq q \leq n/2$. For each $\lambda \in (0, 1)$, it holds that $$\Prob{ \norm{ \mtx{R}_\delta \mtx{A} \mtx{R}_\delta' } \geq
8 \delta^\lambda \max\bigl\{1, 2qn^{\lambda -1/2} \bigr\}
\cdot u }
\leq u^{-2q}
\quad\text{for $u \geq 1$.}$$
Choose an integer $q$ in the range $[13 \log n, n/2]$. Markov’s inequality allows that $$\Prob{ \norm{ \mtx{R}_\delta \mtx{A} \mtx{R}_\delta' } \geq
\left(\Expect \norm{
\mtx{R}_\delta \mtx{A} \mtx{R}_\delta' }^{2q}\right)^{1/2q}
\cdot u } \leq u^{-2q}.$$ Therefore, we may establish the result by obtaining a moment estimate. This estimate is a direct consequence of Lemma \[lem:small-submatrix\] and Proposition \[prop:extrap\]: $$\left(\Expect \norm{\mtx{R}_\delta \mtx{A} \mtx{R}_\delta' }^{2q}\right)^{1/2q}
\leq 8\delta^{\lambda} \max\left\{1, n^{\lambda} \cdot 2qn^{-1/2}\right\}.$$ Combine the two bounds to complete the argument.
The two major results of this paper, Theorem \[thm:both-rand\] and Theorem \[thm:both-rand-norm\], both follow from a simple corollary of Lemma \[lem:tail-bound\].
\[cor:useful-tail\] Suppose that $T$ and $\Omega$ are random sets with cardinalities $\abs{T}$ and $\abs{\Omega}$. Assume $\delta \geq \max\{\abs{T}, \abs{\Omega}\} / n$. For each integer $q$ that satisfies $13 \log n \leq q \leq n/2$ and for $\lambda \in [0, 1]$, it holds that $$\Prob{ \norm{ {\mathbf{F}}_{\Omega T} } \geq
8 \delta^\lambda \max\bigl\{1,
2qn^{\lambda-1/2} \bigr\}
\cdot u }
\leq 4 u^{-2q}
\quad\text{for $u \geq 1$.}$$
Consider the matrix $\mtx{A} = {\mathbf{F}}$. Perform the reductions from Section \[sec:reductions\], Lemma \[lem:square-case\] and Lemma \[lem:rand-coords\]. Then apply the tail bound, Lemma \[lem:tail-bound\].
Proof of Theorem \[thm:both-rand\]
----------------------------------
The content of Theorem \[thm:both-rand\] is to provide a bound on $\delta$ which ensures that $\norm{{\mathbf{F}}_{\Omega T}}$ is somewhat less than one with extremely high probability. To that end, we want to make $\lambda$ close to zero and $q$ large. The following selections accomplish this goal: $$\lambda = \frac{\log 16}{\log (1/\delta)}
\qquad\text{and}\qquad
q = \lfloor 0.5 n^{1/2 - \lambda} \rfloor.$$ Note that we can make $\lambda$ as small as we like by taking $\delta$ sufficiently small. For any value of $\lambda < 0.5$, the number $q$ satisfies the requirements of Corollary \[cor:useful-tail\] as soon as $n$ is sufficiently large.
Now, the bound of Corollary \[cor:useful-tail\] results in $$\Prob{ \norm{ {\mathbf{F}}_{\Omega T} } \geq 0.5 u }
\leq 4 u^{-2q}.$$ For $u = \sqrt{2}$, we see that $$\Prob{ \normsq{ {\mathbf{F}}_{\Omega T} } \geq 0.5 }
\leq 4 \cdot 2^{-q}.$$ If follows that, for any assignable $\eps > 0$, we can make $$\Prob{ \normsq{ {\mathbf{F}}_{\Omega T} } \geq 0.5 }
\leq \exp\bigl\{ - n^{1/2 - \eps} \bigr\}$$ provided that $\delta \leq \econst^{-\cnst{C}/\eps} = \cnst{c}(\eps)$ and that $n \geq N(\eps)$.
Proof of Theorem \[thm:both-rand-norm\]
---------------------------------------
To establish Theorem \[thm:both-rand-norm\], we must make the parameter $\lambda$ as close to $0.5$ as possible. Choose $$\lambda = \frac{1}{2} - \frac{0.1}{\log(1/\delta)}
\qquad\text{and}\qquad
q = \lfloor \cnst{C} \log n \rfloor.$$ where $\cnst{C}$ is a large constant. These choices are acceptable once $\delta$ is sufficiently small and $n$ is sufficiently large.
Corollary \[cor:useful-tail\] delivers $$\Prob{ \norm{ {\mathbf{F}}_{\Omega T} } \geq 8.9 \delta^{1/2} u }
\leq 4u^{- \cnst{C} \log n}.$$ For $u = 90 / 89$, we reach $$\Prob{ \norm{ {\mathbf{F}}_{\Omega T} } \geq 9 \delta^{1/2} }
\leq n^{- \cnst{C}},$$ adjusting constants as necessary. Finally, we transfer the factor $\delta^{1/2}$ to the other side of the inequality and set $\delta = \abs{\Omega} / n$ to complete the proof.
Numerical Experiments {#sec:exper}
=====================
The theorems of this paper provide gross information about the norm of a random submatrix of the DFT. To complement these results, we performed some numerical experiments to give a more detailed empirical view.
The first set of experiments concerns random square submatrices of a DFT matrix of size $n$, where we varied the parameter $n$ over several orders of magnitude. Given a value of $\delta \in (0, 0.5)$, we formed one hundred random submatrices with dimensions $\delta n \times \delta n$ and computed the average spectral norm of these matrices. We did not plot data when $\delta \in (0.5, 1)$ because the norm of a random submatrix equals one.
Figure \[fig:square-unscaled\] shows the raw data for this first experiment. As $n$ grows, one can see that the norm tends toward an apparent limit: $2 \sqrt{\delta (1 - \delta)}$. In Figure \[fig:square-scaled\], we re-scale each matrix by $\delta^{-1/2}$ so its columns have unit norm and then compute the average spectral norm. More elaborate behavior is visible in this plot:
- For $\delta = 1/n$, the norm of a random submatrix is identically equal to one.
- For $\delta = 2/n$, the norm tends toward $1 + 2^{-1/2} = 1.7071\dots$, which can be verified by a relatively simple analytic computation.
- The maximum value of the norm appears to occur at $\delta = 2 / \sqrt{n}$.
- The apparent limit of the scaled norm is $2 \sqrt{1 - \delta}$, in agreement with the first figure.
These phenomena are intriguing, and it would be valuable to understand them in more detail. Unfortunately, the methods of this paper are not refined enough to provide an explanation.
![Sample average of the norm of a random $\delta n \times \delta n$ submatrix drawn from the $n \times n$ DFT.[]{data-label="fig:square-unscaled"}](art/square-unscaled.eps){height="0.4\textheight"}
![Sample average of the norm of a random $\delta n \times \delta n$ submatrix drawn from the $n \times n$ DFT and re-scaled by $\delta^{-1/2}$.[]{data-label="fig:square-scaled"}](art/square-scaled.eps){height="0.4\textheight"}
In the second set of experiments, we studied the norm of a random rectangular submatrix of the $128 \times 128$ DFT matrix. We varied the proportion $\delta_T$ of columns and the proportion $\delta_\Omega$ of rows in the range $(0, 1)$. For each pair $(\delta_T, \delta_\Omega)$, we drew 100 random submatrices and computed the average norm. Figure \[fig:rect-unscaled\] shows the raw data. The apparent trend is that $$\Expect \norm{ \mtx{P}_{\delta_\Omega} {\mathbf{F}}\mtx{P}_{\delta_T}' }
= 2 \sqrt{\delta (1 - \delta)}
\qquad\text{where}\qquad
\delta = \frac{\abs{T} + \abs{\Omega}}{2}.$$ Figure \[fig:rect-scaled\] shows the same data, rescaled by $\max\{\abs{T}, \abs{\Omega}\}^{-1/2}$. As in the square case, this plot reveals a variety of interesting phenomena that are worth attention.
![Sample average of the norm of a random $\delta_\Omega n \times \delta_T n$ submatrix drawn from the $128 \times 128$ DFT matrix.[]{data-label="fig:rect-unscaled"}](art/rect-unscaled.eps){height="0.4\textheight"}
![Sample average of the norm of a random $\delta_\Omega n \times \delta_T n$ submatrix drawn from the $128 \times 128$ DFT matrix and rescaled by $\max\{\abs{T},\abs{\Omega}\}^{-1/2}$.[]{data-label="fig:rect-scaled"}](art/rect-scaled.eps){height="0.4\textheight"}
Further Research Directions {#sec:future}
===========================
The present research suggests several directions for future exploration.
1. It may be possible to improve the constants in Proposition \[prop:extrap\] using a variation of the current approach. Instead of using the Chebyshev polynomial to estimate the coefficients of the polynomial that arises in the proof, one might use the nonnegative polynomial of least deviation from zero on the interval $[0, 1]$. The paper [@BK85:Polynomials-Fixed] is relevant in this connection: its authors identify the nonnegative polynomials with least deviation from zero with respect to $L_p$ norms for $p < \infty$. The $p = \infty$ case appears to be open, and uniqueness may be an issue.
2. Instead of reducing the problem to the square case, it would be valuable to understand the rectangular case directly. Again, it may be possible to adapt Proposition \[prop:extrap\] to handle this situation. This approach would probably require the bivariate polynomials of least deviation from zero identified by Sloss [@Slo65:Chebyshev-Approximation].
3. A harder problem is to determine the limiting behavior of the expected norm of a random submatrix as the dimension grows and the proportion of rows and columns remains fixed. We frame the following conjecture.
A random square submatrix of the $n \times n$ DFT satisfies $$\Expect \norm{ \mtx{P}_\delta {\mathbf{F}}\mtx{P}_\delta' }
\leq 2 \sqrt{ \delta (1 - \delta) }.$$ The inequality becomes an equality as $n \to \infty$.
One can develop a similar statement about random rectangular submatrices. At present, however, these conjectures are out of reach.
4. Finally, one might study the behavior of the lower singular value of a (suitably normalized) random submatrix drawn from the DFT. There are some results available when one set, say $T$, is fixed [@CRT06:Robust-Uncertainty]. It is possible that the behavior will be better when both sets are random. The present methods do not seem to provide much information about this problem.
Acknowledgments {#acknowledgments .unnumbered}
===============
One of the anonymous referees provided a wealth of useful advice that substantially improved the quality of this work. In particular, the referee described a version of Lemma \[lem:small-submatrix\] and demonstrated that it offers a simpler route to the main results than the argument in earlier drafts of this paper.
Chebyshev Extrapolation {#app:extrap}
=======================
One of the major tools in the proof of Theorem \[thm:both-rand\] is Proposition \[prop:extrap\]. This result extrapolates the moments of the norm of a large random submatrix drawn from a fixed matrix, given information about a small random submatrix. An important idea behind the result is to fold information about the spectral norm of the matrix into the estimate. The extrapolation technique is due to Bourgain and Tzafriri [@BT91:Problem-Kadison-Singer]. We require a variant of their result, so we repeat the argument in its entirety. The complete statement of the result follows.
\[prop:extrap-app\] Suppose that $\mtx{A}$ is an $n \times n$ matrix with $\norm{\mtx{A}} \leq 1$. Let $q$ be an integer that satisfies $13 \log n \leq q \leq n/2$. Choose parameters $\varrho \in (0, 1)$ and $\delta \in [\varrho, 1]$. For each $\lambda \in [0, 1]$, it holds that $$\left( \Expect \norm{ \mtx{R}_\delta \mtx{A}\mtx{R}_\delta' }^{2q} \right)^{1/2q}
\leq 8 \delta^{\lambda} \max\left\{1, \varrho^{-\lambda}
\left( \Expect \norm{ \mtx{R}_\varrho \mtx{A}
\mtx{R}_\varrho' }^{2q} \right)^{1/2q} \right\}.$$ The same result holds if we replace $\mtx{R}_{\delta}'$ by $\mtx{R}_{\delta}$ and replace $\mtx{R}_{\varrho}'$ by $\mtx{R}_{\varrho}$.
V. A. Markov observed that the coefficients of an arbitrary polynomial can be bounded in terms of the coefficients of a Chebyshev polynomial because Chebyshev polynomials are the unique polynomials of least deviation from zero on the unit interval. See [@Tim63:Theory-Approximation Sec. 2.9] for more details.
\[prop:markov\] Let $p(t) = \sum_{k = 0}^r c_k t^k$. The coefficients of the polynomial $p$ satisfy the inequality $$\abs{c_k} \leq \frac{r^k}{k!} \max_{\abs{t} \leq 1} \abs{ p(t) }
\leq \econst^r \max_{\abs{t} \leq 1} \abs{ p(t) }.$$ for each $k = 0, 1, \dots, r$.
With Markov’s result at hand, we can prove Proposition \[prop:extrap-app\].
We establish the result when the two diagonal projectors are independent; the other case is almost identical because this independence is never exploited. Define the function $$F(s) = \Expect \norm{ \mtx{R}_s \mtx{A} \mtx{R}_s' }^{2q}
\qquad\text{for $s \in [0, 1]$.}$$ Note that $F(s) \leq 1$ because $\norm{ \mtx{R}_s \mtx{A} \mtx{R}_s' } \leq \norm{\mtx{A}} \leq 1$. Furthermore, $F$ does not decrease.
The function $F$ is comparable with a polynomial. Use the facts that $2q$ is even and that $\mtx{A}$ has dimension $n$ to check the inequalities $$\label{eqn:F-bds}
F(s) \leq
\Expect \trace [(\mtx{R}_s \mtx{A} \mtx{R}_s')^\adj
(\mtx{R}_s \mtx{A} \mtx{R}_s') ]^q
\leq n F( s ).$$ Define a second function $$p(s) = \Expect \trace [(\mtx{R}_s \mtx{A} \mtx{R}_s')^\adj
(\mtx{R}_s \mtx{A} \mtx{R}_s') ]^q
= \Expect \trace (\mtx{A}^\adj \mtx{R}_s \mtx{A} \mtx{R}_s')^q,$$ where we used the cyclicity of the trace and the fact that $\mtx{R}_s$ and $\mtx{R}_s'$ are diagonal matrices with 0–1 entries. Expand the product and compute the expectation using the additional fact that the entries of the diagonal matrices are independent random variables of mean $s$. We discover that $p$ is a polynomial of maximum degree $2q$ in the variable $s$: $$p(s) = \sum\nolimits_{k = 1}^{2q} c_k s^k$$ The polynomial has no constant term because $\mtx{R}_0 = \mtx{0}$.
We can use Markov’s technique to bound the coefficients of the polynomial. First, make the change of variables $s = \varrho t^2$ to see that $$\abs{ \sum\nolimits_{k=1}^{2q} c_k \varrho^k t^{2k} }
= \abs{ p( \varrho t^2) }
\leq n F( \varrho t^2 )
\leq n F( \varrho )
\qquad\text{for $\abs{t} \leq 1$.}$$ The first inequality follows from and the second follows from the monotonicity of $F$. The polynomial $p( \varrho t^2 )$ has degree $4q$ in the variable $t$, so Proposition \[prop:markov\] yields $$\label{eqn:coef-bd1}
\abs{c_k} \varrho^k \leq n \econst^{4q} F( \varrho )
\qquad\text{for $k = 1, 2, \dots, 2q$.}$$ Evaluate this expression at $\varrho = 1$ and recall that $F \leq 1$ to obtain a second bound, $$\label{eqn:coef-bd2}
\abs{c_k} \leq n \econst^{4q}
\qquad\text{for $k = 1, 2, \dots, 2q$.}$$
To complete the proof, we evaluate the polynomial at a point $\delta$ in the range $[\varrho, 1]$. Fix a value of $\lambda$ in $[0,1]$, and set $K = \lfloor 2 \lambda q \rfloor$. In view of and , we obtain $$\begin{aligned}
F(\delta) &\leq \sum\nolimits_{k=1}^{K} \abs{c_k} \delta^k
+ \sum\nolimits_{k=K + 1}^{2q}
\abs{c_k} \delta^k \\
&\leq \sum\nolimits_{k=1}^{K} n \econst^{4q}
F( \varrho ) (\delta/\varrho)^k
+ \sum\nolimits_{k = K + 1}^{2q} n \econst^{4q} \delta^k \\
&\leq n\econst^{4q} \left[ K (\delta/\varrho)^{K} F(\varrho) + (2q - K) \delta^{K + 1} \right] \\
&\leq n\econst^{4q} \delta^{2\lambda q} \left[ K \varrho^{-2\lambda q} F(\varrho) + (2q - K) \right] \\
&\leq n \econst^{4q} \delta^{2\lambda q} \cdot 2q
\max\{ 1, \varrho^{- 2\lambda q} F(\varrho) \}\end{aligned}$$ The third and fourth inequalities use the conditions $\delta/\varrho \geq 1$ and $\delta \leq 1$, and the last bound is an application of Jensen’s inequality. Taking the $(2q)$th root, we reach $$F( \delta )^{1/2q} \leq
(2qn)^{1/2q} \econst^2 \delta^{\lambda}
\max\{1, \varrho^{-\lambda} F(\varrho)^{1/2q} \}.$$ The leading constant is less than 8, provided that $13 \log n \leq q \leq n/2$.
[^1]: The author is with Applied & Computational Mathematics, MC 217-50, California Institute of Technology, 1200 E. California Blvd., Pasadena, CA 91125-5000. E-mail: [[email protected]]([email protected]). Supported by NSF 0503299.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In chemical analysis made by laboratories one has the problem of determining the concentration of a chemical element in a sample. In order to tackle this problem the guide EURACHEM/CITAC recommends the application of the linear calibration model, so implicitly assume that there is no measurement error in the independent variable $X$. In this work, it is proposed a new calibration model assuming that the independent variable is controlled. This assumption is appropriate in chemical analysis where the process tempting to attain the fixed known value $X$ generates an error and the resulting value is $x$, which is not an observable. However, observations on its surrogate $X$ are available. A simulation study is carried out in order to verify some properties of the estimators derived for the new model and it is also considered the usual calibration model to compare it with the new approach. Three applications are considered to verify the performance of the new approach.
.3in
[*Keywords:*]{} linear calibration model, controlled variable, measurement error model, uncertainty, chemical analysis.
---
‘=11 \#1 \#1[[bsphack@filesw [ gtempa[auxout[ ]{}]{}]{}gtempa @nobreak esphack]{} eqnlabel[\#1]{}]{} eqnlabel vacuum \#1
\#1[@underline\#1 $\@@underline{\hbox{#1}}$]{}
‘@=12
by -by -
8.9in
0.5in
6.5in
-.6in
\#1
\#1
===
ß[s/]{} \#1[(\[\#1\])]{} \#1[\[\#1\]]{} \#1 ${\left(}
\def$[)]{}
v
\#1[\^[(R,\#1)]{}]{} ł Ł ø Ø ¶ § 0[\_[+]{}\^[(0)]{}]{} 0[\_[-]{}\^[(0)]{}]{}
\#1\#2\#3 [f\^[\#1\#2]{}\_[\#3]{}]{} 1[[w\_[1+]{}]{}]{} 1[[W\_[1+]{}]{}]{} \#1\#2[r(,)]{} \#1[[O]{}(R\_[\#1]{})]{} \#1[Ad\_[R\_[\#1]{}]{}]{} \#1[Ad\_[R\_[\#1]{}\^]{}]{} \#1[ad\_[R\_[\#1]{}\^]{}]{} \#1["7017[\#1]{}B 23\#1]{} \#1\#2\#3[[*Scandinavian Journal of Statistics*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Applied Statistics*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Metrologia*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Journal of Chemometrics*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Accreditation and Quality Assurance*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Química Nova*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Biometrika*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Technometrics*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Fresenius J Anal Chem*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*International Statistical Review*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Journal of the American Statistical Association*]{} [**\#1**]{} (\#2) \#3]{} ł Ł ø Ø u ${\Big(}
\def$[)]{} $${\Big[}
\def$$[\]]{}
\#1["7017[\#1]{} B 23\#1]{} \#1\#2[[\#1\#2]{}]{}
[**Heteroscedastic controlled calibration model applied to analytical chemistry**]{}
Betsabé G. Blas Achic and Mônica C. Sandoval
.1in Departamento de Estatística, Universidade de São Paulo, São Paulo, Brasil\
.3in
Introduction
============
The usual calibration model [@shukla] is commonly used to estimate the concentration $X_{0}$ of a chemical species in a test sample. It typically assumes that the independent variable is fixed and it is not subject to error. However, in applications in analytical chemistry this variable is subject to error which arises from the preparation process of a standard solution. In many studies, such as [@eura], [@Queenie] and [@lutz], it is attempted to consider the uncertainties due to the preparation process of the standard solutions by application of the error propagation law to the standard error of the estimator of $X_{0}$.
We have that the concentration of the standard solution is pre-fixed by the chemical analyst and a process is carried out attempting to attain it, this process generates errors. Hence, in this case it arises the so called controlled variable [@berk50], where the controlled variable $X$ is defined by the pre-fixed concentration value of the standard solution which is expressed by the equation $X=x+\delta$, where $x$ is the unobserved variable and $\delta$ is the measurement error variable.
In [@bmo] it was proposed the so called homoscedastic controlled calibration model. This model is formulated in the framework of the usual calibration model assuming that the independent variable is a controlled variable and the associated measurement errors have *equal* variances.
In [@ballico] and [@bet], some methods to compute the uncertainties in certain values obtained through measurements are studied. In [@bet], the uncertainties of standard solutions are computed and it is observed that these uncertainties depend on the concentration values, so we can observe that the usual calibration model and the homoscedastic controlled calibration model seem not to be the more suitable ones. This problem motivates us to study a calibration model that considers the errors variability of the preparation of standard solutions. In this work we propose a calibration model that incorporates the errors variability arisen from the preparation process of the standard solution and we call it as *the heteroscedastic controlled calibration model*. This work is a continuation to our previous paper [@bmo] in which it was undertaken the study of the so-called homoscedastic controlled calibration model which assumed *equal* variance errors.
The paper is organized as follows. In Section \[model\], we formulate the heteroscedastic controlled calibration model. In Section \[simulation\], a simulation study to test the new approach is presented. In Section \[applications\], three applications are considered which show that the proposed model seems to be more adequate. Section \[remarks\] presents our concluding remarks. Finally, we present in Appendix \[appa\] the usual calibration model, and in Appendix \[appb\] some tables showing the results of the simulation study.\
The proposed model {#model}
==================
Among the relevant problems in chemical analysis is the one related to the estimation of the concentration $X_0$ of a chemical compound in a given sample. In order to tackle this problem it is used a statistical calibration model, which is defined by a two-step process. This problem has been considered in [@tallis] and [@lwin].
The first stage of the calibration model is given by data points $(X,Y)$ which is determined in an experiment where the independent variable $X$ is the one that the experimenter selects. For instance, the concentrations of the standard solutions that a chemist prepares are independent variables since any concentration may be chosen. The dependent variable $Y$ is a measurable property of the independent variable. For example, the dependent variable may be the amount of intensity supplied by the plasma spectrometry method, since the intensity depends on the concentration.
In the second stage of the calibration model it is prepared a suitable sample related to the unknown concentration $X_0$ in order to obtain the measurements $Y_0$.
We have that the standard concentration $X$ is fixed by the analyst and the process of preparation attempting to get it produces an error $\d$, and the unobserved quantity attained is $x$. Considering the usual calibration model defined by the equations \[m1\] and \[m2\] in the Appendix \[appa\] and the equation $X=x+\d$, we define the heteroscedastic controlled calibration model as $$\begin{aligned}
\label{m11}
Y_{i}&=&\alpha+\beta x_{i}+\epsilon_{i},\,\,\,\,\,i=1,2\cdots,n,\\
\label{22}
X_i&=&x_i+\d_i,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,i=1,2\cdots,n,\\
\label{m33}
Y_{0i}&=&\alpha+\beta X_{0}+\epsilon_{i},\,\,\,\,i=n+1,n+2,\cdots,n+k.\end{aligned}$$ It is considered the usual calibration model assumptions (see Appendix A) in addition to the following conditions
- $\delta_{1},\delta_{2},\cdots,\delta_{n}$ are independent and normally distributed with mean 0.
- the variances $\sigma_{\delta_{i}}^{2},\,(i=1,\cdots,n)$ are supposed to be known.
- $\delta_{i},\,i=1,\cdots,n$ and $\epsilon_{i},\,i=1,\cdots,n+k$ are independent.
Observe that in the model described above we only consider the case when the variances $\sigma_{\delta_{i}}^{2},\,\,\,i=1,\cdots,n$ are known. It is a generalization of the homoscedastic controlled calibration model discussed in [@bmo], when it is considered $\sigma^{2}_{\delta_{i}}=\sigma_{\delta}^{2}$ for all $i$ and the known $\sigma_{\delta}^{2}$ case. This new model is also a generalization of the usual calibration model in which one takes $\delta_{i}=0,\,i=1,\cdots,n$.
For the heteroscedastic controlled calibration model the logarithm of the likelihood function is given by [$$\begin{aligned}
\nonumber
l(\,\alpha,\beta,X_{0},\sigma_{\epsilon}^{2})&&\propto-\frac{1}{2}\sum_{i=1}^{n}log(\gamma_i)-\frac{k}{2}log(\sigma_{\epsilon}^{2})\\
\label{L1}
&&-\frac{1}{2}\left[\sum_{i=1}^{n}\frac{(Y_{i}-\alpha-\beta
X_{i})^{2}}{\gamma_i}+\sum_{i=n+1}^{n+k}\frac{(Y_{0i}-\alpha-\beta
X_{0})^{2}}{\sigma_{\epsilon}^{2}}\right],\end{aligned}$$]{} where $\gamma_i=\sigma_{\epsilon}^{2}+\beta^{2}\sigma_{\delta_{i}}^{2}$, $i=1,\cdots,n$. Solving $\partial l /\partial\alpha=0$ and $\partial l/\partial X_{0}=0$ one can get the maximum likelihood estimator of $\alpha$ and $X_{0}$ given, respectively, by $$\label{axo}
\hat{\alpha}=\bar{Y}-\hat{\beta}\bar{X}\,\,\,\,\,\mbox{and}\,\,\,\,\,\hat{X}_{0}=\frac{\bar{Y}_{0}-\hat{\alpha}}{\hat{\beta}}.$$ From (\[L1\]) and (\[axo\]), it follows that the logarithm of the likelihood function for $(\alpha,\beta,X_0,\s_{\eps}^2)$ can be writen as [$$\begin{aligned}
\label{L2}
l(\alpha,\beta,X_{0},\sigma_{\epsilon}^{2})&&\propto-\frac{1}{2}\sum_{i=1}^{n}log(\gamma_i)-\frac{k}{2}log(\sigma_{\epsilon}^{2})\\
\nonumber
&&-\frac{1}{2}\left[\sum_{i=1}^{n}\frac{[(Y_{i}-\bar{Y})-\beta
(X_{i}-\bar{X})]^{2}}{\gamma_i}+\frac{1}{\sigma_{\epsilon}^{2}}\sum_{i=n+1}^{n+k}(Y_{0i}-\bar{Y_{0}})^{2}\right].\end{aligned}$$]{} Making $\partial l /\partial\beta=0$, $\partial l/\partial\sigma_{\epsilon}^{2}=0$ in the logarithm of the likelihood function (\[L2\]), we have the following equations $$\begin{aligned}
\label{ebh}
\sum_{i=1}^{n}\frac{\beta\sigma_{\delta_{i}}^{2}\left[\gamma_i-(Y_{i}-\alpha-\beta
X_{i})^{2}\right]}{\gamma_i^{2}}=
\sum_{i=1}^{n}\frac{X_{i}(Y_{i}-\alpha-\beta
X_{i})}{\gamma_i}\\
\label{eeh}
\sum_{i=1}^{n}\frac{\gamma_i-(Y_{i}-\alpha-\beta
X_{i})^{2}}{\gamma_i^{2}}=\sum_{i=n+1}^{n+k}\frac{(Y_{0i}-\bar{Y_{0}})^{2}}{\sigma_{\epsilon}^{4}}-\frac{k}{\sigma_{\epsilon}^{2}}.\end{aligned}$$ The estimates of $\beta$ and $\sigma_{\epsilon}^{2}$ can be obtained through some iterative method that solves the equations (\[ebh\]) and (\[eeh\]).
The Fisher expected information $I(\theta)=I(\alpha,\beta,X_{0},\sigma_{\epsilon}^{2})$ is given by\
$I(\theta)= \left(\begin{array}{ccccc}
\sum_{i=1}^{n} \frac{1}{\gamma_{i}}+\frac{k}{\sigma_{\epsilon}^{2}} &\sum_{i=1}^{n} \frac{X_{i}}{\gamma_{i}}+ \frac{kX_{0}}{\sigma_{\epsilon}^{2}} &\frac{k\beta}{\sigma_{\epsilon}^{2}} &0 \\
\sum_{i=1}^{n} \frac{X_{i}}{\gamma_{i}}+ \frac{kX_{0}}{\sigma_{\epsilon}^{2}} & \sum_{i=1}^{n}\frac{X_{i}^{2}}{\gamma_{i}}+2\beta^{2}\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{4}}{\gamma_{i}^{2}}+\frac{kX_{0}^{2}}{\sigma_{\epsilon}^{2}} & \frac{k\beta X_{0}}{\sigma_{\epsilon}^{2}} & \beta\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{2}}{\gamma_{i}^{2}}\\
\frac{k\beta}{\sigma_{\epsilon}^{2}}& \frac{k\beta X_{0}}{\sigma_{\epsilon}^{2}} &\frac{k\beta^{2}}{\sigma_{\epsilon}^{2}} &0 \\
0 & \beta\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{2}}{\gamma_{i}^{2}} & 0& \sum_{i=1}^{n}\frac{1}{2\gamma_{i}^{2}}+\frac{k}{2\sigma_{\epsilon}^{4}}\\
\end{array}\right)$\
When $k=qn$, $q\in Q^{+}$ and $n\rightarrow \infty$, the estimator $\hat{\theta}$ is approximately normally distributed with mean $\theta$ and variance $I(\theta)^{-1}$, thus the approximate variance to order $n^{-1}$ for $\hat{X}_{0}$ is given by $$\label{varhete}
V(\hat{X}_{0})=\frac{\sigma_{\epsilon}^{2}}{\beta^{2}}\left[\frac{1}{n}+\frac{1}{k}-\frac{E_{1}}{n\sigma_{\epsilon}^{2}E_{2}}\right],$$ where $$\begin{aligned}
\nonumber
E_{1}&=&-n\sum_{i=1}^{n}\frac{X_{0}^{2}\sigma_{\epsilon}^{4}}{\gamma_{i}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}^{2}}-nk\sum_{i=1}^{n}\frac{X_{0}^{2}}{\gamma_{i}}-n\sum_{i=1}^{n}\frac{X_{i}^{2}\sigma_{\epsilon}^{4}}{\gamma_{i}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}^{2}}-nk\sum_{i=1}^{n}\frac{X_{i}^{2}}{\gamma_{i}}\\
\nonumber
&&-2n\beta^{2}\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{4}\sigma_{\epsilon}^{4}}{\gamma_{i}^{2}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}^{2}}-2nk\beta^{2}\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{4}}{\gamma_{i}^{2}}+2n\beta^{2}\sigma_{\epsilon}^{4}\left[\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{2}}{\gamma_{i}^{2}}\right]^{2}\\
\nonumber
&&+2nX_{0}\sigma_{\epsilon}^{4}\sum_{i=1}^{n}\frac{X_{i}}{\gamma_{i}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}^{2}}+2nkX_{0}\sum_{i=1}^{n}\frac{X_{i}}{\gamma_{i}}+\sigma_{\epsilon}^{6}\sum_{i=1}^{n}\frac{X_{i}^{2}}{\gamma_{i}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}^{2}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}}\\
\nonumber
&&+k\sigma_{\epsilon}^{2}\sum_{i=1}^{n}\frac{X_{i}^{2}}{\gamma_{i}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}}+2\beta^{2}\sigma_{\epsilon}^{6}\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{4}}{\gamma_{i}^{2}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}^{2}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}}+2k\beta^{2}\sigma_{\epsilon}^{2}\sum_{i=1}^{n}\frac{1}{\gamma_{i}}\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{4}}{\gamma_{i}^{2}}\\
\nonumber
&&-2\beta^{2}\sigma_{\epsilon}^{6}\left[\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{2}}{\gamma_{i}^{2}}\right]^{2}\sum_{i=1}^{n}\frac{1}{\gamma_{i}}-\sigma_{\epsilon}^{6}\left[\sum_{i=1}^{n}\frac{X_{i}}{\gamma_{i}}\right]^{2}\sum_{i=1}^{n}\frac{1}{\gamma_{i}^{2}}-k\sigma_{\epsilon}^{2}\left[\sum_{i=1}^{n}\frac{X_{i}}{\gamma_{i}}\right]^{2}\end{aligned}$$ and $$\begin{aligned}
\nonumber
E_{2}&=&\sigma_{\epsilon}^{4}\sum_{i=1}^{n}\frac{X_{i}^{2}}{\gamma_{i}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}^{2}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}}+k\sum_{i=1}^{n}\frac{X_{i}^{2}}{\gamma_{i}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}}+2\sigma_{\epsilon}^{4}\beta^{2}\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{4}}{\gamma_{i}^{2}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}^{2}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}}\\
\nonumber
&&+2k\beta^{2}\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{4}}{\gamma_{i}^{2}}\sum_{i=1}^{n}\frac{1}{\gamma_{i}}-2\beta^{2}\sigma_{\epsilon}^{4}\left[\sum_{i=1}^{n}\frac{\sigma_{\delta_{i}}^{2}}{\gamma_{i}^{2}}\right]^{2}\sum_{i=1}^{n}\frac{1}{\gamma_{i}}\\
\nonumber
&&-\sigma_{\epsilon}^{4}\left[\sum_{i=1}^{n}\frac{X_{i}}{\gamma_{i}}\right]^{2}\sum_{i=1}^{n}\frac{1}{\gamma_{i}^{2}}-k\left[\sum_{i=1}^{n}\frac{X_{i}}{\gamma_{i}}\right]^{2}.\end{aligned}$$
Note that when $\sigma_{\delta_{i}}^{2}=0,\,i=1,\cdots,n,$ the expression (\[varhete\]) is reduced to the variance of the usual model given in (\[veurachem\]) and when $\sigma_{\delta_{i}}^{2}=\sigma_{\delta}^{2}$ (for all $i$) the expression (\[varhete\]) is also reduced to the variance of the homoscedastic model when $\sigma_{\delta}^{2}$ is known (see eq. (2.12) of ref. [@bmo]).
In order to construct a confidence interval for $X_0$ we consider the interval (\[uinterv\]), where $\hat{V}(\hat{X}_{0C})$ is the estimated variance that follows from (\[varhete\]).
Simulation study {#simulation}
================
We present a simulation study to compare the performance of the estimators obtained from the heteroscedastic controlled calibration model (Proposed-M) with the results obtained by considering the usual model (Usual-M).
It was considered 3000 samples generated from the Proposed-M. In all the samples, the parameters $\alpha$ and $\beta$ take the values 0.1 and 2, respectively. The range of values for the controlled variable was \[0,2\]. The fixed values for the controlled variable were $x_{1}=0,\,x_{i}=x_{i-1}+2/(n-1),\,i=2,\cdots,n,$ and the parameter values $X_{0}$ were 0.01 (extreme inferior value), 0.8 (near to the central value) and 1.9 (extreme superior value). For the first and second stages we consider the sample of sizes $n=5,\,20,\,100,\,5000$ and $k=2,\,20,\,100,\,500$, respectively. We consider $\sigma_{\epsilon}^{2}=0.04$ and the maximum parameter values of $\sigma_{\delta}^{2}$ as $max\{\sigma_{\delta_i}^{2}\}_{i=1}^n$= 0.1. We consider $\sigma_{\delta_i}^{2}=i\times 0.1/n\,\,\,\mbox{for } i=1,\cdots,n$.
The empirical mean bias is given by $\sum_{j=1}^{3000}(\hat{X}_{0}-X_{0})/3000$ and the empirical mean squared error (MSE) is given by $\sum_{j=1}^{3000}(\hat{X}_{0}-X_{0})^{2}/3000$. The mean estimated variance of $\hat{X}_{0}$ is given by $\sum_{j=1}^{3000}\hat{V}(\hat{X}_{0})/3000$. The theoretical variances of $\hat{X}_{0}$ is referred to the expressions (\[veurachem\]) and (\[varhete\]) evaluated on the relevant parameter values. In Appendix B it is presented the simulation results.
In Table \[pd0p\], we observe that, in general, the bias of $\hat{X}_{0}$ from the usual model is smaller than the value supplied by the proposed model, but related to the MSE we have that the outcome from usual model is greater compared with MSE of the proposed model. Also, we observe that the mean estimated variance from the proposed model is closer to the theoretical variance as compared to the outcome from the usual model.
Analyzing Table \[pd1p\], we observe that the amplitude of the proposed model, in most cases, is smaller when compared with the estimate of the usual model. For all $n$ and $X_0$ the amplitude from the usual model greatly decreases as the size of k increases, this behavior is being reflected on the covering percentage decreasing to less than 95%. Adopting the correct model we have that when $k$ increases the confidence interval amplitude decreases and the covering percentage increases approaching 95%.
\[appb\]
------- ------ ----- --------- -------- --------- ------------ ------------------------ ------------------------ ------------------
$X_0$ n k Usual-M Proposed-M Theorical variance
Bias MSE Bias MSE $\hat{V}(\hat{X}_{0})$ $\hat{V}(\hat{X}_{0})$ $V(\hat{X}_{0})$
0.01 5 2 -0.0156 0.0350 -0.0318 0.0334 0.0398 0.0143 0.0257
20 -0.0236 0.0319 -0.0445 0.0278 0.0131 0.0156 0.0211
100 -0.0183 0.0306 -0.0429 0.0276 0.0084 0.0155 0.0207
20 2 -0.0076 0.0119 -0.0049 0.0100 0.0365 0.0053 0.0097
20 -0.0055 0.0074 -0.0073 0.0055 0.0081 0.0036 0.0051
100 -0.0059 0.0068 -0.0101 0.0050 0.0036 0.0033 0.0047
100 2 0.0003 0.0063 0.0020 0.0059 0.0315 0.0047 0.0059
20 -0.0023 0.0019 -0.0020 0.0014 0.0046 0.0011 0.0014
100 -0.0014 0.0015 -0.0013 0.0010 0.0017 0.0007 0.0010
5000 2 0.0008 0.0055 0.0008 0.0055 0.0300 0.0050 0.0050
20 0.0000 0.0005 0.0000 0.0005 0.0030 0.0005 0.0005
100 -0.0003 0.0001 -0.0004 0.0001 0.0006 0.0001 0.0001
500 0.0000 0.0000 0.0000 0.0000 0.0002 0.0000 0.0000
0.8 5 2 0.0061 0.0193 0.0025 0.0202 0.0254 0.0089 0.0167
20 0.0033 0.0135 -0.0019 0.0139 0.0047 0.0081 0.0122
100 0.0014 0.0132 -0.0037 0.0137 0.0029 0.0078 0.0118
20 2 0.0015 0.0077 0.0015 0.0077 0.0291 0.0042 0.0074
20 0.0016 0.0032 0.0009 0.0032 0.0036 0.0020 0.0029
100 -0.0005 0.0026 -0.0018 0.0026 0.0012 0.0016 0.0025
100 2 0.0010 0.0055 0.0014 0.0054 0.0299 0.0044 0.0055
20 0.0006 0.0010 0.0007 0.0010 0.0031 0.0008 0.0010
100 -0.0001 0.0006 -0.0001 0.0006 0.0007 0.0004 0.0006
5000 2 0.0014 0.0051 0.0014 0.0051 0.0300 0.0050 0.0050
20 0.0006 0.0005 0.0006 0.0005 0.0030 0.0005 0.0005
100 0.0001 0.0001 0.0001 0.0001 0.0006 0.0001 0.0001
500 0.0000 0.0000 0.0000 0.0000 0.0001 0.0000 0.0000
1.9 5 2 0.0500 0.0802 0.0582 0.0704 0.0432 0.0275 0.0482
20 0.0314 0.0620 0.0503 0.0562 0.0124 0.0278 0.0435
100 0.0434 0.0645 0.0594 0.0587 0.0085 0.0283 0.0430
20 2 0.0070 0.0213 0.0054 0.0185 0.0351 0.0086 0.0166
20 0.0117 0.0160 0.0127 0.0132 0.0075 0.0067 0.0118
100 0.0099 0.0161 0.0104 0.0130 0.0033 0.0064 0.0114
100 2 0.0016 0.0080 0.0007 0.0076 0.0312 0.0055 0.0074
20 0.0019 0.0035 0.0014 0.0029 0.0043 0.0017 0.0028
100 0.0001 0.0031 0.0009 0.0025 0.0015 0.0013 0.0024
5000 2 -0.0008 0.0051 -0.0009 0.0051 0.0300 0.0050 0.0050
20 -0.0003 0.0006 -0.0004 0.0006 0.0030 0.0005 0.0005
100 -0.0003 0.0002 -0.0002 0.0001 0.0006 0.0001 0.0001
500 0.0000 0.0001 0.0000 0.0001 0.0001 0.0000 0.0001
------- ------ ----- --------- -------- --------- ------------ ------------------------ ------------------------ ------------------
: Empirical bias and mean squared error, the mean estimated variance and theoretical variance of $\hat{X}_{0}$.[]{data-label="pd0p"}
--------- ------ ----- ----- ------ ---- ------
$X_{0}$ $n$ $k$
% $A$ % $A$
0.01 5 2 89 0.35 78 0.22
20 78 0.21 89 0.24
100 70 0.17 88 0.24
20 2 100 0.37 74 0.13
20 95 0.17 90 0.12
100 85 0.12 90 0.11
100 2 100 0.35 84 0.13
20 100 0.13 91 0.06
100 96 0.08 90 0.05
5000 2 100 0.34 94 0.14
20 100 0.11 95 0.04
100 100 0.05 94 0.02
500 100 0.02 94 0.01
0.8 5 2 90 0.28 78 0.18
20 73 0.13 87 0.17
100 63 0.10 87 0.17
20 2 100 0.33 73 0.11
20 95 0.12 87 0.09
100 81 0.07 88 0.08
100 2 100 0.34 86 0.12
20 100 0.11 91 0.05
100 97 0.05 89 0.04
5000 2 100 0.34 95 0.14
20 100 0.11 95 0.04
100 100 0.05 95 0.02
500 100 0.02 93 0.01
1.9 5 2 78 0.35 81 0.31
20 59 0.20 86 0.32
100 51 0.17 87 0.32
20 2 98 0.36 78 0.17
20 81 0.17 84 0.16
100 62 0.11 84 0.16
100 2 100 0.34 87 0.14
20 97 0.13 87 0.08
100 83 0.08 84 0.07
5000 2 100 0.34 95 0.14
20 100 0.11 94 0.04
100 100 0.05 93 0.02
500 99 0.02 89 0.01
--------- ------ ----- ----- ------ ---- ------
: Covering percentage (%) and amplitude (A) of the intervals with a 95% confidence level for the parameter $X_{0}$.[]{data-label="pd1p"}
Application {#applications}
===========
In this section we illustrate the usefulness of the proposed model by applying it to the data supplied by the chemical laboratory of the “Instituto de Pesquisas Tecnológicas do Estado de São Paulo (IPT)" - Brasil. The outcome from the proposed approach are also compared with the results from the usual model. Our main interest is to estimate the unknown concentration value $X_{0}$ of a sample of the chemical elements such as chromium, cadmium and lead.
Table \[cr0\] below presents the fixed values of concentration for the standard solutions with their related uncertainty ($u(X_{i})$) and the corresponding intensities for the chromium, cadmium and lead elements. The uncertainties considered are computed using the method recommended by the ISOGUM guide (see [@isogum]) and the intensities are supplied by the plasma spectrometry method. This data is referred to the first stage of the heteroscedastic controlled calibration model.
Moreover, Table \[cr01\] below presents the intensities of the sample solutions of chromium, cadmium and lead elements. These data are referred to as the second stage of the calibration model.
Observing Tables \[cr0\] and \[cr01\] we verify that the uncertainty values increase with the concentration values.
We consider $\sigma_{\delta_{i}}^{2}=u(X_{i})^{2}$. The expanded uncertainty $U(X_{0})$ is obtained multiplying the squared root of the estimate of variance of $\hat{X}_{0}$ by the value 1.96 (see [@eura] and [@bet]).
We use the *optim* command from R-project program to estimate the parameters $\beta$ and $\sigma_{\epsilon}^{2}$ on the likelihood function of the proposed model (\[L2\]). We use as initial point the estimates from $\hat{\beta}=\sum_{i=1}^{n}(X_i-\bar{x})(Y_i-\bar{Y})/\sum_{i=1}^{n}(X_i-\bar{X})^2$ and $\hat{\sigma}_{\epsilon}^{2}=\sum_{i=1}^{n}(Y_{0i}-\bar{Y_0})^2/n$, which are the estimators from the homoscedastic controlled calibration model when $\sigma_{\delta}^{2}$ is unknow [@bmo].
Table \[cr1\] presents estimates of $\alpha,\,\beta,\,X_{0},$ $V(\hat{X}_{0})$ and the expanded uncertainty, $U(X_{0})$, from the proposed model (Proposed-M) of chromium, cadmium and lead elements. Also, we present the estimates obtained from usual calibration model (Usual-M) to observe the performance of both models.
In Table \[cr1\], for cadmium and lead elements, we observe that the estimates of $\alpha$, $\beta$ and $X_0$ from the Proposed-M and Usual-M are the same. For the chromium element, there are small differences. Also, we observe that for the chromium element there is a small difference between the estimates of $X_{0}$ and $U(X_{0})$ respectively obtained from the usual model and the proposed model. Despite the relevant estimates of $\alpha,\beta$ and $X_0$ from both approaches for cadmium and lead element match, the estimates of $V(X_{0})$ and $U(X_{0})$ differ considerably, the estimates obtained adopting the usual model is greater than the estimates outcome supplied by the proposed model.
--------- ------------ ------------ --------- ------------ ----------- --------- ------------ -----------
$X_{i}$ $u(X_{i})$ Intensity $X_{i}$ $u(X_{i})$ Intensity $X_{i}$ $u(X_{i})$ Intensity
0.05 0.00016 6455.900 0.05 0.00016 4.89733 0.05 0.00015 0.9471
0.11 0.00027 13042.933 0.10 0.00027 9.706 0.10 0.00025 1.46833
0.26 0.00040 32621.733 0.25 0.00041 23.41333 0.26 0.00039 3.09033
0.79 0.00122 97364.500 0.73 0.00122 69.73 0.77 0.00117 8.40533
1.05 0.00161 129178.100 1.01 0.00168 96.85667 1.01 0.00155 10.92667
--------- ------------ ------------ --------- ------------ ----------- --------- ------------ -----------
: Concentration $(mg/g)$, uncertainty($u(X_{i})$) and intensity of the standard solutions of chromium, cadmium and lead elements.[]{data-label="cr0"}
Chromium element Cadmium element Lead element
------------------ ----------------- --------------
10173.6 5.066 1.303
10516.9 5.027 1.290
10352.2 5.085 1.341
: Intensity of the sample solutions of chromium, cadmium and lead elements.[]{data-label="cr01"}
------------------ -------------- -------------- -------------- -------------- -------------- --------------
Parameters Usual-M Proposed-M Usual-M Proposed-M Usual-M Proposed-M
$\alpha$ 134.9469 124.2801 0.454801 0.454801 -0.3822126 -0.3822126
$\beta$ 123003.7 123027.3 10.54381 10.54381 94.29881 94.29881
$X_{0}$ 0.08302691 0.08309769 0.08123556 0.08123556 0.05770535 0.05770535
$V(\hat{X}_{0})$ 4.357870e-06 4.474395e-06 7.898643e-05 4.440342e-06 0.0001181068 7.237226e-08
$U(X_{0})$ 0.004091601 0.004145942 0.01741936 0.004130135 0.02130068 0.000527281
------------------ -------------- -------------- -------------- -------------- -------------- --------------
: Estimates of $\alpha,\,\beta,\,X_{0},$ $V(\hat{X}_{0})$ and $U(X_{0})$ related to usual and heteroscedastic model, for the chromium, cadmium and lead element.[]{data-label="cr1"}
Concluding remarks {#remarks}
==================
The expanded uncertainty of $X_0$ from the proposed model arises from the errors appearing in the both process, the reading of equipment and the heteroscedastic error in the preparation of standard solutions. We observe that, despite the classical model only considers the error originated from equipment reading, there are some applications in which the expanded uncertainty is greater than the one obtained through the new approach.\
Various aspects of the model studied above deserve attention in future research, e.g. it is not considerated the error arisen from the test sample solution preparation, the proposed model can be studied by considering other type of distribution of the errors, such as skew normal distribution [@azzalini]. In particular, one of the drawbacks of the usual model is that it does not consider the error in the independent variable, we believe that despite that this error being very small, it must be considered as an important property of the calibration model. We will concentrate on one of the problems described above in a future work.\
[**Acknowledgments**]{}
The authors are grateful to Prof. Dr. Heleno Bolfarine for carefully reading the manuscript and Dr. Olga Satomi from “Instituto de Pesquisas Tecnológicas” - IPT. Betsabé G. B. Achic has been supported by a grant from CNPq.
Usual calibration model {#appa}
=======================
The first and second stage equations of the usual linear calibration model are given, respectively, by $$\begin{aligned}
\label{m1}
Y_{i}&=&\alpha+\beta x_{i}+\epsilon_{i},\,\,\,\,\,i=1,2\cdots,n,\\
\label{m2}
Y_{0i}&=&\alpha+\beta X_{0}+\epsilon_{i},\,\,\,\,\,i=n+1,n+2,\cdots,n+k.\end{aligned}$$ It is considered the following assumptions:
- $x_{1},x_{2},\cdots,x_{n}$ take fixed values, which are considered as true values.
- $\epsilon_{1},\epsilon_{2},\cdots,\epsilon_{n+k}$ are independent and normally distributed with mean 0 and variance $\sigma_{\epsilon}^{2}$.
The model parameters are $\alpha,\beta,X_{0}$ and $\sigma_{\epsilon}^{2}$ and the main interest is to estimate the quantity $X_{0}$.
The maximum likelihood estimators of the usual calibration model are given by $$\begin{aligned}
\label{xo}
\hat{\alpha}&=&\bar{Y}-\hat{\beta}\bar{x},\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\hat{\beta}=\frac{S_{xY}}{S_{xx}},\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\hat{X}_{0}=\frac{\bar{Y}_{0}-\hat{\alpha}}{\hat{\beta}},\\
\label{se}
\sigma_{\epsilon}^{2}&=&\frac{1}{n+k}\big[\sum_{i=1}^{n}(Y_{i}-\hat{\alpha}-\hat{\beta}x_{i})^{2}+\sum_{i=n+1}^{n+k}(Y_{0i}-\bar{Y}_{0})^{2}\big],\end{aligned}$$ where $$\begin{aligned}
\nonumber \bar{x}&=&\frac{1}{n}\sum_{i=1}^{n}x_{i},\,\,\,\,\bar{Y}=\frac{1}{n}\sum_{i=1}^{n}Y_{i},\,\,\,\,S_{xY}=\frac{1}{n}\sum_{i=1}^{n}(x_{i}-\bar{x})(Y_{i}-\bar{Y}),\\\nonumber
S_{xx}&=&\frac{1}{n}\sum_{i=1}^{n}(x_{i}-\bar{x})^{2},\,\,\,\,\bar{Y}_{0}=\frac{1}{n}\sum_{i=n+1}^{n+k}Y_{0i}.\end{aligned}$$ The approximation of order $n^{-1}$ for the variance of $\hat{X}_{0}$ is given by $$\label{veurachem}
V_{1}(\hat{X}_{0})=\frac{\sigma_{\epsilon}^{2}}{\beta^{2}}\left[\frac{1}{k}+\frac{1}{n}+\frac{(\bar{x}-X_{0})^{2}}{nS_{xx}}\right].$$ In order to construct a confidence interval for $X_{0}$, we consider that $$\label{pibotep}
\frac{\hat{X}_{0}-X_{0}}{\sqrt{\hat{V}(\hat{X}_{0})}}\stackrel{D}{\longrightarrow
}N(0,1),$$ hence, the approximated confidence interval for $X_{0}$ with a confidence level $(1-\alpha)$, is given by $$\label{uinterv}
\left(\hat{X}_{0}-z_{\frac{\alpha}{2}}\sqrt{\hat{V}(\hat{X}_{0})},\hat{X}_{0}+z_{\frac{\alpha}{2}}\sqrt{\hat{V}(\hat{X}_{0})}\right),$$ where $z_{\frac{\alpha}{2}}$ is the quantile of order $(1-\frac{\alpha}{2})$ of the standard normal distribution.\
[\*\*]{} Shukla, G.K. On the problem of calibration. . EURACHEM/CITAC Guide(2000) Quantifying uncertainty in analytical measurement, 2nd edn., Final Draft April 2000. EURACHEM: http://www.measurementuncertainty.org S. H. Chui, Q; Zucchini R. R; and Lichtig J. Qualidade de medições em química analítica. Estudo de caso: determinação de cádmio por espectrometria de absorção atômica com chama.. Bruggemann, L. and Wenrich R. Evaluation of measurement uncertainty for analytical procedures using a linear calibration function. . Berkson, J. Are there two regression?. . Blas, B. G; Sandoval, M. C; Satomi, O. Homoscedastic controlled calibration model. . Ballico, M. Limitations of the Welch-Satterthwaite approximation for measurement uncertainty calculations. . Blas, B. G.(2005). Calibração controlada aplicada à química analítica. São Paulo: IME-USP. Dissertação de Mestrado. Tallis G.M. Note on a Calibration problem.. Lwin, T. and Maritz J. S. A note on the problem of statistical calibration. . International Organization for Standardization (1995). Guide to the expression of uncertainty in measurement (ISOGUM) . Azzalini, A. A class of distributions which includes the normal ones. .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present an exact solution of the Langevin’s equation in the steady state limit in a three dimensional, harmonic crystal of slab geometry whose boundary surfaces along its length are connected to two stochastic, white noise heat baths at different temperatures. We show that the heat transport obeys the Fourier’s law in the continuum limit.'
author:
- Shiladitya Acharya and Krishnendu Mukherjee
title: 'Fourier’s law of heat conduction in a three dimensional harmonic crystal: A retrospection'
---
When a steady temperature gradient is established between the two ends of a piece of solid bar, heat current will flow from high to low temperature end. According to Fourier’s law of heat conduction the current density is proportional to the temperature gradient and mathematically it reads as $${\bf J}({\bf x})=-\kappa\nabla T({\bf x}),
\label{Fourier}$$ where the constant of proportionality $\kappa$ is known as the thermal conductivity of the solid. Conduction of heat in solid by its very nature is a non-equilibrium process. This is an area of Physics, where the idea of non-equilibrium statistical mechanics can be applied to in order to find the underlying physical conditions for the validity of this law in case of solid. Various numerical and analytical studies confirm that the heat transport in one dimensional system exhibits anomalous[@Lepri1] behaviour. It means that the thermal conductivity for such a system is not found to be an intrinsic property of the material. It shows a power law dependence $\kappa \sim N^{\alpha}$, where $N$ be the linear size of the system. There are studies on different models which predict divergent ($0<\alpha<1$) thermal conductivity [@Lepri1; @Dhar1; @Grassberger1; @Narayan; @Mai]. There are also some oscillator models that give non-divergent ($\alpha <0$) thermal conductivity[@Dhar2] in one dimension. The anomalous behaviour of thermal conductivity is also observed in two dimensional system. Numerical study indicates a logarithmic divergence[@Lippi] of thermal conductivity $\kappa\sim \ln{N}$. A power law behaviour[@Grassberger2] is also observed in such a system.
There are strong numerical evidences[@Dhar3] that indicate the validity of Fourier’s law of heat conduction in one and two dimensional systems with pining and anharmonicity. An extensive investigation on heat transport in a three dimensional disordered harmonic crystal has been carried out recently[@Chaudhuri]. The numerical simulation indicates the normal transport of heat when this system is subjected to an external pining potential. Though it is not been verified numerically, but a finite conductivity is predicted for this disordered system from analytical arguments. A more recent simulation study [@Dhar4] establishes for the first time the validity of this law in three dimensional anharmonic crystal. It thus also establishes the fact that the process of heat conduction in three dimensional geometry is diffusive in nature. Apart from bringing in a temperature dependent contribution to the thermal conductivity, which is indeed the case for real systems, it is confirmed that anharmonicity provides a condition which is sufficient for normal heat transport in a solid. In this letter we give an exact analytical derivation of Fourier’s law of heat conduction in three dimensional harmonic crystal. We find that in the continuum limit the thermal conductivity is finite and does not depend on the system size.
We consider a cubic crystal in three dimension. The form of the Hamiltonian $$\begin{aligned}
H = \sum_{\bf n} \frac{\dot x_{\bf n}^2}{2}+\sum_{\bf n,\hat{e}} \frac{1}{2}
(x_{\bf n} - x_{\bf n+\hat{e}})^2. \end{aligned}$$ The displacement field $x_{\bf n}$ is defined on each lattice site ${\bf n}=(n_1,n_2,n_3)$ where $n_1=1,\cdots,N$, $n_2=1,\cdots,W_2$, and $n_3=1,\cdots,W_3$. Here ${\bf \hat{e}}$ denotes the unit vector in the three directions. We choose the value of mass attached to each lattice point and the harmonic spring constant as one. We have Langevin’s type heat baths that are coupled to the surfaces at $n_1=1$ and $n_1=N$ and are maintained at temperatures $T_L$ and $T_R$ ($T_L\,>T_R$) respectively. Hence the equation of motion of a particle at the site ${\bf n}$ reads $$\begin{aligned}
{\ddot{x}}_{\bf n}=&-&\sum_{{\bf \hat{e}}} (x_{\bf n}-x_{\bf n+\hat{e}})-
\gamma (\delta_{n_1,1}+ \nonumber
\delta_{n_1,N}){\dot{x}}_{\bf n}\\
&+&(\delta_{n_1,1} \eta_{\bf n}^L
+\delta_{n_1,N} \eta_{\bf n}^R).
\label{eom}\end{aligned}$$ We have chosen the noises to be white and they are uncorrelated at different sites. Noise strength is specified by $$\langle \eta_{\bf n}^{L,R}(t)
\eta_{\bf n^\prime}^{L,R}(t)\rangle
=2\gamma T_{L,R}\delta(t-t^\prime)\delta_{{\bf{n}},{\bf{n}}^\prime},
\label{noise1}$$ where we have chosen the Boltzmann constant $k_B=1$. We use the periodic boundary conditions for the displacement field and the noises in $n_2$ and $n_3$ directions: $$\begin{aligned}
x_{{\bf n}+(0,W_2,0)}(t) &=& x_{\bf n}(t) = x_{{\bf n}+(0,0,W_3)}(t)\nonumber\\
\eta_{{\bf n}+(0,W_2,0)}^{L,R}(t) &=& \eta_{\bf n}^{L,R}(t)
=\eta_{{\bf n}+(0,0,W_3)}^{L,R}(t)
\label{pbc} \end{aligned}$$ These periodic boundary conditions lead to the following expansion of $x_{\bf n}(t)$ and $\eta_{\bf n}^{L,R}(t)$: $$\begin{aligned}
x_{\bf n}(t)&=&\frac{1}{\sqrt{W_2 W_3}}
\sum_{p_2} \sum_{p_3} y_{n_1}(p_2,p_3,t)
{\rm e}^{i(p_2 n_2+p_3 n_3)a},\nonumber\\
& & \label{ftxn}\\
\eta_{\bf n}^{L,R}(t)&=&\frac{1}{\sqrt{W_2 W_3}}
\sum_{p_2} \sum_{p_3} f_{n_1}(p_2,p_3,t)
{\rm e}^{i(p_2 n_2+p_3 n_3)a},\nonumber\\
& &\label{ftetan}\end{aligned}$$ where $a$ be the lattice constant of the crystal. Upon substitution of Eqn.(\[ftxn\]) and (\[ftetan\]) into Eqn.(\[eom\]) we obtain $$\ddot{y}_j=-V_{jk}y_k-\gamma W_{jk}\dot{y}_k+f_j
\label{eom1}$$ where $$\begin{aligned}
W_{jk} &=& \delta_{j,1}\delta_{k,1}
+\delta_{j,N}\delta_{k,N},\label{Wjk}\\
f_j(p_2,p_3,t) &=& \delta_{j,1}f_L(p_2,p_3,t)
+\delta_{j,N}f_R(p_2,p_3,t)\label{fj},\end{aligned}$$ the $N\times N$ matrix $$V=\left(\begin{array}{ccccc}
2\omega_0^2 & -1 & 0 & 0 & \ldots \\
-1 & 2\omega_0^2 & -1 & 0 & \ldots \\
0 & -1 & 2\omega_0^2 & -1 & \ddots \\
\vdots & \ddots & \ddots & \ddots & \ddots \\
0 & \ldots & 0 & -1 & 2\omega_0^2
\end{array}\right)$$ and $$\omega_0^2(p_2,p_3) = 1+2\sin^2(\frac{p_2 a}{2})
+2\sin^2(\frac{p_3 a}{2}).
\label{w0}$$ Here $j, k=1,\cdots, N$. We have also assumed here that $y_0(p_2,p_3,t)=0=y_{N+1}(p_2,p_3,t)$. To solve Eqn.([\[eom1\]]{}) we diagonalize the matrix $V$. The solution of the $N$ order equation $\left|V-\alpha^2 I\right|=0$ gives the eigenvalues of $V$ as $$\alpha_k^2(p_1,p_2)=2\omega_0^2(p_1,p_2)+
2\cos\left(\frac{k\pi}{N+1}\right).
\label{eval}$$ The $j$-th component of the normalized eigenvector corresponding to the eigenvalue $\alpha_k^2$ is given by $$a^{(k)}_j=\sqrt{\frac{2}{N+1}} (-1)^{j+1} \sin\left(\frac{jk\pi}{N+1}\right).
\label{evec}$$ The diagonalizing matrix $A$ thus reads as $A_{jk}=a^{(k)}_j$ such that $A^TA=I$ and $A^TVA=\alpha^2$, where $(\alpha^2)_{jk}=\alpha_j^2\,\delta_{jk}$. We introduce a new set of coordinates $\xi_j$ as $$y_j(p_2,p_3,t)=A_{jk}\xi_k(p_2,p_3,t).
\label{yxi}$$ The equation of motion of $\xi_j$ in matrix form can be written as $$\ddot{\xi}=-\alpha^2\xi-\gamma Z \dot{\xi} +\tilde{f},
\label{eomxi}$$ where the symmetric matrix $Z=A^TWA$, and $\tilde{f}=A^Tf$. In the steady state limit ($t>>1/\gamma$) we are interested in the particular solution of the set of equations of motion of $\xi$. We use the Fourier transform of $$\xi_j(t)=\int_{-\infty}^\infty\frac{d\omega}{2\pi}\xi_j(\omega)
{\rm e}^{i\omega t}
~{\rm and}~
f_j(t)=\int_{-\infty}^\infty\frac{d\omega}{2\pi}f_j(\omega)
{\rm e}^{i\omega t}
\label{ftxif}$$ in Eqn.(\[eomxi\]) and obtain $$(-\omega^2\delta_{jk}+\alpha_j^2\delta_{jk}+i\gamma\omega Z_{jk})
\xi_k(\omega)=\tilde{f}_j(\omega).
\label{eomxiomega}$$ Since the dynamics of the system in the steady state is governed by the noises, we decompose $\xi_j(\omega)$ as $$\xi_j(\omega)=b(\omega)\tilde{f}_j(\omega)
\label{xiomega}$$ and then using this decomposition into Eqn.(\[eomxiomega\]) we obtain $$b(\omega) =
-\frac{1}{\omega^2-\alpha_j^2-i\gamma\omega}.
\label{b}$$ Now upon substitution of Eqn.(\[xiomega\]) into (\[ftxif\]) along with the use of Eqn.(\[fj\]), (\[b\]) we obtain $$\begin{aligned}
\xi_j(p_2,p_3,t) &=& -\int_{-\infty}^\infty\frac{d\omega}{2\pi}
\frac{{\rm e}^{i\omega t}}{\omega^2-\alpha_j^2-i\gamma\omega}\nonumber\\
& &\times[a^{(j)}_1\,f_L(p_2,p_3,\omega)
+a^{(j)}_N\,f_R(p_2,p_3,\omega)].\nonumber\\
& &
\label{xiss}\end{aligned}$$ Now the use of Eqn.(\[ftetan\]), (\[fj\]) and (\[ftxif\]) into (\[noise1\]) gives $$\begin{aligned}
& &\langle f_{L,R}(p_2,p_3,\omega)f_{L,R}(p_2^\prime,p_3^\prime,\omega^\prime)
\rangle\nonumber\\
&=&4\pi\gamma T_{L,R}\,\delta(\omega+\omega^\prime)\,\delta_{p_2+p_2^\prime,0}
\,\delta_{p_3+p_3^\prime,0}.
\label{noise2}\end{aligned}$$ To compute the correlation between position and velocity we use Eqn.(\[xiss\]) and (\[noise2\]) and after performing a frequency integration using delta function obtain $$\begin{aligned}
& &\langle \xi_{k_1}(p_2,p_3,t)\dot\xi_{k_2}(p_2^{\prime},
p_3^{\prime},t^\prime)\rangle \nonumber\\
&=&2\gamma
\,(a_1^{(k_1)}a_1^{(k_2)}T_L+a_N^{(k_1)}a_N^{(k_2)}T_R)
\,I_c(t-t^\prime)\nonumber\\
& &\times\delta_{p_2+p_2^{\prime},0}
\,\delta_{p_3+p_3^{\prime},0},
\label{correlation}\end{aligned}$$ where $$\begin{aligned}
& &I_c(t-t^\prime)\nonumber\\
&=& -i\int_{-\infty}^\infty\frac{d\omega}{2\pi}
\frac{\omega {\rm e}^{i\omega(t-t^\prime)}}
{(\omega^2-\alpha_{k_1}^2-i\gamma\omega)
(\omega^2-\alpha_{k_2}^2+i\gamma\omega)}.\end{aligned}$$ Performing the integration over $\omega$ we obtain $$\begin{aligned}
I_c(t-t^\prime) &=& \frac{{\rm e}^{-\gamma|t-t^\prime|/2}}
{4\,\Delta_d(\beta_1,\beta_2)}[I_c^>(t-t^\prime)\theta(t-t^\prime)\nonumber\\
& &+\,I_c^<(t-t^\prime)\theta(t^\prime-t)],
\label{ic}\end{aligned}$$ where $$\begin{aligned}
\Delta_d(\beta_1,\beta_2) &=& (\cos\beta_1-\cos\beta_2)^2\nonumber\\
& &+\,\gamma^2(2\omega_0^2
+\cos\beta_1+\cos\beta_2),\label{deld}\\
I_c^>(t-t^\prime) &=& 2(\cos\beta_1-\cos\beta_2)
\cos(\omega_{k_1}|t-t^\prime|)\nonumber\\
& &+\,\frac{\gamma}{\omega_{k_1}}\{(4\omega_0^2+3\cos\beta_1
+\cos\beta_2)\nonumber\\
& &\times\sin(\omega_{k_1}|t-t^\prime|)\},\label{icg}\\
I_c^<(t-t^\prime) &=& 2(\cos\beta_1-\cos\beta_2)
\cos(\omega_{k_2}|t-t^\prime|)\nonumber\\
& &-\,\frac{\gamma}{\omega_{k_2}}\{(4\omega_0^2+\cos\beta_1
+3\cos\beta_2)\nonumber\\
& &\times\sin(\omega_{k_2}|t-t^\prime|)\},\label{icl}\\
\beta_{1,2} &=& \pi k_{1,2}/(N+1),\label{beta12}\\
\omega_{k_{1,2}} &=& \sqrt{\alpha^2_{k_{1,2}}-\gamma^2/4}.
\label{omegak12}\end{aligned}$$ It is clear that $I_c(t-t^\prime)\rightarrow 0$, when $|t-t^\prime|\rightarrow\infty$ and when $t=t^\prime$ $$I_c(0) = \frac{\cos\beta_1-\cos\beta_2}{2\,\Delta_d(\beta_1,\beta_2)}.
\label{ic0}$$ For $1\le|k_1-k_2|\le N-1$, $I_c(0)$ remains finite when $N$ tends to infinity. According to Eqn.(\[evec\]) the factor appeared in Eqn.(\[correlation\]) $(a_1^{(k_1)}a_1^{(k_2)}T_L+a_N^{(k_1)}a_N^{(k_2)}T_R)
=2(T_L+(-1)^{k_1+k_2}T_R)\sin\beta_1\sin\beta_2/(N+1)$. It implies that even for zero momentum modes ($p_{2,3}=0$), which appear owing to the periodic boudary conditions imposed on the displacement field in $n_2$ and $n_3$ directions, the equal time correlation in Eqn.(\[correlation\]) goes as $N^{-\alpha}$ ($1\le\alpha\le 3$) when $N\rightarrow\infty$. The fall of this correlation as a negative power of $N$ in the thermodynamic limit indicates that the ballastic transport remains absent from the conduction process of heat[@Grassberger2].
Heat current density $j_{\bf n}$ from the lattice site ${\bf n}$ to ${\bf n+\hat{e}_1}$, where ${\bf \hat{e}_1}=(1,0,0)$, is given by[@Lepri1] $$\begin{aligned}
j_{\bf n}=\frac{1}{2}\langle (x_{\bf n + \hat{e}_1}-x_{\bf n})
(\dot{x}_{\bf n + \hat{e}_1}+\dot{x}_{\bf n})\rangle\end{aligned}$$ The average heat current density per bond[@Dhar4] $$J = \frac{1}{2W_2W_3 (N-1)}
\sum_{n_1=1}^{N-1}\sum_{n_2=1}^{W_2}\sum_{n_3=1}^{W_3} j_{\bf n}.$$ We substitute Eqn.(\[ftxn\]) and (\[yxi\]) in $J$ and after performing the summations over $n_2$ and $n_3$ obtain the average heat current density per bond in the steady state limit as $$\begin{aligned}
J&=&\frac{1}{2W_2W_3 (N-1)}
\,\sum_{p_2,p_3}\, \sum_{k_1,k_2=1}^N
\,\sum_{n_1=1}^{N-1}(a_{n_1+1}^{(k_1)} - a_{n_1}^{(k_1)})\nonumber\\
& &\times(a_{n_1+1}^{(k_2)} + a_{n_1}^{(k_2)})
\langle \xi_{k_1}(p_2,p_3,t)\dot{\xi}_{k_2}(-p_2,-p_3,t)\rangle.\nonumber\\
& &\end{aligned}$$ We now use Eqn.(\[evec\]) to evaluate the sum $$\begin{aligned}
& &\sum_{n_1=1}^{N-1}(a_{n_1+1}^{(k_1)} - a_{n_1}^{(k_1)})
(a_{n_1+1}^{(k_2)} + a_{n_1}^{(k_2)})\nonumber\\
&=&2(1-(-1)^{k_1+k_2})\sin\beta_1\sin\beta_2\nonumber\\
& &\times\Big[\frac{1}{\cos\beta_2-\cos\beta_1}-1\Big]
\label{n1sum}\end{aligned}$$ and then using (\[correlation\]) and (\[ic0\]) obtain $$\begin{aligned}
J&=&-\frac{2\gamma\,(T_L-T_R)}
{(N+1)^2(N-1)W_2W_3}
\,\sum_{p_2,p_3}\, \sum_{k_1,k_2=1}^N\nonumber\\
& &\times(1-(-1)^{k_1+k_2})
\frac{\sin^2\beta_1 \sin^2\beta_2}
{\Delta_d(\beta_1,\beta_2)}.
\label{hc1}\end{aligned}$$\
The factor $(1-(-1)^{k_1+k_2})$ ensures that the summation over $k_1$ and $k_2$ will be non zero only when $k_1+k_2$ is an odd number and hence we take the factor $(T_L+(-1)^{k_1+k_2}T_R)$ out of the summation as $(T_L-T_R)$. In the continuum limit, when $a\rightarrow 0$ and $W_{2,3}\rightarrow \infty$ keeping $a\,W_{2,3}$ at fixed values, we convert the discrete sums over $p_2$ and $p_3$ into integrals: $$\sum_{p_{2,3}} \rightarrow \frac{a\,W_{2,3}}{2 \pi}
\int_{-\frac{\pi}{a}}^{-\frac{\pi}{a}}\,dp_{2,3}.$$ Evaluation of the integrals[@Gradshteyn] over $p_2$ and $p_3$ gives $$J=-\frac{2\gamma\,(T_L-T_R)}{N-1} I(N,\gamma),$$ where $$\begin{aligned}
& &I(N,\gamma) = \frac{1}{(N+1)^2}\sum_{k_1,k_2=1}^N
(1-(-1)^{k_1+k_2})\nonumber\\
& &\times\frac{\sin^2\beta_1\sin^2\beta_2}{\Delta(\beta_1,\beta_2)}
F\left(\frac{1}{2},\frac{1}{2},1;
\,(4\gamma^2/\Delta(\beta_1,\beta_2))^2\right).
\label{ing}\end{aligned}$$ Here the function $$\begin{aligned}
\Delta(\beta_1,\beta_2) &=& (\cos\beta_1-\cos\beta_2)^2\nonumber\\
& &\,+\gamma^2\,(6+\cos\beta_1+\cos\beta_2).
\label{delta}\end{aligned}$$ $I(N,\gamma)$ is zero if $k_1$ and $k_2$ simultaneously take even integer values or odd integer values. Assuming that $N$ be an even number and using the fact that the summand of Eqn.(\[ing\]) is symmetric in respect of the interchange of $\beta_1$ and $\beta_2$, we rewrite the double sum of $$\begin{aligned}
I(N,\gamma) &=& \frac{4}{(N+1)^2}\sum_{j_1,j_2=1}^{N/2}
\frac{\sin^2\tilde{\beta}_1\sin^2\tilde{\beta}_2}
{\Delta(\tilde{\beta}_1,\tilde{\beta}_2)}\nonumber\\
& &\times F\left(\frac{1}{2},\frac{1}{2},1;
\,(4\gamma^2/\Delta(\tilde{\beta}_1,\tilde{\beta}_2))^2\right),
\label{ing1}\end{aligned}$$ where $\tilde\beta_1=2\pi j_1/(N+1)$ and $\tilde\beta_2=\pi(2j_2-1)/(N+1)$. Again in the continuum limit we convert this double sum into integrals. In this limit $a\rightarrow 0$ and $N\rightarrow\infty$ keeping $Na$ at a fixed value. Defining the integration variables in this limit as $\theta_{1,2}=2\pi j_{1,2}/(N+1)$, we convert the discrete sums into integrals: $$\frac{2}{N+1}\sum_{j_{1,2}=1}^{N/2}\rightarrow
\frac{1}{\pi}\int_0^\pi d\theta_{1,2}.$$ $I(N,\gamma)$ thus takes the form $$\begin{aligned}
g(\gamma) &=& \lim_{N\rightarrow\infty}I(N,\gamma)\nonumber\\
&=&\frac{1}{\pi^2}\int_0^\pi d\theta_1\int_0^\pi d\theta_2
\frac{\sin^2\theta_1\sin^2\theta_2}
{\Delta(\theta_1,\theta_2)}\nonumber\\
& &\times F\left(\frac{1}{2},\frac{1}{2},1;
\,(4\gamma^2/\Delta(\theta_1,\theta_2))^2\right).
\label{gg}\end{aligned}$$ Hence we obtain the steady state current density per bond in the continuum limit $$J = -\kappa\frac{(T_L-T_R)}{N-1},$$ where the conductivity $$\kappa=2\gamma\,g(\gamma).
\label{kpgm}$$ Here $\kappa$ is found to be independent of the size of the system. The variation of the thermal conductivity $\kappa$ as a function of $\gamma$, as given by Eqn.(\[kpgm\]), is plotted in Fig.\[Figkg\].
![(Color online) Plot of $\kappa$ as a function of $\gamma$[]{data-label="Figkg"}](graph1.eps){width="48.00000%"}
-0.45in
Here $\gamma$ appears as a constant in the dissipative force term of the Langevin’s equation. Physically this force term denotes a viscous force experienced by the particles of Brownian like at the boundary surfaces of the crystal owing to collisions with the particles of fluid which seems to constitute the heat baths[@Chaikin]. The increase of $\gamma$, reduces the mobilities of the Brownian particles and thereby reducing their velocities[@Chaikin; @Reif]. Consequently, the velocities of the particles at the surfaces next to the boundaries will also fall because those are connected by springs with the particles at the boundaries. This fall of velocities of the particles at the neighbouring surfaces of the boundaries will reduce the rate of flow of heat from the boundaries to the crystal itself and thereby reducing the thermal conductivity of the system. Hence, it justifies reasonably the nature of variation of $\kappa$ with $\gamma$ as shown in Fig.\[Figkg\].
The average of the square of velocity of a layer at $n_1$ reads $$\begin{aligned}
v_{avg}^2(n_1) &=& \frac{1}{W_2W_3}
\sum_{n_2=1}^{W_2}\sum_{n_3=1}^{W_3}
\langle{\dot{x}}^2_{\bf{n}}\rangle\nonumber\\
&=&\frac{1}{W_2W_3}\sum_{p_2,p_3}\sum_{k_1,k_2=1}^N
a_{n_1}^{(k_1)}a_{n_1}^{(k_2)}\nonumber\\
& &\times\langle\dot{\xi}_{k_1}(p_2,p_3,t)
\dot{\xi}_{k_2}(-p_2,-p_3,t)\rangle.
\label{vav1}\end{aligned}$$ We use Eqn.(\[xiss\]) to compute the velocity-velocity correlation as $$\begin{aligned}
& &\langle\dot{\xi}_{k_1}(p_2,p_3,t)
\dot{\xi}_{k_2}(-p_2,-p_3,t)\rangle \nonumber\\
&=& \frac{2\gamma^2}{N+1}(T_L+(-1)^{k_1+k_2})
\sin\beta_1\sin\beta_2\nonumber\\
& &\times\frac{2\omega_0^2+\cos\beta_1+\cos\beta_2}
{\Delta_d(\beta_1,\beta_2)}
\label{vvcorrelation}\end{aligned}$$ Upon substitution of Eqn.(\[vvcorrelation\]) into Eqn.(\[vav1\]) and evaluation of $p_2$ and $p_3$ sum in the continuum limit along $n_2$ and $n_3$ directions, give $$v_{avg}^2(n_1)=h_L(n_1,N)T_L+h_R(n_1,N)T_R$$ where $$\begin{aligned}
h_L(n_1,N) &=& \frac{4}{(N+1)^2}\sum_{k_1,k_2=1}^N
\frac{\Lambda(\beta_1,\beta_2)}{\Delta(\beta_1,\beta_2)}\nonumber\\
& &\times\sin(n_1\beta_1)\sin(n_1\beta_2)\sin\beta_1\sin\beta_2,
\label{hl}\\
h_R(n_1,N) &=& \frac{4}{(N+1)^2}\sum_{k_1,k_2=1}^N
(-1)^{k_1+k_2}\frac{\Lambda(\beta_1,\beta_2)}
{\Delta(\beta_1,\beta_2)}\nonumber\\
& &\times\sin(n_1\beta_1)\sin(n_1\beta_2)\sin\beta_1\sin\beta_2
\label{hr},\\
\Lambda(\beta_1,\beta_2) &=& \{(\cos\beta_1-\cos\beta_2)^2\nonumber\\
& &\times[1- F(1/2, 1/2, 1; (4\gamma^2/\Delta(\beta_1,\beta_2))^2)]\}
\nonumber\\
& &+\gamma^2(6+\cos\beta_1+\cos\beta_2).
\label{bigl}\end{aligned}$$
![(Color online) Plot of $v_{avg}^2$ as a function of $n_1$[]{data-label="Figavg1"}](graph2.eps){width="48.00000%"}
-0.45in
![(Color online) Plot of $v_{avg}^2$ as a function of $n_1$[]{data-label="Figavg2"}](graph3.eps){width="48.00000%"}
-0.45in
Our evaluation suggests that for $\gamma=0.01$, $h_L$ tends to $0.0396$ and $0$ and $h_R$ tends to $0$ and $0.0396$ at $n_1=1$ and $n_1=N$ respectively when $N\rightarrow\infty$. It indicates that as $h_L$ and $h_R$ are monotonically decreasing and increasing functions of $n_1$ respectively, $v_{avg}^2$ attains a minimum at any layer in the region between $n_1=1$ and $n_1=N$ and it is also evident from our plots given in Fig.\[Figavg1\] and \[Figavg2\]. Since, $v_{avg}^2(n_1)$ is proportional to $T(n_1)$, the temperature of the layer at $n_1$, $T(n_1)$ also exhibits a minimum in the region $1<n_1<N$. This concave upward nature of $T(n_1)$ has also been predicted in Ref.[@Dhar4]
In summary, we have given an exact analytical derivation of Fourier’s law of heat conduction in a three dimensional harmonic crystal. It shows that in three dimensions without introducing any pinning or disorder, harmonicity alone can give rise to a normal transport of heat in the crystal in the continuum limit.
[99]{} For a review on heat conduction see S. Lepri, R. Livi and A. Politi, Phys. Rep. [**[377]{}**]{}, 1 (2003). A. Dhar, Phys. Rev. Lett. [**[86]{}**]{}, 3554 (2001). P. Grassberger, W. Nadler, and L. Yang, Phys. Rev. Lett. [**[89]{}**]{}, 180601 (2002). O. Narayan and S. Ramaswamy, Phys. Rev. Lett. [**[89]{}**]{}, 200601 (2002). T. Mai, A. Dhar and O. Narayan, Phys. Rev. Lett. [**[98]{}**]{}, 184301 (2007). A. Dhar, Phys. Rev. Lett. [**[86]{}**]{}, 5882 (2001). A. Lippi and R. Livi, J. Stat. Phys. [**[100]{}**]{}, 1147 (2000). P. Grassberger and L. Yang, cond-mat/0204247. A. Dhar, Adv. Phys. [**[57]{}**]{}, 457 (2008). A. Chaudhuri, A. Kundu, D. Roy, A. Dhar J. L. Lebowitz and H. Spohn, Phys. Rev. B [**[81]{}**]{}, 064301 (2010). K. Saito and A. Dhar, Phys. Rev. Lett. [**[104]{}**]{}, 040601 (2010). I. S. Gradshteyn and I. M. Ryzhik, *Tables of Integrals, Series and Products*, 6th ed. (Academic Press, New Delhi, 2001). P. M. Chaikin and T. C. Lubensky, *Principles of condensed matter physics*, (Cambridge University Press, New Delhi, 2009). F. Reif, *Fundamentals of Statistical and Thermal Physics*, (McGraw-Hill, Singapore, 1985).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The composite Fermion (CF) picture offers a simple intuitive way of understanding many of the surprising properties of a strongly interacting two-dimensional electron fluid in a large magnetic field. The simple way in which the mean field CF picture describes the low lying bands of angular momentum multiplets for any value of the applied magnetic field is illustrated and compared with the results of exact numerical diagonalization of small systems. The justification of the success of the CF approach is discussed in some detail, and a CF hierarchy picture of the incompressible quantum fluid states is introduced. The CF picture is used to understand the energy spectrum and photoluminescence of systems containing both electrons and valence band holes.'
address:
- ' University of Tennessee, Knoxville, Tennessee 37996, USA'
- |
University of Tennessee, Knoxville, Tennessee 37996, USA\
and Wroclaw University of Technology, 50-370 Wroclaw, Poland
author:
- 'John J. Quinn'
- Arkadiusz Wójs
title: Composite Fermions in Fractional Quantum Hall Systems
---
Introduction {#sec1}
=============
The study of the electronic properties of quasi-two-dimensional (2D) systems has resulted in a number of remarkable discoveries in the past two decades [@ep2ds]. Among the most interesting of these are the integral [@klitzing] and fractional [@tsui] quantum Hall effects. In both of these effects, incompressible states of a 2D electron liquid are found at particular values of the electron density for a given value of the magnetic field applied normal to the 2D layer.
The integral quantum Hall effect (IQHE) is rather simple to understand. The incompressibility results from a cyclotron energy gap, $\hbar\omega_c$, in the single particle spectrum. When all states below the gap are filled and all states above it are empty, it takes a finite energy $\hbar\omega_c$ to produce an infinitesimal compression. Excited states consist of electron–hole pair excitations and require a finite excitation energy. Both localized [@anderson] and extended single particle states are necessary to understand the experimentally observed behavior of the magneto-conductivity [@laughlin1].
The fractional quantum Hall effect (FQHE) is more difficult to understand and more interesting in terms of new basic physics. The energy gap that gives rise to the Laughlin [@laughlin2] incompressible fluid state is completely the result of the interaction between the electrons. The elementary excitations are fractionally charged Laughlin quasiparticles, which satisfy fractional statistics [@halperin1]. The standard techniques of many body perturbation theory are incapable of treating FQH systems because of the complete degeneracy of the single particle levels in the absence of the interactions. Laughlin [@laughlin2] was able to determine the form of the ground state wave function and of the elementary excitations on the basis of physical insight into the nature of the many body correlations. Striking confirmation of Laughlin’s picture was obtained by exact diagonalization of the interaction Hamiltonian within the subspace of the lowest Landau level of small systems [@haldane1]. Jain [@jain1], Lopez and Fradkin [@lopez], and Halperin et al. [@halperin2] have extended Laughlin’s approach and developed a composite Fermion (CF) description of the 2D electron gas in a strong magnetic field. This CF description has offered a simple picture for the interpretation of many experimental results. However, the underlying reason for the validity of many of the approximations used with the CF approach is not completely understood [@wojs1].
The object of this review is to present a simple and understandable summary of the CF picture as applied to FQH systems. Exact numerical calculations for up to eleven electrons on a spherical surface will be compared with the predictions of the mean field CF picture. The CF hierarchy [@sitko1] will be introduced, and its predictions compared with numerical results. It will be shown that sometimes the mean field CF hierarchy correctly predicts Laughlin-like incompressible ground states, and that sometimes it fails.
The CF hierarchy depends on the validity of the mean field approximation. This seems to work well in predicting not only the Laughlin–Jain families of incompressible ground states at particular values of the applied magnetic field, but also in predicting the lowest lying band of states at any value of the magnetic field. The question of when the mean field CF picture works and why [@wojs1] will be discussed in some detail. As first suggested by Haldane [@haldane1], the behavior of the pseudopotential $V(L)$ describing the energy of interaction of a pair of electrons as a function of their total angular momentum $L$ is of critical importance. Some examples of other strongly interacting 2D Fermion systems will be presented, and some problems not yet completely understood will be discussed.
The plan of the paper is as follows. In section \[sec2\] the single particle states for electrons confined to a plane in the presence of an applied magnetic field [@gasiorowicz] are introduced. The integral and fractional quantum Hall effects are discussed briefly. Haldane’s idea [@haldane1] that the condensation of Laughlin quasiparticles leads to a hierarchy containing all odd denominator fractions is discussed. In section \[sec3\] the numerical calculations for a finite number of electrons confined to a spherical surface in the presence of a radial magnetic field are discussed. Results for a ten electron system at different values of the magnetic field are presented. In section \[sec4\] the ideas of fractional statistics and the Chern–Simons transformation are introduced. In section \[sec5\] Jain’s CF approach [@jain1] is outlined. The sequence of Jain condensed states (given by filling factor $\nu=n(1+2pn)^{-1}$, where $n$ is any integer and $p$ is a positive integer) is shown to result from the mean field approximation. The application of the CF picture to electrons on a spherical surface is shown to predict the lowest band of angular momentum multiplets in a very simple way that involves only the elementary problem of addition of angular momenta [@chen1]. In section \[sec6\] the two energy scales, the Landau level separation $\hbar\omega_c$ and the Coulomb energy $e^2/\lambda$ (where $\lambda$ is the magnetic length), are discussed. It is emphasized that the Coulomb interactions and Chern–Simons gauge interactions between fluctuations (beyond the mean field) cannot possibly cancel for arbitrary values the applied magnetic field. The reason for the success of the CF picture is discussed in terms of the behavior of the pseudopotential $V(L)$ and a kind of “Hund’s rule” for monopole harmonics [@wojs1]. In section \[sec7\], a phenomenological Fermi liquid picture is introduced to describe low lying excited states containing three or more Laughlin quasiparticles [@sitko2]. In section \[sec8\] the CF hierarchy picture [@sitko1] is introduced. Comparison with exact numerical results indicates that the behavior of the quasiparticle pseudopotential is of critical importance in determining the validity of this picture at a particular level of the hierarchy. In section \[sec9\] systems containing electrons and valence band holes are investigated [@wojs2]. The photoluminescence and the role of excitons and negatively charged exciton complexes is discussed. The final section is a summary.
Integral and Fractional Quantum Hall Effects {#sec2}
=============================================
The Hamiltonian for an electron confined to the $x$–$y$ plane in the presence of a perpendicular magnetic field $\bi{B}$ is $$H_0={1\over2\mu}\left(\bi{p}+{e\over c}\bi{A}\right)^2.$$ Here $\mu$ is the effective mass, $\bi{p}=(p_x,p_y,0)$ is the momentum operator and $\bi{A}(x,y)$ is the vector potential (whose curl gives $\bi{B}$). For the “symmetric gauge,” $\bi{A}={1\over2}B(y,-x,0)$, the single particle eigenfunctions [@gasiorowicz] are of the form $\psi_{nm}
(r,\theta)=e^{-im\theta}u_{nm}(r)$. The angular momentum of the state $\psi_{nm}$ is $-m$ and its eigenenergy is given by $$E_{nm}={1\over2}\hbar\omega_c(2n+1+|m|-m).$$ In these equations, $\omega_c=eB/\mu c$ is the cyclotron frequency, $n=0$, 1, 2, …, and $m=0$, $\pm1$, $\pm2$, …. The lowest energy states (lowest Landau level) have $n=0$ and $m=0$, 1, 2, … and energy $E_{0m}={1\over2}\hbar\omega_c$. It is convenient to introduce a complex coordinate $z=re^{-i\theta}
=x-iy$, and to write the lowest Landau level wavefunctions as $$\label{eq3}
\psi_{0m}(z)=N_mz^me^{-|z|^2/4},$$ where $N_m$ is a normalization constant. In this expression we have used the magnetic length $\lambda=
\sqrt{\hbar c/eB}$ as the unit of length. The function $|\psi_{0m}|^2$ has its maximum value at a radius $r_m$ which is proportional to $\sqrt{m}$. All single particle states belonging to a given Landau level are degenerate, and separated in energy from neighboring levels by $\hbar\omega_c$.
If the system has a “finite radial range,” then the $m$ values are restricted to being less than some maximum value ($m=0$, 1, 2, …, $N_\phi-1$). The value of $N_\phi$ (the Landau level degeneracy) is equal to the total flux through the sample, $BC$ (where $C$ is the area), divided by the quantum of flux $\phi_0=hc/e$. The filling factor $\nu$ is defined as the ratio of the number of electrons, $N$, to $N_\phi$. When $\nu$ has an integral value, an infinitesimal decrease in the area $C$ requires promotion of an electron across the cyclotron gap $\hbar\omega_c$ to the first unoccupied Landau level, making the system incompressible. This incompressibility together with the existence of both localized and extended states in the system is responsible for the observed behavior of the magneto-conductivity of quantum Hall systems at integral filling factors [@laughlin1].
In order to construct a many electron wavefunction $\Psi(z_1,z_2,\dots,
z_N)$ corresponding to a completely filled lowest Landau level, the product function which places one electron in each of the $N_\phi=N$ orbitals $\psi_{0m}$ ($m=0$, 1, …, $N_\phi-1$) must be antisymmetrized. This can be done with the aid of a Slater determinant $$\label{eq4}
\Psi\propto\left|
\matrix{ 1 & 1 & \dots & 1 \cr
z_1 & z_2 & \dots & z_N \cr
z_1^2 & z_2^2 & \dots & z_N^2 \cr
\vdots & \vdots & & \vdots \cr
z_1^{N-1} & z_2^{N-1} & \dots & z_N^{N-1} }
\right| \exp\left(-{1\over4}\sum_k|z_k|^2\right).$$ The determinant in equation (\[eq4\]) is the well-known Vandemonde determinant. It is not difficult to show that it is equal to $\prod_{i<j}(z_i-z_j)$. Of course, $N_\phi$ is equal to $N$ (since each of the $N_\phi$ orbitals is occupied by one electron) and the filling factor $\nu=1$.
Laughlin noticed that if the factor $(z_i-z_j)$ arising from the Vandemonde determinant was replaced by $(z_i-z_j)^{2p+1}$, where $p$ was an integer, the wavefunction $$\label{eq5}
\Psi_{2p+1}
\propto\prod_{i<j}(z_i-z_j)^{2p+1}
\exp\left(-{1\over4}\sum_i|z_i|^2\right)$$ would be antisymmetric, keep the electrons further apart (and therefore reduce the Coulomb repulsion), and correspond to a filling factor $\nu=(2p+1)^{-1}$. This results because the highest power of $z_i$ in the polynomial factor in $\Psi_{2p+1}$ is $(2p+1)(N-1)$ and it must be equal to the highest orbital index ($m=N_\phi-1$), giving $N_\phi-1=(2p+1)(N-1)$ and $\nu=N/N_\phi$ equal to $(2p+1)^{-1}$ in the limit of large systems. The additional factor $\prod_{i<j}(z_i-z_j)^{2p}$ multiplying $\Psi_{m=1}$ is the Jastrow factor which accounts for correlations between electrons.
It is observed experimentally that states with filling factors $\nu=2/5$, 3/5, 3/7, etc. exhibit FQH behavior in addition to the Laughlin $\nu=(2p+1)^{-1}$ states. Haldane [@haldane1] suggested that a hierarchy of condensed states arose from the condensation of quasiparticles (QP’s) of “parent” FQH states. In his picture, Laughlin condensed states of the electron system occurred when $N_\phi=(2p+1)N_e$, where the exponent $2p+1$ in equation (\[eq5\]) was an odd integer and the symbol $N_e$ denoted the number of electrons. Condensed QP states occurred when $N_e=2qN_{\rm QP}$, because the number of places available for inserting a QP in a Laughlin state was $N_e$. Haldane required the exponent $2q$ to be even “because the QP’s are bosons.” This scheme gives rise to a hierarchy of condensed states which contains all odd denominator fractions. Haldane cautioned that the validity of the hierarchy scheme at a particular level depended upon the QP interactions which were totally unknown.
Numerical Study of Small Systems {#sec3}
=================================
Haldane [@haldane1] introduced the idea of putting a small number of electrons on a spherical surface of radius $R$ at the center of which is a magnetic monopole of strength $2S\phi_0$. The single particle Hamiltonian can be expressed as [@fano] $$H_0={\hbar^2\over2\mu R^2}(\bi{L}-S\hat{R})^2,$$ where $\bi{L}$ is the angular momentum operator (in units of $\hbar$), $\hat{R}$ is the unit vector in the radial direction, and $\mu$ is the mass. The components of $\bi{L}$ satisfy the usual commutation rules $[L_\alpha,L_\beta]=i\epsilon_{\alpha\beta\gamma}L_\gamma$. The eigenstates of $H_0$ can be denoted by $\left|l,m\right>$; they are eigenfunctions of $L^2$ and $L_z$ with eigenvalues $l(l+1)$ and $m$, respectively. The lowest energy eigenvalue (shell) occurs for $l=S$ and has energy ${1\over2}\hbar\omega_c$. The $n$th excited shell has $l=S+n$, and $$E_n={\hbar\omega_c\over2S}\left[l(l+1)-S^2\right]
=\hbar\omega_c\left[n+{1\over2}+{n(n+1)\over2S}\right],$$ where the cyclotron energy is equal to $\hbar\omega_c=S\hbar^2/\mu
R^2$ and the magnetic length is $\lambda=R/\sqrt{S}$. If we concentrate on a partially filled lowest Landau level we have only $N_\phi=2S+1$ degenerate single particle states (since the electron angular momentum $l$ must be equal to $S$ and its $z$-component $m$ can take on values between $-l$ and $l$). The Hilbert space ${\cal H}_{\rm MB}$ of $N$ electrons in these $N_\phi$ single particle states contains $N_{\rm MB}=N_\phi![N!(N_\phi-N)!]^{-1}$ antisymmetric many body states. The single particle configurations $\left|m_1,m_2,\dots,m_N\right>=
c_{m_1}^\dagger c_{m_2}^\dagger\dots c_{m_N}^\dagger\left|{\rm vac}
\right>$ can be chosen as a basis of ${\cal H}_{\rm MB}$. Here $c_m^\dagger$ creates an electron in the single particle state $\left|l=S,m\right>$, and $\left|{\rm vac}\right>$ is the vacuum state. The space ${\cal H}_{\rm MB}$ can also be spanned by the angular momentum eigenfunctions, $\left|L,M,\alpha\right>$, where $L$ is the total angular momentum, $M$ its $z$-component, and $\alpha$ is a label which distinguishes different multiplets with the same $L$. If $\hbar\omega_c\gg e^2/\lambda$, the diagonalization of the interaction Hamiltonian $$H_I=\sum_{i<j}{e^2\over r_{ij}}$$ in the Hilbert space ${\cal H}_{\rm MB}$ of the lowest Landau level gives an excellent approximation to exact eigenstates of an interacting $N$ electron system. The single particle configuration basis is particularly convenient since the many body interaction matrix elements in this basis, $\left<m_1,m_2,\dots,m_N|H_I|m'_1,m'_2,\dots,m'_N\right>$, are expressed through the two body ones, $\left<m_1,m_2|H_I|m'_1,m'_2
\right>$, in a very simple way. On the other hand, using the angular momentum eigenstates $\left|L,M,
\alpha\right>$ allows the explicit decomposition of the total Hilbert space ${\cal H}_{\rm MB}$ into total angular momentum eigensubspaces. Because the interaction Hamiltonian is a scalar, the Wigner–Eckart theorem tells us that $$\left<L',M',\alpha'|H_I|L,M,\alpha\right>=
\delta_{LL'}\delta_{MM'}V_{\alpha\alpha'}(L),$$ where the reduced matrix element $$V_{\alpha\alpha'}(L)=\left<L,\alpha'|H_I|L,\alpha\right>$$ is independent of $M$. The eigenfunctions of $L$ are simpler to find than those of $H_I$, because efficient numerical techniques exist for obtaining eigenfunctions of operators with known eigenvalues. Finding the eigenfunctions of $L$ and then using the Wigner–Eckart theorem considerably reduces dimensions of the matrices that must be diagonalized to obtain eigenvalues of $H_I$. Some matrix dimensions are listed in table \[tab1\], where the degeneracy of the lowest Landau level and the dimensions of the total many body Hilbert space, $N_{\rm MB}$, and of the largest $M$ subspace, $N_{\rm MB}(M=0)$, are given for the Laughlin $\nu=1/3$ state of six to eleven electron systems (the $N$ electron Laughlin $\nu=(2p+1)^{-1}$ state occurs at $N_\phi=(2p+1)(N-1)$).
[@rcrrr]{} $N$ & $N_\phi$ & $N_{\rm MB}$ & $N_{\rm MB}(M=0)$\
6 & 16 & 8,008 & 338\
7 & 19 & 50,388 & 1,656\
8 & 22 & 319,770 & 8,512\
9 & 25 & 2,042,975 & 45,207\
10 & 28 & 13,123,110 & 246,448\
11 & 31 & 84,672,315 & 1,371,535\
\[tab1\]
For example, in the eleven electron system at $\nu=1/3$, the $L=0$ block that must be diagonalized to obtain the Laughlin ground state is only 1160 by 1160, small compared to the total dimension of 1,371,535 for the entire $M=0$ subspace.
------------------------------------------------------------------------
=3.75in
Typical results for the energy spectrum are shown in figure \[fig1\] for $N=10$ and a few different values of $2S$ between 21 and 30. The low energy bands marked with open circles and solid lines will be discussed in detail in the following sections. Frames (a) and (f) show two $L=0$ incompressible ground states: Laughlin state at $\nu=1/3$ and Jain state at $\nu=2/5$, respectively. In other frames, a number of QP’s form the lowest energy bands.
Chern–Simons Transformation and Statistics in 2D Systems {#sec4}
=========================================================
Before discussing the Chern–Simons gauge transformation and its relation to particle statistics, it is useful to look at a system of two particles each of charge $-e$ and mass $\mu$, confined to a plane, in the presence of a perpendicular magnetic field $\bi{B}
=(0,0,B)=\nabla\times\bi{A}(\bi{r})$. Because $\bi{A}$ is linear in the coordinate $\bi{r}=(x,y)$ \[e.g., in the symmetric gauge, $\bi{A}(\bi{r})={1\over2}B(y,-x)$\], the Hamiltonian separates into the center of mass (CM) and relative (REL) coordinate pieces, with $\bi{R}={1\over2}(\bi{r}_1+\bi{r}_2)$ and $\bi{r}=\bi{r}_1-\bi{r}_2$ being the CM and REL coordinates, respectively. The energy spectra of $H_{\rm CM}$ and $H_{\rm REL}$ are identical to that of a single particle of mass $\mu$ and charge $-e$. We have already seen that for the lowest Landau level $\psi_{0m}=
N_mr^me^{-im\phi}e^{-r^2/4\lambda^2}$. For the relative motion $\phi=\phi_1-\phi_2$, and an interchange of the pair, $P\psi(\bi{r}_1,\bi{r}_2)=\psi(\bi{r}_2,\bi{r}_1)$, is accomplished by replacing $\phi$ by $\phi+\pi$. In 3D systems, where two consecutive interchanges must result in the original wavefunction, this implies that $e^{im\pi}$ must be equal to either $+1$ ($m$ even; Bosons) or $-1$ ($m$ odd; Fermions). It is well-known [@leinaas] that for 2D systems $m$ need not be an integer. Interchange of a pair of identical particles can give $P\psi(\bi{r}_1,
\bi{r}_2)=e^{i\pi\theta}\psi(\bi{r}_1,\bi{r}_2)$, where the statistical parameter $\theta$ can assume non-integral values leading to anyon statistics.
A Chern–Simons (CS) transformation is a singular gauge transformation [@halperin2] in which an electron creation operator $\psi_e^\dagger(\bi{r})$ is replaced by a composite particle operator $\psi^\dagger(\bi{r})$ given by $$\psi^\dagger(\bi{r})=\psi_e^\dagger(\bi{r})\exp\left[
i\alpha\int d^2\bi{r}' \arg(\bi{r}-\bi{r}')
\psi^\dagger(\bi{r}')\psi(\bi{r}')\right].$$ Here $\arg(\bi{r}-\bi{r}')$ is the angle the vector $\bi{r}-\bi{r}'$ makes with the $x$-axis and $\alpha$ is an arbitrary parameter. The kinetic energy operator can be written in terms of the transformed operator as $$K={1\over2\mu}\int d^2\bi{r}\; \psi^\dagger(\bi{r})\left[
-i\hbar\nabla
+{e\over c}\bi{A}(\bi{r})
+{e\over c}\bi{a}(\bi{r})
\right]^2 \psi(\bi{r}).$$ Here $$\label{eq9a}
\bi{a}_{\bi{r}'}(\bi{r})
={\alpha\phi_0\over2\pi}\cdot
{\hat{z}\times(\bi{r}-\bi{r}')\over|\bi{r}-\bi{r}'|^2}$$ and $$\label{eq9}
\bi{a}(\bi{r})=\alpha\phi_0\int d^2\bi{r}'\;
\bi{a}_{\bi{r}'}(\bi{r})\;
\psi^\dagger(\bi{r}')\psi(\bi{r}'),$$ where $\hat{z}$ is a unit vector perpendicular to the 2D layer. The CS transformation can be thought of as an attachment to each particle of flux tube carrying a fictitious flux $\alpha\phi_0$ (where $\phi_0=hc/e$ is the quantum of flux) and a fictitious charge $-e$ which couples in the standard way to the vector potential caused by the flux tubes on every other particle. The $\bi{a}_{\bi{r}'}(\bi{r})$ is interpreted as the vector potential at position $\bi{r}$ due to a magnetic flux of strength $\alpha\phi_0$ localized at $\bi{r}'$, and $\bi{a}(\bi{r})$ is the total vector potential at position $\bi{r}$ due to all CS fluxes. The CS magnetic field associated with the particle at $\bi{r}'$ is $\bi{b}(\bi{r})=\nabla\times\bi{a}_{\bi{r}'}(\bi{r})=\alpha
\phi_0\delta(\bi{r}-\bi{r}')\hat{z}$. Because two charged particles cannot occupy the same position, one particle never senses the magnetic field of other particles, but it does sense the vector potential resulting from their CS fluxes. The classical equations of motion are unchanged by the presence of the CS flux, but the quantum statistics of the particles are changed unless $\alpha$ is an even integer.
For the two particle system, the vector potential associated with the CS flux $\bi{a}_{\bi{r}_2}(\bi{r_1})$ depends only on the relative coordinate $\bi{r}=\bi{r}_1-\bi{r}_2$. When $\bi{a}(\bi{r})$ is added to $\bi{A}(\bi{r})$, the vector potential of the applied magnetic field, the Schrödinger equation has a solution $$\tilde{\psi}_m=e^{-i\alpha\phi}\psi_m,$$ where $\psi_m$ is the solution with $\alpha=0$ (i.e. in the absence of CS flux). If $\alpha$ is an odd integer, Boson and Fermion statistics are interchanged; if $\alpha$ is even, no change in statistics occurs and electrons are transformed into composite Fermions with an identical energy spectrum.
The Hamiltonian for the composite particle system (charged particles with attached flux tubes) is much more complicated than the original system with $\alpha=0$. What is gained by making the CS transformation? The answer is that one can use the “mean field” approximation in which $\bi{A}(\bi{r})+\bi{a}(\bi{r})$, the vector potential of the external plus CS magnetic fields, is replaced by $\bi{A}(\bi{r})
+\left<\bi{a}(\bi{r})\right>$, where $\left<\bi{a}(\bi{r})\right>$ is the mean field value of $\bi{a}(\bi{r})$ obtained by simply replacing $\varrho(\bi{r}')=\psi^\dagger(\bi{r}')\psi(\bi{r}')$ by its average value $\varrho_0$ in equation (\[eq9\]). A mean field energy spectrum can be constructed in which the massive degeneracy of the original partially filled electron Landau level disappears. One might then hope to treat both the Coulomb interaction and the CS gauge field interactions among the fluctuations (beyond the mean field) by standard many body perturbation techniques (e.g. by the random phase approximation, RPA). Unfortunately, there is no small parameter for a many body perturbation expansion unless $\alpha$, the number of CS flux quanta attached to each particle, is small compared to unity. However, a Landau–Silin [@silin] type Fermi liquid approach can take account of the short range correlations phenomenologically. A number of excellent papers on anyon superconductivity [@laughlin3] treat CS gauge interactions by standard many body techniques. Halperin and collaborators [@halperin2] have treated the half filled Landau level as a liquid of composite Fermions moving in zero effective magnetic field. Their RPA–Fermi-liquid approach gives a surprisingly satisfactory account of the properties of that state.
The vector potential associated with fluctuations beyond the mean field is given by $\delta\bi{a}(\bi{r})=\bi{a}(\bi{r})-\left<\bi{a}(\bi{r})
\right>$. The perturbation to the mean field Hamiltonian contains both linear and quadratic terms in $\delta\bi{a}(\bi{r})$, resulting in both two body – containing $\varrho(\bi{r}_1)\varrho(\bi{r}_2)$ – and three body – containing $\varrho(\bi{r}_1)\varrho(\bi{r}_2)
\varrho(\bi{r}_3)$ – interaction terms. The three body interaction terms are usually ignored, though for $\alpha$ of the order of unity this approximation is of questionable validity.
Jain’s Composite Fermion Picture {#sec5}
=================================
Jain noted that in the mean field approximation, an effective filling factor $\nu^*$ of the composite Fermions was related to the electron filling factor $\nu$ by the relation $$\label{eq10}
(\nu^*)^{-1}=\nu^{-1}-2p.$$ Remember that $\nu^{-1}$ is equal to the number of flux quanta of the applied magnetic field per electron, and $2p$ is the (even) number of CS flux quanta (oriented opposite to the applied magnetic field) attached to each electron in the CS transformation. Equation (\[eq10\]) implies that when $\nu^*=\pm1$, $\pm2$, … (negative values correspond to the effective magnetic field $\bi{B}^*$ seen by the CF’s oriented opposite to $\bi{B}$) and a non-degenerate mean field CF ground state occurs, then $\nu=\nu^*(1+2p\nu^*)^{-1}$. This Jain sequence of condensed states ($\nu=1/3$, $2/5$, $3/7$, …and $\nu=2/3$, $3/5$, … for $p=1$) is the set of FQH states most prominent in experiment. When $\nu^*$ is not an integer, QP’s of the neighboring Jain state will occur.
It is quite remarkable that the mean field CF picture predicts not only the Jain sequence of incompressible ground states, but the correct band of low energy states for any value of the applied magnetic field. This is very nicely illustrated for the case of $N$ electrons on a Haldane sphere. When the monopole strength seen by an electron has the value $2S$, the effective monopole strength seen by a CF is $2S^*=2S-2p(N-1)$. This equation reflects the fact that a given CF senses the vector potential produced by the CS flux on all other particles, but not its own CS flux. In table \[tab2\] the ten particle system is described for a number of values of $2S$ between 29 and 15.
[@llllllll]{} $2S$ & 29 & 28 & 27 & 26 & 25 & 21 & 15\
$2S^*$ & 11 & 10 & 9 & 8 & 7 & 3 & -3\
$N_{\rm QH}$ & 2 & 1 & 0 & 0 & 0 & 0 & 0\
$N_{\rm QE}$ & 0 & 0 & 0 & 1 & 2 & 6 & 6\
$l_0^*$ &11/2& 5 & 9/2& 4 & 7/2& 3/2& 3/2\
$\nu^*$ & & & 1 & & & 2 & -2\
$\nu$ & & & 1/3& & & 2/5& 2/3\
$L$ &0, 2, 4, 6, 8, 10& 5 & 0 & 5 & 0, 2, 4, 6, 8 & 0 & 0\
\[tab2\]
The Laughlin $\nu=1/3$ state occurs at $2S_3=3(N-1)=27$. For values of $2S$ different from this value, $2S-2S_3=\pm N_{\rm QP}$ (“$+$” corresponds to quasiholes, QH, and “$-$” to quasielectrons, QE). Let us apply the CF description to the ten electron spectra in figure \[fig1\]. At $2S=27$, we take $p=1$ and attach two CS flux quanta each electron. This gives $2S^*=9$ so that the ten CF’s completely fill the $2S^*+1$ states in the lowest angular momentum shell (lowest Landau level). There is a gap $\hbar\omega_c^*=\hbar eB^*/\mu c$ to the next shell, which is responsible for the incompressibility of the Laughlin state. Just as $|S|$ played the role of the angular momentum of the lowest shell of electrons, $l^*=|S^*|$ plays the role of the CF angular momentum and $2|S^*|+1$ is the degeneracy of the CF shell. Thus, the states with $2S=26$ and 28 contain a single quasielectron (QE) and quasihole (QH), respectively. For the QE state, $2S^*=8$ and the lowest shell of angular momentum $l_0^*=4$ can accommodate only nine CF’s. The tenth is the QE in the $l_1^*=l_0^*+1=5$ shell, giving the total angular momentum $L=5$. For the QH state, $2S^*=10$ and the lowest shell can accommodate eleven CF’s each with angular momentum $l_0^*=5$. The one empty state (QH) gives $L=l^*=5$. For $2S=25$ we obtain $2S^*=7$, and there are two QE’s each of angular momentum $l_1^*=9/2$ in the first excited CF shell. Adding the angular momenta of the two QE’s gives the band of multiplets $L=0$, 2, 4, 6, and 8. Similarly, for $2S=29$ we obtain $2S^*=11$, and there are two QH’s each with $l_0^*=11/2$, resulting in the allowed pair states at $L=0$, 2, 4, 6, 8, and 10. At $2S=21$, the lowest shell with $l_0^*=3/2$ can accommodate only four CF’s, but the other six CF’s exactly fill the excited $l_1^*=5/2$ shell. The resulting incompressible ground state is the Jain $\nu=2/5$ state, since $\nu^*=2$ for the two filled shells. A similar argument leads to $\nu^*=-2$ (minus sign means $\bi{B}^*$ oriented opposite to $\bi{B}$) and $\nu=2/3$ at $2S=15$. At $2S=30$, the addition of three QH angular momenta of $l_0^*=6$ gives the following band of low lying multiplets $L=1$, $3^2$, 4, $5^2$, $6^2$, $7^2$, 8, $9^2$, 10, 11, 12, 13, and 15. As demonstrated on an example in figure \[fig1\], this simple mean field CF picture correctly predicts the band of low energy multiplets for any number of electrons $N$ and for any value of $2S$.
Energy Scales and the Electron Pseudopotentials {#sec6}
================================================
The mean field composite Fermion picture is remarkably successful in predicting the low energy multiplets in the spectrum of $N$ electrons on a Haldane sphere. It was suggested originally that this success resulted from the cancellation of the Coulomb and Chern–Simons gauge interactions among fluctuations beyond the mean field. In figure \[fig2\], we show the lowest bands of multiplets for eight non-interacting electrons and for the same number of non-interacting mean field CF’s at $2S=21$.
------------------------------------------------------------------------
=3.75in
The energy scale associated with the CS gauge interactions which convert the electron system in frame (a) to the CF system in frame (b) is $\hbar\omega_c^*\propto B$. The energy scale associated with the electron-electron Coulomb interaction is $e^2/\lambda\propto\sqrt{B}$. The Coulomb interaction lifts the degeneracy of the non-interacting electron bands in frame (a). However, for very large value of $B$ the Coulomb energy can be made arbitrarily small compared to the CS energy (as marked with a shaded rectangle in figure \[fig2\]), i.e. to the separation between the CF Landau levels. The energy separations in the mean field CF model are completely wrong, but the structure of the low lying states (i.e., which angular momentum multiplets form the low lying bands) is very similar to that of the fully interacting electron system and completely different from that of the non-interacting electron system.
Two Fermion Problem
--------------------
An intuitive picture of why this occurs can be obtained by considering the two Fermion problem. The relative (REL) motion of a pair of electrons $(ij)$ is described by a coordinate $z_{ij}=z_i-z_j=r_{ij}e^{-i\phi_{ij}}$, and for the lowest Landau level its wavefunction contains a factor $z_{ij}^m$, where $m=1$, 3, 5, …. If every pair of particles has identical behavior, the many particle wavefunction must contain a similar factor for each pair giving a total factor $\prod_{i<j}z_{ij}^m$. As we have seen, the highest power of $z_i$ in this product is $m(N-1)$. If $m(N-1)$ is equal to $N_\phi-1=2S$, the maximum value of the $z$-component of the single particle angular momentum, the Laughlin $\nu=m^{-1}$ wavefunction results. For electrons, the $m$th cyclotron orbit, whose radius is $r_m$, encloses a flux $m\phi_0$ (i.e. $\pi r_m^2B=m\phi_0$). For a Laughlin $\nu=m^{-1}$ state the pair function must have a radius $r_m=r_1\sqrt{m}$. Let us describe the CF orbits by radius $\varrho_{\tilde{m}}$ and require that the $\tilde{m}$th orbit enclose $\tilde{m}$ flux quanta. It is apparent that if a flux tube carrying two flux quanta (oriented opposite to the applied magnetic field $B$) is attached to each electron in the CS transformation of the $\nu=1/3$ state, the smallest orbit of radius $\varrho_{\tilde{m}=1}$ has exactly the same size as $r_{m=3}$. Both orbits enclose three flux quanta of the applied field, but the CF orbit also encloses the two oppositely oriented CS flux quanta attached to the electrons to form the CF’s. In the absence of electron–electron interactions, the energies of these orbits are unchanged, since they still belong to the degenerate single particle states of the lowest Landau level.
In the mean field approximation the CS fluxes are replaced by a spatially uniform magnetic field, leading to an effective field $B^*=B/m$. The orbits for the CF pair states in the mean field approximation are exactly the same as those of the exact CS Hamiltonian. The smallest orbit has radius $\varrho_{\tilde{m}=1}$ equivalent to the electron orbit $r_{m=3}$. However, in the mean field approximation, the energies are changed (because $\omega_c^*=eB^*/\mu c$ replaces $\omega_c$). This energy change leads to completely incorrect mean field CF energies, but the mean field CF orbitals give the correct structure to the low lying set of multiplets.
In the presence of a repulsive interaction, the low lying energy states will have the largest possible value of $m$. For a monopole strength $2S=m(N-1)$, where $m$ is an odd integer, every pair can have radius $r_m$ and avoid the large repulsion associated with $r_1$, $r_3$, …, $r_{m-2}$. These ideas can be made somewhat more rigorous by using methods of atomic and nuclear physics for studying angular momentum shells of interacting Fermions.
Two Body Interaction Pseudopotential
-------------------------------------
As first suggested by Haldane [@haldane1], the behavior of the interacting many electron system depends entirely on the behavior of the two body interaction pseudopotential, which is defined as the interaction energy $V$ of a pair of electrons as a function of their pair angular momentum. In the spherical geometry, in order to allow for meaningful comparison of the pseudopotentials obtained for different values of $2S$ (and thus different single electron angular momenta $l$), it is convenient to use the “relative” angular momentum ${\cal R}=2l-L_{12}$ rather than $L_{12}$ (the length of $\hat\bi{L}_{12}=\hat\bi{l}_1+\hat\bi{l}_2$). The pair states with a given ${\cal R}=m$ (an odd integer) obtained on a sphere for different $2S$ are equivalent and correspond to the pair state on a plane with the relative (REL) motion described by angular momentum $m$ and radius $r_m$. The pair state with the smallest allowed orbit (and largest repulsion) has ${\cal R}=1$ on a sphere or $m=1$ on a plane, and larger ${\cal R}$ and $m$ means larger average separation. In the limit of $\lambda/R\rightarrow0$ (i.e., either $2S\rightarrow
\infty$ or $R\rightarrow\infty$), the pair wavefunctions and energies calculated on a sphere for ${\cal R}=m$ converge to the planar ones ($\psi_{0m}$ and its energy).
The pseudopotentials $V({\cal R})$ are plotted in figure \[fig3\] for a number of values of the monopole strength $2S$.
------------------------------------------------------------------------
=3.75in
The open circles mark the pseudopotential calculated on a plane (${\cal R}=m$). At small ${\cal R}$ the pseudopotentials rise very quickly with decreasing ${\cal R}$ (i.e. separation). More importantly, they increase more quickly than linearly as a function of $L_{12}(L_{12}+1)$. The pseudopotentials with this property form a class of so-called “short range” repulsive pseudopotentials [@wojs1]. If the repulsive interaction has short range, the low energy many body states must, to the extent that it is possible, avoid pair states the smallest values of ${\cal R}$ (or $m$) and the maximum two body repulsion.
Fractional Grandparentage
--------------------------
It is well-known in atomic and nuclear physics that eigenfunction of an $N$ Fermion system of total angular momentum $L$ can be written as $$\label{eq11}
\left|l^N,L\alpha\right>=
\sum_{L_{12}}\sum_{L'\alpha'}G_{L\alpha,L'\alpha'}(L_{12})
\left|l^2,L_{12};l^{N-2},L'\alpha';L\right>.$$ Here, the totally antisymmetric state $\left|l^N,L\alpha\right>$ is expanded in the basis of states $\left|l^2,L_{12};l^{N-2},
L'\alpha';L\right>$ which are antisymmetric under permutation of particles 1 and 2 (which are in the pair eigenstate of angular momentum $L_{12}$) and under permutation of particles 3, 4, …, $N$ (which are in the $N-2$ particle eigenstate of angular momentum $L'$). The labels $\alpha$ (and $\alpha'$) distinguish independent states with the same angular momentum $L$ (and $L'$). The expansion coefficient $G_{L\alpha,L'\alpha'}(L_{12})$ is called the coefficient of fractional grandparentage (CFGP).
For a simple three Fermion system, equation (\[eq11\]) reduces to $$\left|l^3,L\alpha\right>=
\sum_{L_{12}}F_{L\alpha}(L_{12})\left|l^2,L_{12};l;L\right>,$$ and $F_{L\alpha}(L_{12})$ is called the coefficient of fractional parentage (CFP). In the lowest Landau level, the individual Fermion angular momentum $l$ is equal to $S$, half the monopole strength, and the number of independent multiplets of angular momentum $L$ that can be formed by addition of the angular momenta of three identical Fermions is given in table \[tab3\]
[@rrccccccccccccc]{} $2l$&2S=0&2&4&6&8&10&12&14&16&18&20&22&24&26\
2&1&&&&&&&&&&&&&\
4&&1&&1&&&&&&&&&&\
6&&&1&1&1&&1&&&&&&&\
8&&&&2&1&1&1&1&&1&&&&\
10&&&&&2&1&2&1 &1&1&1&&1&\
12&&&&&&2&2&2 &1&2&1&1&1&1\
14&&& &&& &3&2&2&2&2&1&2&1\
[@rrccccccccccccc]{} $2l$&2S=1&3&5&7&9&11&13&15&17&19&21&23&25&27\
3&&1&&&&&&&&&&&&\
5&&1&1&&1&&&&&&&&&\
7&&&1&1&1&1&&1&&&&&&\
9&&&&1&2&1&1&1&1&&1&&&\
11&&&&&2&2&1&2&1 &1&1&1&&1\
13&&&& &&2&2&2&2&1&2&1&1&1\
\[tab3\]
Low energy many body states must, to the extent it is possible, avoid parentage from pair states with the largest repulsion (pair states with maximum angular momenta $L_{ij}$ or minimum ${\cal R}$). In particular, we expect that the lowest energy multiplets will avoid parentage from the pair state with ${\cal R}=1$. If ${\cal R}=1$, i.e. $L_{12}=2l-1$, the smallest possible value of the total angular momentum $L$ of the three Fermion system is obtained by addition of vectors $\bi{L}_{12}$ (of length $2l-1$) and $\bi{l}_3$ (of length $l$), and it is equal to $|(2l-1)-l|=l-1$. Therefore, the three particle states with $L<l-1$ must not have parentage from ${\cal R}=1$. It is straightforward to show that if $L<l-(2p-1)$, where $p=1$, 2, 3, …, the three electron multiplet at $L$ has no fractional parentage from ${\cal R}\le2p-1$. The multiplets that must avoid one, two, or three smallest values of ${\cal R}$ are underlined with an appropriate number of lines in table \[tab3\] and listed in table \[tab4\]. This gives the results in table \[tab4\], the values of $2L$ that avoid ${\cal R}=1$, 3, and 5 for various values of $2l$.
[@rlll]{} $2l$ & $2L\;({\cal R}\ge3)$ & $2L\;({\cal R}\ge5)$ & $2L\;({\cal R}\ge7)$\
6 & 0 & &\
7 & 3 & &\
8 & 2 & &\
9 & 3, 5 & &\
10 & 0, 4, 6 & 0 &\
11 & 3, 5, 7 & 3 &\
12 & 2, $6^2$, 8 & 2 &\
13 & 3, 5, 7, $9^2$ & 3, 5 &\
14 & 0, 4, 6, $8^2$, 10 & 0, 4, 6 & 0\
\[tab4\]
The $L=0$ states that appear at $2l=6$ (${\cal R}\ge3$), $2l=10$ (${\cal R}\ge5$), and $2l=14$ (${\cal R}\ge7$) are the only states for these values of $2l$ that can avoid one, two, or three largest pseudopotential parameters, respectively, and therefore are the non-degenerate ($L=0$) ground states. They are the Laughlin $\nu=1/3$, 1/5, and 1/7 states.
If only a single multiplet belongs to an angular momentum subspace, its form is completely determined by the requirement that it is an eigenstate of angular momentum with a given eigenvalue $L$. The wavefunction and the type of many body correlations do not depend on the form of the interaction pseudopotential. For interactions that do not have short range, the state that avoids the largest two body repulsion (e.g. the $L=0$ multiplet at $2l=6$) might not have the lowest total three body interaction energy and be the ground state. If more than one multiplet belongs to a given angular momentum eigenvalue (e.g., two multiplets occur at $L=3$ for $2l=8$), the interparticle interaction must be diagonalized in this subspace (two-dimensional for $2l=8$ and $L=3$). Whether the lowest energy eigenstate in this subspace has Laughlin type correlations, i.e. avoids as much as possible largest two body repulsion, depends critically on the short range of the interaction pseudopotential. For the Coulomb interaction, we find that the Laughlin correlations occur and, whenever possible, the CFP of the lowest lying multiplets virtually vanishes (it would vanish exactly for an “ideal” short range pseudopotential which increases infinitely quickly with decreasing ${\cal R}$). For example, for the lower energy eigenstate at $L=3$ and $2l=8$, the CFP for ${\cal R}=1$ is less than $10^{-3}$. A similar thing occurs at $2S=9$ for $L=9/2$, at $2S=10$ for $L=4$ and 6, at $2S=11$ for $L=9/2$, 11/2, and 15/2, at $2S=12$ for $L=5$, 6, 7, and 9, at $2S=13$ for $L=11/2$, 13/2, 15/2, 17/2, and 21/2, and at $2S=14$ for $L=6^2$, 7, 8, 9, 10, and 12. At $2S=14$ for $L=6$ there are three allowed multiplets. The diagonalization of the Coulomb interaction gives the lowest state that avoids ${\cal R}=1$ (CFP $\sim10^{-7}$) and ${\cal R}=3$ (CFP $<10^{-2}$), and the next lowest state that avoids ${\cal R}=1$ (CFP $<10^{-5}$) but orthogonality to the lowest state requires that it has significant parentage from ${\cal R}=3$ (CFP $\approx0.34$).
One can see that the set of angular momentum multiplets $L$ that can be constructed at a given value of $2l$ without parentage from pair states with ${\cal R}=1$ is identical to the set of all allowed multiplets $L$ at $2l^*=2l-4$. For a short range repulsion (e.g. the Coulomb repulsion in the lowest Landau level), these multiplets will be (to a good approximation) the lowest energy eigenstates (the appropriate CFP for the actual eigenstates will be very small although not necessarily zero). More generally, in the lowest Landau level (remember that $l=S$), the set of multiplets $L$ that can be constructed at given $2S$ without parentage from ${\cal R}\le2p-1$ (i.e. with ${\cal R}\ge2p+1$ for all pairs; $p=1$, 2, …) is identical to the set of all allowed multiplets $L$ at $2S^*=2S-2p(N-1)$. The multiplets $L$ forming the lowest Coulomb energy band at a given $2S$ are all multiplets allowed at $2S^*$. But $2S^*=2S-2p(N-1)$ is just the effective magnetic monopole strength in the mean field CF picture! Thus the CF picture with $2p$ attached flux quanta simply picks the subset of angular momentum multiplets which have no parentage from pair states with ${\cal R}\le2p-1$, and neglects the long range part of the pseudopotential, $V({\cal R})$ for ${\cal R}\ge2p+1$.
Definition of the Short Range Pseudopotential
----------------------------------------------
For systems containing more than three Fermions in an angular momentum shell, the simple addition of angular momentum to determine the smallest possible $L$ that has parentage from pair states with $L_{12}=2l-1$ is of no help. Instead, we make use of the following operator identity $$\label{eq13}
\hat{L}^2 + N(N-2)\;\hat{l}^2 = \sum_{i<j}\hat{L}_{ij}^2.$$ Here $\hat{L}=\sum_i\hat{l}_i$ and $\hat{L}_{ij}=\hat{l}_i+\hat{l}_j$. The identity is easily proved by writing out the expression for $\hat{L}^2$ and for $\sum_{i<j}\hat{L}_{ij}^2$ and eliminating $\sum_{i<j}(\hat{l}_i\cdot\hat{l}_j)$ from the pair of equations. Taking matrix elements of equation (\[eq13\]) between states $\left|l^N,L\alpha\right>$ described by equation (\[eq11\]) gives $$\begin{aligned}
\label{eq14}
L(L+1)&+&N(N-2)\,l(l+1)
=\left<l^N,L\alpha\right|
\sum_{i<j}\hat{L}_{ij}^2
\left|l^N,L\alpha\right>
\nonumber\\
&=&{1\over2}N(N-1)\sum_{L_{12}}{\cal G}_{L\alpha}(L_{12})\;
L_{12}(L_{12}+1),\end{aligned}$$ where $${\cal G}_{L\alpha}(L_{12})=
\sum_{L'\alpha'}\left|G_{L\alpha,L'\alpha'}(L_{12})\right|^2.$$ The coefficients of grandparentage satisfy the relation $$\sum_{L_{12}} \sum_{L'\alpha'}
G_{L\alpha,L'\alpha'}(L_{12})\; G_{L\beta,L'\alpha'}(L_{12})
=\delta_{\alpha\beta}.$$ Of course, the energy of the multiplet $\left|l^N,L\alpha\right>$ is given by $$E_\alpha(L)={1\over2}N(N-1)
\sum_{L_{12}} {\cal G}_{L\alpha}(L_{12}) \, V(L_{12}),$$ where $V(L_{12})$ is the electron pseudopotential.
It is important to make the following observations:
1. The expectation value of $\sum_{i<j}\hat{L}_{ij}^2$ in a many body state $\left|l^N,L\alpha\right>$ increases as $L(L+1)$, but it is totally independent of $\alpha$;
2. If the pseudopotential $V_H(L_{12})$ were a linear function of $\hat{L}_{12}^2$ (we refer to $V_H$ as the “harmonic pseudopotential”), all many body multiplets with the same value of $L$ would be degenerate;
3. The difference $\Delta V(L_{12})=V(L_{12})-V_H(L_{12})$ between the actual pseudopotential $V$ and its harmonic part $V_H$ lifts this degeneracy. If $N_L$ many body multiplets of $V_H$ occur at angular momentum $L$, the anharmonic term $\Delta V$ in the pseudopotential causes them to “repel one another” and results in a band of $N_L$ non-degenerate multiplets.
Because the expectation value of $\sum_{i<j}\hat{L}_{ij}^2$ in a many body state of angular momentum $L$ increases as $L(L+1)$, a strict Hund’s rule holds for harmonic pseudopotentials: For $V_H$ that increases as a function of $L_{12}$, the highest energy state is always at the maximum possible value of $L$ equal to $L^{\rm MAX}=Nl-N(N-1)/2$, and the lowest energy state is at the minimum allowed value of $L$ equal to $L^{\rm MIN}$. If $V_H$ decreases as a function of $L_{12}$, the opposite occurs: the lowest energy state is at $L^{\rm MAX}$, and the highest energy state is at $L^{\rm MIN}$ (this is a standard Hund’s rule of atomic physics).
Neither of these Hund’s rules may remain true in the presence of a large anharmonic term $\Delta V$. For example, if the number of multiplets $N_L$ at a value slightly larger than $L^{\rm MIN}$ is very large compared to $N_{L^{\rm MIN}}$, the strong level repulsion due to $\Delta V$ within this $L$ subspace can overcome the difference in the expectation values of $V_H$, and the lowest eigenvalue of $V$ at $L$ can be lower than that at $L^{\rm MIN}$. However, only very few multiplets occur at large values of $L$: $N_{L^{\rm MAX}}=1$ (for $M=L=L^{\rm MAX}$, the only state is $\left|l,l-1,\dots,l-N+1\right>$), $N_{L^{\rm MAX}-1}=0$, $N_{L^{\rm MAX}-2}\le1$, $N_{L^{\rm MAX}-3}\le1$, etc. As a result, breaking of the Hund’s rule that refers to the behavior of energy at large $L$ requires stronger anharmonicity than at small $L$. For the Coulomb pseudopotential in the lowest Landau level we always find that the highest energy indeed occurs at $L^{\rm MAX}$. However, the ability to avoid parentage from pair states having large $L_{ij}$ often favors many body states at small $L>L^{\rm MIN}$ with large $N_L$, as prescribed by the CF picture.
The anharmonicity of the Coulomb pseudopotential in the lowest Landau level (which increases with increasing $L_{12}$) is critical for the behavior of the FQH systems. We have found that the condition for the occurrence of subbands separated by gaps in the energy spectrum, and, in particular, for the occurrence of non-degenerate incompressible fluid ground states at specific values of the filling factor, is that the anharmonic term $\Delta V(L_{12})$ is positive and increases with increasing $L_{12}$. In other words, the total pseudopotential $V(L_{12})$ must increase more quickly than linearly as a function of $L_{12}(L_{12}+1)$.
Hidden Symmetry of the Short Range Repulsion
---------------------------------------------
From our numerical studies we have arrived at the following conjectures:
1. The Hilbert space ${\cal H}_{Nl}$ of $N$ identical Fermions each with angular momentum $l$ contains subspaces ${\cal H}_{Nl}^{(p)}$ of states that have no parentage from ${\cal R}\le2p-1$. The subspaces $\tilde{\cal H}_{Nl}^{(p)}={\cal H}_{Nl}^{(p)}\setminus
{\cal H}_{Nl}^{(p+1)}$ can be defined; they hold states without parentage from ${\cal R}\le2p-1$, but with some parentage from ${\cal R}=2p+1$. Then $${\cal H}_{Nl}=\tilde{\cal H}_{Nl}^{(0)}\oplus
\tilde{\cal H}_{Nl}^{(1)}\oplus
\tilde{\cal H}_{Nl}^{(2)}\oplus\dots.$$
2. For an “ideal” short range repulsive pseudopotential $V_{\rm SR}$, for which $V_{\rm SR}({\cal R})\gg V_{\rm SR}({\cal R}+2)$, the huge difference between energy scales associated with different pair states results in the following (dynamical) symmetry: (i) subspaces $\tilde{\cal H}_{Nl}^{(p)}$ are the interaction eigensubspaces, (ii) $p$ is a good quantum number, (iii) energy spectrum splits into bands (larger $p$ corresponds to lower energy), and (iv) energy gap above the $p$th band scales as $V(2p-2)-V(2p)$.
3. For a finite short range pseudopotential $V$ (increasing more quickly than $V_H$ as a function of $L_{12}$), the above symmetry is only approximate, but the correlation between energy and parentage from highly repulsive pair states persists, and so do the gaps in the energy spectrum. The mixing between neighboring subbands is weak, although the structure of energy levels within each subband depends on the form of $V(L_{12})$ at ${\cal R}\ge2p+1$.
4. The set of angular momentum multiplets in subspace ${\cal H}_{Nl}^{(p)}$ is identical to ${\cal H}_{Nl^*}$, where $l^*=l-p(N-1)$.
Although at present we do not have a general analytic proof for the last conjecture, we have verified it for various small systems and have not found one for which it would fail.
The above conjectures can be immediately translated into the planar geometry. The harmonic pseudopotential $V_H(m)$, used to define the class of short range pseudopotentials, is that of a repulsive interaction potential $V(r)$ which is linear in $r^2$. Then, $${\cal H}_{\nu}=\tilde{\cal H}_{\nu}^{(0)}\oplus
\tilde{\cal H}_{\nu}^{(1)}\oplus
\tilde{\cal H}_{\nu}^{(2)}\oplus\dots,$$ where ${\cal H}_{\nu}$ is the Hilbert space of electrons filling a fraction $\nu$ of an infinitely degenerate Landau level, and subspaces $\tilde{\cal H}_{\nu}^{(p)}$ contain states without parentage from $m\le2p-1$, but with some parentage from $m=2p+1$. The (approximate) dynamical symmetry holds for the Coulomb interaction, and the low energy band ${\cal H}_{\nu}^{(p)}$ contains the same angular momentum multiplets as ${\cal H}_{\nu^*}$, with $\nu^*$ defined by the CF prescription in equation (\[eq10\]).
The validity of our conjectures for systems interacting through the Coulomb pseudopotential is illustrated in figure \[fig4\] for four electrons in the lowest Landau level at $2S=5$, 11, 17, and 23.
------------------------------------------------------------------------
=3.75in
Different symbols mark bands corresponding to (approximate) subspaces ${\cal H}_{Nl}^{(p)}$ with different $p$. The same sets of multiplets reoccur for different $2S$ in bands related by ${\cal H}_{Nl}^{(p)}\sim{\cal H}_{Nl^*}$.
Comparison with Atomic Shells: Hund’s Rule
-------------------------------------------
Our conjectures (verified by the numerical experiments) are based on the behavior of systems of interacting Fermions partially filling a shell of degenerate single particle states of angular momentum $l$. This is a central problem in atomic physics and in nuclear shell model studies of energy spectra. It is interesting to compare the behavior of the spherical harmonics of atomic physics with that of the monopole harmonics considered here. For monopole harmonics $l=S+n$, where $S$ is half of the monopole strength (and can be integral or half integral) and $n$ is a non-negative integer. For the lowest angular momentum shell $l=S$. For spherical harmonics $S=0$ and $l=n$. If in each case electrons are confined to a 2D spherical surface of radius $R$, one can evaluate the pair interaction energy $V$ as a function of the pair angular momentum $L_{12}$. The resulting pseudopotentials, $V({\cal R})$ for the FQH system in the lowest Landau level, and $V(L_{12})$ for atomic shells in a zero magnetic field, are shown in figure \[fig5\] for a few small values of $l$.
------------------------------------------------------------------------
=3.75in
In obtaining these results we have restricted ourselves to spin-polarized shells, so only orbital angular momentum is considered. It is clear that in the case of spherical harmonics the largest pseudopotential coefficient occurs for the lowest pair angular momentum, exactly the opposite of what occurs for monopole harmonics. As a consequence of equation (\[eq13\]), which relates the total angular momentum $L$ to the average pair angular momentum $L_{12}$, the standard atomic Hund’s rule predicts that the energy of a few electron system in an atomic shell will, on the average, decrease as a function of total angular momentum, which is opposite to the behavior of energy of electrons in the lowest Landau level. The difference between the energy spectra of electrons interacting through atomic and FQH pseudopotentials of figure \[fig5\] is demonstrated in figure \[fig6\], where we plot the result for four electrons in shells of angular momentum $l=3$ and 5.
------------------------------------------------------------------------
=3.75in
The solid circles correspond to monopole harmonics and the open ones to spherical harmonics. Note that at $L^{\rm MAX}$ the former give the highest energy and the latter the lowest. Due to anharmonicity of the pseudopotentials, the behavior of energy at low $L$ does not always follow a simple Hund’s rule for either FQH or atomic system. The FQH ground state for $l=3$ occurs at $L=0$ (this is the $\nu=2/3$ incompressible state). However, for $l=5$, the lowest of the three states at $L=2$ has lower energy than the only state $L=0$. This ground state at $L=2$ contains one quasihole in the Laughlin $\nu=1/3$ state and it is the only four electron state at this filling in which electrons can avoid parentage from the ${\cal R}=1$ pair state. Exactly opposite happens for the atomic system at $l=5$, where the anharmonicity is able to push the highest of the three $L=2$ states above the high energy state at $L=0$.
Higher Landau Levels
---------------------
Thus far we have considered only the lowest angular momentum shell (lowest Landau level) with $l=S$ The interaction of a pair of electrons in the $n$th excited shell of angular momentum $l=S+n$ can easily be evaluated to obtain the pseudopotentials $V(L_{12})$ shown in figure \[fig7\].
------------------------------------------------------------------------
=3.75in
Here we compare $V_n(L_{12})$ as a function of $L_{12}(L_{12}+1)$ for $n=0$, 1, and 2. It can readily be observed that $V_{n=0}$ increases more quickly than $L_{12}(L_{12}+1)$ in entire range of $L_{12}$, while $V_{n=1}$ and $V_{n=2}$ do so only up to certain value of $L_{12}$ (i.e., above certain value of ${\cal R}=2l-L_{12}$) For $n=1$, the $V_{n=1}$ has short range for ${\cal R}\ge3$ but is essentially linear in $L_{12}(L_{12}+1)$ from ${\cal R}=1$ to 5. For $n=2$, the $V_{n=2}$ has short range for ${\cal R}\ge5$ but is sublinear in $L_{12}(L_{12}+1)$ from ${\cal R}=1$ to 7. More generally, we find that the pseudopotential in the $n$th excited shell (Landau level) has short range for ${\cal R}\ge2n+1$.
Because the conclusions of the CF picture depend so critically on the short range of the pseudopotential, they are not expected to be valid for all fractional fillings of higher Landau levels. For example, the ground state at $\nu=2+1/3=7/3$ does not have Laughlin type correlations (i.e. electrons in the $n=1$ Landau level do not avoid parentage from ${\cal R}=1$) even if it is non-degenerate ($L=0$) and incompressible (as found experimentally [@willet]).
Fermi Liquid Model of Composite Fermions {#sec7}
=========================================
The numerical results of the type shown in figure \[fig1\] have been understood in a very simple way using Jain’s composite Fermion picture. For the ten particle system, the Laughlin $\nu=1/3$ incompressible ground state at $L=0$ occurs for $2S=3(N-1)=27$. The low lying excited states consist of a single QP pair, with the QE and QH having angular momenta $l_{\rm QE}=11/2$ and $l_{\rm QH}=9/2$. In the mean field CF picture, these states should form a degenerate band of states with angular momentum $L=1$, 2, …, 10. More generally, $l_{\rm QE}=(N+1)/2$ and $l_{\rm QH}=(N-1)/2$ for the Laughlin state of an $N$ electron system, and the maximum value of $L$ is $N$. The energy of this band would be $E=\hbar\omega_c^*=\hbar\omega_c/3$, the effective CF cyclotron energy needed to excite one CF from the (completely filled) lowest to the (completely empty) first excited CF Landau level. From the numerical results, two shortcomings of the mean field CF picture are apparent. First, due to the QE–QH interaction (neglected in the CF picture), the energy of the QE–QH band depends on $L$, and the “magnetoroton” QE–QH dispersion has a minimum at $L=5$. Second, the state at $L=1$ either does not appear, or is part of the continuum (in an infinite system) of higher energy states above the magnetoroton band.
At $2S=27-1=26$ and $2S=27+1=28$, the ground state contains a single quasiparticle (QE or QH, respectively), whose angular momenta $l_{\rm QE}=l_{\rm QH}=N/2=5$ result from the CS transformation which gives $2S^*=2S-2(N-1)=8$ for QE and 10 for QH (and $l_{\rm QE}=S^*+1$ and $l_{\rm QH}=S^*$). States containing two identical QP’s form lowest energy bands at $2S=25$ (two QE’s) and $2S=29$ (two QH’s). The allowed angular momenta of two identical CF QP’s (which are Fermions) each with angular momentum $l_{\rm QP}$ are $L=2l_{\rm QP}
-j$ where $j$ is an odd integer. Of course, $l_{\rm QP}$ depends on $2S$ in the CF picture, and at $2S=25$ we have $l_{\rm QE}=S^*+1=S-(N-1)+1=9/2$ yielding $L=0$, 2, 4, 6, and 8, while at $2S=29$ we have $l_{\rm QH}=S^*=S-(N-1)=11/2$ and $L=0$, 2, 4, 6, 8, 10. More generally, $l_{\rm QE}=(N-1)/2$ and $l_{\rm QH}=(N+1)/2$ in the 2QE and 2QH states of an $N$ electron system, and the maximum values of $L$ are $N-2$ for QE’s and $N$ for QH’s. As for the magnetoroton band at $2S=27$, the CF picture does not account for QP interactions and incorrectly predicts the degeneracy of the bands of 2QP states at $2S=25$ and 27.
The energy spectra of states containing more than one CF quasiparticle can be described in the following phenomenological Fermi liquid picture. The creation of an elementary excitation, QE or QH, in a Laughlin incompressible ground state requires a finite energy, $\varepsilon_{\rm QE}$ or $\varepsilon_{\rm QH}$, respectively. In a state containing more than one Laughlin quasiparticle, QE’s and/or QH’s interact with one another through appropriate QE–QE, QH–QH, and QE–QH pseudopotentials.
An estimate of the QP energies can be obtained by comparing the energy of a single QE (for the $N=10$ electron system, the energy of the ground state at $L=N/2=5$ for $2S=27-1=26$) or a single QH ($L=N/2=5$ ground state at $2S=27+1=28$) with the Laughlin $L=0$ ground state at $2S=27$. There can be finite size effects here, because the QP states occur at different values of $2S$ than the ground state, but using the correct magnetic length $\lambda=R/\sqrt{S}$ ($R$ is the radius of the sphere) in the unit of energy $e^2/\lambda$ at each value of $2S$, and extrapolating the results as a function of $N^{-1}$ to an infinite system should give reliable estimates of $\varepsilon_{\rm QE}$ and $\varepsilon_{\rm QH}$ for a macroscopic system.
The QP pseudopotentials $V_{\rm QP-QP}$ can be obtained by subtracting from the energies of the 2QP states obtained numerically at $2S=25$ (2QE), $2S=27$ (QE–QH), and $2S=29$ (2QH), the energy of the Laughlin ground state at $2S=27$ and two energies of appropriate non-interacting QP’s. As for the single QP, the energies calculated at different $2S$ must be taken in correct units of $e^2/\lambda=\sqrt{S}e^2/R$ to avoid finite size effects. This procedure was carried out in references [@sitko2; @wojs3].
In figure \[fig8\] we plot the QE–QE and QH–QH pseudopotentials for Laughlin $\nu=1/3$ and 1/5 states.
------------------------------------------------------------------------
=3.75in
As we have seen for two electrons (see figure \[fig3\]), the angular momentum $L_{12}$ of a pair of identical Fermions in an angular momentum shell (or a Landau level) is quantized, and the convenient quantum number to label the pair states is ${\cal R}=2l_{\rm QP}-L_{12}$ (on a sphere) or relative (REL) angular momentum $m$ (on a plane). When plotted as a function of ${\cal R}$, the pseudopotentials calculated for systems containing between six to eleven electrons (and thus for different QP angular momenta $l_{\rm QP}$) behave similarly and, for $N\rightarrow\infty$ (i.e., $2S\rightarrow\infty$), they seem to converge to the limiting pseudopotentials $V_{\rm QP-QP}({\cal R}=m)$ describing an infinite planar system.
In figure \[fig9\] we plot the QE–QH pseudopotentials for Laughlin $\nu=1/3$ and 1/5 states.
------------------------------------------------------------------------
=3.75in
As for a conduction electron and a valence hole pair in a semiconductor (an exciton), the motion of a QE–QH pair which does not carry a net electric charge is not quantized in a magnetic field. The appropriate quantum number to label the states is the continuous wavevector $k$ (or momentum), which on a sphere is given by $k=L/R=
L/\sqrt{S}\lambda$ (remember that $L$ is given in units of $\hbar$). When plotted as a function of $k$, the pseudopotentials calculated for systems containing between six to eleven electrons fall on the same curve that describes a continuous magnetoroton dispersion $V_{\rm QE-QH}(k)$ of an infinite planar system (the lines in figure \[fig9\] are only to guide the eye). Only the energies for $L\ge2$ are shown in figure \[fig9\], since the single QE–QH pair state at $L=1$ is either disallowed (hard core) or falls into the continuum of states above the magnetoroton band. The magnetoroton minima for the Laughlin $\nu=1/3$ and 1/5 states occur at about $k_0=1.4$ $\lambda^{-1}$ and $k_0=1.1$ $\lambda^{-1}$, respectively. The magnetoroton band at $\nu=1/3$ is well decoupled from the continuum of higher states because the band width $\sim0.05e^2/\lambda$ is much smaller than the energy gap $\varepsilon_{\rm QE}+\varepsilon_{\rm QH}
=0.1e^2/\lambda$ for additional QE–QH pair excitations. At $\nu=1/5$, the band width $\sim0.015e^2/\lambda$ is closer to the single particle gap $\varepsilon_{\rm QE}+\varepsilon_{\rm QH}=0.021
e^2/\lambda$ and the state of two magnetorotons each with $k\approx
k_0$ can couple to the highest energy QE–QH pair states at $k\le2k_0$.
Knowing the QP–QP pseudopotentials and the bare QP energies allows us to evaluate the energies of states containing three or more QP’s. Typical results are shown in figure \[fig10\].
------------------------------------------------------------------------
=3.75in
In frame (a) we show the energy spectra of three QE’s in the Laughlin $\nu=1/3$ state of eleven electrons. The spectrum in frame (b) gives energies of three QH’s in the nine electron system at the same filling. The exact numerical results obtained in diagonalization of the eleven and nine electron systems are represented by plus signs and the Fermi liquid picture results are marked by solid circles. The exact energies above the dashed lines correspond to higher energy states that contain additional QE–QH pairs. It should be noted that in the mean field CF picture which neglects the QP–QP interactions, all of the 3QP states would be degenerate and the energy gap separating the 3QP states from higher states would be equal to $\hbar\omega_c^*=\hbar\omega_c/3$. Although the fit in figure \[fig10\] is not perfect, it is quite good and justifies the use of the Fermi liquid picture to describe (compressible) states at $\nu\ne(2p+1)^{-1}$.
Composite Fermion Hierarchy {#sec8}
============================
The sequence of Laughlin–Jain states with filling factor $\nu$ given by $\nu=\nu^*(1+2p\nu^*)^{-1}$ where $p=1$, 2, …, and the CF filling factor $\nu^*$ is any non-zero integer, is the most prominent set of condensed states observed experimentally. However, this sequence (together with the conjugate “hole” states, $\nu\rightarrow1-\nu$) does not contain all odd denominator fractions the way the Haldane hierarchy scheme does. The question arises quite naturally of how to treat the CF values of $\nu^*$ which are not integers. The answer leads to the CF approach to the hierarchy of incompressible quantum fluid ground states [@sitko1].
Consider a state of $N_0$ electrons at a monopole strength $2S_0$ with a filling factor $\nu_0$. The CS transformation that attaches to each electron $2p_0$ flux quanta oriented opposite to the applied magnetic field results in the CF system at an effective filling factor $\nu_0^*$ given by $(\nu_0^*)^{-1}=
\nu_0^{-1}-2p_0$ and an effective monopole strength $2S_0^*=2S_0-2p_0
(N_0-1)$. The procedure for handling non-integral values of CF filling factor $\nu_0^*$ is to set it equal to $\nu_0^*=n_1+\nu_1$, where $n_1$ is an integer and $\nu_1$ is the fractional filling of the CF quasiparticle level (same sign as $n_1$ for QE’s and opposite for QH’s). Our problem is then that of placing $N_1$ quasiparticles into $2l_1+1$ available states of a CF shell (Landau level) of angular momentum $l_1$: the QE’s into the lowest empty shell with $l_1=|S_0^*|+n_1+1$, or the QH’s into the highest filled shell with $l_1=|S_0^*|+n_1$, We now ignore all completely filled and completely empty CF shells, and reapply the CS transformation by setting $S_1=l_1$ and attaching $2p_1$ flux quanta to each of the $N_1$ quasiparticles in the partially filled CF shell. This produces a new type of QP’s and a new QP filling factor $\nu_1^*$ given by $(\nu_1^*)^{-1}=\nu_1^{-1}-2p_1$. If $\nu_1^*$ is an integer, we obtain a daughter states in the hierarchy. If it is not, we write $\nu_1^*=n_2+\nu_2$, where $\nu_2$ represents the partial filling of the new QP shell, and repeat the mean field CF procedure. This leads to the set of equations: $$\label{eq17}
\nu_l^{-1}=2p_l+(n_{l+1}+\nu_{l+1})^{-1},$$ where $\nu_l$ is the QP filling factor and $2p_l$ is the number of flux quanta attached to each Fermion at the $l$th level of the CF hierarchy.
As an example, consider a system of $N_0=12$ electrons at $2S_0=30$. We apply the mean field CF approximation by attaching to each electron $2p_0=2$ flux quanta. This gives the effective CF monopole strength $2S_0^*=30-2(12-1)=8$. The lowest CF shell is filled with nine particles, and there are $N_1=3$ quasielectrons in the first excited ($n_1=1$) CF shell of angular momentum $l_1=5$. The filling factor at this level of hierarchy is $\nu_0^*=1+\nu_1$. We now reapply the CF transformation by attaching $2p_1=4$ flux quanta to each of $N_1=3$ QE’s at $2S_1=10$ and obtain $2S_1^*=10-4(3-1)=2$. The lowest CF shell of $l_1=1$ is now completely filled yielding $\nu_1^*=1$. Using equation (\[eq17\]) we obtain $\nu_1^{-1}=4+1^{-1}=5$ and $\nu_0^{-1}=2+(1+1/5)^{-1}=17/6$.
If the mean field CF picture worked on all levels of hierarchy, the twelve electron system at $2S=30$ should have an incompressible $L=0$ ground state corresponding to the filling factor $\nu=6/17$. In figure \[fig11\](a) we show the low energy sector of the spectrum calculated for this system using the Fermi liquid picture (only the lowest energy states containing 3QE’s in the Laughlin $\nu=1/3$ state are calculated).
------------------------------------------------------------------------
=3.75in
Indeed, the $\nu=6/17$ hierarchy ground state at $L=0$ is separated from higher states by a small gap in the twelve electron spectrum (although it is not clear that this small gap will persist in the thermodynamic limit [@wojs3]).
Though the CF hierarchy picture seems to work in some cases, there are others where it is clearly in complete disagreement with numerical results. For example, a CF transformation with $2p_0=2$ applied to an $N_0=8$ electron system at $2S_0=18$ gives $2S_0^*=18-2(8-1)=4$, $n_1=1$, and $N_1=3$ QE’s left in the shell with $l_1=3$. Adding the three QE angular momenta of $l_1=3$ gives a low energy band at $L=0$, 2, 3, 4, and 6. Reapplication of the CF transformation with $2p_1=2$ gives $2S_1^*=
6-2(3-1)=2$, i.e. the completely filled lowest shell, $\nu_1^*=1$ ($n_2=1$ and $\nu_2=0$). From equation (\[eq17\]) we get $\nu_1=1/3$ and $\nu_0=4/11$. In figure \[fig11\](b) we show the spectrum obtained by exact numerical diagonalization of an eight electron system at $2S=18$. It is apparent that the set of multiplets at $L=0$, 2, 3, 4, and 6 form the low energy band. However the reapplication of the mean field CF transformation to the three QE’s in the $l_1=3$ shell (which predicts an $L=0$ incompressible ground state corresponding to $\nu=4/11$) is definitely wrong.
The reason why the CF hierarchy picture does not always work is not difficult to understand. The electron (Coulomb) pseudopotential in the lowest Landau level $V_e({\cal R})$ satisfies the “short range” criterion (i.e. it increases more quickly with decreasing ${\cal R}$ than the harmonic pseudopotential $V_H$) in the entire range of ${\cal R}$, which is the reason for the incompressibility of the principal Laughlin $\nu=(2p+1)^{-1}$ states. However, this does not generally hold for the QP pseudopotentials on higher levels of the hierarchy. In figure \[fig8\] we plotted $V_{\rm QE-QE}({\cal R})$ and $V_{\rm
QH-QH}({\cal R})$ for the $\nu=1/3$ and $\nu=1/5$ Laughlin states of six to eleven electrons. Clearly, the QE and QH pseudopotentials are quite different and neither one decreases monotonically with increasing ${\cal R}$. On the other hand, the corresponding pseudopotentials in $\nu=1/3$ and 1/5 states look similar, only the energy scale is different. The convergence of energies at small ${\cal R}$ obtained for larger $N$ suggests that the maxima at ${\cal R}=3$ for QE’s and at ${\cal R}=1$ and 5 for QH’s, as well as the minima at ${\cal R}=1$ and 5 for QE’s and at ${\cal R}=3$ and 7 for QH’s, persist in the limit of large $N$ (i.e. for an infinite system on a plane). Consequently, the only incompressible daughter states of Laughlin $\nu=1/3$ and 1/5 states are those with $\nu_{\rm QE}=1$ or $\nu_{\rm
QH}=1/3$ and (maybe) $\nu_{\rm QE}=1/5$ and $\nu_{\rm QH}=1/7$. It is clear that no incompressible daughter states of the parent Laughlin $\nu=1/3$ state will form at e.g. $\nu=4/11$ ($\nu_{\rm QE}=
1/3$) or 4/13 ($\nu_{\rm QH}=1/5$), but that they will form (at least, in finite systems [@wojs3]) at $\nu=6/17$ ($\nu_{\rm QE}=1/5$) or 4/13 ($\nu_{\rm QH}=1/7$).
From the CF hierarchy scheme we find the Jain–Laughlin states when the CS transformation is applied directly to electrons (or to holes in a more than half-filled level). These states occur at integral values of $\nu^*$, the effective CF filling factor, and correspond to completely filling a QP shell. For example, the $\nu=2/5$ state occurs when $\nu^*=2$, and the CF’s in the first excited shell (which are Laughlin QE’s of the $\nu=1/3$ state) have $\nu_{\rm QP}=1$. The angular momenta of the two lowest CF shells are $l_0^*=|S^*|$ and $l_1^*=|S^*|+1$, so they contain $2l_0^*+1$ and $2l_1^*+1$ states, respectively. Since $\nu_{\rm QP}=1$, there are $N_{\rm QP}=2l_1^*+1$ CF quasiparticles. The total number of states filled by the $N$ Fermions is $(2l_0^*+1)
+(2l_1^*+1)=2N_{\rm QP}-2$, giving $N=2N_{\rm QP}-2$. For an infinite system this is just Haldane’s relation between the number of quasiparticles and the number of electrons, $N=2qN_{\rm
QP}$, for the integer $q=1$. This demonstrates that integrally filled CF shells correspond to $\nu_{\rm QP}=1$, a completely filled shell of Laughlin QP’s. Adding new Fermions to a system with $\nu_{\rm QP}=1$ requires creating a new type of QP’s, and the counting of available QP states turns out to be exactly the same in the CF hierarchy and Haldane’s Boson hierarchy picrures. Integral CF filling (i.e., $\nu_{\rm QP}=1$) gives a valid mean field picture independent of QP–QP interactions provided that the gap for creating new QP’s is positive. When $\nu^*$ is non-integral, the mean field picture is valid only at values of $L$ for which the “short range” requirement on the pseudopotential $V_{\rm QP-QP}(L)$ is satisfied. The form of the QP–QP interactions obtained from our numerical calculations makes it clear that the mean field approximation is not valid at certain quasiparticle fillings (e.g. for $\nu_{\rm QP}=1/3$ filling of the quasielectron levels of the electron $\nu=1/3$ state).
Systems Containing Electrons and Valence Band Holes {#sec9}
====================================================
There has been a great deal of interest in photoluminescence (PL) of 2D systems in high magnetic fields. An important ingredient in understanding PL is the negatively charged exciton ($X^-$). The $X^-$ consists of two electrons bound to a valence band hole. If the total spin of the pair of electrons, $J_e$, is zero, the $X^-$ is said to be a singlet ($X^-_s$); if $J_e=1$ the $X^-$ is called a triplet ($X^-_t$). Only the $X^-_s$ is bound in the absence of a magnetic field, but in infinite magnetic field (so that only a single Landau level is relevant) only the $X^-_t$ is bound in a 2D system. It often occurs that the photoexcited hole is separated from the plane of the electron system by a small distance (this can happen, e.g., in wide GaAs quantum wells when the electron gas is confined to one GaAs/AlGaAs interface by remote ionized donors, and the photoexcited holes reside close to the other GaAs/AlGaAs interface). Several remarkable effects associated with electron–hole systems and charged excitons can be understood using the composite Fermion picture.
Charged Exciton and the Hidden Symmetry in the Lowest Landau Level
-------------------------------------------------------------------
First let us consider the idealized 2D system at so large a magnetic field that only the lowest electron and hole Landau levels need be considered. The energy spectrum for a two-electron–one-hole system at $2S=20$ is shown in figure \[fig12\].
------------------------------------------------------------------------
=3.75in
The triplet $X^-$ with angular momentum $l_{X^-}=S-1$ is the only bound state, with binding energy $\sim0.05e^2/\lambda$. A pair of (unbound) singlet and triplet states occur at the energy equal exactly to the exciton energy $E_X$. In these so-called “multiplicative” states a neutral exciton $X$ in its ground state is decoupled from the second electron. Addition of exciton and electron angular momenta $L_X=0$ and $l_e=S$ gives a state of total angular momentum $L=S$, and addition of two electron spins of $1/2$ gives both $J_e=0$ and 1 spin configurations.
The occurrence of unbound states at $E=E_X$ and $L=S$ is a manifestation of the following “hidden symmetry:” Because of the exact overlap of electron and hole orbitals in the lowest Landau level (scaled with the same magnetic length $\lambda$), and thus independence of the strength of interaction of the type of particles involved, the commutator of an operator $d_X^\dagger$ that creates an exciton in its $L_X=0$ ground state (on a sphere, $d_X^\dagger=\sum_m
(-1)^mc_m^\dagger h_m^\dagger$, where $c_m^\dagger$ and $h_m^\dagger$ are electron and hole creation operators), with the interaction Hamiltonian $H$ is $[H,d_X^\dagger]=E_Xd_X^\dagger$. As a result, if $\Psi$ is an eigenstate of $N_e$ electrons and $N_h$ holes with an eigenenergy $E$ and angular momentum quantum numbers $L$ and $M$, then the multiplicative state $d_X^\dagger\Psi$ of $N_e+1$ electrons and $N_h+1$ holes is also an eigenstate with energy $E+E_X$ and the same $L$ and $M$. A good quantum number conserved due to the “hidden symmetry” is the number of decoupled excitons, $N_X$. In particular, the ground state for $N_e=N_h=N$ is the totally multiplicative state $(d_X^\dagger)^N\left|{\rm vac}\right>$ with $N_X=N$; for an infinite system this ground state can be viewed as a Bose condensate of non-interacting excitons. It can be readily found that the application of the PL operator that annihilates an optically active exciton ($d_X$) reduces its $N_X$ by one, and therefore that only the multiplicative electron–hole states with $N_X>0$ are optically active (have non-vanishing PL intensity). In figure \[fig12\], the two multiplicative states at $E=E_X$ and $L=S$ have $N_X=1$, and all others have $N_X=0$.
It is essential to realize that two independent symmetries forbid the recombination of a triplet $X^-$ ground state in figure \[fig12\]:
- Due to the 2D translational/rotational space invariance, the PL operator $d_X$ conserves two angular momentum quantum numbers. On a sphere, these are is $L$ and $M$, and the resulting optical selection rule allows only a state with $L=S$ to decay by $e$–$h$ recombination. On a plane, these are the total ($L_{\rm TOT}$) and center-of-mass ($L_{\rm CM}$) angular momenta and the radiative channel for an $X^-$ is that of $L_{\rm REL}\equiv L_{\rm TOT}-L_{\rm CM}=0$. This (geometrical) symmetry can be broken by collisions, but persists in systems with a finite quantum well width, finite electron and hole layer separation, or Landau level mixing.
- Due to the equal strength of $e$–$e$, $h$–$h$, and $e$–$h$ interactions, $N_X$ is a good quantum number. Since $N_X$ is decreased in a PL process, only the multiplicative ($N_X>0$) states are radiative. This (dynamical) symmetry is not broken by collisions, and requires breaking electron–hole orbital symmetry.
Since a number of independent factors are needed to allow for the recombination of a triplet $X^-$, this complex (in narrow and symmetric quantum wells and in high magnetic fields) is expected to be a well defined long-lived quasiparticle. The correlations, optical properties, etc. are expressed more easily in terms of this quasiparticle than in terms of individual electrons and holes. The finite angular momentum of an $X^-$ in spherical geometry (decoupling of the CM excitations from the REL motion on a plane) can be viewed as the formation of a degenerate Landau level of this (charged) quasiparticle. As will be shown later, the interaction of $X^-$ quasiparticles with one another and with electrons can be described using the ideas familiar in the context of FQH systems (Laughlin correlations, composite Fermions, parentage, etc.).
Interaction of Charged Excitons
--------------------------------
The simplest system in which to study $X^-$–$X^-$ interaction contains four electrons and two holes. Its energy spectrum at $2S=17$ is shown in figure \[fig13\].
------------------------------------------------------------------------
=3.75in
The low energy spectrum is characterized by four bands which we have identified as follows:
1. The lowest band taking on all even values between $L=0$ and 12 consists of a pair of charged excitons $X^-$ (each with angular momentum $l_{X^-}=S-1$);
2. The next band contains an electron with $l_e=S$ and a negatively charged biexciton $X_2^-$ (a bound state of an $X$ and an $X^-$) with angular momentum $l_{X_2^-}=S-2$; the allowed $L$ values go from $|l_e-l_{X_2^-}|
=2$ to $l_e+l_{X_2^-}-1=14$;
3. A band of multiplicative states containing an $X$, an $X^-$, and an electron; it begins at $L=|l_e-l_{X^-}|=1$ and goes to $L=l_e
+l_{X^-}-1=15$;
4. A band of multiplicative states containing two neutral excitons and two free electrons; it takes on all even values of $L$ between zero and $2l_e-1=16$.
One interesting feature of figure \[fig13\] is that it gives us the effective pseudopotential $V_{AB}(L)$ for the interaction of the pair of Fermions $AB$ (where $A$ and $B$ can be $e$, $X^-$, $X_2^-$, etc.) as a function of angular momentum. As for electrons, it is convenient to use the relative pair angular momentum ${\cal R}=l_A+l_B-L$. For identical Fermions with angular momentum $l$, the allowed values of $L$ are $2l-j$, where $j$ is an odd integer, i.e., ${\cal R}=1$, 3, 5, …, and ${\cal R}\le2l$. For distinguishable Fermions $A$ and $B$, all values of $L$ between $|l_A-l_B|$ and $|l_A+l_B|$ are expected, i.e., ${\cal R}=0$, 1, 2, …, and ${\cal R}\le2\min(l_A,l_B)$. However, our numerical results display a “hard core” repulsion for composite particles, and one or more of the pair states with the largest values of $L$ (smallest ${\cal R}$) are forbidden (i.e. the corresponding pseudopotential parameters are effectively infinite). For $A=X_n^-$ and $B=X_m^-$, the smallest allowed value of ${\cal R}$ is given by $${\cal R}^{\rm MIN}_{AB}=2\min(n,m)+1.$$ The identification of pair states $AB$ in figure \[fig13\] (as marked with lines) was possible by comparing the displayed $4e$–$2h$ spectrum with the pseudopotentials of point charge particles with appropriate angular momenta $l_A$ and $l_B$ and binding energies $\varepsilon_A$ and $\varepsilon_B$ [@wojs2]. The appropriate values of angular momenta $l_A$ and $l_B$, and of the binding energies $\varepsilon_A$ and $\varepsilon_B$ are obtained by diagonalizing smaller systems (e.g. the $2e$–$1h$ system in figure \[fig12\] for an $X^-$), and the point charge pseudopotentials are used to approximate the $AB$ interaction. The approximate $AB$ energies obtained in this way are rather close to the exact $4e$–$2h$ energies. This implies that, due to different energy scales, the internal dynamics of charged excitons is weakly coupled to their scattering off one another or off electrons, and allows for the interpretation of an electron–hole system in terms of well defined charged excitonic quasiparticles interacting with one another and with excess electrons through Coulomb like forces. Slight difference between the actual pseudopotentials in figure \[fig13\] and the pseudopotentials of point charge particles comes from the larger size of charged excitons and their (nearly frozen) internal degrees of freedom. The latter can be accounted for phenomenologically by attributing each type of composite particles a finite electric polarizability to describe their induced electric dipole moment in the presence of an electric field of other charged particles. Due to an increased charge isotropy, the polarization effects are expected to be greatly reduced in larger systems, and disappear completely in the fluid type states discussed in the following paragraphs.
Generalized Composite Fermion Picture for Charged Excitons
-----------------------------------------------------------
Suppose we have a system of different (distinguishable) charged Fermions ($A$, $B$, …). They can be distinguished either because they are different species (e.g., electrons and charged excitons) or because they are confined to different, spatially separated layers. If all particles in such system repel one another through short range pseudopotentials (as defined for the electron FQH systems), one can think of many body states with Laughlin-type correlations [@laughlin2; @halperin1] given by a generalized (compare equation (\[eq5\])) Laughlin–Jastrow factor $$\label{eqcorrel}
\prod(z_i^{(A)}-z_j^{(B)})^{m_{AB}},$$ where $z_i^{(A)}$ is the complex coordinate for the position of the $i$th Fermion of type $A$, and the product is over all pairs. The restrictions on the integers $m_{AB}$ are that $m_{AA}$ must be odd, $m_{BA}=m_{AB}$, and $m_{AB}$ must not be smaller than certain minimum values ${\cal R}_{AB}^{\rm MIN}$ to avoid the infinite hard cores for all pairs. In a state with correlations given by equation (\[eqcorrel\]), a number of pair states with largest repulsion are avoided for each pair, ${\cal R}_{AB}\ge m_{AB}$. This is equivalent to saying that the high energy collisions (in which any pair of particles would come very close to one another) are forbidden in such state. This intuitive property of the Laughlin fluid states will be very useful in the discussion of collision assisted $X^-$ recombination.
A generalized CF picture can be constructed for a system with Laughlin correlations. In this picture, fictitious flux tubes carrying an integral number of flux quanta $\phi_0$ are attached to each particle. In the multi-component system, each particle of type $A$ carries flux $(m_{AA}-1)\phi_0$ that couples only to charges on all other particles of the same type $A$, and fluxes $m_{AB}\phi_0$ that couple to charges on all particles of other types $B$ ($A$ and $B$ are any of the types of Fermions). On a sphere, the effective monopole strength seen by a CF of type $A$ (CF-$A$) is $$2S_A^*=2S-\sum_b(m_{AB}-\delta_{AB})(N_B-\delta_{AB}).
\label{eq2}$$ For different multi-component systems we expect generalized Laughlin incompressible states (for two components denoted as $[m_{AA},m_{BB},
m_{AB}]$) when all the hard core pseudopotentials are avoided and CF’s of each kind fill completely an integral number of their CF shells (e.g. $N_A=2l_A^*+1$ for the lowest shell). In other cases, the low lying multiplets are expected to contain different kinds of CF quasiparticles (generalized QE’s or QH’s), QP-$A$, QP-$B$, …, in the neighboring incompressible state. It is interesting to realize that the effective monopole strengths $2S^*_A$, i.e. the effective magnetic fields $B^*_A$ seen by particles of different type are not generally equal. One can think of effective CS charges and fluxes of different colors, but the resulting number of different effective CF magnetic fields of different color can no longer be regarded as physical reality, and no cancellation between gauge and Coulomb interactions is possible.
The multi-component (multi-color) CF picture can be applied to electrons and charged excitons in an electron–hole system. We have checked that the pseudopotentials describing interaction of identical composite particles in figure \[fig13\] all satisfy the short range criterion in the entire range of ${\cal R}$. For a pair of different particles, the pseudopotential may increase sufficiently quickly for some values of ${\cal R}$ but not the others and, for example, for $e^-$ and $X^-$ only the correlations described by odd exponents $m_{e^-X^-}$ are expected to occur. As an example, let us consider the $12e$–$6h$ system. In figure \[fig14\] we present its low energy spectrum at $2S=17$, calculated by diagonalizing systems of different combinations of electrons and composite particles interacting through effective pseudopotentials determined in figure \[fig13\].
------------------------------------------------------------------------
=3.75in
The following combinations (groupings of $12e$ and $6h$ into bound complexes) have the highest total binding energy and thus form the lowest energy bands in the $12e$–$6h$ spectrum: (i) $6X^-$, (ii) $e^-$–$5X^-$, (iii) $e^-$–$4X^-$–$X_2^-$, (iv) $2e^-$–$2X^-$–$2X_2^-$, (v) $2e^-$–$3X^-$–$X_3^-$, (vi) $2e^-$–$3X^-$–$X_2^-$, (vii) $2e^-$–$4X^-$. Groupings (ii), (vi), and (vii) also contain neutral excitons that however do not interact with charged particles due to the hidden symmetry. For each of these groupings, the CF transformation can be applied to determine correlations and identify number and type of quasiparticles that occur in the lowest energy states. For example, for groupings (i)–(iii) the generalized CF picture makes the following predictions.
[(i)]{} For $m_{X^-X^-}=3$ we obtain the Laughlin $\nu=1/3$ state with total angular momentum $L=0$. Because of the hard core of $V_{X^-X^-}$, this is the only state of this grouping.
[(ii)]{} We set $m_{X^-X^-}=3$ and $m_{e^-X^-}=1$, 2, and 3. For $m_{e^-X^-}=1$ we obtain $L=1$, 2, $3^2$, $4^2$, $5^3$, $6^3$, $7^3$, $8^2$, $9^2$, 10, and 11. For $m_{e^-X^-}=2$ we obtain $L=1$, 2, 3, 4, 5, and 6. For $m_{e^-X^-}=3$ we obtain $L=1$.
[(iii)]{} We set $m_{X^-X^-}=3$, $m_{e^-X_2^-}=1$, $m_{X^-X_2^-}=3$, and $m_{e^-X^-}=1$, 2, or 3. For $m_{e^-X^-}=1$ we obtain $L=2$, 3, $4^2$, $5^2$, $6^3$, $7^2$, $8^2$, 9, and 10. For $m_{e^-X^-}=2$ we obtain $L=2$, 3, 4, 5, and 6. For $m_{e^-X^-}=3$ we obtain $L=2$.
In groupings (ii) and (iii), the sets of multiplets obtained for higher values of $m_{e^-X^-}$ are subsets of the sets obtained for lower values, and we would expect them to form lower energy bands since they avoid additional small values of ${\cal R}_{e^-X^-}$. However, note that the (ii) and (iii) states predicted for $m_{e^-X^-}=3$ (at $L=1$ and 2, respectively) do not form separate bands in figure \[fig14\]. This is because $V_{e^-X^-}$ increases more slowly than linearly as a function of $L(L+1)$ in the vicinity of ${\cal R}_{e^-X^-}=3$ (see figure \[fig13\]). In such case the CF picture fails [@wojs1; @wojs3].
Our conclusion is that different kinds of long-lived Fermions (electrons and different charged excitonic complexes) formed in an electron–hole plasma in high magnetic fields can exhibit generalized incompressible FQH ground states with Laughlin-type correlations, and that these states can be described using a generalized CF model.
Spatially Separated Electron–Hole System
-----------------------------------------
Even in very high magnetic fields (in the lowest Landau level), an asymmetry between $e$–$e$, $h$–$h$, and $e$–$h$ interactions can be introduced by spatially separating 2D electron and hole layers. Such separation, which occurs for example in asymmetrically doped wide quantum wells, breaks the hidden symmetry and allows for a rich photoluminescence (PL) spectrum, which (unlike that for a co-planar system) can be therefore used as a probe of the low lying electron–hole states.
Let us consider an ideal system, in which electrons and holes occupy 2D parallel planes separated by a distance $d$. The interaction potentials are $V_{ee}(r)=V_{hh}(r)=1/r$ and $V_{eh}(r)=-1/\sqrt{r^2+d^2}$. The energy spectrum of a seven-electron–one-hole system is shown in figure \[fig15\] for $2S=15$ and values of $d$ going from 0 to 5 (measured in units of the magnetic length $\lambda$).
------------------------------------------------------------------------
=3.75in
For $d=5\lambda$, the $e$–$h$ interaction is weak and, as a first approximation, we can say that that the lowest band of states will consist of the lowest CF band of the electron system plus the (constant) hole energy. The allowed angular momenta will be given by $\bi{L}_e$, the angular momenta of the low lying electron states, added to the hole angular momentum $\bi{l}_h$ of length $l_h=S=15/2$. At $2S=15$, the CF picture for the electrons gives $2S^*=2S-2p(N-1)
=15-2(7-1)=3$. The seven electrons fill the $l_0^*=3/2$ shell plus three of the QE states in the shell $l_{\rm QE}=5/2$. The resulting electron angular momenta are $L_e=3/2$, 5/2, and 9/2. This gives three bands of low lying states, with total angular momenta $6\le L\le 9$, $5\le L\le 10$, and $3\le L\le 12$, respectively. These three bands can be clearly distinguished at $d=5\lambda$ and the states within each band become nearly degenerate at $d\sim10\lambda$.
For $d=0$, it is more useful to consider bound excitonic complexes ($X$ and $X^-$) and Laughlin quasiparticles of the $e^-$–$X^-$ fluid. First consider the multiplicative state with a single $X$ and six electrons. At $2S=15$ six electrons have the Laughlin $\nu=1/3$ ground state since $2S^*=15-2(6-1)=5$ gives a CF shell which accommodates all six CF’s. This is the lowest state at $L=0$, marked with a circle in frame (a). For a charge configuration containing one $X^-$ and five electrons, we can use the generalized CF model with $m_{e^-e^-}=m_{e^-X^-}=2$. This gives $2S_e^*=2S-m_{e^-e^-}(N_e-1)-m_{e^-X^-}=5$ and $2S_{X^-}^*
=2S-m_{e^-X^-}N_e=5$, and the angular momenta $l_e^*=S_e^*=5/2$ and $l_{X^-}^*=S_{X^-}^*-1=3/2$. There is one empty state in the lowest CF-$e^-$ shell giving $L_e=5/2$, and the CF-$X^-$ has $L_{X^-}=3/2$. Adding these two angular momenta gives $L=1$, 2, 3, and 4 as the lowest band of $5e^-$–$X^-$ states. The multiplicative state at $L=0$ (open circle) and the band of four multiplets containing an $X^-$ at $L=1$ to 4 (line) can clearly be seen at $d=0$ in frame (a). Although the hidden symmetry is only approximate at $d>0$, these bands can be easily identified at $d=0.5\lambda$ in frame (b).
At an intermediate separation of $d=1.75\lambda$ in frame (c), neither description used for $d<\lambda$ or $d\gg\lambda$ is valid, and it seems that a low energy band occurs at $L=0$, 1, 2, $3^2$, 4, 5, and 6. Most likely, the $X^-$ unbinds but the hole is still able to bind one electron, forming an exciton with a significant electric dipole moment. This dipole moment results in repulsion between the exciton the remaining six electrons, so that the correlations are quite different than at $d=0$, where the exciton decouples.
The PL spectrum can be evaluated from the eigenfunctions obtained in the numerical diagonalization of finite systems. For $d\gg0$, between one and three peaks are observed in the PL spectrum [@chen2]. Their separations are related to the Laughlin gap (for creation of a QE–QH pair) and to the energy of interaction between the valence band hole and the electron system.
Charged Excitons at a Finite Magnetic Field
--------------------------------------------
One final point is worth mentioning. The numerical calculations described so far were performed for an idealized model in which electrons and holes were confined to infinitely thin $2D$ layers, and only the lowest Landau level was considered. For realistic systems, effects due to spin, finite width of the quantum well, and Landau level mixing are very important. The energy spectra of the simple $2e$–$1h$ system calculated at $2S=20$ for parameters appropriate to a 11.5 nm GaAs/AlGaAs quantum well are shown in figure \[fig16\].
------------------------------------------------------------------------
=3.75in
Two frames correspond to the magnetic field of $B=13$ T and 68 T. We used five electron and hole Landau levels ($n\le4$) in the calculation, with the realistic magnetic field dependence of the hole cyclotron mass and the appropriate Zeeman splittings. The interaction matrix elements included finite (and different) effective widths of electron and hole quasi-2D layers.
There are a number of bound $X^-$ states in both frames, in contrast to only one singlet bound state at $B=0$ or only one triplet bound state predicted for an idealized system at infinite $B$. Three of these bound states are of particular importance. The $X^-_s$ and $X^-_{tb}$ ($b$ for “bright”), the lowest singlet and triplet states at $L=S$, are the only well bound radiative states, while $X^-_{td}$ ($d$ for “dark”) has by far the lowest energy of all non-radiative ($L\ne S$) states. The dark triplet state $X^-_{td}$ is the state discussed in the preceding sections; it is the only bound state in the lowest Landau level, but unbinds at low magnetic fields. The bright singlet state $X^-_s$ is the only bound state at $B=0$, but unbinds at very high fields due to the hidden symmetry. These states cross at $B\approx30$ T, as predicted in an earlier calculation [@whittaker]. The bright triplet state $X^-_{tb}$ has been discovered very recently [@wojs4]. It occurs only at intermediate fields and does not cross neither $X^-_s$ or $X^-_{td}$. It has larger PL intensity than the $X^-_s$ state.
Although an isolated $X^-_{td}$ is non-radiative because of the angular momentum selection rule, its collisions with other $X^-$’s or with electrons (which break the translational symmetry) could be expected to allow for $X^-_{td}$ recombination. However, the Laughlin correlations limit high energy collisions at low filling density ($\nu\sim1/5$ or less) and the PL intensity of a dark $X^-_{td}$ remains very low also in a presence of other particles [@wojs4]. In consequence, the $X^-_{td}$ is not seen in PL, and there is no contradiction between experiment [@hayne], which sees recombination of a triplet state at the energy above the singlet state up to 50 T, and theory [@whittaker], which predicts that the lowest triplet state crosses the singlet at roughly 30 T.
Summary {#sec10}
========
We have introduced the Jain CF mean field picture and shown how the low lying states can be understood by simple addition of angular momentum. The mean field CF picture gives the correct spectral structure not because of some cancellation between Chern–Simons and Coulomb interactions beyond the mean field, but because it selects a low angular momentum subset of the allowed multiplets that avoids the largest pair repulsion. The Laughlin correlations, which describe incompressible quantum fluid states, depend critically on the electron pseudopotential being of “short range” (by which we mean that $V(L_{12})$ increases more quickly than $L_{12}(L_{12}+1)$). The validity of Jain’s picture also depends upon $V(L_{12})$ being of short range. The pseudopotential describing quasiparticles of a Laughlin condensed state display short range behavior only at certain values of $L_{12}$. We have used this fact to explain why only certain states in the CF hierarchy give rise to incompressible states of the quasiparticle fluid (or daughter states in the hierarchy). The pseudopotentials $V_n(L_{12})$ for higher Landau levels ($n>0$) do not display short range behavior at all values of $L_{12}$, implying that Laughlin-like correlations will not necessarily result at $\nu'
=2p+\nu$, where $p$ is an integer and $\nu$ is a Laughlin–Jain filling factor. The CF ideas have been applied successfully to multicomponent plasmas containing different types of Fermions with the prediction of possible incompressible fluid states for these systems. Finally, the energy spectrum and PL of electron–hole systems can be interpreted in terms of CF’s and Laughlin correlations.
Acknowledgment {#acknowledgment .unnumbered}
===============
The authors gratefully acknowledge the support of Grant DE-FG02-97ER45657 from the Materials Science Program – Basic Energy Sciences of the US Department of Energy. They wish to thank P. Hawrylak, P. Sitko, I. Szlufarska, and K.-S. Yi for helpful discussions on different aspects of this work.
References {#references .unnumbered}
===========
[99]{}
See, for example, Proc. of Int. Conf. on Electronic Properties of Two-Dimensional Systems 1975–1999 published in [*Surface Sci.*]{}
von Klitzing K, Dorda G and Pepper M 1980 494
Tsui D C, Störmer H L and Gossard A C 1982 1559
Anderson P W 1958 1900
Laughlin R B 1981 B [**23**]{} 5632
Laughlin R B 1983 1395
Halperin B I 1984 1583
Haldane F D M 1983 605 Haldane F D M 1987 [*The Fractional Quantum Hall Effect*]{} (New York: Springer-Verlag) edited by Prange R E and Girvin S M chapt. 8
Jain J K 1989 199
Lopez A and Fradkin E 1991 B [**44**]{}, 5246
Halperin B I, Lee P A and Read N 1993 B [**47**]{}, 7312
Wójs A and Quinn J J 1998 493 Wójs A and Quinn J J 1999 45 Wójs A and Quinn J J 1999 [*Acta Phys. Pol.*]{} A [**96**]{} 593
Sitko P, Yi K S and Quinn J J 1997 B [**56**]{} 12 417
See, for example, Gasiorowicz S 1974 [*Quantum Physics*]{} (New York: John Wiley and Sons)
Chen X M and Quinn J J 1994 865
Sitko P, Yi S N, Yi K S, and Quinn J J 1996 3396
Wójs A, Hawrylak P and Quinn J J 1998 [*Physica*]{} B [**256**]{}–[**258**]{} 490 Wójs A, Hawrylak P and Quinn J J 1999 B [**60**]{} 11 661 Wójs A, Szlufarska I, Yi K S and Quinn J J 1999 B [**60**]{} R11 273
Fano G, Ortolani F and Colombo E 1986 B [**34**]{} 2670
Leinaas J M and Myrheim J 1977 [*Il Nuovo Cimento*]{} B [**37**]{} 1 Wilczek F 1982 1144
Silin V P 1959 [*Sov. Phys.–JETP*]{} [**8**]{} 870
Laughlin R B 1985 [*Science*]{} [**242**]{} 525 Laughlin R B 1988 2677 Fetter A L, Hanna C B and Laughlin R B 1989 B [**39**]{} 9679 Chen Y H, Wilczek F, Witten E and Halperin B I 1989 [*Int. J. Mod. Phys.*]{} B [**3**]{} 1001
Willet R L, Eisenstein J P, Störmer H L, Tsui D C, Gossard A C and English J H 1987 1776
Chen X M and Quinn J J 1994 B [**50**]{} 2354
Wójs A and Quinn J J 2000 B [**61**]{} 2846
Whittaker D M and Shields A J 1997 B [**56**]{} 15 185
Wójs A, Quinn J J and Hawrylak P 2000 Unpublished (http://xxx.lanl.gov/abs/cond-mat/0001327)
Hayne M, Jones C L, Bogaerts R, Riva C, Usher A, Peeters F M, Herlach F, Moshchalkov V V and Henini M 1999 B [**59**]{} 2927
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a user-friendly and versatile experimental technique that generates sub-kilohertz sinusoidal oscillatory flows within microchannels. The method involves the direct interfacing of microfluidic tubing with a loudspeaker diaphragm to generate oscillatory flow in microchannels with frequencies ranging from $10-1000$ Hz and amplitudes ranging from $10 - 600 \ \mu$m. The speaker-based apparatus allows independent control of frequency and amplitude that is unique to the speaker’s manufacturing specifications. The performance of our technique is evaluated by Fourier spectral analysis of oscillatory motion of tracer particles, obtained by particle tracking velocimetry, as well as by comparing oscillatory flow profiles against theoretical benchmarks such as Stokes flow in a square channel and Stokes’ second problem near a solid boundary. Applications that utilize both the oscillatory flow and the associated steady rectified flows are demonstrated in prototypical microfluidic configurations. These include inertial focusing and mixing at low Reynolds numbers, respectively.'
author:
- Giridar Vishwanathan
- Gabriel Juarez
bibliography:
- 'oscflows.bib'
title: 'Generation and application of sub-kilohertz oscillatory flows in microchannels'
---
[^1]
Introduction
============
Oscillatory flows in microfluidic devices have been shown to be useful in a range of applications such as mixing at low Reynolds numbers [@Phelan; @Ahmed; @Frommelt], particle sorting and focusing [@Thameem2016; @Schmid; @Marmottant; @Mutlu], enhancement of heat transfer [@Qu], flow control [@leslie; @phillips], microrheology [@VishwanathanPoF; @VishwanathanJNNFM], and chemical extraction [@Lestari; @Xie]. Nevertherless, the widespread use and study of oscillatory flows in microchannels remains uncommon due to challenges of implementation.
At low frequencies ($0.1 \leq f \leq 10$ Hz), oscillatory flows are usually achieved by a programmed syringe pump, electromechanical relay valves [@Abdolhasani] or a pneumatic pressure controller [@Zhou]. The fidelity of the desired waveform is limited by inertia of the oscillatory driver. For low frequencies, the response time of syringe pumps and actuators in electromechanical valves and pnuematic pressure controllers is on the order of $\mathcal{O}(10 \ \textrm{ms})$, therefore preventing the realization of sinusoidal oscillations at higher frequencies.
At high frequencies ($10^3 \leq f \leq 10^6$ Hz), piezoelectric transducers, which typically possess resonant frequencies in this range, are used [@Rallabandi; @phillips; @Xie; @Lieu; @Morris]. The utility of piezoelectric transducers in the 10 - 1000 Hz range are limited by the small amplitudes generated. The amplitudes may be partially increased through the use of designed features such as membrane cavities in the channel on to which the piezo elements need to be bonded to be used properly [@Vazquez]. More recent designs of microfluidic oscillators primarily aim to achieve oscillatory flows free of external actuators with a focus on miniaturization and integration with other lab-on-chip modules. This is typically done by exploiting non-linear fluid-elastic interaction with a membrane or diaphragm unit as a steady flow is driven through it. Therefore, a time dependent response is obtained even with a steady input at low Reynolds numbers [@Xia; @leslie; @Kim; @Mosadegh]. Other possibilities such as the use of non-Newtonian fluids for switching [@Groisman], generation of oil droplets as an oscillatory source [@Basilio], and the Coanda effect [@Yang] have also been explored. Although these micro-oscillators are highly miniaturized, modular and in some cases, capable of producing frequencies in the audible range [@Xia], they mostly require the fabrication of complex MEMS devices potentially discouraging their use in research attempting to use oscillatory flows. Further, in microfluidic oscillators that function based on fluid elastic interaction the amplitudes and frequencies are coupled and hence cannot be independently controlled.
Here, we describe the operation and performance of a simple plug and play apparatus capable of producing oscillatory flow in microchannels. This setup allows the user to independently control the oscillation frequency in the range of $10-1000$ Hz and amplitude in the range of $10-600 \ \mu$m. The aim of this method is simplicity and accessibility, hence, allowing researchers to implement oscillatory flow at the microscale in a convenient and cost-effective manner without the need for prior design constraints and sophisticated microfabrication.
Experimental setup
==================
{width="\linewidth"}
The apparatus is set up as displayed in the schematic, shown in Figure \[fig:figone\](a). A PDMS microchannel is bonded to a glass slide and observed through an inverted microscope. A loudspeaker (DROK TDA7297B, $15$ W, $90$ dB) is mounted next to the microscope stage. The oscillation frequency and amplitude (volume) of the loudspeaker diaphragm are controlled by a computer via an auxiliary cable. One end of microfluidic tubing (PE60 Intramedic 427416, 0.76 mm ID $\times $ 1.22 mm OD) is directly attached to the diaphragm of loudspeaker while the other end is inserted into the microchannel outlet. The tubing is maintained taut and its boundary conditions correspond to a fixed end at the microchannel outlet and a forced oscillatory displacement at the diaphragm. The microchannel and the tubing are filled with liquid during operation. The oscillatory displacement of the diaphragm is transduced into elastic deformations of the microfluidic tubing at the fixed end of the device outlet. The stress induced by tubing deformation generates a time-varying pressure within, resulting in oscillatory displacement of the liquid at the same frequency ($f$) as the diaphragm. The streamwise displacement of a tracer particle in the channel is described by $x=s \cos(\omega t + \phi)$ and illustrated in Figure \[fig:figone\](b). Here, $s$ oscillation amplitude, $\omega$ is the angular frequency, and $\phi$ is the initial phase.
To characterize the oscillatory flow in microchannels, tracer particles at the midplane of the microchannel were observed using brightfield illumination with objectives of $10\times$ and $20\times$ magnification (depth of field 8.5 $\mu$m and 5.5 $\mu$m respectively). To ensure that the tracers accurately represented the flow, polystyrene tracer particles with a mean diameter of $0.93\ \mu$m and density of $1.08$ g/cm$^{3}$ were suspended in deionized water (unless mentioned otherwise). The response time associated with the tracer particles ($\tau = \rho d^2 /18 \mu \approx 50 \times 10^{-9}$ s) is much smaller than the oscillatory timescales considered in this study ($\tau \ll 1/f$). The particle positions were recorded using a high-speed scientific CMOS camera with frame rates exactly twenty times larger than the driving oscillation frequency ($20f$). The displacement and velocity fields are then obtained from 2D particle tracking velocimetry algorithms.
The microchannel geometries used were the following: (i) a square channel with a width and height of $110 \ \mu$m and length of $5$ cm, (ii) a rectangular channel with a width of $5$ mm, height of $200 \ \mu$m, and length of $2$ cm, and (iii) a cross slot channel with square cross section with a width and height of $110 \ \mu$m.
The dimensionless groups considered here are the Womersley number, the Reynolds number, and the non-dimensional oscillation amplitude. The Womersley number is the ratio of the characteristic channel width to the Stokes boundary layer thickness and is defined as $\alpha = W \sqrt{ \rho \omega / \mu}$ [@Landau]. Here, $W$ is the channel width, $\rho$ and $\mu$ are the liquid density and dynamic viscosity, respectively. The Reynolds number is the ratio of inertial to viscous forces of the fluid within the microchannel and defined as $\textrm{Re}= \rho s\omega W/\mu$ [@Thomas2011]. Above, $s \omega$ is taken to be the characteristic liquid velocity. The non-dimensional oscillation amplitude may be expressed in terms of the Womersley and Reynolds number as $\epsilon=s/W=\text{Re}/\alpha^2$. For operational conditions considered in this study, a microchannel with characteristic width of $200\ \mu$m, and DI water as the working liquid, the values of $\alpha$, $\textrm{Re}$, and $\epsilon$ range from $1.5 < \alpha < 15$, $0.4 < \textrm{Re} < 80$, and $0.1 < \epsilon < 5$, respectively.
Results
=======
Oscillatory displacement in microchannels
-----------------------------------------
Examples of streamwise displacement of individually tracked particles from their mean position over a number of cycles during a $20$ ms period are shown in Figure \[fig:figone\](c). The ratio of sampling frequency (camera framerate) to liquid oscillation frequency is kept constant at $20$. That is, for oscillation at $200$, $400$ and $800$ Hz, a framerate of $4000$, $8000$ and $16000$ Hz is used, respectively. The corresponding amplitudes are $100$, $27$ and $14\ \mu$m.
The independent operational range between amplitude and frequency is shown in Figure \[fig:figtwo\](a). For a given frequency, the displacement of a tracer particle is dependent on the volume setting of the loudspeaker. As an example, three volume settings are considered here: low (30%), intermediate (60%), and high (90%), where the percentages correspond to the maximum speaker volume as determined by the computer. At 100 Hz, for example, the amplitude ranges from $50 \ \mu$m at low to $800 \ \mu$m at high volume setting. The amplitude swept by a tracer particle over a single oscillation period, for a given volume setting, shows a non-monotonic variation with frequency. Owing to the performance characteristics of the speaker, the maximum oscillation amplitude occurs at $200$ Hz, which corresponds to the resonant frequency ($\approx 230$ Hz) of the loudspeaker diaphragm. The horizontal black line indicates the maximum particle oscillation amplitude of $800 \ \mu$m that can be measured due to the field of view limited by the camera when using a $10\times$ microscope objective lens.
A Fourier spectrum analysis of particle trajectories at varying oscillation amplitudes and frequencies is shown in Figure \[fig:figtwo\](b). The spectra have been obtained for oscillation frequencies of $50$, $200$ and $800$ Hz and at volume settings of low, intermediate, and high. For low and intermediate volume settings, monodisperse peaks in the spectral intensity correspond to the input driving frequency of the loudspeaker. The peaks are especially narrow at $200$ Hz, or near the resonance frequency of the diaphragm. For high volume settings, the peaks correspond to the input driving frequency, however, widening of the peak is noticeable. In some cases, such as for $800$ Hz at high volume setting, contributions due to higher harmonics are of considerable strength.
![ (a) The amplitude of oscillatory displacement in microchannels for a range of frequencies and three amplitudes, or speaker volume settings, of low (30%, blue), intermediate (60%, magenta) and high (90%, red). (b) Fourier spectrum analysis of tracer particle displacement in the streamwise direction at three different frequencies (50, 200, and 800 Hz) and amplitudes (low, intermediate, and high). []{data-label="fig:figtwo"}](fig2_05.pdf){width="\linewidth"}
A quantitative measure of harmonic distortion present in the signal as compared to the fundamental driving frequency is obtained by calculating the total harmonic distortion (THD). The THD is defined as $\textrm{THD} =\sqrt{\Sigma^{N}_{i=1} V_i^2}/V_1$, where $V_i$ is the power of the spectral intensity at the $i$th harmonic [@Shmilovitz]. A low THD value is associated with a more accurate representation of the original driving signal. For low volume settings, the THD at $50$, $200$, and $800$ Hz are $3.5\%$, $7.1\%$, and $9.1\%$, respectively. For intermediate volume settings, the THD at $50$, $200$, and $800$ Hz are $5.2\%$, $8.9\%$, and $13.4\%$, respectively.For high volume settings, the THD at $50$, $200$, and $800$ Hz are $7.3\%$, $11.1\%$, and $21.3\%$, respectively. The growing magnitude of higher harmonics with increasing speaker volume typically limits operation at frequencies $>400$ Hz to low or intermediate speaker volumes. At low frequencies however, the maximum amplitude is chosen to avoid damage to the microchannel or unfastening of the outlet tube from the speaker cone. Therefore, sinusoidal oscillations with amplitudes ranging from $10 < s < 200 \ \mu$m can be reliably achieved throughout the entire range of frequency.
The maximum pressure inside the square channel may be estimated from the modified Poiseuille formula: $$\label{eq:pressuredrop}
\Delta P=\frac{64\mu Lsf}{D_h^2} \ ,$$ where $D_h$ and $L$ are the hydraulic diameter and length of the channel. For the square channel ($D_h=110\ \mu$m and $L=5$ cm) filled with DI water ($\mu=1.002$ mPa s) and settings for maximum oscillatory displacement ($f=200$ Hz, $s=600 \ \mu$m), the pressure inside the channel is calculated to be approximately equal to $31$ kPa.
Oscillatory flow in microchannels
---------------------------------
The small length scales of $\mathcal{O}(100 \ \mu \textrm{m})$ associated with microchannels imply that most microscale flows are laminar flows governed by the Stokes equation. An important feature of microscale oscillatory flows in the $10-1000$ Hz range is that transient effects associated with the unsteady Stokes equation become significant. An example of departure from Stokes flow is illustrated by the comparison of the steady Stokes flow velocity profile [@Obrien] at the midplane (black solid curve) against those obtained experimentally for oscillatory flow at different frequencies (symbols), shown in Figure \[fig:figthree\](a).
To obtain the amplitude of velocity in the square channel midplane, $50-200$ particles ($0.93 \ \mu$m diameter) are tracked for one hundred oscillation cycles and their respective velocities are computed. The amplitude of each velocity series ($U_{max}(y)$) is obtained and superposed in the streamwise direction. The resulting spread of speeds is filtered for outliers and averaged. The associated statistical error bars are smaller than the data markers shown.
The results for $100$ Hz and $400$ Hz are similar to the Stokes laminar flow profile (black solid curve). At $800$ Hz, however, there is considerable deviation from the steady velocity profile due to increasing $\alpha$ with frequency. For the cases of $100$, $400$, and $800$ Hz the $\alpha$ values are $2.75$, $5.51$, and $7.78$, respectively. For these $\alpha$ values, the analytical series solution for the amplitude of the midplane oscillatory velocity profile ($U_{max}(y)$) was evaluated correct to one hundred terms [@Obrien], and are shown by the continuous lines to good agreement with experimental data, even at $800$ Hz.
![ (a) Oscillatory velocity flow profile in a square channel (symbols) showing deviation from Stokes flow profile (black solid curve) with increasing frequency. (b) Oscillatory velocity profile with position (symbols) near a solid channel wall in a semi-infinite rectangular channel ($W \gg H$) showing agreement with the theoretical solution (black solid curve) to Stokes’ second problem [@Landau] with increasing frequency. (Inset) Close-up of the region from $1.5 \leq y/\delta \leq 4.5$ to demonstrate agreement with equation (\[eq:stokes2ndprob\]) at different frequencies. []{data-label="fig:figthree"}](fig3_07.pdf){width="\linewidth"}
In contrast, deviations from the unsteady Stokes equation are demonstrated in Figure \[fig:figthree\](b) by comparing the amplitude of measured oscillatory flow (symbols) in the rectangular channel ($W \gg H$), with those obtained theoretically from the solution to oscillatory flow over an infinite flat plate (Stokes’ second problem) [@Landau; @Wang89]. The maximum temporal flow velocity as a function of distance from a flat plate is given by the expression: $$\label{eq:stokes2ndprob}
U_{max}(y)/s\omega=\sqrt{2} e^{-y/(2\delta)} \sqrt{\cosh{(y/\delta)} -\cos{(y/\delta)}} \ ,$$ where $\delta$ is the Stokes boundary layer length and equal to $\sqrt{\mu / \rho \omega}$.
At $800$ Hz, normalized experimental data (symbols) is in good agreement with equation (\[eq:stokes2ndprob\]). At $50$ Hz, and to a lesser extent at $200$ Hz, deviations from the theory occur where the velocity amplitudes are larger than those at the far field and is detailed in the inset. This is due to the relatively short channel height ($200\ \mu$m) and the effect of boundary layers at the top and bottom walls of the channel affecting flow at the midplane. The corresponding values of $\alpha$ for the cases of $50$, $200$, and $800$ Hz are $3.57$, $7.14$, and $14.3$, respectively. Therefore, at the midplane, effects of channel side-walls may be neglected for distances larger than $4\delta$ into the channel. Further, three dimensional flow effects can be ignored for $\alpha \geq 7.5$ when the shorter dimension is used.
Applications of oscillatory flow in microchannels
-------------------------------------------------
{width="\linewidth"}
The application of oscillatory flows in microchannels may be broadly divided into two categories. The first category of flows are those where oscillatory motion enables instantaneous local velocities or shear rates without net displacement, which is usually implemented to reduce device footprint and allow for prolonged observation [@Jo; @Alizadehgiashi; @Mutlu; @Abdolhasani]. The second category of flows are those that utilize steady rectified flows associated with an underlying primary high frequency oscillatory flow [@Riley], which have been shown to be useful in mixing [@Ahmed], hydrodynamic manipulation of particles and cells [@Lutz; @Lieu; @Thameem2016], and more recently, in microrheology [@VishwanathanPoF; @VishwanathanJNNFM]. Here, we demonstrate two specific applications of the oscillatory driver from each category, namely, inertial focusing from the former and mixing from the latter, using simple prototypical microfluidic configurations.
### Inertial focusing
Inertial focusing in microchannels is a passive technique where suspended particles undergoing unidirectional flow migrate across streamlines, due to particle inertia, to equilibrium focus positions [@DiCarlo; @DiCarlo2009; @Martel; @Stoecklein]. In straight channels, of rectangular or circular cross section, the competing forces that lead to particle migration are the wall interaction force, which directs the particle away from the channel wall, and the shear gradient force, which directs the particle toward the channel wall. The summation of these forces is termed the inertial lift and the equilibrium position of the particle is determined once the opposing forces are balanced. Factors influencing the inertial lift force are the channel geometry, flow rate, and particle size.
Since it is a high-throughput method for non-contact manipulation at the microscale, inertial focusing has been utilized in numerous applications ranging from flow cytometry [@Hur; @Bhagat; @Oakey], size sorting [@Kuntaegowdanahalli; @Wu; @Nivedita], mixing [@Amini], and filtration [@Seo]. An important parameter when designing channels for applications is the length required to reach the equilibrium focus position, estimated [@DiCarlo2009] to be equal to $$\label{eq:focuslength}
L_f = \frac{\pi \mu D_h^2}{ \rho U_m a^2 C_{\ell}}\ ,$$ where $U_m$ is the maximum flow velocity and $C_{\ell}$ is the lift coefficient, which typically varies in the range of $0.02 - 0.05$. From this relation, it is apparent that sufficiently high velocities and large particles are required to minimize the focusing length. Recent work, however, demonstrates the use of oscillatory flows for inertial focusing, where oscillatory flow at relatively low frequencies ($<20$ Hz) results in a channel of practically infinite length [@Mutlu]. Thus, it becomes possible to focus particles with far smaller particle Reynolds numbers corresponding to $\textrm{Re}_{\textrm{p}}= (a/W)^2 \textrm{Re} <0.01$.
{width="\linewidth"}
Inertial focusing of a $10\ \mu$m polystyrene particle in oscillatory flow of $400$ Hz with an amplitude of $22\ \mu$m is demonstrated in the square channel, for which, $\textrm{Re}_{\textrm{p}}=0.050$ and $a/W=0.091$. Micrographs of a single polystyrene particle at regular time intervals as it migrates to the equilibrium position is shown in Figure \[fig:figfour\](a). Using stroboscopic imaging, the lateral migration toward the center of the channel is apparent. The corresponding vertical position of the particle centroid, as determined by particle tracking, is shown as a function of the approximate path length traversed by the particle in Figure \[fig:figfour\](b). The marked points (symbols) correspond to the instances shown in the micrographs. The path length is estimated as $4sft$, where $4s$ is the distance covered by the particle at the center of the channel in a single oscillation cycle and $t$ is the time elapsed after the start of oscillations. The focusing behavior of another particle at $25$ Hz and an amplitude of $36\ \mu$m is also shown. The corresponding path length for focusing (or focus length) is found to be about $0.15$ m, comparable to the $400$ Hz case and in good agreement with previous results for similar conditions ($\textrm{Re}_{\textrm{p}}=0.050$ and $a/W=0.125$) [@Mutlu]. The time required for focusing, however, is about $3$ seconds for the $400$ Hz case and $35$ seconds for the $25$ Hz case.
The advantages of oscillatory flow for inertial focusing include decreased channel lengths, lower pressure drops, and lower shear rates. Because there is no net displacement, the particle remains in the field of view as it migrates to the equilibrium position. In contrast with unidirectional flow, the approximate channel length required for the particle to reach it’s equilibrium position as determined by equation (\[eq:focuslength\]), where $U_m$ is given by the characteristic fluid velocity $s \omega$, is $L_f \approx 1.37$ m, which is impractical. The lower pressure drop allows for the convenient fabrication and use of PDMS microchannels. The combination of lower pressure drop and lower shear rates is of particular interest in biomedical application where cells are susceptible to damage induced by fluid stresses.
### Microscale mixing
Low Reynolds number ($\textrm{Re}\leq1$) flow in microchannels present a significant challenge to applications involving mixing [@Ottino]. This is because, the dominant mechanism of mixing is diffusion in the absence of chaotic advection, which is normally associated with high $\textrm{Re}$ mixing. The effectiveness of mixing is quantified by estimating the length required to achieve mixing ($L_m$). For purely diffusive mixing, $$\label{eq:mixlength}
L_m \geq D_h \textrm{Pe} \ .$$ Above, $\textrm{Pe}$ is the Peclet number and defined as the ratio of $U_aW/D$, where $U_a$ is the average flow velocity and $D$ is the diffusion coefficient. For $\textrm{Re} \leq 1$ and diffusion coefficients of $D \approx 100 \ \mu$m$^2$ s$^{1}$, the Peclet number would be $\textrm{Pe} \geq 10^4$, resulting in a channel length for sufficient mixing to be $L_m \geq 1$ m, which is undesirable.
To overcome this challenge, a variety of microscale mixers have been developed to enhance mixing at low Re, and are categorized as either passive or active mixers [@Hessel; @Nguyen; @Lee; @Ward; @lee16; @Cai]. Passive mixers make use of the channel geometry, usually incorporating repeating complex or 3D channel features to enhance mixing of two streams flowing together at a constant rate. On the other hand, active mixers rely on externally applied forces and are further categorized based on the nature of the external actuation [@Ober]. One such category of mixers are acoustic micromixers [@Liu; @Ahmed; @Bachman] that rely on external acoustic actuation to generate steady rectified flows that mix different liquids.
Here, mixing of two aqueous glycerol solutions ($30\%$ w/w, $\mu=2$ mPa s), one with colored dye and one without, is demonstrated in the cross-slot channel using steady rectified flows. As seen from figure \[fig:figfive\](a), the dyed solution enters the cross-slot with a flow rate of $0.3\ \mu$L/min (red arrow) while the undyed solution enters the cross-slot from either side with identical flow rates of $0.15 \ \mu$L/min (green arrows). This results in a total flow rate of $0.6 \ \mu$L/min at the outlet.
When no oscillatory flow is imposed, minimal diffusive mixing at the interface is observed within the field of view. When oscillatory flow of $400$ Hz is imposed on the same configuration, steady vortices are generated near the corners of the cross-slot as seen in Figure \[fig:figfive\](b), which facilitate mass transfer across the interface through advection. A large exposure time of $20$ ms was used for imaging so that variations over signal phase are averaged.
Mixing performance is quantified by first obtaining the intensity profile across the channel at a distance $2.5W$ downstream from the center of the cross slot, indicated by the yellow dashed lines in Figure \[fig:figfive\](a) and (b). The standard deviation of the mixture fraction profile, derived from the intensity values, is used as the mixing index and is defined as $\textrm{MI} = \sqrt{\Sigma(I_i - I_m)^2/N}$, where, $I_i$ is the pixel intensity value and $I_m$ is the pixel intensity of a completely mixed solution, and $N$ is the number of sampling points [@Liu00]. The values of the index range from $\textrm{MI}=0.5$ for completely unmixed to $\textrm{MI}=0$ for completely mixed solutions. A value of $\textrm{MI} \leq 0.1$ indicates sufficient mixing.
The variation of the mixing index as a function of frequency for both constant oscillation amplitude and constant volume setting are shown in Figure \[fig:figfive\](c). For constant amplitude settings, the mixing index decreases monotonically with increasing frequency, implying sufficient mixing for $f \geq 400$ Hz. The improved mixing with increasing frequency is due to the increase in magnitude of the steady rectified flow velocities, which scale as $\mathcal{O}(s^2\omega)$. For a constant volume (intermediate) setting, however, the mixing index is non-monotonic, with sufficient mixing occurring in the range of $100-200$ Hz. Based on the amplitude characteristics, shown in Figure \[fig:figtwo\](a), for constant volume settings, the largest amplitudes occur near the resonance frequency of the loudspeaker diaphragm.
For the specific case presented here, the length demonstrated to achieve good mixing with steady rectified flows ($L_m \leq 250 \ \mu$m) is much less than the length required for mixing according to equation (\[eq:mixlength\]), which is calculated to be $L_m = 2.5$ cm. Although the nature of forces involved are purely hydrodynamic, oscillatory flows with independently controllable amplitude and frequency allow for the decoupling of flow rate from the rate of mixing which is not possible for passive micromixers. Additionally for the range of frequencies and amplitudes achieved here, strong rectified flows are achieved near solid boundaries in the microchannel as opposed to the boundaries of bubbles used with ultrasonic frequencies elsewhere [@Liu; @Ahmed] which are unstable at long operation times. Lastly, the implementation of this method in combination with any other passive or active technique can further enhance mixing at the microscale by increasing the number of passes without affecting flow rate.
Discussion
==========
We have discussed an accessible, effective, and versatile plug and play technique to generate oscillatory flows over a range of amplitudes and frequencies in microchannels. By directly interfacing microfluidic tubing with a loudspeaker diaphragm, sub-kilohertz oscillatory flows in the range of $10-1000$ Hz with amplitudes in the range of $10-600 \ \mu$m are produced. The corresponding wavelengths lie between $1-100$ m and are far larger than the dimensions of a typical microchannel. Thus, nearly unattenuated flows of a uniform phase can be achieved throughout. This is in contrast with flows in the $10^4-10^7$ Hz range where attenuation is significant and effects are usually local to the transducer. The resulting velocity profiles can also be tuned from Stokes-like flow ($\alpha\leq4$) at low frequencies to plug-like unsteady Stokes flow ($\alpha\geq7.5$) at high frequencies allowing for manipulation of the flow profile for a given micro-scale geometry. In addition to coherent oscillatory flows, strong and well defined rectified flows near curved boundaries and interfaces are also made possible in this frequency and amplitude range.
The guiding principles for applications are two-fold. First, oscillatory flows permit an increase the net distance travelled by the fluid without an accompanied increase in flow rate, shear rate, pressure drop, or particle displacement seen in long channel steady flows and high throughput applications. More generally, such conditions are particularly useful when the desired sample or analyte response to the flow environment is directionally invariant. In theory, the resulting effect can be increased indefinitely by simply increasing the frequency of oscillation. Although in practice, attenuation and secondary flows which grow stronger with increasing frequency prevent this from being realized. Owing to the large wavelengths in this range of flows, attenuation is limited while secondary flows usually become significant only for frequencies larger than 100 Hz implying that an optimum frequency is likely encountered in the range of frequencies realized here. The resulting lower steady pressure drop and steady shear rates are particularly useful for cells and other suspended biological matter that are sensitive to or potentially damaged by prolonged exposure to excessive shear rates or pressures. The absence of particle net displacement is useful in decreasing the footprint of microfluidic devices and in situations where continuous observation is needed to track the evolution in processes such as cell growth or chemical synthesis in dynamic environments.
Second, the steady rectified flow speeds are of considerable magnitude despite the maximum frequency considered being much smaller than the typical resonant frequencies of piezoelectric transducers ($1-100$ kHz). This is because of the large amplitudes accessed by the loudspeaker diaphragm and the dependency of flow speed on amplitude, which scales as $s^2\omega$. Further, the Stokes boundary layer thickness is accurately controlled and varies in size from $10 \leq \delta \leq 100 \ \mu$m for the highest and lowest frequencies accessed here, respectively. Therefore, the flow pathlines are less sensitive to manufacturing defects and feature surface quality, making this approach more amenable to precision applications such as sorting, trapping and manipulation of particles and cells, viscometry, and other controlled mass transfer applications.
[^1]: Email address.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- '<span style="font-variant:small-caps;">Kazushi MIMURA</span> [^1]'
title: Linear Complexity Lossy Compressor for Binary Redundant Memoryless Sources
---
Channel coding can be considered as a dual problem of lossy source coding [@Cover2006; @Csiszar1981]. Recent research on error correcting codes and lossy source coding has shown that the statistical mechanical approach can be used to explain such problems [@Nishimori2001]. Lossy compression for memoryless sources has been widely investigated since Matsunaga and Yamamoto showed that it is possible to attain the rate-distortion bound asymptotically using low-density parity-check (LDPC) codes [@Matsunaga2003]． The upper [@Dimakis2007] and lower [@Martinian2006; @Wainwright2007] bounds on its rate-distortion performance of low-density generator-matrix (LDGM) codes for lossy compression with a given check degree are evaluated. Some other lossy compression schemes that have asymptotic optimality have been proposed so far [@Murayama2003; @Hosaka2002; @Hosaka2006; @Ciliberti2005; @Ciliberti2006; @Mimura2006; @Gupta2009; @Miyake2008a; @Muramatsu2008; @Cousseau2008a; @Cousseau2008b; @Mimura2009]．
Efficient compressors, on the other hand, are still in the stage of development. Some efficient encoding algorithms, which have near optimal performance, have been proposed, e.g., the nested binary linear codes [@Wadayama2003], the inertia-term-introduced belief propagation [@Murayama2004; @Hosaka2006]， the survey-propagation-based message passing algorithm [@Wainwright2005]， the bit-flipping-based algorithm [@Gupta2009], the exhaustive search of small words into what an original message is divided [@Gupta2008]， the linear-programming-based algorithm, [@Miyake2008b] and the Markov-Chain Monte-Carlo(MCMC)-based algorithm [@Jalali2008].
For redundant memoryless sources, some low complexity compressors, e.g., the near-linear complexity compressor based on the exhaustive search of small words [@Gupta2008], the quadratic complexity compressor based on the bit-flipping-based algorithm [@Gupta2009], and the MCMC-based compressor [@Jalali2008] whose complexity is independent of the sequence length, have been proposed so far. One of other approaches to obtain near-linear complexity compressors for redundant sources is to apply the inertia-term-introduced belief propagation. Hosaka and Kabashima have proposed an algorithm for redundant sources, whose complexity is $O(N^2)$ [@Hosaka2006]. In this study, we propose a linear complexity lossy compression algorithm based on an inertia-term-introduced belief propagation by using a nonlinear function and a sparse matrix such as low-density generator matrix (LDGM) codes for binary redundant memoryless sources. This proposed algorithm can be regarded as the perceptron-based one [@Hosaka2006] whose edges are extremely deleted to have a finite connectivity and has asymptotic optimality under some constraints. We also show that it has near optimal performance for moderate block lengths.
Let us first provide the concepts of the rate-distortion theory [@Cover2006] and some notations. Let $x$ be a discrete random variable with an alphabet $\mathcal{X}$. A source alphabet, a codeword alphabet, and a reproduced alphabet are $\mathcal{X}$, $\mathcal{S}$ and $\hat{\mathcal{X}}$, respectively. The compressor $\mathcal{F}$ encodes the $M$ bit source sequence ${\boldsymbol{x}}={}^t(x_1,\cdots,x_M) \in \mathcal{X}^M$ into the $N(<M)$ bit codeword ${\boldsymbol{\xi}}={}^t(\xi_1,\cdots,\xi_N)=\mathcal{F}({\boldsymbol{x}})\in\mathcal{S}^N$. The decompressor $\mathcal{G}$ generates the $M$ bit reproduced sequence $\hat{{\boldsymbol{x}}}={}^t(\hat{x}_1,\cdots,\hat{x}_M)={\cal G}({\boldsymbol{\xi}}) \in \hat{\cal X}^M$ from the codeword ${\boldsymbol{\xi}}$. The code rate then becomes $R=N/M(<1)$.
A distortion measure is a map $d : \mathcal{X} \times \hat{\mathcal{X}} \to [0,\infty)$. A distortion between the sequences ${\boldsymbol{x}} = {}^t (x_1, \cdots, x_M) \in \mathcal{X}^M$ and $\hat{{\boldsymbol{x}} } = {}^t (\hat{x}_1, \cdots, \hat{x}_M) \in \hat{\mathcal{X}}^M$ is measured by the averaged single-letter distortion as $d({\boldsymbol{x}},\hat{{\boldsymbol{x}}}) = \frac 1M \sum_{k=1}^M d(x_k,\hat{x}_k)$. A rate distortion pair $(R,D)$ is said to be achievable if there exists a sequence of rate distortion codes $({\cal F},{\cal G})$ with $\mathbb{E}_{{\boldsymbol{x}}} [d({\boldsymbol{x}},\hat{{\boldsymbol{x}}})] \le D$ in the limit $M\to\infty$. The rate distortion function $R(D)$ is the infimum of the rate $R$ such that $(R,D)$ is in the rate distortion region of the source for a given distortion $D$.
We hereafter consider the binary alphabets $\mathcal{X}=\mathcal{S}=\hat{\mathcal{X}}=\{ -1,1 \}$ and a redundant binary memoryless source whose distribution is given by $\mu(X=1)=1-p, \mu(X=-1)=p$. The parameter $p$ is a source bias. We use the Hamming distortion $$d(x,\hat{x}) = \biggl\{
\begin{array}{ll}
0, & \; {\rm if} \; x = \hat{x}, \\
1, & \; {\rm if} \; x \ne \hat{x},
\end{array}$$ as a distortion measure. The rate-distortion function of a Bernoulli($1/2$) i.i.d. source then becomes $R(D)=h_2(p)-h_2(D)$, where $h_2$ denotes the binary entropy function which is defined by $h_2(x) = -x\log_2(x) - (1-x) \log_2(1-x)$.
Using a parameter ${\boldsymbol{w}}=(w_1,w_2)$ $\in (\mathbb{N}\backslash \{0\})^2$, we first introduce the nonlinear function $g : \mathbb{N} \to \{-1,1\}$ defined as $$g_{{\boldsymbol{w}}} (z) = g_{(w_1,w_2)} (z) = \left\{
\begin{array}{rl}
1 & \; \mathrm{if} \; w_1 < |z| < w_2 \\
-1 & \; \mathrm{otherwise}
\end{array}
\right. ,$$ where the operator ${}^tA$ denotes the transpose of $A$. For a vector, this function acts componentwise. We here consider the following decompressor: $$\mathcal{G}({\boldsymbol{\xi}}) = g_{{\boldsymbol{w}}} ( \mathcal{A} {\boldsymbol{\xi}} ),$$ where $\mathcal{A}=(a_{\mu i}) \in \{-1, 0, 1\}^{M \times N}$ denotes a sparse regular matrix whose row weight, i.e., the number of nonzero elements, is $C$. Each nonzero element of the sparse matrix $\mathcal{A}$ takes $\pm 1$ with equiprobability. The function $g_{{\boldsymbol{w}}} (z)$ is adjusted so that the bias of the reproduced sequence is close to that of the original message as much as possible. More complex functions can be considered as the function $g_{{\boldsymbol{w}}} (z)$. As will be discussed later, the complexity of our proposed algorithm is $O(N)$ but it is proportional to $2^C$. So a small $C$ is preferable; therefore, we here consider a simple nonlinear function that can easily be adjusted.
The compressor is then defined by $$\mathcal{F}({\boldsymbol{x}}) = \mathop{\rm argmin}_{{\boldsymbol{s}} \in \{-1,1\}^N} d({\boldsymbol{x}},\mathcal{G}({\boldsymbol{s}})).
\label{eq:def_F}$$
This compressor is identical to $\mathcal{F}({\boldsymbol{x}}) = \mathop{\rm argmax}_{{\boldsymbol{s}} \in \{-1,1\}^N}$ $p({\boldsymbol{s}};{\boldsymbol{x}})$, which is the maximization of the following distribution $$p({\boldsymbol{s}};{\boldsymbol{x}})
\triangleq \frac 1{Z(\beta)} e^{-\beta M d({\boldsymbol{s}};{\boldsymbol{x}})}
= \frac 1{Z(\beta)} \prod_{k=1}^M e^{-\beta G_k({\boldsymbol{s}};x_k)},$$ where $G_k({\boldsymbol{s}};x_k) = \frac 12 (1-x_k \hat{x}_k)$, $\hat{x}_k = g_{{\boldsymbol{w}}} ( \sum_{i \in \mathcal{L}(k)} a_{ki} s_i )$ and $\mathcal{L}(k) = \{i|a_{ki} \ne 0\}$ with the parameter $\beta >0$. Here, $Z(\beta)$ denotes an normalization constant of $p({\boldsymbol{s}};{\boldsymbol{x}})$, which is defined by $Z(\beta) = \sum_{{\boldsymbol{s}}} e^{-\beta M d({\boldsymbol{x}},\mathcal{G}({\boldsymbol{s}}))}$. The function $G_k({\boldsymbol{s}};x_k)$ represents a distortion with respect to the $k$ th bit.
We here consider a large row weight limit, which is a case where $C = N$ holds, to allows us to infer compression performance. In this limit our scheme can be regarded as the perceptron-based code [@Hosaka2002; @Hosaka2006; @Cousseau2008a; @Cousseau2008b; @Mimura2011]. When $w_2>N$, these are equivalent to each other. To make the parameters $w_1$ and $w_2$ be of order unity, we introduce a constant into the decompressor as $\mathcal{G}({\boldsymbol{\xi}}) = {\boldsymbol{g}}_{{\boldsymbol{w}}} ( N^{-1/2} \mathcal{A} {\boldsymbol{\xi}} )$. The achievable distortion $D$ is evaluated as $D = \lim_{\beta \to \infty} \partial [\beta f(\beta)]/\partial \beta$ via the free energy density $f(\beta) = (-\beta M)^{-1} \mathbb{E}_{\mathcal{A},{\boldsymbol{x}}}[ \ln Z(\beta)]$, where $\mathbb{E}$ denotes an expectation operator.
Applying the so called replica method, the free energy density can be evaluated as $f(\beta) = - \beta^{-1} ( p \ln \{ e^{-\beta} + (1-e^{-\beta}) K_{{\boldsymbol{w}}} \} + (1-p) \ln \{ e^{-\beta} + (1-e^{-\beta}) (1-K_{{\boldsymbol{w}}}) \} + R \ln 2)$ within the replica symmetric treatment, where $K_{{\boldsymbol{w}}} = \int_{\{z \in \mathbb{R} |g_{{\boldsymbol{w}}}(z)=-1\}} {(2\pi)}^{-1/2} e^{-z^2/2} {\mathrm{d}}z$. The parameter $K_{{\boldsymbol{w}}}$ is identical to the expectation value $\mathbb{E}_z[g_{{\boldsymbol{w}}}(z)=-1]$ with a random variable $z$ which obeys the standard normal distribution $\mathcal{N}(0,1)$, which originates from the distribution of each element of $N^{-1/2} \mathcal{A} {\boldsymbol{\xi}}$. It can be considered that the compression performance is given using the distribution of $N^{-1/2} \mathcal{A} {\boldsymbol{\xi}}$ and $g_{{\boldsymbol{w}}}$ in this scheme.
The entropy density of $p({\boldsymbol{s}};{\boldsymbol{x}})$ is then obtained as $s(\beta) = \beta ( \partial [\beta f(\beta)]/\partial \beta - f )$. The entropy density $s(\beta)$ must be non-negative owing to the definition of $p({\boldsymbol{s}};{\boldsymbol{x}})$; however it takes negative values in the large $\beta$ region. We therefore evaluate the achievable distortion $D$ at $\beta_c$ which gives a zero entropy density ($s(\beta_c)=0$), so this analysis is equivalent to the Krauth-Mézard approach which is a kind of one-step replica symmetric breaking (RSB) treatment [@Krauth1989].
Using the zero-entropy-density criterion, minimizing the achievable distortion $D=\lim_{\beta \to \beta_c} \partial [\beta f(\beta)]/\partial \beta$ with respect to ${\boldsymbol{w}}$, one can obtain $R(D)=h_2(p)-h_2(D)$, which is identical to the rate-distortion function. From these two conditions, i.e., the zero entropy density and the minimization of the achievable distortion, the following relationships are obtained: $$\begin{aligned}
& & e^{-\beta} = \frac D{1-D}, \label{eq:condition1} \\
& & K_{{\boldsymbol{w}}} = \frac{p-D}{1-2D}. \label{eq:condition2} \end{aligned}$$
The definition of the compressor means that it has exponential complexity. We then utilize a suboptimal algorithm based on message passing to construct the compressor [@Murayama2004]. Instead of the maximization of $p({\boldsymbol{s}};{\boldsymbol{x}})$ we use a symbol MAP encoding scheme, which is maximization of marginal distribution, $$\xi_i = \mathop{\rm argmax}_{s_i \in \{-1,1\} } \sum_{{\boldsymbol{s}} \backslash s_i \in \{-1,1\}^{N-1} } p({\boldsymbol{s}};{\boldsymbol{x}}).$$ To evaluate the marginal distribution we apply the belief propagation. Since $\mathcal{G}({-{\boldsymbol{s}}})=\mathcal{G}({{\boldsymbol{s}}})$ holds for any ${\boldsymbol{s}}$, the expectation value of $s_i$ becomes zero. To avoid this uncertainty we heuristically introduce an inertia term as a prior, which gives the following [*inertia-term-introduced belief propagation*]{} [@Murayama2004]: $$\begin{aligned}
& & \hspace*{-5mm} \hat{\rho}_{ki}^t(s_i) = \sum_{s_{i' \in \mathcal{L}(k) \backslash i}} e^{-\beta G_k({\boldsymbol{s}};x_k)} \prod_{i' \in \mathcal{L}(k) \backslash i} \rho_{i'k}^t(s_{i'}), \\
& & \hspace*{-5mm} \rho_{ik}^{t+1}(s_i) = \alpha_{ik} r_i^t(s_i) \prod_{k' \in \mathcal{M}(i) \backslash k} \hat{\rho}_{k'i}^t(s_i). \end{aligned}$$ A pseudo marginal can be evaluated as $$q_i^{t+1}(s_i) = \alpha_i r_i^t(s_i) \prod_{k \in \mathcal{M}(i)} \hat{\rho}_{ki}^t(s_i),$$ where $\alpha_{ik}$ and $\alpha_i$ denote normalization constants and $\mathcal{M}(i)=\{k|a_{ki} \ne 0 \; \forall \mu \}$. Here, the function $r_i^t(s_i)$ is an introduced prior as the inertia term and the superscript $t$ represents an iteration step. We here consider that $s_i$ is binary, so we can safely put $\rho_{ik}^t(s_i) = \frac12(1 + m_{ik}(t) s_i)$, $\hat{\rho}_{ki}^t(s_i) = \frac12(1 + \hat{m}_{ik}(t) s_i)$, $q_i^t(s_i) = \frac12(1 + m_i(t) s_i)$. We here define a prior as $r_i^t(s_i) = e^{s_i \tanh^{-1} [\gamma m_i(t)]}$, where the parameter $\gamma \; (0\le\gamma<1)$ denotes the amplitude of the inertia term, which is heuristically chosen. When $\gamma=0$ ($r_i^t(s_i)=1$), the inertia-term-introduced belief propagation recovers the conventional belief propagation. It should be noted that the performance does not strongly depend on the detailed shape of the function, if it is an increasing function. It has not yet been investigated how the inertia term works in detail so far; however it is known that the inertia term chooses a single peak in the calculation of the pseudo marginal.
We calculate the equations of the belief propagation, which gives $$\begin{aligned}
\hat{m}_{ki}(t)
&=& \frac{\displaystyle a_{ki} x_k \biggl(\tanh \frac{\beta}2 \biggr) V_{ki}(t)} {\displaystyle 1 + x_k \biggl(\tanh \frac{\beta}2 \biggr) U_{ki}(t)}, \label{eq:algo1} \\
m_{ik}(t+1)
&=& \tanh \Biggl( \sum_{k' \in \mathcal{M}(i) \backslash k} \tanh^{-1} \hat{m}_{k'i}(t) \nonumber \\
& & \qquad\qquad + \tanh^{-1} \gamma m_i(t) \Biggr), \\
m_i(t+1)
&=& \tanh \Biggl( \sum_{k' \in \mathcal{M}(i)} \tanh^{-1} \hat{m}_{ki}(t) \nonumber \\
& & \qquad\qquad + \tanh^{-1} \gamma m_i(t) \Biggr), \end{aligned}$$ where $$\begin{aligned}
U_{ki}(t)
&=& \sum_{s_{i' \in \mathcal{L}(k) \backslash i}}
u_{{\boldsymbol{w}}} \Biggl( \sum_{i' \in \mathcal{L}(k) \backslash i} a_{ki'} s_{i'} \Biggr) \nonumber \\
& & \times \prod_{i' \in \mathcal{L}(k) \backslash i}
\frac{1+m_{i'k}(t)s_{i'}}2, \\
V_{ki}(t)
&=& \sum_{s_{i' \in \mathcal{L}(k) \backslash i}}
v_{{\boldsymbol{w}}} \Biggl( \sum_{i' \in \mathcal{L}(k) \backslash i} a_{ki'} s_{i'} \Biggr) \nonumber \\
& & \times \prod_{i' \in \mathcal{L}(k) \backslash i} \!\!\!\!
\frac{1+m_{i'k}(t)s_{i'}}2, \\
u_{{\boldsymbol{w}}}(x)
&=& \mathbb{I}( -w_2 < x < -w_1 ) + \mathbb{I}( w_1 < x < w_2) \nonumber \\
& & - \mathbb{I}(x<-w_2) - \mathbb{I}(w_2<x) \nonumber \\
& & - \mathbb{I}(-w_1<x<w_1), \\
v_{{\boldsymbol{w}}}(x)
&=& \mathbb{I}(x=-w_2) - \mathbb{I}(x=-w_1) \nonumber \\
& & + \mathbb{I}(x=w_1) - \mathbb{I}(x=w_2), \end{aligned}$$ and $\mathbb{I}(\mathcal{P})$ denotes an indicator function that takes 1 if the proposition $\mathcal{P}$ is true, and 0 otherwise. After $t_m$ iterations, the $i$ th bit of the codeword can be obtained as $\xi_i = {\,\mbox{\rm sgn}\,}[m_i(t_m)]$ using the mean of the pseudo marginal $m_i(t_m)$. To derive these iterative equations, we use the identity $\hat{x}_k = u_{{\boldsymbol{w}}} ( \sum_{i' \in \mathcal{L}(k) \backslash i} a_{ki'} s_{i'} )
+ a_{ki} s_i \; v_{{\boldsymbol{w}}} ( \sum_{i' \in \mathcal{L}(k) \backslash i} a_{ki'} s_{i'} )$, which holds for any $i \in \mathcal{L}(k)$.
The computational cost of the terms $U_{ki}(t)$ and $V_{ki}(t)$ is $O(2^C)$, which depends only on the row weights $C$, namely, it is $O(1)$ with respect to $N$. The complexity of this algorithm is therefore $O(N)$ when the number of iterations $t_m$ is fixed.
Utilizing eqs. (\[eq:condition1\]) and (\[eq:condition2\]) which are obtained in the large-row-weight-limit analysis, we can approximately set all parameters $C$, $w_1$, $w2$, and $\beta$ of our scheme with finite row weights except $\gamma$.
We first consider a setting of the parameter $\beta$. Using eq. (\[eq:condition1\]) and the rate-distortion function, we set $\beta$ as $\beta=\beta_c(p,R)$ for the given the source bias $p$ and the code rate $R$, where $\beta_c (p,R) = \ln ( {[h_2^{-1}(h_2(p)-R)]}^{-1} -1 )$. Here, $h_2^{-1}$ denotes the inverse function of the binary entropy function.
We next consider a setting of the parameters $C$, $w_1$, and $w2$. Each element of the vector $\mathcal{A} {\boldsymbol{s}}$ is the summation of $C$ Bernoulli random variables $1-2Ber(0.5)$, where ${\boldsymbol{s}} \in \{-1,1\}^N$ denotes a candidate of a codeword. This is a similar situation to the row weight limit. To keep the row weight finite, we restrict the row weight as $C \le C_{max}$. Using eq. (\[eq:condition2\]) and the rate-distortion function, we set $(C,w_1,w_2)$ as $(C,w_1,w_2) = \mathop{\rm argmin}_{(C',w_1',w_2') \in D(C_{max})} |\hat{K}(C',w_1',w_2') - K(p,R)|$, where $\hat{K}(C,w_1,w_2)$ $= \sum_{n\in\{0,\cdots,C\}: g_{{\boldsymbol{w}}}(C-2n)=-1} 2^{-C} (_{\, n}^C)$, $K(p,R) = [p-h_2^{-1}(h_2(p)-R)]/[1-2h_2^{-1}(h_2(p)-R)]$, and $D(C_{max})=\{ (C,w_1,w_2)| 2 \le C \le C_{max}, 0 < w_1 \le C-1, w_1<w_2 \le C+1 \}$ for the given $p$ and $R$. Note that the parameter that gives second smallest value might provide better performance.
Lastly, $\gamma$ is determined by trial and error. In this study, we choose $\gamma$ only within $\{0.2, 0.3, 0.4, 0.5\}$.
![ Empirical compression performance against the code rate $R$ for typical source bias $p$. The proposed algorithm (squares) and Hosaka’s algorithm (circles) are shown. The length of the orginal sequences is $N=420$, and all the measurements are averaged over 10 runs. The parameter $\gamma$ is chosen within $\{0.2, 0.3, 0.4, 0.5\}$. The row weight $C$ is chosen within $C \le 8$ ($C_{max}=8$). Top: $p \in \{0.6, 0.8\}$. Bottom: $p \in \{0.7, 0.9\}$. []{data-label="fig:performance"}](fig1a.eps "fig:"){width=".8\linewidth"}\
![ Empirical compression performance against the code rate $R$ for typical source bias $p$. The proposed algorithm (squares) and Hosaka’s algorithm (circles) are shown. The length of the orginal sequences is $N=420$, and all the measurements are averaged over 10 runs. The parameter $\gamma$ is chosen within $\{0.2, 0.3, 0.4, 0.5\}$. The row weight $C$ is chosen within $C \le 8$ ($C_{max}=8$). Top: $p \in \{0.6, 0.8\}$. Bottom: $p \in \{0.7, 0.9\}$. []{data-label="fig:performance"}](fig1b.eps "fig:"){width=".8\linewidth"}\
The empirical compression performance is shown in Fig. \[fig:performance\]. In this figure, the distortion averaged over $10$ runs is plotted as a function of the code rate $R$ for the source bias $p \in \{ 0.6, 0.7, 0.8, 0.9 \}$. The length of codewords is fexed at $N=420$, and the length of original sequence is adjusted. We here choose $C_{max}=8$ and $p>0.5$. In this figure, the time sharing bound is also shown. The time sharing bound is given by $R(D) = ( 1- \frac Dp) h_2(p)$, which denotes the compression performance achieved by the time sharing scheme of a lossless coding ($R(0)=h_2(p)$) and a trivial encoding that always outputs an all-one vector for any input ($R(p)=0$). It can be confirmed that the proposed linear complexity compressor (with $O(N)$ complexity) has slightly better performance than Hosaka’s algorithm (with $O(N^2)$ complexity).
We observed a not-so-good performance for the small-$p$ region that the source bias is less than about $0.2$. In this region $\mathrm{min} |\hat{K}(C',w_1',w_2')-K(p,R)|$ is not much smaller than that of the large-$p$ region. When we compress the original sequence ${\boldsymbol{x}}$ of which bias is $p<0.5$, we can first flip it as $-{\boldsymbol{x}}$ and then compress. The information for determining whether the sequence flips requires one bit. To reduce $\mathrm{min} |\hat{K}(C',w_1',w_2')-K(p,R)|$, it might be helpful to introduce more complex nonlinear functions.
In this study, we have proposed a scheme using a nonlinear function and a sparse matrix, and as well as a linear complexity message passing compressor based on the inertia-term-introduced belief propagation. The proposed method can treat redundant memoryless sources and has near-optimal compression performance for moderate block lengths. The adjustment of the column weight distribution of the sparse matrix might enable us to improve the compression performance. The analysis of this scheme with finite row weights is one of our future studies.
The author would like to thank Tadaaki Hosaka, Tatsuto Murayama, and Jun Muramatsu for their valuable comments and Ankit Gupta, Sergio Verdú, and Tsachy Weissman for providing him valuable preprints. This work was partially supported by a Grant-in-Aid for Scientific Research (C) No. 22500136 from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan.
[99]{}
T. M. Cover and J. A. Thomas: [*Elements of Information Theory, 2nd ed.*]{}, John Wiley and Sons Inc., 2006.
I. Csiszár and J. Körner: [*Information Theory: Coding Theorems for Discrete Memoryless Systems*]{}, Academic Press, 1981.
H. Nishimori: Statistical Physics of Spin Glasses and Information Processing – An Introduction, Oxford University Press, 2001.
Y. Matsunaga and H. Yamamoto: [*IEEE Trans. Inform. Theory*]{}, vol. 49, 2225, 2003.
A. G. Dimakis, M. J. Wainwright, and K. Ramchandran: [*Info. Theory Workshop*]{}, 650, 2005.
E. Martinian and M. J. Wainwright: [*Workshop on Information Theory and its Applications*]{}, 2006.
M. J. Wainwright: [*IEEE Signal Processing Magazine*]{}, 47, Sept. 2007.
T. Murayama and M. Okada: [*J. Phys. A: Math. Gen.*]{}, vol. 36, 11123, 2003.
T. Hosaka, Y. Kabashima, and H. Nishimori: [*Phys. Rev. E*]{}, 66, 066126, 2002.
T. Hosaka and Y. Kabashima: [*Physica A*]{}, 365, 113, 2006.
S. Ciliberti and M. Mézard: [*J Stat. Mech.*]{}, vol. 3, 58, 2006.
S. Ciliberti, M. Mézard, and R. Zecchina: [*Complex Syst. Methods*]{}, vol. 3, 58, 2006.
K. Mimura and M. Okada: [*Phys. Rev. E*]{}, 74, 026108, 2006.
A. Gupta and S. Verdu: [*IEEE Trans. Inform. Theory*]{}, vol. 55, 1961, 2009.
S. Miyake and J. Muramatsu: [*IEICE Trans. Fundamentals*]{}, E91-A, 1488, 2008.
J. Muramatsu: [*Proc. 2008 IEEE Int’l. Sympo. Info. Theory (ISIT2008)*]{}, 424, 2008.
F. Cousseau, K. Mimura, T. Omori, and M. Okada: [*Phys. Rev. E*]{}, 78, 021124, 2008.
F. Cousseau, K. Mimura, and M. Okada: [*Proc. 2008 IEEE Int’l. Sympo. Info. Theory (ISIT2008)*]{}, 509, 2008.
K. Mimura: [*J. Phys. A: Math. Theor.*]{}, 42, 135002, 2009.
T. Wadayama, [*Proc. of 3rd Int’l Sympo. Turbo Codes and Relatd Topics*]{}, 231, 2003.
T. Murayama, [*Phys. Rev. E*]{}, vol. 69, 035105(R), 2004.
M. J. Wainwright and E. Maneva, [*Proc. 2008 IEEE Int’l. Sympo. Info. Theory (ISIT2008)*]{}, 1493, 2005.
A. Gupta, S. Verdu, and T. Weissman, [*Proc. 2008 IEEE Int’l. Sympo. Info. Theory (ISIT2008)*]{}, 847, 2008.
S. Miyake, J. Honda, and H. Yamamoto, [*Proc. 2008 IEEE Int’l. Sympo. Info. Theory and its Applications (ISITA2008)*]{}, 589, 2008.
S. Jalali and T. Weissman, [*Proc. 2008 IEEE Int’l. Sympo. Info. Theory (ISIT2008)*]{}, 852, 2008.
K. Mimura, F. Cousseau, and M. Okada, [*J. Phys. Soc. Jpn.*]{}, 80, 3, 034802, 2011.
W. Krauth and M. Mézard, [*J. Phys. France*]{}, 50, 3057, 1989.
[^1]: E-mail address: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce the notion of directed diagrammatic reducibility which is a relative version of diagrammatic reducibility. Directed diagrammatic reducibility has strong group theoretic and topological consequences. A multi-relator version of the Freiheitssatz in the presence of directed diagrammatic reducibility is given. Results concerning relative asphericity and $\pi_1$-injectivity of subcomplexes are shown. We generalize the Corson/Trace characterization of diagrammatic reducibility to directed diagrammatic reducibility. We compare diagrammatic reducibility of relative presentations to directed diagrammatic reducibility. Classical tools for showing diagrammatic reducibility, such as the weight test, the max/min test, and small cancellation techniques are adapted to directed diagrammatic reducibility. The paper ends with some applications to labeled oriented trees.'
author:
- Jens Harlander and Stephan Rosebrock
title: Directed diagrammatic reducibility
---
Keywords: Diagrammatic reducibility, asphericity, 2-complexes, group-presentations, weight test\
MSC 2010: 57M20, 57M05, 57M35, 20F06
Introduction
============
A 2-complex $K$ is called [*aspherical*]{} if $\pi_2(K)=0$. If $L$ is a subcomplex of a 2-complex $K$ then the pair $(K,L)$ is called [*aspherical*]{} if $\pi_2(K,L)=1$. From the long homotopy sequence for pairs $$\ldots \to \pi_2(L)\to \pi_2(K)\to \pi_2(K,L)\to \pi_1(L)\to \pi_1(K)\to \ldots$$ we see that $\pi_2(K,L)=1$ if and only if the inclusion induced map $$\pi_2(L)\to \pi_2(K)$$ is surjective and the inclusion induced map $$\pi_1(L)\to \pi_1(K)$$ is injective. Asphericity is central to many aspects in low dimensional topology and combinatorial group theory.
There are combinatorial versions of asphericity, the strongest being diagrammatic reducibility, or DR for short. Recall that a map between 2-complexes is called [*combinatorial*]{} if it maps open cells homeomorphically to open cells. For a 2-complex $K$, a combinatorial map $f\colon C \to K$ is called a [*spherical diagram*]{}, if $C$ is a 2-sphere with cell structure. A 2-complex $K$ is [*diagrammatically reducible*]{}, DR for short, if every spherical diagram $f\colon C \to K$ contains an edge $e$ so that the 2-cells in $C$ that contain $e$ in their boundary map to the same 2-cell in $K$ with opposite orientations by folding over $e$. The edge $e$ is called a [*folding edge*]{} in $C$. Diagrammatic reducibility is a 2-dimensional version of free reductions of cycles in a graph. It was first considered by Sieradski [@Sie83]. See Gersten [@Ger87] for a good overview of diagrammatic techniques with applications to combinatorial topology and group theory.\
A presentation $P$ defines a 2-complex $K(P)$ and a group $G(P)$ which is the fundamental group of $K(P)$ by Van Kampen’s theorem. A presentation $P$ is called DR if its associated 2-complex $K(P)$ is DR. For a set $X$ call a subset $S$ [*proper*]{} if $S\ne X$ ($S$ may be empty). If $S$ is a proper subset of the set of generators of $P$ let $P_S$ be the sub-presentation of $P$ carried by $S$. We say [*$P$ is DR directed away from $S$*]{} if every spherical diagram $f\colon C\to K(P)$ that is not already a diagram over $K(P_S)$ contains a folding edge that is mapped to an edge in $K(P)$ that is not an edge in $K(P_S)$.\
Here is an outline of the paper. In Section \[sec:basics\] we show that if $P$ is DR directed away from $S$ then $\pi_2(K(P),K(P_S))=1$. In particular the inclusion $K(P_S)\to K(P)$ is $\pi_1$-injective and $\pi_2$-surjective. We provide basic examples and give, as our first application, a multi-relator Freiheitssatz for presentations $P$ that are DR in all directions. In Section \[sec:corsontrace\] we characterize directed diagrammatic reducibility in terms of finite subcomplexes of $\tilde K(P)$, the universal covering of $K(P)$. In the standard DR setting this was done by Corson and Trace [@CT00]. In Section \[sec:relpres\] diagrammatic reducibility of relative presentations introduced by Bogley and Pride in [@BP92] is compared with directed diagrammatic reducibility. We show that the two concepts are equivalent. However, the viewpoints are very different. In the Bogley/Pride setting, the starting point is a relative presentation where the generators are a free product $H*F({\bf x})$, where $H$ is fixed. We start with an ordinary presentation on generators ${\bf x}$ and consider DR directed away from a proper subset $S\subseteq {\bf x}$, where $S$ is not necessarily fixed but can vary over ${\bf x}$. This enables us to consider the idea of diagrammatic reducibility in all directions. We can formulate a natural multi-relator Freiheitssatz in Corollary \[cor:freiheit\]. Retaining combinatorial information about the subpresentation $P_S$ carried by the subset $S$ of generators of $P$ rather than only keeping its fundamental group and passing to a relative presentation, might give an advantage in designing tools for detecting DR in the two settings. Such tools, a max/min and a weight test for showing directed diagrammatic reducibility are given in Section \[sec:tests\]. Theorem \[s44\] and the examples we exhibit at the end of Section \[sec:appl\] might have been overlooked from a purely relative presentation point of view.\
We conclude this introduction with some remarks about diagrams. In this paper we consider surface diagrams where the surface can be different from a sphere. A [*surface diagram over a 2-complex $K$*]{} is a combinatorial map $f\colon F\to K$, where $F$ is a compact surface, possibly with boundary, with a cell-structure. If $c$ is a cell in $F$, we call $f(c)$ the [*label*]{} on $F$. The labeled cell-complex $F$ contains all the information of the combinatorial map $f$, and we sometimes refer to the labeled $F$ as a surface diagram over $K$. However, keeping the map point of view is useful as we see in Section \[sec:corsontrace\].
We are primarily concerned with 2-complexes and groups defined by a finite presentation $$P=\langle x_1,\ldots ,x_n\mid r_1,\ldots ,r_m\rangle.$$ Recall that $K(P)$ is the 2-complex with one vertex, edges and 2-cells in one-to-one correspondence with the generators and relators respectively. The 2-cell $\Delta(r)$ for the relator $r$ is attached to the one skeleton according to the word $r$. The universal cover $\tilde K(P)$ has as vertex set the elements of $G(P)$, oriented edges $(g,x_i)$, $g\in G(P)$, $1\le i \le n$, and 2-cells $(g,r_j)$, $g\in G(P)$, $1\le j \le m$. The 1-skeleton of $\tilde K(P)$ is also referred to as the Cayley graph of $G(P)$ associated with the set of generators. The boundary of $(g,r_j)$ is the lift of the path $r_j$ to the Cayley graph, starting (and ending) at $g$.
Directed diagrammatic reducibility {#sec:basics}
==================================
As was mentioned in the introduction, a 2-complex $K$ is called diagrammatically reducible (or DR for short) if each spherical diagram $f\colon C\to K$ contains a folding edge.
Let $P=\langle x_1,\ldots ,x_n\mid r_1,\ldots ,r_m\rangle$ be a presentation and $S$ a proper subset (possibly empty) of the set of generators. We say that $P$ is
- [*DR directed away from $S$*]{} if every spherical diagram $f\colon C\to K(P)$ that contains an edge with label not from $S$ also contains a folding edge with label not from $S$;
- [*DR in all directions*]{} if every spherical diagram $f\colon C\to K(P)$ that contains an edge labeled $x_i$ also contains a folding edge with label $x_i$, $i=1,\ldots ,n$. Note that this implies that $P$ is DR directed away from all proper subsets $S$.
If $S=\emptyset$ then DR directed away from $S$ simply means DR.\
[**Examples and Comments:**]{}\
1) Just like DR, directed DR is a hereditary property: Let $T$ be a sub-presentation of a presentation $P$ and let $S$ be a subset of the set of generators of $T$. Then if $P$ is DR directed away from $S$ so is $T$. If $P$ is DR in all directions, then so is $T$.
However, a presentation $P$ can be DR directed away from $S$ but not DR directed away from some $S_0\subset S$. Here is an example. Suppose that $$P=\langle a, b, c\mid aba^{-1}b^{-2}, bab^{-1}a^{-2}, cab\rangle.$$ Note that $G(P)$ is the trivial group. $P$ is DR away from $S=\{a,b\}$. This is because if a spherical diagram $f\colon C\to K(P)$ contains an edge $e$ with label not in $S$, then the label on $e$ is $c$. Since $c$ is a free edge in $K(P)$, this implies that $e$ is a folding edge in $C$. Note that $P$ is not DR directed away from $S_0=\{ a \}$ by Theorem \[RelPiS\] (2) below. The presentation is also not DR because the subpresentation $P_0=\langle a, b, c\mid aba^{-1}b^{-2}, bab^{-1}a^{-2}\rangle$ is not DR. The 2-complex $K(P_0)$ is simply connected but does not have a free edge (see Corson/Trace [@CT00], and also Section \[sec:corsontrace\] in this article).\
2) Subdivisions preserve DR but not directed DR. Consider $P=\langle a, b \mid aba^{-1}b^{-1}\rangle$ and $P'=\langle a, b, c \mid abc, c^{-1}a^{-1}b^{-1}\rangle$. Then $P$ is DR in all directions (see the next example) but $P'$ is not DR directed away from $S=\{ a, b\}$: the two triangles coming from the relators can be glued together to form a disc-diagram that contains an edge labeled $c$ but not a folding edge labeled $c$. Thus $P'$ is not DR directed away from $S$ by Theorem \[RelPiS\].\
3) Direct products of graphs are DR in all directions. Consider $$P=\langle x_1,\ldots ,x_m, y_1,\ldots ,y_n \mid [x_i,y_j], 1\le i\le m, 1\le j\le n \rangle.$$ The presentation $P$ is DR in all directions. Let $f\colon C\to K(P)$ be a spherical diagram that contains an edge labeled by $y_k$. We will show that there is a folding edge labeled by $y_k$. If $d$ is a 2-cell in $C$ with two of its edge labels $y_k$, then draw a red line across $d$ connecting the midpoints of these two edges. The red graph in $C$ we draw in this way consists of red circles. Take one of these red circles and consider the gallery of 2-cells that comes with it. This gallery forms an annulus $A$, the two boundary circles are labeled by a word $w$ in the $x$’s. Edges connecting the two boundary circles are all labeled by $y_k$. Since the subgroup of $G(P)$ generated by $x_1,\ldots ,x_m$ is free, the word $w$ must contain a cancelling pair $x_i^{\epsilon}x_i^{-\epsilon}$. Thus the annulus $A$ contains a cancelling pair of 2-cells labelled by $[x_i,y_k]^{\epsilon} [x_i,y_k]^{-\epsilon}$ with a folding edge labeled by $y_k$. The argument for an $x_k$ is analogues.\
4) Orientable closed surfaces are DR in all directions. Let $$P=\langle x_1, x_2,\ldots ,x_{2g-1}, x_{2g}\mid [x_1,x_2]\ldots [x_{2g-1}, x_{2g}]\rangle$$ be the standard presentation of the fundamental group of an orientable surface of genus $g$. Then $P$ is DR in all directions. Let $f\colon C\to K(P)$ be a spherical diagram that contains an edge labeled by $x_1$, say. We will show that there is a folding edge labeled by $x_1$. Just as in the previous example, $C$ contains an annulus $A$ with the two boundary components labeled by words $u$ and $v$ in $\{ x_2,\ldots ,x_{2g}\}$. Edges connecting the boundary components are all labeled with $x_1$. Since the subgroup of $G(P)$ generated by $x_2,\ldots ,x_{2g}$ is free, the words $u$ and $w$ must contain a cancelling pair of letters. Hence the annulus $A$ contains a cancelling pair of 2-cells with folding edge labeled by $x_1$.\
Let $P=\langle x_1,\ldots ,x_n\mid r_1,\ldots ,r_m\rangle$ be a presentation and $S$ a subset of the generators. Let $P_S$ be the subpresentation of $P$ carried by $S$.
\[RelPiS\] Suppose that $P$ is DR directed away from the proper subset $S$ of the generators. Then every disc diagram $g\colon D\to K(P)$ with boundary labeled by a word in $S$, that contains a label not from $S$, has a folding edge with label not from $S$. Consequently
1. $\pi_2(K(P),K(P_S))=1$;
2. The inclusion $K(P_S)\to K(P)$ is $\pi_1$-injective;
3. The inclusion $K(P_S)\to K(P)$ is $\pi_2$-surjective;
Proof. Suppose $g\colon D\to K(P)$ is a disc diagram as in the statement of the theorem. If $S\ne \emptyset$ glue $D$ to $-D$ to obtain a spherical diagram $g'\colon C\to K(P)$ containing a label not from $S$. If $S=\emptyset$ then $D$ is already a spherical diagram and $D=C$. Since $P$ is DR away from $S$ this spherical diagram contains a folding edge with label not from $S$. Thus the original disc diagram has to contain a folding edge with label not from $S$. In particular every reduced disc diagram with boundary label a word in $S$ is already a disc diagram over $P_S$. This directly implies (2) and (1). The statement (3) follows from (1) by the long homotopy sequence of pairs.
We recall the Freiheitssatz for 1-relator groups: Suppose $P=\langle x_1,\ldots ,x_n\mid r \rangle$ is a 1-relator presentation, where $r$ is a cyclically reduced word that contains all the generators. Then any proper subset $S$ of $\{x_1,\ldots ,x_n\}$ generates a free subgroup of $G(P)$ with basis $S$. The following three results should be viewed as multi-relator versions of this celebrated result.
Let $P=\langle x_1,\ldots ,x_n\mid r_1,\ldots ,r_m\rangle$ be a presentation with cyclically reduced relators. Assume that $P$ is DR in all directions. Then the inclusion induced homomorphism $G(P_S)\to G(P)$ is injective for every subset $S$ of the generators.
Proof. If $S$ is the set of generators of $P$ then $G(P_S)=G(P)$ and the statement is true. If $S$ is a proper subset of the set of generators then the result follows from Theorem \[RelPiS\] using the fact that $P$ is DR directed away from $S$.
Let $P=\langle x_1,\ldots ,x_n\mid r_1,\ldots ,r_m\rangle$ be a presentation with cyclically reduced relators and $S$ a proper subset of the generators. Assume that each $r_i$ contains a generator not from $S$. If $P$ is DR directed away from $S$, then $S$ generates a free subgroup of $G(P)$ with basis $S$.
Proof. Since each $r_i$ is a reduced word that contains a generator not from $S$ we have that $P_S=\langle S\mid\ \rangle $ and $G(P_S)$ is free. Now Theorem \[RelPiS\] (2) gives the desired result.\
\[cor:freiheit\] Let $P=\langle x_1,\ldots ,x_n\mid r_1,\ldots ,r_m\rangle$ be a presentation where each $r_i$ is a reduced word that contains all the generators. If $P$ is DR in all directions, then any proper subset $S$ of $\{x_1,\ldots ,x_n\}$ generates a free subgroup of $G(P)$ with basis $S$.
Corson/Trace for directed diagrammatic reducibility {#sec:corsontrace}
===================================================
If $K(P)$ is the 2-complex build from the presentation $P$ let $\tilde K(P)$ be its universal cover and $\tilde K(P)^{(1)}$ the 1-skeleton of the universal cover. In [@CT00] Corson and Trace have shown the following result:
\[thm:classicCT\] A presentation $P$ is DR if and only if every finite subcomplex of $\tilde K(P)$ collapses into the 1-skeleton $\tilde K(P)^{(1)}$.
In this section we retell the Corson/Trace story in the context of directed DR.
\[RedSurDiaS\] Let $P$ be a finite presentation and let $S_0$ be the subset of the generators which correspond to the free edges in $K(P)$. Let $x$ be a generator contained in a relator but not in $S_0$. Then there exists a closed oriented surface diagram $f\colon F\to K(P)$ that contains an edge labeled by $x$, which is reducible only at edges labeled by elements from $S_0$.
Proof. To avoid an excessive amount of subscripts we use letters $a, b, c,\ldots $ for the set of generators in $P$. For every relator $r$ draw two $|r|$-gons $\Delta(r)$ and $\Delta(r^{-1})$ in the plane. Orient the edges of these $|r|$-gons and label them according to the relator $r$, using subscripts. For example if $r=abab^{-1}ac$ then label the edges of $\Delta(r)$ clockwise by $$a_1(r),b_1(r),a_2(r),b_2^{-1}(r),a_3(r),c_1(r)$$ and $\Delta(r^{-1})$ counter-clockwise by $$a_1(r^{-1}),b_1(r^{-1}),a_2(r^{-1}),b_2^{-1}(r^{-1}),a_3(r^{-1}),c_1(r^{-1}).$$ Note that the number of edges on the $2m$ discs $\Delta(r_i)$, $\Delta(r_i^{-1})$, $i=1,...,m$, with labels of the form $a_{*}(*)$ is even, equal to 2 if $a$ is a free edge of $K(P)$, otherwise greater or equal to $4$. The same holds true for every other generator letter. We now match edges of the $2m$ discs in pairs. If $a\in S_0$ is a free edge that occurs only once among the set of relators, say in $r$, then match $a_1(r)$ and $a_1(r^{-1})$. If $a$ is not a free edge then match $a_i(r_k)$ with any $a_j(r_l^{\pm 1})$, but avoid the match with $a_i(r_k^{-1})$. One concrete way to do this is the following: Define $$\begin{aligned}
L_+(a) & : & \ a_1(r_1), a_2(r_1),\ldots ,a_1(r_2), a_2(r_2),\ldots ,a_1(r_3),a_2(r_3),\ldots\\
L_{-}(a) & : & \ a_1(r_1^{-1}), a_2(r_1^{-1}),\ldots ,a_1(r_2^{-1}), a_2(r_2^{-1}),\ldots ,a_1(r_3^{-1}),a_2(r_3^{-1}),\ldots\end{aligned}$$ Cyclically match item $N$ on the $L_+(a)$ list with item $N+1$ on the $L_-(a)$ list. Proceed in this fashion with all generators $a, b, c,\ldots$. This way we build a surface, possibly with boundary, possibly consisting of more than one component. Let $F_0$ be a component that contains an edge $x_i(r_k)$. If we erase from the edge labels all information except the letter (for example $a_i(r_j^{\pm 1})$ becomes $a$), we obtain a closed surface diagram $f_0\colon F_0\to K(P)$ that contains the label $x$. Passing to the 2-fold orientable covering $F$ of $F_0$ we obtain a closed orientable surface diagram $f\colon F\to K(P)$ that contains the label $x$. By construction $F$ is reducible only at edges with labels from $S_0$.\
Here is a topological version of this lemma.
\[TopRedSurDiaS\] Let $X$ be a 2-complex and let $E_0$ be the set of boundary edges of $X$. Let $e$ be a fixed edge that is part of a 2-cell of $X$ and is not a boundary edge. Then there exists a closed oriented surface diagram $F\to X$ so that $F$ contains an edge labeled by $e$ and $F$ is reducible only at edges which carry labels from $E_0$.
Let $P$ be a presentation and $w$ be a word in the generators of $P$ that represents the trivial element in $G(P)$. Then there exists a Van Kampen digram $D_{w}\to K(P)$ with boundary word $w$. This is a combinatorial map where $D_{w}$ is a [*tree of discs*]{}. Here is the Corson/Trace result for directed DR. The original Theorem \[thm:classicCT\] is covered by the case $S=\emptyset$.
\[thm:GenCT\]The presentation $P$ is DR directed away from $S$ if and only if every finite subcomplex $X \subseteq \tilde K(P)$ can be collapsed into $p^{-1}(K(P_S))\cup \tilde K(P)^{(1)}$.
Proof. Suppose first that $P$ is DR directed away from $S$. Let $X$ be a finite subcomplex of $\tilde K(P)$ that is not already contained in $p^{-1}(K(P_S))\cup \tilde K(P)^{(1)}$. We may assume every edge of $X$ is part of a 2-cell. It suffices to show that $X$ has a boundary edge that is not contained in $p^{-1}(K(P_S))$. We will show this by contradiction. Let $\tilde E_0$ be the set of boundary edges of $X$ and assume that $\tilde E_0\subseteq p^{-1}(K(P_S))$. Then $p(\tilde E_0)=S_0\subseteq S$. By Lemma \[TopRedSurDiaS\] there exists a closed oriented surface diagram $\tilde f\colon F\to X\subseteq \tilde K(P)$ so that $\tilde f(F)$ is not contained in $p^{-1}(K(P_S))$ and folding edges in $F$ (should there be any) are mapped to $E_0$. We consider the composition $$f\colon F\to X\subseteq \tilde K(P) \stackrel{p}\to K(P).$$ Note that $f\colon F\to K(P)$ is a closed oriented surface diagram so that $f(F)$ is not contained in $K(P_S)$ and folding edges in $F$ (should there be any) are labeled by elements from $S_0\subseteq S$. Since $\tilde f$ is a lift of $f$, closed edge paths in $F$ are labeled by words $w$ in the generators of $P$ that represent the trivial element in $G(P)$. We now proceed as in the proof of Lemma 2.1 of Corson/Trace [@CT00]. Choose a complete set of closed cutting curves $\gamma_1,\ldots,\gamma_k$ in $F$ (cutting along all $\gamma_i$ would transforms $F$ into a simply-connected region of the plane). Since the $f(\gamma_i)$ are trivial in $G(P)$ we have a tree of disc-diagram $D_{\gamma_i}$ for each $\gamma_i$. Form a 2-complex $L$ by attaching $D_{\gamma_i}$ to $F$ for every $i$. We have
1. $L\subseteq S^3$ is a simply connected 2-skeleton of a cell decomposition of the 3-sphere $S^3$;
2. we have combinatorial maps $F\stackrel{\alpha}{\rightarrow}L\stackrel{\beta}{\rightarrow}K(P)$ so that $f=\beta\circ \alpha$.
Let $L'$ be a 2-complex that has the minimal number of 2-cells among all 2-complexes that satisfy the two conditions just stated. Note that $L'$ can not contain a 2-cell $d'$ with a free edge $e'$ with label $x\notin S$. Suppose it does. Then there is no edge $e$ in $F$ so that $\alpha(e)=e'$, because such an edge $e$ would be a folding edge in $F$ with label $x\notin S$, which we know does not exist. Thus we can collapse $d'$ in $L'$ and contain a 2-complex that satisfies our two conditions and contains fewer 2-cells than $L'$. A contradiction to minimality. We should note that we do not rule out free edges in $L'$ with label in $S$.
Consider a spherical diagram $h'\colon C\to L'$ that arises as an attaching map of a 3-cell of $S^3$ so that the composition $h=\beta\circ h'\colon C\to K(P)$ does not entirely map into $K(P_S)$. Such an attaching map exists because otherwise $\beta(L')$ and hence $f(F)$ would be entirely contained in $K(P_S)$, which is not the case. Note that the spherical diagram $h\colon C\to K(P)$ does not contain a folding edge with label $x\notin S$. If it does, a fold could be performed in $L'$ to obtain a 2-complex that satisfies our conditions and contains fewer 2-cells than $L'$. A contradiction to minimality. The existence of the diagram $h\colon C\to K(P)$ contradicts the hypothesis that $P$ is DR directed away from $S$. So the assumption $p(E_0)=S_0\subseteq S$ is false.
Now we prove the other direction. Let $f\colon C\to K(P)$ be a spherical diagram so that $f(C)$ is not contained in $K(P_S)$. Let $\tilde f\colon C\to \tilde K(P)$ be a lift and let $X=\tilde f(C)$. Then $X$ is not contained in $p^{-1}(K(P_S))\cup \tilde K(P)^{(1)}$ and hence contains a boundary edge $\tilde e$ not in $p^{-1}(K(P_S))$. The edge $e$ in $C$ such that $\tilde f(e)=\tilde e$ is a folding edge in $f\colon C\to K(P)$ and $f(e)=x\notin S$.
Suppose $P$ presents a finite group. Then $P$ is $DR$ directed away from $S$ if and only if $K(P)$ collapses into $K(P_S)$.
Proof: Since we assume that $G(P)$ is finite $\tilde K(P)$ is a finite complex. By Theorem \[thm:GenCT\] $P$ is DR directed away from $S$ if and only if $\tilde K(P)$ collapses into $p^{-1}(K(P_S))\cup \tilde K(P)^{(1)}$. A collapsing sequence in the universal cover gives a collapsing sequence downstairs that collapses $K(P)$ into $K(P_S)\cup K(P)^{(1)}$. Since $K(P)^{(1)}$ is a wedge of circles and $G(P)$ is assumed to be finite, we have collapsed $K(P)$ into $K(P_S)$.\
Relative presentations {#sec:relpres}
======================
In this section we compare diagrammatic reducibility of relative presentations to directed diagrammatic reducibility. The concept of asphericity for relative presentations was introduced by Bogley and Pride [@BP92]. See also Bogley, Edjvet, Williams [@BEW18] for a recent survey.\
A [*relative presentation*]{} $\hat P=\langle H, {\bf x}\mid {\bf \hat r} \rangle$ consists of a group $H$, a generating set ${\bf x}$ and relator set ${\bf \hat r}\subseteq H*F({\bf x})$. Let $\bar P$ be the presentation obtained from $\hat P$ by forgetting all the $H$-information. That is $\bar P=\langle {\bf x}\mid \bar {\bf r} \rangle$ where $\bar r\in \bar {\bf r}$ is the image of $\hat r\in {\bf \hat r}$ under the projection $H*F({\bf x})\to F({\bf x})$. The 2-complex $K(\hat P)$ is the 2-complex $K(\bar P)$ with corners marked by elements from $H$: If $\hat r=x_{1}h_1x_{2}h_2\ldots x_{t}h_t\in {\bf \hat r}$, where $x_{i}\in {\bf x}^{\pm 1}$ and $h_i\in H$, then $\bar r=x_{1}\ldots x_{t}$ and we mark the oriented corner from the $i$-th edge to the $(i+1)$-th edge in the boundary of the 2-cell $\Delta(\bar r)$ (oriented clockwise) by $h_i$. An example is given in Figure \[arelrel\].
In a surface diagram $\hat f\colon \hat F \to K(\hat P)$ the corners in the 2-cells are marked by elements from $H$. The corner marking is obtained by pulling back the corner marking of the cells of $K(\hat P)$. We can assign group elements $h(v)$ to the vertices $v$ in the diagram in the following way: If $c_1(v)\ldots c_l(v)$ is the clockwise corner cycle (or path, in case $v$ is a boundary vertex) at the vertex $v$, and $c_i(v)$ is marked with $h_i$, then $h(v)=h_1\ldots h_l$. Note that $h(v)$ is defined only up to cyclic permutation in case $v$ is an interior vertex.
(0,0) circle (2pt); (4,0) circle (2pt); (2,3.2) circle (2pt);
(4,0) – (0,0) node\[midway, below\][$x_1$]{}; (2,3.2) – (0,0) node\[midway, left\][$x_2$]{}; (2,3.2) – (4,0) node\[midway, right\][$x_3$]{};
at (0.5,0.25) [$h_1$]{}; at (3.5,0.25) [$h_3$]{}; at (2,2.65) [$h_2$]{};
(0.47,0.7) arc (62:0:0.8); (1.6,2.5) arc (239:301:0.8); (3.1,0) arc (180:120:0.8);
A spherical diagram $\hat f\colon \hat C \to K(\hat P)$ is [*admissible*]{} if $h(v)=1$ for all but possibly one vertex $v_0$ of $\hat C$. A relative presentation is defined to be DR if every admissible spherical diagram has a folding edge.
*Bogley and Pride [@BP92] defined DR for relative presentations in terms of pictures instead of diagrams. A picture is a connected planar graph labeled in a certain way. It gives rise to a cell structure of a 2-sphere whose dual yields a spherical diagram. The regions in a picture correspond to the vertices of the spherical diagram. Given a picture, a connected planar graph, the outer region is special and singles out one vertex in the corresponding spherical diagram. The special nature of the outer region is lost when passing to the diagram.*
Given a relative presentation $\hat P=\langle H, {\bf x}\mid {\bf \hat r} \rangle$ one can construct its [*cellular model*]{}, a cell complex whose homotopy type is uniquely determined by $\hat P$. Choose a presentation $P_0=\langle {\bf x_0}\mid {\bf r_0} \rangle $ for $H$. For each $\hat r=x_{1}h_1x_{2}h_2\ldots x_{t}h_t\in {\bf \hat r}$ choose an element $r=x_{1}u_1x_{2}u_2\ldots x_{t}u_t$, where the $u_i$ are words in ${\bf x_0}^{\pm 1}$ so that $r$ maps to $\hat r$ under $F({\bf x_0}\cup{\bf x})\to H*F({\bf x})$. Let ${\bf r}$ be the set of words obtained in this fashion from the set ${\bf \hat r}$ and define $P=\langle {\bf x_0}\cup {\bf x}\mid {\bf r_0}\cup {\bf r} \rangle $. Note that $K(P_0)$ is a subcomplex of $K(P)$. Attach higher dimensional cells to $K(P_0)\subseteq K(P)$ to arrive at a complex $L(\hat P)$ that contains a $K(H,1)$ complex (the Eilenberg-MacLane space) as a subcomplex. $L(\hat P)$ is a cellular model of $\hat P$.
If $K$ is a 2-complex a [*tree-disc diagram*]{} is a combinatorial map $f\colon E\to K$ where $E$ is a simply-connected planar region with one boundary component consisting of discs joined by arcs.
\[lemma:shrinkblow\] (1) Assume $H=G(P_0)\to G(P)$ is injective. If there exists a spherical diagram $f\colon C\to K(P)$ that carries the label $x\in {\bf x}$ and all of whose folding edges carry labels from ${\bf x_0}$, then there exists a reduced admissible spherical diagram $\hat f\colon \hat C\to K(\hat P)$ that carries the label $x$.\
(2) Conversely, in case $H\to G(\hat P)$ is injective, an admissible reduced spherical diagram over $K(\hat P)$ that carries the label $x$ gives rise to a spherical diagram over $K(P)$ that carries the label $x$ and whose folding edges (in case there are such) all carry labels from ${\bf x_0}$.
Proof. We prove (1). We first construct a subdivision of $K(P)$. Consider a clockwise oriented 2-cell $\Delta(r)$, where $r=x_1u_1x_2u_2\ldots x_ku_k\in {\bf r}$, $x_i\in { \bf x}^{\pm 1}$ and the $u_i$ are words in ${\bf x_0}^{\pm 1}$ (which can be empty). Connect the endpoint of the edge labeled $x_l$ with the beginning of the edge labeled $x_{l+1}$ ($l$ mod $k$) by an oriented red corner edge $c_l(r)$ and label it with $h_l(r)$, the element of $G(P_0)$ represented by $u_l$. Denote by $K(P)'$ this subdivided version of $K(P)$.
Assume $f\colon C\to K(P)$ is a spherical diagram as in the statement. We can subdivide $C$ to obtain a spherical diagram $f'\colon C'\to K(P)'$. The red corner edges in $C'$ form a directed red graph $\Gamma$ that consists of circles. Let $F$ be the closure of the component of $C'-\Gamma$ that contains an edge with label $x$. Note that $F$ is a planar surface with red boundary, made up of red corner edges. All interior edges carry labels not in ${\bf x_0}$. If $c_1\ldots c_t$ is the edge loop that makes up a boundary circle in $F$, then $h_1\ldots h_t=1$ in $G(P_0)$, where $h_i$ is the element of $H$ written on the red corner edge $f'(c_i)$. This is because $c_1\ldots c_t$ is a loop in $C'$ that is homotopic to an edge loop in $C'$ with label $u_1\ldots u_t$, a word in ${\bf x_0}^{\pm 1}$, where $u_i$ represents $h_i$, and the assumption that $G(P_0)\to G(P)$ is injective. Now if we contract each boundary circle in $F$ to a point we obtain an admissible spherical diagram $\hat f\colon \hat C \to K(\hat P)$. It is reduced because $F$ does not contain folding edges, since we assume that no folding edge of $C$ carries a label outside of ${\bf x_0}$. This concludes the proof of (1).
We will now prove (2). Let $\hat f\colon \hat C\to K(\hat P)$ be an admissible reduced spherical diagram. Let $v$ be a vertex in $\hat C$. Note that the corners around $v$ are oriented and marked by elements from $H$. We remove a small open disc from $\hat C$ with center $v$. The boundary we create is made up of the oriented corners. Let $c$ be one of these corners and suppose it is contained in a 2-cell $d$ with label $\hat r$. Suppose $\hat r=x_1h_1x_2h_2\ldots x_kh_k\in {\bf \hat r}$, $x_i\in {\bf x}^{\pm 1}$, $h_i\in H$. Suppose further $c$ is the corner between edges labeled $x_l$ and $x_{l+1}$ with label $h_l$. We have a corresponding relator $r=x_1u_1x_2u_2\ldots x_ku_k\in {\bf r}$, and $u_l$ represents $h_l$. We subdivide $c$ into an edge path, orient and label edges by elements from $\bf x_0$ so that we obtain $u_l$ as the label on that path. We proceed in this manner at all corners at $v$, and all vertices $v\in \hat C$. We have built a reduced surface diagram $f_0\colon F\to K(P)$, where $F$ is a planar surface whose boundary components are labeled by words in ${\bf x_0}^{\pm 1}$ that represent elements $h(v)$ of $G(P_0)=H$. Since we assume all $h(v)=1$ in $G(\hat P)$ at all but possibly one vertex $v_0$ and $H\to G(\hat P)$ is injective, we have that $h(v)=1$ in $H$ for all $v\ne v_0$. We can cap the boundary components of the surface diagram that correspond to vertices $v\ne v_0$ by reduced tree-disc diagrams over $K(P_0)$ to obtain a disc diagram $f_1\colon D \to K(P)$. The boundary of $D$ is labeled by a word $u$ in ${\bf x_0}$ and hence represents the trivial element in $G(P)$. Since we assume that $H\to G(\hat P)$ is injective it follows $G(P_0)\to G(P)$ is injective. Thus $u$ represents the trivial element of $G(P_0)$. We cap off the remaining boundary component by reduced tree-disc diagrams over $K(P_0)$ and produce a spherical diagram $f\colon C\to K(P)$ that contains the label $x$. Since we started out with $\hat f\colon \hat C\to K(\hat P)$ reduced, folding edges in $f\colon C\to K(P)$ have to be edges on the boundary of $F$, and hence they all carry labels from ${\bf x_0}$.\
\[thm:equivalence\] Let $\hat P=\langle H, {\bf x}\mid {\bf \hat r} \rangle$ be a relative presentation and $P=\langle {\bf x_0}\cup {\bf x}\mid {\bf r_0}\cup {\bf r} \rangle $ be the presentation obtained from the 2-skeleton of a cellular model $L(\hat P)$. Then $\hat P$ is DR if and only if $P$ is DR directed away from ${\bf x_0}$.
Proof. Assume first that $P$ is DR directed away from ${\bf x_0}$. Assume we have an admissible reduced spherical diagram $\hat C \to K(\hat P)$. Since it is not empty it must carry a label $x\in {\bf x}$. Since $P$ is DR directed away from ${\bf x_0}$ we know from Theorem \[RelPiS\] (2) that $G(P_0)=H$ injects into $G(P)$ and thus $H$ injects into $G(\hat P)$. By Lemma \[lemma:shrinkblow\] (2) there exists a spherical diagram $C \to K(P)$ that carries the label $x\in {\bf x}$ whose folding edges (should there be such) all carry labels from ${\bf x_0}$. This contradicts the assumption that $P$ is DR away from ${\bf x_0}$.
Now assume that the relative presentation $\hat P$ is DR. It follows that $H\to G(\hat P)$ is injective (see statement (0.1) of [@BP92]). Let $C\to K(P)$ be a spherical diagram that carries an edge label $x\in {\bf x}$, all of whose folding edge labels are from ${\bf x_0}$. By Lemma \[lemma:shrinkblow\] (1) there exists an admissible reduced spherical diagram $\hat C\to K(\hat P)$ that carries the edge label $x$. But this contradicts the DR assumption.\
Any result about directed diagrammatic reducibility has a translation into a result about diagrammatically reducible relative presentations and vice versa. We will next give a version of our generalized Corson/Trace Theorem \[thm:GenCT\] for relative presentations. Let $L(\hat P)$ be a cellular model of a relative presentation $\hat P$ that is DR. Let $p\colon \tilde L(\hat P) \to L(\hat P)$ be the universal covering. Since $H\to G(\hat P)$ is injective, the preimage $p^{-1}(K(H,1))$ is a disjoint union of complexes, each being the universal covering of $K(H,1)$ and hence a disjoint union of contractible complexes. We contract each component of $p^{-1}(K(H,1))$ to a point and arrive at a simply connected 2-complex $Y$. Denote the quotient map by $q\colon \tilde L(\hat P) \to Y$.
The relative presentation $\hat P$ is DR if and only if every finite subcomplex of $Y$ collapses into $Y^{(1)}$.
This result is due to Bogley, Edjvet and Williams and can be found in the proof of Theorem 3.6 in [@BEW18]. It also follows directly from our results: Let $\hat P=\langle H, {\bf x}\mid {\bf \hat r} \rangle$ be a relative presentation and $P=\langle {\bf x_0}\cup {\bf x}\mid {\bf r_0}\cup {\bf r} \rangle $ be the presentation obtained from the 2-skeleton of a cellular model $L(\hat P)$. Now $\hat P$ is DR if and only if $P$ is DR directed away from ${\bf x_0}$ by Theorem \[thm:equivalence\], and that is the case if and only if every finite subcomplex of $\tilde K(P)$ collapses into $p^{-1}(K(P_0))\cup \tilde K(P)^{(1)}$ by Theorem \[thm:GenCT\]. It is not difficult to see that the last statement is equivalent to saying that every finite subcomplex of $Y$ collapses into $Y^{(1)}$. Note that $\tilde K(P)$ is the 2-skeleton of $\tilde L(\hat P)$, so the quotient map restricts to $q\colon \tilde K(P)\subseteq \tilde L(\hat P)\to Y$. If $Y'$ is a finite subcomplex of $Y$ then $q^{-1}(Y')$ might not be finite, but it is finite outside of $p^{-1}(K(P_0))$. So there exists a finite subcomplex $X$ of $\tilde K(P)$ so that $q(X)=Y'$. A collapsing strategy of $X$ yields one for $Y'$ and vice versa.\
We end this section with a simple lemma that will be useful for establishing directed DR in concrete situations. Let $\phi\colon H_1\to H_2$ be a group homomorphism and suppose $\hat P_1=\langle H_1, {\bf x}\mid {\bf \hat r}\rangle $ is a relative presentation. If $\hat r=x_1h_1x_2h_2\ldots x_kh_k$, where the $x_i\in {\bf x}^{\pm 1}$, $h_i\in H$, define $\phi(\hat r)=x_1\phi(h_1)x_2\phi(h_2)\ldots x_k\phi(h_k)$. Then we have a relative presentation $\hat P_2=\langle H_2, {\bf x}\mid \phi({\bf \hat r})\rangle$. We write $$\phi\colon \hat P_1=\langle H_1, {\bf x}\mid {\bf \hat r}\rangle \to \hat P_2=\langle H_2, {\bf x}\mid \phi({\bf \hat r})\rangle.$$
\[lemma:phi\] Suppose we have $\phi\colon \hat P_1 \to \hat P_2$. If $\hat P_2$ is DR, then $\hat P_1$ is so as well.
Proof. If $\hat f_1\colon \hat C \to K(\hat P_1)$ is an admissible spherical diagram then the spherical diagram $\hat f_2=\phi\circ \hat f_1\colon \hat C \to K(\hat P_2)$ is admissible as well. Topologically the two diagrams are the same, but a corner marked $h$ in the first diagram is marked $\phi(h)$ in the second. Since we assume that $\hat P_2$ is DR, the second diagram contains a folding edge. But this folding edge is also present in the first diagram.\
Tests for directed DR {#sec:tests}
=====================
Let $T_1=\langle x_1,\ldots ,x_k\mid r_1,\ldots ,r_l\rangle$ and $T_2=\langle y_1,\ldots ,y_p\mid s_1,\ldots ,s_q\rangle$ be presentations. Suppose we have a map $\phi_0\colon F( x_1,\ldots ,x_k)\to F(y_1,\ldots ,y_p)$ that induces a group homomorphism $G(T_1)\to G(T_2)$. We can extend $\phi_0$ to $\phi\colon F(x_1,\ldots ,x_n)\to F(y_1,\ldots ,y_p, x_{k+1},\ldots ,x_n)$ by defining $\phi(x_i)=x_i$ for $i>k$. Now let $P_1$ be a presentation of the form $$P_1=\langle x_1,\ldots ,x_k, x_{k+1},\ldots ,x_n\mid r_1,\ldots ,r_l, r_{l+1},\ldots ,r_m\rangle.$$ We assume each relator $r_j$, $j>l$, contains a generator $x_i$, $i>k$, so that $P_{S_1}=T_1$ for $S_1=\{ x_1,\ldots ,x_k\}$. Define $$P_2=\langle y_1,\ldots ,y_p, x_{k+1},\ldots ,x_n \mid s_1,\ldots ,s_q, \phi(r_{l+1}),\ldots ,\phi(r_m)\rangle.$$ We use $\phi\colon P_1\to P_2$ as shorthand for the situation just described. Here is Lamma \[lemma:phi\] in the directed DR setting.
\[sbarP\] Suppose we have $\phi\colon P_1 \to P_2$. If $P_2$ is DR directed away from $S_2=\{ y_1,\ldots ,y_p\}$, then $P_1$ is DR directed away from $S_1=\{ x_1,\ldots ,x_k\}$.
Proof. We have $\phi\colon \hat P_1\to \hat P_2$, where $\hat P_i$ is the relative presentation obtained from $P_i$ in the manner described in section \[sec:relpres\]. Since $P_2$ is DR directed away from $S_2$ it follows from Theorem \[thm:equivalence\] that $\hat P_2$ is DR. It follows from Lemma \[lemma:phi\] that $\hat P_1$ is DR. Hence, by Theorem \[thm:equivalence\] $P_1$ is DR directed away from $S_1$.\
The simplest choices for $T_2$ is the empty presentation $T_2=\langle \ |\ \rangle$, that is $G(T_2)$ is trivial.
*Consider $$P_1=\langle x_1,\ldots ,x_k, a, b\mid u_1au_2bu_3a^{-1}u_4b^{-1} \rangle \to P_2= \langle a, b\mid aba^{-1}b^{-1}\rangle,$$ where the $u_i$ are words in $x_1^{\pm 1},\ldots ,x_k^{\pm 1}$. Since $P_2$ is $DR$ (directed away from $\emptyset$), it follows that $P_1$ is DR directed away from $S=\{ x_1,\ldots ,x_k \}$. This can also be seen directly as in Example (3) of Section \[sec:basics\]. More general: Take any DR presentation $P_2$. Add generators $S=\{ x_1,\ldots ,x_k \}$ and insert words in $S^{\pm 1}$ into the relators of $P_2$ and one obtains a presentation $P_1$ which is DR directed away from $S$.*
A transformation of spherical diagrams over $K(P_1)$ to diagrams over $K(P_2)$ can be seen directly. For every relation $r$ in $T_1$ choose a Van Kampen diagram $D_{r}\to K(T_2)$ that expresses $\phi_0(r)$ in terms of the $s$-relations in $T_2$. This is possible because we assumed that $\phi_0$ induces a group homomorphism $G(T_1)\to G(T_2)$. Now given a spherical diagram $f\colon C\to K(P_1)$ we can produce a diagram $\phi\circ f: C'\to K(P_2)$ in the following way. If $e$ is an edge in $C$ labeled with $x$, then subdivide it and label it with $\phi(x)$. Note that if $x$ is a generator of $T_1$, then $\phi(x)$ is a word in the $y$ generators of $T_2$. If $d$ is a 2-cell in $C$ labeled with an $r_i$, $1\le i\le k$ from $T_1$, then remove the interior of $d$ and insert the Van Kampen diagram $D_r$. If $d$ is a 2-cell in $C$ labeled with $r_i$, $k+1\le i \le n$, and insert a relator disc labeled $\phi(r_i)$. The resulting diagram over $K(P_2)$ is $\phi\circ f$. Using this transformation we can give a direct proof of Lemma \[sbarP\] without reference to relative presentations. If $f\colon C\to K(P_1)$ is a spherical diagram that contains an edge labeled $x_i$, $k+1\le i\le n$ then so does $\phi\circ f$. Since $P_2$ is DR away from $S_2$, $\phi\circ f$ contains a folding edge with label $x_j$, $k+1\le j\le n$ (not in $S_2$). This edge is present in $f$ and is a folding edge.\
A [*cycle*]{} in a graph $\Gamma $ is a closed edge path in $\Gamma $. If $P=\langle x_1,\ldots ,x_n\mid r_1,\ldots ,r_m\rangle$ is a presentation then the [*Whitehead graph*]{} $W(P)$ is the boundary of a regular neighborhood of the only vertex of $K(P)$. It is a non-oriented graph on vertices $\{ x_i^+, x_i^- \mid i=1,\ldots ,n\}$, where $x_i^+$ is a point of the oriented edge $x$ of $K(P)$ close to the beginning of that edge, and $x_i^-$ is a point close to the ending of that edge. The [*positive graph*]{} $W^+(P)\subset W(P)$ is the full subgraph on the vertex set $\{ x_i^+\mid i=1,\ldots ,n\}$, the [ *negative graph* ]{} $W^-(P)\subset W(P)$ is the full subgraph on the vertex set $\{x_i^- \mid i=1,\ldots ,n\}$. Let $C$ be a cell decomposition of an oriented 2-sphere with oriented edges. A [*sink*]{} is a vertex in $C$ with all adjacent edges pointing towards it, a [*source*]{} is a vertex in $C$ with all adjacent edges pointing away from it.
\[sminmax\] Let $Q=\langle x_{1},\ldots ,x_n\mid r_{1},\ldots ,r_m\rangle$ be a finite presentation with cyclically reduced relators of exponent sum 0. Assume that $W^+(Q)$ or $W^-(Q)$ is a forest. Then $Q$ is DR directed away from $\{ x_k\}$ for each generator $x_k$.
Proof. Assume that $W^-(Q)$ is a forest. Let $f\colon C\to K(Q)$ be a spherical diagram. Since the relators are assumed to be of exponent sum zero, it follows that $C$ has a sink and a source vertex. Suppose $v$ is a sink vertex. The link at $v$ gives a cycle $z$ in a connected component of $W^-(Q)$, which is a tree. Thus, the edges covered by $z$ form a subtree $\Gamma$ of $W^-(Q)$, which contains two (or more) distinct vertices of valency one of the set $\{ x_{1}^-,\ldots ,x_n^-\}$. Thus the edges at $v$ in $C$, all directed towards $v$, will contain folding edges labelled by two elements of $\{ x_{1},\ldots ,x_n\}$. Thus one of them is labelled by a generator distinct from $x_k$. This implies that $Q$ is DR directed away from $\{ x_k\}$.
Analogous arguments apply if $W^+(Q)$ is a forest for a source vertex of $C$.\
It was known previously (Gersten [@Ger87]) that presentations satisfying the assumptions of Theorem \[sminmax\] are DR. Lemma \[sbarP\] together with Theorem \[sminmax\] provide a tool for showing directed DR for presentations. We will discuss this in greater detail in the next section.\
If $w$ is a real valued function on the edges of a graph $\Gamma $ and $z=e_1\ldots e_p$ is an edge path in $\Gamma$, we write $$w(z)=\sum_{1\le i\le p}w(e_i).$$
The following is a weight test for directed diagrammatic reducibility. It is a generalized version of Gerstens weight test (see [@Ger87]).
\[sWT\] Let $Q=\langle y_1,\ldots ,y_p,x_1,\ldots ,x_n\mid r_1,\ldots ,r_m\rangle$ be a presentation with cyclically reduced relators. Suppose we can assign non-negative weights $\omega(e)\ge 0$ to the edges $e$ in $W=W(Q)$, such that the following conditions hold:
1. If $e$ is an edge of $W$ that connects (for some $i,j\le p$) $y_i^\epsilon$ with $y_j^\delta$, ($\epsilon ,\delta =\pm$) then $\omega(e)\ge 1$;
2. If $e$ is an edge of $W$ and exactly one of its boundary points is $y_i^+$ or $y_i^-$ (for some $i\le p$), then $\omega(e)\ge 1/2$;
3. If $z$ is a reduced cycle in $W$, then $\omega(z)\ge 2$;
4. Let $r$ be a relator of length $d(r)$ from ${Q}$, then $\sum_{c\in r}\omega(c)\le d(r)-2$.
Then $Q$ is DR directed away from $\{ y_1,\ldots ,y_p\}$.
Proof. Note that conditions (3) and (4) in the statement of the theorem imply that $Q$ satisfies the standard weight test and it follows that $Q$ is DR. Let $h\colon C\to K(Q)$ be a spherical diagram. DR implies that there are folding edges in $C$. Suppose only the generators $y_1,\ldots ,y_p$ occur as labels on folding edges.
Let $z$ be a cycle in $W$ obtained from reading around a vertex in $C$. If $z$ is reduced then $\omega(z)\ge 2$ by condition (3). Suppose $z$ is not reduced. Then $z$ contains consecutive reducing edge pairs $e^{\epsilon}e^{-\epsilon}$, where $\epsilon=\pm 1$. For convenience we assume $\epsilon=1$. The terminal vertex of $e$ must be $y_i^+$ or $y_i^-$ for some $y_i$. If the other vertex of $e$ (note that $e$ connects two distinct vertices in $W$ because we assume that relators in $P$ are cyclically reduced) comes also from a generator $y_j$, then the pair $ee^{-1}$ contributes at least $2$ to the weight of $z$, by condition (1). Since weights are positive we then have $\omega(z)\ge 2$.
So suppose the other vertex of $e$ does come from a generator $x_i$. Then, by condition (2), the edge pair $ee^{-1}$ contributes at least $1$ to the weight of $z$. So if $z$ contains a second such edge pair, then the two edge pairs contribute at least $2$ to the weight of $z$, and we again have $\omega(z)\ge 2$. Suppose that $z$ contains only one consecutive reducing edge pair $ee^{-1}$. Then there is a reduced cycle $z'$ which consists of edges which also appear in $z$. By condition (3) we have $\omega(z)\ge \omega(z')\ge 2$.
Since $\omega(z)\ge 2$ for every cycle in $W$ obtained by reading around a vertex in $C$, the combinatorial curvature at every vertex in the spherical diagram $C$ is less or equal to $0$. The combinatorial curvature of the 2-cells in $C$ is also less or equal to $0$ by condition (4). The combinatorial Gauss-Bonnet Theorem now implies that the Euler characteristic of $C$ is less or equal to $0$, which contradicts the fact that $C$ is a sphere. So in $h$ must be a folding edge labelled by a generator different from one of $\{ y_1,\ldots ,y_p\}$.\
Again, Lemma \[sbarP\] together with Theorem \[sWT\] provide a method for proving directed DR for presentations. Details are discussed in the next section.
Applications {#sec:appl}
============
The small cancellation conditions C($p$), T($q$) are defined for instance in the book of Lyndon and Schupp (see [@LS77]).
\[s44\] Let $Q$ be a finite presentation with cyclically reduced relators and $S$ a subset of the generators. Assume that $Q$ is C(4), T(4) or C(6), T(3) and that no two consecutive letters in a (cyclically read) relator of $Q$ are generators in $S$. Then $Q$ is DR directed away from $S$.
Proof. We define a weight function $w\colon E\to\mathbb R$ where $E$ are the edges of the Whiteheadgraph $W(Q)$, such that the weight test of Theorem \[sWT\] is satisfied with $S=\{ y_1,\ldots ,y_p\}$.
Give each edge $e\in E$, which is in a cycle of length 2 weight $1$. Give all other edges in case C(4), T(4) weight $1/2$ and in case C(6), T(3) weight $2/3$. So condition (2) of Theorem \[sWT\] is satisfied. T(4) (or T(3)) implies condition (3) and C(4) (or C(6)) implies condition (4). Since no two consecutive letters in a (cyclically read) relator of $Q$ are generators of $S$ there will be no edges in $E$ connecting $y_i^\pm $ with $y_k^\pm $ for some $y_i,y_k\in S$. So condition (1) is satisfied also.\
At last we present some applications to labelled oriented trees. A [*labeled oriented tree*]{} (LOT) $\mathcal P$ consists of an oriented graph which is a tree, whose oriented edges are labeled by its vertices. From a labeled oriented tree $\mathcal P$ we obtain a LOT presentation $P$: The generators in $P$ are the vertices of $\mathcal P$, and for every oriented edge from a vertex $u_1$ to a vertex $u_2$, labeled by $u_3$ we have a relation $u_1u_3(u_3u_2)^{-1}$. Wirtinger presentations read from knot diagrams are examples of LOT presentations. Much information on labeled oriented trees and their importance to questions in combinatorial topology and group theory can be found in Rosebrock [@Ro18].
In Rosebrock [@Ro94] it is described how to check whether a LOT is C(4), T(4). If a LOT is C(4), T(4) then Theorem \[s44\] implies that it is DR away from any of its generators. But there is more:
Consider the LOT $Q$ of Figure \[alc4t4\] with any orientation of its edges. This LOT is C(4), T(4). If you choose $S$ to be one of the sets $$\{ x_1,x_2,x_5\} , \{ x_1,x_2\} , \{ x_1,x_4\} , \{ x_1,x_5\} , \{ x_2,x_5\} , \{ x_2,x_6\} , \{ x_2,x_7\} , \{ x_3,x_6\} , \{ x_3,x_7\} , \{ x_4,x_7\}$$ then Theorem \[s44\] implies that $Q$ is DR away from $S$.
(1,0)–(2,0)–(3,0)–(4,0)–(5,0)–(6,0)–(7,0); (1,0) circle (2pt); (2,0) circle (2pt); (3,0) circle (2pt); (4,0) circle (2pt); (5,0) circle (2pt); (6,0) circle (2pt); (7,0) circle (2pt); at (1,0) [$x_1$]{}; at (1.5,0) [$x_3$]{}; at (2,0) [$x_2$]{}; at (2.5,0) [$x_4$]{}; at (3,0) [$x_3$]{}; at (3.5,0) [$x_5$]{}; at (4,0) [$x_4$]{}; at (4.5,0) [$x_6$]{}; at (5,0) [$x_5$]{}; at (5.5,0) [$x_7$]{}; at (6,0) [$x_6$]{}; at (6.5,0) [$x_1$]{}; at (7,0) [$x_7$]{};
A [*sub-LOT*]{} of a labeled oriented tree $\mathcal P$ is a connected subgraph $\mathcal T$ (containing at least one edge) such that each edge label of $\mathcal T$ is a vertex of $\mathcal T$. A LOT is called [*compressed*]{} if each relator consists of three different generators. The relators of a LOT presentation of a compressed LOT are cyclically reduced.
Let $\mathcal P$ be a compressed LOT and $\mathcal T$ be a maximal proper sub-LOT with vertex set $S$. Let $\bar{\mathcal P}$ be the LOT obtained from $\mathcal P$ by collapsing all of $\mathcal T$ to a vertex $y$. Note that $\bar{\mathcal P}$ is compressed also. Every occurrence of a vertex $x$ of $\mathcal T$ in $\mathcal P-\mathcal T$ is replaced by $y$ in $\bar{\mathcal P}$. Let $T$, $P$, $\bar P$ be the corresponding LOT presentations.
\[sLOT\] If $\bar P$ is DR directed away from $y$ then $P$ is DR directed away from $S$, the generators of $T$. In particular if
1. $W^+(\bar P)$, or $W^-(\bar P)$ is a tree, or
2. the Whitehead graph $W(\bar P)$ does not contain cycles of length less than four,
then $P$ is DR directed away from $S$.
Proof. Let $\bar T=\langle y \mid - \rangle$. We can apply Lemma \[sbarP\] to the situation $T\subseteq P \to \bar T\subseteq \bar P$, where all generators of $T$ go to $y$ in $\bar T$. Here are the details. Let $S=\{ x_1,\ldots , x_k\}$ be the generators of $T$ (the vertices in $\mathcal T$) and let $\{ x_1,\ldots ,x_n\}$ be the generators of $P$ (the vertices in $\mathcal P$). Define $\phi_0\colon F(x_1,\ldots ,x_k)\to F(y)$ to be the map that sends each $x_i$ to $y$, and $\phi\colon F(x_1,\ldots ,x_n)\to F(y, x_{k+1},\ldots ,x_n)$ the extension of $\phi_0$ that sends $x_i$ to $x_i$ for $k+1\le i\le n$. Then $$\langle y, x_{k+1},\ldots ,x_n \mid \phi(r_{l+1}),\ldots , \phi(r_{m})\rangle$$ is the LOT presentation $\bar P$. Here the $r_j$, $l+1\le j \le m$ are the LOT relations coming from the edges in $\mathcal P-\mathcal T$.
Now (1) implies that $\bar P$ is DR away from $y$ by Theorem \[sminmax\] and (2) implies that $\bar P$ is DR away from $y$ by Theorem \[s44\]. It follows from Lemma \[sbarP\] that $P$ is DR directed away from $S$.\
[*Figure \[aex1\] below shows a compressed LOT $\mathcal P$ with a sub-LOT $\mathcal T$ with vertices $S=\{ x_1,\ldots ,x_5\}$. Below $\mathcal P$ we see the LOT $\bar{\mathcal P}$ obtained from $\mathcal P$ by collapsing $\mathcal T$ to the vertex $y$.*]{}\
at (-1,0) [$\mathcal P$]{}; at (5,0.75) [$\mathcal T$]{};
(0,0) circle (2pt); at (0,0) [$u_3$]{}; (1,0) circle (2pt); at (1,0) [$u_2$]{}; (2,0) circle (2pt); at (2,0) [$u_1$]{}; (3,0) circle (2pt); at (3,0) [$x_1$]{}; (4,0) circle (2pt); at (4,0) [$x_2$]{}; (5,0) circle (2pt); at (5,0) [$x_3$]{}; (6,0) circle (2pt); at (6,0) [$x_4$]{}; (7,0) circle (2pt); at (7,0) [$x_5$]{}; (8,0) circle (2pt); at (8,0) [$u_4$]{};
(3,0) – (4,0) node\[midway, above\][$x_3$]{}; (4,0) – (5,0) node\[midway, above\][$x_1$]{}; (5,0) – (6,0) node\[midway, above\][$x_5$]{}; (7,0) – (8,0) node\[midway, above\][$u_3$]{};
(0,0) – (1,0) node\[midway, above\][$u_1$]{}; (1,0) – (2,0) node\[midway, above\][$u_3$]{}; (2,0) – (3,0) node\[midway, above\][$u_4$]{}; (6,0) – (7,0) node\[midway, above\][$x_1$]{};
at (-1,0) [$\bar{\mathcal P}$]{};
(0,0) circle (2pt); at (0,0) [$u_3$]{}; (1,0) circle (2pt); at (1,0) [$u_2$]{}; (2,0) circle (2pt); at (2,0) [$u_1$]{}; (3,0) circle (2pt); at (3,0) [$y$]{}; (4,0) circle (2pt); at (4,0) [$u_4$]{};
(3,0) – (4,0) node\[midway, above\][$u_3$]{};
(0,0) – (1,0) node\[midway, above\][$u_1$]{}; (1,0) – (2,0) node\[midway, above\][$u_3$]{}; (2,0) – (3,0) node\[midway, above\][$u_4$]{};
Note that $W^+(\bar{P})$ is a tree, so by Theorem \[sLOT\] $P$ is DR directed away from $S$. Also observe that $W^-(T)$ is a tree, so $K(T)$ is DR, which implies that $K(P)$ is aspherical. Note that neither $W^+(P)$ nor $W^-(P)$ is a tree.\
The process of collapsing a sub-LOT in a given labeled oriented tree can also be reversed: If $\bar {\mathcal P}$ is a labeled oriented tree, $y$ is a vertex in $\bar {\mathcal P}$, and $\mathcal T$ is a labeled oriented tree, we remove $y$ from $\bar {\mathcal P}$ and insert $\mathcal T$ to obtain a labeled oriented tree $\mathcal P$ that contains $\mathcal T$. Collapsing $\mathcal T$ in $\mathcal P$ to a vertex $y$ brings us back to $\bar {\mathcal P}$. So the previous theorem can also be stated as follows: If $\bar{\mathcal P}$ is a labeled oriented tree that satisfies either condition (1) or (2) of Theorem \[sLOT\], then inserting any LOT $\mathcal T$ into $\bar{\mathcal P}$ results in a labeled oriented tree $\mathcal P$ for which the LOT presentation $P$ is DR directed away from the set $S$ of vertices of $\mathcal T$.\
LOT presentations $P$ where $W^+({P})$ or $W^-({P})$ is a tree abound. If a LOT $\mathcal P'$ is obtained from a LOT $\mathcal P$ by changing some edge orientations, we call $\mathcal P'$ a [*reorientation*]{} of $\mathcal P$. In In [@HR01] Proposition 5.1, Huck and Rosebrock show that each LOT $\mathcal P$ has a reorientation $\mathcal P'$ such that its positive Whitehead graph $W^+(P')$ is a tree. Theorem \[sminmax\] now implies
Each LOT $\mathcal P$ has a reorientation $\mathcal P'$ so that $P'$ is DR away from any one of its generators.
In [@Ro94] Rosebrock gives conditions on a labeled oriented tree so that condition (2) of Theorem \[sLOT\] holds. A concrete example is shown in Figure \[aex2\].
[*Figure \[aex2\] below shows a labelled oriented tree $\mathcal P$ (orientations can be chosen at will) with a sub-LOT $\mathcal T$ between $u_4$ and $u_4'$ (which can be filled in at will).*]{}
at (-1,0) [$\mathcal P$]{}; at (5,0.75) [$\mathcal T$]{};
(0,0) circle (2pt); at (0,0) [$u_1$]{}; (1,0) circle (2pt); at (1,0) [$u_2$]{}; (2,0) circle (2pt); at (2,0) [$u_3$]{}; (3,0) circle (2pt); at (3,0) [$u_4$]{}; (7,0)\[red\] circle (2pt); at (7,0) [$u'_4$]{}; (8,0) circle (2pt); at (8,0) [$u_5$]{}; (9,0) circle (2pt); at (9,0) [$u_6$]{}; (10,0) circle (2pt); at (10,0) [$u_7$]{};
(3,0) – (4,0); (4,0) – (5,0); (5,0) – (6,0); (7,0) – (8,0); (8,0) – (9,0); (9,0) – (10,0); (0,0) – (1,0) node\[midway, above\][$u_3$]{}; (1,0) – (2,0) node\[midway, above\][$u_4$]{}; (2,0) – (3,0) node\[midway, above\][$u_5$]{}; (6,0) – (7,0); (7,0) – (8,0) node\[midway, above\][$u_6$]{}; (8,0) – (9,0) node\[midway, above\][$u_7$]{}; (9,0) – (10,0) node\[midway, above\][$u_1$]{};
Notice that if we collapse the red sub-LOT $\mathcal T$ to the vertex $y=u_4$, we obtain a labeled oriented tree $\bar{\mathcal P}$ for which $W(\bar{\mathcal P})$ does not contain cycles of length less than four. It follows from Theorem \[sLOT\] that $P$ is DR away from the vertex set of $\mathcal T$.
[99]{}
W. Bogley and S. Pride. Proc. of the Edinburgh Math. Society 35, (1992), pp. 1-39.
W. Bogley, M. Edjvet and G. Williams. To appear in Groups St. Andrews,(2017).
J. M. Corson and B. Trace, [*Diagrammatically reducible complexes and Haken manifolds*]{}, J. Austral. Math. Soc. (Series A) 69, (2000), pp. 116-126.
S. M. Gersten, [*Reducible diagrams and equations over groups*]{}, in: Essays in Group Theory (S. M. Gersten editor), MSRI Publications 8 (1987), pp. 15-73.
J. Harlander and S. Rosebrock, [*Injective labelled oriented trees are aspherical*]{}, Mathematische Zeitschrift 287 (1), (2017), pp. 199-214.
G. Huck and S. Rosebrock, [*Aspherical Labelled Oriented Trees and Knots*]{}, Proceedings of the Edinburgh Math. Soc. 44, (2001), pp. 285-294.
R. Lyndon and P. Schupp. , Springer Verlag, Berlin, (1977).
S. Rosebrock, [*Labelled Oriented Trees and the Whitehead Conjecture*]{}; Advances in Two-Dimensional Homotopy and Combinatorial Group Theory; Cambridge University Press, LMS Lect. Notes 446, editors W. Metzler, S. Rosebrock; (2018), pp. 72-97.
S. Rosebrock, [*On the realization of Wirtinger presentations as knot groups*]{}, Journal of Knot Theory and its Ramifications, Vol. 3 (1994), 211-222.
A. J. Sieradski. ; Quart. J. Math. Oxford (2) 34 (1983), pp. 97-106.
Prof. Dr. Jens Harlander\
Department of Mathematics\
Boise State University\
Boise, ID 83725-1555\
USA\
email: [email protected]\
Dr. Stephan Rosebrock\
P[ä]{}dagogische Hochschule Karlsruhe\
Bismarckstr. 10\
76133 Karlsruhe\
Germany\
email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Under normal conditions, bulk crystals of BiScO$_3$, BiCrO$_3$, BiMnO$_3$, BiFeO$_3$, and BiCoO$_3$ present three very different variations of the perovskite structure: an antipolar phase, a rhombohedral phase with a large polarization along the space diagonal of the pseudocubic unit cell, and a supertetragonal phase with even larger polarization. With the aim of understanding the causes for this variety, we have used a genetic algorithm to search for minima in the surface energy of these materials. Our results show that the number of these minima is very large when compared to that of typical ferroelectric perovskites like BaTiO$_3$ and PbTiO$_3$, and that a fine energy balance between them results in the large structural differences seen. As byproducts of our search we have identified charge-ordering structures with low energy in BiMnO$_3$, and several phases with energies that are similar to that of the ground state of BiCrO$_3$. We have also found that a inverse supertetragonal phase exists in bulk, likely to be favored in films epitaxially grown at large values of tensile misfit strain.'
author:
- 'Akansha Singh,$^{1,2}$ Viveka N. Singh,$^{1}$ Enric Canadell,$^{3}$ Jorge Íñiguez,$^{4}$ and Oswaldo Diéguez$^{1,2}$'
title: 'Polymorphism in Bi-based perovskite oxides: a first-principles study'
---
.
Introduction {#sec:intro}
============
Thousands of research articles about BiFeO$_3$ have been published since Wang [*et al.*]{}[@Wang2003S] reported that epitaxial films of this material were multiferroic at room temperature (Ref. reviews the properties of this oxide). This effort has unvealed potential applications based on the possibility to control the electric polarization with a magnetic field[@Hur2004N] and the magnetization with an electric field.[@Chu2008NM] Other phenomena such as domain wall conductivity,[@Seidel2009NM] novel photovoltaic effects,[@Yang2010NN] and the presence of a feature akin to a morphotropic phase boundary in thin films[@Zeches2009S] have further fuelled research on BiFeO$_3$.
BiFeO$_3$ has a simple crystal structure in bulk at room temperature—a perovskite oxide with 10 atoms in a rhombohedral unit cell and space group $R3c$.[@Michel1969SSC] The cations are displaced along the space diagonal of the perovskite pseudocubic unit cell, and the O$_6$ octahedra rotate in antiphase about this same three-fold axis, as shown in Fig. \[fig\_groundstates\](a); the pseudocubic lattice constant is 3.965 Å and the pseudocubic angle is 89.45$^\circ$.[@Kubel1990AC] The large displacements of the cations give rise to a polarization[@Catalan2009AM] of around 100 $\mu$C/cm$^2$, while a slight canting of the spins of the Fe$^{+3}$ ions is responsible for the tiny ferromagnetic moment experimentally measured in an otherwise strong antiferromagnet of the G type[@Catalan2009AM] (the type in which the spins of two nearest Fe$^{+3}$ ions are antiparallel; in bulk BiFeO$_3$ there is in addition a spiral spin wave of period 640 Å).
![Structures of insulating Bi$X$O$_3$ crystals reported experimentally, labelled according to their corresponding space groups: (a) BiFeO$_3$ at normal conditions, (b) BiCoO$_3$ at normal conditions, (c) BiScO$_3$, BiCrO$_3$, and BiMnO$_3$ at normal conditions, and (d) BiFeO$_3$, BiScO$_3$, BiCrO$_3$, BiMnO$_3$, and BiCoO$_3$ at high temperature or high pressure. We plot the smallest pseudocubic supercell compatible with each phase; here and in later figures of this kind we show views along the $x$, $y$, and $z$ pseudocubic axes that are inequivalent (when only two pictures are included, the left picture corresponds to the two views that are equivalent).[]{data-label="fig_groundstates"}](fig_groundstates.pdf){width="80mm"}
None of the other bismuth $3d$ transition-metal perovskite insulators displays this $R3c$ structure in bulk.[@Belik2012JSSC; @Guennou2015CRP] BiCoO$_3$ crystallizes in what is called a supertetragonal structure—a tetragonal structure of $P4mm$ space group where the $c/a$ ratio is so large (1.27) that the O$_6$ octahedra become O$_5$ pyramids, as depicted in Fig. \[fig\_groundstates\](b); every plane of Co$^{3+}$ ions perpendicular to the $c$ axis has its spin aligned antiparallel to that of its nearest neighbors, while these planes are stacked along $c$ in ferromagnetic fashion (BiCoO$_3$ is therefore a C-type antiferromagnet). BiScO$_3$ (a non-magnetic material), BiCrO$_3$ (a G-type antiferromagnet, like BiFeO$_3$), and BiMnO$_3$ (the only ferromagnet in this family) have been experimentally reported to crystallize in a centrosymmetric phase with $C2/c$ space group, like the one shown in Fig. \[fig\_groundstates\](c). In contrast with this rich behavior under normal conditions, when the temperature or the pressure becomes large enough all these oxides (BiScO$_3$, BiCrO$_3$, BiMnO$_3$, BiFeO$_3$, and BiCoO$_3$) display the same structural phase:[@Belik2012JSSC] the paraelectric structure with space group $Pnma$ commonly found in perovksites,[@Lufaso2001ACB; @Chen2018PRB] represented in Fig. \[fig\_groundstates\](d).
First-principles studies have helped to characterize these structures for each material involved. In particular, they have confirmed that they correspond to special points of the energy surface: the $R3c$ structure of BiFeO$_3$ was identified as its ground state according to first-principles calculations already in Ref. ; McLeod [*et al.*]{}[@McLeod2010PRB] carried out first-principles optimizations of the $C2/c$ structure of BiScO$_3$ and BiCrO$_3$, which showed excellent agreement with experimental results regarding structural parameters; Baettig, Seshadri, and Spaldin[@Baettig2007JACS] confirmed the $C2/c$ structure as the ground state of BiMnO$_3$; and the supertetragonal structure of bulk BiCoO$_3$ was first analyzed from the first-principles point of view in Ref. , where a large value of its polarization was predicted. Computational studies also probed phases other than the ground state; early examples of these concern BiCrO$_3$[@Baettig2005PRB] (showing that it is possible to optimize a phase with the structure of bulk BiFeO$_3$) and BiFeO$_3$[@Ravindran2006PRB] (where the $R3c$ and four other additional phases were optimized). In our previous work, we reported that the energy surface of bulk BiFeO$_3$ shows many local minima, including several of supertetragonal type.[@Dieguez2011PRB] We also showed that in BiMnO$_3$ there are local minima of the bulk energy in addition to the experimental $C2/c$ ground state,[@Dieguez2015PRB] and that in BiCoO$_3$ the orthorhombic $Pnma$ phase and a $R3c$-like phase have low energy.[@Dieguez2011PRL] Finally, let us note that there is a considerable amount of work on solid solutions, particularly lanthanide-doped BiFeO$_{3}$.[@Arnold2015TUFFC] Beyond providing alternatives to tune the properties of bismuth ferrite, e.g. by inducing morphotropic phases boundaries[@Karpinsky2013JAC; @Shi2016JAP; @Gonzalez2012PRB] these solid solutions often feature long-period structures involving unusual rotation patterns of the O$_{6}$ octahedra.[@Prosandeev2013AFM]
It is thus abundantly clear that Bi-based perovskite oxides display a polymorphism that renders them a unique and very attractive materials family. In view of this, it is natural to wonder whether these materials have already revealed to us all the structures they have in store, or whether they may present even more phases with unexpected features. At the same time, it is not yet clear whether the polymorphism affects all Bi-based perovskite oxides, whether the presence of specific transition metal cations may play a critical role, whether a similar structural richness is present (hiding) in other perovskites, etc. In this article we address these issues by running a systematic search for bulk metastable phases of Bi$X$O$_3$ compounds ($X={\rm Sc}, {\rm Cr}, {\rm Mn}, {\rm Fe}, {\rm Co}$). We do this in an automatic and unbiased way with the help of the [uspex]{}[@uspex1; @uspex2; @uspex3] evolutionary algorithm for the search of crystal structures. We show that these oxides show very similar features in their energy surfaces, with the same structures appearing as minima in many of the materials, and that small energy differences between those minima are responsible for the rich variety of polymorphs displayed. We also compare this situation with what we obtain for some of the most used ferroelectric perovskite oxides (BaTiO$_3$ and PbTiO$_3$) where very few minima exist in their energy surface. In Section II we describe in detail the Methodology that we have used. In Section III we present and discuss our results. Finally, we state our conclusions in Section IV.
Method {#sec:methods}
======
We used the [vasp]{} software[@vasp] to carry out calculations based on density-functional theory (DFT).[@Hohenberg1964PR; @Kohn1965PR] Since DFT does not reproduce well the localization of $3d$ electrons in some of these oxides, a correction inspired in the Hubbard model was added; this methodology is usually refered to as LDA+$U$ or DFT+$U$,[@Himmetoglu2014IJQC] (details about it are given in Table \[tab\_dftu\]). We approximated the exchange-correlation functional following the work of Perdew, Burke, and Ernzerhof adapted to solids (PBEsol)\cite{}; different exchange-correlation approximations can affect the energy difference between minima of the energy surface, but in most cases they agree regarding whether a crystal structure is a minimum of the energy or not.[@Dieguez2011PRB] To capture the interaction between the valence electrons and the ion cores we used projector-augmented wave (PAW) potentials. [@paw1; @paw2] The electrons treated as valence ones were: 3[*p*]{}, 3[*d*]{}, and 4[*s*]{} (Sc, Ti, Cr, Mn, Fe, and Co); 5$s$, 5$p$, 6$s$ (Ba); 5$d$, 6$s$, and 6$p$ (Pb); 5[*d*]{} and 6[*s*]{} (Bi); and 2[*s*]{} and 2[*p*]{} (O).
BiCrO$_3$ BiMnO$_3$ BiFeO$_3$ BiCoO$_3$
----------- ----------- ----------- ----------- -----------
Formalism Ref. Ref. Ref. Ref.
$U$ 2.2 eV 4.0 eV 4.0 eV 6.0 eV
$J$ 0.0 eV 1.0 eV 0.0 eV 0.0 eV
: Details of the DFT+$U$ methodology used to treat the $d$ electrons of the transition-metal atoms: formalism references, values of the effective on-site Coulomb interactions $U$, and values of the effective on-site exchange interactions $J$. Those values are chosen based on previous experience; our tests showed that varying these parameters within a few eV has a small effect on whether a configuration is a minimum of the energy or not, and on the variation of the value of bond lengths and lattice parameters.[@Dieguez2017PRB] Regular DFT (no $U$) was used to treat BiScO$_3$, BaTiO$_3$, and PbTiO$_3$.
\[tab\_dftu\]
In order to carry out an unbiased search for energy minima we have used the [uspex]{} [@uspex1; @uspex2; @uspex3] code, which implements an evolutionary algorithm to find crystal structures with low energy. We worked with structures that have 10 or 20 atoms in their unit cell; the initial space group and atomic positions were chosen randomly by the code. For every [uspex]{} run we generated 50 of such random structures, and each of them was optimized by [vasp]{} using first a conjugate-gradient algorithm and then a quasi-Newton algorithm.[@Pulay1980CPL] The energy of the resulting configurations was taken as a fitness parameter to qualify for the next generation of trial structures, composed of 30 of them (this was also the number of structures used in subsequent generations). From then on, the population in every generation was computed following [uspex]{}’ specially designed variational operators (heredity, mutation, and soft-mutation).[@uspex3] This procedure was stopped at a maximum of 35 generations, or when the energy of the best structure did not change for 15 consecutive generations. The calculations were performed at different degrees of numerical convergence, with the most accurate level of calculations using a plane-wave energy cutoff of 500 eV, a reciprocal space resolution of $0.06 \times 2\pi$ Å$^{-1}$, and a force stopping criterion of 0.01 eV/Å. All the structures obtained in this way were reoptimized using an energy cutoff of 600 eV, a self-consistent cycle energy threshold of $10^{-8}$ eV, and a force stopping criterion of 0.001 eV/Å. The zone centered ($\Gamma$-point) frequencies were also calculated using finite differences, to ensure that the obtained structures are local minima of the energy in the simulation box used.
Once a structure was identified by the [uspex]{} search as a local minimum of the energy for one of the materials, we optimized it also for the other materials (starting with a configuration where the lattice parameters were rescaled to take into account that the ion sizes are differents for different oxides). While in some cases [uspex]{} found the same structures in different materials, in a few cases these new optimizations identified new energy minima.
As mentioned earlier, the Bi-based oxides studied here show different magnetic orderings: ferromagnetism (BiMnO$_3$), G-type antiferromagnetism (BiCrO$_3$ and BiFeO$_3$), C-type antiferromagnetism (BiCoO$_3$), and zero magnetic moments (BiScO$_3$). In our study with [uspex]{} we used antiferromagnetic 10- and 20-atom unit cells where the initial spins were assigned randomly. After we found crystal structures that were a minimum of the energy, we enlarged (if needed) and reoptimized the cell, so that the energy of other typical magnetic orderings was also computed. In our previous studies in some these perovskites we found that once one type of magnetic ordering is a minimum of the energy, others also exist as minima of the energy.[@Dieguez2011PRB; @Dieguez2011PRL; @Dieguez2015PRB]
Finally, we resorted to hybrid calculations for the few cases where we required a more accurate calculation of the energy differences within a group of phases, or a more accurate estimation of bandgaps. We used the modified Heyd-Scuseria-Ernzerhof approach[@Krakau2006JCP] (HSE06), which predicts bandgaps in good agreement with experiment for perovskite oxides[@Stroppa2010PCCP]; contrary to DFT+$U$, it also predicts the ground state of BiMnO$_3$ to be the experimentally identified one.[@Dieguez2015PRB] Calculations with this method are two orders of magnitude slower than with DFT+$U$, but, apart from the better agreement with experiment, they do not require differente fitting parameters for each material (HSE06 uses a fraction of exact exchange equal to 25%, and a range-separation parameter equal to 0.20 Å$^{-1}$).
Note that, because we choose to work with relatively small cells, this investigation does not address the long-period polymorphs that, as mentioned above, are believed to occur in some BiFeO$_{3}$-based solid solutions. Yet, as we will see, even if restricted to relatively small cell sizes, our simulations reveal an incredible structural richness that clearly single out Bi-based perovskites from the rest.
Results {#sec:results}
=======
First, we report in detail the BiFeO$_3$ polymorphs identified during our evolutionary-algorithm search. Despite of the extensive previous studies carried out for this oxide, we have identified several structures with low energy that had not been reported so far as energy minima of BiFeO$_3$. We then move on to the results of a similar search for other Bi-based transition-metal oxides; these harbor similar polymorphs to those of BiFeO$_3$, but slight differences in energy between those polymorphs lead to global minima that are structurally very different. We also report similar searches for prototype perovskites BaTiO$_3$ and PbTiO$_3$ for comparison. In the following three subsections we give more details about novel results of our search: structures with energies very close to that of the ground state of BiScO$_3$, BiCrO$_3$, and BiMnO$_3$; charge-ordering phases of BiMnO$_3$; and details about a polar phase expected to be stable as an epitaxial film is some cases. Finally, we report bandgap trends for the different phases found, and we discuss the reasons for the variety of polymorphs predicted for Bi-based perovskite oxides.
Local minima of BiFeO$_3$
-------------------------
Our searches for polymorphs of BiFeO$_3$ with [uspex]{} following the procudure describe in the Methods Section produced hundreds of structures, a result that is consistent with our initial exploration based on a more rudimentary approach.[@Dieguez2011PRB] In order to sort out which of those structures were unique, we used the crystal fingerprint method described in Ref. ; this allowed us to identify duplicated structures that differ slightly in their atomic coordinate positions due to numerical noise, and structures that differ only because of the use of different unit cell shapes and sizes by [uspex]{}. The structures that remained were re-optimized in the G- and C-type antiferromagnetic patterns. Only the lowest-energy magnetic ordering is reported here.
In this way we found 17 structurally unique structures within less than 200 meV/f.u. of the ground state. These local minima within 10- and 20-atom unit cells are listed in Table I. The ground state is, in agreement with experiment, the polar $R3c$ phase with G-type antiferromagnetism. The next configuration in energy is the non-polar $Pnma$ phase; as mentioned earlier, this is a phase that appears at high temperature or pressure for Bi-based transition-metal perovskite oxides.
----------------- ------------ ----- ------------ ------- ------- ------- ---------- --------- ---------- ----------------- ---------------------------------------- --------------- --
\# Group AFM $\Delta E$
$a$ $b$ $c$ $\alpha$ $\beta$ $\gamma$ Bi (polar) Bi (antipolar) O (rotation)
1 $R3c$ G 0 3.945 3.945 3.945 89.58 89.58 89.58 $(P_x,P_x,P_x)$ – ($a^-a^-a^-$)
2 $Pnma$ G 28 3.914 3.914 3.885 90 90 87.64 – $(2\pi/a)(0,0,1/2)$, $[aa0]$ ($a^-a^-c^+$)
3 $C2/c$ G 68 3.950 3.950 3.902 89.31 89.31 89.60 – $(2\pi/a)(1/4,1/4,1/4)$, $[aa0]$ ($a^-a^-c^0$)
4 $P\bar{1}$ G 93 3.924 3.897 3.957 89.56 89.33 89.94 – $(2\pi/a)(1/4,1/4,1/4)$, $[abc]$ ($a^-b^-c^-$)
5 $Cc$ C 97 3.760 3.760 4.721 88.02 88.02 89.99 $(p_x,p_x,P_z)$ $(2\pi/a)(1/2,1/2,1/2)$, $[a\bar{a}0]$ ($a^-a^-c^0$)
6 $Pna2_1$ C 100 3.756 3.756 4.722 90 90 89.98 $(0,0,P_z)$ $(2\pi/a)(0,0,1/2)$, $[aa0]$ ($a^-a^-c^0$)
7 $Cm$ C 103 3.689 3.803 4.769 86.59 90 90 $(0,p_y,P_z)$ $(2\pi/a)(1/2,0,1/2)$, $[0a0]$ ($a^0b^+c^0$)
8 $Cc$ C 105 3.747 3.753 4.741 90 87.97 90 $(p_x,0,P_z)$ $(2\pi/a)(0,0,1/2)$, $[0a0]$ ($a^0b^-c^+$)
9 $Pc$ C 106 3.750 3.750 4.744 88.09 88.09 89.74 $(p_x,p_x,P_z)$ $(2\pi/a)(1/2,1/2,0)$, $[a\bar{a}0]$ ($a^0a^0c^+$)
10 $Pmn2_1$ C 106 3.690 3.799 4.762 90 90 90 $(0,0,P_z)$ $(2\pi/a)(0,0,1/2)$, $[0a0]$ ($a^0b^+c^0$)
11 $Cm$ C 109 3.742 3.742 4.760 87.97 87.97 89.68 $(p_x,p_y,P_z)$ – –
12 $Pc$ G 109 3.987 3.987 3.971 89.06 89.06 89.92 $(P_x,P_x,P_z)$ $(2\pi/a)(1/2,1/2,0)$, $[a\bar{a}0]$ ($a^0a^0c^+$)
13 $Cm$ G 111 3.915 3.928 3.982 90 89.38 90 $(P_x,0,P_z)$ $(2\pi/a)(0,1/2,1/2)$, $[a0c]$ ($a^+b^+c^+$)
14 $Fmm2$ C 116 3.747 3.753 4.675 90 90 90 $(0,0,P_z)$ $(2\pi/a)(1/2,1/2,1/2)$, $[a00]$ ($a^-b^0c^0$)
15 $Pmc2_1$ G 119 4.189 4.189 3.706 90 90 88.65 $(P_x,P_x,0)$ $(2\pi/a)(1/2,1/2,0)$, $[a\bar{a}0]$ ($a^0a^0c^+$)
16 $Cm$ G 125 3.987 4.031 3.871 90 90 89.18 $(P_x,P_y, 0)$ $(2\pi/a)(0,1/2,1/2)$, $[a00]$ ($a^+b^0c^0$)
17 $R\bar{3}$ G 164 3.934 3.934 3.934 89.57 89.57 89.57 – $(2\pi/a)(1/2,1/2,1/2)$, $[aaa]$ ($a^-a^-a^-$)
\[tab\_bifeo3\]
----------------- ------------ ----- ------------ ------- ------- ------- ---------- --------- ---------- ----------------- ---------------------------------------- --------------- --
A first surprising result appears next in Table \[tab\_bifeo3\], as we identified a polymorph of BiFeO$_3$ with relatively low energy that corresponds to the experimentally reported crystal structure of bulk BiScO$_3$, BiCrO$_3$, and BiMnO$_3$ in normal conditions. This is a structure with $C2/c$ space group in which the pattern of antiparallel Bi displacements in the planes perpendicular to the inequivalent axis resemble that of Pb atoms in PbZrO$_3$ (space group $Pbam$), the prototype of antiferroelectric material.[@Trolier2008JAP] In both structures there is an underlying pattern of O$_6$ rotations in antiphase along the $[110]$ direction ($a^-a^-c^0$ rotations in Glazer’s notation[@GlazerPapers]). However, as Figure \[fig\_afe\] shows, the pattern of $A$ cation displacements in the direction perpendicular to the page is different in the $C2/c$ structure and in the PbZrO$_3$ structure—the former is associated with point $(2 \pi/a)(1/4,1/4,1/4)$ of the Brillouin zone of the simple cubic perovskite cell, while the latter it is associated with the $(2 \pi/a)(1/4,1/4,0)$ point; for $C2/c$ BiFeO$_3$ the movement of the Bi atoms distorts further the O$_6$ octahedra.
The similarity with the antiferroelectric phase of PbZrO$_{3}$ suggest an intriguing possibility: that, if stabilized, the non-polar polymorphs of BiFeO$_{3}$ would probably behave as antiferroelectrics,[@Rabe2013book] as it seems likely that they could be transformed, by application of an electric field, into the polar $R3c$ state. This is in fact a behavior that is being explored in BiFeO$_{3}$ solid solutions in which the $Pnma$ structure exists at ambient conditions;[@Xu2017NC; @Kan2010AFM] our results suggest that other antiferroelectric phases could be similarly identified. (A precise definition of an antiferroelectric phase is a matter of some debate; it is not our purpose here to contribute to that debate, and we simply adopt the practical view point just introduced. Also, in the following, when referring to atomic displacement patterns like those of Pb in the PbZrO$_{3}$ structure, we use the terms antiferroelectric and antipolar indistinctly.)
When we optimized BiFeO$_3$ starting with a structure like that of PbZrO$_3$ we did not obtain a minimum of the energy—instead, the relaxed structure shows a soft mode that involves additional antiferroelectric displacements of the Bi atoms along the inequivalent axis; if we distort the structure along that mode, the energy indeed lowers, and re-optimization leads to a 40-atom unit cell phase around 41 meV/fu above the ground state with $P2_1/c$ symmetry (which could not have been found in our [uspex]{} search since it has a 40-atom primitive cell).
![Comparison between the $C2/c$ structure that is a local energy minimum of 20-atom cell BiFeO$_3$ (left) and the $Pbam$ structure that is a global energy minimum of bulk PbZrO$_3$ (right). Each image shows the distribution of O$_6$ octahedra in each plane perpendicular to the inequivalent axis of each structure, together with arrows that represent the displacement of Bi atoms in the nearest parallel plane to those octahedra. []{data-label="fig_afe"}](fig_afe.pdf){width="70mm"}
Complex patterns of antiferroelectric Bi displacements appear also in the next most stable structure found, of space group $P\bar{1}$ (Figure \[fig\_struct2\](a)). Like the $C2/c$ phase, this low-symmetry non-polar structure shows an antiferroelectric pattern tied to the $(2 \pi/a)(1/4,1/4,1/4)$ point of the first Brillouin zone, and a $a^-b^-c^-$ rotation pattern for the O$_6$ octahedra. Another difference with the $C2/c$ phase is that now the Bi atoms move substantially along the three pseudocubic axes.
The four lowest energy structures of Table \[tab\_bifeo3\] have in common that their three pseudocubic lattice parameters are similar in value. As discussed in our previous work,[@Escorihuela2013PRL] this implies that the favoured antiferromagnetic ordering is of the G type. Polar structures 12, 13, 15, and 16, and antipolar structure 17 also fit this pattern, although Fig. \[fig\_struct2\] shows how diverse they are in their geometries.
For example, in structure 12 (space group $Pc$, Fig. \[fig\_struct2\](b)) we find a mix of a polar $\Gamma_4^-$ mode along (111) with and antipolar $M_5^-$ mode along $(1 \bar{1} 0)$, resulting in large equal Bi displacements along the $z$ pseudocubic axis, and displacements of Bi atoms in the perpendicular plane that involve half the atoms moving along the $x$ pseudocubic axis and the other half moving along the $y$ pseudocubic axis; the associated polarization lies close to the (111) pseudocubic axis, while the O$_6$ octahedra show a $a^0a^0c^+$ pattern.
Structure 13 (space group $Cm$, Fig. \[fig\_struct2\](c)) also mixes a polar and an antipolar mode: a $\Gamma_4^-$ mode along with components along the $x$ and $z$ directions, and a $M_5^-$ mode almost perpendicular to it; this results in a complex pattern of Bi displacements along two different pseudocubic axes and the corresponding in-plane polarization, together with a $(a^+b^+c^+)$ O$_6$ rotation.
Structure 15 (space group $Pmc2_1$, Fig. \[fig\_struct2\](d)) is far from a typical perovskite; the coexistence of a $\Gamma_4^-$ component along (110) with a strong $M_5^-$ mode perpendicular to it results in the breaking of the O$_6$ octahedra into O$_5$ pyramids, but these pyramids form a zig-zag pattern different to the one in supertetragonal structures. This structure has been reported before[@Yang2012PRL] as a plausible phase to appear in epitaxial films under large tensile strains; now we see that it is a minimum in [*bulk*]{} BiFeO$_3$, giving support to its possible experimental realization even at the nominally large strains required if comparing to the $R3c$ ground state. This $Pmc2_1$ phase shows the smallest $c/a$ ratio among the BiFeO$_3$ polymorphs, so in the following we refer to it as an [*inverse supertetragonal*]{} structure.
Structure 16 (space group $Cm$, Fig. \[fig\_struct2\](e)) mixes a polar $\Gamma_4^-$ mode in the $xy$ plane with an antipolar $M_3^-$ mode along $x$, resulting in a net polarization off the (110) direction.
Finally, structure 17 (space group $R3$, Fig. \[fig\_struct2\](e)) is the antipolar equivalent of the $R3c$ ground state in that for every Bi that moves along $(111)$, its six nearest Bi neighbors move along $(\bar1 \bar1 \bar1)$, while the O$_6$ octahedra pattern is still $(a^-a^-a^-)$.
The rest of phases found by [uspex]{} are of supertetragonal type. Structures 5, 6, 7, and 8 where already found in our previous search,[@Dieguez2011PRB] while structures 9, 10, 11, and 14 are of a similar kind. All these supertetragonal structures show a large $\Gamma_4^-$ soft mode distortion along $z$, and some of them show polar Bi displacements in the $xy$ plane (5, 7, 8, 9, and 11). They differ also regarding slight O$_5$ pyramid rotations.
Bi$X$O$_3$ Phase Search {#sec:other_compounds}
-----------------------
To help understand the complexity of BiFeO$_3$’s potential energy landscape we compared it to that of related oxides Bi$X$O$_3$, with $X = {\rm Sc}, {\rm Cr}, {\rm Mn}, {\rm Co}$. We run [uspex]{} on 10-atom and 20-atom unit cells of BiScO$_3$ (spin unpolarized calculations), BiCrO$_3$ and BiCoO$_3$ (antiferromagnetic), and BiMnO$_3$ (both ferromagnetic and antiferromagnetic). In order to find which magnetic order is favored, we re-optimized all minima identified by imposing ferromagnetic and and A-type antiferromagnetic for BiMnO$_3$, and G- and C-type for BiCrO$_3$ and BiCoO$_3$ (these are expected to be the competitive cases[@Dieguez2011PRB; @Dieguez2015PRB]).
This search resulted in 22 new minima, to add to the 17 initial ones of BiFeO$_3$. The main information about these 39 local minima (in their 10- or 20-atom unit cell) is contained in Table \[tab\_bixo3\], where their energy is underlined; this energy is computed with respect to the latest experimental ground states reported. Once a structure was found for one of the transition metals, we re-optimized it for the others too; if this led us to a energy minimum of similar geometric characteristics, we quote its energy, space group, and magnetic ordering in Table \[tab\_bixo3\] (not underlined). For all calculations reported in this Table we extended the unit cell of the structures to a 40-atom unit cell—the $2 \times 2 \times 2$ pseudocubic one when possible, and the conventional cell of the $C2/c$ phase for the other cases. The existence of negative eigenvalues of the dynamical matrix was checked for those extended 40-atom unit cells (and not only for the cell in which the minima was found, as in Table \[tab\_bifeo3\]). If soft modes exist, we give the energy of the structure in brackets. Thus, for example, structures 3, 11, and 16 are reported as local minima in Table \[tab\_bifeo3\] because they are so in their respective 20-, 10-, and 20-atom unit cells where they were found by [uspex]{}; however, when we used 40-atom unit cells soft modes appeared in those three cases. The structure that results when we follow the soft mode of the $C2/c$ phase is the subject of next subsection.
---------------- ---- ------------ ------- ------------ ---- ------------ ------- ------------ ------ ------------ ------- ------------ ---- ------------ ------- ------------ ---- ------------ -------
S.G. M. $\Delta E$ $c/a$ S.G. M. $\Delta E$ $c/a$ S.G. M. $\Delta E$ $c/a$ S.G. M. $\Delta E$ $c/a$ S.G. M. $\Delta E$ $c/a$
$P4mm$ 0 (409) 1.296 $P4mm$ A (390) 1.264 $P4mm$ C (140) 1.282 $P4mm$ C 1.257
$R3c$ 0 1.000 $R3c$ G 1.000 $Cc$ FM 0.975 $R3c$ G 1.000 $Cc$ G 1.026
$Pnma$ 0 0.994 $Pnma$ G -59 0.996 $Pnma$ FM -17 0.962 $Pnma$ G 0.992 $Pnma$ G 47 1.014
$C2/c$ FM\* 0.985
$P2_1/c$ FM\* 0.990
$C2/c$ 0 0 0.984 $C2/c$ G (0) 0.993 $C2/c$ FM (0) 1.017 $C2/c$ G () 0.988 $C2/c$ G 53 0.979
$R3$ FM\* 1.000
$R\bar{3}$ FM\* 1.000
$Pna2_1$ 0 1.009
$P\bar{1}$ 0 28 1.011 $P\bar{1}$ G (26) 0.994 $P\bar{1}$ FM (37) 1.019 $P\bar{1}$ G 1.012 $P\bar{1}$ G 88 0.964
$R\bar{3}$ 0 1.000 $R\bar{3}$ G 43 1.000 $P\bar{1}$ FM 64 0.952 $R\bar{3}$ G 1.000 $P\bar{1}$ G 1.023
$Cm$ 0 0.988 $Cm$ G (95) 1.003 $Cm$ FM 97 0.949 $Cm$ G 1.015 $Cm$ G (139) 1.025
$Cc$ C 1.255
$Pna2_1$ A 139 1.149 $Pna2_1$ C 1.257
$Cm$ C 1.273
$Cc$ C 1.264
$Pc$ A 212 1.220 $Pc$ C 1.265
$Pmn2_1$ C 1.272
$Cc$ 0 1.022 $Cc$ G 107 1.039
$Pc$ 0 0.991
$Cm$ C () 1.272
$Pc$ 0 69 1.004 $Pc$ G (122) 1.015 $Pc$ FM (127) 1.000 $Pc$ G 0.996 $Pc$ G (148) 1.025
$Pna2_1$ 0 0.997 $Pna2_1$ G 1.001 $Pna2_1$ FM 122 0.990 $Pna2_1$ G 126 0.992 $Pna2_1$ G 136 1.020
$Ama2$ 0 () 0.996 $Ama2$ FM (131) 0.988
$Pmc2_1$ 0 (208) 0.879 $Pmc2_1$ FM (146) 0.907 $Pmc2_1$ G 0.885
$Pmc2_1$ 0 (236) 0.970 $Pmc2_1$ G (169) 1.014 $Pmc2_1$ G (163) 0.969 $Pmc2_1$ G () 1.014
$Fmm2$ C 1.247
$Cm$ 0 (81) 0.968 $Cm$ G (118) 1.024 $Cm$ FM (121) 1.039 $Cm$ G () 0.966 $Cm$ G (117) 1.031
$Cc$ G 1.041
$Cm$ 0 1.009
\[tab\_bixo3\]
---------------- ---- ------------ ------- ------------ ---- ------------ ------- ------------ ------ ------------ ------- ------------ ---- ------------ ------- ------------ ---- ------------ -------
Another example will help to clarify how information is reported in Table \[tab\_bixo3\]. For BiCoO$_3$, [uspex]{} found the supertetragonal $P4mm$ phase that is indeed the known experimental ground state of the material; so an underlined value of 0 meV/fu is added to Table \[tab\_bixo3\] in the BiCoO$_3$ column. When computing energies of different magnetic arrangements, antiferromagnetism of the C type is favored (also in agreement with experiment), so this is too reflected, together with the large $c/a$ ratio. We then run the same structure for the other four materials (we started optimizations by simply changing the $B$ cation and using a initial cell volume adjusted proportionally to the size of the new cation). In three cases (BiScO$_3$, BiMnO$_3$, and BiFeO$_3$) the optimization converged, but there were negative dynamical matrix eigenvalues in the $2 \times 2 \times 2$ pseudocubic cell, so the energy values are in brackets. In the other case, BiCrO$_3$, the supertetragonal $P4mm$ phase would just not be a special point of the energy surface.
Once again, we see a richness of polymorphs that is unique to bismuth perovskite oxides. The phases here include the previously reported ones for these five materials, most notably their ground state form according to experiment and the $Pnma$ phase that they all reach at high enough temperature or pressure. We can establish three groups of structures: phases with similar pseudocubic lattice constants, phases with one pseudocubic lattice constant significantly larger than the other two (supertetragonal phases), and phases with one pseudocubic lattice constant significantly smaller than the other two (inverse supertetragonal phases). Among the first group, a phase that appears for all five materials as a local minimum is the $R3c$ phase (global minimum of BiFeO$_3$). In BiMnO$_3$ and BiCoO$_3$ this phase is slightly distorted because Mn$^{+3}$ and Co$^{+3}$ are Jahn-Teller active ions, so their space group becomes actually $Cc$; however, the main distortions in the structures are still the large cation displacement along the (111) pseudocubic direction and a O$_6$ rotation similar to $a^-a^-a^-$, so we group these minima in the same line of the Table. The supertetragonal polar phases appear only in BiMnO$_3$, BiFeO$_3$ and BiCoO$_3$, with A-antiferromagnetic ordering (BiMnO$_3$) and C-antiferromagnetic ordering (BiFeO$_3$ and BiCoO$_3$). On the other hand, non-supertetragonal polar and non-polar phases can be seen in all five materials; they favour G-antiferromagnetic ordering in BiCrO$_3$, BiFeO$_3$, and BiCoO$_3$, while ferromagnetic ordering and A-antiferromagnetic ordering are very close in BiMnO$_3$, as previously reported.[@Dieguez2015PRB] The inverse supertetragonal phases are analyzed in more detail in subsection E.
Apart from the high number of structures listed, a few other facts deserve to be commented in relation to Table \[tab\_bixo3\]. First, there are several instances of negative energies for BiScO$_3$, BiCrO$_3$, and BiMnO$_3$. This implies that according to our calculations, those phases have energies below that of the assumed experimental ground state. In a previous article we reported that for BiMnO$_3$ DFT+$U$ does not agree with experiment in that the $C2/c$ phase is the one with the lowest energy, but that when using hybrid methods reconciliation with experiment is achieved. Now we see that the situation is extended to BiScO$_3$ and BiCrO$_3$. In next subsection we look in more depth at this issue. Second, we have found that BiMnO$_3$ shows some unique phases where charge separation is present—Mn$^{+2}$ and Mn$^{+4}$ ions coexist. This is further discussed in subsection D.
As a point of comparison, we also run analogous [uspex]{} optimization calculations for prototypical perovskites BaTiO$_3$ and PbTiO$_3$. Full runs as those described in the Methods section produced just the known ground states of these materials: the rhombohedral $R3m$ phase for BaTiO$_3$ and the tetragonal $P4mm$ phase for PbTiO$_3$. No other local minima was found in 20-atom unit cells (the tetragonal $P4mm$ and orthorhombic $Amm2$ phases of BaTiO$_3$ are saddle points, not minima).
To give a graphical idea of the richness of Bi$X$O$_3$ phases found, Fig. \[fig\_tableII\_cVSa\] shows the value of the $c$ lattice parameter versus the average $a$ and $b$ parameters for those phases of Table \[tab\_bixo3\] that are local minima. The lattice parameters are sorted so that $a$ and $b$ are as close as possible to each other; we do this with the idea in mind that a possible way to access local minima of these materials is to grow them as epitaxial films on square perovskite lattices—then, the in-plane lattice parameter would likely be around $(a+b)/2$ and the out-of-plane lattice parameter would be around $c$. It is apparent from the graph that one cluster of supertetragonal phases appear at the top left of the figure, while most of the rest of the phases have $c/a$ ratios not far from one. The outlier at the bottom right part of the graph is phase 15 of BiFeO$_3$ (in this graph only phases that are minima in the 40-atom unit cells described earlier are represented, so the inverse supertetragonal phases of other materials are absent).
![ Out-of-plane lattice parameter $c$ versus in-plane lattice parameter $a$ (the average of the two closest lattice parameters in the pseudocubic setting), for the structures listed in Table \[tab\_bixo3\] that are local minima of the energy. The discontinuous line corresponds to $c=a$. []{data-label="fig_tableII_cVSa"}](tableII_cVSa.pdf){width="60mm"}
This adaptability of the Bi ion to different environments of perovskite variations implies not only that phases other than the ground state might be stabilized by strain or pressure, but also that these materials should have richer surfaces and interfaces than typical perovskites. For example, some BiFeO$_3$ domain walls can be seen as narrow regions in which the structure corresponds to a diffferent polymorph[@Dieguez2013PRB; @Prosandeev2013AFM] and the same it is true for its surfaces.[@Marti2011PRL]
Ground State of BiScO$_3$, BiCrO$_3$, and BiMnO$_3$
---------------------------------------------------
We move on now to the issue of the phases of BiScO$_3$, BiCrO$_3$, and BiMnO$_3$ that show lower energies than the one of the experimental ground state of these materials, the phase with $C2/c$ symmetry. In our previous study of BiMnO$_3$[@Dieguez2015PRB] we reported that DFT+$U$ calculations very similar to the ones presented here indeed stablish that the $Pnma$ phase has a somewhat lower energy than the $C2/c$ phase. We also showed that when using hybrid functionals, this ordering is reversed, and the $C2/c$ phase becomes the lowest-energy one, in accordance to the latest experiments about the structure of this material.[@Belik2012JSSC] Now we see that for BiScO$_3$ and BiCrO$_3$ we also obtain that the $Pnma$ phase has lower energy than the $C2/c$ phase. Moreover, in this case this is also true for the $R3c$ phase.
More puzzingly, the data in Table \[tab\_bixo3\] tells that the $C2/c$ phase of BiCrO$_3$ and BiMnO$_3$ shows a soft mode when the conventional 40-atom unit cell is used. When we perturb slightly the atomic positions of these structures by following the eigenvector of the soft mode, indeed the energy of BiCrO$_3$ and BiMnO$_3$ (and also BiFeO$_3$) goes down, while this does not happen in BiScO$_3$ (and BiCoO$_3$). This is consistent with the structure being a minimum in BiScO$_3$ (and BiCoO$_3$), but a saddle point in BiCrO$_3$ and BiMnO$_3$ (and BiFeO$_3$). What is the structure that appears if we keep following the eigenvector of the soft mode and then let the atoms relax? This structure has 40 atoms in its unit cell (and therefore could not have been found in our [uspex]{} search over 10-atom and 20-atom cells), and space group $P2_1/c$. In addition to the distortions quoted in Table \[tab\_bifeo3\] for the $C2/c$ phase, this $P2_1/c$ phase has two other prominent ones: an antipolar Bi displacement along the $(001)$ pseudocubic axis, and a $(a^0 a^0 c^+)$ O$_6$ rotation. In all, it shows a similar O$_6$ rotation pattern to that of the $Pnma$ phase, and a complicated antipolar Bi pattern associated mainly to the $(2\pi/a)(1/4,1/4,1/4)$ point of the simple cubic first Brillouin zone.
Figure \[fig\_c2c\_to\_p21c\] (left panels) shows the energy as we interpolate linearly the atomic positions and lattice vectors from the $C2/c$ phase to the $P2_1/c$ phase, for all five Bi-based perovskites. It is apparent that the curvature at the $C2/c$ point is much larger for those materials where this phase is a minimum (BiScO$_3$ and BiCoO$_3$) than for the others. When the $C2/c$ phase is a saddle point of the energy surface, then the $P2_1/c$ phase corresponds to a minimum with lower energy; when instead the $C2/c$ phase is a minimum, then the $P2_1/c$ phase is also a minimum, but it might exist at a lower energy (BiScO$_3$) or at a higher one (BiCoO$_3$).
![ Left panels: energy of the structures that result from a linear interpolation of atomic positions and lattice vectors between the $C2/c$ phase and the $P2_1/c$ phases of the materials quoted. Right panels: square of the frequency of the mode that connects the $C2/c$ and $P2_1/c$ phases as a function of the volume of the $C2/c$ unit cell; the vertical lines mark the volume of the cell with optimal lattice parameters. []{data-label="fig_c2c_to_p21c"}](fig_c2c_to_p21c.pdf){width="77mm"}
What is the reason for the different character of the $C2/c$ phases in the different oxides analyzed here? To help us answer this question, Figure \[fig\_c2c\_to\_p21c\] (right panels) contains the value of the frequency of the mode that becomes soft in BiCrO$_3$ and BiMnO$_3$, for different values of unit cell volume—i.e., we increased the length of the $a$, $b$, and $c$ lattice parameter of the $C2/c$ conventional unit cell by some amount, relaxed the atoms with those fixed lattice vectors, computed the eigenvalues of the dynamical matrix, and plotted the square of the eigenvalue of interest in the graphs. What results is consistent with a similar behavior of this frequency for all five materials: at large values of volume there is room for the O$_6$ cages without the need for the extra rotation of the $P2_1/c$ phase, so the eigenvalues are positive, while at small enough volumes this is not the case and the structure adds those rotations as a way to avoid compressing the bonds between transition metals and oxygens further. In between those regimes, the rotations and the Bi displacements compete, giving rise to the non-monotonic behaviour seen in the panels. Depending on what the optimal value of the volume is for a given material, different parts of the curve are accessed, and the different behaviours listed earlier result.
The previous discussion points out that the relative ordering of the $C2/c$ and $P2_1/c$ phases is a subtle effect. To further investigate this issue, and the one of the negative energies of Table \[tab\_bixo3\], we have used hybrid functionals. These computationally demanding calculations represent today one of the most accurate types of first-principles calculations that can be performed to compute energies for unit cells of 40 atoms.[@Krakau2006JCP; @Stroppa2010PCCP] The results we obtained are quoted in Table \[tab\_hybrid\]. Some of the BiMnO$_3$ data had been reported before,[@Dieguez2015PRB] showing that the $C2/c$ phase was lower in energy than the $R3c$ and $Pnma$ phases; now we see that the $P2_1/c$ phase has very close energy to the $C2/c$ phase, but $C2/c$ is still the ground state according to hybrid calculations. Also in BiScO$_3$ the $C2/c$ phase is the ground state according to our HSE06 calculations, in agreement with experiment. For BiCrO$_3$, however, the hybrid calculations put the analyzed structures in a small bracket of energies where the $C2/c$ phase is not the lowest-energy one (whether this is an artifact or not of the calculations is difficult to assess, given the small differences in energy involved). Computing possible phonon instabilities of these phases with the HSE06 hybrid functional would require ten to twenty times more computer power than the calculations reported in Table \[tab\_hybrid\], which are already computationally demanding. Instead, the results reported in Fig. \[fig\_c2c\_to\_p21c\] give evidence that when the $C2/c$ phase is lower in energy than the $P2_1/c$ phase both correspond to energy minima. When using hybrids we see that the $C2/c$ phase in BiScO$_3$ and BiMnO$_3$ is lower in energy than the $P2_1/c$ phase, so we expect both structures to be minima ($C2/c$ being the global one, in agreement with experiment).
----------------- ---- ------------ ----- ---- ------------ ----- ---- ------------ -----
Phase M. $\Delta E$ Gap M. $\Delta E$ Gap M. $\Delta E$ Gap
$C2/c$ 0 0 4.3 G 0 3.1 FM 0 1.8
$P2_1/c$ 0 14 4.0 G -2 3.1 FM 1 1.9
$R3c$ or $Cc$ 0 4 4.2 G -13 3.3 A 33 2.5
$Pnma$ 0 64 4.0 G -8 2.7 FM 53 1.7
\[tab\_hybrid\]
----------------- ---- ------------ ----- ---- ------------ ----- ---- ------------ -----
: Results of HSE06 calculations of the most stable structures of bulk BiScO$_3$, BiCrO$_3$, and BiMnO$_3$. For each material we report its lowest magnetic ordering, the energy difference with the $C2/c$ phase (in meV/f.u.), and its band gap (in eV).
Charge-Ordering Phases in BiMnO$_3$
-----------------------------------
Another interesting feature reported in Table \[tab\_bixo3\] is the presence of four phases of BiMnO$_3$ that possess charge ordering. These correspond, in order of increasing energy, to variations of the $C2/c$, $Pnma$, $R3c$, and $R\bar{3}$ phases, where a rocksalt pattern of Mn$^{2+}$ and Mn$^{4+}$ ions exists in the $B$ site of the perovksite (unit cells of two of the phases found are pictured in Fig. \[fig\_co\]). Charge-ordering phases of perovskites have been reported experimentally before, with the ordering taking place either in the $B$ site (e.g., Mn$^{+3}$ and Mn$^{+4}$ in Ln$_{1-x}$$A_x$MnO$_3$, where Ln $=$ rare earth and $A =$ Ca, Sr[@Rao1998CM]) or in the $A$ site (e.g., Bi$^{+3}$ and Bi$^{5}$ in BiNiO$_3$[@Ishiwata2002JMC]). However, no $AB$O$_3$ perovskite with charge ordering in $B$ seems to have been found so far.
![ 40-atom unit cells for two of the charge-ordering phases of BiMnO$_3$ found in this work: a $R3c$-like phase (left) similar to the ground state of BiFeO$_3$ but with a rock-salt pattern imposed on the perovksite $B$ sites, and a $Pnma$-like phase (right) similar to the high-pressure phase of these materials but with the same rock-salt pattern. The larger (smaller) octahedra enclose Mn$^{+2}$ (Mn$^{+4}$) ions. []{data-label="fig_co"}](fig_co_1.pdf "fig:"){width="40mm"} ![ 40-atom unit cells for two of the charge-ordering phases of BiMnO$_3$ found in this work: a $R3c$-like phase (left) similar to the ground state of BiFeO$_3$ but with a rock-salt pattern imposed on the perovksite $B$ sites, and a $Pnma$-like phase (right) similar to the high-pressure phase of these materials but with the same rock-salt pattern. The larger (smaller) octahedra enclose Mn$^{+2}$ (Mn$^{+4}$) ions. []{data-label="fig_co"}](fig_co_2.pdf "fig:"){width="40mm"}
To check how robust the prediction done with DFT+$U$ is, we also carried out hybrid calculations for these phases (assuming that the spins are oriented parallel to each other). Table \[tab\_co\] contains the results of a comparison of the structure of these variations and the one of the regular phases. In particular, we show the average Mn–O distance obtained for each individual Mn ion of the structure; these results agree well with the picture of ionic bonding that emerges from the radii given by Shannon[@Shannon1976AC] (1.35 Å for O$^{-2}$, 0.53 Å for Mn$^{+4}$, 0.65 Å for Mn$^{+3}$, 0.82 Å for Mn$^{+2}$).
[rccccccc]{} & & &\
Phase & $\Delta E$ & $a$ (Å) & $b$ (Å) & $c$ (Å) & Mn$^{4+}$ & Mn$^{3+}$ & Mn$^{2+}$\
\
$C2/c$ & 0 & 3.903 & 3.903 & 3.975 & – & 2.02 & –\
$R3c$-like & 36 & 4.026 & 4.026 & 3.772 & – & 2.02 & –\
$Pnma$ & 53 & 3.949 & 3.949 & 3.764 & – & 2.03 & –\
$R\bar{3}$-like & 102 & 3.986 & 4.072 & 3.749 & – & 2.02 & –\
\
$C2/c$-like & 97 & 3.929 & 3.929 & 3.857 & 1.91 & – & 2.13\
$R3c$-like & 152 & 3.919 & 3.919 & 3.919 & 1.92 & – & 2.09\
$Pnma$-like & 160 & 3.923 & 3.923 & 3.870 & 1.92 & – & 2.12\
$R\bar{3}$-like & 145 & 3.918 & 3.918 & 3.918 & 1.91 & – & 2.13\
\[tab\_co\]
As seen in Table \[tab\_co\], the energy differences between the phases with charge ordering and the ground state are larger than the ones reported in Table \[tab\_bixo3\] using DFT+$U$. Still, these differences are similar to the ones between experimentally grown supertetragonal phases of BiFeO$_3$ and its ground state. It might therefore be feasible to stabilize some of these phases; a possible route could be to grown BiMnO$_3$ epitaxially on a perovskite substrate that shows a checkerboard pattern compatible with the distortions, e.g., using as substrate one of the many ordered double perovskites that are known.[@King2010JMC]
Highly Strained Epitaxial Film Phase
------------------------------------
In Fig. \[fig\_tableII\_cVSa\] we plotted the behaviour of the lattice parameter that most differ from the other two ($c$) as a function of the average of those other two ($a$). This is to connect with the topic of epitaxial films that are grown on a square substrate, a common situation for perovskites grown on perovskites or on other materials of cubic symmetry—the bulk forms are likely to adapt more easily to the substrate in the orientation that matches best two of its lattice parameters to the substrate square parameter. The points in that figure can be divided in three groups. In most cases, $c/a$ is between 0.9 and 1.1, and we find there typical phases of perovskites, such as the $R3c$, $C2/c$, and $Pnma$ ones. When $c/a$ is larger than 1.1 we obtain the supertetragonal phases similar to the ground state of BiCoO$_3$ that we have already mentioned. There is a third set of points, with $c/a$ below 0.9, corresponding to an orthorhombic $Pmc2_1$ phase first reported by Yang [*et al.*]{}[@Yang2012PRL] after they found computationally as favorable epitaxial phase in the region of large tensile strains of BiFeO$_3$ and PbTiO$_3$ (they quote a strain of 5%).
Now we show that this phase is a local minimum [*of the bulk*]{} of BiFeO$_3$, and a saddle point in some of the related materials. This is relevant because it implies that, like for its supertetragonal phase, BiFeO$_3$ can stand nominally large epitaxial strains without creating a myriad of defects—the material grows into a phase whose lattice parameters match very well those of the substrate, and therefore very small stresses are present in the film.
There is an interesting detail in Table \[tab\_bixo3\] related to this inverse supertetragonal phase (initially found in our BiFeO$_3$ [uspex]{} search): another phase with the same set of main distortions and the same $Pmc2_1$ space group exists, but with a $c/a$ ratio much closer to 1. This similar phase appeared in the [uspex]{} search of BiCoO$_3$ (111 eV/fu above the $P4mm$ ground state). Apart from $c/a$, the difference between the two phases is that the $(a^0 a^0 c^+)$ rotation occurs in opposite directions with respect to the mix of polar and antipolar Bi displacements, as shown in Fig. \[fig\_tensile\] (a). As seen in the Table, both phases are present in BiFeO$_3$ and BiScO$_3$, but we have only confirmed one of them for the other three materials.
With the goal of further understanding these phases and of assesing their suitability as possible ground states in epitaxial films at large tensile strains, Fig. \[fig\_tensile\](b) shows the results of constraining them to square substrates. In particular, the bottom graph shows that it costs little energy to adapt the bulk phases to the elastic constraint imposed by the substrate (the energy difference between the minimum of the open symbols curves and the small filled symbols is very small). It also reveals that these two isosymmetric structures correspond to two actual different phases, as it can be seen more clearly for the cases of BiFeO$_3$ and BiScO$_3$, where at intermediate values of $a$ the two types can be optimized. The middle graph shows how the $c/a$ ratio is almost linear with $a$, and almost identical for all cases, but when the larger Sc$^{+3}$ ion is present. The top graph gives the clearest evidence of the values of $a$ for which one or the other phase is the favored one. Finally, the energy curves help to understand why in some cases our bulk calculations find only one of the two phases: they are very close in the search space of crystal configurations, and when free bulk optimizations are performed these end up in the lowest special point available—the inverse supertetragonal one (BiMnO$_3$) or its companion (BiCoO$_3$ and BiCrO$_3$). For BiFeO$_3$ and BiScO$_3$ the two phases are further away from each other, and optimizations can access either of them.
In all, our calculations related to this inverse supertetragonal phase reinforce the point made by Yang [*et al.*]{}[@Yang2012PRL] that materials such as BiFeO$_3$ (and perhaps other Bi-based perovskites if they can be stabilized) might exist in epitaxial polymorphs with a large in-plane polarization, and with other interesting properties that arise due to the special network of square pyramids present.
Band Gaps
---------
In recent times, the search for clean energy has spurred the optimization of materials for converting solar light into electricity. Optimal materials have bandgaps that are smaller than those typical of ferroelectric perovskite oxides, but efforts have been made to reduce those bandgaps, and in this way couple the ferroelectric functionalities to those of an energy material.[@Kreisel2012NM; @Grinberg2013N] This motivated us to analyze the bandgaps predicted for the phases identified in our search.
Experimentally, the optical bandgap of BiFeO$_3$ has been measured by several groups, with results ranging from 2 to 3 eV.[@Catalan2009AM] For the other Bi-based materials of this study much less is known about the value of their bandgap. McLeod [*et al.*]{}[@McLeod2010PRB] estimated them from soft X-ray emission and X-ray absorption spectroscopy measurements, and obtained values between 0.9 eV (BiMnO$_3$ and BiFeO$_3$) and 2.6 eV (BiScO$_3$)—the value for BiFeO$_3$ is significantly lower than what others have measured. Computationally, it is well known that standard DFT implementations predict bandgaps that are systematically smaller than experimentally reported ones, and that hybrid implementations such as HSE06 produce much better estimations of the gap.[@Krakau2006JCP] Because of this, we show in Fig. \[fig\_gap\] (top) a comparison between our DFT+$U$ methodology and our HSE06 methodology; this shows that the former predicts bandgaps that are around 1.5 eV lower than the latter. With this in mind, Fig. \[fig\_gap\] (bottom) shows our DFT+$U$ computations of the band gap of the phases of Table \[tab\_bixo3\] (shifted by 1.5 eV, and done in a reciprocal space grid that is twice more dense in each direction that the one used to optimize the structures). From these results we conclude that, for a given composition, the band gap is not strongly dependent on the particular polymorph. Further, BiMnO$_3$ clearly appears as the most promising material in this family for photovoltaic applications.
![ Top panel: comparison between band gaps computed with DFT+$U$ and with HSE06; the discontinuous line shows that the predictions by the first approach is systematically lower than the one by the second approach, by a shift of around 1.5 eV. Bottom panel: value of DFT+$U$ bandgap for all phases reported in Table \[tab\_bixo3\] that are energy minima (shifted by 1.5 eV to reflect the most accurate hybrid calculations) as a function of the average between the two closest lattice parameters of their pseudocubic unit cell. []{data-label="fig_gap"}](bandgaps.pdf){width="60mm"}
Why So Many Polymorphs?
-----------------------
As mentioned earlier, when similar [upsex]{} phase searches are applied to prototype ferroelectric perovskites BaTiO$_3$ and PbTiO$_3$ and to Bi-based perovskites Bi$X$O$_3$, few minima appear in the former cases and many minima appear in the latter one. How can we understand the difference between these crystals based on the characteristics of the $A$ and $B$ cations?
One particularity of the Bi-based perovskites is the lone pair of $6s$ electrons of Bi$^{3+}$ that overlaps with the O orbitals to attain a lobe-shape form, as shown by, e.g., electron-localization function analyses.[@Seshadri2001CM; @Ravindran2006PRB; @Dieguez2011PRB] The presence of these lobes breaks the spherical symmetry of the Bi$^{3+}$ ions, and it causes them to move out of high-symmetry positions—this is how ferroelectricity originates in BiFeO$_3$ and related materials. But those lobes can be accommodated in many other ways (paying small energy prices) that give rise to local minima of the energy illustrated in Table \[tab\_bifeo3\] for BiFeO$_3$.
Ba is an spherical ion, and therefore in BaTiO$_3$ it will not add variety to the structures of this perovskite. However, Pb does have a $6s$ electron pair, so the case of PbTiO$_3$ must be analyzed further. There are two main differences between Bi and Pb: (i) Bi$^{3+}$ is smaller (Shannon radius of 1.17 Å at maximum coordination reported[@Shannon1976AC]) than Pb$^{2+}$ (Shannon radius of 1.49 Å); and (ii) Bi$^{3+}$ is more electronegative than Pb$^{2+}$ (2.02 versus 1.87 in the Pauling scale[@Allred1961JINC]). The $s$/$p$ mixing leading to the active lone pair arises from the interaction with the $p$ orbitals of oxygen; since Bi$^{3+}$ is more electronegative there should be a better match in this respect, and since Bi$^{3+}$ is smaller it should adapt better to the electronic requirements of the metal-oxygen lattice to distort. In any case, it is an oversimplification to assume that the energy surface of PbTiO$_3$ has one unique feature (its tetragonal minimum); for example, it is known that its cubic phase has tilt instabilities, which only decay into an untilted phase as strain coupling occurs.[@Ghosez1999PRB; @Wojdel2013JPCM]
The differences between BaTiO$_3$, PbTiO$_3$, and Bi$X$O$_3$ can be therefore attributed mainly to the uniqueness of Bi$^{+3}$’s lone pair, but the properties of the transition-metal cation also play a role. In particular, they must be responsible for the variations in the structures of the Bi$X$O$_3$ oxides, where the formal $d$ electron count is the main difference between transition-metal atoms that in some cases are very similar in size (the Shannon radii[@Shannon1976AC] of Sc$^{3+}$, Cr$^{3+}$, Mn$^{3+}$, Fe$^{3+}$, and Co$^{3+}$ are 0.745 Å, 0.615 Å, 0.645 Å, 0.645 Å, and 0.61 Å, respectively).
For a $d^0$ perovskite (BaTiO$_3$, PbTiO$_3$, BiScO$_3$) distortions away for the ideal cubic structure can stabilize the system because only the metal-oxygen bonding levels are filled, and none of the metal-based orbitals (which are metal-oxygen antibonding) are occupied. In BaTiO$_3$ atoms move most favorably towards a triangular face of a O$_6$ octahedron, resulting in its $R3m$ rhombohedral phase. The presence of a lone pair and a large Pb$^{+2}$ cation changes this balance and makes a tetragonal phase more stable in PbTiO$_3$. A smaller ion as Bi$^{+3}$ results, as mentioned above, in more freedom for the ion to explore its surroundings, and consequently in more phases.
Once the $d$ electron count increases, more possibilities open. For example, vanadates ($d^1$) have a tendency to exhibit a square pyramidal coordination,[@Zavalij1999AC] and they do so in a perovskite like PbVO$_3$,[@Belik2005CM] giving rise to a supertetragonal structure. BiCoO$_3$ assumes a related high-spin d$^6$ configuration (a spherically symmetric high-spin $d^5$ configuration plus a $d^1$ one) and therefore exhibits the same type of $P4mm$ structure in its ground state. Now of course the other cation is Bi$^{3+}$, and thus many more possibilities arise.
Systems like BiScO$_3$ ($d^0$), BiFeO$_3$ (high spin $d^5$), and even BiCrO$_3$ (high spin $d^3$, occupying all $t_{2g}$ levels with one electron) contain a transition-metal atom with some sort of electronic closed shell, so in principle they should be quite similar; indeed we see some of the structures in Table \[tab\_bixo3\] being shared for these materials. However, finer effects are also at play, involving O$_6$ rotations that are mainly combinations of three types: $(a^- a^- a^0)$, $(a^0 a^0 c^+)$, and $(a^0 a^0 c^-)$. These add variety to the structures, and result in phases like $C2/c$ being also quite low in energy. This is because $C2/c$ combines the very energetically favourable $(a^- a^- c^+)$ rotations with parallel and antiparallel Bi displacements to accommodate the lone pair lobes. In this respect, it represents a compromise between the paraelectric $Pnma$ phase and the strongly polar $R3c$ phase.
The Jahn-Teller active ions Mn$^{+3}$ ($d^4$) and Co$^{+3}$ ($d^6$) add extra possible distortions, reflected for example in the breaking of symmetry that affects the phase that is the ground state of $R3c$ and energetically competitive in BiCoO$_3$ and BiMnO$_3$, but with space group $Cc$.
Conclusions
===========
In this article we report tens of metastable structures of Bi$X$O$_3$ ($X={\rm Sc}, {\rm Cr}, {\rm Mn}, {\rm Fe}, {\rm Co}$) compounds. These are minima of the energy within 200 meV/fu or less of the ground state, as computed using methods based on DFT (more minima exist at higher energy differences). This large degree of polymorphism is related in part to the same mechanism that is responsible for large values of the polarization in many of them: the lone electron pair of the Bi$^{3+}$ ion.[@Seshadri2001CM; @Ravindran2006PRB; @Dieguez2011PRB] These materials can accommodate those pairs in many combinations of polar and antipolar displacements that result in many energy minima. In addition to this, three other factors add variety to these structures: (1) several available O$_6$ rotation patterns, mainly based on combinations of $(a^- a^- c^0)$, $(a^0 a^0 c^+)$, and $(a^0 a^0 c^-)$; (2) the possibility of breaking one of the $X$–O bonds to create stable supertetragonal phases in BiMnO$_3$, BiFeO$_3$, and BiCoO$_3$, and inverse supertetragonal phases in BiFeO$_3$; and (3) Jahn-Teller distortions in BiMnO$_3$ and BiCoO$_3$.
In this way, we see that the variety of lowest-energy states in this family is just a reflection of the fine balance between the energies involved in the cation displacements and the octahedra rotations. Our calculations show that the reported experimental structures of these materials actually exist in all of them (with the exception of the variations of the supertetragonal structure of BiCoO$_3$, which do not exist in BiScO$_3$ and BiCrO$_3$ as minima). When the materials are grown, the small energy differences involved translate into the dramatic structural differences seen. Epitaxial strain and pressure may be used to shift this balance and favor different polymorphs (including polar ones), as it has been done already in supertetragonal BiFeO$_3$.[@Bea2009PRL; @Zeches2009S]
While many calculations based on DFT agree that the ground states of BiCoO$_3$ and BiFeO$_3$ are the ones found experimentally,\cite{} the situation for BiScO$_3$, BiCrO$_3$, and BiMnO$_3$ is not so simple. Our calculations using the accurate HSE06 hybrid functional show that the experimentally reported $C2/c$ phase is in competition with a similar phase where the primitive unit cell doubles and an extra $(a^0 a^0 c^+)$ O$_6$ rotation appears. In the case of BiCrO$_3$, this new phase is slightly lower in energy, and so are the $R3c$ and $Pnma$ phases.
As part of this research, we have also identified local minima of the energy in BiMnO$_3$ where charge ordering is present—Mn$^{2+}$ and Mn$^{+4}$ ions alternate in a rocksalt pattern inside the O$_6$ octahedra. These phases are not far in energy from the ground state, and they might be stabilized by growing BiMnO$_3$ on double perovskites that favor this charge ordering.
A detailed study of the $Pmc2_1$ inverse supertetragonal phase found here as a bulk minimum adds to the evidence that BiFeO$_3$ might grow into this polymorph on square-symmetry substrates with in-plane lattice parameter around 4.2 Å; this was proposed earlier by Yang [*et al.*]{},[@Yang2012PRL] who identified this polymorph as a favored configuration in epitaxial films under large tensile strains. In other materials of the family this is also a special point of the energy surface, and therefore amenable to stabilization.
Finally, calculations of band gaps for these structures show that BiMnO$_3$ is the most promising of these materials for optical applications, since it has the lowest band gap of the family—between 1.5 eV and 2.0 eV according to our HSE06 hybrid calculations.
Acknowledgements {#acknowledgements .unnumbered}
================
[ O.D. acknowledges funding from the Israel Science Foundation through Grants 1814/14 and 2143/14. Work in Bellaterra was supported by MINECO (Spain) through Grant FIS2015-64886-C5-3-P as well as the Severo Ochoa Centers of Excellence Programs under Grant SEV-2015-0496, and by Generalitat de Catalunya (2017SGR1506). J.I. was supported by the Luxembourg National Research Fund through Grant No. P12/4853155 COFERMAT. ]{}
[99]{}
J. Wang, J.B. Neaton, H. Zheng, V. Nagarajan, S.B. Ogale, B. Liu, D. Viehland, V. Vaithyanathan, D.G. Schlom, U.V. Waghmare, N.A. Spaldin, K.M. Rabe, M. Wuttig, and R. Ramesh, Science [**299**]{}, 1719 (2003).
G. Catalan and J. F. Scott, Adv. Mater. [**21**]{}, 2463 (2009).
N. Hur, S. Park, P.A. Sharma, J.S. Ahn, S. Guha, and S.-W. Cheong, Nature 429, [**392**]{} (2004).
Y.-H. Chu, L.W. Martin, M.B. Holcomb, M. Gajek, S.-J. Han, Q. He, N. Balke, C.-H. Yang, D. Lee, W. Hu, Q. Zhan, P.-L. Yang, A. Fraile-Rodriguez, A. Scholl, S.X. Wang, and R. Ramesh, Nat. Mater. [**7**]{}, 478 (2008).
J. Seidel, L.W. Martin, Q. He, Q. Zhan, Y.-H. Chu, A. Rother, M.E. Hawkridge, P. Maksymovych, P. Yu, M. Gajek, N. Balke, S.V. Kalinin, S. Gemming, F. Wang, G. Catalan, J.F. Scott, N.A. Spaldin, J. Orenstein, and R. Ramesh, Nat. Mater. [**8**]{}, 229 (2009).
S.Y. Yang, J. Seidel, S.J. Byrnes, P. Shafer, C.-H. Yang, M.D. Rossell, P. Yu, Y.-H. Chu, J.F. Scott, J.W. Ager, L.W. Martin, and R. Ramesh, Nature Nanotechnology [**5**]{}, 143 (2010).
R.J. Zeches, M.D. Rossell, J.X. Zhang, A.J. Hatt, Q. He, C.H. Yang, A. Kumar, C.H. Wang, A. Melville, C. Adamo, G. Sheng, Y.-H. Chu, J.F. Ihlefeld, R. Erni, C. Ederer, V. Gopalan, L.Q. Chen, D.G. Schlom, N.A. Spaldin, L.W. Martin, and R. Ramesh, Science [**326**]{}, 977 (2009).
C. Michel, J.M. Moreau, G.D. Achenbach, R. Gerson, and W.J. James, Solid State Communications [**7**]{}, 701 (1969).
F. Kubel and H. Schmid, Acta Crystallogr. B [**46**]{}, 698 (1990).
A.A. Belik, Journal of Solid State Chemistry [**195**]{}, 32 (2012).
M. Guennou, M. Viret, and J. Kreisel, C. R. Physique [**16**]{}, 182 (2015).
M.W. Lufaso and P.M. Woodward, Acta Crystallogr. Sect. B [**57**]{}, 725 (2001).
P. Chen, M.N. Grisolia, H. J. Zhao, O.E. González-Vázquez, L. Bellaiche, M. Bibes, B.-G. Liu, and J. Íñiguez, Phys. Rev. B [**97**]{}, 024113 (2018).
J.A. McLeod, Z.V. Pchelkina, L.D. Finkelstein, E.Z. Kurmaev, R.G. Wilks, A. Moewes, I.V. Solovyev, A.A. Belik, E. Takayama-Muromachi, Phys. Rev. B [**81**]{} 144103 (2010).
P. Baettig, R. Seshadri, and N.A. Spaldin, J. Am. Chem. Soc. [**129**]{} 9854 (2007).
Y. Uratani, T. Shishidou, F. Ishii, and T. Oguchi, Jpn. J. Appl. Phys. [**44**]{}, 7130 (2005).
P. Baettig, C. Ederer, and N. Spaldin, Phys. Rev. B [**72**]{}, 214105 (2005).
P. Ravindran, R. Vidya, A. Kjekshus, H. Fjellvåg, and O. Eriksson, Phys. Rev. B [**74**]{}, 224412 (2006).
O. Diéguez, O. E. González-Vázquez, J. C. Wojdeł, and J. Íñiguez, Phys. Rev. B [**[83]{}**]{}, 094105 (2011).
O. Diéguez and J. Íñiguez, Phys. Rev. B [**91**]{}, 184113 (2015).
O. Diéguez and J. Íñiguez, Phys. Rev. Lett. [**107**]{}, 057601 (2011).
D.C. Arnold, IEEE Trans. Ultrason. Ferroelectr. Freq. Control, [**62**]{}, 62 (2015).
D.V. Karpinsky, I.O. Troyanchuk, M. Tovar, V. Sikolenko, V. Efimov, and A.L. Kholkin, Journal of Alloys and Compounds [**555**]{}, 101 (2013).
X.X. Shi, X.Q. Liu, and X.M. Chen, Journal of Applied Physics [**119**]{}, 064104 (2016).
O.E. González-Vázquez, J.C. Wojdeł, O. Diéguez, and J. Íñiguez, Phys. Rev. B [**85**]{}, 064119 (2012).
S. Prosandeev, W. Dawei, R. Wei, J. Íñiguez, and L. Bellaiche, Adv. Funct. Mater. [**23**]{}, 234 (2013).
A.R. Oganov and C.W. Glass, J. Chem. Phys., [**124**]{}, 244704 (2006).
C.W. Glass, A.R. Oganov and N. Hansen, Comp. Phys. Comm., [**175**]{}, 713, (2006).
A.O. Lyakhov, A.R. Oganov, H.T. Stokes and Q. Zhu, Comp. Phys. Comm., [**184**]{}, 1172-1182, (2013).
G. Kresse and J. Hafner, Phys. Rev. B [**47**]{} 558 (1993); G. Kresse and J. Hafner, Phys. Rev. B [**49**]{}, 14251 (1994); G. Kresse and J. Furthmüller , J Comput. Mat. Sci. [**6**]{}, 15 (1996); G. Kresse and J. Furthmüller , Phys. Rev. B [**54**]{}, 11169 (1996); G. Kresse, Thesis, 1993, Technische Universit at Wien.
P. Hohenberg and W. Kohn, Phys. Rev. [**136**]{}, B864 (1964).
W. Kohn and L. J. Sham, Phys. Rev. [**140**]{}, A1133 (1965).
B. Himmetoglu, A. Floris, S. de Gironcoli, and M. Cococcioni, Int. J. Quantum Chem. [**114**]{}, 14 (2014).
S.L. Dudarev, G.A. Botton, S.Y. Savrasov, C.J. Humphreys, and A.P. Sutton, Phys. Rev. B [**[57]{}**]{}, 1505094105 (1998).
A.I. Liechtenstein, V.I. Anisimov and J. Zaane, Phys. Rev. B [**52**]{}, R5467 (1995).
O. Diéguez and J. Íñiguez, Phys. Rev. B [**95**]{}, 085129 (2017).
P. E. Blöchl, Phys. Rev. B. [**50**]{}, 17953 (1994).
G. Kresse and D. Joubert, Phys. Rev. B. [**59**]{}, 1758 (1999).
P. Pulay, Chem. Phys. Lett. [**73**]{}, 393 (1980).
A. V. Krakau, O. A. Vydrov, A. F. Izmaylov, and G. E. Scuseria, J. Chem. Phys. [**125**]{}, 224106 (2006).
A. Stroppa and S. Picozzi, Phys. Chem. Chem. Phys. [**12**]{}, 5405 (2010).
A.R. Oganov and M. Valle, J. Chem. Phys [**130**]{}, 104504 (2009).
S. Trolier-McKinstry, M.D. Biegalski, J. Wang, A.A. Belik, E. Takayama-Muromachi, and I. Levin, J. Appl. Phys. [**104**]{}, 044102 (2008).
A.M. Glazer, Acta Cryst. [**B28**]{}, 3384 (1972); A.M. Glazer, Acta Cryst. [**A31**]{}, 756 (1974).
K.M. Rabe, in [*Functional Metal Oxides: New Science and Novel Applications*]{}, edited by S. Ogale, V. Venkateshan, and M.G. Blamire (Wiley, Hoboken, NJ, 2013)
B. Xu, J. Íñiguez, and L. Bellaiche, Nature Communications [**8**]{}, 15682 (2017).
D. Kan, L. Pálová, V. Anbusathaiah, C.J. Cheng, S. Fujino, N. Valanoor, K.M. Rabe, and I. Takeuchi, Adv. Funct. Mater. [**20**]{}, 1108 (2010).
The [Isotropy]{} software suite is available at [http://iso.byu.edu]{}.
C. Escorihuela-Sayalero, O. Diéguez, and J. Íñiguez, Phys. Rev. Lett. [**109**]{}, 247202 (2012).
Y. Yang, W. Ren, M. Stengel, X.H. Yan, and L. Bellaiche, Phys. Rev. Lett. [**109**]{}, 057602 (2012).
O. Diéguez, P. Aguado-Puente, J. Junquera, and J. Íñiguez, Phys. Rev. B [**87**]{}, 024102 (2013).
S. Farokhipoor, C. Magén, S. Venkatesan, J. Íñiguez, C.J.M. Daumont, D. Rubi, E. Snoeck, M. Mostovoy, C. de Graaf, A. Müller, M. Döblinger, C. Scheu, and B. Noheda, Nature [**515**]{}, 379 (2014).
X. Martí, P. Ferrer, J. Herrero-Albillos, J. Narvaez, V. Holy, N. Barrett, M. Alexe, and G. Catalan, Phys. Rev. Lett. [**106**]{}, 236101 (2011).
C.N.R. Rao, A. Arulraj, P.N. Santosh, and A.K. Cheetham, Chem. Mater. [**10**]{}, 2714 (1998).
S. Ishiwata, M. Azuma, M. Takano, E. Nishibori, M. Takata, M. Sakata, and K. Kato, J. Mater. Chem. [**12**]{}, 3733 (2002).
R. Seshadri and N.A. Hill, Chem. Mater. [**13**]{}, 2892 (2001).
R.D. Shannon, Acta Cryst. [**A32**]{}, 751 (1976).
G. King and P.M. Woodward, J. Mater. Chem. [**20**]{}, 5785 (2010).
J. Kreisel, M. Alexe, and P.A. Thomas, Nature Materials [**11**]{}, 260 (2012).
I. Grinberg, D.V. West, M. Torres, G. Gou, D.M. Stein, L. Wu, G. Chen, E.M. Gallo, A.R. Akbashev, P.K. Davies, J.E. Spanier, and A.M. Rappe, Nature 503, 509 (2013).
A.L. Allred, J. Inorg. Nucl. Chem. [**17**]{}, 215 (1061).
Ph. Ghosez, E. Cockayne, U.V. Waghmare, and K.M. Rabe, Phys. Rev. B [**60**]{}, 836 (1999).
J.C. Wojdeł, P. Hermet, M.P. Ljungberg, Ph. Ghosez, and J. Íñiguez, J. Phys.: Condens. Matter [**25**]{}, 305401 (2013).
P.Y. Zavalij and M.S. Whittingham, Acta Cryst. B [**55**]{}, 627 (1999).
A.A. Belik, M. Azuma, T. Saito, Y. Shimakawa, and M. Takano, Chem. Mater. [**17**]{}, 269 (2005).
H. Béa, B. Dupé, S. Fusil, R. Mattana, E. Jacquet, B. Warot-Fonrose, F. Wilhelm, A. Rogalev, S. Petit, V. Cros, A. Anane, F. Petroff, K. Bouzehouane, G. Geneste, B. Dkhil, S. Lisenkov, I. Ponomareva, L. Bellaiche, M. Bibes, and A. Barthélémy, Phys. Rev. Lett. 102, 217603 (2009).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A method for approximate solution of spectral problems for Sturm-Liouville equations based on the construction of the Delsarte transmutation operators is presented. In fact the problem of numerical approximation of solutions and eigenvalues is reduced to approximation of a primitive of the potential by a finite linear combination of generalized wave polynomials introduced in [@KKTT], [@KT; @Transmut]. The method allows one to compute both lower and higher eigendata with an extreme accuracy.'
author:
- |
Vladislav V. Kravchenko and Sergii M. Torba\
[Departamento de Matemáticas, CINVESTAV del IPN, Unidad Querétaro, ]{}\
[Libramiento Norponiente No. 2000, Fracc. Real de Juriquilla, Querétaro, Qro. C.P. 76230 MEXICO]{}\
[e-mail: [email protected], [email protected] [^1]]{}
title: Analytic approximation of transmutation operators and applications to highly accurate solution of spectral problems
---
Introduction
============
Solution of Sturm-Liouville equations and of a wide range of related direct and inverse spectral problems is at the core of modern mathematical physics and its numerous applications. Since the work of J. Fourier on the theory of the heat and his method of separation of variables the properties and methods for solving different kinds of Sturm-Liouville spectral problems were studied in thousands of publications. One of the important mathematical tools for approaching problems related to Sturm-Liouville equations was introduced in 1938 by J. Delsarte [@Delsarte] and called [DelsarteLions1956]{} the transmutation operator. It relates two linear differential operators and allows one to transform a more complicated equation into a simpler one. Nowadays the transmutation operator is widely used in the theory of linear differential equations (see, e.g., [BegehrGilbert]{}, [@Carroll], [@HrynivMykytyuk2003], [@LevitanInverse], [@Marchenko], [@Sitnik], [@Trimeche]). Very often in literature the transmutation operators are called the transformation operators. It is well known that under certain regularity conditions the transmutation operator transmuting the operator $A=-\frac{d^{2}}{dx^{2}}+q(x)$ into $B=-\frac{d^{2}}{dx^{2}}$ is a Volterra integral operator with good properties. Its integral kernel can be obtained as a solution of a certain Goursat problem for the Klein-Gordon equation with a variable coefficient. In spite of their attractive properties and importance there exist very few examples of the transmutation kernels available in a closed form (see [@KrT2012]).
In the recent work [@KT; @Transmut] it was observed that the kernel of the transmutation operator relating the Schrödinger operator $A$ with $B$ is a complex component of a bicomplex-valued pseudoanalytic function of a hyperbolic variable [@APFT], a solution of a hyperbolic Vekua equation of a special form. The other component of that pseudoanalytic function is the transmutation kernel corresponding to a Darboux associated Schrödinger operator [@KrT2012]. This observation combined with some new results concerning such hyperbolic pseudoanalytic functions allowed us to obtain [@KT; @Transmut] a new and extremely convenient representation for the transmutation kernel in terms of so-called generalized wave polynomials [KKTT]{}.
In the present work we develop this result into a practical method for solving a wide spectrum of initial value, boundary value and spectral problems for the one-dimensional Schrödinger equation $Au=\lambda u$. The mentioned above convenience of the representation of the kernel consists in the fact that with respect to the variable of integration the kernel results to be a polynomial. Since to obtain solutions of the Schrödinger equation the transmutation operator is applied to the functions $\sin \sqrt{\lambda }t$ and $\cos \sqrt{\lambda }t$, solutions of the simplest such equation $Bv=\lambda v$, all the involved integrals are calculated explicitly. The coefficients of the polynomial are functions of another independent variable $x$. Due to the developed theory they can be found as a result of approximation of the pair of functions $g_{1}(x)=\frac{h}{2}+\frac{1}{4}\int_{0}^{x}q(s)ds$ and $g_{2}(x)=\frac{1}{4}\int_{0}^{x}q(s)ds$ in terms of a specially constructed family of functions. Here $h$ is a constant defined below. The family of functions is the family of traces of the generalized wave polynomials on the lines $x=t$ and $x=-t$ in the plane $(x,t)$.
Thus, the method developed and presented in this paper for solving spectral problems for the Schrödinger operator consists in the following steps.
1. Construction of a particular solution for the equation $Au=0$.
2. Construction of certain recursive integrals which serve for calculating the family of traces of the generalized wave polynomials.
3. Approximation of the pair of functions $g_{1}$ and $g_{2}$ by a linear combination of the traces.
4. Computation of the solution of the Schrödinger equation or of a characteristic function of the spectral problem in a domain of interest as a function of the spectral parameter $\lambda $.
5. Localization of the eigenvalues as zeros of the characteristic function.
Several advantages of the method should be emphasized.
- The error of the computed eigendata does not increase for higher eigenvalues. One can compute, e.g., the $1000$th eigenvalue and eigenfunction with roughly the same accuracy as the first ones.
- A quite simple a-priori control of the accuracy of the computed eigendata is available. As we show, the accuracy of the computed eigendata is entirely linked to the accuracy of approximation of the functions $g_{1}$ and $g_{2}$ on step 3. This error of approximation is easily calculated. If it is inadmissible, more approximating functions should be taken.
- The method works equally well in the case of complex-valued coefficients and spectral parameter dependent boundary conditions.
- The method allows one to obtain highly accurate eigendata. For example, for a standard test problem (the Paine problem [@Pryce]) we present eigendata corresponding to $\lambda _{1},\ldots,
\lambda _{10000}$ computed with the accuracy of the order $10^{-105}$. For the Coffey-Evans problem well known for being extremely difficult for numerical computation due to neatly clustered eigenvalues the achieved accuracy is of the order $10^{-65}$. Note that even for $\lambda =0$ the solution of the Coffey-Evans equation delivered by the built-in Mathematica’s function NDSolve was computed at best with an absolute error greater than $61$.
The main results presented in this paper are aimed to give a rigorous justification of the developed method. We prove that the central object of the transmutation kernel representation, the set of traces of the generalized wave polynomials is complete in the required functional spaces and therefore the set can be used for approximating the functions $g_{1}$ and $g_{2}$. We obtain corresponding estimates for the accuracy of approximation of the transmutation kernel. Further estimates concerning the accuracy of the resulting approximation of solutions of the Schrödinger equation and the independence of the accuracy of largeness of $\lambda $ are proved as well. We show that the same approximation coefficients obtained on step 3 can be used to obtain a related approximation for the transmutation kernel for the Darboux associated Schrödinger operator. This is important for considering problems involving derivatives in boundary conditions.
The numerical experiments were performed in Mathematica with multiple-precision arithmetic. The main reason for not restricting our computations to the machine precision consists in a fuller exploration of the method, its capabilities and features. For example, the results of involved computations obtained with a high precision allow us to study more completely the link between the accuracy of the approximation (step 3) and of the resulting computed eigendata. This study reveals that the absolute error of the approximation essentially coincides with the final error in the eigendata and decays exponentially with respect to the number $N$ of traces used, which for practical purposes means that an accurate approximation of the functions $g_{1}$ and $g_{2}$ by the system of traces of generalized polynomials guarantees the computation of arbitrarily high eigendata with the same accuracy. However the complete description of the class of potentials for which the approximation and eigendata errors decay exponentially and rigorous proofs of the observed relation between approximation errors and eigendata errors remain for the future research.
Due to the difficulties with construction of the kernels of transmutation operators there were very few attempts for their practical use in numerical solution of spectral problems. In this relation we mention the paper [@Boumenir2006] where certain analytic approximation formulas for the transmutation kernels were obtained. To our knowledge the present paper is a first publication offering an efficient and highly accurate (and to our opinion clearly promising) numerical algorithm based on the transmutation operators for solving spectral problems for the Schrödinger operator. One of the direct applications requiring a large number of eigendata computed with a considerable and non-decreasing accuracy arises from the Fourier method of separation of variables. According to the method the solution admits an analytic expression in the form of a series which in practice is known to be slowly convergent. The method presented here allows one to calculate partial sums of such series containing large numbers of terms in a fast and accurate manner.
In the next Section \[Section2\] we recall some definitions and properties concerning the transmutation operators. In Section \[Section3\] we introduce the system of generalized wave polynomials. In Section \[Section4\] we prove the completeness of their traces in appropriate functional spaces. In Section \[SectApprox\] we construct the approximate kernels of the transmutation operators and obtain corresponding estimates for their accuracy. In Section \[Section6\] we obtain the main result of the paper, the formulas for approximate solutions of the Schrödinger equation $Au=\lambda
u$ as well as for their derivatives and prove corresponding estimates for their accuracy. Section \[Section7\] is dedicated to the description of the algorithm and its numerical implementation, the proof of the uniform error bounds for the approximate zeros of characteristic functions of Sturm-Liouville spectral problems as well as to the presentation of numerical results.
Transmutation operators {#Section2}
=======================
We give a definition of a transmutation operator from [@KT; @Obzor] which is a modification of the definition proposed by Levitan [@LevitanInverse], adapted to the purposes of the present work. Let $E$ be a linear topological space and $E_{1}$ its linear subspace (not necessarily closed). Let $A$ and $B$ be linear operators: $E_{1}\rightarrow E$.
\[DefTransmut\] A linear invertible operator $T$ defined on the whole $E$ such that $E_{1}$ is invariant under the action of $T$ is called a transmutation operator for the pair of operators $A$ and $B$ if it fulfills the following two conditions.
1. Both the operator $T$ and its inverse $T^{-1}$ are continuous in $E$;
2. The following operator equality is valid $$AT=TB \label{ATTB}$$ or which is the same $$A=TBT^{-1}.$$
Our main interest concerns the situation when $A=-\frac{d^{2}}{dx^{2}}+q(x)$, $B=-\frac{d^{2}}{dx^{2}}$, and $q$ is a continuous complex-valued function. Hence for our purposes it will be sufficient to consider the functional space $E=C[a,b]$ with the topology of uniform convergence and its subspace $E_{1}$ consisting of functions from $C^{2}\left[ a,b\right] $. One of the possibilities to introduce a transmutation operator on $E$ was considered by Lions [@Lions57] and later on in other references (see, e.g., [@Marchenko]), and consists in constructing a Volterra integral operator corresponding to a midpoint of the segment of interest. As we begin with this transmutation operator it is convenient to consider a symmetric segment $[-b,b]$ and hence the functional space $E=C[-b,b]$. It is worth mentioning that other well known ways to construct the transmutation operators (see, e.g., [@LevitanInverse], [@Trimeche]) imply imposing initial conditions on the functions and consequently lead to transmutation operators satisfying (\[ATTB\]) only on subclasses of $E_{1}$. We introduce such transmutation operators below.
Thus, consider the space $E=C[-b,b]$. In [@CKT] and [@KrT2012] a parametrized family of transmutation operators for the defined above $A$ and $B$ was studied. Operators of this family can be realized in the form of the Volterra integral operator $$\mathbf{T}_{h} u(x)=u(x)+\int_{-x}^{x}\mathbf{K}(x,t;h)u(t)dt \label{Tmain}$$ where $\mathbf{K}(x,t;h)=\mathbf{H}\big(\frac{x+t}{2},\frac{x-t}{2};h\big)$, $h$ is a complex parameter, $|t|\le|x|\le b$ and $\mathbf{H}$ is the unique solution of the Goursat problem $$\frac{\partial^{2}\mathbf{H}(u,v;h)}{\partial u\,\partial v}=q(u+v)\mathbf{H}(u,v;h), \label{GoursatTh1}$$$$\mathbf{H}(u,0;h)=\frac{h}{2}+\frac{1}{2}\int_{0}^{u}q(s)\,ds,\qquad \mathbf{H}(0,v;h)=\frac{h}{2}. \label{GoursatTh2}$$ If the potential $q$ is continuously differentiable, the kernel $\mathbf{K}$ itself is a solution of the Goursat problem $$\label{GoursatKh1}
\left( \frac{\partial^{2}}{\partial x^{2}}-q(x)\right) \mathbf{K}(x,t;h)=\frac{\partial^{2}}{\partial t^{2}}\mathbf{K}(x,t;h),$$ $$\label{GoursatKh2}
\mathbf{K}(x,x;h)=\frac{h}{2}+\frac{1}{2}\int_{0}^{x}q(s)\,ds,\qquad\mathbf{K}(x,-x;h)=\frac{h}{2}.$$ If the potential $q$ is $n$ times continuously differentiable, the kernel $\mathbf{K}(x,t;h)$ is $n+1$ times continuously differentiable with respect to both independent variables.
In the case $h=0$ the operator $\mathbf{T}_{h}$ coincides with the transmutation operator studied in [@Marchenko Chap. 1, Sect. 2]. In [@LevitanInverse], [@Lions57], [@Trimeche] it was established that in the case $q\in C^{1}[-b,b]$ the Volterra-type integral operator is a transmutation in the sense of Definition \[DefTransmut\] on the space $C^{2}[-b,b]$ if and only if the integral kernel $\mathbf{K}(x,t)$ satisfies the Goursat problem , .
The following proposition shows that to be able to construct transmutation operators $\mathbf{T}_{h}$ or their integral kernels $\mathbf{K}_{h}$ for arbitrary values of the parameter $h$ it is sufficient to know the transmutation operator $\mathbf{T}_{h_{1}}$ or its integral kernel $\mathbf{K}_{h_{1}}$ for some particular parameter $h_{1}$.
\[PropChangeOfH\] The operators $\mathbf{T}_{h_{1}}$ and $\mathbf{T}_{h_{2}}$ are related by the equality $$\mathbf{T}_{h_{2}}u=\mathbf{T}_{h_{1}}\bigg[u(x)+\frac{h_{2}-h_{1}}{2}\int_{-x}^{x}u(t)\,dt\bigg]$$valid for any $u\in C[-b,b]$.
The corresponding integral kernels $\mathbf{K}(x,t;h_{1})$ and $\mathbf{K}(x,t;h_{2})$ are related as follows $$\mathbf{K}(x,t;h_{2})=\frac{h_{2}-h_{1}}{2}+\mathbf{K}(x,t;h_{1})+\frac{h_{2}-h_{1}}{2}\int_{t}^{x}\big(\mathbf{K}(x,s;h_{1})-\mathbf{K}(x,-s;h_{1})\big)\,ds.$$
The following theorem states that the operators $\mathbf{T}_{h}$ are indeed transmutations in the sense of Definition \[DefTransmut\].
\[Th Transmutation\] Let $q\in C[-b,b]$. Then the operator $\mathbf{T}_{h}$ defined by satisfies the equality $$\left( -\frac{d^{2}}{dx^{2}}+q(x)\right) \mathbf{T}_{h}[u]=\mathbf{T}_{h}\left[ -\frac{d^{2}}{dx^{2}}(u)\right] $$ for any $u\in C^{2}[-b,b]$.
\[RemTh\]$\mathbf{T}_{h}$ maps a solution $v$ of the equation $v^{\prime
\prime }+\omega ^{2}v=0$, where $\omega $ is a complex number, into a solution $u$ of the equation $$u^{\prime \prime }-q(x)u+\omega ^{2}u=0 \label{SLomega2}$$with the following correspondence of the initial values $
u(0)=v(0)$, $u^{\prime }(0)=v^{\prime }(0)+hv(0)$.
Following [@Marchenko] we introduce the notations$$K_{c}(x,t;h)=\mathbf{K}(x,t;h)+\mathbf{K}(x,-t;h)$$ where $h$ is a complex number, and $K_{s}(x,t;\infty )=\mathbf{K}(x,t;h)-\mathbf{K}(x,-t;h)$.
\[TcTsMapsSolutions\] Solutions $c(\omega ,x;h)$ and $s(\omega ,x;\infty )$ of equation satisfying the initial conditions $$\begin{gathered}
c(\omega ,0;h)=1,\qquad c_{x}^{\prime }(\omega ,0;h)=h \label{ICcos}\\
s(\omega ,0;\infty )=0,\qquad s_{x}^{\prime }(\omega ,0;\infty )=1
\label{ICsin}\end{gathered}$$can be represented in the form $$c(\omega ,x;h)=\cos \omega x+\int_{0}^{x}K_{c}(x,t;h)\cos \omega t\,dt
\label{c cos}$$and $$s(\omega ,x;\infty )=\frac{\sin \omega x}{\omega }+\int_{0}^{x}K_{s}(x,t;\infty )\frac{\sin \omega t}{\omega }\,dt. \label{s sin}$$
Denote by $$T_{c}u(x)=u(x)+\int_{0}^{x}K_{c}(x,t;h)u(t)dt $$ and$$T_{s}u(x)=u(x)+\int_{0}^{x}K_{s}(x,t;\infty )u(t)dt $$ the corresponding integral operators.
Recursive integrals and generalized wave polynomials {#Section3}
====================================================
Let $f\in C[a,b]$ be a complex valued function and $f(x)\neq 0$ for any $x\in \lbrack a,b]$. The interval $(a,b)$ is assumed being finite. Let us consider two sequences of recursive integrals$$X^{(0)}(x)\equiv 1,\qquad X^{(n)}(x)=n\int_{x_{0}}^{x}X^{(n-1)}(s)\left(
f^{2}(s)\right) ^{(-1)^{n}}\,\mathrm{d}s, \qquad x_{0}\in \lbrack a,b],\quad n=1,2,\ldots \label{Xn}$$and $$\widetilde{X}^{(0)}\equiv 1,\qquad \widetilde{X}^{(n)}(x)=n\int_{x_{0}}^{x}\widetilde{X}^{(n-1)}(s)\left( f^{2}(s)\right) ^{(-1)^{n-1}}\,\mathrm{d}s, \qquad
x_{0}\in \lbrack a,b],\quad n=1,2,\ldots . \label{Xtiln}$$
Define two families of functions $\left\{ \varphi _{k}\right\}
_{k=0}^{\infty }$ and $\left\{ \psi _{k}\right\} _{k=0}^{\infty }$ constructed according to the rules $$\varphi _{k}(x)=\begin{cases}
f(x)X^{(k)}(x), & k\text{\ odd}, \\
f(x)\widetilde{X}^{(k)}(x), & k\text{\ even},\end{cases}
\label{phik}$$and $$\psi_{k}(x)=\begin{cases}
\dfrac{\widetilde{X}^{(k)}(x)}{f(x)}, & k\text{\ odd,} \\
\dfrac{X^{(k)}(x)}{f(x)}, & k\text{\ even}.\end{cases}
\label{psik}$$
The following result obtained in [@KrCV08] (for additional details and simpler proof see [@APFT] and [@KrPorter2010]) establishes the relation of the system of functions $\left\{ \varphi _{k}\right\}
_{k=0}^{\infty }$ and $\left\{ \psi _{k}\right\} _{k=0}^{\infty }$ to the Sturm-Liouville equation.
\[ThGenSolSturmLiouville\] Let $q$ be a continuous complex valued function of an independent real variable $x\in \lbrack a,b]$ and $\lambda $ be an arbitrary complex number. Suppose there exists a solution $f$ of the equation $$f^{\prime \prime }-qf=0 \label{SLhom}$$on $(a,b)$ such that $f\in C^{2}(a,b)\cap C^{1}[a,b]$ and $f(x)\neq 0$ for any $x\in \lbrack a,b]$. Then the general solution $y\in C^{2}(a,b)\cap
C^{1}[a,b]$ of the equation $$y^{\prime \prime }-qy=\lambda y \label{SLlambda}$$on $(a,b)$ has the form $y=c_{1}y_{1}+c_{2}y_{2}$ where $c_{1}$ and $c_{2}$ are arbitrary complex constants, $$y_{1}=\sum_{k=0}^{\infty }\frac{\lambda ^{k}}{(2k)!}\varphi _{2k}\qquad
\text{and}\qquad y_{2}=\sum_{k=0}^{\infty }\frac{\lambda ^{k}}{(2k+1)!}\varphi _{2k+1} \label{u1u2}$$and both series converge uniformly on $[a,b]$ together with the series of the first derivatives which have the form$$\begin{gathered}
y_{1}^{\prime }=f^{\prime }+\sum_{k=1}^{\infty }\frac{\lambda ^{k}}{(2k)!}\left( \frac{f^{\prime }}{f}\varphi _{2k}+2k\,\psi _{2k-1}\right) \qquad
\text{and} \label{du1du2} \\
y_{2}^{\prime }=\sum_{k=0}^{\infty }\frac{\lambda ^{k}}{(2k+1)!}\left( \frac{f^{\prime }}{f}\varphi _{2k+1}+\left( 2k+1\right) \psi _{2k}\right) .\end{gathered}$$The series of the second derivatives converge uniformly on any segment $[a_{1},b_{1}]\subset (a,b)$.
Representations and , also known as the SPPS method (Spectral Parameter Power Series), present an efficient and highly competitive technique for solving a variety of spectral and scattering problems related to Sturm-Liouville equations. The first work implementing Theorem \[ThGenSolSturmLiouville\] for numerical solution was [KrPorter2010]{} and later on the SPPS method was used in a number of publications (see [@CKOR], [@ErbeMertPeterson2012], [@KKB], [KKRosu]{}, [@KiraRosu2010], [@KT; @Obzor] and references therein).
\[RemInitialValues\]It is easy to see that by definition the solutions $y_{1}$ and $y_{2}$ from (\[u1u2\]) satisfy the following initial conditions $$\begin{aligned}
y_{1}(x_{0})& =f(x_{0}), & y_{1}^{\prime }(x_{0})& =f^{\prime }(x_{0}), \\
y_{2}(x_{0})& =0, & y_{2}^{\prime }(x_{0})& =1/f(x_{0}).\end{aligned}$$
\[RemarkNonVanish\] It is worth mentioning that in the regular case the existence and construction of the required $f$ presents no difficulty. Indeed, let $q$ be real valued and continuous on $[a,b]$. Then (\[SLhom\]) possesses two linearly independent real-valued solutions $f_{1}$ and $f_{2} $ whose zeros alternate. Thus, one may choose $f=f_{1}+if_{2}$. Moreover, for the construction of $f_{1}$ and $f_{2}$ in fact the same SPPS method may be used [@KrPorter2010]. In the case of complex-valued coefficients the existence of a non-vanishing solution was shown in [@KrPorter2010 Remark 5].
In what follows we choose $x_{0}=0$.
\[Th Transmute\]Let $q$ be a continuous complex valued function of an independent real variable $x\in \lbrack -b,b]$ for which there exists a particular solution $f$ of such that $f\in C^{2}[-b,b]$, $f\neq 0$ on $[-b,b]$ and normalized as $f(0)=1$. Denote $h:=f^{\prime
}(0)\in \mathbb{C}$. Suppose $\mathbf{T}_{h}$ is the operator defined by and $\varphi _{k}$, $k\in \mathbb{N}_{0}$ are functions defined by . Then $$\mathbf{T}_{h}x^{k}=\varphi _{k}(x)\qquad \text{for any}\ k\in \mathbb{N}_{0}. \label{mapping powers 1}$$
The mapping property of the transmutation operator allows one to see that the SPPS representations from Theorem \[ThGenSolSturmLiouville\] are nothing but the images of Taylor expansions of the functions $\cosh \sqrt{\lambda }x$ and $\frac{1}{\sqrt{\lambda }}\sinh \sqrt{\lambda }x$ under the action of $\mathbf{T}_{h}$. Moreover, equality is behind a new method for solving Sturm-Liouville problems proposed in [@KT; @MMET; @2012] and based on the use of Tchebyshev polynomials for approximating trigonometric functions, combined with .
In what follows we assume that $f\in C^{2}[-b,b]$, $f\neq 0$ on $[-b,b]$, $f(0)=1$ and denote $h:=f^{\prime }(0)\in \mathbb{C}$. Any such function is associated with an operator $\mathbf{T}_{h}$. For convenience, from now on we will write $T_{f}$ instead of $\mathbf{T}_{h}$ and the integral kernel of $T_{f}$ will be denoted by $\mathbf{K}_{f}$. The kernel $K_{c}(x,t;h)$ will be denoted by $C_{f}(x,t)$ and the kernel $K_{s}(x,t;\infty )$ by $S_{f}(x,t) $.
Notice that $$C_{f}(x,t)=\mathbf{K}_{f}(x,t)+\mathbf{K}_{f}(x,-t) \label{Cf}$$and $$S_{f}(x,t)=\mathbf{K}_{f}(x,t)-\mathbf{K}_{f}(x,-t). \label{Sf}$$
The following functions introduced in [@KKTT]$$u_{0}=\varphi _{0}(x),\quad u_{2m-1}(x,t)=\sum_{\text{even }k=0}^{m}\binom{m}{k}\varphi _{m-k}(x)t^{k},\quad u_{2m}(x,t)=\sum_{\text{odd }k=1}^{m}\binom{m}{k}\varphi _{m-k}(x)t^{k}, \label{um}$$are called generalized wave polynomials (the wave polynomials are introduced below, in Example \[Ex Wave polynomials\]). The following parity relations hold for the generalized wave polynomials. $$u_{0}(x,-t)=u_{0}(x,t),\qquad u_{2n-1}(x,-t)=u_{2n-1}(x,t),\qquad
u_{2n}(x,-t)=-u_{2n}(x,t). \label{umParity}$$
For the values of the generalized wave polynomials on the characteristics $x=t$ and $x=-t$ we introduce the additional notations$$\mathbf{c}_{m}(x)=u_{2m-1}(x,x)=\sum_{\text{even }k=0}^{m}\binom{m}{k}x^{k}\varphi _{m-k}(x),\quad m=1,2,\ldots \text{ and }\mathbf{c}_{0}(x)=u_{0}(x,x)=f(x), \label{cm}$$$$\mathbf{s}_{m}(x)=u_{2m}(x,x)=\sum_{\text{odd }k=1}^{m}\binom{m}{k}x^{k}\varphi _{m-k}(x),\quad m=1,2,\ldots . \label{sm}$$As we show in Sections \[SectApprox\] and \[Section6\] the systems of functions $\left\{
\mathbf{c}_{m}\right\} _{m=0}^{\infty }$, $\left\{ \mathbf{s}_{m}\right\}
_{m=1}^{\infty }$ and $\left\{ u_{m}\right\} _{m=0}^{\infty }$ play a crucial role in the construction of the transmutation kernels and hence of the solutions of equation (\[SLlambda\]).
\[Ex Wave polynomials\]In a special case when $f\equiv 1$ we obtain that $\varphi _{k}(x)=x^{k}$, $k\in \mathbb{N}_{0}$ and $u_{k}(x,t)=p_{k}(x,t)$ where $p_{k}$ are wave polynomials [@KKTT Proposition 1] defined by the equalities $$p_{0}(x,t)=1,\quad p_{2m-1}(x,t)=\mathcal{R}\bigl((x+\mathbf{j}t)^{m}\bigr),\quad p_{2m}(x,t)=\mathcal{I}\bigl((x+\mathbf{j}t)^{m}\bigr),\ m\geq 1.$$Here $\mathbf{j}$ is a hyperbolic imaginary unit (see, e.g., [Lavrentyev and Shabat]{}, [@MotterRosa], [@Sobczyk] and [@APFT]): $\mathbf{j}^{2}=1$ and $\mathbf{j}\neq \pm 1$, $\mathcal{R}$ and $\mathcal{I}
$ are the real and the imaginary parts respectively of a corresponding hyperbolic number. The wave polynomials may also be written as follows $$p_{0}(x,t)=1,\quad p_{2m-1}(x,t)=\sum_{\mathrm{even}\text{ }k=0}^{m}\binom{m}{k}x^{m-k}t^{k},\quad p_{2m}(x,t)=\sum_{\mathrm{odd}\text{ }k=1}^{m}\binom{m}{k}x^{m-k}t^{k}. $$ We see that in this special case $\mathbf{c}_{m}(x)=\mathbf{s}_{m}(x)=2^{m-1}x^{m}$, $m=1,2,\ldots $.
Goursat-to-Goursat transmutation operators and completeness of the systems $\{\mathbf{c}_n\}$ and $\{\mathbf{s}_n\}$ {#Section4}
====================================================================================================================
By $\overline{\mathbf{S}}$ we denote a closed square with a diagonal joining the endpoints $(b,b)$ and $(-b,-b)$. Let $\square :=\partial
_{x}^{2}-\partial _{t}^{2}$ and the functions $\widetilde{u}$ and $u$ be solutions of the equations $\square \widetilde{u}=0$ and $\left( \square
-q(x)\right) u=0$ in $\overline{\mathbf{S}}$, respectively such that $u=T_{f}\widetilde{u}$. Following [@KT; @Transmut] we consider the operator $T_{G}$ mapping the Goursat data corresponding to $\widetilde{u}$ into the Goursat data corresponding to $u$, $$T_{G}:\quad \binom{\widetilde{u}(x,x)}{\widetilde{u}(x,-x)}\quad \longmapsto
\quad \binom{u(x,x)}{u(x,-x)}.$$The operator $T_{G}$ is well defined on the linear space $V$ of vector functions $\begin{pmatrix}
\varphi \\
\psi
\end{pmatrix}$ from $C^{1}[-b,b]\times C^{1}[-b,b]$ such that $\varphi (0)=\psi
(0)$, equipped with the maximum norm.
\[Prop Goursat to Goursat\] The operator $T_{G}$ together with $T_{G}^{-1}$ are bounded on $V$ and the action of the operator $T_{G}$ is defined by the following relation $$T_{G}
\begin{pmatrix}
\varphi (x) \\
\psi (x)
\end{pmatrix}=
\begin{pmatrix}
\varphi (x)+2\int_{0}^{x}\mathbf{K}_{f}(x,2t-x)\varphi (t)dt+\psi
(0)+2\int_{-x}^{0}\mathbf{K}_{f}(x,2t+x)\psi (t)dt-\varphi (0)f(x) \\
\varphi (0)+2\int_{-x}^{0}\mathbf{K}_{f}(x,2t+x)\varphi (t)dt+\psi
(x)+2\int_{0}^{x}\mathbf{K}_{f}(x,2t-x)\psi (t)dt-\varphi (0)f(x)\end{pmatrix}.$$In particular, $$\begin{aligned}
T_{G}:\ 2^{n-1}
\begin{pmatrix}
x^{n}\\
x^{n}\end{pmatrix}& \longmapsto \begin{pmatrix}
\mathbf{c}_{n}(x)\\
\mathbf{c}_{n}(x)\end{pmatrix},\\ T_{G}:\ 2^{n-1}\binom{x^{n}}{-x^{n}}& \longmapsto \begin{pmatrix}
\mathbf{s}_{n}(x)\\
-\mathbf{s}_{n}(x)\end{pmatrix} $$ for $n\in \mathbb{N}$ and $T_{G}: \left(\begin{smallmatrix}
1\\
1\end{smallmatrix}\right) \longmapsto \left(\begin{smallmatrix}
\mathbf{c}_{0}(x)\\
\mathbf{c}_{0}(x)\end{smallmatrix}\right)$.
Using the properties of the operator $T_{G}$ the following proposition was obtained in [@KT; @Transmut].
\[Prop Unif Conv u\] Let $u$ be a regular solution of the equation $$\left( \square -q(x)\right) u=0 \label{KG}$$in $\overline{\mathbf{S}}$ such that its Goursat data admit the following series expansions$$\frac{1}{2}\left( u(x,x)+u(x,-x)\right) =\sum_{n=0}^{\infty }a_{n}\mathbf{c}_{n}(x), $$ and $$\frac{1}{2}\left( u(x,x)-u(x,-x)\right) =\sum_{n=1}^{\infty }b_{n}\mathbf{s}_{n}(x), $$ both uniformly convergent on $[-b,b]$. Then for any $(x,t)\in \overline{\mathbf{S}}$, $$u(x,t)=a_{0}u_{0}(x,t)+\sum_{n=1}^{\infty }\left(
a_{n}u_{2n-1}(x,t)+b_{n}u_{2n}(x,t)\right) $$ and the series converges uniformly in $\overline{\mathbf{S}}$.
It is not difficult to prove the linear independence as well as the completeness of the families of functions $\left\{ \mathbf{c}_{n}\right\}
_{n=0}^{\infty }$ and $\left\{ \mathbf{s}_{n}\right\} _{n=1}^{\infty }$ in appropriate functional spaces. For this together with the operator $T_{G}$ it is convenient to consider the following its modification,$$G:=UT_{G}U \label{opG}$$where $$U:=\frac{1}{\sqrt{2}}
\begin{pmatrix}
1 & 1 \\
1 & -1
\end{pmatrix}.$$Notice that $U^{2}=
\begin{pmatrix}
1 & 0 \\
0 & 1
\end{pmatrix}$ and hence $U=U^{-1}$.
Let us consider the operator $G$ on the space $C^{1}[-b,b]\times
C_{0}^{1}[-b,b]$ where $C_{0}^{n}[-b,b]$ denotes a subspace of $C^{n}[-b,b]$ consisting of functions vanishing in the origin. The operator $G$ transforms the half-sum and the half-difference of Goursat data for solutions of the wave equation into their counterpart for solutions of (\[KG\]), $$G:\ \frac{1}{2}\begin{pmatrix}
\widetilde{u}(x,x)+\widetilde{u}(x,-x)\\
\widetilde{u}(x,x)-\widetilde{u}(x,-x)
\end{pmatrix} \longmapsto \frac{1}{2}\begin{pmatrix}
u(x,x)+u(x,-x)\\
u(x,x)-u(x,-x)\end{pmatrix}.$$It is bounded together with its inverse. Note that from Proposition [Prop Goursat to Goursat]{} we have$$\begin{aligned}
G:\ \begin{pmatrix}1\\0\end{pmatrix} &\longmapsto \begin{pmatrix}\mathbf{c}_{0}(x)\\0\end{pmatrix},
\label{map c0}\\
G:\ 2^{n-1}\begin{pmatrix}x^{n}\\
0\end{pmatrix} &\longmapsto \begin{pmatrix}\mathbf{c}_{n}(x)\\
0\end{pmatrix} \label{map cn}\end{aligned}$$and $$G:\ 2^{n-1}\begin{pmatrix}0\\
x^{n}\end{pmatrix} \longmapsto \begin{pmatrix}0\\
\mathbf{s}_{n}(x)\end{pmatrix}. \label{map sn}$$
1. The operator $G$ defined by on $C^{1}[-b,b]\times
C_{0}^{1}[-b,b]$ admits the following representation $$G\begin{pmatrix}\eta (x)\\
\xi (x)\end{pmatrix}=\begin{pmatrix}G_{+}\left[ \eta (x)-\frac{\eta (0)
}{2}\right] +\frac{\eta (0)}{2}\\
G_{-}\left[ \xi (x)\right] \end{pmatrix} $$ where $G_{+}$ and $G_{-}$ have the form$$G_{\pm }\eta (x)=\eta (x)+\int_{-x}^{x}\mathcal{K}_{\pm }(x,t)\eta (t)dt
\label{Gform1}$$with the kernels given by the equalities$$\mathcal{K}_{\pm }(x,t):=\begin{cases}
\pm 2\mathbf{K}_{f}(x,2t+x), & -x\leq t<0, \\
2\mathbf{K}_{f}(x,2t-x) & 0\leq t\leq x.
\end{cases}$$The operators $G_{+}$ and $G_{-}$ can also be written in the form $$G_{\pm }\eta (x)=\eta (x)+\int_{-x}^{x}\mathbf{K}_{f}(x,t)\left( \eta \left(\frac{t+x}{2}\right)\pm \eta \left(\frac{t-x}{2}\right)\right) dt. \label{Gform2}$$
2. Both operators $G_{+}$ and $G_{-}$ preserve the value of the function in the origin and $G_{+}:C^{1}[-b,b]\rightarrow C^{1}[-b,b]$, $G_{-}:C_{0}^{1}[-b,b]\rightarrow C_{0}^{1}[-b,b]$.
3. There exist the inverse operators $G_{+}^{-1}$ and $G_{-}^{-1}$ defined on $C^{1}[-b,b]$ and $C_{0}^{1}[-b,b]$, respectively, and the inverse operator for $G$ admits the representation $$G^{-1}\begin{pmatrix}\eta (x)\\
\xi (x)\end{pmatrix}=\begin{pmatrix}G_{+}^{-1}\left[ \eta (x)-\frac{\eta
(0)}{2}\right] +\frac{\eta (0)}{2}\\
G_{-}^{-1}\left[ \xi (x)\right] \end{pmatrix}.
\label{G-1}$$
1\) Let us observe that (\[Gform2\]) is obtained from (\[Gform1\]) by a simple change of variables. From Proposition \[Prop Goursat to Goursat\] we obtain $$G\binom{\eta (x)}{\xi (x)}=\binom{G_{+}\left[ \eta (x)\right] +\eta
(0)(1-f(x))}{G_{-}\left[ \xi (x)\right] }$$and observe that if $c$ is a complex constant then $G_{+}\left[ c\right]
=2c(f(x)-1)+c$. Indeed, from (\[Gform2\]) we have that $$G_{+}\left[ c\right] =c+2c\int_{-x}^{x}\mathbf{K}_{f}(x,t)dt=2cT_{f}\left[ 1\right] -c=2cf(x)-c \label{G+c}$$where we used (\[mapping powers 1\]). Hence $G_{+}\left[ \frac{\eta (0)}{2}\right] =\eta (0)(f(x)-1)+\frac{\eta (0)}{2}$.
The proof of 2) follows directly from (\[Gform2\]).
3\) The existence of the inverse operators $G_{+}^{-1}$ and $G_{-}^{-1}$ follows from (\[Gform1\]), the fact that the kernels $\mathcal{K}_{\pm }$ are bounded measurable functions and a well known result on the invertibility of Volterra operators with such kernels (see [@KolmFom Ch. 9, Subsect. 4.5]). The representation (\[G-1\]) can be verified directly using 2) and (\[G+c\]).
Observe that $G$ (and hence $G^{-1}$) results to be a diagonal operator, $$G=\begin{pmatrix}
G_{1} & 0 \\
0 & G_{2}\end{pmatrix}$$with $G_{1}=G_{+}(I-\frac{1}{2}\delta )+\frac{1}{2}\delta $ and $G_{2}=G_{-}$ where $I$ is the identity operator and $\delta $ is the functional acting as follows $\delta \left[ \eta (x)\right] =\eta (0)$.
From (\[map c0\]), (\[map cn\]) and (\[map sn\]) we have $G_{1}\left[ 1\right] =\mathbf{c}_{0}(x)=f(x)$, $2^{n-1}G_{1}\left[ x^{n}\right] =\mathbf{c}_{n}(x)$ and $2^{n-1}G_{2}\left[ x^{n}\right] =\mathbf{s}_{n}(x)$ for $n\in
\mathbb{N}$.
The systems of functions $\left\{ \mathbf{c}_{n}\right\} _{n=0}^{\infty }$ and $\left\{ \mathbf{s}_{n}\right\} _{n=1}^{\infty }$ are linearly independent and complete in $C^{1}[-b,b]\ $and $C_{0}^{1}[-b,b]$ respectively.
To prove the linear independence of functions $\mathbf{c}_{n}$ consider a nontrivial linear combination $\mathbf{c}:=a_{0}\mathbf{c}_{0}+\ldots +a_{N}\mathbf{c}_{N}$ and suppose that $\mathbf{c}\equiv 0$ on $[-b,b]$. We have then $G_{1}^{-1}\left[ \mathbf{c}(x)\right] =a_{0}+\ldots +2^{N-1}a_{N}x^{N}$ and hence $a_{0}+\ldots +2^{N-1}a_{N}x^{N}\equiv 0$ on $[-b,b]$ which is a contradiction. The linear independence of functions $\mathbf{s}_{n}$ is proved analogously.
The completeness of $\left\{ \mathbf{c}_{n}\right\} _{n=0}^{\infty }$ in $C^{1}[-b,b]\ $follows from the completeness of the powers $\left\{
x^{n}\right\} _{n=0}^{\infty }$ in $C^{1}[-b,b]\ $and properties of $G_{1}$. Analogously, the completeness of $\left\{ \mathbf{s}_{n}\right\}
_{n=1}^{\infty }$ in $C_{0}^{1}[-b,b]\ $follows from the completeness of the powers $\left\{ x^{n}\right\} _{n=1}^{\infty }$ in $C_{0}^{1}[-b,b]\ $and properties of $G_{2}$.
Approximate construction of integral kernels\[SectApprox\]
==========================================================
\[Th Kapprox\]Let the complex numbers $a_{0},\ldots ,a_{N}$ and $b_{1},\ldots ,b_{N}$ be such that$$\label{KxxErr}
\left\vert \frac{h}{2}+\frac{1}{4}\int_{0}^{x}q(s)ds-\sum_{n=0}^{N}a_{n}\mathbf{c}_{n}(x)\right\vert <\varepsilon _{1}$$and $$\label{KxmxErr}
\left\vert \frac{1}{4}\int_{0}^{x}q(s)ds-\sum_{n=1}^{N}b_{n}\mathbf{s}_{n}(x)\right\vert <\varepsilon _{2}$$for every $x\in \lbrack -b,b]$. Then the kernel $\mathbf{K}_{f}(x,t)$ is approximated by the function$$K_{f,N}(x,t)=a_{0}u_{0}(x,t)+\sum_{n=1}^{N}a_{n}u_{2n-1}(x,t)+\sum_{n=1}^{N}b_{n}u_{2n}(x,t) \label{K(x,t)}$$in such a way that for every $(x,t)\in \overline{\mathbf{S}}$ the inequality holds $$\bigl|\mathbf{K}_{f}(x,t)-K_{f,N}(x,t)\bigr|\leq 3\Vert T_{f}\Vert \cdot
\Vert T_{G}^{-1}\Vert (\varepsilon _{1}+\varepsilon _{2}). \label{estim}$$
Notice that $g_{1}(x):=\frac{h}{2}+\frac{1}{4}\int_{0}^{x}q(s)ds=\frac{1}{2}\left( \mathbf{K}_{f}(x,x)+\mathbf{K}_{f}(x,-x)\right) $ and $g_{2}(x):=\frac{1}{4}\int_{0}^{x}q(s)ds=\frac{1}{2}\left( \mathbf{K}_{f}(x,x)-\mathbf{K}_{f}(x,-x)\right) $ and hence the theorem establishes that if the half-sum and the half-difference of the Goursat data corresponding to $\mathbf{K}_{f}(x,t)$ are approximated by linear combinations of the functions $\mathbf{c}_{n}$ and $\mathbf{s}_{n}$ respectively, the function (\[K(x,t)\]) approximates uniformly the kernel $\mathbf{K}_{f}(x,t)$. Indeed, consider the functions $\widetilde{\mathbf{K}}_{f}=T_{f}^{-1}\mathbf{K}_{f}$ and $\widetilde{K}_{f,N}=T_{f}^{-1}K_{f,N}$. Then by the definition of the Goursat-to-Goursat transmutation operator $$\begin{pmatrix}
\widetilde{\mathbf{K}}_{f}(x,x) \\
\widetilde{\mathbf{K}}_{f}(x,-x)\end{pmatrix}=T_{G}^{-1}\begin{pmatrix}
\mathbf{K}_{f}(x,x) \\
\mathbf{K}_{f}(x,-x)\end{pmatrix}\quad \text{and}\quad
\begin{pmatrix}
\widetilde{K}_{f,N}(x,x) \\
\widetilde{K}_{f,N}(x,-x)\end{pmatrix}=T_{G}^{-1}\begin{pmatrix}
K_{f,N}(x,x) \\
K_{f,N}(x,-x)\end{pmatrix},$$hence due to the boundedness of the operator $T_{G}^{-1}$ $$\begin{gathered}
\max \biggl(\max_{x\in \lbrack -b,b]}\bigl|\widetilde{\mathbf{K}}_{f}(x,x)-\widetilde{K}_{f,N}(x,x)\bigr|,\max_{x\in \lbrack -b,b]}\bigl|\widetilde{\mathbf{K}}_{f}(x,-x)-\widetilde{K}_{f,N}(x,-x)\bigr|\biggr) \\
\leq \Vert T_{G}^{-1}\Vert \max \biggl(\max_{x\in \lbrack -b,b]}\bigl|\mathbf{K}_{f}(x,x)-K_{f,N}(x,x)\bigr|,\max_{x\in \lbrack -b,b]}\bigl|\mathbf{K}_{f}(x,-x)-K_{f,N}(x,-x)\bigr|\biggr) \displaybreak[2]\\
\leq \Vert T_{G}^{-1}\Vert \max_{x\in \lbrack -b,b]}\biggl(\biggl|\frac{1}{2}\bigl(\mathbf{K}_{f}(x,x)+\mathbf{K}_{f}(x,-x)\bigr)-\frac{1}{2}\bigl(K_{f,N}(x,x)+K_{f,N}(x,-x)\bigr)\biggr| \\
+\biggl|\frac{1}{2}\bigl(\mathbf{K}_{f}(x,x)-\mathbf{K}_{f}(x,-x)\bigr)-\frac{1}{2}\bigl(K_{f,N}(x,x)-K_{f,N}(x,-x)\bigr)\biggr|\biggr)<\Vert
T_{G}^{-1}\Vert (\varepsilon _{1}+\varepsilon _{2}),\end{gathered}$$where we have used the equalities $\frac{1}{2}\bigl(K_{f,N}(x,x)+K_{f,N}(x,-x)\bigr)=a_{0}u_{0}(x,x)+\sum_{n=1}^{N}a_{n}u_{2n-1}(x,x)$ and $\frac{1}{2}\bigl(K_{f,N}(x,x)-K_{f,N}(x,-x)\bigr)=\sum_{n=1}^{N}b_{n}u_{2n}(x,x)$. We obtain from the proof of [@KKTT Theorem 3] that for every $(x,t)\in \overline{\mathbf{S}}$ $$\bigl|\widetilde{\mathbf{K}}_{f}(x,t)-\widetilde{K}_{f,N}(x,t)\bigr|\leq
3\Vert T_{G}^{-1}\Vert (\varepsilon _{1}+\varepsilon _{2}),$$hence for every $(x,t)\in \overline{\mathbf{S}}$ $$\bigl|\mathbf{K}_{f}(x,t)-K_{f,N}(x,t)\bigr|\leq 3\Vert T_{f}\Vert \cdot
\Vert T_{G}^{-1}\Vert (\varepsilon _{1}+\varepsilon _{2}).\qedhere$$
It is not difficult to make the estimate (\[estim\]) more explicit. For example, let the function $\mathbf{H}(u,v)$ be a solution of the Goursat problem for equation (\[GoursatTh1\]) with the conditions $\mathbf{H}(u,0)=\varepsilon _{1}(u)$ and $\mathbf{H}(0,v)=\varepsilon _{2}(v)$, $\varepsilon_1(0)=\varepsilon_2(0)=\varepsilon_0$, where $\varepsilon _{1}(u)=\frac{h}{2}+\frac{1}{2}\int_{0}^{u}q(s)ds-
\sum_{n=0}^{N}a_{n}\mathbf{c}_{n}(u)-\sum_{n=1}^{N}b_{n}\mathbf{s}_{n}(u)$ and $\varepsilon _{2}(v)=\frac{h}{2}-\sum_{n=0}^{N}a_{n}\mathbf{c}_{n}(v)+\sum_{n=1}^{N}b_{n}\mathbf{s}_{n}(v)$ are the differences between the exact and the approximate transmutation kernels on the characteristics. Then similarly to [@Vladimirov Subsect. 15.1] one can see that the Goursat problem is equivalent to the integral equation $\mathbf{H}(u,v)=\int_0^u\int_0^v q(u'+v')\mathbf{H}(u',v')\,du'\,dv'+\varepsilon_1(u)+\varepsilon_2(v)-\varepsilon_0$. Applying the successive approximations technique one obtains for $\mathbf{H}$ the following estimate $\max|\mathbf{H}(u,v)|\le m I_0(2\kappa\sqrt{|uv|})$, where $I_0$ is the modified Bessel function of the first kind, $\kappa =\sqrt{\max \left\vert q\right\vert}$ and $m= \max \left\vert \varepsilon _{1}\right\vert +\max \left\vert
\varepsilon _{2}\right\vert+\max \left\vert \varepsilon
_{0}\right\vert$. Thus, $\bigl|\mathbf{K}
_{f}(x,t)-K_{f,N}(x,t)\bigr|\leq mI_0 (\kappa x)$.
\[Rm HalfSegment\] Let us notice that the approximation of the transmutation kernel $\mathbf{K}_{f}(x,t)$ in the form (\[K(x,t)\]) implies the following approximations of the kernels $$S_{f}(x,t)\cong S_N(x,t):=2\sum_{n=1}^{N}b_{n}u_{2n}(x,t) \label{Sapprox}$$and $$C_{f}(x,t)\cong C_N(x,t):=2\left(
a_{0}u_{0}(x,t)+\sum_{n=1}^{N}a_{n}u_{2n-1}(x,t)\right) . \label{Capprox}$$This is a direct corollary of Theorem \[Th Kapprox\], formulas (\[Cf\]), (\[Sf\]) and (\[umParity\]). Notice that to obtain the coefficients $a_{n}$ and $b_{n}$ in (\[Sapprox\]) and (\[Capprox\]) the approximation of $g_{1}$ and $g_{2}$ can be performed on $[0,b]$ only and hence for $q\in
C[0,b]$.
The approximation of the functions $g_{1}$ and $g_{2}$ by the corresponding combinations of the functions $\mathbf{c}_{n}$ and $\mathbf{s}_{n}$ can be done in several ways. For example, the least squares method can be used to obtain a reasonably good approximation. Even though its not clear how to verify whether the systems of functions $\mathbf{c}_n$ and $\mathbf{s}_n$ are Tchebyshev systems, in the case when all the involved functions are real valued the Remez algorithm can be used, see [@KKTT Section 6] and references therein. Another alternative is to reformulate the approximation problem as a linear programming problem and solve it.
Approximate solution {#Section6}
====================
Let us explain the special convenience of the approximations of the transmutation kernels in terms of the generalized wave polynomials. Consider, for example, the kernel $S_{f}(x,t)$ which by (\[s sin\]) transforms the function $\frac{\sin \omega x}{\omega }$ into a solution of (\[SLomega2\]) satisfying the initial conditions (\[ICsin\]) (see Theorem \[TcTsMapsSolutions\]). Now instead of the exact kernel $S_{f}(x,t)$ let us substitute into (\[s sin\]) its approximation $S_N(x,t)$. We obtain an approximate solution $$\begin{split}
s(\omega ,x;\infty ) \cong s_{N}(\omega ,x) &=\frac{\sin \omega x}{\omega }+\int_{0}^{x}S_N(x,t)\frac{
\sin \omega t}{\omega }\,dt \\
&=\frac{\sin \omega x}{\omega }+2\int_{0}^{x}\sum_{n=1}^{N}b_{n}u_{2n}(x,t)\frac{\sin \omega t}{\omega }\,dt.
\end{split}$$ By the definition of the generalized wave polynomials we have $$\label{sN}
s_{N}(\omega ,x)=\frac{1}{\omega }\left( \sin \omega
x+2\sum_{n=1}^{N}b_{n}\sum_{\text{odd }k=1}^{n}\binom{n}{k}\varphi
_{n-k}(x)\int_{0}^{x}t^{k}\sin \omega t\,dt\right) .$$ The integrals here are, of course, easily calculated explicitly. For example, the following formula can be used [@GradRizh 2.633] $$\label{Int_tsin}
\int_{{}}^{{}}t^{k}\sin (\omega t)dt=-\sum\limits_{j=0}^{k}j!\binom{k}{j}\frac{t^{k-j}}{\omega ^{j+1}}\cos \left( \omega t+\frac{j\pi }{2}\right),$$or alternatively the integrals can be calculated recursively.
Analogously, the approximation of the solution $c(\omega ,x;h)$ is calculated as follows $$\label{cN}
\begin{split}
c(\omega ,x;h) \cong c_{N}(\omega ,x) &=\cos \omega x+\int_{0}^{x}C_N(x,t)\cos \omega t\,dt \\
&=\cos \omega x+2\sum_{n=0}^{N}a_{n}\sum_{\text{even }k=0}^{n}\binom{n}{k}\varphi
_{n-k}(x)\int_{0}^{x}t^{k}\cos \omega t\,dt
\end{split}$$ where the integrals can be calculated exactly using the formula [GradRizh]{} $$\label{Int_tcos}
\int_{{}}^{{}}t^{k}\cos (\omega t)dt=\sum\limits_{j=0}^{k}j!\binom{k}{j}\frac{t^{k-j}}{\omega ^{j+1}}\sin \left( \omega t+\frac{j\pi }{2}\right) .$$Thus, the problem of approximate solution of equation (\[SLomega2\]) can be reduced to the problem of approximation of the functions $g_{1}$, $g_{2}$ in terms of the functions $\mathbf{c}_{n}$ and $\mathbf{s}_{n}$ respectively.
For a real $\omega $ we have that the accuracy of the approximate solution does not deteriorate when $\omega $ increases. Indeed, considering, e.g., $\left\vert
c(\omega ,x;h)-c_{N}(\omega ,x)\right\vert $ under the assumption $\left\vert C_{f}(x,t)-C_N(x,t)\right\vert \leq \varepsilon $ we have$$\left\vert c(\omega ,x;h)-c_{N}(\omega ,x)\right\vert \leq
\int_{0}^{x}\left\vert C_{f}(x,t)-C_N(x,t)\right\vert \left\vert \cos \omega
t\,\right\vert dt\leq \varepsilon \int_{0}^{x}\left\vert \cos \omega
t\,\right\vert dt\leq \varepsilon \left\vert x\right\vert .$$
A similar observation is true also for complex values of the parameter $\omega $. Indeed, for an arbitrary complex valued potential $q\in C[0,b]$ and arbitrary nondegenerate boundary conditions the asymptotic formulas for the square roots of eigenvalues (see [@Marchenko Chapter 1, Sect.5]) tell us that all the square roots of eigenvalues are located in a strip on a complex plane parallel to the real axis. For example, in the case of the problem for (\[SLomega2\]) with the boundary conditions $u(0)=u(\pi )=0$ one has that $\omega _{k}=k+\theta (k)$ where $\theta (k)\rightarrow 0$ when $k\rightarrow \infty $ [@Marchenko p. 69]. Analogous asymptotic formulas are available for all other nondegenerate boundary conditions. Thus, solution of a nondegenerate Sturm-Liouville problem implies consideration of solutions $c(\omega ,x;h)$ and $s(\omega ,x;\infty )$ for $\omega $ with $\operatorname{Im}\omega $ belonging to a finite interval. The following statement establishes that similarly to the case $\operatorname{Im}\omega
=0$, when $\left\vert \operatorname{Im}\omega \right\vert \leq \mathrm{Const}$ the accuracy of approximation does not depend on $\omega $.
Let the parameter $\omega $ belong to the strip $\left\vert \operatorname{Im}\omega
\right\vert \leq C$ where $C$ is a positive number. Suppose that $\max_{\overline{S}}\left\vert C_{f}(x,t)-C_N(x,t)\right\vert \leq \varepsilon $. Then $$\left\vert c(\omega ,x;h)-c_{N}(\omega ,x)\right\vert \leq \varepsilon
\sinh (Cx)/C. \label{estimate c}$$
Consider $$\begin{aligned}
\left\vert c(\omega ,x;h)-c_{N}(\omega ,x)\right\vert &\leq \varepsilon
\int_{0}^{x}\left\vert \cos \omega t\,\right\vert dt=\frac \varepsilon 2
\int_{0}^{x}\left\vert e^{i\operatorname{Re}\omega\cdot t}e^{-\operatorname{Im}\omega\cdot t}+e^{-i\operatorname{Re}\omega\cdot t}e^{\operatorname{Im}\omega\cdot t} \right\vert dt \\
&\leq \frac \varepsilon 2 \int_{0}^{x}\left( e^{\operatorname{Im}\omega\cdot t}
+ e^{-\operatorname{Im}\omega\cdot t} \right) dt =
\varepsilon \int_{0}^{x}\cosh \left( |\operatorname{Im}\omega| \cdot t\right) dt =\frac{\varepsilon\sinh\left(|\operatorname{Im}\omega|\cdot x\right)}{|\operatorname{Im}\omega|}.\end{aligned}$$ Since the function $\sinh (\xi x)/\xi $ is monotonically increasing with respect to both variables when $\xi ,x\geq 0$, we obtain the required inequality (\[estimate c\]).
A similar statement is true for the solution $s(\omega ,x;\infty )$.
Let the parameter $\omega $ belong to the strip $\left\vert \operatorname{Im}\omega
\right\vert \leq C$ where $C$ is a positive number and $\left\vert \omega
\right\vert >1$. Suppose that $\max_{\overline{S}}\left\vert
S_{f}(x,t)-S_N(x,t)\right\vert \leq \varepsilon $. Then $$\left\vert s(\omega ,x;\infty )-s_{N}(\omega ,x)\right\vert \leq
\varepsilon \sinh (Cx)/C. \label{estimate s}$$If $\left\vert \omega \right\vert \leq 1$, then $$\left\vert s(\omega,x;\infty)-s_{N}(\omega,x)\right\vert \leq\varepsilon
c_{b}x$$ where the constant $c_{b}$ depends only on $b$. For any $\omega\neq0$ the following estimate holds $$\left\vert s(\omega,x;\infty)-s_{N}(\omega,x)\right\vert \leq\frac
{\varepsilon}{\left\vert \omega\right\vert }\sinh(Cx)/C.
\label{estimate s/omega}$$
Let $\left\vert \operatorname{Im}\omega \right\vert \leq C$ and $\left\vert \omega
\right\vert >1$. Then $$\left\vert s(\omega ,x;\infty )-s_{N}(\omega ,x)\right\vert \leq
\left\vert \omega \right\vert \left\vert s(\omega ,x;\infty )-s_{N}(\omega
,x)\right\vert \leq \varepsilon \int_{0}^{x}\left\vert \sin \omega t\,\right\vert dt.$$Now following the reasoning from the proof of the preceding proposition we obtain (\[estimate s\]).
Considering the case $\left\vert \omega \right\vert \leq 1$ we observe that the function $\sin \left( \omega t\right) /\omega $ is analytic with respect to $\omega $ and hence $\max_{\left\vert \omega \right\vert \leq
1}\left\vert \sin \left( \omega t\right) /\omega \right\vert
=\max_{\left\vert \omega \right\vert =1}\left\vert \sin \left( \omega
t\right) /\omega \right\vert =\max_{\left\vert \omega \right\vert
=1}\left\vert \sin \left( \omega t\right) \right\vert $. Denote this number by $c(t)$. Again, due to the maximum principle we obtain $c(t)\leq
c(b)=:c_{b}$. Thus, $\left\vert s(\omega ,x;\infty )-s_{N}(\omega
,x)\right\vert \leq \varepsilon \int_{0}^{x}\left\vert \frac{\sin \omega t}{\omega }\,\right\vert dt\leq \varepsilon c_{b}x$. The estimate (\[estimate s/omega\]) is proved in a complete analogy with (\[estimate c\]).
In order to be able to consider problems for equation (\[SLlambda\]) with boundary conditions involving the derivative of the solution we need to obtain a convenient representation for it as well. Let $u_{N}$ be an approximation of a solution $u$ of (\[SLlambda\]), defined by the formula$$u_{N}(x)=\widetilde{u}(x)+\int_{-x}^{x}K_{f,N}(x,t)\widetilde{u}(t)dt
\label{uN}$$where $\widetilde{u}$ is a linear combination of the functions $\sin (\omega
x)/\omega $ and $\cos (\omega x)$, and $K_{f,N}$ has the form (\[K(x,t)\]). The corresponding exact solution has the form $$u(x)=\widetilde{u}(x)+\int_{-x}^{x}\mathbf{K}_{f}(x,t)\widetilde{u}(t)dt.$$From (\[uN\]) it is quite easy to obtain a corresponding expression for $u_{N}^{\prime }$ observing that the differentiation of (\[K(x,t)\]) where the functions $u_{k}(x,t)$ are defined by (\[um\]) does not present any difficulty. Nevertheless one still has to prove a convenient estimate for the difference $\left\vert u^{\prime }-u_{N}^{\prime }\right\vert $ which is an additional task. To make it easier we choose here another way which involves the transmutation operator for the Darboux associated Schrödinger equation which in our notations is the operator $T_{1/f}$. In [KrT2012]{} this operator was studied in detail and two explicit formulae for the corresponding integral kernel $\mathbf{K}_{1/f}(x,t)$ in terms of $\mathbf{K}_{f}(x,t)$ were obtained (a third formula was presented in [KT Transmut]{}).
Consider the function $$v:=f\left( \frac{u}{f}\right) ^{\prime }=u^{\prime }-\frac{f^{\prime }}{f}u
\label{v}$$which is a result of the Darboux transformation applied to $u$ (more on the Darboux transformation see, e.g., [@Matveev] and [@Rosu]). The function $v$ is a solution of the equation $$v^{\prime \prime }-q_{D}(x)v=\lambda v$$with $q_{D}=-q+2\left( f^{\prime }/f\right) ^{2}$. As was shown in [KrT2012]{}, $v$ can be represented in the form $$v=T_{1/f}\widetilde{v} \label{vT1/f}$$where $\widetilde{v}$ is a solution of the equation $\widetilde{v}^{\prime
\prime }=\lambda \widetilde{v}$, that is, $\widetilde{v}$ is a linear combination of the functions $\sin (\omega x)/\omega $ and $\cos (\omega x)$ with $\omega ^{2}=-\lambda $, $$\widetilde{v}=\alpha \cos (\omega x)+\beta \sin (\omega x)/\omega .$$Assuming that $$u(x)=ac(\omega ,x;h)+bs(\omega ,x;\infty ) \label{sol u}$$ or, what is the same $\widetilde{u}=a\cos (\omega x)+b\sin (\omega x)/\omega
$, let us find the coefficients $\alpha $ and $\beta $ in terms of $a$ and $b $. From (\[v\]) we have $$v(x)=a\left( c^{\prime }(\omega ,x;h)-\frac{f^{\prime }}{f}c(\omega
,x;h)\right) +b\left( s^{\prime }(\omega ,x;\infty )-\frac{f^{\prime }}{f}s(\omega ,x;\infty )\right) \label{v_ab}$$meanwhile from (\[vT1/f\]) we obtain $$v(x)=T_{1/f}\left[ \alpha \cos (\omega x)+\frac{\beta }{\omega }\sin (\omega
x)\right] .$$The last relation can obviously be written as follows $$v(x)=T_{1/f}\left[ \left( \frac{\alpha }{\omega }\sin (\omega x)-\frac{\beta
}{\omega ^{2}}\cos (\omega x)\right)'\right] . \label{vT1/fd}$$In [@KrT2012] the following useful operator equality was obtained $$\frac{1}{f}T_{1/f}\frac{d}{dx}=\frac{d}{dx}\frac{1}{f}T_{f}$$which is true on $C^{1}[-b,b]$. Applying it to (\[vT1/fd\]) we obtain$$\begin{aligned}
v(x) &=&f(x)\frac{d}{dx}\left( \frac{1}{f(x)}T_{f}\left[ \frac{\alpha }{\omega }\sin (\omega x)-\frac{\beta }{\omega ^{2}}\cos (\omega x)\right]
\right) \\
&=&f(x)\frac{d}{dx}\left( \frac{1}{f(x)}\left[ \alpha s(\omega ,x;\infty )-\frac{\beta }{\omega ^{2}}c(\omega ,x;h)\right] \right) .\end{aligned}$$Comparison of this result with (\[v\_ab\]) gives us the relations $\alpha =b$ and $\beta =-\omega ^{2}a$. Hence $$v(x)=T_{1/f}\left[ b\cos (\omega x)-a\omega \sin (\omega x)\right] .
\label{vT1}$$Notice that $v(0)=b$ and $v^{\prime }(0)=-a\omega ^{2}-bh$. This follows from the properties of the operator $T_{1/f}$ (Remark \[RemTh\]) and observation that the value of $\left( 1/f\right) ^{\prime }$ in the origin is $-h$.
From (\[vT1\]) we obtain a convenient representation for the derivative of the solution (\[sol u\]),$$u^{\prime }=v+\frac{f^{\prime }}{f}u=T_{1/f}\left[ b\cos (\omega x)-a\omega
\sin (\omega x)\right] +\frac{f^{\prime }}{f}T_{f}\left[ a\cos (\omega x)+\frac{b}{\omega }\sin (\omega x)\right] . \label{uprime}$$
An approximation $K_{1/f,N}$ of the kernel $\mathbf{K}_{1/f}$ of the operator $T_{1/f}$ can be done repeating the general scheme of Theorem \[Th Kapprox\] for the Darboux associated potential $q_D$ and the particular solution $1/f$. However it is possible to omit the solution of another approximation problem, an approximation $K_{1/f,N}$ can be taken in the form $$K_{1/f,N}=-\left( b_{0}v_{0}+\sum_{n=1}^{N}\left(
a_{n}v_{2n}+b_{n}v_{2n-1}\right) \right) \label{K1/f}$$with $b_{0}=-\mathbf{K}_{1/f}(0,0)=h/2$ and the coefficients $a_{n}$, $b_{n}$, $n=1,\ldots ,N$ from (\[K(x,t)\]). Here $v_{k}$ are introduced as follows$$v_{0}=\psi _{0}(x),\quad v_{2n-1}(x,t)=\sum_{\text{even }k=0}^{n}\binom{n}{k}\psi _{n-k}(x)t^{k},\quad v_{2n}(x,t)=\sum_{\text{odd }k=1}^{n}\binom{n}{k}\psi _{n-k}(x)t^{k}. $$The fact that a representation for $K_{1/f,N}$ is a linear combination of the terms $v_{k}$ follows from Theorem \[Th Kapprox\] where instead of $q$ and $f$ one should consider $q_{D}$ and $1/f$ respectively. Then the corresponding generalized wave polynomials $u_{k}$ result to be precisely $v_{k}$. In (\[K1/f\]) we state additionally that taking in the representation the coefficients from (\[K(x,t)\]) we obtain an approximation of the kernel $\mathbf{K}_{1/f}$. In order to obtain this result one needs to consider the kernels $K_{f,N}$ and $
K_{1/f,N} $ as scalar components of a single bicomplex function and take into account that the generalized wave polynomials (see [@KKTT]) $u_{k}$ and $v_{k}$ are nothing but scalar components of hyperbolic pseudoanalytic formal powers (for the corresponding details we refer to [@KT; @Transmut]). Thus, the expression (\[K1/f\]) is in fact a metaharmonic conjugate of (\[K(x,t)\]). We formulate here the result stating that if $\left\Vert \mathbf{K}_{f}-K_{f,N}\right\Vert
<\varepsilon $ then necessarily there is an appropriate estimate for $\left\Vert \mathbf{K}_{1/f}-K_{1/f,N}\right\Vert $ where $K_{1/f,N}$ is defined by (\[K1/f\]) and $\left\Vert \mathbf{\cdot }\right\Vert $ is the maximum norm. We give its proof in the Appendix A.
\[Th Estimate for K1/f\]Let $\max_{\overline{S}}\left\vert \mathbf{K}
_{f}-K_{f,N}\right\vert <\varepsilon $ where $K_{f,N}$ has the form . Then $\max_{\overline{S}}\left\vert \mathbf{K}_{1/f}-K_{1/f,N}
\right\vert <\varepsilon C$ where $K_{1/f,N}$ is defined by and the constant $C$ depends only on $f$ and $b$.
The estimates for the kernels $\mathbf{K}_{f}$ and $\mathbf{K}_{1/f}$ imply corresponding estimates for the kernels $S_{f}$, $C_{f}$ and $S_{1/f}$, $C_{1/f}$.
Theorem \[Th Estimate for K1/f\] together with the equality (\[uprime\]) suggests to approximate the derivatives of the solutions $c^{\prime}(\omega,x;h)$ and $s^{\prime}(\omega,x;\infty)$ by the functions $$\overset{\circ}{c}_{N}(\omega,x) :=-\omega T_{1/f,N}\left[\sin\omega
x\right]+\frac{f^{\prime}}{f}T_{f,N}\left[\cos\omega x\right] =-\omega^{2}s_{1/f,N}(\omega,x)+\frac{f^{\prime}}{f}c_{f,N}(\omega,x)$$ and $$\overset{\circ}{s}_{N}(\omega,x) :=T_{1/f,N}\left[\cos\omega x\right]+\frac
{f^{\prime}}{f}T_{f,N}\left[\frac{\sin\omega x}{\omega}\right] =c_{1/f,N}(\omega,x)+\frac{f^{\prime}}{f}s_{f,N}(\omega,x),$$ respectively. Notice that $c_{1/f,N}(\omega,x)$ is an approximation of $c_{1/f}(\omega,x;-h)$.
Let us emphasize that $\overset{\circ}{c}_{N}(\omega,x)$ and $\overset{\circ
}{s}_{N}(\omega,x)$ do not coincide in general with the derivatives of $\overset{}{c}_{N}(\omega,x)$ and $\overset{}{s}_{N}(\omega,x)$.
Approximation of the transmutation kernels corresponding to $f$ and $1/f$ imply approximations for the solutions $c(\omega,x;h)$, $s(\omega,x;\infty)$, $c_{1/f}(\omega,x;-h)$ and $s_{1/f}(\omega,x;\infty)$. From the corresponding estimates it is easy to obtain estimates for the approximations of $c^{\prime
}(\omega,x;h)$ and $s^{\prime}(\omega,x;\infty)$ by $\overset{\circ}{c}_{N}(\omega,x)$ and $\overset{\circ}{s}_{N}(\omega,x)$.
We note that due to (\[K1/f\]), $\overset{\circ}{c}_{N}(\omega,x)$ and $\overset{\circ}{s}_{N}(\omega,x)$ can be written in the following form $$\begin{aligned}
\overset{\circ}{c}_{N}(\omega,x) & =-\omega\sin\omega x+2\omega\sum_{n=1}^{N}a_{n}\sum_{\text{odd }k=1}^{n}\binom{n}{k}\psi_{n-k}(x)\int_{0}^{x}t^{k}\sin\omega t\,dt\\
& \qquad +\frac{f^{\prime}}{f}\left( \cos\omega x+2\sum_{n=0}^{N}a_{n}\sum_{\text{even }k=0}^{n}\binom{n}{k}\varphi_{n-k}(x)\int_{0}^{x}t^{k}\cos\omega t\,dt\right)\end{aligned}$$ and $$\begin{aligned}
\overset{\circ}{s}_{N}(\omega,x) & =\cos\omega x-2\sum_{n=0}^{N}b_{n}\sum_{\text{even }k=0}^{n}\binom{n}{k}\psi_{n-k}(x)\int_{0}^{x}t^{k}\cos\omega
t\,dt\\
& \qquad +\frac{f^{\prime}}{\omega f}\left( \sin\omega x+2\sum_{n=1}^{N}b_{n}\sum_{\text{odd }k=1}^{n}\binom{n}{k}\varphi_{n-k}(x)\int_{0}^{x}t^{k}\sin\omega t\,dt\right)\end{aligned}$$ where $b_{0}=h/2$.
Numerical results {#Section7}
=================
General scheme and implementation details {#SubsectAlgorithm}
-----------------------------------------
Consider a Sturm-Liouville equation $$\label{SLMain}
-y''+q(x)y=\lambda y$$ on a segment $[0,b]$ and a corresponding initial value problem $$\label{SLIC}
y(0)=y_0\qquad\text{and}\qquad y'(0)=y_1$$ or a spectral problem $$\begin{aligned}
\alpha_0 y(0)+\beta_0 y'(0) &= 0,\label{SLBC0}\\
\alpha_b y(b)+\beta_b y'(b) &= 0,\label{SLBCb}\end{aligned}$$ where we allow for the coefficients $\alpha_0$, $\beta_0$, $\alpha_b$ and $\beta_b$ to be not only constants but also entire functions of the square root $\omega$ of the spectral parameter $\lambda$ satisfying $|\alpha_0|+|\beta_0|\ne 0$ and $|\alpha_b|+|\beta_b|\ne 0$ (for every $\lambda$).
Based on the results of the previous sections we can formulate the following algorithm for solving initial value and spectral problems and – for equation .
1. Find a non-vanishing on $[0,b]$ solution $f$ of the equation $$-f''+q(x)f=0.$$ Let $f$ be normalized as $f(0)=1$ and define $h:=f'(0)$.
2. Compute the functions $\varphi_k$ and $\psi_k$, $k=0,\ldots,N$ using and .
3. Compute the functions $\mathbf{c}_k$ and $\mathbf{s}_k$, $k=0,\ldots,N$ using and .
4. Find coefficients $a_0,a_1,\ldots,a_N$ and $b_1,\ldots,b_N$ of an approximation of the functions $\frac h2+\frac 14 \int_0^x q(s)\,ds$ and $\frac 14 \int_0^x q(s)\,ds$ by linear combinations $\sum_{n=0}^N a_n\mathbf{c}_n(x)$ and $\sum_{n=1}^N b_n\mathbf{s}_n(x)$ as in Theorem \[Th Kapprox\]. Set $b_0=a_0$. Note that also one can take $a_0=\frac h2$ and approximate the function $\frac h2(1-f(x))+\frac 14 \int_0^x q(s)\,ds$ by a linear combination $\sum_{n=1}^N a_n\mathbf{c}_n(x)$ in order to find coefficients $a_1,\ldots,a_N$.
5. Calculate the approximations $s_N(\omega,x)$ and $c_N(\omega,x)$ of the solutions $s(\omega,x;\infty)$ and $c(\omega,x;h)$ by and . If necessary, calculate the approximations of the derivatives of the solutions using and . Recall that the expressions $T_{1/f} \cos \omega t$ and $T_{1/f} \frac{\sin \omega t}\omega$ can be computed similarly to and using the coefficients $\tilde a_n:=-b_n$ and $\tilde b_n:=-a_n$ and the functions $\psi_n$ instead of the functions $\varphi_n$, c.f., and .
6. According to and the approximation of the solution of the initial problem has the form $$y=y_0 c_N(\omega,x)+(y_1-y_0h)s_N(\omega, x).$$ The eigenvalues of the problem – coincide with the squares of the zeros of the entire function $$\label{SLCharEq}
\Phi(\omega) := \alpha_b\bigl(\beta_0 c(\omega, b;h)-(\alpha_0+\beta_0 h)s(\omega,b;\infty)\bigr)+\beta_b\bigl(\beta_0 c'(\omega,b;h)-(\alpha_0+\beta_0 h)s'(\omega,b;\infty)\bigr)$$ and are approximated by squares of zeros of the function $$\label{SLCharEqApprox}
\Phi_N(\omega) := \alpha_b\bigl(\beta_0 c_N(\omega, b)-(\alpha_0+\beta_0 h)s_N(\omega,b)\bigr)+\beta_b\bigl(\beta_0 c_N'(\omega,b)-(\alpha_0+\beta_0 h)s_N'(\omega,b)\bigr).$$ Note that despite the division by $\omega$ in and the singularity at zero of the function $\Phi_N(\omega)$ is removable and $\Phi_N(\omega)$ can be considered as an entire function.
7. The eigenfunction $y_\lambda$ corresponding to the eigenvalue $\lambda=\omega^2$ can be taken in the form $$\label{SLEigenfunction}
y_\lambda=\beta_0 c(\omega,x;h)-(\alpha_0+\beta_0 h)s(\omega,x;\infty).$$ Hence once the eigenvalues are calculated the computation of the corresponding eigenfunctions can be done using formulas and .
The results of the previous section allow us to prove the uniform error bound for all approximate zeros of the characteristic function (at least when the coefficients in the boundary conditions and are independent of the spectral parameter) obtained by the proposed algorithm and that neither spurious zeros appear nor zeros missed whenever inequalities and are satisfied with sufficiently small $\varepsilon_1$ and $\varepsilon_2$. For not going into too much detail in the present paper we consider only the case of Dirichlet boundary conditions, i.e., when the characteristic equation reduces to $s(\omega, b;\infty)=0$. We also refer the reader to [@HrynivMykytyuk2009] where similar questions are discussed.
\[Prop Uniform Errors\] Suppose that the boundary conditions and are the Dirichlet boundary conditions, i.e., $\alpha_0=\alpha_b\equiv 1$, $\beta_0=\beta_b\equiv 0$. Then for every $\varepsilon>0$ there exist such $\varepsilon_{1,2}>0$ that if inequalities and are satisfied for some $N$ with these $\varepsilon_{1}$ and $\varepsilon_{2}$ respectively then all the zeros (including multiplicities) of the characteristic function of the problem , , are approximated by the (complex) zeros of the function $\Phi_N(\omega)$ with errors uniformly bounded by $\varepsilon$ and no spurious zeros appear.
Even in the case when the problem , , possesses a purely real spectrum we need to consider complex zeros of the function $\Phi_N(\omega)$ for Proposition \[Prop Uniform Errors\] to hold. For zeros of multiplicity greater than one the proposition establishes that in an $\varepsilon$-neighbourhood of such zero there is a corresponding number of zeros of the approximate characteristic function $\Phi_N$.
For the Dirichlet boundary conditions the characteristic function has the form $\Phi(\omega)=s(\omega,b;\infty)$. Consider the function $\widetilde \Phi(\omega):=\omega \Phi(\omega)$. It is known that $\widetilde \Phi(\omega)$ is an entire function (see, e.g., [@Marchenko §1.3]), has a countable set of zeros, all of finite multiplicity. Denote this set of zeros by $\Omega$.
Let $0<\varepsilon<\frac 1{2b}$ is given. Define a number $$\widetilde\varepsilon=\min\left\{\varepsilon,\inf_{\omega_1,\omega_2\in\Omega,\ \omega_1\ne \omega_2}\frac{|\omega_1-\omega_2|}2\right\}.$$ Since $\Omega$ has no finite accumulation point, $\widetilde\varepsilon>0$. Note that disks of radiuses $\widetilde\varepsilon$ and centers in different zeros of the function $\widetilde\Phi$ do not have common interior points.
Now we show that $$\label{Eq Inf Char Eq Val}
m:=\inf\left\{\widetilde\Phi(z): z\in\mathbb{C},\ |z-\omega|=\widetilde\varepsilon,\ \omega\in\Omega\right\}>0.$$ Recall that $$\label{Eq Char Fun}
\widetilde\Phi(\omega) = \sin\omega b+\int_0^b S_f(b,t)\sin\omega t\,dt$$ and that the zeros (excluding 0) of $\widetilde \Phi(\omega)$ after reordering satisfy the following asymptotics [@Marchenko Lemma 1.3.3] $$\label{Eq Eigenvalue Asympt}
\omega_n=\frac{\pi}b n+\frac{\alpha_n}n,\qquad \text{where } \sup |\alpha_n|<\infty.$$ Hence for large values of $n$ the circles $\{\omega:|\omega-\omega_n|=\widetilde\varepsilon,\ \omega_n\in\Omega\}$ belong to the rings $R_n:=\{\omega: \frac{\widetilde\varepsilon}2\le|\omega-\frac{\pi}b n|\le\frac{3\widetilde\varepsilon}2\}$. Since the function $|\sin\omega b|$ is periodic with the period $\frac{\pi}b$ and does not vanish on $\{\omega: \frac{\widetilde\varepsilon}2\le |\omega|\le \frac{3\widetilde\varepsilon}2\}$, there exists $$\label{Eq Min Sin}
m_1:=\min\left\{ |\sin\omega b|: \frac{\widetilde\varepsilon}2<|\omega|<\frac{3\widetilde\varepsilon}2\right\} >0.$$ Due to [@Marchenko Lemma 1.3.1] if $|\omega|\to\infty$ with $|\operatorname{Im}\omega|$ remaining bounded then $\int_0^b S_f(b,t)\sin\omega t\,dt\to 0$. Hence we obtain from , and that, e.g., $|\widetilde \Phi(\omega)|\ge \frac{m_1}{2}$ when $|\omega-\omega_n|=\widetilde\varepsilon$ for all sufficiently large $|n|$. For all remaining values of $n$ the function $\widetilde\Phi(\omega)$ does not vanish on the circles $|\omega-\omega_n|=\widetilde\varepsilon$, which finishes the proof of the positivity of the constant $m$ in .
Due to the asymptotics all zeros $\omega_n$ belong to a strip $|\operatorname{Im}\omega|\le M$. Let $\varepsilon_1$ be such that $$\varepsilon_1\frac{\sinh ( (M+\widetilde\varepsilon)b)}{M+\widetilde\varepsilon}\le m$$ and for some $N$ the inequalities and are satisfied with this $\varepsilon_1$. Consider the approximate solution $s_N$. Then it follows from that on all circles $|\omega-\omega_n|=\widetilde\varepsilon$, $\omega_n\in\Omega$ we have $$|\widetilde \Phi(\omega)-\omega \Phi_N(\omega)|=|\omega s(\omega,b;\infty)-\omega s_N(\omega,b)|<m.$$ Hence by the Rouche theorem the functions $\widetilde \Phi(\omega)$ and $\omega \Phi_N(\omega)$ have the same number of zeros in the disks $|\omega-\omega_n|<\widetilde\varepsilon$, $\omega_n\in\Omega$.
The statement that no spurious zeros appear follows from the results of [@Marchenko §1.3] where it is shown that the functions $s(\omega,b;\infty)$ and $\sin \omega b$ possess the same number of zeros in $\{\omega: |\operatorname{Re} \omega|<2n+1/2\}$ for sufficiently large $n$, and it can be seen that this statement holds for the function $s_N$ as well.
An analogues statement can be proved for all other boundary conditions of the form , , at least whenever they are spectral parameter independent. The scheme of the proof remains the same and should involve corresponding asymptotic relations similar to which can also be found in [@Marchenko §1.3].
Some remarks should be made regarding the implementation of the described algorithm.
The non-vanishing solution of equation can be constructed using the SPPS representation, see, e.g., [@KrPorter2010] for details. In the case when an exact particular solution is known we compared the obtained approximated solution against the exact one.
The accuracy and speed of the calculation of the recursive integrals play a crucial role for the accuracy and speed of the proposed algorithm. Previously we applied two different approaches for the integration. One is based on a modification of the 6 point Newton-Cottes formula [@CKT2013], second uses spline approximation and integration of the obtained splines [@KKB]. For the first method we can easily use several millions subdivision points, meanwhile the computation time required to construct the approximating splines limits the maximal number of subdivision points for the second methods to tens of thousands. Computation based on the first approach can be highlighted as an especially recommendable option. In all numerical tests reported recently (see, e.g., [@CKT2013]) it delivered fast and accurate results. However for the present work for all but one example we opted for another approach. The main reason is that both methods possess the saturation property, i.e., their accuracy depends polynomially on the used step size and are not suited well enough for really high-precision calculation. For example, for the 6 point Newton-Cottes formula the final accuracy is of the order $O(h^7)$, where $h$ is the step size.
Further choice between available methods is limited by the requirement that the integrals should be calculated recursively, which leads to the following simple condition. Either the integration method should take the values of the function $g$ defined in some predefined abscissas $x_0<x_1<\ldots<x_M$ and return the values of the indefinite integral in the same set of abscissas, or the integration method should determine the set of abscissas $x_0<x_1<\ldots<x_{M'}$ analyzing the given function, and after that provide the value of the indefinite integral in an arbitrary point of interest using only the values $g(x_0),\ldots,g(x_{M'})$.
From several known methods of evaluation of indefinite integrals with high accuracy and suitable for computing the recursive integrals, e.g., Clenshaw–Curtis, Sinc and double exponential methods [@DavisRabinovich Section 2.13.1], [@Haber], [@MuhammadMori], [@Stenger], [@TakahasiMori], [@TSM], [@Wright], we chose the Clenshaw–Curtis quadrature scheme based on the approximation of the integrand by a partial sum of its expansion into a series in terms of Tchebyshev polynomials and termwise integration of the approximation. The method is described in detail in [@DavisRabinovich Section 2.13.1] and has the advantage that restricting all calculation to the Tchebyshev nodes $\frac b2\left(1+\cos\frac{k\pi}M\right)$, $k=0,\ldots, M$ it reduces to the discrete cosine transform (DCT), a simple transformation of the obtained coefficients and the inverse DCT, and hence has a near-linear complexity with respect to the number $M+1$ of used abscissas. Another advantage is that the method works for an arbitrary differentiable function, analyticity is not required. It should be mentioned that the smallest computation time is achieved when $M=2^m$ and that in some cases to obtain a good accuracy of the calculated recursive integrals we used extra digits for intermediate calculations, see the following examples for the details. Another possibility is to split the interval of integration into several subintervals.
To find the coefficients $a_1,\ldots,a_n$ and $b_1,\ldots,b_n$ of the approximations from Theorem \[Th Kapprox\] we applied the least squares method. There exist other methods providing more accurate uniform approximations, however as a rule they are slower, and in our implementation of the described algorithm in Mathematica software even the build-in function `LeastSquares` required a computation time comparable to the time of the calculation of all recursive integrals.
For the calculation of the integrals and we used the recurrent formulas [@AbramovitzStegun 4.3.119 and 4.3.123].
We do not discuss in detail the step of finding the eigenvalues. The main purpose of the numerical examples is to illustrate that the final approximation of a characteristic equation contains all the information required to evaluate accurate approximations of the eigenvalues. The problem reduces to the search of zeros of some analytic function with the only possible pole at $\omega=0$. The derivative of this function is easily obtainable and in the most complicated cases, say clusters of closely located eigenvalues as, e.g., in the Coffey-Evans problem (Example \[ExCE\]), well-known theorems of complex analysis like the argument principle are useful, see, e.g., [@YingKatz], [@DelnitzEtAl]. It is possible that the calculation of the closest to zero eigenvalues by the proposed method may present difficulties due to the pole of $\Phi_N$ at $\omega=0$. One possible solution is to perform a spectral shift, that is, to consider equation in the form $-y''+(q(x)+\lambda_\ast)y=(\lambda+\lambda_\ast)y$, where $|\lambda_\ast|$ is sufficiently separated from zero. Another solution is to use the SPPS representation [@KrPorter2010]. The approximation of the characteristic equation given by the SPPS representation works especially well near the origin, and all required functions are calculated on the step 3 of the described algorithm.
By the described algorithm the functions $\varphi_n$ are calculated only in $M+1$ points coinciding with the Tchebyshev nodes allowing us to compute the eigenfunctions $u_\lambda$ directly by formulas , and only in these $M+1$ points. If for some applications such subset of points is insufficient, the functions $\varphi_n$ can be easily interpolated to arbitrary subset of the segment $[0,b]$. One of the best ways to perform the interpolation is by using the partial sums of approximations of the functions $\varphi_n$ by their expansions into series in terms of Tchebyshev polynomials. The expansion coefficients can be obtained using the DCT, and final interpolations are obtained by summing up corresponding partial sums. Since all the functions $\varphi_n$ are computed applying similar approximation procedure, described interpolation does not deteriorate significantly the accuracy. We illustrate such approach in Example \[ExPaine1\] where we show that even large index highly oscillating eigenfunctions can be accurately approximated.
Sturm-Liouville spectral problems
---------------------------------
\[ExPaine1\] Consider the following spectral problem (the first Paine problem, [@Paine]) $$\begin{cases}
-u''+e^x u=\lambda u, & 0\le x\le \pi,\\
u(0,\lambda)=0, & u(\pi,\lambda)=0.
\end{cases}$$ With the help of Mathematica software we found a non-vanishing particular solution $$\label{Ex1PS}
u_0(x)=I_0\big(2e^{x/2}\big)$$ and the characteristic function $$\Phi(\omega)=I_{2 i \omega }(2)I_{-2 i \omega }\big(2
\sqrt{e^\pi}\big)-I_{-2 i \omega }(2) I_{2 i
\omega }\big(2 \sqrt{e^\pi}\big),$$ where $\omega^2=\lambda$ and $I$ is the modified Bessel function of the first kind.
In this example we performed calculations in Matlab in machine precision and in Mathematica using high precision arithmetic. For the Matlab program we used $N=30$ for the approximations and and computed all involved recursive integrals using Newton-Cottes integration formula with $M=20000$. This experiment is similar to [@KT; @Transmut Example 9], however the runtime improved due to the different integration method used. The approximation errors achieved in and were $5.5\cdot 10^{-11}$ and $9.3\cdot 10^{-11}$ respectively. 500 eigenvalues were calculated and the maximal absolute error of the approximated eigenvalues was $1.95\cdot 10^{-9}$. The overall time required for the calculation was 5 seconds for the approximation part (Steps 1–5 of the algorithm from Subsection \[SubsectAlgorithm\]) and 6.5 seconds was required to finish Step 6, such time was necessary because we constructed a spline approximating the characteristic function and used Matlab function `fnzeros` to find its zeros. A personal computer equipped with Intel i7-3770 processor was used for this and following computations.
For the second experiment we performed all the numerical calculations with 200-digit arithmetic in Mathematica. We used $M=256$ for the calculation of all involved recursive integrals. The particular solution was computed using the SPPS representation with $150$ formal powers and compared with the solution to verify the precision of the approximate solution. The maximal difference between the approximate and the exact solutions was $3.7\cdot 10^{-187}$ showing an excellent accuracy achievable by the combination of the SPPS representation and the Clenshaw–Curtis integration procedure.
Using the obtained particular solution we calculated the functions $\mathbf{c}_n$ and $\mathbf{s}_n$ for $n\le 150$. After that for each $N=8,10,\ldots,150$ we found coefficients $a_0,\ldots,a_N$, $b_1,\ldots,b_N$ for and , calculated first 500 eigenvalues as zeros of the approximate characteristic function and compared them to the exact ones. The function `FindRoot` from Mathematica was used to find both exact and approximate eigenvalues. On Figure \[ExPaineApproxError\] we present both the obtained approximation errors in the inequalities and and the maximal resulted relative error of the first 500 eigenvalues. We would like to point out that the error decays exponentially with respect to the number $N$ of functions used and that the resulted error of the eigenvalues is bounded by the error of the approximations on the characteristics.
Note that on the final step from $N=148$ to $N=150$ the approximation error increases, which can be explained by the fact that we passed a limit where the used precision and the number of points work well. Hence we used the value $N=148$ to verify the accuracy of the first 500 eigenvalues and 500 corresponding eigenfunctions. We calculated the eigenfunctions $u_\lambda$ satisfying the initial condition $u'_\lambda(0)=\sqrt{\lambda}$ on the uniform mesh of 2000 points from $[0,\pi]$ using the interpolated values of the functions $\varphi_n$ and compared them to the exact ones. The approximation errors in and were $3.9\cdot 10^{-106}$ and $6.1\cdot 10^{-106}$ respectively, while the largest error of the computed eigenvalues was $4.0\cdot 10^{-104}$ for the eigenvalue number 49. The computation time was 56 seconds for the approximation part (Steps 1–5 of the algorithm from Subsection \[SubsectAlgorithm\]) and 213 seconds was required for build-in Mathematica function `FindRoot` to finish Step 6. On Figure \[ExPaineEigError\] we show the absolute errors of the first 500 eigenvalues together with the distances between exact and approximate eigenfunctions in the uniform norm. Observe that even though the eigenvalues grow, the absolute errors remain essentially of the same order. The computed eigendata accuracy does not deteriorate even for eigenvalues with larger indices. For example, absolute errors of $\lambda_{1000}$, $\lambda_{2500}$ and $\lambda_{10000}$ are $1.3\cdot 10^{-105}$, $2.1\cdot 10^{-105}$ and $2.9\cdot 10^{-106}$, and errors of corresponding eigenfunctions (evaluated on a mesh of 25000 points) are $4.7\cdot 10^{-108}$, $1.3\cdot 10^{-108}$ and $1.8\cdot 10^{-109}$, respectively.
\[ExPaine2\] Consider the following spectral problem (the second Paine problem, [@Paine; @Pryce]) $$\begin{cases}
-u''+\frac{1}{(x+0.1)^2} u=\lambda u, & 0\le x\le \pi,\\
u(0,\lambda)=0, & u(\pi,\lambda)=0.
\end{cases}$$ With the help of Mathematica software we found a non-vanishing particular solution $$\label{Ex2PS}
u_0(x)=(1+10x)^{(1+\sqrt{5})/2}$$ and the characteristic function $$\Phi(\omega)=M_{0,-\frac{\sqrt{5}}{2}}\left(\frac{\omega}{5}\right)
W_{0,-\frac{\sqrt{5}}{2}}\left(\frac{\omega}{5}(1+10 \pi )\right)-M_{0,-\frac{\sqrt{5}}{2}}\left(\frac{\omega}{5}
(1+10 \pi )\right)
W_{0,-\frac{\sqrt{5}}{2}}\left(\frac{\omega}{5}\right),$$ where $\omega^2=\lambda$, $M$ and $W$ are the Whittaker functions [@AbramovitzStegun].
We calculated an approximate particular solution in Mathematica using the SPPS representation with 300 formal powers and performed integrations with $M=1024$ and 200-digits arithmetic. The larger number of points compared to Example \[ExPaine1\] was necessary for an accurate evaluation of the recursive integrals and possibly can be explained by the fact that the accuracy of the Clenshaw–Curtis integration depends on the decay rate of coefficients of the function expansion into a series in terms of Tchebyshev polynomials, which in turn is related to the size of an ellipse on a complex plane with foci at the points $z_1=-1$ and $z_2=1$ to which the potential $q$ possesses the analytic continuation, see, e.g., [@RiessJohnson], [@Trefethen]. The maximum error of the approximate solution compared with the exact one was less than $4\cdot 10^{-164}$.
Using the obtained particular solution we calculated the functions $\mathbf{c}_n$ and $\mathbf{s}_n$ for $n\le 320$. For the calculation we used 400-digit arithmetic. Such precision appeared to be a necessity for the build-in Mathematica function `LeastSquares` to be able to produce approximation coefficients for and for large values of $N$. After that for values of $N$ in the range $[8,320]$ taken with step sizes increasing from $4$ to $16$ we found coefficients $a_0,\ldots,a_N$, $b_1,\ldots,b_N$ for and . It turned out that finding the first several eigenvalues as zeros of the approximated characteristic function $\Phi_N$ for large $N$ is not possible directly with Mathematica’s function `FindRoot` (see the explanation at the end of subsection \[SubsectAlgorithm\]), so we applied the following procedure. First we calculate roots of the polynomial obtained as a truncation of the SPPS representation of the characteristic equation, see, e.g., [@KrPorter2010]. These roots are known to give an excellent accuracy especially for the eigenvalues close to the origin. Then we find zeros of $\Phi_N$. To combine the two obtained sets we find the two closest values in these two sets and take the smaller ones from the roots of the SPPS polynomial and the larger ones from the zeros of $\Phi_N$. Such strategy worked well for all values of $N$. On Figure \[ExPaine2ApproxError\] we present both the obtained approximation errors in the inequalities and and the maximal resulted absolute and relative errors of the first 500 eigenvalues as functions of $N$. We would like to point out that the error decay exponentially with respect to the number $N$ of functions used and that the slopes of the graphs are close.
Note that the slope of the approximation error graph changes around $N=270$, and the absolute and relative errors decrease slower starting from this value of $N$. Again we can explain such behavior by the fact that we are close to the limit where the used precision and the number of points still work. Hence we used the value $N=280$ to present the graph of the errors of the first 500 eigenvalues. The approximation errors in and were $2.5\cdot 10^{-43}$ and $2.3\cdot 10^{-43}$ respectively, while the largest error of the computed eigenvalues was $2.3\cdot 10^{-42}$ for the eigenvalue number 237. On Figure \[ExPaine2EigError\] we show the absolute errors of the first 500 eigenvalues. Again we see that the absolute errors of the eigenvalues remain at the same level. Moreover, the accuracy of the computed eigenvalues does not deteriorate for even larger eigenvalues. For example, absolute errors of $\lambda_{1000}$, $\lambda_{2500}$ and $\lambda_{10000}$ are $1.4\cdot 10^{-42}$, $1.1\cdot 10^{-42}$ and $2.2\cdot 10^{-43}$ respectively.
While high precision arithmetic is necessary in this example to produce accurate eigenvalues, it is possible to use smaller parameters $N$ and $M$ and lower precision arithmetic if one looks for the eigenvalues accurate to 13–15 digits (i.e., with the precision expected from double-precision machine arithmetics), leading to faster runtime, comparable with other codes available. For example, we used $M=256$, $N=120$ and 128-digit arithmetic and the algorithm finished Steps 1–5 in 33 seconds and in 103 seconds found 500 eigenvalues with the largest absolute error of $2\cdot 10^{-13}$.
\[ExCE\] Consider the Coffey-Evans problem [@ChildCmabers] $$\label{CEeqn}\begin{cases}
-u''+\bigl(\beta^2\sin^2 2x-2\beta \cos 2x\bigr) u=\lambda u, & -\frac\pi 2\le x\le \frac\pi 2,\\
u\bigl(-\frac\pi 2,\lambda\bigr)=u\bigl(\frac\pi 2,\lambda\bigr)=0.
\end{cases}$$ This problem is considered as a standard test case for numerical methods for solving Sturm-Liouville spectral problems, see, e.g., [@Pryce], [@PruceFulton], [@AliciTaceli], [@KrPorter2010], [@Ledoux2010], and presents the challenge of distinguishing eigenvalues within the triple clusters which form as the parameter $\beta$ increases. The equation in is a particular case of the Whittaker-Hill equation and its particular solution with the initial conditions $u(-\frac\pi 2)=1$, $u'(-\frac\pi 2)=0$ is known [@HemeryVeselov] and is given by $$\label{ExCEps}
u_0(x)=e^{\beta\cos 2x}.$$ To our best knowledge the most accurate eigenvalues of are reported in [@AliciTaceli], where the table of the first 18 eigenvalues for the case $\beta=50$ correct to 24 decimal places is included. It is worth mentioning that the method used in [@AliciTaceli] is suitable only for a special subclass of Sturm-Liouville equations. For the numerical example we also chose $\beta=50$.
An approximation of the particular solution was calculated using the SPPS representation with 1200 formal powers. We used 800-digit arithmetic and $M=2048$ for the calculation of the formal powers. The resulted error of the approximate solution compared with was less than $2.7\cdot 10^{-643}$. We would like to point out such remarkable accuracy. The built-in Mathematica’s function `NDSolve` was not able to achieve any acceptable precision calculating the approximate solution. As it can be seen from the solution $u_0$ is symmetric and satisfies $u_0(\frac\pi 2)=1$, meanwhile the best result we were able to obtain using `NDSolve` function was $u_0(\frac\pi 2)\approx 62$.
To apply the described algorithm we transformed the problem to the interval $[0,1]$. We found that the Clenshaw–Curtis integration requires a lot of extra precision to evaluate the iterative integrals. The following simple test was used. It follows from the definition of the transmutation operator and Theorem \[Th Transmute\] that $\frac{\varphi_k(x)}{x^k}\to 1$, $k\to\infty$. Hence, any weird behavior of the quantity $\frac{\varphi_k(x)}{x^k}$ when $k$ increases indicates serious errors in the calculated formal powers. On Figure \[ExCEphiError\] we present the graphs of $\varphi_k(1)$ evaluated with different values of $M$ and different precision. We tried formulas , as well as recently discovered formulas from [@KrTNewSPPS] for the calculation of the formal powers. As one can see from the presented graphs, for the same value of the parameter $M$ and the same arithmetic precision the formulas from [@KrTNewSPPS] allow one to roughly double the number of calculated formal powers. Moreover, a further increase of the number of calculated formal powers can be achieved by increasing the arithmetic precision.
For the first experiment we calculated the functions $\mathbf{c}_n$ and $\mathbf{s}_n$ for $n\le 256$ using $M=8192$ and 2500-digit arithmetic for the intermediate calculations, final values of the functions were stored at 1025 points with 400 digit precision. Our results were compared with those from [@AliciTaceli]. We calculated the error of the eigenvalues separately for the isolated eigenvalues and for each of the first four clusters. The obtained errors are presented on Figure \[ExCEEigError\]. Note that the errors of the isolated eigenvalues decrease much faster than the errors of approximation in and , meanwhile the slope of graph of the error for the clustered eigenvalues is an agreement with the slope of the error of approximation. Note that the lower points of the error graphs correspond to the limit of precision of the values presented in [@AliciTaceli].
For the second experiment we calculated the functions $\varphi_n$, $\mathbf{c}_n$ and $\mathbf{s}_n$ for $n\le 400$ using $M=16384$ and 4000-digit arithmetic for the intermediate calculations. It is known [@KrTNewSPPS] that for the first eigenvalues the SPPS method achieves an especially remarkable accuracy. Hence we computed the eigenvalues of the first cluster using the SPPS representation with all functions $\varphi_n$, $n\le 400$ and used the obtained values to verify the accuracy of the proposed algorithm. We observed an exponential decay of approximation errors in and as well as of the error of the computed eigenvalues with respect to $N$. The slopes of both lines are nearly equal, see Figure \[ExCEEigError2\]. For $N=380$ the errors of approximation were $2.12\cdot 10^{-63}$ and $2.16\cdot 10^{-63}$. In Table \[ExCETable1\] we present the obtained eigenvalues. We are sure that they are exact for all 65 decimal places, evaluation with higher $N$ confirmed the presented digits. The eigenvalue with the index 0 is taken from the roots of the SPPS polynomial. On Figure \[ExCEEigenfunctions\] we illustrate that our method allows one to obtain eigenfunctions as well.
[cr@l]{}$n$ &\
0 & 4.&712 683 501 976 174 806 164 70 $\cdot 10^{-42}$\
1 & 197.&968 726 516 507 291 450 189 104 613 631 137 680 282 238 501 965 516 499 678 457 445 33\
2 & 391.&808 191 489 053 841 050 234 434 838 960 967 152 793 679 673 029 474 438 776 558 053 94\
3 & 391.&808 191 489 053 841 832 241 250 450 567 879 442 934 703 990 750 980 000 552 054 926 43\
4 & 391.&808 191 489 053 842 614 248 066 062 174 820 764 145 024 196 402 821 083 449 116 736 40\
5 & 581.&377 109 231 579 654 864 715 898 934 768 731 234 409 061 366 366 202 824 138 138 986 73\
6 & 766.&516 827 285 532 616 579 817 794 300 693 795 455 315 745 010 536 896 934 624 541 778 33\
7 & 766.&516 827 285 535 505 431 430 237 728 556 528 324 964 223 414 154 644 684 143 065 548 50\
8 & 766.&516 827 285 538 394 283 042 681 500 534 232 365 256 146 570 623 086 585 717 773 157 61\
9 & 947.&047 491 585 860 179 592 142 658 200 615 670 883 560 237 084 089 403 375 253 394 471 82\
10 &1122.&762 920 067 901 205 616 045 550 505 249 660 804 795 577 778 366 617 496 311 372 260 47\
11 &1122.&762 920 071 056 526 891 891 942 465 507 589 782 179 421 839 584 709 544 874 366 683 38\
12 &1122.&762 920 074 211 848 168 115 209 545 412 485 203 977 238 174 036 843 957 138 054 891 76\
13 &1293.&423 567 331 707 081 413 958 872 197 134 275 865 126 916 380 700 329 000 293 361 954 74\
14 &1458.&746 557 025 357 659 317 371 063 260 166 052 216 899 792 667 117 964 891 413 575 933 50\
15 &1458.&746 558 472 128 708 810 534 887 553 090 428 313 351 825 225 479 463 874 997 007 515 49\
16 &1458.&746 559 918 899 832 786 248 167 778 046 441 588 242 318 698 298 075 300 043 185 188 61\
17 &1618.&391 008 042 643 345 932 885 816 053 496 039 220 613 799 984 685 321 316 858 406 202 35\
18 &1771.&934 971 252 995 278 016 339 167 903 087 106 095 945 385 769 959 120 071 823 242 633 83\
19 &1771.&935 290 604 372 265 020 106 948 865 313 215 142 762 994 141 327 857 394 637 946 848 95\
20 &1771.&935 609 959 205 928 887 875 652 226 707 319 017 267 239 862 605 818 421 687 555 219 86\
25 &2189.&490 124 838 400 777 638 432 025 527 998 260 500 900 985 813 188 172 572 445 106 470 28\
50 &3928.&016 942 351 712 838 529 915 885 833 769 553 216 441 127 779 708 418 407 221 179 291 87\
100&11470.&288 862 210 604 335 996 473 673 114 332 968 420 756 176 847 415 047 788 545 642 847 41\
\
\
Complex potential and spectral parameter dependent boundary conditions
----------------------------------------------------------------------
\[ExBoumenir\] First we consider a problem with a complex potential for which the eigenvalues are known explicitly: $$\label{BoumenirEqn}
\begin{cases}
-u''+(3+4i) u=\lambda u, & 0\le x\le \pi,\\
u'(0)=u'(\pi)=0.
\end{cases}$$ A similar problem was treated in [@Boumenir2001], [@Chanane2007]. We pose the Neumann boundary conditions to illustrate the performance of the proposed method in the case when one has to use approximations of both transmutation operators $T_f$ and $T_{1/f}$ and there is no zero coefficient in the expression .
The eigenvalues of the problem are given by $\lambda_n=n^2+3+4i$, $n=0,1,2\ldots$. As a particular solution we took $f(x)=e^{(2+i)x}$, then $h=f'(0)=2+i$, and the solution of the equation in satisfying the left boundary condition can be taken in the form $u(x,\omega):=c(\omega,x;h)-h s(\omega,x;\infty)$, see and . Equation $u'(\pi,\omega)=0$ gives us the characteristic equation of the problem .
For the numerical experiment we computed the functions $\mathbf{c}_n$ and $\mathbf{s}_n$ for $n\le 100$ using $M=256$ and 200-digit arithmetic for the calculation of the iterative integrals. After that we found the coefficients $a_0,a_1,\ldots,a_N$ and $b_1,\ldots,b_N$ for and and approximated the integral kernel $K_{1/f}$ by . On Figure \[ExBoumenirError\] we present the absolute error of the first 500 eigenvalues obtained by the proposed algorithm with different values of $N$. Note that the absolute errors remain at the same level and are in an excellent agreement with the approximation errors.
\[ExChanane\] Consider the following problem with a complex potential and a spectral parameter dependent boundary condition [@Chanane2007] $$\label{ChananeEqn}
\begin{cases}
-u''+e^{2ix} u=\mu^2 u, & 0\le x\le 1,\\
u(0)+\mu u(1)=0, & u'(0)=0.
\end{cases}$$ For this example the characteristic equation is given by the equality $$\pi \mu \left(J_{-\mu }\left(e^i\right) (\mu J_{\mu
}(1)-J_{\mu -1}(1))+(J_{-\mu -1}(1)+\mu J_{-\mu
}(1)) J_{\mu }\left(e^i\right)\right)=2 \sin (\pi
\mu ),$$ where $J$ is the Bessel function of the first kind. As a particular solution we took $f(x)=\frac{Y_0(e^{ix})}{Y_0(1)}$ with $f'(x)=-\frac{ie^{ix}Y_1(e^{ix})}{Y_0(1)}$ and hence $h:=f'(0)=-\frac{iY_1(1)}{Y_0(1)}$, where $Y$ is the Bessel function of the second kind.
We checked that the proposed algorithm was able to produce accurate results even using small values of the parameters $N$ and $M$. We used $N=20$, $M=96$ and 24-digit arithmetic. The approximation errors in and were less than $2\cdot 10^{-18}$. The computation time was 1.6 seconds for the approximation part (Steps 1–5 of the algorithm from Subsection \[SubsectAlgorithm\]) and 15 seconds was required to finish Step 6. In Table \[ExChananeTable\] we present the exact eigenvalues of the problem together with the absolute errors obtained by our method and those reported in [@Chanane2007].
[cr@[+]{}lcc]{}$n$ & &
-------------
Abs. error,
our method
-------------
: The eigenvalues of the problem (Example \[ExChanane\]) and the absolute eigenvalue errors obtained by our method and by method from [@Chanane2007].[]{data-label="ExChananeTable"}
&
----------------------------
Abs. error,
method from [@Chanane2007]
----------------------------
: The eigenvalues of the problem (Example \[ExChanane\]) and the absolute eigenvalue errors obtained by our method and by method from [@Chanane2007].[]{data-label="ExChananeTable"}
\
1 & 4.9685430929323576232 & 0.3906545895360696300$i$ & $7.45\cdot 10^{-19}$ & $5.549\cdot 10^{-15}$\
2 & 20.602710348893372907 & 0.750232523531540313$i$ & $4.15\cdot 10^{-19}$ & $3.393\cdot10^{-14}$\
3 & 64.140382448045471607 & 0.684228375311332294$i$ & $3.84\cdot 10^{-19}$ & $3.977\cdot10^{-13}$\
4 & 119.34792168887388950 & 0.714972404794013828$i$ & $3.88\cdot 10^{-19}$ & $8.004\cdot10^{-13}$\
5 & 202.31443747778733950 & 0.70057212586524954$i$ & $2.43\cdot 10^{-19}$ & $2.064\cdot10^{-13}$\
7 & 419.44558800598640866 & 0.70446189520144488$i$ & $2.39\cdot 10^{-17}$ & $4.582\cdot10^{-12}$\
10 & 889.18520034251622114 & 0.70898948206981412$i$ & $1.05\cdot 10^{-17}$ & $2.734\cdot10^{-11}$\
15 & 2077.5390063282081426 & 0.7073452595732362$i$ & $9.8\cdot 10^{-19}$ & $8.757\cdot10^{-10}$\
20 & 3751.3714273572505215 & 0.7082902747558125$i$ & $6.69\cdot 10^{-18}$ & $2.410\cdot10^{-9}$\
25 & 5926.6847018611521726 & 0.7078182266190013$i$ & $5.9\cdot 10^{-18}$ & $0.0003165$\
50 & 24181.452786752952659 & 0.708107048845456$i$ & $3.6\cdot 10^{-18}$ &\
75 & 54781.226477096285949 & 0.708045812501978$i$ & $4.5\cdot10^{-18}$ &\
100 & 97710.005609281581148 & 0.708081740843471$i$ & $5.0\cdot 10^{-18}$ &\
Quantum wells
-------------
In this subsection we use notations and recall some results from [@CKOR].
Consider the eigenvalue problem for the one dimensional Schrödinger operator $$\label{eqQuantWell}
Hu:=-u''+Q(x)u=\lambda u,\qquad x\in\mathbb{R},$$ where $$Q(x)=\begin{cases}
\alpha_1, & x<0\\
q(x), & 0\le x\le\ell,\\
\alpha_2,& x>\ell
\end{cases}$$ $\alpha_1$ and $\alpha_2$ are complex constants and $q$ is a continuous function on the segment $[0,\ell]$. The spectral problem consists in finding values of the spectral parameter $\lambda\in\mathbb{C}$ for which equation possesses a nontrivial solution $u$ belonging to the Sobolev space $H^2(\mathbb{R})$.
In the selfadjoint case, i.e., when $Q$ is a real-valued function, the operator $H$ has a continuous spectrum $\left[\min\{\alpha_1,\alpha_2\},+\infty\right)$ and a discrete spectrum located on the set $$\label{DiscrSpectrQW}
\Bigl[\min_{x\in[0,\ell]} q(x),\min\{\alpha_1,\alpha_2\}\Bigr).$$
It was shown in [@CKOR] that finding the eigenvalues of the operator $H$ is equivalent to the Sturm-Liouville spectral problem for equation on the segment $[0,\ell]$ with the boundary conditions $$\begin{aligned}
u'(0)-\mu u(0)&=0, \label{bcQWleft}\\
u'(\ell)+\nu u(\ell) &=0, \label{bcQWright}\end{aligned}$$ where $\mu=+\sqrt{\alpha_1-\lambda}$, $\nu=\sqrt{\alpha_2-\lambda}$ and the eigenvalues are sought on the interval .
We consider a particular case $\alpha_1=\alpha_2=0$. Let $f$ be a particular solution of for $\lambda=0$, non-vanishing on $[0,\ell]$ and satisfying $f(0)=1$. Define $h:=f'(0)$. Introducing a new parameter $\lambda=-\beta^2$ we obtain that $\mu=\nu=\beta$ in and , $$\beta\in \Bigl[0,\sqrt{\max_{x\in[0,\ell]} (-q(x))}\Bigr),$$ and the spectral parameter $\omega=i\beta$. It follows from , that the function $u(x)=c(\omega, x;h)-(i\omega+h)s(\omega, x;\infty)$ satisfies , and from we obtain the characteristic equation $$ c'(\omega,\ell;h)-(i\omega+h)s'(\omega,\ell;\infty)-i\omega c(\omega,\ell;h)+i\omega(i\omega+h)s(\omega,\ell;\infty)=0.$$
\[ExSquareWell\] Consider the square-well potential $$Q(x)=\begin{cases}
-U, & |x|\le a\\
0, & \text{elsewhere}.
\end{cases}$$ The eigenvalue $\lambda_n=-\beta_n^2$ is a solution of equation $$\arctan \frac{\sqrt{U-\beta_n^2}}{\beta_n}+a\sqrt{U-\beta_n^2}=\frac{n\pi}2,$$ and the number of eigenvalues for each value of $U$ is equal to the smallest integer greater or equal to $2a\sqrt{U}/\pi$.
We have chosen the numerical values $U=15$, $a=1$, shifted the problem to the segment $[0,2]$ and taken $f(x)=e^{i\sqrt U x}$ as a particular solution. We used $N=32$, $M=96$ and 32-digits arithmetic. The errors in and were of the magnitude $5.5\cdot 10^{-19}$. The exact eigenvalues of the problem and the absolute errors of the approximated eigenvalues are listed in Table \[ExSquareWellTable\].
[cccccc]{}$n$ & $\beta_{n}$ (our method) & Absolute error & $\beta_{n}$ ([@CKOR]) & Absolute error\
1 & 1.54436716376282718435 & $2\cdot 10^{-20}$ & 1.544367170 & $6\cdot 10^{-9}$\
2 & 2.99547074607315853471 & $1.2\cdot 10^{-19}$ & 2.995470748 & $2\cdot 10^{-9}$\
3 & 3.66781322275488144840 & $9\cdot 10^{-20}$ & 3.667813223 & $<3\cdot 10^{-10}$\
\[ExSechSquare\] Consider the sech-squared potential defined by the expression $Q(x)=-m(m+1)\operatorname{sech}^2x$, $x\in(-\infty,\infty)$, $m\in\mathbb{N}$. An attractive feature of the potential $Q$ is that its eigenvalues can be calculated explicitly. The eigenvalue $\lambda_n$ is given by the formula $\lambda_n=-(m-n)^2$, $n=0,1,\ldots,m-1$.
The potential $Q$ is not of a finite support, nevertheless its absolute value decreases rapidly when $x\to\pm\infty$. We approximate the original problem by a problem with a finitely supported potential $\widehat Q$ defined by the equality $$ \widehat Q(x)=\begin{cases}
-m(m+1)\operatorname{sech}^2x, & |x|\le a\\
0, & \text{elsewhere},
\end{cases}$$ where $a$ is chosen in such way that $Q(a)$ is sufficiently small.
For the numerical experiment we took $a=10$. Again in this example the recently discovered formulas from [@KrTNewSPPS] produced much more accurate results than formulas and . Using $N=70$, $M=2096$ and 128-digit arithmetic for the case $m=3$ we obtained the results presented in Table \[ExSechSqWellTable\].
[cccc]{}$n$ & Exact values & $\lambda_{n}$ (our method) & $\lambda_{n}$ ([@CKOR])\
0 & $-9$ & $-8.99999999999999999980$ & $-8.999628656$\
1 & $-4$ & $-4.00000000000000000020$ & $-3.999998053$\
2 & $-1$ & $-0.99999999999999877643$ & $-0.999927816$\
Proof of Theorem \[Th Estimate for K1/f\]
=========================================
The proof is based on several results from [@KT; @Transmut], so we preserve notations from [@KT; @Transmut]. Consider the bicomplex function $W:=\mathbf{K}_{f}-\mathbf{j}\mathbf{K}_{1/f}$. Here $\mathbf{j}$ is the hyperbolic imaginary unit: $\mathbf{j}^{2}=1$. Using (\[K(x,t)\]) and (\[K1/f\]) we observe that $W_{N}:=K_{f,N}-\mathbf{j}K_{1/f,N}=\sum_{n=0}^{N}Z^{(n)}(\alpha _{n},0;z)$ where $\alpha _{n}:=a_{n}+\mathbf{j}b_{n}$ and $Z^{(n)}$ are hyperbolic pseudoanalytic formal powers admitting the representation [@KT; @Transmut]$$\begin{aligned}
Z^{(0)}(\alpha _{0},0;z)& =a_{0}u_{0}(x,t)+\mathbf{j}b_{0}v_{0}(x,t)=\frac{h}{2}\left( f(x)+\frac{\mathbf{j}}{f(x)}\right) , & & \\ Z^{(n)}(\alpha _{n},0;z)& =a_{n}u_{2n-1}(x,t)+b_{n}u_{2n}(x,t)+\mathbf{j}\bigl(a_{n}v_{2n}(x,t)+b_{n}v_{2n-1}(x,t)\bigr), & & n\geq 1.
$$ Here the hyperbolic variable $z$ has the form $z=x+\mathbf{j}t$ with the corresponding conjugate $\overline{z}=Cz:=x-\mathbf{j}t$.
Denote $\widetilde{\mathbf{K}}_{f}=T_{f}^{-1}\left[ \mathbf{K}_{f}\right] $ and $\widetilde{K}_{f,N}=T_{f}^{-1}\left[ K_{f,N}\right] $. For the corresponding Goursat data we introduce the notations$$\begin{pmatrix}
\varphi (x) \\
\psi (x)\end{pmatrix}:=\begin{pmatrix}
\widetilde{\mathbf{K}}_{f}(x,x) \\
\widetilde{\mathbf{K}}_{f}(x,-x)\end{pmatrix}\quad \text{and}\quad
\begin{pmatrix}
\varphi _{N}(x) \\
\psi _{N}(x)\end{pmatrix}:=\begin{pmatrix}
\widetilde{\mathbf{K}}_{f,N}(x,x) \\
\widetilde{\mathbf{K}}_{f,N}(x,-x)\end{pmatrix}.$$Then $\left\vert \varphi -\varphi _{N}\right\vert \leq \varepsilon
\Vert T_{f}^{-1}\Vert $ and $\left\vert \psi -\psi
_{N}\right\vert \leq \varepsilon \Vert T_{f}^{-1}\Vert $. For an estimate of the uniform norm $\Vert T_{f}^{-1}\Vert $ we refer to [@KKTT]. It depends on $f$ and $b$ only.
Let us consider the functions $\Phi (x):=2\varphi (x)$, $\Psi (x)=2\psi
(x)-h $, $\Phi _{N}(x):=2\varphi _{N}(x)$, $\Psi _{N}(x)=2\psi _{N}(x)-h$. We have that $\widetilde{W}:=\widetilde{\mathbf{K}}_{f}-\mathbf{j}\widetilde{\mathbf{K}}_{1/f}$ and $\widetilde{W}_{N}:=\widetilde{K}_{f,N}-\mathbf{j}\widetilde{K}_{1/f,N}$ are the unique solutions of the following corresponding Goursat problem [@KT; @Transmut]$$\partial _{\bar{z}}\widetilde{W}=0,$$$$\widetilde{W}(x,x)=P^{+}\Phi \left( x\right) +P^{-}\Psi \left( 0\right)
\quad \text{and\quad }\widetilde{W}(x,-x)=P^{+}\Phi \left( 0\right)
+P^{-}\Psi \left( x\right)$$and $$\partial _{\bar{z}}\widetilde{W}_{N}=0,$$$$\widetilde{W}_{N}(x,x)=P^{+}\Phi _{N}\left( x\right) +P^{-}\Psi _{N}\left(
0\right) \quad \text{and\quad }\widetilde{W}_{N}(x,-x)=P^{+}\Phi _{N}\left(
0\right) +P^{-}\Psi _{N}\left( x\right)$$where the idempotents $P^{+}$ and $P^{-}$ are defined by $P^{\pm }=\frac{1}{2}(1\pm \mathbf{j})$. Moreover, $\widetilde{W}=\mathbf{V}_{1}^{-1}[
\mathbf{K}_{f}-\mathbf{jK}_{1/f}] $ where the operators $\mathbf{V}_{1} $ and $\mathbf{V}_{1}^{-1}$ were introduced in [@KT; @Transmut] as $\mathbf{V}_{1}^{\pm 1}=T_{f}^{\pm 1}\mathcal{R}+\mathbf{j}T_{1/f}^{\pm 1}\mathcal{I}$ with $\mathcal{R}$ and $\mathcal{I}$ being projection operators projecting a bicomplex valued function onto the respective scalar components, $\mathcal{R=}\frac{1}{2}(I+C)$ and $\mathcal{I}=\frac{1}{2\mathbf{j}}(I-C)$ where $I$ is the identity operator. Thus, $\widetilde{W}=T_{f}^{-1}\left[ \mathbf{K}_{f}\right] -\mathbf{jT}_{1/f}^{-1}\left[
\mathbf{K}_{1/f}\right] $ meanwhile for $\widetilde{W}_{N}$ we have $\widetilde{W}_{N}=\mathbf{V}_{1}^{-1}\left[ \sum_{n=0}^{N}Z^{(n)}(\alpha
_{n},0;z)\right] =\sum_{n=0}^{N}\alpha _{n}z^{n}$ (see [@KT; @Transmut Sect. 5]). Now we consider $$\Vert W-W_{N}\Vert =\Vert \mathbf{V}_{1}\widetilde{W}-\mathbf{V}_{1}\widetilde{W}_{N}\Vert \leq \Vert \mathbf{V}_{1}\Vert \Vert \widetilde{W}-\widetilde{W}_{N}\Vert
\label{estimate1}$$ where $\Vert \widetilde{W}-\widetilde{W}_{N}\Vert =\max_{\overline{S}}\vert \widetilde{W}-\widetilde{W}_{N}\vert _{\mathbb{B}}$ and the norm $\left\vert w\right\vert _{\mathbb{B}}$ of a bicomplex number $w$ is defined as in [@KT; @Transmut Sect. 5]: $\left\vert
w\right\vert _{\mathbb{B}}=\frac{1}{2}\left( \left\vert w^{+}\right\vert
+\left\vert w^{-}\right\vert \right) $ and $w^{+}$, $w^{-}\in \mathbb{C}$ are such that $w=P^{+}w^{+}+P^{-}w^{-}$. In [@KT; @Transmut Sect. 5] it was shown that $\mathbf{V}_{1}$ and $\mathbf{V}_{1}^{-1}$ are bounded in the space of continuous bicomplex valued functions on $\overline{S}$ with the norm $\left\Vert \cdot \right\Vert $ and hence $\left\Vert \mathbf{V}_{1}\right\Vert $ is a finite number depending on $f$ and $b$ only.
In a full analogy with the proof of Proposition 2.10 from [@KT; @Transmut Sect. 5] we obtain $\Vert \widetilde{W}-\widetilde{W}_{N}\Vert
<\varepsilon \Vert T_{f}^{-1}\Vert $. Then from (\[estimate1\]) we have $$\Vert \mathbf{K}_{f}-\mathbf{jK}_{1/f}-K_{f,N}+\mathbf{j}K_{1/f,N}\Vert <\varepsilon \Vert T_{f}^{-1}\Vert
\Vert \mathbf{V}_{1}\Vert .$$Finally, since $\left\vert \mathcal{I}w\right\vert \leq \left\vert
w\right\vert _{\mathbb{B}}$ (see [@CKr2012 Proposition 2]), we obtain $$\vert \mathbf{K}_{1/f}-K_{1/f,N}\vert \leq \Vert \mathbf{K}_{f}-\mathbf{jK}_{1/f}-K_{f,N}+\mathbf{j}K_{1/f,N}\Vert <\varepsilon
\Vert T_{f}^{-1}\Vert \Vert \mathbf{V}_{1}\Vert$$ which finishes the proof.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank our colleague R. Michael Porter for his valuable help with Wolfram Mathematica.
[99]{} =-1pt M. Abramovitz, I. A. Stegun, *Handbook of mathematical functions*, New York: Dover, 1972.
H. Al[i]{}c[i]{} and H. Taşeli, *Pseudospectral methods for solving an equation of hypergeometric type with a perturbation*, Journal of Computational and Applied Mathematics, 234 (2010), 1140–1152.
H. Begehr and R. Gilbert, *Transformations, transmutations and kernel functions*, vol. 1–2, Longman Scientific & Technical, Harlow, 1992.
A. Boumenir, *Sampling and eigenvalues of non-self-adjoint Sturm-Liouville problems*, SIAM J. Sci. Comput., 23 (2001), no. 1, 219–229.
A. Boumenir, *The approximation of the transmutation kernel*, J. Math. Phys., 47 (2006), 013505.
H. Campos and V. V. Kravchenko, *Fundamentals of bicomplex pseudoanalytic function theory: Cauchy integral formulas, negative formal powers and Schrödinger equations with complex coefficients*, Complex Anal. Oper. Theory 7 (2013), no. 2, 485–518.
H. Campos, V. V. Kravchenko and S. M. Torba, *Transmutations, L-bases and complete families of solutions of the stationary Schrödinger equation in the plane*, J. Math. Anal. Appl., 389 (2012), no. 2, 1222–1238.
R. W. Carroll, *Transmutation theory and applications*, Mathematics Studies, Vol. 117, North-Holland, 1985.
R. Castillo, V. V. Kravchenko, H. Oviedo and V. S. Rabinovich, *Dispersion equation and eigenvalues for quantum wells using spectral parameter power series*, J. Math. Phys., 52 (2011), 043522 (10 pp.).
R. Castillo-Pérez, V. V. Kravchenko and S. M. Torba, *Spectral parameter power series for perturbed Bessel equations*, Applied Mathematics and Computation, 220 (2013), 676–694.
B. Chanane, *Computing the spectrum of non-self-adjoint Sturm-Liouville problems with parameter-dependent boundary conditions*, Journal of Computational and Applied Mathematics, 206 (2007), 229–237.
M. S. Child and A. V. Chambers, *Persistent Accidental Degeneracies for the Coffey-Evans Potential*, J. Phys. Chem., 92 (1988), 3122–3124.
P. J. Davis and P. Rabinowitz, *Methods of numerical integration*, Second edition, New York: Dover Publications, 2007.
M. Dellnitz, O. Schütze and Q. Zheng, *Locating all the zeros of an analytic function in one complex variable*, Journal of Computational and Applied Mathematics, 138 (2002), 325–333.
J. Delsarte, *Sur certaines transformations fonctionnelles relatives aux équations linéaires aux dérivées partielles du second ordre*, C. R. Acad. Sc. 206 (1938), 178–182.
J. Delsarte and J. L. Lions, *Transmutations d’opérateurs différentiels dans le domaine complexe*, Comment. Math. Helv., 32 (1956), 113–128.
L. Erbe, R. Mert and A. Peterson, *Spectral parameter power series for Sturm–Liouville equations on time scales*, Applied Mathematics and Computation, 218 (2012), 7671–7678.
I. S. Gradshtein and I. M. Rizhik, *Table of integrals, series, and products*, Moscow: FizMatLit, 1963 (in Russian).
S. Haber, *Two formulas for numerical indefinite integration*, Math. Comp., 60 (1993), 279–296.
A. D. Hemery and A. P. Veselov, *Whittaker-Hill equation and semifinite-gap Schrodinger operators*, J. Math. Phys., 51 (2010), 072108 (17pp).
R. O. Hryniv and Ya. V. Mykytyuk, *Inverse spectral problems for Sturm-Liouville operators with singular potentials*, Inverse Problems, 19 (2003), no. 3, 665–684.
R. O. Hryniv and Ya. V. Mykytyuk, *On zeros of some entire functions*, Trans. Amer. Math. Soc. 361 (2009), 2207–2223.
K. V. Khmelnytskaya, V. V. Kravchenko and J. A. Baldenebro-Obeso, *Spectral parameter power series for fourth-order Sturm-Liouville problems*, Applied Mathematics and Computation, 219 (2012), 3610–3624.
K. V. Khmelnytskaya, V. V. Kravchenko and H. C. Rosu, *Eigenvalue problems, spectral parameter power series, and modern applications*, Submitted, available at arXiv:1112.1633.
K. V. Khmelnytskaya, V. V. Kravchenko, S. M. Torba and S. Tremblay, *Wave polynomials and Cauchy’s problem for the Klein-Gordon equation*, J. Math. Anal. Appl., 399 (2013), 191–212.
K. V. Khmelnytskaya and H. C. Rosu, *A new series representation for Hill’s discriminant*, Annals of Physics, 325 (2010), 2512–2521.
A. N. Kolmogorov and S. V. Fomin, *Elements of function theory and functional analysis*, Moscow: Fizmatlit, 2004 (in Russian).
V. V. Kravchenko, *A representation for solutions of the Sturm-Liouville equation*, Complex Variables and Elliptic Equations, 53 (2008), 775–789.
V. V. Kravchenko, *Applied pseudoanalytic function theory*, Basel: Birkhäuser, Series: Frontiers in Mathematics, 2009.
V. V. Kravchenko and R. M. Porter, *Spectral parameter power series for Sturm-Liouville problems*, Mathematical Methods in the Applied Sciences, 33 (2010), 459–468.
V. V. Kravchenko and S. Torba, *Transmutations for Darboux transformed operators with applications*, J. Phys. A: Math. Theor., 45 (2012), \# 075201 (21 pp.).
V. V. Kravchenko and S. M. Torba, *Transmutations and spectral parameter power series in eigenvalue problems*, Operator Theory: Advances and Applications, 228 (2013), 209–238.
V. V. Kravchenko and S. M. Torba, *Spectral problems in inhomogeneous media, spectral parameter power series and transmutation operators*, in 2012 International Conference on Mathematical Methods in Electromagnetic Theory (MMET), IEEE Conference Publications, 2012, 18–22.
V. V. Kravchenko and S. M. Torba, *Construction of transmutation operators and hyperbolic pseudoanalytic functions*, Complex Anal. Oper. Theory, 2014, doi:10.1007/s11785-014-0373-3.
V. V. Kravchenko and S. Torba, *Modified spectral parameter power series representations for solutions of Sturm-Liouville equations and their applications*, Applied Mathematics and Computation, 2014, doi:10.1016/j.amc.2014.03.153.
M. A. Lavrentyev and B. V. Shabat, *Hydrodynamics problems and their mathematical models*, Nauka: Moscow, 1977 (in Russian).
V. Ledoux, M. Van Daele and G. V. Berghe, *Efficient numerical solution of the one-dimensional Schrödinger eigenvalue problem using Magnus integrators*, IMA Journal of Numerical Analysis, 30 (2010), 751–776.
B. M. Levitan, *Inverse Sturm-Liouville problems*, VSP, Zeist, 1987.
J. L. Lions, *Solutions élémentaires de certains opérateurs différentiels à coefficients variables*, Journ. de Math. 36 (1957), fasc 1, 57–64.
V. Matveev and M. Salle, *Darboux transformations and solitons*, New York, Springer, 1991.
V. A. Marchenko, *Sturm-Liouville Operators and Applications: Revised Edition*, AMS Chelsea Publishing, 2011.
A. F. Motter and M. A. F. Rosa, *Hyperbolic calculus*, Adv. Appl. Clifford Algebras, 8 (1998), no 1, 109–128.
M. Muhammad, M. Mori, *Double exponential formulas for numerical indefinite integration*, Journal of Computational and Applied Mathematics, 161 (2003), 431–448.
J. W. Paine, F. R. de Hoog and R. S. Anderssen, *On the correction of finite difference eigenvalue approximations for Sturm-Liouville problems*, Computing, 26 (1981), 123–139.
J. D. Pryce, *Numerical solution of Sturm-Liouville problems*, Oxford: Clarendon Press, 1993.
Pruess, S. and C. T. Fulton, *Mathematical software for Sturm-Liouville problems*, ACM Trans. Math. Software, 19(3) (1993), 360–376.
R. D. Riess and L.W. Johnson, *Error Estimates for Clenshaw-Curtis Quadrature*, Numer. Math. 18 (1972), 345–353.
H. Rosu, *Short survey of Darboux transformations*, Proceedings of “Symmetries in Quantum Mechanics and Quantum Optics”, Burgos, Spain, 1999, 301–315.
S. M. Sitnik, *Transmutations and applications: a survey*, arXiv:1012.3741v1, originally published in the book: *Advances in Modern Analysis and Mathematical Modeling*, Editors: Yu.F.Korobeinik, A.G.Kusraev, Vladikavkaz: Vladikavkaz Scientific Center of the Russian Academy of Sciences and Republic of North Ossetia–Alania, 2008, 226–293.
G. Sobczyk, *The hyperbolic number plane*, Coll. Maths. Jour., 26 (1995), no. 4, 268–280.
F. Stenger, *Numerical methods based on Sinc and analytic functions*, Springer-Verlag, New York, 1993.
H. Takahasi and M. Mori, *Double Exponential Formulas for Numerical Integration*, Publications of RIMS, Kyoto University, 9 (1974), 721–741.
K. Tanaka, M. Sugihara and K. Murota, *Numerical indefinite integration by double exponential sinc method*, Math. Comp., 74 (2005), 655–679.
L. N. Trefethen, *Is Gauss quadrature better than Clenshaw-Curtis?*, SIAM Review, 50 (2008), no. 1, pp. 67–87.
K. Trimeche. *Transmutation operators and mean-periodic functions associated with differential operators*, London: Harwood Academic Publishers, 1988.
V. S. Vladimirov, *Equations of mathematical physics*, Moscow: Nauka, 1984 (in Russian); Engl. transl.: of the first edition: N.Y.: Marcel Dekker, 1971.
K. Wright, *Series Methods for Integration*, The Computer Journal, 9(2) (1966), 191–199.
X. Ying and N. Katz, *A Reliable Argument Principle Algorithm to Find the Number of Zeros of an Analytic Function in a Bounded Domain*, Numer. Math. 53 (1988), 143–163.
[^1]: Research was supported by CONACYT, Mexico via the project 166141.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The phase diagram of the staggered six vertex, or body centered solid on solid model, is investigated by transfer matrix and finite size scaling techniques. The phase diagram contains a critical region, bounded by a Kosterlitz-Thouless line, and a second order line describing a deconstruction transition. In part of the phase diagram the deconstruction line and the Kosterlitz-Thouless line approach each other without merging, while the deconstruction changes its critical behaviour from Ising-like to a different universality class. Our model has the same type of symmetries as some other two-dimensional models, such as the fully frustrated XY model, and may be important for understanding their phase behaviour. The thermal behaviour for weak staggering is intricate. It may be relevant for the description of surfaces of ionic crystals of CsCl structure.'
address:
- |
Instituut voor Theoretische Fysica, Universiteit Utrecht, Postbus 80006, 3508 TA Utrecht, Nederland\
and HLRZ, Forschungszentrum Jülich, D-52425 Jülich, Deutschland
- |
Dipartimento di Fisica, Università di Genova, and Istituto Nazionale per la Fisica della Materia (INFM),\
via Dodecaneso 33, 16146 Genova, Italia
- |
Instituut voor Theoretische Fysica, Universiteit Utrecht, Postbus 80006, 3508 TA Utrecht, Nederland\
and HLRZ, Forschungszentrum Jülich, D-52425 Jülich, Deutschland
author:
- 'Enrico Carlon [@newaddress]'
- Giorgio Mazzeo
- Henk van Beijeren
title: A Transfer Matrix Study of the Staggered BCSOS Model
---
[2]{}
INTRODUCTION {#sec:intro}
============
Six vertex models were introduced by Slater [@Slter] to describe ferroelectricity in two dimensional networks. Placing arrows on the bonds of a square lattice one can define the sixteen possible arrangements of arrows pointing towards and away from a lattice point as vertices. In six vertex models only those six vertex configurations are kept (see Fig. \[FIG01\]) that satisfy the [*ice rule*]{}, i.e. they have two arrows pointing in and two pointing out at each vertex. Assigning energies $\epsilon_1$, …, $\epsilon_6$ to these vertices one obtains a class of exactly solved models [@LiebWubook; @Baxter].
Six vertex models can also be interpreted as surface models, by mapping them to the so-called body centered solid on solid (BCSOS) models [@HvBPRL], defined as limiting cases of a lattice gas, or Ising model, on a body centered cubic lattice. Therefore the phase structure of the six vertex model as function of its vertex weights can be translated directly to the surface phase structure of the corresponding BCSOS model. The mapping turned out to be important in understanding the properties of the [*roughening transition*]{} [@HenkIm]. Using the exact solution of the six vertex model it was found that roughening is a transition of infinite order of Kosterlitz-Thouless (KT) type, confirming previous renormalisation group results [@ChuiWeeks].
Experimental situations often are too complex to allow even a qualitative description by the exactly solved BCSOS models. Various extensions of the standard six vertex model have been proposed to deal with these cases. Two main classes may be identified: one where interactions between vertices are added, and another one in which the vertex lattice is split into two sublattices with different vertex energies. These modifications, however, lead to models which, apart from some particular cases [@BaxFreeFer; @FreeFer], lose the property of being exactly solvable. Other techniques (e.g. numerical ones) have to be adopted. Models in the first class have been proposed to account for further neighbour interactions between surface atoms, which may change the symmetry of the ground state and give rise to phase transitions other than the roughening transition. Vertex interactions were introduced to reproduce the $(2 \times 1)$ reconstruction of the (110) face of fcc noble metals like Au and Pt [@LeviTouz]. This led to investigations on equilibrium phase transitions on these surfaces as well as on surfaces of lighter metals like Ag, Rh, etc... [@MJLT]. A model of the (100) surface of an fcc crystal exhibiting a $(2 \times 2)$ reconstructed ground state [@BastKnops] has recently extended the list. The second class of models, with vertex weights alternating on the two sublattices, are known as [*staggered six vertex models*]{}. A staggering only involving the weights of vertices 5 and 6 corresponds to the imposition of a “staggered field”, i.e. a field coupled to the arrow directions that changes sign between neighbouring arrows. This gives rise to an inverse roughening transition in part of the phase diagram [@Erik]. Alternating the values for the energies of the vertices 1, 2 and 3, 4 on the two sublattices leads to a model known as “the staggered six vertex model” (or staggered BCSOS model) in the literature. In a large part of its parameter space it can be mapped onto the Ashkin-Teller model [@atdual]. Using this transformation Knops investigated the phase structure of the staggered BCSOS model in part of its phase diagram by renormalisation group methods [@Knops], but until recently a large region of the phase diagram has remained unexplored.
In this paper we present a complete account of our investigations, over the full range of parameters, of the staggered BCSOS model. A preliminary description has been given already in Ref. [@ourPRL], here we present further details as well as a number of new results.
In the unexplored region of the phase diagram the model has a ground state which is twofold degenerate, therefore it has a symmetry of Ising type. The twofold degeneracy is lost at a second order transition line which approaches another line of KT roughening transitions. The interplay between the two is particularly interesting, especially since a similar interplay between a KT and a second order transition has been found for several different models, among which other models for reconstructed surfaces [@MJLT; @denNijs3], but also the fully frustrated XY model [@ffXY1; @ffXY2; @Olsson] and coupled XY – Ising models [@XYising1; @XYising2]. They have received a great deal of attention in recent years and till now their critical behaviour is not fully understood. The strong interplay between Ising and KT degrees of freedom may lead to several possible scenarios where, in a certain region of the phase diagram, either the two transitions occur close to each other but remain separate, or they merge into a single phase transition, which may perhaps belong to a new universality class.
Apart from these more theoretical aspects the model is likely to be relevant for the study of the equilibrium properties of a certain class of crystal surfaces, e.g. the (001) surface of ionic crystals of the CsCl structure. This too will be discussed in some detail.
The paper is structured as follows. In Section II we give a description of the model. In Section III we present its full phase diagram. In Section IV we review the techniques employed in our studies, i.e. the transfer matrix method and finite size scaling, and discuss the correlation functions and free energies we calculated to derive our results. In Section V we discuss the critical exponents of the model and some possible scenarios for the changes in the critical behaviour along the deconstruction line. In Section VI we conclude with a brief discussion of related models.
The staggered six vertex model
==============================
The partition function of the six vertex model is given by $$\begin{aligned}
Z=\sum_{\{\cal C\}} \,e^{\textstyle{- \beta \sum_{i=1}^6
n_i({\cal C}) \, \epsilon_i}}
\label{partfun}\end{aligned}$$ where the sum runs over the set of all allowed vertex configurations $\{{\cal C} \}$ and $n_i({\cal C})$ denotes the number of vertices of type $i$ in the configuration $\cal C$ ($\beta = 1 /k_B T $, with $k_B$ Boltzmann’s constant and $T$ the temperature). The model has been solved exactly [@LiebWu] for any choice of values of the energies $\epsilon_i$ ($i=1 \ldots 6$). A relatively simple choice of the vertex energies is given by $\epsilon_1 = \epsilon_2 = \epsilon_3 = \epsilon_4 = \epsilon$ and $\epsilon_5 = \epsilon_6 = 0$ which defines, for $\epsilon
>0$, the so-called F model. The ground state is twofold degenerate and is composed of vertices 5 and 6 arranged alternatingly in a chess board configuration. The low temperature phase is usually called an “antiferroelectric" phase, since along both horizontal and vertical rows the arrows predominantly alternate in direction. From the exact solution it is known that this system undergoes an infinite order phase transition to a disordered paraelectric state at $\beta \epsilon = \ln 2$.
As already pointed out in the introduction, the six vertex models are isomorphic to a class of solid on solid (SOS) models called BCSOS models [@HvBPRL]. Microscopic configurations of an SOS model are given in terms of discrete heights $h_i$ of surface atoms with respect to a reference plane. All lattice sites up to these heights are occupied and all sites above them are empty. In the BCSOS model the height variables are placed on the dual lattice of the six vertex lattice. This is subdivided into an even and an odd sublattice, which are intertwined in a chess board pattern and on which the surface heights assume even respectively odd values only. The even sites will be referred to as black (B) sites and the odd ones as white (W) sites. In addition the height differences between neighbouring sites are restricted to the values $\pm 1$. The mapping of six vertex configurations to corresponding configurations of a BCSOS model is very simple. The height differences between neighbouring sites are put in a one-to-one correspondence with the arrow directions in the six vertex configuration. The convention is that the higher of the two surface sites is at the right side of the arrow. Given a configuration of vertices, the configuration of heights is fixed uniquely once the height of a reference atom has been fixed (see Fig. \[FIG01\]).
The vertex energies can be reinterpreted in terms of bond energies between the atoms. When periodic boundary conditions are applied along the (say) horizontal direction of the vertex lattice the number of vertices 5 and 6 per row is equal, therefore with no loss of generality one can always choose $\epsilon_5 = \epsilon_6 = 0$, fixing the point of zero energy. The vertices 5 and 6 describe local configurations in which the height variables on either diagonal are equal (see Fig. \[FIG01\]). Vertices 1, 2, 3 and 4 correspond to configurations where the height variables along either of the two diagonals are different, therefore $\epsilon_1$, $\epsilon_2$, $\epsilon_3$ and $\epsilon_4$ can be viewed as energies needed to break a next nearest neighbour bond and produce a height difference of two vertical lattice units between neighbouring sites of equal colour.
In the ordinary BCSOS model the distinction between B and W atoms has been introduced only for convenience of description, but the two sublattices are equivalent and are treated exactly on the same footing. Knops [@Knops] extended the model to a two component system where the B and W atoms are physically different. While energy zero is still attributed to all vertices 5 and 6, Knops assigned two different energies, $\epsilon$ and $\epsilon^{\prime}$, to broken bonds between W-W and B-B atoms respectively. In terms of the six vertex representation also the vertex lattice is divided into two alternating sublattices I and II on which the vertices assume different energies as follows:
$$\begin{aligned}
\left\{\begin{array}{cccr}
& \mbox{on sublattice I:} \; &\epsilon_1=\epsilon_2=\epsilon;\;\;
\epsilon_3=\epsilon_4=\epsilon^{\prime};\;\;
\epsilon_5=\epsilon_6=0 &
\\
& \mbox{on sublattice II:} \; &\epsilon_1=\epsilon_2=\epsilon^{\prime};\;\;
\epsilon_3=\epsilon_4=\epsilon;\;\;
\epsilon_5=\epsilon_6=0 & \;.
\end{array}\right.
\label{energies}\end{aligned}$$
This choice defines the staggered six vertex model. In the BCSOS representation the model is described by the hamiltonian: $$\begin{aligned}
H &=& \frac{\epsilon}{2} \sum_{\langle ij \rangle}
\left\vert h_i^W - h_j^W \right\vert +
\frac{\epsilon^\prime}{2}\sum_{\langle kl \rangle}
\left\vert h_k^B - h_l^B \right\vert
\label{hamiltonian}\end{aligned}$$
[2]{} subject to the constraint that the height difference between neighbouring B and W sites is $\pm 1$. The first sum in (\[hamiltonian\]) runs over all pairs of neighbouring W sites on the surface and the second sum over the corresponding B pairs. Throughout this article we will also use the parameter $\delta$, defined by the relation $\epsilon^{\prime}=\epsilon + 2 \delta$. As mentioned already in the Introduction the model defined here will be referred to as the staggered six vertex (or BCSOS) model. Obviously when $\epsilon = \epsilon^{\prime}$ ($\delta =0$) one recovers the usual F model.
For negative values of the vertex energies $\epsilon$ and $\epsilon^{\prime}$, the system may model ionic crystals of bcc structure as for instance CsCl [@simple]. The constraint of minimal height difference between neighbouring surface sites reflects the effects of the strong attraction between oppositely charged ions, while neighbouring pairs of the same colour, having equal charges, repel each other. It is further assumed that on top of the Coulombic repulsion other interactions, as for instance spin exchange, generate a slight difference in the energies for broken bonds between B-B and W-W pairs ($\epsilon \neq \epsilon^{\prime}$). In the staggered BCSOS model the interaction range is limited to next nearest neighbours and to have a more realistic representation of ionic crystals one needs to extend the interactions to further neighbours. Yet we expect the phase structure described here for the staggered BCSOS model may be encountered in real ionic crystals.
THE PHASE DIAGRAM {#sec:phase}
=================
We have investigated the phase diagram of our model by means of transfer matrix and finite size scaling techniques, which will be the subject of Section \[sec:numerical\]. Here we present the main results. Since the model shows a trivial symmetry upon exchange of $\epsilon$ and $\epsilon^{\prime}$, corresponding to the replacement $(\delta,\epsilon)
\longleftrightarrow (-\delta,\epsilon+2\delta)$, we can restrict ourselves to the region $\delta \ge 0$. The phase diagram naturally divides into three sectors of globally different behaviour, though smoothly connected to each other. These are described in the three following subsections.
The range
----------
For positive values of $\epsilon$ and $\epsilon^{\prime}$ Knops [@Knops] investigated the phase diagram through a mapping onto the Ashkin-Teller model [@atdual]; the phase diagram of the latter had been obtained before by renormalisation group methods [@ATRG]. On the $\beta \epsilon$ axis the (001) surface in the corresponding BCSOS model is in a flat phase for $\beta \epsilon > \ln2$, whereas the interval $\beta\epsilon \le \ln2$ represents the temperature region in which the surface is rough. The infinite order transition occurring at the KT point $\beta \epsilon = \ln 2$, $\beta \delta = 0$ corresponds to the roughening transition of this surface.
Roughening is a phase transition which can be characterized by the vanishing of the free energy of a [*step*]{}, separating two surface regions of different average height. The roughening transition results into a proliferation of steps leading to a delocalisation of the surface position and to a logarithmic divergence of the mean square height difference at large distances:
$$\begin{aligned}
G(R_{ij})=\langle \left( h_i - h_j \right)^2 \rangle \; \sim 2 a_V^2 K(T)
\ln R_{ij}
\;\;\;\;\mbox{for}\;\;\;\; \mbox{$R_{ij} \rightarrow \infty$}
\;\;\;\;\mbox{and}\;\;\;\; \mbox{$T \ge T_R$}
\label{gofr}\end{aligned}$$
[2]{} with $a_V$ the vertical lattice spacing, $K(T)$ a temperature dependent prefactor, $R_{ij}$ the distance between the lattice sites $i$ and $j$ and $T_R$ the roughening temperature. Below $T_R$, $G(R)$ saturates for large $R$ at a temperature dependent constant value. Renormalisation group calculations [@Ohta] show that at $T_R$ the prefactor assumes the universal value $$\begin{aligned}
K(T_R)=\frac{1}{\pi^2} \, .
\label{koft}\end{aligned}$$ In the particular case of the exactly solved F model, $K(T)$ is known for every temperature above $T_R$ [@axe], that is $$\begin{aligned}
K(T)=\frac{1}{\pi\arccos\Delta}
\label{univjump}\end{aligned}$$ where $\Delta = 1 - e^{2 \beta \epsilon}/2$. In fact Eqs. (\[gofr\]) and (\[univjump\]) are valid not only in the high temperature phase of the F model, but also for $\delta =0$, $\beta \epsilon < 0$, which defines the so-called inverted F model [@Erik]: all along the negative $\beta \epsilon$ axis the surface is in a rough state.
For $\delta \neq 0$ Knops found two critical lines originating from the KT point and running into the regions $\delta > 0$ and $\delta < 0$. The lines represent phase transitions of Ising type from an ordered low temperature phase to a disordered flat (DOF) phase, similar to the phase introduced by Rommelse and Den Nijs [@DOF].
The ground state of the model is twofold degenerate. At higher temperatures the more weakly bound sublattice fluctuates above and below the more strongly bound sublattice, which remains almost localised at a given level. In the limit $\delta \rightarrow \infty$ the model can be mapped [*exactly*]{} onto the two dimensional Ising model, which is critical at $\beta \epsilon = \ln(1 + \sqrt 2)$; the strong sublattice is “frozen" to height (say) zero, the only freedom left for the heights of the other sublattice is to take the values $\pm 1$ just below or above that of the strong sublattice (see Fig. \[FIG02\]). According to the renormalisation group results obtained for the Ashkin-Teller model [@ATRG] the phase transition remains in the universality class of the two dimensional Ising model all along the critical line down to $\delta = 0$.
Starting from the low temperature phase and increasing the temperature, the system undergoes an Ising transition to the DOF phase, while roughening is pushed up to infinite temperature [@Knops]. We reinvestigated this part of the phase diagram with the transfer matrix methods to be described in Section \[sec:numerical\] and obtained results in full agreement with those of Knops.
The range
----------
When $\epsilon$ becomes negative the ground state of the system changes drastically (irrespective of the sign of $\epsilon^{\prime}$). Breaking bonds between white atoms now lowers the energy, so that at zero temperature one finds the black sublattice unbroken (provided $\delta >0$) while atoms of the white sublattice are found alternatingly above and below the black sublattice (see Fig. \[FIG03\]). This surface configuration is commonly referred to as a $c(2\times 2)$ reconstructed surface. In the equivalent six vertex representation the ground state is formed by columns of vertical arrows running alternately all upwards and all downwards, and by rows of horizontal arrows running alternately all right and all left. Such an arrangement of directed paths is known as a [*Manhattan lattice*]{}, due to its resemblance to the one-way street pattern of Manhattan. As the energy is invariant under the reversal of all arrows, the ground state is twofold degenerate, just as in the case $\epsilon>0$. Indeed in the limit $\delta \to \infty$ the model can be mapped [*exactly*]{} onto an antiferromagnetic Ising model leading to the value $\beta\epsilon=-\ln(1+\sqrt{2})$ for the critical temperature. This constitutes a horizontal asymptote, as in the case $\epsilon > 0$, for a second order transition line, whose existence can be deduced again from the mapping of the staggered six vertex model onto the Ashkin-Teller model [@critfan]. It separates a low temperature $c(2\times 2$) reconstructed phase from a high temperature DOF phase, where the reconstruction order is lost but the surface is still globally flat: it is the same DOF phase found for $\epsilon > 0$; no singularities are met in crossing the $\beta \delta $-axis. Our transfer matrix calculations confirm the existence of this critical line and show it exhibits Ising type critical behaviour throughout sector B.
The range {#sec:subC}
----------
For $\epsilon^{\prime} < 0 $ the mapping of the staggered six vertex model to the Ashkin-Teller model leads to negative Boltzmann weights in the latter. It loses its physical relevance and cannot be used any more to make predictions on the phase behaviour of the staggered six vertex model. In spite of this Kohmoto [*et al.*]{} [@critfan] have made some conjectures, which have proven to be correct, on the physical situation beyond the “horizon" $\epsilon + 2 \delta =0$.
Our transfer matrix analysis shows the existence of three phases: a low temperature $c(2\times 2)$ reconstructed phase and a DOF phase, which are present already in sector B, and a rough phase, which is found only in the present sector. Two critical lines separate these phases, as shown in Fig. \[FIG04\]: the first one is just the continuation of the second order line beyond the horizon. It still separates the $c(2\times 2)$ region from the DOF region and asymptotically approaches the axis $\beta\delta =0$. We have strong indications that, within at least a major part of the sector $\epsilon^{\prime} < 0 $, this line does not belong to the Ising universality class. We will present the evidence for this in Section \[sec:exponents\]. The other critical line is a line of KT points separating the rough region (or [*critical fan*]{}, as predicted already by Kohmoto [*et al.*]{} [@critfan]) from the DOF region.
The point where the KT line meets the vertical axis can be determined from the exact solution of the F model as the point where the prefactor of the logarithmic term in the mean square height difference is four times as large as its universal value assumed at the ordinary roughening temperature of the F model $$\begin{aligned}
K(T) &=& 4 K(T_R) = \frac{4}{\pi^2}
\label{factfour}\end{aligned}$$ from which one obtains, inverting (\[univjump\]): $$\begin{aligned}
\beta \epsilon &=& \frac{1}{2} \ln\left(2 - \sqrt{2} \right)
\approx -0.2674 \, .\end{aligned}$$ The factor four in (\[factfour\]) stems from the fact that for $\delta \neq 0$ the roughening transition is driven by steps of a height of two vertical lattice units (as in Ref. [@Erik]), due to the inequivalence between the two atomic sublattices.
A simple estimate of the roughening transition temperature based on a random walk approximation (see Ref. [@ourPRL]) yields
$$\begin{aligned}
e^{-2\beta\delta}\,+\,e^{\beta\epsilon} &=& 1 \, .
\label{pottsline}\end{aligned}$$
This line has been drawn in Fig. \[FIG04\]. Indeed, for large and negative $\beta\epsilon$ it is seen to run very close to the KT line, which we could determine with great accuracy by the methods described in the next section.
A most remarkable feature of our phase diagram is the apparent merging of the second order and the KT line into a single line (see Fig. \[FIG04\]).
Their horizontal distance $d$ as a function of $\beta\epsilon$ can be well described by a curve of the form $d(\beta \epsilon) = C e^{\alpha \left( \beta \epsilon \right)} $, with $\alpha \simeq 12$ [@ourPRL]. This exponential fit suggests that although the two lines are coming rapidly closer together as $\beta \epsilon$ is decreasing, they do never actually merge. Other fits, of the form $d(\beta\epsilon) = C \left\vert \beta\epsilon - \beta\epsilon_0
\right\vert ^ {\alpha}$, which would be expected to work in the case of a merging of the lines at $\epsilon=\epsilon_0$, could not be stabilised against changes in the fitting range.
The apparent non-crossing of the two critical lines at first looks very surprising. At low temperatures a domain wall between two different Ising phases mainly consists of diagonal sequences of vertices 5 and 6, as depicted in Fig. \[FIG05\]; its energy per unit length approximately equals $-\epsilon/\sqrt{2}$. On the other hand a step consists mainly of long horizontal and vertical chains of overturned arrows and has an approximate energy per unit length of $2\delta$. To a first approximation steps do not couple with the Ising order, since the reconstructed phase remains the same at both sides of the step (see Fig. \[FIG06\]). Hence one would expect the KT line (characterized by vanishing step free energy) and the Ising line (vanishing Ising domain wall free energy) to cross near $\epsilon+2\sqrt{2} \, \delta=0$. We think that the actual non-crossing of the two lines can be explained as follows. When temperature is raised, more and more closed steps will be formed on the surface as one approaches the roughening temperature $T_R$. On these steps the direction of the arrows is reversed. In this way the Ising order parameter becomes more and more diluted, which will, eventually, strongly reduce the free energy of a domain wall. If in the end the closed steps become so prolific that they cover on average half of the surface, without becoming of infinite length, the Ising order disappears without roughening of the surface.
For $2\delta \lesssim 0.4|\epsilon|$ the thermal behaviour implied by our phase diagram is quite intricate and remarkable. At low temperatures the surface is in a $c(2 \times 2)$ reconstructed flat phase, then on raising the temperature there is a second order transition to a DOF phase, rapidly followed by a KT transition to a rough phase. Next there is a reentrant KT transition to the DOF phase. This is an inverted roughening transition similar to the one described in [@Erik]. Finally, as temperature approaches infinity, the system asymptotically approaches a rough phase again. Instead for $2\delta \gtrsim 0.4 \vert \epsilon \vert$ the system goes through a single phase transition from the ordered to a DOF phase and remains flat for all finite temperatures.
TRANSFER MATRIX AND FINITE SIZE SCALING METHODS {#sec:numerical}
===============================================
Transfer matrix techniques are frequently used in studies of the critical properties of two dimensional systems with short range interactions. The construction of the transfer matrix (TM) follows a standard procedure and the interested reader is referred to the existing literature [@Baxter] for details.
We use two different transfer matrices, one oriented parallel to the axes of the vertex lattice and another one tilted of $45^\circ$ with respect to these axes. We refer to the former as [*vertical*]{} TM and to the latter as [*diagonal*]{} TM (Fig. \[FIG07\] shows a configuration of the diagonal TM). We consider a lattice of width $N$ and height $M$, with periodic boundary conditions in the horizontal direction. For the vertical TM the subdivision of the lattice into a white and a black sublattice, combined with the periodic boundary conditions, restricts $N$ to even values. For the diagonal TM the horizontal and vertical axis are chosen along the diagonals of the vertex lattice and $N$ can be odd as well as even. The element $T_{ij}$ of the matrix is defined as the Boltzmann weight of a row of $N$ vertices generated by arrow configurations labeled by the indices $i$ and $j$. One has $T_{ij}=0$ if this row of vertices does not satisfy the ice rule. For the vertical TM, if $i$ and $j$ are identical there are in fact two possible configurations of rows of vertices: in this case the transfer matrix simply sums their Boltzmann weights.
There are $2^N$ different arrow configurations for the vertical TM, whereas for the diagonal TM this number is $2^{2N}$. The largest values of $N$ we could treat numerically were $N=22$ for the vertical and $N=12$ for the diagonal TM. Actually, due to the rotation of the lattice over $45^\circ$ the latter should be compared to $12 \sqrt{2} \approx 17$ for the vertical TM.
In the limit $M \rightarrow \infty$ the partition function per row becomes: $$\begin{aligned}
\lim_{M \rightarrow \infty} \left( Z_{N \times M} \right)^{1/M}
=\lambda_0(N)\end{aligned}$$ with $\lambda_0(N)$ the largest eigenvalue of $T$, from which the free energy per row follows as $$\begin{aligned}
\beta f(N) = - \ln \lambda_0(N) \, .
\label{free}\end{aligned}$$
To each state $i$ we associate a polarisation $P_i = N_{i \uparrow}
- N_{i \downarrow}$, with $N_{i\uparrow}$ and $N_{i \downarrow}$ the total numbers of up and down [@note2] arrows in the state $i$. By virtue of the periodic boundary conditions in the horizontal direction the transfer matrix can be reduced to blocks of fixed polarisation, since $T_{ij} = 0$ if $P_i \neq P_j$ (see, for instance, Ref. [@Baxter]). The so-called central block is the one corresponding to zero polarisation and describes a flat surface. The blocks with polarisation $P_i = \pm 2$ (subcentral blocks) describe a surface with a step.
The difference between the free energies of a surface with a step and of a flat surface gives the step free energy, which, per unit of length, on an $N \times \infty$ strip can be expressed as: $$\begin{aligned}
\beta f_S(N) &=& - \left( \ln \lambda_1(N) - \ln \lambda_0(N) \right)
\label{STEPF}\end{aligned}$$ where $\lambda_{0}(N)$ and $\lambda_{1}(N)$ are the largest eigenvalue of the central and the subcentral block respectively [@fnlambda]. The study of this quantity will allow us to determine the roughening temperature.
The deconstruction transition can be studied by considering two correlation lengths, which are both defined within the central block. We define the inverse correlation length $\xi_D^{-1}$ as: $$\begin{aligned}
\xi_{D}^{-1}(N) &=& -(\ln \lambda_2(N) - \ln \lambda_0(N) )
\label{defcsiN}\end{aligned}$$ where $\lambda_{2}(N)$ is the second largest eigenvalue of the central block. The other correlation length can be calculated from the diagonal TM as the inverse of the domain wall free energy per unit length $f_W^{-1}(N)$, where $f_W(N)$ is given by $$\begin{aligned}
f_W(N) &=& f(N) - \frac{f(N+1) + f(N-1)}{2}
\label{deffw}\end{aligned}$$ with $N$ odd. Indeed in the diagonal TM an $N \times \infty$ strip, with $N$ odd, is partially frustrated since it cannot accommodate the Manhattan ground state without creating a domain wall (see Fig. \[FIG07\]). $f_W^{-1}(N)$ can be interpreted as the correlation length connected to the correlation function between two disorder variables [@KadCeva].
Conformal invariance [@Cardybook] predicts that, at a critical point, the correlation lengths scale as $N$, so the deconstruction transition can be located at the crossing point of the curves representing the scaled quantities $N/\xi_D$ and $N \beta f_W$ as functions of the temperature for different sizes. In reality, as shown in Figs. \[FIG08\](a) and (b), no perfect crossing is found. Instead, pairs of curves obtained for sizes $N$ and $N+2$ intersect each other in a sequence of points, $(\beta \delta_D(N),\beta \epsilon_D(N))$, respectively $(\beta \delta_W(N),\beta \epsilon_W(N)) $, which converges to the infinite system critical point $(\beta \delta_D,\beta \epsilon_D)$, respectively $(\beta \delta_W,\beta \epsilon_W)$. An extrapolation procedure requiring several iterations [@iteratedfit] is then used to estimate $\beta\delta_D$ or $\beta\delta_W$. Of course the two independent estimates of the critical point have to coincide, which provides a good check on the internal consistency and accuracy of our procedures.
To locate the roughening temperature one has to employ a different method. The scaling $f_S(N) \sim 1/N$ holds not only at the KT transition but also inside the rough region, where the surface is in a critical state. There the curves $N f_S(N)$, plotted as functions of temperature for different values of N, coalesce in the limit $N \rightarrow \infty$ and the point where they detach from each other can be identified as the KT point (see Fig. \[FIG09\]). For an accurate location of $T_R$ one has to use the universal properties of the KT transition which give rise to the scaling prediction [@Erik; @logcorr]
$$\begin{aligned}
N \beta f_S(N) = \frac{\pi}{4} + \frac{1}{A + B \ln N}
\label{ktscaling}\end{aligned}$$
which holds exactly at $T = T_R$, with $A$ and $B$ nonuniversal constants. The constant $\pi/4$ is characteristic for steps with a height of two vertical lattice spacings. The free energy of such a step corresponds to the line tension between a vortex-antivortex pair with vorticity 2 in the dual representation [@Erik]. The KT transition temperature is determined by requiring that a three-point fit of the form $N \beta f_S(N) = A_0+1/\left(A+B\ln N\right)$ yields $A_0 = \pi/4$. For the extrapolation we used iterated fits in the spirit of Ref. [@iteratedfit]. We performed this procedure along different lines across the phase diagram, scanning lines with $\beta \delta$ fixed, lines with $\beta \epsilon$ fixed and thermal trajectories.
CRITICAL EXPONENTS AND CENTRAL CHARGE {#sec:exponents}
=====================================
As we noted in the previous section, the critical line separating the flat from the rough region can be well characterized as a KT line. As we will see, the critical properties of the second order line are less well determined, especially in the region $\epsilon^\prime < 0$. We will calculate critical exponents and central charge pertaining to the deconstruction transition using finite-size scaling methods. The two exponents $\alpha$ and $\nu$ are related to the behaviour of the singular part of the surface free energy $f_{sing} \sim
t^{2-\alpha}$ and of the domain wall free energy $f_W \sim t^{\nu}$ [@othernu] (where $t = (T-T_D)/T_D$, $T_D$ the deconstruction temperature). They satisfy the finite size scaling predictions $$\begin{aligned}
\frac{1}{N}
\frac{\partial^2 f(N)}{\partial t^2} \sim N^{\frac{\alpha}{\nu}}
\label{specheat}\end{aligned}$$ and $$\begin{aligned}
N \frac{\partial f_W(N)}{\partial t} \sim N^{\frac{1}{\nu}}
\label{oneovernu}\end{aligned}$$ respectively, valid at the critical point $T=T_D$. Two other critical indices we will calculate are $$\begin{aligned}
x & = & \left. \frac{1}{2 \pi} \lim_{N\to\infty} \frac{N}{\xi_D(N)}
\right\vert_{T=T_D}
\label{xandx}
\\ \nonumber
\\
x^{\prime} & = & \left. \frac{1}{2 \pi} \lim_{N\to\infty} N
\beta f_W\left(N\right)
\right\vert_{T=T_D}
\label{xandxprime}\end{aligned}$$ which represent the exponent of the spin-spin correlation function [@eta] and that of the correlation function between disorder variables [@KadCeva] respectively. The numerical errors on the values assumed by these quantities are obtained as follows. We first evaluate the error on the determination of the critical temperature $\Delta T_D$ from the quality of the extrapolation to $N \to \infty$ of our finite size data [@iteratedfit]. Subsequently, we extract the values of the exponents, again by iterated fits, at three different temperatures: $T_D - \Delta T_D$, $T_D$ and $T_D + \Delta T_D$. This procedure allows to determine the maximum possible variation on the values of $\alpha/\nu$, $\nu$, $x$ and $x^\prime$, thus assigning them an error bar. Notice these errors are typically small if the critical temperature is determined accurately enough.
Finally, from conformal invariance [@Cardybook; @BloCarAffl] it follows that the leading finite size correction to the free energy per site of an infinite system with periodic boundary conditions, $\widetilde f_{\infty}$, is determined by the central charge (or conformal anomaly) $c$ as $$\begin{aligned}
\frac{f(N)}{N} \approx \widetilde f_{\infty} + \frac{\pi c}{6 N^2} \, .
\label{cc}\end{aligned}$$ In fact we analyzed the central charge using the finite size approximation
$$\begin{aligned}
c(N,N+2) = \frac{3}{2 \pi}
\frac{N^2 \left( N+2 \right)^2}{\left(N+1\right)}
\left(\frac{f(N)}{N} - \frac{f(N+2)}{N+2}\right)
\label{ccc}\end{aligned}$$
which converges to $c$ in the limit $N \to \infty$.
With the techniques described above we find that the deconstruction line for $\epsilon < 0$ belongs no doubt to the Ising universality class in sector B of our phase diagram. Good convergence with increasing size is obtained for the critical exponents as well as for the central charge, the values of which are: $$\begin{aligned}
\alpha = 0 \,\,\,\,\,\,\,\, \nu = 1
\,\,\,\,\,\,\,\, x = x^{\prime} = \frac{1}{8}
\,\,\,\,\,\,\,\, c = \frac{1}{2} \; .
\label{Isingexp}\end{aligned}$$ In the region $\epsilon^{\prime} < 0$ the situation is less clear. The convergence of the data with increasing system size is worse, the values of some of the critical exponents seem to vary along the critical line and the central charge cannot be determined with any great accuracy. Yet our results seem to clearly rule out the possibility that the critical line remains in the Ising universality class. We present the results for the various exponents and for the central charge below and then draw some more general conclusions.
[The exponent]{} {#sec:espx}
----------------
In part of sector C of the phase diagram we find difficulties in convergence for the quantities extracted from the correlation length $\xi_D(N)$. Fig. \[FIG10\](a) shows the behaviour of the exponent $x$ obtained from Eq. (\[xandx\]), only along part of the deconstruction line. The extrapolation procedure to infinite size is in fact far from trivial close to the horizon $\epsilon + 2 \delta = 0$, where we find non-monotonic behaviour with increasing size for $N/\xi_D(N)$ and even for the sequence $\beta \delta_D(N)$. In order to give an estimate of the exponent $x$ nonetheless, we looked at the quantity $x(N) \equiv N/(2 \pi \widetilde \xi_D(N))$, where $\widetilde \xi_D(N)$ is the correlation length evaluated now at the intersection points $(\beta \delta_D(N),\beta \epsilon_D(N))$.
Fig. \[FIG10\](b) shows some plots of $x(N)$ vs. $N$ along the deconstruction line. The curves 1, 2 and 3 refer to critical points in sector B located on the deconstruction line at $\beta \delta = 0.88$, $\beta \delta = 0.60$ and $\beta \delta = 0.45$. They show a good convergence to the Ising exponent $x =1/8$. The other curves, (4-10), refer to the values $\beta \delta=
0.37$, $0.31$, $0.28$, $0.25$, $0.23$, $0.21$, $0.20$, $0.19$ in sector C of the phase diagram. As the system size increases the curves (4-7) show a reentrant behaviour towards the value $x = 1/8$. At values of $\beta \delta \lesssim 0.20$ we find monotonic convergence again as function of system size, but to values which vary continuously as shown in Fig. \[FIG10\](a). The behaviour of this set of curves suggests that along any thermal scan in sector C of the phase diagram the quantity $x(N)$ will show an asymptotic decrease after a maximum. The position of the maximum gradually shifts to higher values of size until it exceeds the largest value accessible to our calculations and eventally disappears from sight. As already mentioned, no accurate fit can be performed on curves 4, 5, 6 and 7 of Fig. \[FIG10\](a), though a rough estimate provides values of $x$ below $1/8$. When a fitted value can be extracted again (at smaller values of $\beta \delta$) and drawn in Fig. \[FIG10\](a), one should thus be cautioned against the possibility of missing a maximum and a decreasing part. This would provide values of $x$ possibly below $1/8$ and more in accordance with those of $x^{\prime}$ given in the following sub-section. However, another difficulty may arise: in the vicinity of the roughening transition the correlation lenght $\xi_D$ may also be strongly influenced by steplike excitations [@privdennijs]. A better quantity to look at is represented by the exponent $x^{\prime}$.
[The exponent]{} {#sec:espxp}
----------------
The quantity $N \beta f_W(N)$, converges monotonically as function of the system size $N$ all along the deconstruction line. For $0.25 \lesssim \beta \delta \lesssim 0.40$ the convergence is slow, but it is still possible to give an estimate of the exponent $x^{\prime}$ using Eq. (\[xandxprime\]). However the error bars are fairly large. We notice a change in the direction of convergence: $N \beta f_W(N)/2 \pi$ converges to $x^{\prime}$ from above in sector B of the phase diagram but from below in sector C. Around the line $\epsilon^{\prime} = 0$ finite size effects are very small. For $0.3 \lesssim \beta \delta \lesssim 0.4$ the exponent is still compatible, within error bars, with the Ising value of $1/8$, as shown in Fig. \[FIG11\], but for $\beta \delta \lesssim 0.3$ the exponent shifts towards values well below this.
[The exponents]{} [and]{}
-------------------------
Fig. \[FIG12\] shows the exponents $\alpha$ and $\nu$ calculated with the diagonal transfer matrix along the deconstruction line, with the aid of standard extrapolation methods based on the scaling relations (\[specheat\],\[oneovernu\]) [@Night]. It is almost impossible to obtain these exponents using the vertical transfer matrix, due to difficulties in convergence with increasing size. These problems are much less severe with the diagonal transfer matrix (see also the Appendix), even though the maximum available system size is smaller. The values thus obtained are not compatible with Ising exponents when $\epsilon^{\prime} < 0$. They do satisfy the hyperscaling relation $2 \nu = 2 - \alpha$ within error bars.
The central charge
------------------
In general the central charge $c$ vanishes in non-critical phases (here the flat reconstructed phase and the DOF phase) and assumes finite values at critical points or inside critical regions (like the rough phase). As in the determination of the exponents $\alpha$ and $\nu$, $c$ is calculated with the diagonal TM, as this leads to better convergence and smaller finite size effects than calculations with the vertical TM. Fig. \[FIG13\](a) shows finite size approximations of $c$ along vertical lines in the phase diagram based on Eq. (\[ccc\]). The left part of Fig. \[FIG13\](a) refers to a scan with $\beta
\delta =0.55$, which crosses the deconstruction line in a point of sector B, where we find exponents in the Ising universality class. In this case the central charge at the transition shows good convergence towards the Ising value ($c = 1/2$). The right part of Fig. \[FIG13\](a) refers to a scan which crosses the deconstruction line in a point of sector C with $\beta \delta = 0.25$. Fig. \[FIG13\](b) shows two other plots of central charges along vertical lines with $\beta \delta =0.22$ (left) and $\beta \delta
=0.20$ (right). In this part of the phase diagram the central charge increases markably beyond the Ising value $c = 1/2$. Due to strong finite size effects, slow convergence and nearness of the KT line we cannot give a reasonable estimate for its actual value.
Fig. \[FIG13\](c) shows the central charge calculated along the thermal trajectory $\epsilon + 10 \delta = 0$, starting from the rough region (at small $\beta \delta$) towards the reconstructed phase at larger $\beta \delta$. According to our numerical results the line $\epsilon + 10 \delta = 0$ crosses the roughening and the deconstruction line in two points very close in temperature. In the infinite system limit the central charge should be 1 in the rough region, drop abruptly from $1$ to $0$ at the KT point, remain $0$ in the DOF region, assume a non-zero value at the single point where the trajectory crosses the deconstruction line and remain $0$ beyond that. In finite systems this behaviour is smeared out, as is the case also in the other plots in Fig. \[FIG13\]. Hence, since the KT point and the deconstruction transition are extremely close on this trajectory, one expects to see an apparent convergence of $c$ to the sum of the KT value $1$ and that of the deconstruction transition. For a deconstruction of Ising type this would yield $c=3/2$. From conformal invariance [@Cardybook] it follows that for unitary models with central charge smaller than unity $c$ can only assume the values $$\begin{aligned}
c &=& 1 - \frac{6}{M(M+1)} \;\;\;\; \mbox{with $M=3,4,\ldots \;\;$ .}
\label{discretec}\end{aligned}$$ The Ising value $c=1/2$ is the lowest possible value, obtained with $M=3$. Higher values of $M$ correspond to phase transitions in different universality classes. From Fig. \[FIG13\](c) it is apparent that $c$ converges to a value larger than $3/2$, which we estimate around $c = 1.7$ –$1.8$.
Deconstruction of non-Ising type?
---------------------------------
The results presented above strongly suggest that the deconstruction transition is not in the Ising universality class in, at least, part of the region $\epsilon^{\prime}<0$. One cannot entirely exclude the possibility that the observed deviations of critical exponents and central charge from their Ising values are due to strong cross-over effects, induced by the vicinity of the KT line [@privdennijs], rather than being a genuine feature of the deconstruction transition; in view of our numerical results however, we believe this is quite unlikely.
Of course, the next intriguing question is: what, if not Ising, is the universality class of this reconstruction line? The answer to this question is not easy and our numerical results are not conclusive.
In general the exponents vary along the deconstruction line, although some vary less than others. The exponent $x$ shows generally worse convergence than the exponent $x^\prime$ and extrapolation of the values of $x$ in part of the phase digram turned out impossible due to the non-monotonic behaviour of the finite size data as function of the system size $N$.
The exponent $x^\prime$ varies along the deconstruction line as well, but it remains roughly constant in a limited region around the value of $\beta \delta \approx 0.2$, with small error bars thanks to rapid convergence of the finite size data. At smaller values of $\beta \delta$ its value increases as well as its error bars. This may be due to the vicinity of the KT line or to the finite size effects caused by the increasing length of straight step segments. In general finite size effects increase at smaller values of $\beta \delta$ (see also the Appendix); in this part of the phase diagram the most important excitations consist of closed loops of reversed arrows which may become very elongated as the energy per unit of length for a straight segment is proportional to $2 \delta$. One should expect that finite size effects are particularly strong when the typical size of a loop becomes of the same order of magnitude as the width of the strip, $N$. Slow convergence also is present in a region to the left of the line $\epsilon^\prime = 0$, as can be seen from the large error bars around $\beta \delta \approx 0.3$ in Fig. \[FIG11\]. This is due to a poor determination of the value of the deconstruction temperature $T_D$.
The exponents $\alpha$ and $\nu$, as calculated from Eqs. (\[specheat\],\[oneovernu\]), vary along the deconstruction line in sector C. However, the hyperscaling relation $2 \nu = 2 - \alpha$ is always satisfied within error bars. In general, as shown in Fig. \[FIG12\], $\alpha$ tends to have larger error bars than $\nu$. In the region $\beta \delta \approx 0.2$, the convergence is rapid in the sense that a two parameter fit is sufficient to extract $\alpha$ and $\nu$ from (\[specheat\]) and (\[oneovernu\]). At smaller values of $\beta \delta$ one in general needs to consider corrections to scaling using a three parameter fit.
Unfortunately our numerical results do not allow an inequivocal identification of the critical behaviour of the deconstruction transition in the region $\epsilon^\prime < 0$. We notice however that the exponent $x^\prime$ remains constant in a region around $\beta \delta
\approx 0.2$, where the error bars are smallest. In this region also $\alpha$ and $\nu$ converge rapidly with increasing size, compared to other parts of the deconstruction line in sector C. One possible candidate for the observed exponents in this region could be that of the four state Potts model, for which $\alpha = \nu = 2/3$, compatible with our calculated values of $\alpha$ and $\nu$.
Conformally invariant models are classified according to the value of their central charge, which can assume only discrete values depending on some integer $M$, as given in Eq. (\[discretec\]). At fixed values of $M$ conformal invariance [@Cardybook] predicts also the possible values for the exponents of correlation functions at the critical point. For the four state Potts model, the predicted exponents are of the type $x, x^{\prime} = 2 p^2/q^2$ with $p$ and $q$ integers, as pointed out in Ref. [@NienKnops]. For $p=1$ and $q=4$ one indeed obtains the well-known magnetic exponent $1/8$, instead for $p=1$, $q=5$ one obtains the value $2/25$. Both values are shown as horizontal dashed lines in Fig. \[FIG11\]; the exponent $x^{\prime}=2/25$ seems to fit the measured values of the exponent very well for $\beta \delta \approx 0.2$. For the two dimensional Ising model conformal invariance predicts the exponents $x=1/8$ (magnetic) and $x=1$ (thermal) only. Thus a measured exponent of value $x^{\prime} \approx 2/25$ is a quite clear sign of non-Ising critical behaviour.
Further, the central charge clearly shifts away from its Ising value $c=1/2$. For the four state Potts model we should expect a central charge equal to $1$ ($M \to \infty$ in Eq. (\[discretec\])). The central charge markably increases in the region $\epsilon^\prime < 0$. However, like for the critical exponents, this increase goes smoothly from the Ising value, $c=1/2$, towards higher values. The central charge calculated along the line $\epsilon + 10
\delta = 0$, where the deconstruction and roughening line are almost coinciding in temperature, extrapolates to $c = 1.7-1.8$, well above the Ising plus KT value $c = 1/2 + 1 = 3/2$. As pointed out above, this is another indication of non-Ising behaviour of the deconstruction transition, though not quite compatible with that of the four state Potts model, which would imply a central charge equal to $c = 1 + 1 = 2$.
Finally, also the possibility of having a line with continuously varying exponents, as the behaviour of especially the exponent $x$ in the sector C suggests (Fig. \[FIG10\](a)), should be considered. In this case the central charge would equal unity, as in the four state Potts model.
Anyhow, as discussed in Section V A - B, the convergence of $x$ is much poorer than that of $x^{\prime}$. The slow shift of $x^{\prime}$ away from the Ising value as $\beta \delta$ decreases in sector C (Fig. \[FIG11\]) is known to be a common feature of finite size scaling in the vicinity of points where a change of universality class occurs. Moreover, as $\beta \delta$ becomes very small, the nearness of the KT line is seen to influence the convergence of the exponents of the deconstruction line. In conclusion, it seems quite plausible to have in practice only a window of $\beta \delta$ values where constant critical exponents are found. To enlarge this window one would have to consider larger system sizes.
Bastiaansen and Knops [@BastKnops] recently studied a six vertex model with an extended range of interactions. They also found a phase diagram with a second order line approaching a KT line. The exponents of the second order line clearly deviate from their Ising values and the authors suggested they might be explained as tricritical Ising exponents. Applied to the staggered BCSOS model this would mean a deconstruction line of Ising type with a tricritical point, continuing beyond this point as a first order line, which is the phase behaviour of the annealed diluted Ising (or Blume-Capel [@BlumeCapel]) model. The exponents at the tricritical point would be $\alpha = 8/9$, $\nu = 5/9$ and the central charge $c = 7/10$. For the exponent $x^\prime $ conformal invariance predicts $x^\prime = 3/40$. Around $\beta \delta \approx 0.2$ the extrapolated value of $x^\prime$ would also be compatible with this value, but $\alpha$ and $\nu$ are far away from their tricritical values. At smaller values of $\beta \delta$ we do find exponents which approach those of the tricritical Ising model, but this happens in a region where the values we obtain for $x^\prime$ clearly shift away from $3/40$ and where in general finite size effects are quite strong. These same finite size effects also make it impossible to tell whether at sufficiently small $\beta \delta$ the deconstruction line becomes first order or not.
The point along the deconstruction line where the change of universality class occurs is not sharply determined by our numerical results. We do not observe an abrupt jump of the exponents at a given point, rather a continuous shift. A reasonable candidate for the point separating the two regions (i.e. Ising and non-Ising), could be the point where the deconstruction line crosses the line $\epsilon^{\prime} = \epsilon + 2 \delta =0$. Crossing this line, we find changes in the type of convergence of the exponents $x$ and $x^{\prime}$ (Sections V-A, B), although without an abrupt change in their values. We recall that in the surface representation of the model, in one region the coupling constants between the atoms in the two sublattices are both negative ($\epsilon < 0$, $\epsilon^{\prime}<0$); in the other (where the deconstruction transition is of Ising type) one of the two coupling constants is positive ($\epsilon^{\prime}>0$). In terms of the vertex lattice, in the region $\epsilon^{\prime}>0$, vertices 5 and 6 are the excited vertices with the lower energy above the ground state value; at $\epsilon^{\prime}<0$, vertices 5 and 6 get the higher excitation energy.
All these considerations suggest that the properties of the system may change between the two regions $\epsilon^\prime > 0$ and $\epsilon^\prime < 0$ and make it more plausible that the shifts in the exponents are not just due to cross-over, but also result from a real change of universality class of the deconstruction transition [@footnote].
DISCUSSION AND CONCLUSION {#sec:discuss}
=========================
In this article we studied the critical properties of the staggered BCSOS model. Using transfer matrix techniques we found two critical lines describing the deconstruction and the roughening of the $(001)$ surface of a two component bcc crystal.
The two lines approach each other in part of the phase diagram, apparently without merging. According to our results the deconstruction line in part of the phase diagram changes its universality class from Ising to non-Ising, although further investigations are needed to make this point more convincing. On the basis of the exponents we find, we conclude that a possible universality class matching these exponents reasonably well, in the region where the best convergence is found, is that of the four state Potts model. Another possible scenario is that proposed by Bastiaansen and Knops [@BastKnops]. In their six vertex model with interactions extended to further neighbours it is hard too to distinguish between a single critical line and two lines approaching each other, but remaining separate. They find critical exponents for the deconstruction transition clearly deviating from the Ising values and conjecture that the observed criticality could be explained as tricritical Ising behaviour. The idea of a diluted Ising model is particularly attractive in our case where, as we have seen, the deconstruction transition is, to all likeliness, the consequence of the dilution of the Ising order in the system caused by the formation of a large number of closed steps of finite length. Unfortunately we find little numerical evidence for this scenario. Finally, also the possibility of having a line with continuously varying exponents cannot be completely excluded. Other models of reconstructed surfaces have been studied by several authors. Den Nijs [@denNijs3] introduced a model that describes $(110)$ missing-row reconstructed surfaces of some fcc metals ($Au$, $Pt$, …). He found a deconstruction and a roughening line merging into a single critical line, whith Ising and KT behaviour simply superimposed. From his data, as presented in the literature, it is not possible to really distinguish between actual merging or mere rapid approach of the lines. A clear distinction to our model is that in Den Nijs’ model the deconstruction transition remains of Ising type throughout. Another class of models for the same metal surfaces has been developed and extensively studied by the Trieste group [@MJLT; @Santoro]. Again a deconstruction line and a KT roughening line are seen to approach each other. Depending on the microscopic details of the model, the deconstruction line keeps its Ising character either all along, or up to a tricritical point where it changes to a first order line.
As mentioned in the Introduction there are several other two dimensional models with KT and Ising degrees of freedom. One that has received a lot of attention, starting from the beginning of the last decade [@Teitel], is the fully frustrated XY model, which describes certain two dimensional Josephson junction arrays. The study of its critical behaviour has led to several different conjectures about its universality class and critical exponents. Several papers [@ffXY1; @ffXY2] report non-Ising exponents and it has been suggested that the model would belong to a novel type of universality class. To our knowledge, whether this type of universality class would or would not coincide with that of some known models has not been established yet. In the most recent study concerning the fully frustrated XY model Olsson [@Olsson] presents evidence of two separate transitions, a KT and an Ising one where the former occurs at somewhat lower temperature than the latter: $T_{KT} < T_{IS}$. This would be in agreement with our results since the XY model can be mapped onto a solid-on-solid model via a duality transformation [@KnopsXY], which maps the low temperature phase of one model onto the high temperature phase of the other and vice versa. Olsson’s work suggests that the non-Ising exponents observed by other groups are due to the failure of some finite-size scaling hypothesis used in previous works. In the staggered BCSOS model instead, we find clear evidence of non-Ising exponents. Unfortunately there exists no exact mapping between this model and the fully frustrated XY model, therefore they may well be in different universality classes. Yet, we hope that some of the ideas developed in this paper to study the staggered BCSOS model, will be generalised to other models so as to reach a deeper understanding of their critical properties.
It is a pleasure to thank Paul Bastiaansen, Henk Blöte, Hubert Knops, Bernard Nienhuis and Marcel den Nijs for stimulating discussions. Financial support permitting several meetings between the authors of this paper is gratefully acknowledged. In particular H. v. B. thanks the Centro di Fisica delle Superfici e delle Basse Temperature (CNR), Genova, while G. M. thanks the Instituut voor Theoretische Fysica, Utrecht, and acknowledges the kind hospitality of prof. Dietrich Wolf at HLRZ, Jülich.
FINITE SIZE EFFECTS {#finite-size-effects .unnumbered}
===================
We show here how the diagonal and the vertical transfer matrix have different finite size effects in part of the phase diagram. Consider first the phase point $\delta = 0$, $\epsilon
\rightarrow -\infty$. Vertices 5 and 6 are absent and the partition function can be calculated easily. Consider lattices of size $N \times M$ with cylindrical geometry, that is the $N$ vertices along a horizontal row are connected to each other through periodic boundary conditions. In the vertical transfer matrix there are in total $2^{N+M}$ configurations, and the free energy per site is given by: $$\begin{aligned}
-\beta \widetilde f = \frac{N+M}{N M} \ln 2
\label{finit1}\end{aligned}$$ (the symbol $\widetilde f$ is used to distinguish the free energy per site from $f$, the free energy per row). In the thermodynamic limit $N$,$M \rightarrow \infty$ the free energy per site vanishes.
However in transfer matrix calculations one takes the limit $M \rightarrow \infty$ keeping $N$ finite; this gives a free energy per site equal to: $$\begin{aligned}
-\beta \widetilde f = \frac{1}{N} \ln 2 \, .\end{aligned}$$
With the diagonal transfer matrix the total number of configurations available is $2^{2N}$, since once the arrows on a row are fixed the whole configuration is fixed. Repeating the same calculation as done above one finds a free energy per site: $$\begin{aligned}
-\beta \widetilde f = 0
\label{finit2}\end{aligned}$$ independent of the value of $N$. The conclusion is that the free energy shows finite size corrections of the order $1/N$ in the vertical transfer matrix, while there are no finite size effects for the diagonal transfer matrix. In Fig. \[FIG14\] we plot the free energy calculated along the line $e^{\beta \epsilon} + e^{-2 \beta \delta} =1$ with the vertical and diagonal transfer matrices. The endpoint $e^{-2 \beta \delta} =1$ corresponds to the free energy which we calculated in (\[finit1\]) and (\[finit2\]). As can be seen from the figure, there is a wide area to the left of this point where the free energies calculated from the vertical transfer matrix show large finite size effects, while the convergence is much faster for the diagonal transfer matrix. In both cases the convergence is faster for $e^{-2 \beta \delta} < 1/2$, that is in the region of the phase diagram where only a second order line is present. Obviously, for small values of $\delta$ the boundary effects are very strong, due to closed loops of reversed arrows winding around the cylinder, i.e. the ribbon which constitutes our system with periodic boundary conditions in the horizontal direction (see Fig. \[FIG15\]). These closed loops are more frequent in the vertical transfer matrix, since one can reverse the arrows along a horizontal line with a cost in energy of $2 \delta N$. Closed loops in the diagonal transfer matrix require at least a vertex 5 and a vertex 6, and therefore occur less frequently.
[99]{}
Present address: Instituut voor Theoretische Fysica, Katholieke Universiteit Leuven, Celestijnenlaan 200D, B-3001 Leuven, Belgium.
J. C. Slater, J. Chem. Phys. [**9**]{}, 16 (1941).
E. H. Lieb and F. Y. Wu, in [*Phase Transitions and Critical Phenomena*]{}, edited by C. Domb and M. S. Green (Academic Press, London, 1972), Vol. 1, pp. 331–490.
R. J. Baxter, [*Exactly Solved Models in Statistical Mechanics*]{} (Academic Press, London, 1982).
H. van Beijeren, Phys. Rev. Lett., 993 (1977).
H. van Beijeren and I. Nolden in [*Structure and Dynamics of Surfaces*]{}, edited by W. Schommers and P. von Blanckenhagen (Springer-Verlag, Berlin Heidelberg, 1987), Vol. 2, pp. 259–300.
S. T. Chui and J. D. Weeks, Phys. Rev. B [**14**]{}, 4978 (1976).
R. J. Baxter, Phys. Rev. B [**1**]{}, 2199 (1970).
F. Y. Wu and K. Y. Lin, Phys. Rev. B [**12**]{}, 419 (1975).
A. C. Levi and M. Touzani, Surf. Sci. [**218**]{}, 223 (1989).
G. Mazzeo, G. Jug, A. C. Levi and E. Tosatti, Phys. Rev. B [**49**]{}, 7625 (1994).
P. J. M. Bastiaansen and H. J. F. Knops, Phys. Rev. B [**53**]{}, 126 (1996).
E. Luijten, H. van Beijeren and H. W. J. Blöte, Phys. Rev. Lett. [**73**]{}, 456 (1994).
J. Ashkin and E. Teller, Phys. Rev. [**64**]{}, 178 (1943).
H. J. F. Knops, Phys. Rev. B [**20**]{}, 4670 (1979).
G. Mazzeo, E. Carlon and H. van Beijeren, Phys. Rev. Lett. [**74**]{}, 1391 (1995); Surf. Sci. [**352-354**]{}, 960 (1996). M. den Nijs, Phys. Rev. Lett. [**66**]{}, 907 (1991); Phys. Rev. B [**46**]{}, 10386 (1992).
G. Ramirez-Santiago and J. V. José, Phys. Rev. Lett. [**68**]{}, 1224 (1992).
Y. M. M. Knops, B. Nienhuis, H. J. F. Knops and H. W. J. Blöte, Phys. Rev. B [**50**]{}, 1061 (1994).
P. Olsson, Phys. Rev. Lett. [**75**]{}, 2758 (1995).
E. Granato and J. M. Kosterlitz, Phys. Rev. B [**33**]{}, 4767 (1986).
E. Granato, J. M. Kosterlitz, J. Lee and M. P. Nightingale, Phys. Rev. Lett. [**66**]{}, 1090 (1991).
E. H. Lieb, Phys. Rev. Lett. [**18**]{}, 1046 (1967).
Crystallographically $CsCl$ has a simple cubic structure; however, if one ignores the identities of the different ions one obtains a bcc lattice.
F. Y. Wu and K. Y. Lin, J. Phys. C [**7**]{}, L181 (1974); H. J. F. Knops, J. Phys. A [**8**]{}, 1508 (1975); S. E. Ashley, J. Phys. A [**11**]{}, 2015 (1978).
T. Ohta and K. Kawasaki, Prog. Theor. Phys. [**60**]{}, 365 (1978).
R. W. Youngblood and J. D. Axe, Phys. Rev. B, [**23**]{}, 232 (1981).
K. Rommelse and M. den Nijs, Phys. Rev. Lett., 2578 (1987); M. den Nijs and K. Rommelse Phys. Rev. B, [**40**]{}, 4709 (1989).
The thermal trajectories are those trajectories in the phase diagram obtained by fixing the values of the energies and varying the temperature.
M. Kohmoto, M. den Nijs and L. P. Kadanoff, Phys. Rev. B [**24**]{}, 5229 (1981).
In the diagonal transfer matrix $N_{i \uparrow}$ and $N_{i \downarrow}$ refer to the number of arrows with vertical components pointing up or down respectively.
The exponent $x$ is related to the more standard exponent $\eta$ as $x=\eta/2$.
L. P. Kadanoff and H. Ceva, Phys. Rev. B [**3**]{}, 3918 (1971).
J. L. Cardy, in [*Phase Transitions and Critical Phenomena*]{}, edited by C. Domb and J. L. Lebowitz (Academic Press, London, 1987), Vol. 11, pp. 55–126.
M. P. Nightingale, J. Appl. Phys. [**53**]{}, 7927 (1982).
H. W. J. Blöte and B. Nienhuis, J. Phys. A [**22**]{}, 1415 (1989).
H. W. J. Blöte and P. Nightingale, Phys. Rev. B [**47**]{}, 15046 (1993).
With $\nu$ we indicate both the exponent of the correlation length $\xi$ and of the interface free energy $f_W$. We recall that the general scaling hypothesis predicts that for all systems near the critical point $f_W \xi \sim k_B T$ holds. For more details the reader may consult: B. Widom, J. Chem. Phys. [**43**]{}, 3892 (1965) and [**43**]{}, 3898 (1965), and D. B. Abraham, Phys. Rev. B [**19**]{}, 3833 (1979). In Ref. [@ourPRL] we adopted $\xi_D$ to extract $\nu$, here better convergence is achieved by using $f_W$, anyway the finite size scaling analysis on these two different quantitites lead to the same estimate for $\nu$ when they both converge.
The largest eigenvalue of the transfer matrix is always found in the central block.
H. W. J. Blöte, J. L. Cardy and M. P. Nightingale Phys. Rev. Lett. [**56**]{}, 742 (1986); I. Affleck, Phys. Rev. Lett. [**56**]{}, 746 (1986).
M. den Nijs (private communication).
B. Nienhuis and H. J. F. Knops, Phys. Rev. B [**32**]{}, 1872 (1985) and references therein.
M. Blume, Phys. Rev. [**141**]{}, 517 (1966); H. W. Capel, Physica [**32**]{}, 966 (1966).
Of course, if the second scenario, of a diluted Ising line, would apply, the tricritical point is far removed from the point $\epsilon^{\prime}=0$ and there would have to be strong cross-over effects.
G. Santoro and M. Fabrizio, Phys. Rev. B [**49**]{}, 13886 (1994); G. Santoro, M. Vendruscolo, S. Prestipino and E. Tosatti, Phys. Rev. B [**53**]{}, 13169 (1996).
S. Teitel and C. Jayaprakash, Phys. Rev. B [**27**]{}, 598 (1983).
H. J. F. Knops, Phys. Rev. Lett. [**39**]{}, 766 (1977).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A semi-dualizing module over a commutative noetherian ring $A$ is a finitely generated module $C$ with $\RHom_A(C,C) \simeq A$ in the derived category $\sD(A)$.
We show how each such module gives rise to three new homological dimensions which we call $C$–Gorenstein projective, $C$–Gorenstein injective, and $C$–Gorenstein flat dimension, and investigate the properties of these dimensions.
address:
- ' [Matematisk]{} Afdeling, Universitetsparken 5, DK–2100 København Ø, Denmark, e-mail address: [`[email protected]`]{.nodecor} '
- ' [Department]{} of Pure Mathematics, University of Leeds, Leeds LS2 9JT, United Kingdom, e-mail address: [`[email protected]`]{.nodecor}, homepage: [`http://www.maths.leeds.ac.uk/popjoerg`]{.nodecor} '
author:
- Henrik Holm
- Peter Jørgensen
title: 'Semi-dualizing modules and related Gorenstein homological dimensions'
---
Introduction {#introduction .unnumbered}
============
It is by now a well-established fact that over any associative ring $A$, there exists a Gorenstein injective, Gorenstien projective and Gorenstein flat dimension defined for complexes of $A$–modules. These are usually denoted $\Gid_A(-)$, $\Gpd_A(-)$ and $\Gfd_A(-)$, respectively. Some references are [@Wintherbook], [@CFH], [@HHGorensteinHomDim], and [@Veliche].
In this paper, we need to consider *semi-dualizing* $A$–modules $C$ (see Definition \[def:semi-dual\]), and in order to make things less technical, we only consider commutative and noetherian rings.
For any semi-dualizing module (in fact, complex) $C$ over $A$, and any complex $Z$ with bounded and finitely generated homology, Christensen [@Wintherpaper] introduced the dimension $\GdimC Z$, and developed a satisfactory theory for this new invariant.
If $C$ is a semi-dualizing $A$–module and $M$ is any $A$–complex, then we suggested in [@HHPJ2] the viewpoint that one should change rings from $A$ to $A \ltimes C$ (the *trivial extension* of $A$ by $C$; see Definition \[def:trivial\_extension\]), and then consider the three “ring changed” Gorenstein dimensions: $$\CGid M \quad , \quad \CGpd M \quad \textnormal{ and } \quad \CGfd
M.$$ The usefulness of this viewpoint was demonstrated as it enabled us to introduce three new *Cohen-Macaulay dimensions*, which characterize Cohen-Macaulay rings in a way one could hope for.
In this paper, we define for every semi-dualizing $A$–module $C$, three new Gorenstein dimensions: $$ \GIdim(-) \quad , \quad \GPdim(-) \quad \textnormal{ and } \quad
\GFdim(-),$$ which are called the *$C$–Gorenstein injective*, *$C$–Gorenstein projective* and *$C$–Gorenstein flat dimension*, respectively (see Definition \[def:new\_dimensions\]).
It is worth pointing out that the, say, $C$–Gorenstein injective dimension is defined in terms of resolutions consisting of so-called *$C$–Gorenstein injective $A$–modules* (see Definition \[def:Cmodules\]); and it does not involve a change of rings. The $C$–Gorenstein dimensions have at least five nice properties:
1. For complexes with bounded and finitely generated homology, our $\GPdim(-)$ agrees with Christensen’s $\GdimC(-)$ (Proposition \[pro:GdimC\]).
2. The three $C$–Gorenstein dimensions always agree with the “ring changed” dimensions $\CGid(-)$, $\CGpd(-)$ and $\CGfd(-)$, which were so important in [@HHPJ2] (Theorem \[thm:main\_theorem\]).
3. If $C=A$, the $C$–Gorenstein dimensions agree with the classical Gorenstein dimensions $\Gid_A(-)$, $\Gpd_A(-)$ and $\Gfd_A(-)$.
If $A$ admits a dualizing complex $D$; cf. [@CFH Definition (1.1)], then finiteness of the $C$–Gorenstein dimensions can be interpreted in terms of *Auslander and Bass categories* (see Remark \[rmk:AB\]):
1. If we define $C^{\dagger} = \RHom_A(C,D)$, then for all (appropriately homologically bounded) $A$–complexes $M$ and $N$, we have the following implications (Theorem \[thm:AB\]): $$\begin{aligned}
M \in \sA_{C^{\dagger}}(A) &\,\Leftrightarrow\, \GPdim M<\infty
\,\Leftrightarrow\, \GFdim M<\infty; \\
N \in \sB_{C^{\dagger}}(A) &\,\Leftrightarrow\, \GIdim N<\infty.
\end{aligned}$$
This generalizes the main results in [@CFH Theorems (4.3) and (4.5)].
Finally, each of the three $C$–Gorenstein dimensions has a related *proper* variant, giving us three additional dimensions (Definitions \[dfn:ProperDimensions\] and \[dfn:SpecialProperDimensions\]): $$ \ProperGIdim(-) \quad , \quad \ProperGPdim(-) \quad \textnormal{ and
} \quad \ProperGFdim(-).$$ It turns out that the best one could hope for really happens, as we in Theorems \[thm:proper\_CGid\], \[thm:proper\_CGpd\] and \[thm:proper\_CGfd\] prove:
1. The proper $C$–Gorenstein dimensions (whenever these are defined) agree with the ordinary $C$–Gorenstein dimensions.
The paper is organized as follows:
In Section \[sec:AxC\] we have collected some fundamental facts about the trivial extension $A \ltimes C$, which will be important later on. Section \[sec:CGdims\] defines the three new $C$–Gorenstein dimensions and proves how they are related to the “ring changed” Gorenstein dimensions over $A \ltimes C$. Section \[sec:Christensen\] compares our $\GPdim(-)$ with Christensen’s $\GdimC(-)$. In Section \[sec:AB\] we interpret the $C$–Gorenstein dimensions in terms of Auslander and Bass categories. Finally, Section \[sec:proper\_dimensions\] investigates the proper $C$–Gorenstein dimensions.
[**Setup and notation.**]{} Throughout this paper, $A$ is a fixed commutative and noetherian ring with unit, and $C$ is a fixed semi-dualizing $A$–module; cf. Definition \[def:semi-dual\] below.
We work within the derived category $\sD(A)$ of the category of $A$–modules; cf. e.g. [@HartsResDual Chapter I] and [@Weibel Chapter 10]; and complexes $M \in \sD(A)$ have differentials going to the right: $$M = \xymatrix{\cdots \ar[r] & M_{i+1} \ar[r]^-{\partial^M_{i+1}} &
M_i \ar[r]^-{\partial^M_i} & M_{i-1} \ar[r] & \cdots}.$$ We consistently use the hyper-homological notation from [@Wintherbook Appendix], in particular we use $\RHom_A(-,-)$ for the right derived Hom functor, and $-\LTensor_A-$ for the left derived tensor product functor.
A few results about the trivial extension {#sec:AxC}
=========================================
In this section we collect some fundamental results about the trivial extension, which will be important later on.
\[def:semi-dual\] A finitely generated $A$–module $C$ with $\RHom_A(C,C) \simeq A$ in $\sD(A)$ is called *semi-dualizing* ($C=A$ is such an example).
\[def:trivial\_extension\] If $C$ is any $A$–module, then the direct sum $A \oplus C$ can be equipped with the product: $$(a,c)\cdot(a',c') \,=\, (aa',ac'+a'c).$$ This turns $A \oplus C$ into a ring which is called the *trivial extension* of $A$ by $C$ and denoted $A \skewtimes C$.
There are canonical ring homomorphisms, $\!\!\xymatrix{A \ar@<0.5ex>[r]
& A \ltimes C, \ar@<0.5ex>[l]}\!$ which enable us to view $A$–modules as $(A \ltimes C)$–modules, and vice versa. This will be done frequently.
We import from [@HHPJ2 Lemma 3.2] the following facts about the interplay between the rings $A$ and $A \ltimes C$:
\[lem:extension\_formulae\] Let $A$ be a ring with a semi-dualizing module $C$.
1. There is an isomorphism in $\sD(A\ltimes C)$: $$\RHom_A(A \skewtimes C,C) \cong A \skewtimes C.$$
2. There is a natural equivalence of functors on $\sD(A)$: $$\RHom_{A \skewtimes C}(-,A \skewtimes C) \simeq \RHom_A(-,C).$$
3. If $M$ is in $\sD(A)$ then the two biduality morphisms: $$\begin{array}{c}
M \longrightarrow \RHom_A(\RHom_A(M,C),C)
\ \text{ and } \medskip \\
M \longrightarrow \RHom_{A \skewtimes C}(\RHom_{A \skewtimes
C}(M,A \skewtimes C),A \skewtimes C)
\end{array}$$ are equal.
4. There is an isomorphism in $\sD(A \ltimes C)$:
[3]{} & & \_[A C]{}(A,A C) C. & &&
Furthermore, we have the next result [@HHPJ2 Lemma 3.1] about injective modules over $A$ and $A \ltimes C$:
\[lem:induced\_injectives\] The following two conclusions hold:
1. If $I$ is a (faithfully) injective $A$–module then $\Hom_A(A
\skewtimes C,I)$ is a (faithfully) injective $(A \skewtimes
C)$–module.
2. Each injective $(A \skewtimes C)$–module is a direct summand in a module $\Hom_A(A \skewtimes C,I)$ where $I$ is some injective $A$–module. $\qed$
Using the same methods, we obtain:
\[lem:induced\_projectives\] The following two conclusions hold:
1. If $P$ is a projective $A$–module then $(A \skewtimes
C)\ten_AP$ is a projective $(A \skewtimes C)$–module.
2. Each projective $(A \skewtimes C)$–module is a direct summand in a module $(A \skewtimes C)\ten_AP$ where $P$ is some projective $A$–module. $\qed$
$C$–Gorenstein homological dimensions {#sec:CGdims}
=====================================
Let $M$ be an (appropriately homologically bounded) $A$–complex. In [@HHPJ2] we demonstrated the usefulness of changing rings from $A$ to $A \ltimes C$, and then considering the “ring changed” Gorenstein dimensions: $$\CGid M \quad , \quad \CGpd M \quad \textnormal{ and } \quad \CGfd
M.$$ This point of view enabled us to introduce three *Cohen-Macaulay dimensions* which characterize Cohen-Macaulay local rings in a way one could hope for. The next result is taken from [@HHPJ2 Lemma 4.6].
\[prop:dual\] If $E$ is a faithfully injective $A$–module, and $M$ is any homologically right-bounded $A$–complex, then:
[3]{} & & \_A(M,E) = M. & &&
\[lem:calculation\] Let $J$ be an injective $A$–module and $Q$ a projective $A$–module. Then we have a natural equivalence of functors on $\sD(A \ltimes C)$:
1. $\RHom_{A \ltimes C}(\Hom_A(A \ltimes C,J),-) \,\simeq\,
\RHom_A(\Hom_A(C,J),-)$.
2. $\RHom_{A \ltimes C}(-,(A \ltimes C)\ten_AQ) \,\simeq\,
\RHom_A(-,C\ten_AQ)$.
\(1) is [@HHPJ2 Lemma 3.4], and (2) is proved similarly.
\[cor:Ext\] For any $A$–module $M$, and integer $n$ we have:
1. $\Ext^n_A(\Hom_A(C,J),M)=0$ for all injective $A$–modules $J$ if and only if $\Ext^n_{A \ltimes C}(U,M)=0$ for all injective $(A
\ltimes C)$–modules $U$.
2. $\Ext^n_A(M,C\ten_AP)=0$ for all projective $A$–modules $P$ if and only if $\Ext^n_{A \ltimes C}(M,S)=0$ for all projective $(A \ltimes C)$–modules $S$.
\(1) follows from Lemmas \[lem:calculation\](1) and \[lem:induced\_injectives\], while (2) is a consequence of Lemmas \[lem:calculation\](2) and \[lem:induced\_projectives\].
We need to recall the next result from [@HHPJ2 Lemma 4.1]. Its proof uses, in fact, Lemmas \[lem:calculation\](1) and \[lem:induced\_injectives\].
\[lem:CGorInj/A<=GorInj/AxC\] Let $M$ be an $A$–module which is Gorenstein injective over $A
\skewtimes C$. Then there exists a short exact sequence of $A$–modules, $$0 \rightarrow M^{\prime} \longrightarrow \Hom_A(C,I)
\longrightarrow M \rightarrow 0,$$ where $I$ is injective over $A$ and $M^{\prime}$ is Gorenstein injective over $A \skewtimes C$. Furthermore, the sequence stays exact if one applies to it the functor $\Hom_A(\Hom_A(C,J),-)$ for any injective $A$–module $J$. $\qed$
“Dualizing” the proof of Lemma \[lem:CGorInj/A<=GorInj/AxC\]; this time using Lemmas \[lem:calculation\](2) and \[lem:induced\_projectives\], we establish the next:
\[lem:CGorProj/A<=GorProj/AxC\] Let $M$ be an $A$–module which is Gorenstein projective over $A
\skewtimes C$. Then there exists a short exact sequence of $A$–modules, $$0 \rightarrow M \longrightarrow C\ten_AP \longrightarrow M'
\rightarrow 0,$$ where $P$ is projective over $A$ and $M^{\prime}$ is Gorenstein projective over $A \skewtimes C$. Furthermore, the sequence stays exact if one applies to it the functor $\Hom_A(-,C\ten_A Q)$ for any projective $A$–module $Q$. $\qed$
The last result we will need to get started is [@HHPJ2 Lemma 3.3]:
\[lem:Proj/A=>GorProj/AxC\] The $A$–modules $A$ and $C$ are Gorenstein projective over $A
\skewtimes C$. If $I$ is an injective $A$–module, then $\Hom_A(A,I)
\cong I$ and $\Hom_A(C,I)$ are Gorenstein injective over $A \ltimes
C$. $\qed$
Next, we introduce three new classes of modules:
\[def:Cmodules\] An $A$–module $M$ is called *$C$–Gorenstein injective* if:
1. $\Ext_A^{\geqslant 1}(\Hom_A(C,I),M)=0$ for all injective $A$–modules $I$.
2. There exist injective $A$–modules $I_0,I_1,\ldots$ together with an exact sequence: $$\cdots \to \Hom_A(C,I_1) \to \Hom_A(C,I_0) \to M \to 0,$$ and also, this sequence stays exact when we apply to it the functor $\Hom_A(\Hom_A(C,J),-)$ for any injective $A$–module $J$.
$M$ is called *$C$–Gorenstein projective* if:
1. $\Ext_A^{\geqslant 1}(M,C\ten_AP)=0$ for all projective $A$–modules $P$.
2. There exist projective $A$–modules $P^0,P^1,\ldots$ together with an exact sequence: $$0 \to M \to C\ten_AP^0 \to C\ten_AP^1 \to \cdots,$$ and furthermore, this sequence stays exact when we apply to it the functor for any projective $A$–module $Q$.
Finally, $M$ is called *$C$–Gorenstein flat* if:
1. $\Tor^A_{\geqslant 1}(\Hom_A(C,I),M)=0$ for all injective $A$–modules $I$.
2. There exist flat $A$–modules $F^0,F^1,\ldots$ together with an exact sequence: $$0 \to M \to C\ten_AF^0 \to C\ten_AF^1 \to \cdots,$$ and furthermore, this sequence stays exact when we apply to it the functor $\Hom_A(C,I)\ten_A-$ for any injective $A$–module $I$.
\[exa:injective\_is\_CGinjective\] (a) If $I$ is an injective $A$–module, then $\Hom_A(C,I)$ and $I$ are $C$–Gorenstein injective because:
It is easy to see that $\Hom_A(C,I)$ is $C$–Gorenstein injective. Concerning $I$ itself it is clear that condition (I1) of Definition \[def:Cmodules\] is satisfied. From Lemma \[lem:Proj/A=>GorProj/AxC\] it follows that $I$ is Gorenstein injective over $A \skewtimes C$, so iterating Lemma \[lem:CGorInj/A<=GorInj/AxC\] we also get condition (I2).
\(b) Similarly, if $P$ is a projective $A$–module, then $C\ten_AP$ and $P$ are $C$–Gorenstein projective. The last claim uses Lemmas \[lem:Proj/A=>GorProj/AxC\] and \[lem:CGorProj/A<=GorProj/AxC\].
\(c) If $F$ is a flat $A$–module, then $C\ten_AF$ and $F$ are $C$–Gorenstein flat. The last claim uses (a) together with Propositions \[prop:dual\], \[prp:inj\_proj\_module\_case\](1), \[prp:flat\_module\_case\] (the last two can be found below).
\[def:new\_dimensions\] By Example \[exa:injective\_is\_CGinjective\](a), there exists for every homologically left-bounded complex $N$ a left-bounded complex $Y$ of $C$–Gorenstein injective modules with $Y \simeq N$ in $\sD(A)$ (as one could take $Y$ to be an injective resolution of $N$). Every such $Y$ is called a *$C$–Gorenstein injective resolution* of $N$.
*$C$–Gorenstein projective* and *$C$–Gorenstein flat resolutions* of homologically right-bounded complexes are defined in a similar way, and they always exist by Examples \[exa:injective\_is\_CGinjective\](b) and (c). Thus, we may define:
For any homologically left-bounded $A$–complex $N$ we introduce: $$\GIdim N \,=\, \underset{Y}{\inf} \Big( \sup
\big\{n\in\BZ \,|\, Y_{-n} \neq 0 \big\} \Big),$$ where the infimum is taken over all $C$–Gorenstein injective resolutions $Y$ of $N$. For a homologically right-bounded $A$–complex $M$ we define: $$\GPdim M \,=\, \underset{X}{\inf} \Big( \sup
\big\{n\in\BZ \,|\, X_n \neq 0 \big\} \Big),$$ where the infimum is taken over all $C$–Gorenstein projective resolutions $X$ of $M$. Finally, we define $\GFdim M$ anologously to $\GPdim M$.
Note that when $C=A$ in Definition \[def:Cmodules\], we recover the categories of ordinary Gorenstein injective, Gorenstein projective, and Gorenstein flat $A$–modules.
Thus, $\AGIdim(-)$, $\AGPdim(-)$, and $\AGFdim(-)$ are the usual Gorenstein injective, Gorenstein projective and Gorenstein flat dimensions over $A$, which one usually denotes $\Gid_A(-)$, $\Gpd_A(-)$ and $\Gfd_A(-)$, respectively.
\[lem:CGorInj/A=>GorInj/AxC\] Let $M$ be an $A$–module which is $C$–Gorenstein injective. Then there exists a short exact sequence of $(A \skewtimes C)$–modules, $$0 \rightarrow M^{\prime} \longrightarrow U \longrightarrow M
\rightarrow 0,$$ where $U$ is injective over $A \skewtimes C$ and $M^{\prime}$ is $C$–Gorenstein injective over $A$. Furthermore, the sequence stays exact if one applies to it the functor $\Hom_{A \skewtimes C}(V,-)$ for any injective $(A \skewtimes C)$–module $V$.
Since $M$ is $C$–Gorenstein injective, we in particular get a short exact sequence of $A$–modules: $$0 \rightarrow N \longrightarrow \Hom_A(C,I) \longrightarrow M
\rightarrow 0,$$ where $I$ is injective and $N$ is $C$–Gorenstein injective, which stays exact under $\Hom_A(\Hom_A(C,J),-)$ when $J$ is injective. Applying the functor $\Hom_A(-,I)$ to the exact sequence: $$ \tag{\text{$*$}}
0 \rightarrow C \longrightarrow A \skewtimes C \longrightarrow A
\rightarrow 0$$ gives an exact sequence of $(A \skewtimes C)$–modules: $$ \tag{\text{$**$}}
0 \rightarrow I \longrightarrow \Hom_A(A \skewtimes C,I)
\longrightarrow \Hom_A(C,I) \rightarrow 0.$$ If viewed as a sequence of $A$–modules then this is split, because the same holds for $(*)$. Combining these data gives a commutative diagram of $(A \skewtimes C)$–modules with exact rows: $$\xymatrix{0 \ar[r] & M' \ar[r] \ar[d] & \Hom_A(A \ltimes
C,I) \ar[r] \ar@{->>}[d] & M \ar[r] \ar@{=}[d] & 0 \\
0 \ar[r] & N \ar[r] & \Hom_A(C,I) \ar[r] & M \ar[r] & 0. }$$ We will prove that the upper row here has the properties claimed in the lemma:
First, $\Hom_A(A \skewtimes C,I)$ is an injective $(A \skewtimes
C)$–module by Lemma \[lem:induced\_injectives\](1). Secondly, using the Snake Lemma on the diagram embeds the vertical arrows into exact sequences. The leftmost of these is: $$0 \rightarrow I \longrightarrow M^{\prime} \longrightarrow N
\rightarrow 0,$$ proving that as $A$–modules, $M^{\prime} \cong I \oplus N$. Here $N$ is $C$–Gorenstein injective by construction, and $I$ is by Example \[exa:injective\_is\_CGinjective\](a). So $M^{\prime}$ is clearly also $C$–Gorenstein injective.
Finally, by construction, the lower row in the diagram stays exact under $\Hom_A(\Hom_A(C,J),-)$ when $J$ is injective. If viewed as a sequence of $A$–modules then the sequence $(**)$ is split, so the surjection $\Hom_A(A \skewtimes C,I) \longrightarrow \Hom_A(C,I)$ is split, and therefore the upper row in the diagram also stays exact under $\Hom_A(\Hom_A(C,J),-)$.
By aplying $\H_0(-)$ to Lemma \[lem:calculation\](1), we see that the upper row in the diagram stays exact under $\Hom_{A \skewtimes
C}(\Hom_A(A \skewtimes C,J),-)$ when $J$ is an injective $A$–module. Thus, it also stays exact under $\Hom_{A \skewtimes
C}(V,-)$ for any injective $(A \skewtimes C)$–module $V$, because of Lemma \[lem:induced\_injectives\](2).
By a similar argument we get:
\[lem:CGorProj/A=>GorProj/AxC\] Let $M$ be an $A$–module which is $C$–Gorenstein projective. Then there exists a short exact sequence of $(A \skewtimes C)$–modules, $$0 \rightarrow M \longrightarrow R \longrightarrow M^{\prime}
\rightarrow 0,$$ where $R$ is projective over $A \skewtimes C$ and $M^{\prime}$ is $C$–Gorenstein projective over $A$. Furthermore, the sequence stays exact if one applies to it the functor $\Hom_{A \skewtimes C}(-,S)$ for any projective $(A \skewtimes C)$–module $S$.
\[prp:inj\_proj\_module\_case\] For any $A$–module $M$ the two conclusions hold:
1. $M$ is $C$–Gorenstein injective if and only if $M$ is Gorenstein injective over $A \ltimes C$.
2. $M$ is $C$–Gorenstein projective if and only if $M$ is Gorenstein projective over $A \ltimes C$.
\(1) If $M$ is $C$–Gorenstein injective, then Lemma \[lem:CGorInj/A=>GorInj/AxC\] gives the “left half” of a complete injective resolution of $M$ over $A \ltimes C$.
Conversely, if $M$ is Gorenstein injective over $A \ltimes C$, then Lemma \[lem:CGorInj/A<=GorInj/AxC\] gives the existence of a sequence like the one in Definition \[def:Cmodules\] (I2). Now, to finish the proof we only need to refer to Corollary \[cor:Ext\](1).
\(2) Similar, but using Lemmas \[lem:CGorProj/A=>GorProj/AxC\], \[lem:CGorProj/A<=GorProj/AxC\] and Corollary \[cor:Ext\](2).
Before turning to $C$–Gorenstein flat modules, we need to recall the notion of *Kaplansky classes* from [@EnochsLopezRamos Definition 2.1], which is reformulated in Definition \[def:Kaplansky\], Section \[sec:proper\_dimensions\]. The following lemma will be central:
\[lem:C\_tensor\_flad\_Kaplansky\] The class ${\sf F} = \{ C\ten_AF \,|\ F \,\text{flat }
A\text{--module}\}$ is Kaplansky, and furthermore it is closed under direct limits.
Every homomorphism $\varphi \colon C\ten_AF_1 \to C\ten_AF_2$, where $F_i$ is flat, has the form $\varphi = C\ten_A\psi$ for some homomorphism $\psi \colon F_1 \to F_2$; namely $\psi =
\Hom_A(C,\varphi)$, because $\Hom_A(C,C\ten_AF_i) \cong F_i$.
With this observation in mind it is clear that $\sF$ is closed under direct limits, since the class of flat modules has this property.
To see that $\sF$ is Kaplansky, we first note that a finitely generated $A$–module has cardinality at most $\kappa =
\max\{|A|,\aleph_0\}$.
Now, assume that $x$ is an element of $G = C\ten_AF$, where $F$ is a flat $A$–module. Write $x=\sum_{i=1}^n c_i\ten x_i$ for some $c_1,\ldots,c_n \in C$ and $x_1,\ldots,x_n \in F$. Let $S$ be the $A$–submodule of $F$ generated by $x_1,\ldots,x_n$, and then use [@Xu Lemma 2.5.2] (or [@EnochsJendabook Lemma 5.3.12]) to enlarge $S$ to a pure submodule $F'$ in $F$ with cardinality: $$|F'| \,\leqslant\, \max\{|S|\!\cdot\!|A|,\aleph_0\} \,\leqslant\,
\kappa.$$ Since $F$ is flat and $F' \subseteq F$ is a pure submodule, then $F'$ and $F/F'$ are flat as well. Furthermore, exactness of: $$0 \to C\ten_AF' \to C\ten_AF \to C\ten_A(F/F') \to 0$$ shows that $G' = C\ten_AF'$ is a submodule of $G = C\ten_AF$ which contains $x$. Clearly, $G'$ and $G/G' \cong C\ten_A(F/F')$ belong to ${\sf F}$, and: $$|G'| \,=\, |C\ten_AF'| \,\leqslant\, |\mathbb{Z}^{(C \times F')}|
\,\leqslant\, |(2^{\mathbb{Z}})^{(C
\times F')}| \,=\, |2^{(\mathbb{Z} \times C \times F')}|
\,\leqslant\, 2^{\kappa}.$$ The last inequality comes from the fact that all three cardinal numbers $|\mathbb{Z}|$, $|C|$ and $|F'|$ are less than $\kappa$. Note that the cardinal number $2^{\kappa}$ only depends on the ring $A$.
The next proof is modelled on that of [@Wintherbook Theorem (6.4.2)].
\[prp:flat\_module\_case\] Let $M$ be an $A$–module. Then $M$ is $\,C$–Gorenstein flat if and only if $M$ is Gorenstein flat over $A \ltimes C$. In the affirmative case, $M$ has the next property, which implies Definition \[def:Cmodules\] [(F2)]{.nodecor}:
1. There exist flat $A$–modules $F^0,F^1,\ldots$ together with an exact sequence: $$0 \to M \to C\ten_AF^0 \to C\ten_AF^1 \to \cdots,$$ and furthermore, this sequence stays exact when we apply to it the functor $\Hom_A(-,C\ten_AG)$ for any flat $A$–module $G$.
For the first statement, it suffices by Propositions \[prop:dual\] and \[prp:inj\_proj\_module\_case\](1) to show that if $E$ is a faithfully injective $A$–module, then: $$\begin{aligned}
\text{$M$ is $C$--Gorenstein flat $\,\Leftrightarrow\,$
$\Hom_A(M,E)$ is $C$--Gorenstein injective.}
\end{aligned}$$ For any injective $A$–module $I$ we have (adjointness) isomorphisms: $$\begin{aligned}
\Ext_A^i(\Hom_A(C,I),\Hom_A(M,E)) \,\cong\, \\
\Hom_A(\Tor^A_i(\Hom_A(C,I),M),E).
\end{aligned}$$ Thus, Definition \[def:Cmodules\] (F1) for $M$ is equivalent to (I1) for $\Hom_A(M,E)$. The rest of the proof will concern the conditions (F2) for $M$ and (I2) for $\Hom_A(M,E)$ in Definition \[def:Cmodules\].
If $\,\mathbb{S} = 0 \to M \to C\ten_AF^0 \to C\ten_AF^1 \to \cdots$ is a sequence for $M$ like the one in Definition \[def:Cmodules\] (F2), then, using adjointness, it is easy to see that $\Hom_A(\mathbb{S},E)$ is a sequence for $\Hom_A(M,E)$ like the one in (I2). Therefore, we have proved the implication “$\Rightarrow$”
To show “$\Leftarrow$”, we assume that $\Hom_A(M,E)$ is $C$–Gorenstein injective. As already noted, we only have to show Definition \[def:Cmodules\] (F2) for $M$. First note that (F2’) really implies Definition \[def:Cmodules\] (F2), since: $$\begin{aligned}
\Hom_A(\Hom_A(C,I)\ten_A-,E) &\,\simeq\,
\Hom_A(-,\Hom_A(\Hom_A(C,I),E)) \\
&\,\simeq\, \Hom_A(-,C\ten_A\Hom_A(I,E)),
\end{aligned}$$ and when $I$ is injective, then $G=\Hom_A(I,E)$ is flat. In order prove (F2’), it suffices to show the existence of a short exact sequence: $$\tag{\text{$\dagger$}}
0 \to M \to C\ten_AF \to M' \to 0,$$ satisfying the following three conditions:
1. $F$ is flat,
2. $\Hom_A(M',E)$ is $C$–Gorenstein injective,
3. is exact for any flat $A$–module $G$.
Because then one obtains the sequence in (F2’) by iterating $(\dagger)$. By Lemma \[lem:C\_tensor\_flad\_Kaplansky\], the class of $A$–modules: $${\sf F} \,=\, \big\{ C\ten_AF \,|\ F \text{ flat }
A\text{--module} \big\}.$$ is Kaplansky. Furthermore, it is closed under arbitrary direct products; since $C$ is finitely generated and $A$ is noetherian, and hence [@EnochsLopezRamos Theorem 2.5] implies that every $A$–module has an ${\sf F}$–preenvelope.
Note that since $\Hom_A(M,E)$ is $C$–Gorenstein injective, there in particular exists an epimorphism $\Hom_A(C,I) \twoheadrightarrow
\Hom_A(M,E)$, where $I$ is injective. Applying $\Hom_A(-,E)$, we get a monomorphism: $$\begin{aligned}
M &\,\hookrightarrow\, \Hom_A(\Hom_A(M,E),E) \\
&\,\hookrightarrow\, \Hom_A(\Hom_A(C,I),E) \,\cong\,
C\ten_A\Hom_A(I,E) \,\in\, {\sf F}.
\end{aligned}$$ Thus, $M$ can be embedded into a module from ${\sf F}$. Therefore, taking an ${\sf F}$–preenvelope $\varphi \colon M \to C\ten_AF$ of $M$, it is automaticly injective; and defining $M'=\text{Coker}\,\varphi$, we certainly get an exact sequence $(\dagger)$ satisfying (1) and (3).
Finally, we argue that (2) is true. Keeping Proposition \[prp:inj\_proj\_module\_case\](1) in mind we must prove that $\Hom_A(M',E)$ is Gorenstein injective over $A \ltimes C$. Applying $\Hom_A(-,E)$ to $(\dagger)$ we get: $$\tag{\text{$\ddagger$}}
0 \to \Hom_A(M',E) \to \Hom_A(C,J) \to \Hom_A(M,E) \to 0,$$ where $J \cong \Hom_A(F,E)$ is injective. $\Hom_A(C,J)$ and $\Hom_A(M,E)$ are both Gorenstein injective over $A \ltimes C$ — the last module by assumption. Hence, if we can prove that $\Ext^1_{A\ltimes C}(U,\Hom_A(M',E))=0$ for every injective $(A
\ltimes C)$–module $U$, then [@EnochsJenda Theorem 2.13] gives the desired conclusion. Using Corollary \[cor:Ext\](1), we must prove that: $$\tag{\text{$\natural$}}
\Ext_A^1(\Hom_A(C,I),\Hom_A(M',E))=0$$ for all injective $A$–modules $I$. Consider the commutative diagram with exact columns: [ $$\xymatrix{0 & {} \\
\Ext^1_A(\Hom_A(C,I),\Hom_A(M',E)) \ar[u] & 0 \\
\Hom_A(\Hom_A(C,I),\Hom_A(M,E)) \ar[u] &
\Hom_A(\Hom_A(C,I)\ten_AM,E) \ar[l]_-{\cong} \ar[u] \\
\Hom_A(\Hom_A(C,I),\Hom_A(C,J)) \ar[u] &
\Hom_A(\Hom_A(C,I)\ten_A(C\ten_AF),E) \ar[l]_{\cong} \ar[u]
}$$ ]{} [The]{} first column is the induced long exact sequence which comes from applying $\Hom_A(\Hom_A(C,I),-)$ to $(\ddagger)$. We get another monomorphism when we apply $\Hom_A(C,I)\ten_A-$ to the one $0
\to M \to C\ten_AF$ from $(\dagger)$; this follows from the property (3) which $(\dagger)$ satisfies together with the calculation preceding $(\dagger)$. Turning this into an epimorphism with $\Hom_A(-,E)$ we get the second column. The vertical isomorphisms are by adjointness. The diagram implies that the module in $(\natural)$ is zero.
\[thm:main\_theorem\] For any (appropriately homologically bounded) $A$–complex $M$, we have the following equalities: $$\begin{aligned}
\GIdim M & = & \CGid M, \\
\GPdim M & = & \CGpd M, \\
\GFdim M & = & \CGfd M.
\end{aligned}$$
The proof uses Propositions \[prp:inj\_proj\_module\_case\](1),(2) and \[prp:flat\_module\_case\] in combination with [@CFH Theorems (2.5), (2.2) and (2.8)]. We only prove that $\GIdim M = \CGid
M$, since the proofs of the other two equalities are similar:
From Proposition \[prp:inj\_proj\_module\_case\](1) we get that every $C$–Gorenstein injective $A$–module is also Gorenstein injective over $A \ltimes C$, and this give us the inequality “$\geqslant$”.
For the opposite inequality “$\leqslant$”, we may assume that $n=\CGid M$ is finite. Pick a left-bounded complex $I$ of injective $A$–modules such that $I \simeq M$ in $\sD(A)$. By Lemma \[lem:Proj/A=>GorProj/AxC\] the modules $I_i$ are Gorenstein injective over $A \ltimes C$, and therefore [@CFH Theorem (2.5)] implies that the $A$–module $\textnormal{Z}^I_{-n}$ is Gorenstein injective over $A \ltimes C$.
Now, Proposition \[prp:inj\_proj\_module\_case\](1) shows that $\textnormal{Z}^I_{-n}$ is $C$–Gorenstein injective. By Example \[exa:injective\_is\_CGinjective\](a), the complex consists of $C$–Gorenstein injective $A$–modules, and since we see that $\GIdim M \leqslant n$.
\[cor:C=A\] For any (appropriately homologically bounded) $A$–complex $M$, we have the following equalities: $$\begin{array}{lclcl}
\AGid M & = & \Gid_{A[x]/(x^2)}M & = & \Gid_AM, \\
\AGpd M & = & \Gpd_{A[x]/(x^2)}M & = & \Gpd_AM, \\
\AGfd M & = & \Gfd_{A[x]/(x^2)}M & = & \Gfd_AM.
\end{array}$$
This follows immediately from Theorem \[thm:main\_theorem\]; we only have to note that $A \ltimes A \cong A[x]/(x^2)$ (sometimes refered to as the *dual numbers* over $A$).
Having realized that, on the level of $A$–complexes, the three (classical) Gorenstein dimensions can not distinguish between $A$ and $A \ltimes A$, we can reap a nice result from the work of [@HHPJ2]:
\[thm:GorThm\] If $(A,\mathfrak{m},k)$ is local, then the following conditions are equivalent:
1. $A$ is Gorenstein.
2. There exists an $A$–complex $M$ such that all three numbers $\,\fd_AM$, $\Gid_AM$ and $\textnormal{width}_AM$ are finite.
3. There exists an $A$–complex $N$ such that all three numbers $\,\id_AN$, $\Gpd_AN$ and $\textnormal{depth}_AN$ are finite.
4. There exists an $A$–complex $N$ such that all three numbers $\,\id_AN$, $\Gfd_AN$ and $\textnormal{depth}_AN$ are finite.
It is well-known that over a Gorenstein ring, every homologically bounded complex has finite Gorenstein injective, Gorenstein projective and Gorenstein flat dimension, and thus $(1) \Rightarrow
(2), (3), (4)$.
Of course, $(3) \Rightarrow (4)$; and using Corollary \[cor:C=A\], the remaining implications $(2) \Rightarrow (1)$ and $(4)
\Rightarrow (1)$ follow immediately from [@HHPJ2 Propositions 4.5 and 4.7].
There already exist special cases of this result in the litterature: If $A$ admits a dualizing complex, then [@Wintherbook (3.3.5)] compared with [@CFH Theorems (4.3) and (4.5)] gives Theorem \[thm:GorThm\]. If one drops the assumption that a dualizing complex should exists, then Theorem \[thm:GorThm\] is proved in [@HHRWFGID Corollary (3.3)], but only for modules.
Comparison with Christensen’s $\GdimC(-)$ {#sec:Christensen}
=========================================
In [@Wintherpaper Definition (3.11)], Christensen introduced the number $\GdimC Z$ for any semi-dualizing complex $C$, and any complex $Z$ with bounded and finitely generated homology. When $C = A$ (and $Z$ is a module), we recover Auslander–Bridger’s $G$–dimension by [@Wintherbook Theorem (2.2.3)].
\[pro:GdimC\] If $\,C$ is a semi-dualizing $A$–module, and $M$ an $A$–complex with bounded and finitely generated homology, then: $$\GPdim M \,=\, \GdimC M.$$
By Theorem \[thm:main\_theorem\], the proposition amounts to: $$\tag{\text{$*$}}
\Gpd_{A \skewtimes C} M \,=\, \GdimC M.$$ The homology of $M$ is bounded and finitely generated over $A$, and hence it is also bounded and finitely generated over $A \skewtimes
C$. So by e.g. [@CFH Theorem (2.12)(b)] or [@Wintherbook Theorem (4.2.6)], the left hand side in $(*)$ equals $\GdimAxC
M$ (Auslander–Bridger’s $G$–dimension over the ring $A \ltimes
C$). We must therefore prove that: $$\tag{\text{$**$}}
\GdimAxC M \,=\, \GdimC M.$$ The left hand side is finite precisely if the biduality morphism: $$M \longrightarrow \RHom_{A \skewtimes C}(
\RHom_{A \skewtimes C}(M,A \skewtimes C),A \skewtimes C)$$ is an isomorphism, and the right hand side is finite precisely when $$M \longrightarrow \RHom_A(\RHom_A(M,C),C)$$ is an isomorphism. But these two morphisms are equal by Lemma \[lem:extension\_formulae\](3), so the left hand side and right hand side of $(**)$ are simultaneously finite. When the left hand side of $(**)$ is finite, it equals: $$-\inf \RHom_{A \skewtimes C}(M,A \skewtimes C),$$ and when the right hand side is finite, it is equal to: $$-\inf \RHom_A(M,C)$$ But these two numbers are equal by Lemma \[lem:extension\_formulae\](2).
Christensen’s $\GdimC(-)$ only works when the argument has bounded and finitely generated homology, but it has the advantage that $C$ is allowed to be a semi-dualizing *complex*.
By Theorem \[thm:main\_theorem\], we get that for $A$–complexes $M$, the $C$–Gorenstein projective dimension $\GPdim M$ agrees with the “ring changed” Gorenstein projective dimension $\CGpd M$.
It is not immediately clear how one should make either of these dimensions work when $C$ is a semi-dualizing *complex*. Because in this case, $A \ltimes C$ becomes a differential graded algebra, and the $C$–Gorenstein projective objects in Definition \[def:Cmodules\] (from which we build our resolutions) become complexes.
In [@Wintherthesis Page 28] we find an interesting comment, which makes it even more clear why we run into trouble when $C$ is a complex:
“*On the other hand, let $C$ be a semi-dualizing complex with $\textnormal{amp}\,C = s >0$. We are free to assume that $\inf C = 0$, and it is then immediate from the definition that $\GdimC C = 0$; but a resolution of $C$ must have length at least $s$, so the $G$–dimension with respect to $C$ can not be interpreted in terms of resolutions.*”
It is notable that the number $\Gpd_A \RHom_A(C,N)$, $N \in
\sB_C(A)$, occuring in Theorem \[thm:CGpd\_and\_BC\] below makes perfect sense even if $C$ is a complex.
Interpretations via Auslander and Bass categories {#sec:AB}
=================================================
In this section, we interpret the $C$–Gorenstein homological dimensions from Section \[sec:CGdims\] in terms of Auslander and Bass categories.
\[rmk:AB\] Let $C$ be a semi-dualizing $A$–complex. In [@Wintherpaper Section 4] is considered the adjoint pair of functors: $$\xymatrix{ \sD(A) \ar@<0.5ex>[rrr]^-{C\LTensor_A-} & {} & {} &
\sD(A) \ar@<0.5ex>[lll]^-{\RHom_A(C,-)} }$$ and the full subcategories (where $\sD_{\b}(A)$ is the full subcategory of $\sD(A)$ consisting of homologically bounded complexes): [ $$\sA_C(A) = \left\{ M \in \sD(A) \:
\left|
\begin{array}{l}
\mbox{$M$ and $C \LTensor_A M$ are in $\sD_{\b}(A)$ and} \\
\mbox{$M \rightarrow \RHom_A(C,C \LTensor_A M)$ is an
isomorphism}
\end{array}
\right. \!\!
\right\}$$ and $$\sB_C(A) = \left\{ N \in \sD(A) \:
\left|
\begin{array}{l}
\mbox{$N$ and $\RHom_A(C,N)$ are in $\sD_{\b}(A)$ and} \\
\mbox{$C \LTensor_A \RHom_A(C,N) \rightarrow N$ is an
isomorphism}
\end{array}
\right. \!\!
\right\}.$$ ]{} [It]{} is an exercise in adjoint functors that the adjoint pair above restricts to a pair of quasi-inverse equivalences of categories: $$\xymatrix{ \sA_C(A) \ar@<0.5ex>[rrr]^-{C\LTensor_A-} & {} & {} &
\sB_C(A). \ar@<0.5ex>[lll]^-{\RHom_A(C,-)} }$$
\[thm:CGid\_and\_AC\] For any complex $M \in \sA_C(A)$ we have an equality: $$\GIdim M \,=\, \Gid_A(C\LTensor_AM).$$
Throughout the proof we make use of the nice desciptions of the *modules* in $\sA_C(A)$ and $\sB_C(A)$ from [@Wintherpaper Observation (4.10)].
<span style="font-variant:small-caps;">Step</span> 1: In order to prove the equality $\GIdim M =
\Gid_A(C\LTensor_AM)$, we first justify the (necessary) bi-implication: $$\begin{aligned}
\tag{\text{$\natural$}}
&M \textnormal{ is } C\textnormal{--Gorenstein injective} \quad
\iff \quad \\ &C\ten_AM \textnormal{ is Gorenstein injective}
\end{aligned}$$ for any *module* $M \in \sA_C(A)$.
“$\Rightarrow$”: By Definition \[def:Cmodules\](I2) there is an exact sequence: $$\tag{\text{$*$}}
\cdots \to \Hom_A(C,I_1) \to \Hom_A(C,I_0) \to M \to 0,$$ where $I_0,I_1,\ldots$ are injective $A$–modules. Furthermore, we have exactness of $\Hom_A(\Hom_A(C,J),(*))$ for all injective $A$–modules $J$.
$M$ belongs to $\sA_C(A)$, and so does $\Hom_A(C,I)$ for any injective $A$–module $I$, since $I \in \sB_C(A)$ by [@Wintherpaper Proposition (4.4)]. In particular, $C$ is Tor-independent with both of the modules $M$ and $\Hom_A(C,I)$ (two $A$–modules $U$ and $V$ are Tor-independent if $\Tor_{\geqslant
1}^A(U,V)=0$). Hence the sequence $(*)$ stays exact if we apply to it the functor $C\ten_A-$, and doing so we obtain: $$\tag{\text{$**$}}
\cdots \to I_1 \to I_0 \to C\ten_AM \to 0.$$ By similar arguments we see that if we apply $\Hom_A(C,-)$ to the sequence $(**)$, then we get $(*)$ back. If $J$ is any injective $A$–module, then we have exactness of $\Hom_A(J,(**))$ because: $$\begin{aligned}
\Hom_A(J,(**)) &\,\cong\, \Hom_A(C\ten_A\Hom_A(C,J),(**)) \\
&\,\cong\, \Hom_A(\Hom_A(C,J),\Hom_A(C,(**))) \\
&\,\cong\, \Hom_A(\Hom_A(C,J),(*)).
\end{aligned}$$ Thus, $(**)$ is a “left half” of a complete injective resolution of the $A$–module $C\ten_AM$. We also claim that $\Ext^i_A(J,C\ten_AM)=0$ for all $i>0$ and all injective $A$–modules $J$. First note that: $$\begin{aligned}
\tag{\text{$\diamond$}}
\Ext^i_A(J,C\ten_AM) &\,\stackrel{\rm (a)}{=}\,
\H^i\RHom_A(C\LTensor_A\RHom_A(C,J),C\LTensor_AM) \\
&\,\stackrel{\rm (b)}{\cong}\,
\H^i\RHom_A(\RHom_A(C,J),\RHom_A(C,C\LTensor_AM)) \\
&\,\stackrel{\rm (c)}{\cong}\, \H^i\RHom_A(\RHom_A(C,J),M) \\
&\,\cong\, \Ext_A^i(\Hom_A(C,J),M).
\end{aligned}$$ Here (a) is follows as $J \in \sB_C(A)$ by [@Wintherpaper Proposition (4.4)]; (b) is by adjointness; and (c) is because $M
\in \sA_C(A)$. This last module is zero because $M$ is $C$–Gorenstein injective. These considerations prove that $C\ten_AM$ is Gorenstein injective over $A$.
“$\Leftarrow$”: If $C\ten_AM$ is Gorenstein injective over $A$, we have by definition an exact sequence: $$\tag{\text{$\dagger$}}
\cdots \to I_1 \to I_0 \to C\ten_AM \to 0,$$ where $I_0,I_1,\ldots$ are injective $A$–modules. Furthermore, we have exactness of $\Hom_A(J,(\dagger))$ for all injective $A$–modules $J$.
Since $I_0,I_1,\ldots$ and $C\ten_AM$ are modules from $\sB_C(A)$, then so are all the kernels in $(\dagger)$, as $\sB_C(A)$ is a triangulated subcategory of $\sD(A)$. If $N \in \sB_C(A)$, then $\Ext^{\geqslant 1}_A(C,N)=0$, and consequently, the sequence $(\dagger)$ stays exact if we apply to it the functor $\Hom_A(C,-)$. Doing so we obtain: $$\tag{\text{$\ddagger$}}
\cdots \to \Hom_A(C,I_1) \to \Hom_A(C,I_0) \to M \to 0.$$ If $J$ is any injective $A$–module, then we have exactness of the complex $\Hom_A(\Hom_A(C,J),(\ddagger))$ because: $$\begin{aligned}
\Hom_A(\Hom_A(C,J),(\ddagger)) &\,\cong\,
\Hom_A(\Hom_A(C,J),\Hom_A(C,(\dagger))) \\
&\,\cong\, \Hom_A(C\ten_A\Hom_A(C,J),(\dagger)) \\
&\,\cong\, \Hom_A(J,(\dagger)).
\end{aligned}$$ Furthermore, $(\diamond)$ above gives that: $$\begin{aligned}
\Ext_A^{\geqslant 1}(\Hom_A(C,J),M) \,\cong\, \Ext^{\geqslant
1}_A(J,C\ten_AM) \,=\, 0,
\end{aligned}$$ for all injective $A$–modules $J$. The last zero is because $C\ten_AM$ is Gorenstein injective over $A$. Hence $M$ is $C$–Gorenstein injective.
<span style="font-variant:small-caps;">Step</span> 2: To prove the inequality $\GIdim M \geqslant
\Gid_A(C\LTensor_AM)$ for any complex $M \in \sA_C(A)$, we may assume that $m=\GIdim M= \CGid M$; cf. Theorem \[thm:main\_theorem\], is finite. Since $C\LTensor_AM$ is homologically bounded, there exists a left-bounded injective resolution $I$ of $C\LTensor_AM$, that is, $I \simeq C\LTensor_AM$ in $\sD(A)$.
We wish to prove that the $A$–module $\textnormal{Z}^I_{-m}$ is Gorenstein injective. Since $M$ belongs to $\sA_C(A)$, we get isomorphisms: $$M \,\simeq\, \RHom_A(C,C\LTensor_AM) \,\simeq\, \RHom_A(C,I)
\,\simeq\, \Hom_A(C,I).$$ Now, $\Hom_A(C,I)$ is a complex of Gorenstein injective $A \ltimes
C$–modules, and thus the $A$–module $L :=
\textnormal{Z}_{-m}^{\Hom_A(C,I)}$ is Gorenstein injective over $A
\ltimes C$ by [@CFH Theorem (2.5)]. By Proposition \[prp:inj\_proj\_module\_case\](1), $L$ is also $C$–Gorenstein injective. Note that: $$-m \,=\, -\Gid_{A \ltimes C}M \,\leqslant\, \inf M
\,\stackrel{\rm (a)}{=}\, \inf(C\LTensor_AM) \,=\, \inf I,$$ where the equality (a) comes from [@Wintherpaper Lemma(4.11)(b)]. Therefore, $0 \to
\textnormal{Z}^I_{-m} \to I_{-m} \to I_{-m-1}$ is exact, and applying the left exact functor $\Hom_A(C,-)$ to this sequence we get an isomorphism of $A$–modules: $$\tag{\text{$\flat$}}
L \,=\, \textnormal{Z}_{-m}^{\Hom_A(C,I)} \,\cong\,
\Hom_A(C,\textnormal{Z}_{-m}^I).$$ We have a degreewise split exact sequence of complexes: $$0 \rightarrow \Sigma^{-m}\textnormal{Z}^I_{-m} \longrightarrow
I_{-m}\!\!\supset \ \longrightarrow I_{-m+1}\!\!\sqsupset
\ \longrightarrow 0,$$ where we have used the notation from [@Wintherbook Appendix (A.1.14)] to denote soft and hard truncations. Since has finite injective dimension it belongs to $\sB_C(A)$ by [@Wintherpaper Proposition (4.4)], and furthermore, $$I_{-m}\!\!\supset \ \, \simeq\, I \,\simeq\,
C\LTensor_AM \,\in\, \sB_C(A).$$ Thus, the module $\textnormal{Z}^I_{-m}$ is also in $\sB_C(A)$, as $\sB_C(A)$ is a triangulated subcategory of $\sD(A)$. Consequently, the module $L$ from $(\flat)$ belongs to $\sA_C(A)$ and has the property that $C\otimes_AL \cong \textnormal{Z}^I_{-m}$. Therefore, the implication “$\Rightarrow$” in $(\natural)$ gives that $\textnormal{Z}^I_{-m}$ is Gorenstein injective over $A$, as desired.
<span style="font-variant:small-caps;">Step</span> 3: To prove the opposite inequality $\GIdim M
\leqslant \Gid_A(C\LTensor_AM)$ for any complex $M \in \sA_C(A)$, we assume that $n= \Gid_A(C\LTensor_AM)$ is finite. Pick any left-bounded injective resolution $I$ of $C\LTensor_AM$. Then the $A$–module $\textnormal{Z}^I_{-n}$ is Gorenstein injective by [@CFH Theorem (2.5)].
As in <span style="font-variant:small-caps;">Step</span> 2 we get $M \simeq \Hom_A(C,I)$, and thus it suffices to show that the module: $$N \,:=\, \textnormal{Z}_{-n}^{\Hom_A(C,I)} \,\cong\,
\Hom_A(C,\textnormal{Z}_{-n}^I).$$ is $C$–Gorenstein injective, because then shows that $\GIdim M \leqslant n$. As before we get that $N$ is a module in $\sA_C(A)$ with $C\ten_AN
\cong \textnormal{Z}_{-n}^I$, which this time is Gorenstein injective over $A$. Therefore, the implication “$\Leftarrow$” in $(\natural)$ gives that $N$ is $C$–Gorenstein injective.
Using Proposition \[prp:inj\_proj\_module\_case\](2), a similar argument gives:
\[thm:CGpd\_and\_BC\] For any complex $N \in \sB_C(A)$ we have an equality:
[3]{} & & N = \_A \_A(C,N). & &&
From Theorems \[thm:CGid\_and\_AC\] and \[thm:main\_theorem\], and Proposition \[prop:dual\] we can easily derive:
\[thm:CGfd\_and\_BC\] For any complex $N \in \sB_C(A)$ we have an equality: $$\GFdim N \,=\, \Gfd_A \RHom_A(C,N).$$
Let $E$ be a faithfully injective $A$–module. Since $N \in
\sB_C(A)$ it is easy to see that $\RHom_A(N,E) \simeq \Hom_A(N,E)$ is in $\sA_C(A)$. Hence: $$\begin{aligned}
\GFdim N &\,\stackrel{\rm (a)}{=}\, \GIdim \RHom_A(N,E) \\
&\,\stackrel{\rm (b)}{=}\, \Gid_A\big(C \LTensor_A
\RHom_A(N,E)\big) \\
&\,\stackrel{\rm (c)}{=}\, \Gid_A \RHom_A(\RHom_A(C,N),E) \\
&\,\stackrel{\rm (d)}{=}\, \Gfd_A \RHom_A(C,N).
\end{aligned}$$ Here (a) is by Proposition \[prop:dual\] and Theorem \[thm:main\_theorem\]; (b) is by Theorem \[thm:CGid\_and\_AC\]; (c) is by the isomorphism [@Wintherbook (A.4.24)]; and finally, (d) is by Proposition \[prop:dual\] and Corollary \[cor:C=A\].
In the rest of this section, we assume that $A$ admits a *dualizing complex* $D^A$; cf. [@CFH Definition (1.1)]. The canonical homomorphism of rings, $A \to A \skewtimes C$, turns $A
\skewtimes C$ into a finitely generated $A$–module, and thus $$D^{A \skewtimes C} \,=\, \RHom_A(A \skewtimes C,D^A)$$ is a dualizing complex for $A \skewtimes C$.
\[lem:extension\_formulae\_2\] There is an isomorphism over $A$, $$D^{A \skewtimes C} \LTensor_{A \skewtimes C} A \,\cong\,
\RHom_A(C,D^A).$$
This is a computation: $$\begin{aligned}
D^{A \skewtimes C} \LTensor_{A \skewtimes C} A
& = & \RHom_A(A \skewtimes C,D^A) \LTensor_{A \skewtimes C} A \\
& \stackrel{\rm (a)}{\cong}
& \RHom_A(\RHom_{A \skewtimes C}(A,A \skewtimes C),D^A) \\
& \stackrel{\rm (b)}{\cong}
& \RHom_A(C,D^A),
\end{aligned}$$ where (a) holds because $D^A$ has finite injective dimension over $A$ and (b) is by Lemma \[lem:extension\_formulae\](4).
By [@Wintherpaper Corollary (2.12)], the complex $C^{\dagger} =
\RHom_A(C,D^A)$ is semi-dualizing for $A$. We now have the following generalization of the main results in [@CFH Theorems (4.3) and (4.5)]:
\[thm:AB\] Let $M$ and $N$ be $A$–complexes such that the homology of $M$ is right-bounded and the homology of $N$ is left-bounded. Then:
1. $M \in \sA_{C^{\dagger}}(A)$ $\iff$ $\GPdim M < \infty$ $\iff$ $\GFdim M < \infty$.
2. $N \in \sB_{C^{\dagger}}(A)$ $\iff$ $\GIdim N < \infty$.
Recall that $D^{A \skewtimes C} = \RHom_A(A \skewtimes C,D^A)$ is a dualizing complex for $A \skewtimes C$. If $M$ is a complex of $A$–modules then $$\begin{aligned}
C^{\dagger} \LTensor_A M
& = & \RHom_A(C,D^A) \LTensor_A M \\
& \stackrel{\rm (a)}{\cong}
& (D^{A \skewtimes C} \LTensor_{A \skewtimes C} A) \LTensor_A M \\
& \cong & D^{A \skewtimes C} \LTensor_{A \skewtimes C} M
\end{aligned}$$ and $$\begin{aligned}
\RHom_A(C^{\dagger},M)
& = & \RHom_A(\RHom_A(C,D^A),M) \\
& \stackrel{\rm (b)}{\cong}
& \RHom_A(D^{A \skewtimes C} \LTensor_{A \skewtimes C} A,M) \\
& \stackrel{\rm (c)}{\cong}
& \RHom_{A \skewtimes C}(D^{A \skewtimes C},\RHom_A(A,M)) \\
& \cong & \RHom_{A \skewtimes C}(D^{A \skewtimes C},M),
\end{aligned}$$ where (a) and (b) are by Lemma \[lem:extension\_formulae\_2\] and (c) is by adjunction. So using the adjoint pair: $$\xymatrix{ \sD(A) \ar@<0.5ex>[rrr]^-{C^{\dagger}\LTensor_A-} & {} &
{} & \sD(A) \ar@<0.5ex>[lll]^-{\RHom_A(C^{\dagger},-)} }$$ on complexes of $A$–modules is the same as viewing these complexes as complexes of $(A \skewtimes C)$–modules and using the adjoint pair: $$\xymatrix{ \sD(A \ltimes C) \ar@<0.5ex>[rrrr]^-{D^{A \skewtimes C}
\LTensor_{A \ltimes C}-} & {} & {} & {} & \sD(A \ltimes C)
\ar@<0.5ex>[llll]^-{\RHom_{A \skewtimes C}(D^{A \skewtimes C},-)}
}$$ Hence a complex $M$ of $A$–modules is in $\sA_{C^{\dagger}}(A)$ if and only if it is in $\sA_{D^{A \skewtimes C}}(A \ltimes C)$ when viewed as a complex of $(A \skewtimes C)$–modules. If $M$ has right-bounded homology, this is equivalent both to $\Gpd_{A
\skewtimes C}M < \infty$ and $\Gfd_{A \skewtimes C}M < \infty$ by [@CFH Theorem (4.3)], and by Theorem \[thm:main\_theorem\] this is the same as $\GPdim M < \infty$ and $\GFdim M < \infty$.
So part (1) of the theorem follows, and a similar method using [@CFH Theorem (4.5)] deals with part (2).
Proper dimensions {#sec:proper_dimensions}
=================
In this section, we define and study the *proper* variants of the dimensions from Theorem \[thm:main\_theorem\]. The results to follow depend highly on the work in [@EnochsLopezRamos].
In Definition \[def:new\_dimensions\] we introduced the dimensions $\GIdim(-)$, $\GPdim(-)$ and $\GFdim(-)$ for $A$–complexes. When $M$ is an $A$–module it is not hard to see that these dimensions specialize to: [$$\begin{aligned}
\GIdim M &\,=\, \inf \left\{ n \in \mathbb{N}_0
\left|
\begin{array}{l}
\mbox{$0 \to M \to I^0 \to \cdots \to I^n \to 0$ is exact} \\
\mbox{and $I^0,\ldots,I^n$ are $C$--Gorenstein injective}
\end{array}
\right. \!\!
\right\},\end{aligned}$$ ]{} [and]{} similarly for $\GPdim M$ and $\GFdim M$.
Let $\sQ$ be a class of $A$–modules (which contains the zero-module), and let $M$ be any $A$–module. A *proper left $\sQ$–resolution* of $M$ is a complex (not necessarily exact): $$\tag{\text{$\dagger$}}
\cdots \to Q_1 \to Q_0 \to M \to 0,$$ where $Q_0, Q_1, \ldots \in \sQ$ and such that $(\dagger)$ becomes exact when we apply to it the functor $\Hom_A(Q,-)$ for every $Q \in
\sQ$. A *proper right $\sQ$–resolution* of $M$ is a complex (not necessarily exact): $$\tag{\text{$\ddagger$}}
0 \to M \to Q^0 \to Q^1 \to \cdots,$$ where $Q^0, Q^1, \ldots \in \sQ$ and such that $(\ddagger)$ becomes exact when we apply to it the functor $\Hom_A(-,Q)$ for every $Q \in
\sQ$.
\[dfn:ProperDimensions\] Let $\sQ$ be a class of $A$–modules, and let $M$ be any $A$–module. If $M$ has a proper left $\sQ$–resolution, then we define the *proper left $\sQ$–dimension* of $M$ by: $$\cL\textnormal{-dim}_{\footnotesize \sQ}M \, = \, \inf \left\{ n
\in \mathbb{N}_0
\left|
\begin{array}{l}
\mbox{$0 \to Q_n \to \cdots \to Q_0 \to M \to 0$ is} \\
\mbox{a proper left $\sQ$--resolution of $M$}
\end{array}
\right. \!\!
\right\}.$$ Similarly, if $M$ has a proper right $\sQ$–resolution, then we define the *proper right $\sQ$–dimension* of $M$ by: $$\cR\textnormal{-dim}_{\footnotesize \sQ}M \, = \, \inf \left\{ n
\in \mathbb{N}_0
\left|
\begin{array}{l}
\mbox{$0 \to M \to Q^0 \to \cdots \to Q^n \to 0$ is} \\
\mbox{a proper right $\sQ$--resolution of $M$}
\end{array}
\right. \!\!
\right\}.$$
\[dfn:SpecialProperDimensions\] We use $\sG\sI_C(A)$, $\sG\sP_C(A)$ and $\sG\sF_C(A)$ to denote the classes of $C$–Gorenstein injective, $C$–Gorenstein projective and $C$–Gorenstein flat $A$–modules, respectively.
A proper right $\sG\sI_C(A)$–resolution is called a *proper $C$–Gorenstein injective resolution*, and a proper left $\sG\sP_C(A)$/$\sG\sF_C(A)$–resolution is called a *proper $C$–Gorenstein projective/flat resolution*.
Finally, we introduce the (more natural) notation:
1. $\ProperGIdim(-)$ for the proper right $\sG\sI_C(A)$–dimension,
2. $\ProperGPdim(-)$ for the proper left $\sG\sP_C(A)$–dimension,
3. $\ProperGFdim(-)$ for the proper left $\sG\sF_C(A)$–dimension,
whenever these dimensions are defined.
The next definition is taken directly from [@EnochsLopezRamos Definition 2.1]:
\[def:Kaplansky\] Let ${\sf F}$ be a class of $A$–modules. Then ${\sf F}$ is called *Kaplansky* if there exists a cardinal number $\kappa$ such that for every module $M \in {\sf F}$ and every element $x \in M$ there is a submodule $N \subseteq M$ satisfying $x \in N$ and $N,
M/N \in {\sf F}$ with $|N| \leqslant \kappa$.
\[lem:C-Gor-inj\_Kaplansky\] The class of $\,C$–Gorenstein injective $A$–modules is Kaplansky.
The class of Gorenstein injective $(A \ltimes C)$–modules is Kaplansky by [@EnochsLopezRamos Proposition 2.6]. Let $\kappa$ be a cardinal number which implements the Kaplansky property for this class.
Now assume that $M$ is a $C$–Gorenstein injective $A$–module, and that $x \in M$ is an element. By Proposition \[prp:inj\_proj\_module\_case\](1), $M$ is Gorenstein injective over $A \ltimes C$, and thus there exists a Gorenstein injective $(A
\ltimes C)$–submodule $N \subseteq M$ with $x \in N$ and $|N|
\leqslant \kappa$, and such that the $(A \ltimes C)$–module $M/N$ is Gorenstein injective.
Since $M$ is an $A$–module, when we consider it as a module over $A
\ltimes C$, it is annihilated by the ideal $C \subseteq A \ltimes
C$. Consequently, the two $(A \ltimes C)$–modules $N$ and $M/N$ are also annihilated by $C$. This means that $N$ and $M/N$ really are $A$–modules which are viewed as $(A \ltimes C)$–modules. Hence Proposition \[prp:inj\_proj\_module\_case\](1) implies that $N$ and $M/N$ are $C$–Gorenstein injective $A$–modules; and we are done.
\[thm:proper\_CGid\] Every $A$–module $M$ has a proper $C$–Gorenstein injective resolution, and we have an equality: $$\ProperGIdim M \,=\, \GIdim M.$$
By Lemma \[lem:C-Gor-inj\_Kaplansky\] above, the class of $C$–Gorenstein injective $A$–modules is Kaplansky, and it is obviously also closed under arbitrary direct products. Therefore, [@EnochsLopezRamos Theorem 2.5 and Remark 3] implies that every $A$–module admits a proper $C$–Gorenstein injective resolution.
Every injective $A$–module is also Gorenstein injective by Example \[exa:injective\_is\_CGinjective\](a), and hence a proper $C$–Gorenstein injective resolution is exact. Consequently, we immediately get the inequality: $$\ProperGIdim M \,\geqslant\, \GIdim M.$$ To show the opposite inequality, we may assume that $n = \GIdim M$ is finite. Let $0 \to M \to E^0 \to E^1 \to \cdots$ be a proper $C$–Gorenstein injective resolution of $M$. Defining $D^n =
\textnormal{Coker}(E^{n-2} \to E^{n-1})$ we get an exact sequence: $$0 \to M \to E^0 \to \cdots \to E^{n-1} \to D^n \to 0,$$ which also stays exact when we apply to it the (left exact) functor $\Hom_A(-,E)$ for every $C$–Gorenstein injective $A$–module $E$. Since $\GIdim M = \CGid M = n$, we get by [@HHGorensteinHomDim Theorem 2.22] and Proposition \[prp:inj\_proj\_module\_case\](1) that $D^n$ is $C$–Gorenstein injective, so $\ProperGIdim M \leqslant n$.
Sometimes, nice proper $C$–Gorenstein injective resolutions exist:
\[prp:nice\_CGI-res\] If $M$ is module in $\sA_C(A)$ such that $n = \GIdim M$ is finite, then there exists a proper $C$–Gorenstein injective resolution of the form: $$\tag{\text{$*$}}
0 \to M \to H^0 \to \Hom_A(C,I^1) \to \cdots \to \Hom_A(C,I^n)
\to 0,$$ where $H^0$ is $C$–Gorenstein injective and $I^1, \ldots, I^n$ are injective.
As in the proof of Theorem \[thm:CGid\_and\_AC\], the assumption $M
\in \sA_C(A)$ gives the existence of an exact sequence of $A$–modules: $$0 \to M \to \Hom_A(C,J^0) \to \ldots \to \Hom_A(C,J^{n-1}) \to
D^n \to 0,$$ where $J^0,\ldots,J^{n-1}$ are injective, and $D^n$ is Gorenstein injective over $A \ltimes C$. Applying Lemma \[lem:CGorInj/A<=GorInj/AxC\] to $D^n$ we get a commutative diagram of $A$–modules with exact rows: [ $$\xymatrix{0 \ar[r] & M \ar@{..>}[d] \ar[r] & \Hom_A(C,J^0)
\ar@{..>}[d] \ar[r] & \cdots \ar[r] & \Hom_A(C,J^{n-1})
\ar@{..>}[d] \ar[r] & D^n \ar@{=}[d] \ar[r] & 0 \\
0 \ar[r] & D^0 \ar[r] & \Hom_A(C,U^0) \ar[r] & \cdots
\ar[r] & \Hom_A(C,U^{n-1}) \ar[r] & D^n \ar[r] & 0 }$$ ]{} $\!\!\!\!$ where $U^0,\ldots,U^{n-1}$ are injective and $D^0$ is $C$–Gorenstein injective. The mapping cone of this chain map is of course exact, and furthermore, it has $0 \to D^n
\stackrel{=}{\longrightarrow} D^n \to 0$ as a subcomplex.
Consequently, we get the exact sequence $(*)$, where $I^i = U^{i-1}
\oplus J^i$ for $i = 1, \ldots, n-1$ together with $I^n = U^{n-1}$ are injective; and $H^0 = D^0 \oplus \Hom_A(C,J^0)$ is $C$–Gorenstein injective.
We claim that the sequence $(*)$ remains exact when we apply to it the functor $\Hom_A(-,N)$ for any $C$–Gorenstein injective $A$–module $N$ (and this will finish the proof):
Splitting $(*)$ into short exact sequences, we get sequences of the form $0 \to X \to Y \to Z \to 0$, where $Z$ has the property that it fits into an exact sequence: $$0 \to Z \to \Hom_A(C,E^0) \to \Hom_A(C,E^m) \to 0,$$ where $E^0,\ldots,E^m$ are injective. Therefore, it suffices to prove that every such module $Z$ satisfies $\Ext_A^1(Z,N)=0$ for all $C$–Gorenstein injective modules $N$. But as $\Ext_A^{\geqslant
1}(\Hom_A(C,E^i),N)=0$ for $i=0,\ldots,m$, this follows easily.
We do not know if every module has a proper $C$–Gorenstein projective resolution. However, in the case where $A$ admits a dualizing complex and where $C=A$, then the answer is positive by [@PJgorproj Theorem 3.2].
“Dualizing” the proof of Theorem \[thm:proper\_CGid\] (except the first part about existence of proper resolutions) and Proposition \[prp:nice\_CGI-res\], we get:
\[thm:proper\_CGpd\] Assume that $M$ is an $A$–module which has a proper $C$–Gorenstein projective resolution. Then we have an equality:
[3]{} & & M = M. & &&
\[prp:nice\_CGP-res\] If $M$ is module in $\sB_C(A)$ such that $n = \GPdim M$ is finite, then there exists a proper $C$–Gorenstein projective resolution of the form: $$0 \to C\ten_AP_n \to \cdots \to C\ten_AP_1 \to G_0 \to M \to 0$$ where $G_0$ is $C$–Gorenstein projective and $P_1, \ldots, P_n$ are projective. Furthermore, if $M$ is finitely generated, then $G_0,
P_1, \ldots, P_n$ may be taken to be finitely generated as well. $\qed$
The $C$–Gorenstein flat case is more subtle. We begin with the next:
\[lem:C-Gor-flat\_Kaplansky\] The class of $C$–Gorenstein flat $A$–modules is Kaplansky, and closed under direct limits.
As in the proof of Lemma \[lem:C-Gor-flat\_Kaplansky\]; this time using [@EnochsLopezRamos Proposition 2.10], we see that the class of $C$–Gorenstein flat $A$–modules is Kaplansky.
By Proposition \[prp:flat\_module\_case\], a module $M$ is $C$–Gorenstein flat if and only if $M$ satisfies conditions (F1) in Definition \[def:Cmodules\] and (F2’) in Proposition \[prp:flat\_module\_case\]. Clearly, the condition (F1) is closed under direct limits.
Concerning condition (F2’), we recall from Lemma \[lem:C\_tensor\_flad\_Kaplansky\] that the class of $A$–modules ${\sf F} = \{ C\ten_AF \,|\ F \,\text{flat } A\text{--module}\}$ is closed under direct limits. Condition (F2’) states that $M$ admits an infinite proper right $\sF$–resolution, or in the language of [@EnochsJendaOyonarte; @EnochsLopezRamos], that $\mu_{\sf F}(M) =
\infty$. Hence [@EnochsLopezRamos Theorem 2.4] implies that also (F2’) is closed under direct limits.
\[thm:proper\_CGfd\] Every $A$–module $M$ has a proper $C$–Gorenstein flat resolution, and there is an equality: $$\ProperGFdim M \,=\, \GFdim M.$$
The class $\sG\sF_C(A)$ of $C$–Gorenstein flat modules contains the projective (in fact, flat) modules by Example \[exa:injective\_is\_CGinjective\](c), and furthermore, it is closed under extensions by [@HHGorensteinHomDim Theorem 3.7] and Proposition \[prp:flat\_module\_case\].
Thus, by Lemma \[lem:C-Gor-flat\_Kaplansky\] above and [@EnochsLopezRamos Theorem 2.9] we conclude that the pair $(\sG\sF_C(A),\sG\sF_C(A)^{\perp})$ is a *perfect cotorsion theory* according to [@EnochsLopezRamos Definition 2.2]. In particular, every module admits a $C$–Gorenstein flat (pre)cover, and hence proper $C$–Gorenstein flat resolutions always exist.
The equality $\ProperGFdim M = \GFdim M$ follows as in Theorem \[thm:proper\_CGid\]; this time using [@HHGorensteinHomDim Theorem 3.14] instead of [@HHGorensteinHomDim Theorem 2.22].
[99]{}
L. W. Christensen, “Functorial dimensions”, Ph.D. thesis, 1999, available from `http://www.math.ku.dk/winther/papers.html`.
, “Gorenstein dimensions”, Lecture Notes in Math., Vol. 1747, Springer-Verlag, Berlin, 2000.
, [*Semi-dualizing complexes and their Auslander categories*]{}, Trans. Amer. Math. Soc. [**353**]{} (2001), 1839–1883.
, A. Frankild, and H. Holm, [*On Gorenstein projective, injective and flat dimensions — a functorial description with applications*]{}, preprint (2003), available from `http://arXiv.org/abs/math.AC/0403156`.
E. E. Enochs and O. M. G. Jenda, [ *Gorenstein injective and projective modules*]{}, Math. Z. [**220**]{} (1995), 611–633.
and O. M. G. Jenda, “Relative homological algebra”, de Gruyter Exp. Math., Vol. 30, de Gruyter, Berlin, 2000.
, O. M. G. Jenda and L. Oyonarte, [*$\lambda$ and $\mu$-dimensions of modules*]{}, Rend. Sem. Math. Univ. Padova [**105**]{} (2001), 111–123.
and J. A. López-Ramos, [ *Kaplansky classes*]{}, Rend. Sem. Math. Univ. Padova [**107**]{} (2002), 67–79.
R. Hartshorne, “Residues and duality”, Lecture Notes in Math., Vol. 20, Springer-Verlag, Berlin, 1966. Notes of a seminar on the work of A. Grothendieck, given at Harvard 1963/64. With an appendix by P. Deligne.
H. Holm, [*Gorenstein homological dimensions*]{}, J. Pure Appl. Algebra. [**189**]{} (2004), 167–193
, [*Rings with finite Gorenstein injective dimension*]{}, Proc. Amer. Math. Soc. [**132**]{} (2004), 1279–1283
and P. Jørgensen, [*Cohen-Macaulay injective, projective, and flat dimensions*]{}, preprint (2004).
P. Jørgensen, [*The Gorenstein projective modules are precovering*]{}, preprint (2003), available from [ http://arXiv.org/abs/math.RA/0312263]{}.
O. Veliche, [*Gorenstein projective dimension for complexes*]{}, Trans. Amer. Math. Soc. (to appear), available from [ http://www.math.purdue.edu/oveliche]{}
C. A. Weibel, “An introduction to homological algebra”, Cambridge Studies in Advanced Math., Vol. 38, Cambridge University Press, Cambridge, 1994.
J. Xu, “Flat covers of modules”, Lecture Notes in Math., Vol. 1634, Springer-Verlag, Berlin, 1996.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The phase diagram of a driven two-dimensional vortex lattice in the presence of dense quasi-point pins is investigated. The transition from the crystal to the liquid is found continuous at intermediate inductions. The correlations in the pseudo random force that allow for an uncomplete unbinding of the dislocations is proposed as a key mechanism to account for the continuous transition.'
author:
- 'L. Fruchter'
title: 'Continuous melting of a driven two-dimensional flux lattice with strong pins'
---
Introduction
============
It has long been noticed that a driven elastic lattice driven at zero temperature may experience pinning as an effective shaking temperature, due to randomly induced displacements of the lattice nodes. This powerful analogy allows for the prediction of some properties of the driven lattice from the common phase diagram of particles with a repulsive interaction. In particular, a dynamic melting transition is predicted and observed numerically[@koshelev94]. However, strongly disordered systems, as obtained in the presence of strong quasi-point pins, may move aside from this picture. Indeed, the thermal analogy may break down for the moving crystal due to temporal correlations of the pseudo thermal Langevin force, a situation which is encountered in the case of heterogeneous pinning involving plastic flow channels[@balents98]. As a consequence, while this analogy accounts for the existence of a first order-like dynamic transition, the driven phases may differ from their thermodynamic analogues. Several examples of such exotic phases have been given in refs. [@moon96; @ryu96; @olson98; @fangohr01]. The anisotropy of the pinning potential, once tilted by the external force, is essential to the formation of these phases[@balents98; @ledoussal98]. Here, the dynamic transition is examined in detail for a simple system of dense quasi-point pins with positional disorder. For part of the phase diagram, it is found that there is a continuous transition between the crystal and the liquid, through what might be called a liquid crystal, whereas the transition is first order like for the rest of the phase diagram. In the case of the continuous transition, disclinations tend to form chains, which likely arise from the correlations of the pseudo random force which are specific to the driven lattice.
Experiment
==========
The numerical sample used here is the one in ref. [@fruchter02]. It is the one of two dimensional particles with a repulsive interaction, interacting also with a random attractive potential. The sample mimics a two dimensional vortex lattice, or a three dimensional rigid vortex lattice, in the presence of strong pins, as can be created by heavy ions irradiation. Adopting the terminology of superconductors, the vortex density is set by the magnetic induction, $B$, as $n =
a_0^{-2} = B / \Phi_0 $ - $\Phi_0$ being the flux quantum carried by each vortex ($2\;10^{-7}\;G cm^2$). The repulsive force between vortices is taken as :
$$f_{vv}(r)=\left(A_{V}/\lambda\right) K_{1}\left(r/\lambda\right)
\label{fvv}$$
where $K_{1}$ is a Bessel function, behaving as $\ln\;r^{-1}$ at short distance and $r^{-1/2}\exp(-r)$ at large distance. To keep computation tractable, the repulsive force is cut smoothly at a distance $11 a_0$, which insures that each particle interacts with many of its closest neighbors.
The short range potential originates from strong pins randomly distributed in the sample, each creating the attractive force in the range $r_{P}$:
$$f_{p}(r) = \left(2\;A_{P}/r_{P}\right)\:(r/r_{P}),\: \textrm{for
$r \leq r_{P}$};\;
0 \;\textrm{for $r > r_{P}$}
\label{fp}$$
All pins are identical and the randomness of the potential originates from the pins position only.
In the rest, driving current densities are normalized to the single vortex critical current density, $J_c =
2\;A_{P}/r_{P}\;\Phi_0$. The density of the pinning sites, relative to that of the vortices, $B_\Phi/B$, is constant and equal to $12$. The pinning potential range, relative to the vortex average separation, is also constant and equal to $r_P/a_0
= 5.5 \;10^{-2}$, as well as the reduced force magnitude, $\lambda A_{P}/r_{P} A_{V} = 20$. As a consequence, using $a_0$ as the length scale, the different numerical experiments made for different values of the induction $B$ only differ by the reduced vortex interaction length, $\lambda/a_0$, where it was set $\lambda=1400\;\AA$. As in [@fruchter02], the total force on each vortex, originating from its neighbors, a possible pinning site at the vortex location and the uniform external force is computed at each time step. Vortices which are not pinned are then moved on a time interval small enough so that their motion is small compared to all characteristic lengths. The boundary conditions are periodic along the driving force direction. A large area free from any pinning site is kept at the sample edges parallel to the vortex motion, where a perfect hexagonal lattice is obtained under the action of the external magnetic pressure[@fruchter02]. In this way, the measurements actually sample the driven phase embbeded in the crystal. Whereas such an interface may promote the formation of the ordered phase in the case of a first order transition and for finite samples, in the case of a continuous transition, as will be discussed later, the interface probably induces an interfacial layer only. In the following, samples far enough from the interface are considered and their uniformity is an indication that finite size effects are not playing a major role when a continuous transition is observed.
Experiments are carried out for different values of the induction and of the external force. After a stationary state is obtained (characterized by a steady average velocity), a snapshot of the moving lattice is recorded, on which a Delaunay triangulation is performed. Positive and negative disclinations (vortices with coordinance 5 and 7), either free or forming dislocations by pairs[@nelson79] are counted. Samples typically enclose $7000$ vortices and $4\;10^4$ pins.
\[results\]
results and discussion
======================
As shown in ref.[@fruchter02], as the driving force decreases, the system evolves from a moving crystal to an amorphous phase. Contrasting with the results in [@moon96; @ryu96; @olson98], the high velocity phase does not show here smectic ordering, as evidenced from the diffraction pattern : this comes from the small ratio $r_p/a_0$ and from the fact that the tilted pinning potential shows here a moderate anisotropy on the scale of $a_0$. There is no attractive interaction between the vortices, which would allow for a transition between a liquid and a gas. However, considering the comparable densities of the crystal and the less ordered phase, as well as the strong interactions between the vortices in the amorphous phase, it must obviously be called a ’liquid phase’. As evidenced in Fig.\[159\] and \[155\] and the inspection of the average hexatic parameter,$|<\Psi_{6}>|=
|<1/c_\alpha\;\sum_{\beta=1,c_\alpha}
e^{6\;i\;\theta_{\alpha,\beta}}>_{\alpha}|$ where is $c_\alpha$ the coordination number for vortex $\alpha$ and $\theta_{\alpha,\beta}$ is the angle of the bond between neighboring vortices $\alpha$ and $\beta$, some residual orientational correlation is retained for low $j$ ($|<\Psi_{6}>|
\simeq 0.1$), which justifies to call the low $j$ phase an ’hexatic liquid’[@ryu96].
I now examine in more detail the transition between the crystal and the liquid. For all systems, the concentration of defects exhibits a clear onset upon decreasing the driving force, similar to the one reported in [@koshelev94]. However, depending on the magnetic pressure, a discontinuous or gradual rise of this concentration is observed. This may be seen in Figs. \[159\] and \[155\] obtained for two different magnetic inductions, which clearly exhibit respectively a gradual and a step increase of the number of defects. In order to quantify this observation, the defects concentration was fitted with an exponential, $n_d
\propto 1 - \exp [(j_o -j)/\delta]\; (j < j_o) $, yielding the onset, $j_o$, and a width for the transition to the liquid phase, $\delta$. A phase diagram similar to the temperature-density representation for the thermodynamics may be obtained, using the theory for the equivalent ’shaking temperature’[@koshelev94]. It should be stressed that this representation is qualitative only, considering the reservations made in ref.[@koshelev94; @balents98] (mainly, the perturbative approach from the uniform velocity which rules out plasticity, and the observation that the effective temperature differs for the fluidlike motion and the coherent one). Also, the equivalent temperature in [@koshelev94], must be modified to account for the proximity of the flux flow to the flux creep crossover. Imposing for the equivalent temperature to be proportional to the potential well depth when $j_o \rightarrow 1$ and $B \rightarrow
0$, $T \propto (1-j)$, and using $T \propto j^{-1}$ from [@koshelev94], a phase diagram is obtained in the $B$ vs $(1-j)/j$ representation. The onset for the defects concentration ($j_o$), as well as the location where it saturates ($j_o-\delta$), are plotted in this way in Fig.\[phasediag\]. Clearly, there is a range of magnetic induction for which a regime, intermediate between the moving crystal and the hexatic liquid, can be found. The existence of such a regime was already pointed out in [@fruchter02].
In order to characterize the continuous transition, let us examine some autocorrelation functions which are classical tools for the study of solids and liquids. The average hexatic order parameter does not provide an accurate characterization of the intermediate regime: as may be seen in Fig. \[159\], following a sharp drop at $j=j_o$, there is no significative change at lower $j$ where the density of defects however still exhibits significant variations. The *spatial correlations* of the hexatic parameter carry more useful information[@nelson79]. The correlator $<\Psi_{6}(0)\;\Psi_{6}^{*}(r)>_{r}$ for the data in Fig.\[159\] is shown in Fig.\[psi6\]. Besides the existence of a non zero background related to the non zero averaged value $|<\Psi_{6}>|$, it reveals some additional short range correlations of the orientational order, which extends to a larger range as the system is closer to $j_o$. The oscillations for small $r$ reflect the existence of a crystalline order within this range: they are associated with the fluctuations of the density autocorrelation function which come with the translational symmetry breaking of the crystal order. This is confirmed by the examination of the growth of the displacement field (actually a positional correlation function): $<\mathbf{u}^{2}(r=n\; a_0)> =
<\sum_{j}
\mathbf{u}^{2}(n\;[\mathbf{r}_{j}-\mathbf{r}_{i}])>_{i}$ where $j$ denotes one of the nearest neighbors of vortex $i$ in a Delaunay triangulation and $\mathbf{u}$ is the displacement field from the periodic arrangement. As may be seen in Fig. \[u2\], there is an exponential decay of the positional correlations, with a diverging correlation length, $\xi(j)$, as one approaches $j_0$ from below (i.e. from larger ’temperatures’). Following ref. [@nelson79], one may then call this regime an ’hexatic liquid crystal’. It is possible to track the positional correlation length in Fig.\[u2\] as one approaches $j_0$. The result is diplayed in Fig.\[length\] showing a divergence as $\xi =
\xi_0\;(1-j/j_0)^{-1}$, with the bare correlation length $\xi_0
\simeq 0.15\;a_0$. Recalling the Lindeman melting criterion, $<u^{2}> = c_L^{2}\;a_0^{2}$ and the exponential increase of the displacement field, $<u^{2}> = <u^{2}(\infty)> (1 -
\exp(-r/\xi))$, one may write an equivalent melting criterion for the present case as $a_{0}/\xi =
\ln(1-c_{L}^{2}\;a_{0}^{2}/<u^{2}(\infty)>)$. The result obtained using $c_{L}\approx 0.2$ and $<u^{2}(\infty)> \simeq 0.14$, $a_{0}/\xi \approx0.3$, is displayed in Fig.\[159\] and \[length\]. Although this quantitative result should be considered with caution, due to the uncertainty on the effective Lindeman number, this confirms that the solid has not melted in the conventional way below the threshold value $j_0$.
The observation that the transition is continuous at intermediate induction and involves a proliferation of defects appeals for a comparison with the KTHNY extension of the Kosterliz-Thouless theory [@nelson79; @young79]. The theory of dislocation mediated melting of two-dimensional solids accounts for a continuous transition, involving first unbinding of the dislocations leading to the hexatic liquid, and then unbinding of the disclinations leading to the regular liquid. Here, dislocations do not first dissociate at $j_0$ to form an homogeneous ’plasma’. Rather, they tend to form chains of alternating positive (five-coordinated) and negative (seven-coordinated) disclinations which proliferate in the liquid phase. As a result, unbounded dislocations and disclinations remain marginal (Fig. \[chains\]). Correlations between dislocations were also reported in ref.[@ryu96] where free dislocations, although not bound in chains, formed quenched patterns moving with the average flux flow. In order to explain the formation of these chains, the examination of the early creation of defects in a driven crystal may be useful. Snapshots of the earlier defects detected in a sample driven in the intermediate region in Fig.\[159\] are displayed in Fig.\[early\]. After a dislocation pair with opposite Burger vectors has been created by the pinning of one vortex (a), it is seen that the dislocations quickly arrange to form rings of diameter $\sim 2\; a_0$ (c) and then larger loops (d). Remarkably, the composite defects reflect the external force anisotropy as soon as the dislocations dissociate (b): this results from the plastic mechanism at work to create these defects. This is also a direct evidence that the correlations in the pseudo random force cannot be neglected in their formation. The relation between these initial stages and the formation of chains is not completely clear. A possible mechanism is the stretching of elementary loops as in Fig.\[early\]d, as the vortices making disclinations appear to become more easily pinned than the regular ones. This would make the long-range ’random force’ correlations a key ingredient in the chain formation again. Equivalent rates for the growth and the annihilation of the chains would then account for the existence of a stationary regime intermediate between the crystal and the liquid.
In conclusion, it is found that a vortex lattice driven on dense quasi-point pins shows a continuous transition between the crystal and the liquid at intermediate induction, while first order otherwise. The binding of the disclinations in chains is proposed as a key mechanism to account for the existence of the continuous transition.
Simulations have been performed on the cluster of the Centre de Ressources Informatiques de l’Universite Paris-Sud (CRI).
A.E. Koshelev and V.M. Vinokur, Phys. Rev. Lett.[**73**]{}, 3580 (1994).
L. Balents, M.C. Marchetti, and L. Radzihovsky, Phys. Rev. B [**57**]{}, 7705 (1998).
K. Moon, R.T. Scalettar, G.T. Zimanyl, Phys. Rev. Lett. 77, 2778 (1996).
R. Seungoh, M. Hellerqvist, S. Doniach, A. Kapitulnik, D. Stroud, Phys. Rev. Lett. [**77**]{}, 5114 (1996).
C.J. Olson, C. Reichhardt, F. Nori, Phys. Rev. Lett. 81,3757 (1998).
H. Fangohr, S.J. Cox, P.A.J. de Groot, Phys. Rev. B [**64**]{}, 064505 (2001).
P. Le Doussal and T. Giamarchi, Phys. Rev. B [**57**]{}, 11356 (1998).
L. Fruchter, Eur. Phys. J. B [**25**]{}, 313 (2002).
D.R. Nelson and B.I. Halperin, Phys. Rev. B [**19**]{}, 2457 (1979).
A.P. Young, Phys. Rev. B [**19**]{}, 1855 (1979).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recently, within the framework of the Composite Operator Method, it has been proposed a three-pole solution for the two-dimensional Hubbard model [@Avella_14], which is still considered one of the best candidate model to microscopically describe high-$T_{c}$ cuprate superconductors. The operatorial basis comprise the two Hubbard operators (complete fermionic local basis) and the electronic operator dressed by the nearest-neighbor spin fluctuations. The effectiveness of the approximate solution has been proved through a positive comparison with different numerical methods for various quantities. In this article, after recollecting the main analytical expressions defining the solution and the behavior of basic local quantities (double occupancy and chemical potential) and of the quasi-particle energy dispersions, we resolve and analyze the momentum components of relevant quantities: filling (i.e. the momentum distribution function), double occupancy and nearest-neighbor spin correlation function. The analysis is extended to COM(2p) solutions that will be used as primary reference. Thanks to this, the role played by the third field, with respect to the two Hubbard ones, in determining the behavior of many relevant quantities and in allowing the extremely good comparison with numerical results is better understood giving a guideline to further improve and, possibly, optimize the application of the COM to the Hubbard model.'
author:
- Adolfo Avella
title: 'COM(3p) solution of the 2D Hubbard model: momentum resolved quantities'
---
Introduction
============
The cuprate high-$T_{c}$ superconductors [@Bednorz_86] still lack a widely accepted and unifying microscopic description of their anomalous behavior experimentally observed, mainly in the underdoped region, in almost all experimentally measurable physical properties [@Timusk_99; @Basov_99; @Orenstein_00; @Damascelli_03; @Shen_05; @Eschrig_06; @Kanigel_06; @Lee_06; @Valla_06; @Doiron-Leyraud_07; @LeBoeuf_07; @Hossain_08; @Sebastian_08; @Meng_09; @Laliberte_11; @Ramshaw_11; @Riggs_11; @Sebastian_11a; @Sebastian_12a]. Non-Fermi-liquid response, quantum criticality, pseudogap formation, ill-defined Fermi surface, kinks in the electronic dispersion, $\ldots$ cannot be explained by standard many-body theory within the Fermi-liquid framework by means of diagrammatic expansions and remain controversially debated [@Lee_06; @Tremblay_06; @Sebastian_12a]. Strong electronic correlations, competition between localization and itinerancy, Mott physics, and low-energy spin excitations are considered key ingredients necessary to explain these anomalous features and the Hubbard model [@Hubbard_63] contain all of them by construction.
The Hubbard model[@Hubbard_63] together with its relevance to real materials, in particular cuprate high-$T_{c}$ superconductors, has always raised a more fundamental and theoretical interest as it is universally considered the prototypical model for strongly correlated systems. Unfortunately, although many trials have been made, no analytical approximation method can be considered to have given a clear and definitive answer to the very many relevant issues raised by this very simple model. Numerical approaches [@Avella_13a] are fundamental for benchmarking and fine tuning analytical theories and for establishing which are those capable to deal with the quite complex phenomenology of the Hubbard model. Unfortunately, numerical techniques cannot explore, because of their limited resolution in frequency and momentum, the most relevant regime of model parameters (small doping, low temperature and large on-site Coulomb repulsion) where one expects strong electronic correlations to dominate the physics of the system. As regards analytical and semi-analytical (i.e. embedding a numerical core) theories [@Avella_12], a few are definitely worth mentioning: the work of Mori [@Mori_65], Hubbard [@Hubbard_63; @Hubbard_64; @Hubbard_64a], Rowe [@Rowe_68], Roth [@Roth_69], Tserkovnikov [@Tserkovnikov_81; @Tserkovnikov_81a], the Gutzwiller approximation [@Gutzwiller_63; @Gutzwiller_64; @Gutzwiller_65; @Brinkman_70], the slave boson method [@Barnes_76; @Coleman_84; @Kotliar_86], the spectral density approach [@Kalashnikov_69; @Nolting_72], the two-particle self-consistent approach [@Tremblay_06], the RPA and equations-of-motion based techniques [@Chubukov_04; @Prelovsek_05; @Plakida_06], the dynamical mean-field theory (DMFT) [@Metzner_89; @Georges_92; @Georges_96], the DMFT$+\Sigma$ approach [@Sadovskii_05; @Kuchinskii_05; @Kuchinskii_06] as well as all cluster-DMFT-like theories[@Maier_05] (the cellular-DMFT [@Kotliar_01a], the dynamical cluster approximation [@Hettler_98] and the cluster perturbation theory [@Senechal_00]).
We have also been developing a systematic approach, the composite operator method (COM) [@Theory; @Avella_11a], to study highly correlated systems. In the last fifteen years, COM has been applied to several models and materials: Hubbard [@Hub-redux; @Odashima_05; @Avella_14], $p$-$d$ [@p-d], $t$-$t'$-$U$ [@ttU-redux], extended Hubbard ($t$-$U$-$V$) [@tUV-redux], Kondo [@Villani_00], Anderson [@Anderson-redux], two-orbital Hubbard [@Plekhanov_11], Ising [@Ising-redux], $J_{1}-J_{2}$ [@Avella_08a; @J1J2-redux], Cuprates [@Cuprates-NCA; @Avella_07; @Avella_07a; @Avella_08; @Avella_09], etc The Composite Operator Method (COM) [@Theory; @Avella_11a] has the advantage to be completely microscopic, exclusively analytical, and fully self-consistent. COM recipe uses three main ingredients [@Theory; @Avella_11a]: *composite* operators*, algebra* constraints and *residual* self-energy. Composite operators are products of electronic operators and describe the new elementary excitations appearing in the system owing to strong correlations. According to the system under analysis [@Theory; @Avella_11a], one has to choose a set of composite operators as operatorial basis and rewrite the electronic operators and the electronic Green’s function in terms of this basis. Algebra Constraints are relations among correlation functions dictated by the non-canonical operatorial algebra closed by the chosen operatorial basis [@Theory; @Avella_11a]. Other ways to obtain algebra constraints rely on the symmetries enjoined by the Hamiltonian under study, the Ward-Takahashi identities, the hydrodynamics, etc [@Theory; @Avella_11a]. Algebra Constraints are used to compute unknown correlation functions appearing in the calculations. Interactions among the elements of the chosen operatorial basis are described by the residual self-energy, that is, the propagator of the residual term of the current after this latter has been projected on the chosen operatorial basis [@Theory; @Avella_11a]. According to the physical properties under analysis and the range of temperatures, dopings, and interactions you want to explore, one has to choose an approximation to compute the residual self-energy. In the last years, we have been using the $n-$pole Approximation [@Hub-redux; @Odashima_05; @p-d; @ttU-redux; @tUV-redux; @Plekhanov_11; @Ising-redux; @Cuprates-NCA; @Avella_14], the Asymptotic Field Approach [@Villani_00; @Anderson-redux] and the Non-Crossing Approximation (NCA) [@Avella_07; @Avella_07a; @Avella_08; @Avella_09].
In this article, we first recollect the main analytical expressions defining the COM(3p) approximation for the 2D Hubbard model (Sec. \[sec:Theory\]). More details can be found in [@Avella_14]. Then, we set the stage by reporting both (i) basic local quantities (Sec. \[sec:Results:D\_mu\]: double occupancy and chemical potential), comparing them with numerical and semi-analytical methods to asses the solution and characterize it, and (ii) the quasi-particle dispersions (Sec. \[sec:Results:Ek\]). These latter, in particular, together with the comparison to COM(2p) solutions, will allow to analyze and understand the behavior of the momentum resolved components of relevant quantities: filling (i.e. the momentum distribution function), double occupancy and nearest-neighbor spin correlation function (Sec. \[sec:Results:nk\_Dk\_chiask\]). Finally, in Sec. \[sec:Summary\], we draw some conclusions.
Theory\[sec:Theory\]
====================
Hamiltonian
-----------
The Hamiltonian of the two-dimensional Hubbard model reads as $$\begin{gathered}
H=-4t\sum_{\mathbf{i}}c^{\dagger}\left(i\right)\cdot c^{\alpha}\left(i\right)\\
+U\sum_{\mathbf{i}}n_{\uparrow}\left(i\right)n_{\downarrow}\left(i\right)-\mu\sum_{\mathbf{i}}n\left(i\right)\label{eq:Ham}\end{gathered}$$ where $c^{\dagger}\left(i\right)=\begin{pmatrix}c_{\uparrow}^{\dagger}\left(i\right) & c_{\downarrow}^{\dagger}\left(i\right)\end{pmatrix}$ is the electronic field operator in spinorial notation and Heisenberg picture ($i=\left(\mathbf{i},t_{i}\right)$). $\cdot$ and $\otimes$ stand for the inner (scalar) and the outer products, respectively, in spin space. $\mathbf{i}=\mathbf{R_{i}}=\left(i_{x},i_{y}\right)$ is a vector of the two-dimensional square Bravais lattice, $n_{\sigma}\left(i\right)=c_{\sigma}^{\dagger}\left(i\right)c_{\sigma}\left(i\right)$ is the particle density operator for spin $\sigma$ at site $\mathbf{i}$, $n\left(i\right)=\sum_{\sigma}n_{\sigma}\left(i\right)=c^{\dagger}\left(i\right)\cdot c\left(i\right)$ is the total particle density operator at site $\mathbf{i}$, $\mu$ is the chemical potential, $t$ is the hopping integral and the energy unit hereafter, $U$ is the Coulomb on-site repulsion and $\alpha_{\mathbf{ij}}$ is the projector on the nearest-neighbor sites $$\begin{aligned}
\alpha_{\mathbf{ij}} & =\frac{1}{N}\sum_{\mathbf{k}}\mathrm{e}^{\mathrm{i}\mathbf{k}\cdot(\mathbf{R_{i}}-\mathbf{R_{j}})}\alpha\left(\mathbf{k}\right)\label{eq:alpha-ij}\\
\alpha\left(\mathbf{k}\right) & =\frac{1}{2}\left[\cos\left(k_{x}a\right)+\cos\left(k_{y}a\right)\right]\label{eq:alpha-k}\end{aligned}$$ where $\mathbf{k}$ runs over the first Brillouin zone, $N$ is the number of lattice sites and $a$ is the lattice constant, which will be set to one for the sake of simplicity. For any operator $\Phi\left(i\right)$, we use the notation $\Phi^{\kappa}\left(i\right)=\sum_{\mathbf{j}}\kappa_{\mathbf{ij}}\Phi\left(\mathbf{j},t_{i}\right)$ where $\kappa_{\mathbf{ij}}$ can be any function of the two sites $\mathbf{i}$ and $\mathbf{j}$ and, in particular, a projector over the cubic harmonics of the lattice: e.g. $c^{\alpha}\left(i\right)=\sum_{\mathbf{j}}\alpha_{\mathbf{ij}}c\left(\mathbf{j},t_{i}\right)$.
Basis and equations of motion
-----------------------------
According to COM prescription [@Theory; @Avella_11a], we have chosen as composite basic field $$\psi\left(i\right)=\begin{pmatrix}\psi_{1}\left(i\right)\\
\psi_{2}\left(i\right)\\
\psi_{3}\left(i\right)
\end{pmatrix}=\begin{pmatrix}\xi\left(i\right)\\
\eta\left(i\right)\\
c_{s}\left(i\right)
\end{pmatrix}\label{eq:basis}$$ where $\eta\left(i\right)=n\left(i\right)c\left(i\right)$ and $\xi\left(i\right)=c\left(i\right)-\eta\left(i\right)$ are the Hubbard operators and $c_{s}\left(i\right)=n_{k}\left(i\right)\sigma_{k}\cdot c^{\alpha}\left(i\right)$ is the electronic operator dressed by the nearest-neighbor spin fluctuations, which are expected to be the most relevant fluctuations, compared to charge and pair ones, in determining the fundamental response and the important features of the system under analysis [@Avella_14]. This assumption has been proved to be definitely valid [@Avella_14] in the parameter regime where the electronic correlations are expected to be very strong: large $U$, small doping $\delta=1-n$ and low temperature $T$. In absence of correlations, or for the very weak ones, no type of fluctuations is relevant. $n_{\mu}\left(i\right)=c^{\dagger}\left(i\right)\cdot\sigma_{\mu}\cdot c\left(i\right)$ is the charge- ($\mu=0$) and spin- ($\mu=1,2,3=k$) density operator, $\sigma_{\mu}=\left(1,\vec{\sigma}\right)$, $\sigma^{\mu}=\left(-1,\vec{\sigma}\right)$, $\sigma_{k}$ with $\left(k=1,2,3\right)$ are the Pauli matrices.
The field $\psi(i)$ satisfies the following equation of motion $$\mathrm{i}\frac{\partial}{\partial t}\psi\left(i\right)=\begin{pmatrix}-\mu\xi\left(i\right)-4tc^{\alpha}\left(i\right)-4t\pi\left(i\right)\\
\left(U-\mu\right)\eta\left(i\right)+4t\pi\left(i\right)\\
-\mu c_{s}\left(i\right)+4t\kappa_{s}\left(i\right)+U\eta_{s}\left(i\right)
\end{pmatrix}$$ where the higher-order composite fields $\pi\left(i\right)$, $\kappa_{s}\left(i\right)$ and $\eta_{s}\left(i\right)$ are defined as $$\begin{aligned}
& \pi\left(i\right)=\frac{1}{2}n_{\mu}(i)\sigma^{\mu}\cdot c^{\alpha}\left(i\right)+c^{\dagger\alpha}\left(i\right)\cdot c\left(i\right)\otimes c\left(i\right)\\
& \kappa_{s}\left(i\right)=-n_{k}\left(i\right)\sigma_{k}\cdot c^{\alpha^{2}}\left(i\right)\nonumber \\
& +\left(c^{\alpha\dagger}\left(i\right)\cdot\sigma_{k}\cdot c\left(i\right)-c^{\dagger}\left(i\right)\cdot\sigma_{k}\cdot c^{\alpha}\left(i\right)\right)\sigma_{k}\cdot c^{\alpha}\left(i\right)\\
& \eta_{s}\left(i\right)=n_{k}\left(i\right)\sigma_{k}\cdot\eta^{\alpha}\left(i\right)\end{aligned}$$ It is clear now that $c_{s}\left(i\right)$ has been chosen proportional to the *spin* component of $\pi\left(i\right)$. Accordingly, we define $\bar{\pi}\left(i\right)=\pi\left(i\right)-\frac{1}{2}c_{s}\left(i\right)$.
Current projection (pole approximation)
---------------------------------------
The current $J\left(i\right)=\mathrm{i}\frac{\partial}{\partial t}\psi\left(i\right)=\left[\psi\left(i\right),H\right]$ of the basis $\psi(i)$ can be approximated $$J\left(i\right)\cong\sum_{\mathbf{j}}\varepsilon\left(\mathbf{i},\mathbf{j}\right)\psi\left(\mathbf{j},t\right)\label{eq:J}$$ projecting the current $J\left(i\right)$ on the basis $\psi\left(i\right)$. $\varepsilon\left(\mathbf{i},\mathbf{j}\right)$ is named energy matrix and can be computed by means of the equation $$\begin{gathered}
\left\langle \left\{ J\left(\mathbf{i},t\right),\psi^{\dagger}\left(\mathbf{j},t\right)\right\} \right\rangle \\
=\sum_{\mathbf{j}}\varepsilon\left(\mathbf{i},\mathbf{j}\right)\left\langle \left\{ \psi\left(\mathbf{i},t\right),\psi^{\dagger}\left(\mathbf{j},t\right)\right\} \right\rangle \label{eq:dJ}\end{gathered}$$ where $\left\langle \cdots\right\rangle $ stands for the thermal average taken in the grand-canonical ensemble: $$\varepsilon(\mathbf{k})=m(\mathbf{k})I^{-1}(\mathbf{k})\label{eq:m-I}$$ where $$\begin{aligned}
\varepsilon\left(\mathbf{i},\mathbf{j}\right) & =\frac{1}{N}\sum_{\mathbf{k}}\mathrm{e}^{\mathrm{i}\mathbf{k}\cdot(\mathbf{R_{i}-R_{j}})}\varepsilon\left(\mathbf{k}\right)\\
I\left(\mathbf{i},\mathbf{j}\right) & =\left\langle \left\{ \psi\left(\mathbf{i},t\right),\psi^{\dagger}\left(\mathbf{j},t\right)\right\} \right\rangle \nonumber \\
& =\frac{1}{N}\sum_{\mathbf{k}}\mathrm{e}^{\mathrm{i}\mathbf{k}\cdot(\mathbf{R_{i}-R_{j}})}I\left(\mathbf{k}\right)\\
m\left(\mathbf{i},\mathbf{j}\right) & =\left\langle \left\{ J\left(\mathbf{i},t\right),\psi^{\dagger}\left(\mathbf{j},t\right)\right\} \right\rangle \nonumber \\
& =\frac{1}{N}\sum_{\mathbf{k}}\mathrm{e}^{\mathrm{i}\mathbf{k}\cdot(\mathbf{R_{i}-R_{j}})}m\left(\mathbf{k}\right)\end{aligned}$$ Since $\psi\left(i\right)$ is made up of composite operators, the normalization matrix $I\left(\mathbf{k}\right)$ is not the identity matrix as it happens for the original electronic field operator. Hereafter, we will use the very convenient notation $I_{\phi\varphi}\left(\mathbf{i},\mathbf{j}\right)=\left\langle \left\{ \phi\left(\mathbf{i},t\right),\varphi^{\dagger}\left(\mathbf{j},t\right)\right\} \right\rangle $, which generalizes the definition of the normalization matrix ($I\left(\mathbf{i},\mathbf{j}\right)=I_{\psi\psi}\left(\mathbf{i},\mathbf{j}\right)$) and of the $m$-matrix ($m\left(\mathbf{i},\mathbf{j}\right)=I_{J\psi}\left(\mathbf{i},\mathbf{j}\right)$) and provide the operator space of a scalar product.
Green’s and correlation functions
---------------------------------
By using the projection of the source (\[eq:J\]), that is, by working in the framework of a three-pole approximation, and by introducing the Fourier transform $\mathcal{F}_{\mathbf{k}\omega}\left[\cdots\right]$, the retarded thermodynamic Green’s functions $$\begin{gathered}
G\left(i,j\right)=\left\langle \mathcal{R}\left[\psi\left(i\right)\psi^{\dagger}\left(j\right)\right]\right\rangle \\
=\theta\left(t_{i}-t_{j}\right)\left\langle \left\{ \psi\left(i\right),\psi^{\dagger}\left(j\right)\right\} \right\rangle \label{eq.G}\end{gathered}$$ has the following expression $$\begin{gathered}
G\left(\mathbf{k},\omega\right)=\frac{1}{\omega-\varepsilon\left(\mathbf{k}\right)+\mathrm{i}\delta}I\left(\mathbf{k}\right)\\
=\sum_{m=1}^{3}\frac{\sigma^{\left(m\right)}\left(\mathbf{k}\right)}{\omega-E_{m}\left(\mathbf{k}\right)+\mathrm{i}\delta}\label{eq:Gk}\end{gathered}$$ where $E_{m}\left(\mathbf{k}\right)$ are the eigenvalues of the energy matrix $\varepsilon(\mathbf{k})$ and, as poles of the Green’s function, serve as main excitation bands of the system. $\sigma^{\left(m\right)}\left(\mathbf{k}\right)$ are the spectral density weights per band and can be computed as $$\sigma_{ab}^{(m)}\left(\mathbf{k}\right)=\sum\limits _{c=1}^{3}\Omega_{am}\left(\mathbf{k}\right)\Omega_{mc}^{-1}\left(\mathbf{k}\right)I_{cb}\left(\mathbf{k}\right)\label{eq:sigmak}$$ where the matrix $\Omega\left(\mathbf{k}\right)$ contains the eigenvectors of $\varepsilon\left(\mathbf{k}\right)$ as columns.
The correlation functions of the fields of the basis $C_{ab}\left(i,j\right)=\langle\psi_{a}\left(i\right)\psi_{b}^{\dagger}\left(j\right)\rangle$ can be easily determined in terms of the Green’s function by means of the spectral theorem and their Fourier transforms have the general expression $$\begin{aligned}
C_{ab}\left(\mathbf{k},\omega\right) & =2\pi\sum\limits _{m=1}^{3}C_{ab}^{\left(m\right)}\left(\mathbf{k}\right)\delta\left(\omega-E_{m}\left(\mathbf{k}\right)\right)\label{eq:Ck}\\
C_{ab}^{\left(m\right)}\left(\mathbf{k}\right) & =\left[1-f_{\mathrm{F}}\left(E_{m}\left(\mathbf{k}\right)\right)\right]\sigma_{ab}^{\left(m\right)}\left(\mathbf{k}\right)\label{eq:Ck_m}\end{aligned}$$ where $f_{\mathrm{F}}\left(\omega\right)=\left(\mathrm{e}^{\frac{\omega}{k_{\mathrm{B}}T}}+1\right)^{-1}$ is the Fermi function and $C_{ab}^{\left(m\right)}\left(\mathbf{k}\right)$ is the band component per momentum of the corresponding same-time correlation function $C_{ab}\left(\mathbf{k}\right)=\sum\limits _{m=1}^{3}C_{ab}^{\left(m\right)}\left(\mathbf{k}\right)$.
Normalization $I$ matrix
------------------------
In a paramagnetic and homogeneous system, the normalization $I\left(\mathbf{k}\right)$ matrix has the following entries $$\begin{aligned}
I_{11}\left(\mathbf{k}\right) & =I_{11}=1-\frac{n}{2}\label{eq:I11k}\\
I_{12}\left(\mathbf{k}\right) & =0\label{eq:I12k}\\
I_{13}\left(\mathbf{k}\right) & =3C_{\xi c}^{\alpha}+\frac{3}{2}\alpha\left(\mathbf{k}\right)\chi_{s}^{\alpha}\label{eq:I13k}\\
I_{22}\left(\mathbf{k}\right) & =I_{22}=\frac{n}{2}\label{eq:I22k}\\
I_{23}\left(\mathbf{k}\right) & =3C_{\eta c}^{\alpha}-\frac{3}{2}\alpha\left(\mathbf{k}\right)\chi_{s}^{\alpha}\label{eq:I23k}\\
I_{33}\left(\mathbf{k}\right) & =4C_{c_{s}c}^{\alpha}+\frac{3}{2}C_{\eta\eta}+3\alpha\left(\mathbf{k}\right)\left(f_{s}+\frac{1}{4}C_{cc}^{\alpha}\right)\nonumber \\
& +\frac{3}{2}\beta\left(\mathbf{k}\right)\chi_{s}^{\beta}+\frac{3}{4}\eta\left(\mathbf{k}\right)\chi_{s}^{\eta}\label{eq:I33k}\end{aligned}$$ where $n=\left\langle n\left(i\right)\right\rangle $ is the filling, $\chi_{s}^{\kappa}=\frac{1}{3}\left\langle n_{k}^{\kappa}\left(i\right)n_{k}\left(i\right)\right\rangle $ is the spin-spin correlation function at distances determined by the projector $\kappa$ and $f_{s}=\frac{1}{3}\left\langle c^{\dagger}\left(i\right)\cdot\sigma_{k}\cdot c^{\alpha}\left(i\right)n_{k}^{\alpha}\left(i\right)\right\rangle $ is a higher-order (up to three different sites are involved) spin-spin correlation function. We have also introduced the following definitions, which is based on those related to the correlation functions of the fields of the basis (\[eq:Ck\]): $C_{\phi\varphi}=\left\langle \phi_{\sigma}\left(i\right)\varphi_{\sigma}^{\dagger}\left(i\right)\right\rangle $ and $C_{\phi\varphi}^{\kappa}=\left\langle \phi_{\sigma}^{\kappa}\left(i\right)\varphi_{\sigma}^{\dagger}\left(i\right)\right\rangle $, where no summation over sigma is intended. $\beta\left(\mathbf{k}\right)$ and $\eta\left(\mathbf{k}\right)$ are the projectors onto the second-nearest-neighbor sites along the main diagonals and the main axes of the lattice, respectively.
$m$-matrix
----------
In a paramagnetic and homogeneous system, the $m$-matrix has the following entries $$\begin{aligned}
m_{11}\left(\mathbf{k}\right) & =-\mu I_{11}-4t\left[\Delta+\left(p+I_{11}-I_{22}\right)\alpha\left(\mathbf{k}\right)\right]\label{eq:m11k}\\
m_{12}\left(\mathbf{k}\right) & =4t\left[\Delta+\left(p-I_{22}\right)\alpha\left(\mathbf{k}\right)\right]\label{eq:m12k}\\
m_{13}\left(\mathbf{k}\right) & =-\left(\mu+4t\alpha\left(\mathbf{k}\right)\right)I_{13}\left(\mathbf{k}\right)-4t\alpha\left(\mathbf{k}\right)I_{23}\left(\mathbf{k}\right)\nonumber \\
& -2tI_{33}\left(\mathbf{k}\right)-4t\alpha\left(\mathbf{k}\right)I_{\bar{\pi}c_{s}}^{\alpha}\label{eq:m13k}\\
m_{22}\left(\mathbf{k}\right) & =\left(U-\mu\right)I_{22}-4t\left[\Delta+p\alpha\left(\mathbf{k}\right)\right]\label{eq:m22k}\\
m_{23}\left(\mathbf{k}\right) & =\left(U-\mu\right)I_{23}\left(\mathbf{k}\right)+2tI_{33}\left(\mathbf{k}\right)+4t\alpha\left(\mathbf{k}\right)I_{\bar{\pi}c_{s}}^{\alpha}\label{eq:m23k}\\
m_{33}\left(\mathbf{k}\right) & =-\mu I_{33}\left(\mathbf{k}\right)+2dtI_{\kappa_{s}c_{s}^{\dagger}}\left(\mathbf{k}\right)+UI_{\eta_{s}c_{s}^{\dagger}}\left(\mathbf{k}\right)\label{eq:m33k}\end{aligned}$$ where $\Delta=C_{\xi\xi}^{\alpha}-C_{\eta\eta}^{\alpha}$ is the difference between upper and lower intra-Hubbard-subband contributions to the kinetic energy and $p=\frac{1}{4}\left(\chi_{0}^{\alpha}+3\chi_{s}^{\alpha}\right)-\chi_{p}^{\alpha}$ is a combination of the nearest-neighbor charge-charge $\chi_{0}^{\alpha}=\left\langle n^{\alpha}\left(i\right)n\left(i\right)\right\rangle $, spin-spin $\chi_{s}^{\alpha}$ and pair-pair $\chi_{p}^{\alpha}=\left\langle \left[c_{\uparrow}\left(i\right)c_{\downarrow}\left(i\right)\right]^{\alpha}c_{\downarrow}^{\dagger}\left(i\right)c_{\uparrow}^{\dagger}\left(i\right)\right\rangle $ correlation functions.
Self-consistency and Algebra constraints
----------------------------------------
By restricting $I_{\kappa_{s}c_{s}^{\dagger}}\left(\mathbf{k}\right)$ and $I_{\eta_{s}c_{s}^{\dagger}}\left(\mathbf{k}\right)$ to just the local and the nearest-neighbor terms [@Avella_14], we have $$m_{33}\left(\mathbf{k}\right)\cong-\mu I_{33}\left(\mathbf{k}\right)+\bar{m}_{33}^{0}+\alpha\left(\mathbf{k}\right)\bar{m}_{33}^{\alpha}\label{eq:m33k_approx}$$ and we can use a couple of Algebra constraints[@Theory; @Avella_11a; @Avella_14] to compute $\bar{m}_{33}^{0}$ and $\bar{m}_{33}^{\alpha}$. $I_{\bar{\pi}c_{s}}^{\alpha}$ can be fixed in the very same manner [@Avella_14]. For the sake of consistency, we also neglect the $\beta\left(\mathbf{k}\right)$ and $\eta\left(\mathbf{k}\right)$ terms in $I_{33}\left(\mathbf{k}\right)$ [@Avella_14]. More details can be found in [@Avella_14]. We can recognize the following Algebra Constraints $$\begin{aligned}
C_{\xi\xi} & =1-n+D\label{eq:Cxixi}\\
C_{\eta\eta} & =\frac{n}{2}-D\label{eq:Cetaeta}\\
C_{\xi\eta} & =0\label{eq:Cxieta}\\
C_{\xi c_{s}} & =3C_{\xi c}^{\alpha}\label{eq:Cxics}\\
C_{\eta c_{s}} & =0\label{eq:Cetacs}\end{aligned}$$ where $D=\left\langle n_{\uparrow}\left(i\right)n_{\downarrow}\left(i\right)\right\rangle $ is the double occupancy. These relations lead to the following very relevant ones $$\begin{aligned}
n & =2\left(1-C_{\xi\xi}-C_{\eta\eta}\right)\label{eq:n}\\
D & =1-C_{\xi\xi}-2C_{\eta\eta}\label{eq:D}\end{aligned}$$
On the other hand, we can compute $\chi_{0}^{\alpha}$, $\chi_{s}^{\alpha}$, $\chi_{p}^{\alpha}$ and $f_{s}$ by operatorial projection, which is equivalent to the well-established one-loop approximation [@Theory; @Avella_11a] for same-time correlations functions $$\begin{aligned}
\chi_{0}^{\alpha} & \approx n^{2}-2\frac{I_{11}\left(C_{c\eta}^{\alpha}\right)^{2}+I_{22}\left(C_{c\xi}^{\alpha}\right)^{2}}{C_{\eta\eta}}\label{eq:chia0}\\
\chi_{s}^{\alpha} & \approx-2\frac{I_{11}\left(C_{c\eta}^{\alpha}\right)^{2}+I_{22}\left(C_{c\xi}^{\alpha}\right)^{2}}{2I_{11}I_{22}-C_{\eta\eta}}\label{eq:chias}\\
\chi_{p}^{\alpha} & \approx\frac{C_{c\xi}^{\alpha}C_{\eta c}^{\alpha}}{C_{\eta\eta}}\label{eq:chiap}\\
f_{s} & \approx-\frac{1}{2}C_{c\xi}^{\alpha}-\frac{3}{4}\chi_{s}^{\alpha}\left(\frac{C_{c\xi}^{\alpha}}{I_{11}}-\frac{C_{c\eta}^{\alpha}}{I_{22}}\right)\nonumber \\
& -2\frac{C_{c\xi}^{\alpha}}{I_{11}}\left(C_{c\xi}^{\alpha^{2}}-\frac{1}{4}C_{c\xi}\right)-2\frac{C_{c\eta}^{\alpha}}{I_{22}}\left(C_{c\eta}^{\alpha^{2}}-\frac{1}{4}C_{c\eta}\right)\label{eq:fs}\end{aligned}$$ Summarizing, we can fix the unknowns $I_{\bar{\pi}c_{s}}^{\alpha}$, $\bar{m}_{33}^{0}$, $\bar{m}_{33}^{\alpha}$, $\mu$, $\chi_{0}^{\alpha}$, $\chi_{s}^{\alpha}$, $\chi_{p}^{\alpha}$ and $ $$f_{s}$ through the set of equations (\[eq:Cxieta\]), (\[eq:Cxics\]), (\[eq:Cetacs\]), (\[eq:n\]), (\[eq:chia0\]), (\[eq:chias\]), (\[eq:chiap\]) and (\[eq:fs\]).
Results\[sec:Results\]
======================
Double occupancy and chemical potential: solution assesment\[sec:Results:D\_mu\]
--------------------------------------------------------------------------------
![Double occupancy $D$ (left) and scaled chemical potential $\mu-\nicefrac{U}{2}$ (right) as functions of the filling $n$ for $U=4$ and $T=\nicefrac{1}{6}$ for COM(3p) (black lines), COM(2p,$p>0$) (dashed red line), COM(2p,$p<0$) (dotted blue line), Hubbard I (dot-dashed green line), Roth (dot-dot-dashed magenta line) and DMFT [@Capone] (dash-dotted purple line). Analytical results are compared with $12\times12$-site qMC [@Moreo_90] and $2$-site DCA [@Sangiovanni] numerical data (red and blue hollow circles, respectively).\[fig:1\_2\]](fig1){width="7cm"}
![Double occupancy $D$ (left) and scaled chemical potential $\mu-\nicefrac{U}{2}$ (right) as functions of the filling $n$ for $U=4$ and $T=\nicefrac{1}{6}$ for COM(3p) (black lines), COM(2p,$p>0$) (dashed red line), COM(2p,$p<0$) (dotted blue line), Hubbard I (dot-dashed green line), Roth (dot-dot-dashed magenta line) and DMFT [@Capone] (dash-dotted purple line). Analytical results are compared with $12\times12$-site qMC [@Moreo_90] and $2$-site DCA [@Sangiovanni] numerical data (red and blue hollow circles, respectively).\[fig:1\_2\]](fig2){width="7cm"}
In Fig. \[fig:1\_2\], we report the behavior of the double occupancy $D$ (left panel) and of the scaled chemical potential $\mu-\nicefrac{U}{2}$ (right panel) as functions of the filling $n$ for $U=4$ and $T=\nicefrac{1}{6}$. It is evident the very good agreement in the whole range of filling $n$ between COM(3p) and the $12\times12$-site qMC [@Moreo_90] and $2$-site DCA [@Sangiovanni] numerical data. The double occupancy $D$ features a very elaborated behavior presenting a continuos, but well defined, change of slope on approaching half filling. COM(3p) correctly catches this feature, while all other presented solutions do not manage to achieve the same level of agreement over the whole range of filling. Hubbard I and Roth solutions report values of the $D$ extremely far from the numerical ones and always much smaller than these latter, showing a tendency to an excess of correlations present in such solutions. DMFT [@Capone] performs extremely well, with respect to numerical data, at low-intermediate values of filling, but at intermediate-high ones features values of $D$ larger than the numerical ones. This is a clear evidence of a lack of correlations for this value of $U$. COM(2p, $p<0$) performs really very well too at low-intermediate values of filling, but on increasing $U$ it shows an excess of correlations close to half filling (it is actually insulating for any finite value of $U$ at half filling). In COM(2p, $p>0$), it is evident a complete suppression of $D$ at low values of the filling as well as a small, but visible, discrepancy in the slope close to half filling. COM(3p) evidently has the capability to correctly interpolate between the two COM(2p) solutions sticking to COM(2p, $p<0$) at low-intermediate values of filling and even improving on COM(2p, $p>0$) at intermediate-high values of filling. The DCA data for the chemical potential show a concavity in proximity of half filling that is correctly caught by COM(3p) and COM(2p, $p>0$) and not by COM(2p, $p<0$), Hubbard I and Roth solutions. Roth solution actually reports a rather evident region of thermodynamic instability, $\frac{d\mu}{dn}<0$, close to half filling. As a matter of fact, $U=4$ induces already quite strong electronic correlations: the chemical potential gets ready to open a gap for higher values of $U$ and $n=1$. COM(2p, $p<0$), Hubbard I and Roth solutions place themselves always on the strongly correlated side and report values of $\mu$ quite far from the numerical ones: their particle counting - actual effective filling - is definitely far from the exact one. DMFT [@Capone] solution does not catch the correct concavity again showing a lack of correlations for this value of $U$, but it features values of $\mu$ very close to the numerical ones in the whole range of filling $n$ although not so close as COM(3p) ones in proximity of half-filling, which is the most interesting region.
Quasi-particle energy dispersions: solution characterization\[sec:Results:Ek\]
------------------------------------------------------------------------------
In Fig. \[fig:3\_4\], we report the energy bands $E_{m}\left(\mathbf{k}\right)$ along the principal directions of the first Brillouin zone ($\Gamma=(0,0)$ $\to$ $S=(\nicefrac{\pi}{2},\nicefrac{\pi}{2})$ $\to$ $M=(\pi,\pi)$ $\to$ $X=(\pi,0)$ $\to$ $Y=(0,\pi)$ $\to$ $\Gamma=(0,0)$) at $T=\nicefrac{1}{6}$, $U=4$ and two different values of the filling $n=0.2$ (left panel) and $n=0.9$ (right panel). At $n=0.2$, it is evident that the occupied bands are almost identical across all reported COM solutions. Actually, COM(3p) is characterized by a small, but finite, occupation of its LHB, besides the occupation of its central band (CB), which is the band coinciding with the COM(2p) LHBs. This can be understood in terms of the proximity of COM(3p) LHB to the chemical potential at the $M$ point. LHB is the only occupied band in COM(2p,$p<0$) at all finite values of $U$. At $n=0.9$, the occupied region in energy-momentum space across the three COM solutions is instead quite different, although some similarities can still be found. In particular, as regards the regions close to the chemical potential at the $\Gamma$ point and along the main anti-diagonal (the $X-Y$ line). COM(3p) CB, which was the main actor at low fillings, tends to systematically lose occupation in favor of the LHB. Close to half filling, this latter eventually exceeds the former in occupation and collects more and more of it on increasing $U$ while the CB depletes on approaching the metal-insulator transition. As regards COM(2p,$p>0$) instead, UHB plays a minor role all the way up to the metal-insulator transition. It collects a small fraction of the electronic occupation and only above a certain intermediate value of the filling. It is evident that COM(3p) CB is still almost pinned to the chemical potential along the main anti-diagonal (the $X-Y$ line); the van Hove singularity lies little below the Fermi level. Accordingly, changing the filling in this region of low doping (from $n=0.85$ to $n=1$) has mainly the effect to induce a transfer of spectral weight between the bands and between their components in terms of fields of the basis, as one would expect in a strongly correlated regime, rather than shifting the chemical potential more or less rigidly within the bands, as it could be expected at small fillings and weak interactions. It is also evident that the LHB has still a minor role with respect to the CB, which collects the vast majority of the occupied states. It is worth noting that the spin-spin correlations are already present, but not yet so strong to determine the reduction of the bandwidth in the energy-momentum space region shared by CB and LHB. It is worth noticing that COM(3p) bands are quite close to COM(2p,$p<0$) ones.
![Energy bands $E_{m}\left(\mathbf{k}\right)$ along the principal directions of the first Brillouin zone ($\Gamma=(0,0)$ $\to$ $S=(\nicefrac{\pi}{2},\nicefrac{\pi}{2})$ $\to$ $M=(\pi,\pi)$ $\to$ $X=(\pi,0)$ $\to$ $Y=(0,\pi)$ $\to$ $\Gamma=(0,0)$) at $T=\nicefrac{1}{6}$, $U=4$ and two different values of the filling $n=0.2$ (left) and $n=0.9$ (right) for COM(3p) (black line), COM(2p,$p>0$) (red line) and COM(2p,$p<0$) (blue line).\[fig:3\_4\]](fig3){width="8cm"}
![Energy bands $E_{m}\left(\mathbf{k}\right)$ along the principal directions of the first Brillouin zone ($\Gamma=(0,0)$ $\to$ $S=(\nicefrac{\pi}{2},\nicefrac{\pi}{2})$ $\to$ $M=(\pi,\pi)$ $\to$ $X=(\pi,0)$ $\to$ $Y=(0,\pi)$ $\to$ $\Gamma=(0,0)$) at $T=\nicefrac{1}{6}$, $U=4$ and two different values of the filling $n=0.2$ (left) and $n=0.9$ (right) for COM(3p) (black line), COM(2p,$p>0$) (red line) and COM(2p,$p<0$) (blue line).\[fig:3\_4\]](fig4){width="8cm"}
Momentum resolved quantities\[sec:Results:nk\_Dk\_chiask\]
----------------------------------------------------------
----------------------------------- ----------------------------------- ------------------------------------
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}
----------------------------------- ----------------------------------- ------------------------------------
Given the decomposition of the momentum-dependent correlation functions per band reported in Eqs. \[eq:Ck\] and \[eq:Ck\_m\], it is possible to define a similar decomposition for any quantity that can be expressed in terms of correlation functions of the chosen operatorial basis, that is for any quantity computable within the reported approximation. In particular, we have the following expressions for the filling $n$, the double occupancy $D$ and the nearest-neighbor spin correlation function $\chi_{s}^{\alpha}$ $$\begin{aligned}
n & =2\sum\limits _{m=1}^{3}\frac{1}{N}\sum_{\mathbf{k}}n^{\left(m\right)}\left(\mathbf{k}\right)\end{aligned}$$ $$\begin{aligned}
D & =\sum\limits _{m=1}^{3}\frac{1}{N}\sum_{\mathbf{k}}D^{\left(m\right)}\left(\mathbf{k}\right)\end{aligned}$$ $$\begin{aligned}
\chi_{s}^{\alpha} & =\sum\limits _{m=1}^{3}\frac{1}{N}\sum_{\mathbf{k}}\chi_{s}^{\alpha\left(m\right)}\left(\mathbf{k}\right)\end{aligned}$$ where $$\begin{aligned}
n^{\left(m\right)}\left(\mathbf{k}\right) & =\frac{1}{2}f_{\mathrm{F}}\left(E_{m}\left(\mathbf{k}\right)\right)\sigma_{cc}^{\left(m\right)}\left(\mathbf{k}\right)\\
D^{\left(m\right)}\left(\mathbf{k}\right) & =f_{\mathrm{F}}\left(E_{m}\left(\mathbf{k}\right)\right)\sigma_{22}^{\left(m\right)}\left(\mathbf{k}\right)\\
\chi_{s}^{\alpha\left(m\right)}\left(\mathbf{k}\right) & =\frac{2\alpha\left(\mathbf{k}\right)f_{\mathrm{F}}\left(E_{m}\left(\mathbf{k}\right)\right)}{2I_{11}I_{22}-C_{\eta\eta}}\times\nonumber \\
& \times\left(I_{11}C_{c\eta}^{\alpha}\sigma_{2c}^{\left(m\right)}\left(\mathbf{k}\right)+I_{22}C_{c\xi}^{\alpha}\sigma_{1c}^{\left(m\right)}\left(\mathbf{k}\right)\right)\end{aligned}$$
In Fig. \[fig:5\_10\], we report the momentum-distribution function per band and spin $n^{\left(m\right)}\left(\mathbf{k}\right)$ (left column), the double-occupancy components per band $D^{\left(m\right)}\left(\mathbf{k}\right)$ (center column) and the nearest-neighbor spin correlation function components per band $\chi_{s}^{\alpha\left(m\right)}\left(\mathbf{k}\right)$ (right column) along the principal directions of the first Brillouin zone ($\Gamma=(0,0)$ $\to$ $S=(\nicefrac{\pi}{2},\nicefrac{\pi}{2})$ $\to$ $M=(\pi,\pi)$ $\to$ $X=(\pi,0)$ $\to$ $Y=(0,\pi)$ $\to$ $\Gamma=(0,0)$) at $T=\nicefrac{1}{6}$, $U=4$ and two different values of the filling $n=0.2$ (top row) and $n=0.9$ (bottom row). Components from not reported bands are zero or definitely negligible.
At $n=0.2$, $n^{\left(m\right)}\left(\mathbf{k}\right)=\frac{1}{2}\mathcal{F}_{\mathbf{k}}\left[\left\langle c^{\dagger}(\mathbf{i})\cdot c(\mathbf{j}\right\rangle )\right]_{m}$ shows that reported COM bands have a similar and quite *ordinary* occupations except for the region close to $M$ that is occupied only for COM(3p) and COM(2p,$p<0$) LHBs. This is the result of the peculiar shape of such bands (see Fig. \[fig:3\_4\]) that closely recalls the bending driven by antiferromagnetic fluctuations and the simultaneous occupation of $\Gamma$ and $M$ points. What is really surprising is the fact that the major, almost the only, contribution to the double occupancy ($D^{\left(m\right)}\left(\mathbf{k}\right)=\frac{1}{2}\mathcal{F}_{\mathbf{k}}\left[\left\langle \eta^{\dagger}(\mathbf{i})\cdot\eta(\mathbf{j}\right\rangle )\right]_{m}$) comes just from the this last region in momentum (close to $M$ point). This is really counterintuitive as one expects those branches of energy-momentum dispersion to have such a shape because of the strong antiferromagnetic fluctuations and to be the main seat of electrons contributing to single occupation and, therefore, with well-formed spin momenta. As a matter of fact, the contribution to $D^{\left(m\right)}\left(\mathbf{k}\right)$ close to $M$ should be correctly read as the main (actually the only) contribution to the kinetic energy coming from the $\eta$ Hubbard operators, that is from those electrons moving between double occupied sites. In fact, this situation well explains the almost identical value of the double occupancy for COM(3p) and COM(2p,$p<0$) at this filling, where the motion between doubly occupied sites is allowed, as well as the almost negligible value for COM(2p,$p>0$), where the motion between doubly occupied sites is definitely negligible as it is confined to the UHB, which is empty for this value of the filling. The decomposition of $\chi_{s}^{\alpha\left(m\right)}\left(\mathbf{k}\right)$ shows the expected negative contribution by the electrons close to $\Gamma$ point and an almost negligible positive contribution by those close to $M$ point. At such low value of the filling, we can expect very weak spin correlations and the contribution close to $\Gamma$ point is not so large as well as the whole momentum dependence very little structured.
At $n=0.9$, $n^{\left(m\right)}\left(\mathbf{k}\right)$ shows how much COM bands differ in occupation between the three reported solutions for a value of the filling where quite intense correlations are expected. COM(3p) features an occupation of the CB close to $\Gamma$ point quite reduced with respect to that sported at $n=0.2$, even if we take into account that it is now spanning a quite larger region in momentum and the main anti-diagonal (the $X-Y$ line) is somewhat filled too. This can be explained by noting that the LHB, which at $n=0.2$ was filled only close to $M$ point, features now a quite relevant occupation spanning all over the first Brillouin zone and, in particular, at the $M$ point and along the main anti-diagonal (the $X-Y$ line). Accordingly, we expect the physics of COM(3p) solution to be determined by both bands at the same time and on almost equal footing. Overall, COM(2p,$p<0$) occupation is very similar to the COM(3p) one although concentrated in the only occupied band, the LHB. As a matter of fact, COM(2p,$p<0$) LHB seems to interpolate between the LHB and the CB of COM(3p) showing once more the very strict connections between these two solutions. COM(2p,$p>0$) features instead similar occupations for the two bands except for the extension towards the main anti-diagonal (the $X-Y$ line) of the almost completely filled LHB. The region in momentum close to $M$ remains anyhow empty marking the greatest difference to the other two solutions. Coming to $D^{\left(m\right)}\left(\mathbf{k}\right)$, COM(3p) solution features again a complementary presence between CB and LHB with the exception of the main anti-diagonal (the $X-Y$ line) where the more marked difference reported for $n^{\left(m\right)}\left(\mathbf{k}\right)$ is greatly reduced. Bare looking at the values of $D^{\left(m\right)}\left(\mathbf{k}\right)$ reported along these principal direction, COM(2p,$p<0$) should have an overall value of $D$ quite similar to that of COM(3p), but this is quite not right and can be explained by looking at the region in momentum along the main anti-diagonal (the $X-Y$ line) and close to it (even along the main diagonal: the $\Gamma-M$ line). Although, COM(2p,$p<0$) LHB is lying over the chemical potential in this region of momentum, the double occupancy contribution is definitely negligible marking a huge difference to the CB of COM(3p) that occupies the same region in momentum-frequency space. This is one of the net improvements of the COM(3p) solution over COM(2p,$p<0$) one; an improvement that reflects in many other physical quantities. COM(2p,$p>0$) contributions come both (for LHB and UHB) from the region in momentum close to $\Gamma$ clearly showing that although the overall value of $D$ is quite similar, clearly by accident, between COM(3p) and COM(2p,$p>0$), its physical origin, in terms of quasi-particle contributions and of their momentum-frequency dispersions, is very different and explains the quite different behavior in terms of slope of $D$ as a function of the filling $n$. Let us come to $\chi_{s}^{\alpha\left(m\right)}\left(\mathbf{k}\right)$. As regards COM(3p), the CB brings a much larger contribution, with respect to $n=0.2$, that extends along the main anti-diagonal (the $X-Y$ line). LHB contribution is much smaller, but definitely larger than at $n=0.2$ and negative. Accordingly, it is just the CB, which originates from the third basic field describing the nearest-neighbor antiferromagnetic fluctuations, to bring the larger contribution as one would have expected. Therefore, having such a field in the basis results as one of the main ingredients in order to get such a good performance in comparing this solution with numerical ones [@Avella_14]. COM(2p,$p<0$) contributions to $\chi_{s}^{\alpha\left(m\right)}\left(\mathbf{k}\right)$ all come from the only occupied band, the LHB, and once more seem to mime the overall behavior of COM(3p). COM(2p,$p>0$) has a completely different behavior. In particular, the contribution of the LHB close to the $\Gamma$ point is quite difficult to understand: instead of increasing in absolute value towards the $\Gamma$ point, it decreases leading to the presence of a maximum absolute value for a value of momentum that coincide with the Fermi surface of the related COM(2p,$p>0$) UHB.
Summary\[sec:Summary\]
======================
In this manuscript, we have first recollected the main analytical expressions defining a recently proposed, within the framework of the Composite Operator Method[@Theory; @Avella_11a], three-pole solution for the two-dimensional Hubbard model [@Avella_14]. Together with the two Hubbard fields, well describing the physics at the energy scale of $U$, the presence of a third field, embedding the strong antiferromagnetic fluctuations, has enormously boosted the performance of COM(3p) solution with respect to COM(2p) ones. The extremely positive comparison with the data obtained by different numerical methods for momentum-integrated quantities (e.g. local properties) as functions of all model parameters (filling, on-site Coulomb repulsion and temperature) as well as for the energy bands of the system [@Avella_14] makes this solution extremely interesting to be analyzed further. Here, we have reported a summary of the behavior of the basic local quantities - the double occupancy and the chemical potential -, together with the quasi-particle energy dispersions definitely necessary to guide the subsequent analysis, which is the main focus of the present manuscript: the study of the momentum-resolved components of filling (i.e. the momentum distribution function), double occupancy and nearest-neighbor spin correlation function. The analysis has been extended to COM(2p) solutions that have been used as primary reference as in the main paper [@Avella_14]. Analyzing the momentum-resolved quantities, it emerges very clearly the role played by the third field with respect to the two Hubbard ones in determining the behavior of many relevant quantities and allowing to get the extremely good comparison with numerical results. In particular, the proximity between COM(3p) and COM(2p,$p<0$) solutions is further reinforced and better understood giving a guideline to further improve and, possibly, optimize the application of the COM to the Hubbard model with the choice of a fourth field solving the few remaining issues with COM(3p)[@Avella_14].
The author wishes to thank Gerardo Sica for many insightful discussions.
[75]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, ****, ().
, , , , , , , , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , , , ****, ().
, , , , , , , , , ****, ().
, ****, ().
, ****, ().
, , , , , , , , , ****, ().
, ****, ().
, ****, ().
, , , , , , , ****, ().
, , , , , , , , , , ****, ().
, , , , , , , ****, ().
, , , ****, ().
, , , ****, (), .
, ****, ().
, eds., **, vol. of ** (, ).
, eds., **, vol. of ** (, ).
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, (), .
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, , , , , ****, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , , ****, ().
, , , ****, ().
, in **, edited by (, ), vol. of **, p. , <http://dx.doi.org/10.1007/978-3-642-21831-6_4>.
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, .
, , , , , ****, ().
, .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a machine learning based information retrieval system for astronomical observatories that tries to address user defined queries related to an instrument. In the modern instrumentation scenario where heterogeneous systems and talents are simultaneously at work, the ability to supply with the right information helps speeding up the detector maintenance operations. Enhancing the detector uptime leads to increased coincidence observation and improves the likelihood for the detection of astrophysical signals. Besides, such efforts will efficiently disseminate technical knowledge to a wider audience and will help the ongoing efforts to build upcoming detectors like the LIGO-India etc even at the design phase to foresee possible challenges. The proposed method analyses existing documented efforts at the site to intelligently group together related information to a query and to present it on-line to the user. The user in response can further go into interesting links and find already developed solutions or probable ways to address the present situation optimally. A web application that incorporates the above idea has been implemented and tested for LIGO Livingston, LIGO Hanford and Virgo observatories.'
author:
- Nikhil Mukund
- Saurabh Thakur
- Sheelu Abraham
- 'A. K. Aniyan'
- Sanjit Mitra
- Ninan Sajeeth Philip
- Kaustubh Vaghmare
- 'D. P. Acharjya'
bibliography:
- 'Mukund\_HeyLIGO\_2017.bib'
title: Information Retrieval and Recommendation System for Astronomical Observatories
---
Introduction {#sec:intro}
============
Data mining in the big data framework often encounters difficulty in both extracting the relevant information from the data and in coming up with meaningful interpretations in a highly reliable fashion [@fan2014challenges; @wu2014data; @khan2014big]. In many situations, the data comes in a format which is not suitable to store in relational databases of coherent hierarchy [@stephens2015big]. The methods in which the data is stored and associated with different entities also pose the challenge in mining required information from it. For example, in a gravitational wave observatory, there will be a core science dataset with plenty of meta-data on the observation and a variety of other auxiliary datasets collected from various sensors and actuators.
Even though data mining methods like association analysis, clustering and other machine learning techniques exist, the presentation of unstructured data into these algorithms and inference generation is not a trivial task [@han2011data]. Generation of insights from big data with recommendation systems which are based on learning from unstructured text data [@pazzani2007content] tackle these challenges at large [@lavalle2011big; @hu2014toward]. Descriptive recommendations and information retrieval [@sigurbjornsson2008flickr; @gretzel2004tell] have recently have recently gained popularity and have been applied to areas like travel recommendation systems [@gretzel2004tell] and content personalization systems [@liang2006personalized]. Besides commercial applications, text summarization based content recommendation [@hassan2009content] is an interesting area which has high level of applicability in different areas of science and research [@miner2012practical; @2017arXiv170505840K]. Unlike conventional rank based search systems, these do not perform topical modeling and rank topics of recurring interest [@zoghbi2013words]. Topical modeling is usually done for retrieving information from a single website with multiple topics. The challenge is when different topics in a single site may be weakly linked to each other [@cointet2010local]. While there could be already known relations among different entities, the process of data mining and better data representation can reveal the latent unforeseen linkages among different topical entities [@behrens2006information].
Large science projects, especially astronomical observatories, have plenty of data about telescope operations, scheduling, maintenance and general observational activities all logged in text form. Over the years, these logbook entries will accumulate almost all the aspects of the instruments in the observatory. Although the key technologies are changing rapidly, the fundamental principles involved in construction and maintenance at observatories are getting altered at a less rapid rate. This fact necessitates the need for keeping a record of activities carried out over the years for prompt diagnostics. Projects like SKA, TMT, LIGO, SALT, JWST also require extensive internal coordination. These typically are a collaboration consisting of thousands of scientists whose research can span area like instrument fabrication, installation, commissioning, characterization, maintenance, upgrade, data analysis and parameter estimation. Often their time span spread across few decades and thus generate information whose volume and complexity cannot be handled effectively by traditional search engine backed information processing tools. On the positive side, analysis of such big data volumes can yield powerful insights into the inherent trends and fluctuations within the concerned project.
In this paper, we demonstrate natural language processing(NLP) backed knowledge rediscovery [@ricci2011introduction] by making use of the open source logbook data from the Laser Interferometric Gravitational Observatory (LIGO). This is a novel approach to observational astronomy, and the developed software is made available for the public through a web application named **Hey LIGO[^1]**. We also show the application of descriptive content based recommendations to compare common issues among multiple observatories. These methods are generally scalable and will be very useful in the event of upcoming projects like the Square Kilometer Array (SKA) and upcoming LIGO-India detector.
We have organized the paper as follows: in section two we describe the methodology adopted to convert raw data to useful and representable information. Section three provides the details of data used in our analysis. Features of the recommendation system are outlined in section four. In the last part, we apply the scheme to various gravitational wave observatories around the world and discuss the results obtained.
Contextual learning of unstructured Data
========================================
Structured data is highly organized and usually resides in a relational database schema. But unstructured data refer to information that does not follow the traditional database scheme. For example, e-mails, web pages, business documents, FAQ’s, etc. are some examples of unstructured data. They include text and even multimedia content. So processing of such information is an energy and time consuming task.
This section briefly describes the manner in which the unstructured textual data is acquired, processed and finally given structure. Moreover, it also enumerates the various steps involved in the development of a machine learning model which is used to differentiate between the available textual data points contextually. Finally, the model is used to perform clustering over all the textual data, thus, adding a structure for ease of access. Figure \[fig:flowchart\] shows the schematic representation of our web interface that is used to implement the scheme.
The unstructured dataset that we use is in the form of textual web pages. These pages have an identical HTML structure and defined attributes for every data point. Due to this same structure and open source nature of the web pages, it is possible to write a script which can extract each attribute from the HTML source code and organize the complete information into a data frame. A data frame is a tabular structure with columns as attributes and rows as individual data points. The first part of our algorithm does data acquisition using the python package “Beautiful Soup” to retrieve information from web pages by searching through all the new posts and related data. It saves it into relevant files for later utilization when the need arises.
Once the data is stored locally, non-essential attributes are removed, textual time stamps are converted to system timestamps, duplicate data points are removed/combined, and the resulting data is passed on to the text processing unit. A vocabulary for our data is generated by converting the unstructured data into stem words. For that, we have removed all special characters and punctuations such as $!, @, \#, \$, \%, * ,\&, ',",(,) $ etc. All non-English words and other HTML tags, URLs are also excluded from the data. The text is then tokenized [@huang2007rethinking] by splitting the strings of text into a list of words called tokens. To reduce the redundancy in the vocabulary, it is important to convert the related token forms and their derivatives to a common base stem by a process known as stemming [@smirnov2008overview].
In the process of embedding, the textual data is converted into vectors which can be easily handled by the computer [@li2015word]. There exist various embedding algorithms like One-Hot Encoding [@harris2012digital] and Term Frequency Inverse Document Frequency (TFIDF) [@leskovec2014mining] etc for this process. Not all methods can capture the contextual differences between the words. However, a recent breakthrough in the field of Natural Language Processing incorporates neural networks that can learn the vector values for each word by iterating over the text multiple times using a gradient based algorithm [@2013arXiv1301.3781M; @2013arXiv1310.4546M]. [@bengio2003neural] have coined the term word embeddings with a neural language model to train them with the model’s parameter.
One of the commonly used tools to convert words into vectors is word2vec described in @2013arXiv1301.3781M. Word2vec has a single hidden layer, fully connected neural network that takes a large text corpus as input and produces a higher dimensional vector for each unique word in the corpus. Words which share common contexts in the corpus are located close to each other in the vector space. Word2vec models do not consider word order and can capture semantic information between words in a very efficient way [@ling2015two]. With the help of Word2vec embeddings, a computer can differentiate between words of different types. Word2vec implements two computationally less expensive models known as Continuous Bag of Words (CBOW) and a Skip Gram model [@2013arXiv1301.3781M] to learn word embeddings. The representation of a corpus of text or an entire document in the form of a list of words (Multiset) is referred to as Bag of Words representation [@markov2007data]. The algorithm essentially tries to predict the target based on a set of context words [@2013arXiv1301.3781M; @2013arXiv1310.4546M].
![Skip-Gram Model[]{data-label="fig:skipgm"}](Fig2.pdf)
The model that we have used in this work is the Skip Gram model. The basic architecture of the skip gram model is shown in Figure \[fig:skipgm\]. This representation is similar to CBOW model, but instead of predicting the target word, it predicts the context words based on a given target word [@2013arXiv1310.4546M]. Thus, the model maximizes the probability for classification of a word based on another word in the same sentence [@2013arXiv1301.3781M]. Thus the vector representation is capable of capturing the semantic meanings of the words from a sequence of training words $w1, w2, ..., wT$ and their contexts $c$. The algorithm can be briefed as follows. First the words are applied to as an input to a log-linear classifier where the objective is to maximize the average log probability given by,
$$\label{eq:skpg}
\mathcal{L} = \frac{1}{T}\sum_{t=1}^{T} \sum_{-c\leq j \leq c,j\not\equiv 0} log \ P(w_{t+j} | wt)$$
Larger value of $c$ can result in higher accuracy but requires more training time [@2013arXiv1310.4546M]. To obtain the output probability, $\mathit{P(w_{o}|w_{i})}$, the model estimates a matrix which maps the embeddings into $V$-dimensional vector $O_{w_{i}}$. Thus the probability of predicting the word $w_{o}$ given the word $w_{i}$ is defined using the softmax function:
$$\mathit{P(w_{o}|w_{i})} = \frac{\exp{(O_{w_{i}}(w_{o}))}}{\sum_{w \epsilon \mathit{V}} \exp{(O_{w_{i}}(w))}}$$
where $V$ is the number of words in the vocabulary [@ling2015two; @2013arXiv1310.4546M]. But this formulation is computationally intensive for larger vocabularies. This is solved in word2vec by using the hierarchical softmax function [@morin2005hierarchical] or with negative sampling approach [see this for more details; @2014arXiv1402.3722G].
After embedding all words, every data point is represented as the average of all the word vectors of the words present in it. A Nearest Neighbor Algorithm [@Andoni:2006:NHA:1170136.1170526] is then used to cluster these data point vectors to respective clusters efficiently. The optimal number of clusters is estimated iteratively until it is observed that the accuracy peaked, which in our case was found to be $1/5^{th}$ of the vocabulary of our model. We used the Python implementation of $Scikit-learn$ package for doing the nearest neighbor algorithm.
Even after NLP classification, we observed that quite a number of relevant posts were left out unobserved and so we added one more layer of processing by analyzing the overall emotional content of the reports. We used the AFINN lexicon [@AFINN] consisting of a collection of 2477 words each with an associated integer value ranging between -5 to +5 representing transition from negative to positive sentiment. Modifying the word valence and appending the lexicon with technical words that better represent the associated sentiments was found to provide better results. For example, LIGO specific application will associate terms like ’lockloss’ and ’scatter noise’ with negative sentiment while ’new filter installed’ would mean something positive.
The prototype is designed so that the users can query for information through a web interface. The stem words in the query are identified and the vector is projected into the previously modeled word vector space. The nearest neighbor model retrieves the top neighbors for the query vector, and are then displayed as the search results on the web interface. The search result is then filtered to check for the presence of the query words in either the title or in its content to weed out false positives. In Figure \[fig:respg\], we have shown a simple search query displayed on the web interface. The different features incorporated in the web page is described in Section \[sec:HFunc\].
Gravitational Wave Observatories {#sec:Test}
=================================
GW interferometers (IFOs) have been in operation for the past few decades and have made the first direct detection of merging binary black holes [@PhysRevLett.116.061102]. The complex nature of this multi-physics experiment requires scientists from multiple domain of expertise to work together and share information. Rigorous commissioning and characterization efforts have been carried over a span of two decades to reach the current level of sensitivity. LIGO, Virgo, GEO600 and KAGRA archives most of the activities happening at the sites through their logbooks. These may be complete from installation activities to noise hunting and mitigation works carried out during the lifespan of the observatory. Although there are site specific issues, often they encounter problems of similar nature and employing solutions that worked at the other sites may be a good strategy to start with. Also it is not uncommon to see previously fixed issues to reappear at a later time where the time scales could be of few months to years. This happens due to recurring environmental fluctuations and configuration changes in the detector. Since the current GW detectors aim at coincident detection of events, joint uptime of the instruments are crucial. This is more significant because the probability of detection scales linearly with observation time and cubically with sensitivity of the instrument.
The effective information extraction and processing of logbook information as envisaged here is expected to help in making better decisions pertaining to detector maintenance. For example, identifying the subsystems that could possibly get affected during instrument upgrade will be vital in scheduling and coordinating tasks among sub-groups involved. Similarly, long term tracking of a issue can be carried out to see if the various overhauling attempts indeed lead to an improvement in performance which correlates with lesser number of related posts.
In the case of GW interferometers, day-to-day activities are recorded using web interfaces known as Logbooks. It is mandatory for the reports to have a title, section, task, details and author details. Although anyone can view the reports, only users with valid credentials can login to add logs and additional supporting files like measurement figures, sensor data, codes etc. It is also possible to to add comments and carry out further discussion on any of the log book entries. Details of retrieved information are given below in Table \[tab:logs\]
Observatory Logbook Entries Contributors Timespan Dictionary Size Clusters
--------------------- ----------------- -------------- ----------- ----------------- ----------
LIGO Livingston[^2] 24351 261 2010-2017 2273 455
LIGO Hanford[^3] 24968 237 2010-2017 2713 543
Virgo[^4] 34592 660 2010-2017 5026 1005
Hey LIGO Functionalities {#sec:HFunc}
========================
An open access NLP based web application implementation named **Hey LIGO** is developed and deployed to support the commissioning and characterization efforts at the GW observatories. It relies on the logbook data recorded since 2010 by scientists specialized in different aspects of the detector. Every query is answered by matching it with most relevant logbook entries sorted as per their closeness to the query term in the word-vector space. We further analyze the sentiment of the post and color code so that green indicates a positive outcome and red corresponds to something undesirable in the context of activities carried out at the detector. An image retrieval facility displays thumbnail of the figures attached to the sorted data simplifying the knowledge discovery process. Contextual data visualization across multiple detectors is carried out as shown in Fig \[timeline\] and Fig \[scatter\_pie\]. This feature lets the user to compare and see the trends in the searched keyword across different observatories.
Automatic check for new data entries is done periodically so that the NLP models are regularly updated. We track the volume of discussions happening on various topics and hence identify and rank the trending issues on a daily basis. Scientists involved with the project will mostly be interested in getting notified about specific issues that correlates with their domain of expertize and so the application only issues alerts to registered participants with matching interests. This targeted delivery will remove clutter and will ensure proper dissemination of information to the concerned people.
Code development is usually a tedious procedure wherein significant amount of time is spent on readability and re-usability so as to benefit a wider research community. Our application makes better use of this idea by auto-detecting and notifying the user about the presence of codes in the searched content. This feature we believe would simplify the procedure involved in result reproduction and its consequent independent verification.
{width="16cm"}
To check the performance of our application, we analyzed six months of logbook data from LIGO Livingston and compared the NLP results and the actual relevant entries. Table \[tab:acur\] gives the performance of the our implementation for certain set of randomly chosen keywords. In most cases,the false alarms occur at the tail end of the search results which represent neighbors of least relevance in the k-NN search. These can be removed either by setting a threshold on the similarity distance measure or by post-filtering the results by additionally comparing the content of each post. Currently, we have implemented the process of post-filtering to remove such post in the final web application. In the future we are planning to incorporate a mechanism that will make use of feedback received from the users and utilize it to improve the accuracy in retrieving relevant posts.
[|m[5em]{}|m[1.2cm]{}|m[1.2cm]{}|m[1.2cm]{}|m[1.2cm]{}|]{} Keyword & [Log book Entries]{} &\
& & & &\
Lock loss & 108 & 108 & 89 & 19\
Earthquake & 83 & 94 & 80 & 14\
& 62 & 65 & 58 & 7\
Guardian & 55 & 65 & 55 & 10\
Oplev & 63 & 61 & 48 & 13\
& 55 & 52 & 45 & 7\
Inferring from logbook entries
===============================
Once the relevant logbook entries are identified using the techniques mentioned above, their associated meta-data can be utilized to obtain several quantitative information about the topic of interest.
Trends within detectors
-----------------------
Below we briefly compare the trends obtained for few test search queries and briefly discuss the observed patterns. Although of similar configuration, the effect of various noises on each detector seems to be of a different nature. Variation in instrumental behavior and environmental effects due to geographical location will also influence the efficiency of implemented mitigation measures. Figuring out such details can positively speed up the commissioning activities of future detectors like LIGO-India.
1. ****
First plot from figure \[timeline\] shows the trends in posts related to installation work at each of the observatories. Activities picked up momentum in 2010 at LIGO and continued till the mid of 2014 after which testing and commissioning tasks started. Advanced Virgo seems to have started such activities in 2014 and carried on till the end of 2016.
2. ****
Jitter noise arising out of laser pointing fluctuations [@PhysRevD.93.112004], is sensitive to cavity alignments and angular mirror motions. It has been partly caused by the pre-stabilized laser (PSL) periscope motion induced by chiller water flow around PSL’s high power oscillator. Various efforts to understand it’s possible origin and subsequent efforts to subtract it from the data stream is reflected through the increased number of alogs at the Hanford detector compared to other sites. Commissioners performed online feed-forward noise subtraction using auxiliary witness channels which reduced the coupling significantly (@alogLLO30412 [@alogLLO30473; @alogLLO34631]).
3. **** Noise from scattered light is currently one of the factors that limit the sensitivity in the frequency bin from 50 Hz to 200 Hz (see Fig. \[scatter\_modelled\]) especially during periods of high microseism. Off-axis beam scattered laser beam could hit a reflecting surface like camera mirror mount or beam tube and reenter the cavity. Nonlinear features are seen in the gravitational wave spectrum when this beam picks up resonances from reflecting surfaces which then get upconverted or phase modulated by low-frequency seismic-like motion. It’s effect at LLO is more pronounced as compared to LHO as the former is vulnerable to microseismic activity [@Ottaway:12].
![ Noise due to scattered light observed at LIGO Livingston. Noise gets amplified and upconverted during periods of high microseism and limits the sensitivity range of the GW detectors. []{data-label="scatter_modelled"}](Fig4.png){width="10cm"}
Figure \[scatter\_modelled\] shows the effect of acoustic excitation on 82 Hz peak seen in gravitational wave differential arm motion (DARM) data. The acoustic injections carried out at LIGO Y-end station are reconstructed using the proposed model [@VirgoScatterPaper] Scatter Noise = A sin(4\*pi\* (n\*Yrms + Yac)/lambda) where Yac = B\*sin(2pi fo t) where Yrms is the ground motion and Yac is the chamber motion with (A,n,B) being the tunable parameters. Model parameters using are fine tuned using pattern search. The scatter noise projection to DARM from ambient motion is obtained by scaling down the chamber motion based on the accelerometer signal before and after injection.
4. **Bounce and Roll Modes: Effect of tuned mass damper installation**
During the first phase of Advanced LIGO observation run (September 2015-January 2016) bounce and roll modes of the suspension lead to excess noise in 9 Hz to 20 Hz affecting the sensitivity of the instrument. As a remedy mechanical dampers tuned to these resonant frequencies were installed at various seismic isolation platforms within the detectors. The positive effect of these is evident from the decreased number of related reports written at the both the detectors.
5. ****
Glitches often show up in the strain data leading to false alarms in the various search pipelines that look for astrophysical signals. Triggers are also observed in badly functioning instruments and are witnessed in auxiliary sensor channels. Some of them are also reported to cause loss of lock of the interferometer. The general operation of all the three detectors have been affected by such transients ever since their beginning of operation \[timeline\]. The report generation feature of our application provides the following glitch distribution (Fig. \[scatter\_pie\]) across multiple subsystems based on their tags in the data. It is interesting to note the subtle variations in the noise sources between LHO and LLO. The origin of many of them have been studied and reported in the logbooks while a vast majority are still not well understood.
{width="14cm"} {width="14cm"} {width="14cm"} {width="14cm"}
Visualizing Observatory as a Complex Network
--------------------------------------------
Behavior of an observatory and the elements that lead to changes within the system behavior can be studied through its representation as a complex network. Complexity is expressed through nodes and links within the network. Here, the nodes can be either subsystems or specific instruments or even subgroups within the observatory and edges between them provide the probability of each one them being connected to the other as inferred from logbook entries. We first create a dictionary of subsystem keywords and for each one, find the frequency of their joint occurrence with everyone else. This information is then used to form the adjacency matrix whose diagonal elements are all zero and the off-diagonal value representing the linkage is given by the ratio of joint occurrence frequency divided by total occurrence of the keyword. Adjacency matrix being non-symmetric leads to a directed graph. Number of incident edges determines the node size while the edge width is given by the associated connection probability. To better aid visualization, we adopt Force Atlas 2 layout [@jacomy2014forceatlas2] with repulsion being approximated using Barnes Hut optimization [@barnes1986hierarchical] which is well suited for larger graphs. The interconnectedness information within the observatory revealed through these networks may help in identifying the critical nodes in the system and the makes it easier to identify the vulnerable connections. These representations could possibly be useful during large scale repair and maintenance as they reveal the other subsystems that can get affected in the process.
In Fig \[VirgoCombo\] we show the network connection for a few prominent nodes of Virgo observatory. It differs from real-world networks in terms of its degree distribution (degree refers to the number of edges connected to each node). Sparse networks are characterized by a degree distribution which takes form of a power law and are commonly seen in biological networks and computer networks. [@barabasi2016network]. For the case of Virgo network, this distribution deviates from such a power law indicating dense connection between the nodes. Further research is needed to analyze the network and study the instrument’s robustness to random sub-system failures.
{width="14cm"}
Discussions & Conclusion
========================
We have demonstrated how information retrieval and recommendation systems could be useful for LIGO like astronomical observatories. Compared with conventional search associated with the existing sites, our web application incorporates a natural language processing based information retrieval system which can also do visualization of the user queried data. Involving a wider science community in big science projects can alleviate some of the issues related to lack of sufficient man power within the project. The developed interface identifies the major issues based on the discussions done within LIGO community and recognizes the trending issues. It is plausible that someone outside the project has already seen and solved these before. Hence proper dissemination of information will help in technical experts outside the project collaboration to contribute improving the overall performance of the instrument.
Coordinated efforts are being undertaken worldwide to carry out electromagnetic follow-up searches looking for counter parts to coalescing binaries sources[@abbott2016localization]. During the instance of GW candidate event alert, astronomers may be able to take advantage of our application and know more about the instrument.
Future improvement in the application would be to include capabilities wherein an identified issue will be provided with possible fixes making use past attempts which fixed an identical issue. This would require text abstraction and summarization, quite challenging when the data has ample amount of technical terms. Efforts to add other GW detectors like GEO600 and KAGRA is currently under progress and will enhance the effectiveness of our application.
This kind of system has a lot of potential applications with the commissioning and running of large science projects like the SKA and future LIGO observatories. In this project the data source was more unstructured and had few tags related to the status of different activities. At present institutions like SKA South Africa which is in charge of building MeerKAT [^5] telescope which is one of the precursors to SKA, use a more structured systems like JIRA [^6] for issue tracking and log keeping[^7]. Scaling our present system to such databases can improve the efficiency of topical modeling. This also enables auto-update of the learning database as more and more information is logged into the system finally making it a robust.
The aforementioned feature will eventually be very useful when the organization grows with the number of participants increasing over time. Availability of such systems will make the re-usability of information much easier and efficient. This will also help in resolving instrument issues much easier and faster. Enhanced analytics of key components and recurring issues can help improving the fault tolerance of different subsystems and could provide insights on how modify them for better performance.
Acknowledgements
================
We would like to thank detector characterization group and machine learning sub-group of the LIGO Scientific Collaboration for their comments and suggestions. NM acknowledges the Council for Scientific and Industrial Research (CSIR), India for providing financial support as a Senior Research Fellow and Navajbai Ratan Tata Trust (NRTT) grant for supporting his visit to LIGO Livingston. A. K. Aniyan would like to thank the SKA South Africa postgraduate bursary program. SM acknowledges the support of the Science and Engineering Research Board (SERB), India through the fast track Grant No. SR/FTP/PS-030/2012. Authors express thanks to Arnaud Pele, Anamaria Effler and Ajit K Kembhavi for their valuable comments and suggestions. Authors wish to thank Malathi Deenadayalan and Santosh Jagade for technical support. LIGO was constructed by the California Institute of Technology and Massachusetts Institute of Technology with funding from the National Science Foundation and operates under Cooperative Agreement No.PHY-0757058. This paper has been assigned LIGO Document No. LIGO-P1700250.
[^1]: [heyligo.gw.iucaa.in](heyligo.gw.iucaa.in)
[^2]: <https://alog.ligo-la.caltech.edu/aLOG/>
[^3]: <https://alog.ligo-wa.caltech.edu/aLOG/>
[^4]: <https://logbook.virgo-gw.eu/virgo/>
[^5]: <http://www.ska.ac.za/science-engineering/meerkat/>
[^6]: <https://www.atlassian.com/software/jira>
[^7]: <https://indico.skatelescope.org/event/402/material/1/6.pdf>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'As many as 5 ice giants—Neptune-mass planets composed of $\sim$90% ice and rock and $\sim$10% hydrogen—are thought to form at heliocentric distances of $\sim$10–25 AU on closely packed orbits spaced $\sim$5 Hill radii apart. Such oligarchies are ultimately unstable. Once the parent disk of planetesimals is sufficiently depleted, oligarchs perturb one another onto crossing orbits. We explore both the onset and the outcome of the instability through numerical integrations, including dynamical friction cooling of planets by a planetesimal disk whose properties are held fixed. To trigger instability and the ejection of the first ice giant in systems having an original surface density in oligarchs of $\Sigma \sim 1 \gm/{\rm cm}^2$, the disk surface density $\sigma$ must fall below $\sim$$0.1\gm/{\rm cm}^2$. Ejections are predominantly by Jupiter and occur within $\sim$$10^7\yr$. To eject more than 1 oligarch requires $\sigma \lesssim 0.03 \gm/{\rm cm}^2$. For certain choices of $\sigma$ and initial semi-major axes of planets, systems starting with up to 4 oligarchs in addition to Jupiter and Saturn can readily yield solar-system-like outcomes in which 2 surviving ice giants lie inside 30 AU and have their orbits circularized by dynamical friction. Our findings support the idea that planetary systems begin in more crowded and compact configurations, like those of shear-dominated oligarchies. In contrast to previous studies, we identify $\sigma \lesssim 0.1 \Sigma$ as the regime relevant for understanding the evolution of the outer solar system, and we encourage future studies to concentrate on this regime while relaxing our assumption of a fixed planetesimal disk. Whether evidence of the instability can be found in Kuiper belt objects (KBOs) is unclear, since in none of our simulations do marauding oligarchs excite as large a proportion of KBOs having inclinations $\gtrsim 20^{\circ}$ as is observed.'
author:
- 'Eric B. Ford and Eugene I. Chiang'
title: |
The Formation of Ice Giants in a Packed Oligarchy:\
Instability and Aftermath
---
INTRODUCTION {#sec_intro}
============
Without gravitational focussing, [*in situ*]{} coagulation of Uranus and Neptune takes too long to complete. In a minimum-mass disk at heliocentric distances of 20–30 AU, timescales to assemble the ice giants exceed the age of the solar system by 2 orders of magnitude, if growth is unfocussed (e.g., Goldreich, Lithwick, & Sari 2004, hereafter GLS04).[^1] N-body coagulation simulations that do not damp relative velocities between planetesimals, either by dynamical friction, inelastic collisions, or gas drag, fail to form Uranus and Neptune (Levison & Stewart 2001; see also Lissauer et al. 1995). The ice giants contain 10–20% hydrogen by mass, a fraction so large that such gas must originate from the solar nebula. The outer planets must therefore form within a few $\times$ $10^7\yr$, before all of the nebular hydrogen photoevaporates (Shu, Johnstone, & Hollenbach 1993; Matsuyama, Johnstone, & Hartmann 2003).
One way to alleviate (but not necessarily eliminate) the timescale problem is to form Uranus and Neptune closer to the Sun, where material densities and collision rates are greater. Thommes, Levison, & Duncan (1999, 2002) explore a scenario in which the two planets form at distances of 5–10 AU, between the cores of Jupiter and Saturn. Once the gas giant cores amass their envelopes, they scatter the ice giants outward onto eccentric orbits. These orbits subsequently circularize by dynamical friction with planetesimals at 15–30 AU. Tsiganis et al. (2005) propose an alternative history in which Uranus and Neptune accrete at 12 and 17 AU, are thrown outward by Jupiter and Saturn, and have their orbits circularized by dynamical friction. According to their story, the outward scattering of ice giants is triggered by having Jupiter and Saturn divergently migrate across their mutual 2:1 resonance.
Another approach to solving the timescale problem is to consider how gravitational focussing can be amplified. GLS04 adopt this route by appealing to a massive disk of sub-km-sized planetesimals, similar to those produced by coagulation simulations set in the outer solar system (Kenyon & Luu 1999). The disk envisioned by GLS04 has a mass several times the minimum-mass value in condensates so that the “isolation mass”—the mass to which a protoplanet grows by consuming all material within its annulus of influence—equals Neptune’s mass. The small bodies comprising the disk collide so frequently that their velocity dispersion damps to values less than the Hill velocity of a Neptune-mass planet. Accretion rates then enjoy maximal enhancement by gravitational focussing, and proto-Neptune can accrete the last half of its mass in $\sim$$10^5\yr$ (see eqn. 105 of GLS04). Gas drag supplies another means to damp planetesimal velocity dispersions (Rafikov 2004; Chambers 2006; see also appendix A of GLS04).
Strongly focussed, [*in situ*]{} assembly of planets from a dynamically cold disk carries, however, a potential embarrassment of riches: The disk can spawn more ice giants than the solar system’s current allotment of 2 (Uranus and Neptune). We estimate that about 5 isolation masses or “oligarchs,” each having the mass of Neptune, can form between 15 and 25 AU (see eqn. \[eqn\_define\_a\] below). These planets comprise a “shear-dominated oligarchy,” so-called because the encounter velocities between planets and planetesimals are given by their minimum values set by Keplerian shear. Initially, the oligarchs’ nested orbits would be stabilized by dynamical friction with the disk. GLS04 suggest that excess oligarchs would be purged from the outer solar system by an eventual dynamical instability. According to their order-of-magnitude analysis, this “velocity instability” occurs once the mass of the disk becomes less than the mass in oligarchs, whereupon dynamical friction ceases to stabilize the system against mutual gravitational stirring (a.k.a. “viscous stirring”). In the ensuing chaos, several oligarchs would be ejected, either by other oligarchs or by Jupiter or Saturn, possibly leaving two survivors whose orbits could circularize by dynamical friction at 15–30 AU.
Despite their disparate perspectives on the timescale problem and different motivations, the scenarios of Thommes et al. (1999), Tsiganis et al. (2005), and GLS04 share quite a few features. In their simplest forms, each theory starts with a more crowded configuration for solar system planets than is observed today; each is characterized by an intermediate period of dynamical chaos during which Uranus and Neptune execute highly eccentric orbits; and each invokes final regularization of ice giant orbits by dynamical friction with an ambient disk. Thommes et al. (2002) and Chiang et al. (2006, hereafter C06) point out that these violent histories might be encoded in Kuiper belt objects (KBOs). In particular, so-called scattered KBOs possess large eccentricities, inclinations, and perihelion distances which might reflect gravitational stirring by marauding ice giants.
The notion that planets originate in compact and crowded configurations is bolstered by the study of extra-solar systems as well. To explain the striking preponderance of large orbital eccentricities observed among extra-solar giant planets (Butler et al. 2006), multiple planets each having on the order of a Jupiter mass are imagined to have once resided on orbits sufficiently close that the planets scatter one another onto elliptical trajectories (Marzari & Weidenschilling 2002; Ford, Rasio, & Yu 2003; Ford, Lystad, & Rasio 2005).
GLS04 outline a possible formation history for Uranus and Neptune in a packed oligarchy, and C06 expand upon its consequences for the Kuiper belt, by making many simplifying assumptions and order-of-magnitude approximations. In this paper, we test some of their ideas by numerical simulations. In particular, we seek answers to the following questions:
1. For a shear-dominated oligarchy containing more than two Neptune-mass oligarchs beyond Saturn’s orbit, what is the critical value of the disk surface density below which the velocity instability occurs?
2. What is the likelihood that the instability will result in the survival of two oligarchs whose final orbits resemble those of Uranus and Neptune?
3. To what degree is the Kuiper belt dynamically excited by velocity-unstable oligarchs?
Our methods are described and tested in §\[sec\_method\]. That section contains empirical determinations of how rapidly 5 oligarchs viscously stir one another, with and without the gas giants Jupiter and Saturn. Comparisons are made with analytic theory. Results of hundreds of simulations designed to provide statistical answers to the above questions are presented in §\[sec\_result\]. We summarize and offer an outlook in §\[sec\_sum\].
METHOD AND TESTS {#sec_method}
================
To guide the reader, we provide a condensed description of our method in §\[sec\_overview\]. Details are elaborated upon in §§\[sec\_visc\_stir\]–\[sec\_dyn\_fric\].
Overview {#sec_overview}
--------
We simulate the final stages of oligarchy by numerically integrating the trajectories of 5 closely packed Neptune-mass oligarchs, together with those of 2 gas giants resembling Jupiter and Saturn. Oligarchs and gas giants are referred to as planets. We employ the hybrid orbit integrator MERCURY6 (Chambers 1999), which combines a conventional Bulirsch-Stoer integrator to handle close encounters between planets, with the fast symplectic algorithm invented by Wisdom & Holman (1991). Viscous stirring of an oligarch by other planets is simulated as accurately as the orbit integrator solves the gravitational equations of motion. Case studies of viscous stirring are described in §\[sec\_visc\_stir\].
To model dynamical friction between a planet and the surrounding disk of planetesimals, we introduce a perturbative force on each planet. The force damps the component of the planet’s velocity that differs from the local disk (circular) velocity. For simplicity, we take the disk to have a constant surface density between an inner and an outer radius. Disk parameters are held fixed. In our simple scheme, the planets respond to the disk through dynamical friction, but the disk does not respond to the planets. The details of the perturbation force are provided in §\[sec\_dyn\_fric\]. The validity of our fixed disk approach is briefly considered in §\[sec\_reduce\].
Finally, to investigate how oligarchs might excite the Kuiper belt, an ensemble of test particles is included in a subset of the simulations. We use the terms “test particle” and “Kuiper belt object (KBO)” interchangeably. These test particles are intended to represent large KBOs like those observed today, having sizes on the order of 100 km. This size is small enough that we can neglect dynamical friction between KBOs and the disk, yet also large enough that we can ignore damping of KBOs’ velocities by physical collisions with the disk (C06). Thus, in our simulations, KBOs (test particles) feel directly only the gravity of the Sun and of the planets.
In the sub-sections below, we explore separately the processes of viscous stirring (§\[sec\_visc\_stir\]) and dynamical friction (§\[sec\_dyn\_fric\]), in isolation from one another. We present full-fledged simulations, in which the two processes are combined, in §\[sec\_result\].
Viscous Stirring {#sec_visc_stir}
----------------
We first study how multiple oligarchs gravitationally stir one another. For this sub-section, we ignore dynamical friction with the disk, but include the gas giants, Jupiter and Saturn. We compare our findings to those of GLS04. The results of this sub-section will be applied in §\[sec\_result\] to understanding the threshold conditions required for velocity instability.
### Initial Conditions {#sec_init_con}
We consider $N_{\rm olig} = 5$ oligarchs, each having the mass of Neptune ($\mu = m/m_{\odot} = m_{\rm
N}/m_{\odot} = 5.1 \times 10^{-5}$). The oligarchs are initially spaced 5 Hill radii apart in semi-major axis ($a$); that is, the difference between semi-major axes of nearest-neighboring oligarchs is
$$a_{j+1} - a_j = 2.5 (\mu/3)^{1/3}a_j + 2.5 (\mu/3)^{1/3}a_{j+1} \,,
\label{eqn_define_a}$$
where $j$ ranges from 1 to $N_{\rm olig}$. The coefficient of 2.5 is inspired by numerical studies by Greenberg et al. (1991) of the width of a protoplanet’s feeding zone, for the case where Keplerian shear dominates the relative velocity between a protoplanet and a planetesimal (see their eqn. 9; see also Ida & Makino 1993). While the coefficient of 2.5 is the standard value for our study, the coefficient in reality can be somewhat different, depending on the accretion and migration histories of the planets. We explore some consequences of varying the coefficient in §\[sec\_stir\_curve\] and §\[sec\_instab\_ss\].
We assume the semi-major axis of the innermost oligarch $a_1 = 15$ AU. Then according to eqn. (\[eqn\_define\_a\]), semi-major axes of the next four oligarchs equal 17.1, 19.4, 22.1, and 25.1 AU. Initial eccentricities, and initial inclinations relative to an arbitrary reference (x-y) plane, are such that $e_{j} = \sin i_{j} = 10^{-4}$. All orbital elements in this paper are osculating and referred to a barycentric coordinate system. That is, when computing the osculating Kepler elements for a given body, the position and velocity of the body are referred to the system barycenter (calculated using all massive bodies), while the central mass of the assumed Kepler potential equals the mass of the Sun plus that of the given body (alone). For each oligarch, the initial longitude of ascending node, longitude of pericenter, and mean anomaly are randomly generated from uniform distributions between 0 and $2\pi$.
Jupiter and Saturn are assigned initial masses and semi-major axes equal to their current values: $\mu_{\rm J} = m_{\rm J}/m_{\odot} = 9.5
\times 10^{-4}$, $\mu_{\rm S} = m_{\rm S}/m_{\odot} = 2.9 \times
10^{-4}$, $a_{\rm J}= 5.18$ AU, and $a_{\rm S} = 9.54$ AU. Initial eccentricities equal $e_{\rm J}=e_{\rm S} =0.05$, and inclinations relative to the reference plane are such that $\sin
i_{\rm J} = \sin i_{\rm S} = 0.01$. Orbital longitudes are randomly generated, just as they are for the oligarchs.
These initial conditions, particularly our choices for $a_1 = 15 \AU$ and $N_{\rm olig} = 5$, are somewhat arbitrary. They are intended to represent qualitatively the final stages of shear-dominated oligarchy in the outer solar system. We will adjust starting parameters (e.g., $a_1$, $N_{\rm olig}$) in later sections to achieve simulation outcomes in better agreement with observed properties of the solar system.
A total of $N_{\rm real} = 200$ orbital integrations (“realizations”) are performed with the hybrid integrator MERCURY6 (Chambers 1999), each characterized by a unique set of starting longitudes and each lasting $10^7 \yr$. The timestep for the symplectic integrator is set to 130 days. Timesteps for the conventional Bulirsch-Stoer integrator are as short as necessary to achieve an accuracy parameter of $10^{-10}$. The changeover distance that separates the symplectic regime from the close encounter regime is set to $\Delta r_{\rm crit} = 3$ Hill radii.
A “collision” between two bodies occurs when their mutual separation becomes less than the sum of their physical radii. The physical radius of each oligarch is computed using Neptune’s bulk density of $1.6 \gm/{\rm cm}^{3}$. Physical radii for Jupiter and Saturn are computed using densities of $1.3$ and $0.7 \gm/{\rm cm}^{3}$, respectively. We assume that bodies that collide merge completely.
An “ejection” occurs when an oligarch’s distance from the Sun exceeds 10000 AU and when its total kinetic plus potential energy (evaluated in barycentric coordinates with the potential energy set to zero at infinity) becomes positive. Ejected planets are dropped from the simulation.
### Results: Outcomes After $t = 10^7 \yr$ {#sec_fates}
At $t = 10^7\yr$, the outcomes for all $N_{\rm all} = N_{\rm real} \times N_{\rm olig} = 200 \times 5 = 1000$ oligarchs divide into the following mutually exclusive categories, in order of decreasing frequency of incidence:
1. Ejection but no collision (463)
2. No collision and no ejection (439)
3. Collision with another oligarch but no ejection (42; i.e., 21 oligarchs remain, each with mass twice that of an original oligarch)
4. Collision with Sun only (19)
5. Collision with Jupiter only (14)
6. Collision with another oligarch, and subsequent ejection or subsequent collision with Jupiter or Saturn (13)
7. Collision with Saturn only (10)
The dominant outcome is ejection. In 50% of the realizations (i.e., 100 out of $N_{\rm real}=200$), at least one oligarch is ejected before $t=1.6 \times 10^6\yr$. By $t = 3.2 \times 10^6 \yr$, 85% of all realizations experience a first ejection. All but 5 out of 200 realizations experience at least one ejection of an oligarch by $t = 10^7\yr$.
Jupiter and Saturn are responsible for the preponderance of ejections. When we repeat the experiment with Jupiter and Saturn omitted, outcomes at $t = 10^7 \yr$ are as follows: 870 out of 1000 oligarchs experience neither an ejection nor a collision, 118 collide with another oligarch (so that 59 remain), and 12 collide with two other oligarchs (so that 4 remain). No ejection is observed to occur by $t = 10^7 \yr$ when the gas giants are absent.
These outcomes are consistent with timescale estimates by GLS04. Neglecting Jupiter and Saturn, GLS04 predict that the oligarch ejection timescale is $\sim$$10^9\yr$ (see their eqn. 114). This is consistent with our finding that no ejection occurs within $t = 10^7\yr$ in the absence of gas giants. GLS04 mention the possibility that Jupiter and Saturn hasten ejections. We confirm this possibility. When gas giants are present, we find the ejection timescale is $\sim$$10^6\yr$.
### Results: Eccentricity and Inclination Growth (“Stirring Curves”) {#sec_stir_curve}
Fig. \[fig\_stir\] tracks the median eccentricity, $e_{50}(t)$, of all oligarchs. The sample from which the median is drawn always contains $N_{\rm all} = 1000$ objects, regardless of whether oligarchs collide or are ejected. When computing the median, we adopt the following rules: ejected oligarchs have their eccentricities set equal to 1 (but remain part of the sample); an oligarch that collides with either the Sun, Jupiter, or Saturn has its eccentricity set equal to 1; and oligarchs that collide with other oligarchs are still counted as separate objects and are each assigned an eccentricity equal to the current eccentricity of the merged body. Experiments with alternative sets of rules produced no qualitative changes to our results.
The resultant “stirring curves” of Fig. \[fig\_stir\] exhibit a variety of behaviors. We first discuss the case when Jupiter and Saturn are omitted from the integrations (Fig. \[fig\_stir\]b). As annotated in Fig. \[fig\_stir\]b, we distinguish four phases of viscous stirring:
1. [*Distant Conjunctions:*]{} At early times $t \lesssim 500 \yr$, planetary orbits do not cross and $e_{50}$ grows roughly linearly with time. A linear dependence is expected from viscous stirring by distant conjunctions, i.e., conjunctions between oligarchs that are not nearest neighbors. To derive the $t^1$ scaling, we estimate using the impulse approximation that a conjunction between two oligarchs separated by distance $x < a$ imparts eccentricities on the order of $$\Delta e \sim \mu \left( \frac{a}{x} \right)^2$$ to both bodies, provided they have eccentricities less than $\Delta
e$ prior to conjunction. For a given oligarch, a total of $N
\sim \Sigma a x / m$ oligarchs all reside about the same distance $x$ away, where $\Sigma$ is the mass surface density of oligarchs. Conjunctions with these oligarchs occur over the synodic period $$t_{\rm syn} \sim \frac{a}{x}\, t_{\rm orb}\,,$$ where $t_{\rm orb}$ is the orbital period. Then $$\frac{de}{dt} \sim N \frac{\Delta e}{t_{\rm syn}} \sim \frac{\Sigma
a^2}{m_{\odot}} \frac{1}{t_{\rm orb}} \sim {\rm constant}
\label{eqn_distant}$$ as roughly observed in Fig. \[fig\_stir\]. As time elapses, ever closer neighbors at smaller $x$ drive the stirring. This reasoning matches that given by GLS04 in their treatment of viscous stirring in the shear-dominated regime; their eqn. (33) is identical in form to our eqn. (\[eqn\_distant\]). The $t^1$ scaling is also derived by Collins & Sari (2006) and Collins, Schlichting, & Sari (2007). These latter studies concentrate on the limit $N_{\rm olig}\gg 1$.
2. [*Conjunctions with Nearest Neighbors:*]{} At intermediate times $500 \lesssim t ({\rm yr}) \lesssim 1.5\times 10^4$, planetary orbits remain non-crossing but the eccentricity distribution hardly changes. Since the synodic period between nearest neighboring oligarchs is $t_{\rm
syn} \sim 500\yr$, a given oligarch during this phase experiences repeated conjunctions with its nearest neighbor. Such repeated close encounters might be expected to produce chaotic motion and to cause eccentricities to random walk, in which case $e_{50} \propto t^{1/2}$. That this scaling is not observed implies that our 5 oligarchs do not behave in a strongly chaotic manner despite their close spacing. Indeed, we observe in our simulations that the epicyclic phases (true anomalies) of a given oligarch at successive conjunctions with its nearest neighbor do not vary completely randomly. Perturbations from conjunctions with a nearest neighbor apparently tend to cancel out during this second phase.
3. [*Onset of Orbit Crossing:*]{} The cancellations characterizing the preceding phase are not perfect, however. Eventually, from $t \sim
1.5\times 10^4\yr$ to $t\sim 5 \times 10^4\yr$, eccentricities surge as oligarchs start crossing orbits. The median eccentricity $e_{50}$ surpasses the orbit-crossing value, $e \approx 0.06$, during this third phase.
Our finding that orbits cross in a few $\times$ $10^4\yr$ is consistent with numerical experiments by Chambers, Wetherill, & Boss (1996), who measure times required for close encounters to occur in initially circular, co-planar, multi-planet systems as a function of planet mass and orbital spacing. For reference, our spacing of $5 R_{\rm H}$ corresponds to 4 “mutual Hill sphere radii” as defined by those authors.
4. [*Orbit Crossing:*]{} At late times $t \gtrsim 5 \times 10^4
\yr$, oligarchs routinely cross orbits and we observe $e_{50}
\propto t^{0.25}$. We can reproduce this scaling using the following particle-in-a-box argument. An oligarch’s random velocity $v$ at time $t$ is determined largely by its closest encounter with another oligarch up until that time. We call the impact parameter characterizing this closest encounter $b_{\rm min}$. From kinetic theory, $n b_{\rm min}^2 v t \sim 1$, where $n \sim \Sigma
\Omega / m v$ is the number density of oligarchs and we have assumed that the random velocity distribution of oligarchs is isotropic. It follows that $b_{\rm min} \propto 1/t^{1/2}$ and $v \sim (Gm/b_{\rm min}^2)^{1/2} \propto t^{1/4}$. This scaling agrees with that of eqn. (49) of GLS04 and that of eqn. (14) of C06.
When we restore Jupiter and Saturn to the integrations (Fig. \[fig\_stir\]a), we can still discern the four phases enumerated above. However, compared to the case without giants, some phase boundaries are shifted to earlier times, and $e_{50}$ rises more quickly during some phases. Phase 1 transitions to phase 2a at $t \approx 30 \yr$; at this time, all oligarchs have undergone their first conjunctions with Jupiter and Saturn. Phase 2a transitions to phase 2b at $t \approx 500 \yr$; as in the case without giants, this transition marks the time when every oligarch has experienced about one conjunction with its nearest neighboring oligarch. During phase 2b, we witness the same remarkable near-constancy of $e_{50} \approx 0.01$. Finally, during phase 4 at $t \gtrsim 3 \times 10^4 \yr$, when oligarchs are on crossing orbits, $e_{50} \propto
t^{0.38}$. Such growth outpaces that observed in the absence of the gas giants.
What about oligarch inclinations? In simulations without gas giants, we observe that the median inclination $i_{\rm 50}$ remains fairly constant at the initial value of $10^{-4}$ until phase 3. As orbits cross, $\sin i_{50}$ surges up to $\sim$$0.05$, and thereafter grows as $t^{0.25}$ during phase 4, just as $e_{50}$ does. By $t= 10^7\yr$, $i_{50} \approx 10^{\circ}$. When Jupiter and Saturn are included, $\sin i_{50} \propto t^{0.28}$ during phase 4. By $t = 10^7\yr$, $i_{50} \approx 10^{\circ}$, as was the case without the gas giants. The modest growth of oligarchs’ inclinations will limit the degree to which inclinations of KBOs are stirred (§\[sec\_stir\]).
Not all of the different phases of viscous stirring that we observe are anticipated from the study of GLS04, which documents only the $t^1$ scaling characterizing shear-dominated oligarchy (phase 1) and the $t^{1/4}$ scaling characterizing the super-Hill, orbit-crossing regime (phase 4, no giants). Their analysis misses the intermediate phase 2 of slow-to-no growth just prior to orbit crossing, and the significant roles that Jupiter and Saturn play in accelerating viscous stirring during phase 4 ($t^{0.38}$ vs. $t^{1/4}$). That differences exist is not too surprising, as their analysis is rooted in the large $N_{\rm olig} \gg 1$ limit, whereas for our system, $N_{\rm olig} = 5$. More importantly, we take nearest-neighboring oligarchs to be separated by 5 Hill sphere radii, as dictated by the extent of an oligarch’s feeding zone in a shear-dominated disk (Greenberg et al. 1991), whereas the order-of-magnitude equations of GLS04 governing shear-dominated oligarchy assume the separation is closer to $\sim$1 Hill sphere radius. In this regard, we present in Figure \[fig\_stir2\] viscous stirring curves for cases where the oligarch separation is $3 R_{\rm H}$ and $7 R_{\rm H}$ (corresponding to coefficients in eqn. (\[eqn\_define\_a\]) of 1.5 and 3.5, respectively). Without gas giants, for a separation of $3R_{\rm H}$, phases 2 and 3 disappear, leaving only phases 1 and 4 as originally envisioned by GLS04. The time to orbit crossing varies from $\sim$$300 \yr$ to $\sim$$2 \times 10^6\yr$ as the spacing changes from 3 to 7 Hill radii. This extreme sensitivity to spacing was also found by Chambers et al. (1996). Including gas giants, however, reduces this sensitivity, as Fig. \[fig\_stir2\]a shows.
The actual oligarchic spacing might only be determined by careful numerical simulations of accretion and orbital migration. We adopt in this paper a standard value of $5R_{\rm H}$, identical to that assumed by GLS04, and motivated by studies of shear-dominated accretion by Greenberg et al. (1991). Shorter spacings seem less attractive insofar as they will produce smaller isolation masses for a given disk surface density.
The intermediate phase of slow-to-no growth of eccentricity that characterizes oligarchic spacings $\geq 5 R_{\rm H}$ will prove important in determining the threshold disk surface density below which dynamical friction cooling cannot balance viscous heating, i.e., the threshold surface density for the velocity instability (§\[sec\_require\]).
Dynamical Friction {#sec_dyn_fric}
------------------
Oligarchs grow from a disk of planetesimals. Those planetesimals that are not accreted exert dynamical friction on oligarchs. The conditions for velocity instability, and the ease with which survivors of the instability return to low-eccentricity, low-inclination orbits, depend on the strength of dynamical friction. We describe how we implement dynamical friction in our simulations in §\[sec\_prescrip\], present a test case in §\[sec\_testcase\], and show that our implementation is compatible with the formulae of GLS04 in §\[sec\_connect\_gls\].
### Prescription {#sec_prescrip}
Consider a planet having an eccentricity and an inclination much greater than those of disk planetesimals. Dynamical friction reduces the planet’s random (peculiar) velocity: the difference $\vec{v} \equiv v
\hat{v}$ between the orbital velocity of the planet and that of the mean disk flow. From Binney & Tremaine (1987, their eqn. 7-17),
$$\frac{d\vec{v}}{dt} = - \frac{2\pi G^2 m \rho}{v^2 } \ln (1+\Lambda^2) \,
\hat{v} \,,
\label{eqn_binney_tremaine}$$
where $\rho$ is the local mass density of the disk, $m$ is the mass of the planet, $G$ is the gravitational constant, and
$$\Lambda = \frac{b_{\rm max}(v^2 + 2\langle \sin^2i_{\rm disk}\rangle
v_{\rm circ}^2)}{Gm}
\label{eqn_lambda}$$
is the Coulomb parameter appropriate for dynamical friction in a Keplerian disk (Stewart & Ida 2000). For $b_{\rm max}$, the maximum impact parameter between the planet and a disk planetesimal, we adapt the expression of Stewart & Ida (2000; see the discussion following their eqn. 2-17):
$$b_{\rm max} = R_{\rm H} + r\left( \langle \sin^2i_{\rm disk}\rangle +
\sin^2i\right)^{1/2} \,,
\label{eqn_b_max}$$
where $r$ is the instantaneous distance between the planet and the system barycenter, $R_{\rm H} = (\mu/3)^{1/3}r$ is the Hill sphere radius, and $\langle \sin^2 i_{\rm disk} \rangle^{1/2} \ll 1$ is the inclination dispersion of disk planetesimals, held constant for each simulation (more on its precise value later). The term $2\langle \sin^2 i_{\rm disk} \rangle v_{\rm circ}^2$ in eqn. (\[eqn\_lambda\]) approximates the square of the velocity dispersion of disk planetesimals, where $v_{\rm circ}$ is the local mean disk speed. We take ${v}_{\rm circ}$ to equal the speed that the planet would have on a circular orbit about the Sun.
Usually it is assumed in writing eqn. (\[eqn\_binney\_tremaine\]) that $\Lambda \gg 1$. We do not make this assumption. In fact, we use eqns. (\[eqn\_binney\_tremaine\])–(\[eqn\_b\_max\]) regardless of the magnitude of $\Lambda$. When $\Lambda \ll 1$, dynamical friction is in the shear-dominated regime. In §\[sec\_connect\_gls\], we justify our universal application of (\[eqn\_binney\_tremaine\])–(\[eqn\_b\_max\]) by showing that these equations correctly reduce to forms appropriate to the shear-dominated case when $\Lambda \ll 1$.
We implement dynamical friction as follows. We are interested in the case where the oligarchs are so dynamically excited that each plunges through a vertically thin, dynamically cold disk of planetesimals twice per orbit. Specifically, we assume that the time a planet spends immersed in the disk, $\Delta
t \approx h/|v_z|$, where $h$ is the full vertical thickness of the disk and $|v_z|$ is the vertical component of $\vec{v}$ at the moment of disk crossing, is short compared to the orbital period, $t_{\rm orb}
=2\pi/\Omega$. Equivalently, $\sin i \gg \langle \sin^2i_{\rm
disk}\rangle^{1/2}$. At every disk crossing, a planet receives a specific impulse of
$$\begin{aligned}
\Delta \vec{v} \,\, \approx \,\, \frac{d\vec{v}}{dt} \Delta t & \approx &
- \frac{2\pi G^2m}{v^2} \ln (1+\Lambda^2) \rho \Delta t \, \hat{v}
\nonumber \\
& \approx & - \frac{2\pi G^2 m}{v^2} \ln (1+\Lambda^2) \frac{\sigma}{|v_z|}
\,\hat{v} \,, \label{eqn_prescrip}\end{aligned}$$
where $\sigma$ is the disk surface density (height-integrated $\rho$). At every timestep of the integration, we check whether the planet crosses through the disk, which is fixed to lie in the x-y plane. At moments of disk crossing, we apply a kick according to eqn. (\[eqn\_prescrip\]): we increment the velocity of the planet by $\Delta \vec{v}$ but do not change the planet’s position. We compute the difference velocity $\vec{v}$ by subtracting the barycentric velocity of the planet from ${v}_{\rm circ} \hat{\phi}$, where $\hat{\phi}$ is the unit vector in the azimuthal direction. The kick is applied in the subroutine MDT\_HY.FOR in the MERCURY6 code, after the positions are advanced but before the velocities are updated for the second time by the interaction Hamiltonian.
For all our simulations, we fix $\langle \sin^2i_{\rm
disk}\rangle^{1/2} = 10^{-3}$, a value sufficiently small that $\sin i
\gg \langle \sin^2 i_{\rm disk} \rangle^{1/2}$ for all but a tiny fraction of the time. In other words, the strength of dynamical friction in our simulations depends much more strongly on the planet’s random speed $v$ than on the much smaller random speeds of planetesimals (see eqns. \[eqn\_lambda\]–\[eqn\_b\_max\]). Planetesimals can maintain low velocity dispersions by inelastic collisions or by gas drag.
Our scheme for dynamical friction damps orbital inclinations relative to the x-y (disk) plane. The inclination may become so small that $\Delta t \propto 1 / \sin i$ exceeds $t_{\rm orb}$, at which point the planet is immersed within the disk and our impulse approximation breaks down. To account for this possibility, we arbitrarily set $\Delta t = \min (h/|v_z|, 0.025/\Omega)$. Our softening prescription slows but does not stop the damping of inclination and eccentricity for $\sin i \lesssim 0.004$. The softening might represent slight misalignments between the planet’s orbital plane and the disk midplane, which in reality will be warped. We have verified that our principal findings, described in §\[sec\_result\], do not depend sensitively upon the details of this prescription. While the precise values of the inclinations that we compute are clearly not very meaningful, we expect that our results are still qualitatively correct, i.e., the code correctly identifies when mutual inclinations between planetary orbits are large $(> 1 \,{\rm rad})$ and small $(\ll 1\,{\rm
rad})$.
The main virtue of our prescription for dynamical friction is its simplicity. We need only specify the disk surface density $\sigma
(r)$, not the volumetric density $\rho (r,z)$, and we need only apply dynamical friction at disk crossings. The main shortcoming of our prescription is that it does not account for the response of the disk to the planets. The clearing of gaps and the generation of time-dependent, non-axisymmetric structures (see, e.g., Goldreich & Tremaine 1982) will alter the gravitational back-reaction of the disk onto embedded planets in ways that could be significant but that we (and GLS04) do not capture. For ways to model the response of the planetesimal disk more realistically, see Lithwick & Chiang (2007) and Levison & Morbidelli (2007).
### Test with Single Planet {#sec_testcase}
We test our prescription for dynamical friction in the case of a single planet interacting with a disk. The disk has constant surface density $\sigma = 1 {\,{\rm g}}/{{\rm cm}}^{2}$. A Neptune-mass oligarch is placed on an orbit having initial semi-major axis $a_{\rm init} = 20$ AU and initial eccentricity and inclination such that $e_{\rm init} = \sin i_{\rm
init} = 0.3$. Two cases are considered, one where the initial argument of perihelion $\omega_{\rm init} = 0$ and another where $\omega_{\rm init} = \pi/2$. The evolution of $e(t)$ and $i(t)$ depends on $\omega_{\rm init} \bmod \pi$.
Fig. \[fig\_DynFricTest\] displays the resultant evolution. The planet’s eccentricity and inclination both drop precipitously toward zero after a time on the order of $10^6\yr$; the exact time varies by a factor of 3 between our choices for $\omega_{\rm init}$. The semi-major axis can increase or decrease. It changes by 3–13%, on the order of but less than the starting eccentricity.
We check our numerical results by comparing them to the following approximate analytic solution. Since the kick $\Delta \vec{v}$ is applied twice per orbit, we write
$$\begin{aligned}
\frac{d\vec{v}}{dt} & \approx & - \frac{2 \Delta \vec{v}}{t_{\rm orb}}
\nonumber \\
& \approx & - 4\pi \ln(1+\Lambda^2) \frac{G^2 \sigma m}{t_{\rm orb}}
\frac{\vec{v}}{v^3 |v_z|} \,. \label{eqn_two_kicks}\end{aligned}$$
We set $e=\sin i$ and make the following approximations: $\vec{v} = (e\hat{p}+i\hat{z})\Omega a$, $v = \sqrt{2} e \Omega a$, and $|v_z| = i \Omega a$. Here $\hat{p}$ and $\hat{z}$ are unit vectors that lie parallel and perpendicular to the disk, respectively. Then (\[eqn\_two\_kicks\]) simplifies to
$$\frac{de}{dt} = -\frac{\ln (1+\Lambda^2)}{\sqrt{2}} \frac{G\sigma}{\Omega a}
\frac{m}{m_{\odot}} \frac{1}{e^3} \,,
\label{eqn_dedt}$$
with an analogous equation for $i$. For fixed $\Lambda$ and $a$, eqn. (\[eqn\_dedt\]) integrates to
$$e = \left( e_{\rm init}^4 - \frac{4 \ln (1+\Lambda^2)}{\sqrt{2}}
\frac{m}{m_{\odot}} \frac{G\sigma}{\Omega a} t \right)^{1/4} \,.
\label{eqn_analyt}$$
The eccentricity (equivalently, inclination) vanishes in a finite time
$$t_{\rm vanish} = \frac{\sqrt{2}}{4\ln (1+\Lambda^2)} \frac{m_{\odot}}{m}
\frac{\Omega a}{G\sigma} e_{\rm init}^4 \,.
\label{eqn_vanish}$$
We overlay eqn. (\[eqn\_analyt\]) in Fig. \[fig\_DynFricTest\], taking as representative values $a = a_{\rm init} = 20$ AU and $\ln (1+\Lambda^2) = \ln (1 + \Lambda_{\rm init}^2) = 13$. The analytic solution lies between the two numerical solutions. We consider the agreement acceptable.
### Connecting Our Prescription to GLS04 {#sec_connect_gls}
Our equations can be re-cast into the same forms as those of GLS04, under the assumption $e=i$. We start with eqn. (\[eqn\_two\_kicks\]) and substitute $|v_z| = v/\sqrt{2}$, $t_{\rm orb}
= 2\pi/\Omega$, and $m = (4\pi/3)\rho_{\rm p} R_{\rm p}^3$, where $\rho_{\rm p}$ and $R_{\rm p}$ are the internal density and physical radius of the planet:
$$\frac{1}{v}\frac{dv}{dt} = -2^{3/2} \ln(1+\Lambda^2) \, \sigma \Omega \,
\frac{4\pi G^2 \rho_{\rm p} R_{\rm p}^3}{3 v^4} \,.
\label{eqn_start}$$
We next recognize that $4\pi G^2 \rho_{\rm p}^2 R_{\rm p}^4 / 3 = 3 v_{\rm
esc,p}^4 / 16 \pi$, where $v_{\rm esc,p}$ is the escape velocity from the surface of the planet. Then eqn. (\[eqn\_start\]) simplifies to
$$\frac{1}{v} \frac{dv}{dt} = - \frac{3\sqrt{2}}{8\pi} \ln (1 + \Lambda^2)
\frac{\sigma \Omega}{\rho_{\rm p} R_{\rm p}}
\left( \frac{v_{\rm esc,p}}{v} \right)^4 \,. \label{eqn_recast}$$
When $\ln (1+\Lambda^2)$ is a constant of order unity, eqn. (\[eqn\_recast\]) matches the form of eqn. (45) of GLS04, evaluated using the first line of their eqn. (46), with their planetesimal random velocity $u$ replaced by $v$ (since $v>u$; see their section 5.5, end of first paragraph). This formula describes dynamical friction in the dispersion-dominated regime, where $v$ exceeds the Hill velocity $v_{\rm H} = \Omega
R_{\rm H}$.
On the other hand, it is possible for $\Lambda \ll 1$. This happens, according to (\[eqn\_lambda\])–(\[eqn\_b\_max\]), when $v \sim
\sqrt{2} i \Omega \ll \sqrt{Gm/R_{\rm H}} \sim v_{\rm H}$ (terms proportional to $\langle \sin^2 i_{\rm disk} \rangle^{1/2} = 10^{-3}$ are negligible). In this shear-dominated regime,
$$\Lambda \approx \frac{R_{\rm H} v^2}{Gm} \ll 1 \,,$$
$\ln (1+ \Lambda^2) \approx \Lambda^2$, and eqn. (\[eqn\_recast\]) reduces to
$$\begin{aligned}
\frac{1}{v} \frac{dv}{dt} & = & - \frac{3}{\pi \sqrt{2}}
\frac{\sigma \Omega}{\rho_{\rm p} R_{\rm p}}
\left( \frac{R_{\rm H}}{R_{\rm p}} \right)^2 \nonumber \\
& = & - \frac{3}{\pi \sqrt{2}} \frac{\sigma \Omega}{\rho_{\rm p} R_{\rm p}}
\frac{1}{\alpha^2} \,, \label{eqn_shear}\end{aligned}$$
where we have defined, following GLS04, $\alpha \equiv
R_{\rm p}/R_{\rm H}$. Eqn. (\[eqn\_shear\]) matches, to within a numerical constant, eqn. (45) of GLS04, evaluated using the second line of their eqn. (46).
We conclude that our treatment of dynamical friction is compatible with that of GLS04.
RESULTS {#sec_result}
=======
We present the results of simulations that combine viscous stirring due to multiple oligarchs, with dynamical friction due to a planetesimal disk.
Initial Conditions and Integration Times {#sec_init}
----------------------------------------
Each system begins with either $N_{\rm olig}=5$, 4, or 3 oligarchs, together with the gas giants, Jupiter and Saturn. Initial conditions are the same as those described in §\[sec\_init\_con\], except that initial eccentricities and inclinations of oligarchs are such that $e_j = \sin i_j = 10^{-2}$. Initial inclinations are therefore ten times larger than the assumed inclination dispersion of disk planetesimals ($\langle \sin^2i_{\rm disk} \rangle^{1/2} = 10^{-3}$). Initial eccentricities are the same as those that characterize phase 2b of the viscous stirring curves (Fig. \[fig\_stir\]a). Planetary orbits initially do not cross.
Dynamical friction is exerted by a disk of constant surface density $\sigma$ which extends from a barycentric radius of 12.5 AU to 45 AU. The outer boundary coincides with the location of the classical Kuiper belt (C06). The inner boundary is less well motivated. It is chosen so that the oligarchs reside initially inside the disk while the gas giants do not. We explore values for $\sigma$ ranging from 0.4 to $0.001 \gm /{\rm cm}^2$. For reference, the initial surface density in oligarchs is $\Sigma \approx 1.5 \gm /{\rm cm}^2$.
In a subset of runs, we include 400 test particles representing large KBOs. These feel the gravity of the planets but do not feel dynamical friction from the disk (see §\[sec\_overview\]). Initial semi-major axes of test particles range from $a = 40$ to 45 AU, and initial eccentricities and inclinations are such that $e = \sin i = 10^{-2}$. For all planets and KBOs, initial longitudes of ascending node, longitudes of pericenter, and mean anomalies are randomly chosen from uniform distributions between 0 and $2\pi$.
Settings for the MERCURY6 code are the same as those given in §\[sec\_init\_con\], except for the duration of integration. The integration automatically halts when there are a catastrophic number of ejections, i.e., when the only massive bodies remaining include the Sun, Jupiter, Saturn, and one oligarch (which may have collided with other oligarchs). In the absence of such an event, each system is integrated first to $t = 2\times 10^7\yr$. If the planets seem to have stabilized at that time—i.e., their eccentricities no longer grow—then we stop the integration and record the outcome as final. Otherwise, we repeat this test as necessary at $5 \times 10^7\yr$ and $1 \times 10^8 \yr$.
By $t = 1 \times 10^8 \yr$, most but not all realizations stabilize. Excluding systems that are stopped abruptly once only three planets remain, we find that $\lesssim 10\%$ of systems have undergone a close encounter (here defined to occur when the distance between any two planets is less than 1 Hill radius) within the last $10^7\yr$ of the integration, for $N_{\rm olig}=5$ and all values of $\sigma$ tested. For $N_{\rm olig}=4$ and 3, the corresponding fractions are $\lesssim 10\%$ and $\lesssim 1\%$. Those realizations that do not stabilize by $t = 1 \times 10^8\yr$ are typically characterized by small values of $\sigma \lesssim 0.006 \gm/{\rm cm}^2$ (i.e., relatively weak dynamical friction) and one oligarch remaining on an eccentric orbit that extends well past the outer edge of the disk. For these low values of $\sigma$, the number of ejected oligarchs that we report will be underestimated, but not in a way that changes our qualitative conclusions.
The initial conditions just summarized apply to all results in the following two sections, §§\[sec\_require\]–\[sec\_two\]. An alternative set of initial conditions, motivated by the findings in those sections, and results pertaining thereto are presented in §\[sec\_compact\].
Threshold Disk Surface Densities for Instability {#sec_require}
------------------------------------------------
Oligarchs cross orbits when the disk surface density $\sigma$ is so low that dynamical friction cooling cannot compete with viscous stirring. GLS04 estimate the critical surface density for instability to be on the order of the surface density of oligarchs: $\sigma_{\rm crit} \sim \Sigma$ (see also Chiang & Lithwick 2005 for a correction in the derivation of this result). How well does this criterion predict the onset of instability for our system of $N_{\rm olig} = 5$ oligarchs?
As noted in §\[sec\_visc\_stir\], the rates of eccentricity growth (viscous stirring) exhibited in our N-body integrations differ from those estimated by GLS04. Specifically, our rates are slower, as evidenced by the period of slow-to-no growth of eccentricity (phase 2b) in Fig. \[fig\_stir\]a. Overestimating the vigor of viscous stirring leads to overestimates for $\sigma_{\rm crit}$. We try to predict $\sigma_{\rm crit}$ ourselves by drawing from the numerical results of §\[sec\_visc\_stir\]. We observe in Fig. \[fig\_stir\]a that after a time $t_{\rm unstable} \sim 5\times 10^3 \yr$, eccentricities surge rapidly to crossing values.[^2] Therefore for oligarchs to cross orbits, eccentricities must not be allowed to vanish by dynamical friction before $t_{\rm unstable}$:
$$t_{\rm vanish} > t_{\rm unstable} \,.
\label{eqn_unstable}$$
Using (\[eqn\_vanish\]) for $t_{\rm vanish}$, we find that (\[eqn\_unstable\]) translates into $\sigma < \sigma_{\rm crit}$, where
$$\begin{aligned}
\sigma_{\rm crit} & \sim & \frac{\sqrt{2}}{4\ln (1+\Lambda^2)}
\frac{m_{\odot}}{m} \frac{\Omega a e_{\rm init}^4}{G t_{\rm unstable}}
\label{eqn_semi_empirical}\\
\nonumber \\
& \sim & 0.2 \gm /{\rm cm}^2 \sim 0.1 \Sigma \,.
\label{eqn_crit}\end{aligned}$$
The numerical evaluation takes $a = 19.4 \AU$, $e_{\rm init} = 0.01$ (the value appropriate to phase 2b of the viscous stirring curves in Fig. \[fig\_stir\]a; the median eccentricity does not rise above 0.01 for $t < t_{\rm unstable}$), and $\ln (1+ \Lambda^2) = \ln (1 + \Lambda_{\rm init}^2) = 0.02$.
Our semi-empirical estimate for $\sigma_{\rm crit}$ finds support in Fig. \[fig\_branching\]a, which documents, for all runs starting with $N_{\rm olig} = 5$ oligarchs, the frequency of incidence of outcomes (“branching ratios”) as a function of $\sigma$. Instability and the subsequent ejection of 1 and only 1 oligarch is the dominant outcome for $\sigma = 0.1 \gm /{\rm cm}^2 \approx 0.07 \Sigma$. For $\sigma \gtrsim 0.3\Sigma$, more than 90% of realizations produce no ejection. Fig. \[fig\_sample\_1\] displays a sample simulation for $\sigma \approx 0.07\Sigma$ in which 1 oligarch escapes before the system stabilizes.
Figs. \[fig\_branching\]b and \[fig\_branching\]c supply branching ratios for $N_{\rm olig} = 4$ and 3. To produce the same number of ejections with smaller $N_{\rm olig}$ (less viscous stirring) requires smaller $\sigma$ (less dynamical friction). For example, for $N_{\rm olig} = 3$, the ejection of 1 and only 1 oligarch is the dominant outcome for $\sigma \approx 0.03\Sigma$, occurring in about 60% of realizations. The corresponding $\sigma$ for $N_{\rm olig} = 4$ is $0.05\Sigma$.
Runs with Two Surviving Oligarchs (Solar-System-Like Outcomes) {#sec_two}
--------------------------------------------------------------
While $\sigma_{\rm crit} \sim 0.1 \Sigma$ roughly characterizes the onset of instability and the subsequent ejection of a single oligarch, the disk surface density must be reduced below $\sigma_{\rm crit}$ to produce more than 1 ejection in a large fraction of runs. To generate an outcome reminiscent of our solar system starting with $N_{\rm olig}=5$ requires 3 ejections and the survival of 2 oligarchs. According to Fig. \[fig\_branching\]a, such an outcome occurs with a maximum probability of $\sim$50% for $\sigma \approx 0.01 \gm /{\rm cm}^2 \approx 0.007 \Sigma$. The probability exceeds 20% for all values of $\sigma \lesssim 0.03 \Sigma$ that we tested.
Figs. \[fig\_branching\]b and \[fig\_branching\]c indicate that for $N_{\rm olig}=4$ and 3, the values of $\sigma$ most likely to produce 4-planet systems (Jupiter, Saturn, plus 2 oligarchs) are $\sim$$0.01\Sigma$ and $\sim$$0.03\Sigma$, respectively. The probabilities for generating 4-planet systems starting with $N_{\rm olig} = 4$ or 3 reach large maximum values of about 50%, and remain above 20% over a considerable range in $\sigma$, up to $\sim$$0.07\Sigma$ in the case $N_{\rm olig}=3$.
The vast majority of the resultant 4-planet systems are correctly ordered; they contain, in order of increasing semi-major axis, Jupiter, Saturn, and 2 oligarchs. Moreover, in most of these systems, the surviving planets have not experienced a collision. In the following sub-sections we further quantify the properties of these correctly ordered, collisionally unmodified, 4-planet systems, comparing them to those of the solar system. We refer to the 2 surviving oligarchs in each of these systems as Uranus and Neptune.
### Final Semi-Major Axis Distributions {#sec_semi}
Because packed oligarchies evolve chaotically, we can only meaningfully compute probability distributions for their final semi-major axes. Fig. \[fig\_finalA\_sigma\] illustrates how these distributions depend on $\sigma$, for realizations starting with $N_{\rm olig} = 5$. Increasing $\sigma$ increases dynamical friction and therefore tends to pull Saturn and Neptune, whose orbits lie near disk boundaries, into the disk. For example, if Saturn’s orbital apocenter intersects the disk while its pericenter remains outside the disk, then dynamical friction will circularize the orbit by raising the pericenter closer to apocenter. The kinks in the distribution functions for Saturn in Figs. \[fig\_finalA\_sigma\]a and \[fig\_finalA\_sigma\]b are located at $a=12.5$ AU, exactly at the inner disk edge. The kink vanishes in Fig. \[fig\_finalA\_sigma\]c. For the simulations in Fig.\[fig\_finalA\_sigma\]c, dynamical friction is strong enough to pull Saturn’s orbit wholly into the disk at $a \geq 12.5 \AU$.
To improve statistics, we ignore these small artifacts of our disk boundary conditions and pool together $N_{\rm real} = 438$ realizations, all of which start with $N_{\rm olig} = 5$ and produce 4-planet systems, but have a variety of $\sigma$’s between 0.001 and $0.1 \gm/{\rm cm}^2$. From this pool we construct the distribution of final semi-major axes shown in Fig. \[fig\_finalA\]a. Clearly, most realizations end with Saturn, Uranus, and Neptune on orbits too large compared to their actual counterparts in the solar system. Furthermore, Jupiter typically migrates inward as a consequence of ejecting several oligarchs outward. The excessively large orbits of Saturn, Uranus, and Neptune are a consequence of those planets having scattered oligarchs inward to Jupiter.
Though they only comprise (given our assumed initial conditions) a few percent of outcomes for $\sigma = 0.02 \gm/{\rm cm}^2$, some realizations better resemble the solar system insofar as Uranus and Neptune have final semi-major axes less than 30 AU. Fig. \[fig\_sample\_3\] showcases an example. Even for this simulation, however, Saturn’s final orbit is 3 AU larger than its actual one.
Reducing the number of starting oligarchs significantly lessens the problem of excessive migration. Figs. \[fig\_finalA\]b and \[fig\_finalA\]c show final semi-major axis distributions corresponding to $N_{\rm olig} = 4$ and 3, respectively. Outcomes for $N_{\rm olig} = 3$ are most solar-system-like. In §\[sec\_compact\] we will experiment with initial conditions to demonstrate that a level of agreement comparable to that displayed in Fig. \[fig\_finalA\]c for $N_{\rm olig}=3$ can also be obtained for $N_{\rm olig}=4$.
### Stirring of KBOs {#sec_stir}
Fig. \[fig\_KBOs\_3\] describes how the test particles (KBOs), initially distributed in a dynamically cold ring at $a=40$–45 AU, are stirred by oligarchs, for the same $N_{\rm olig}=5$ simulation (Fig. \[fig\_sample\_3\]) which places Uranus and Neptune on final orbits inside 30 AU. Simulation data are collected at $t = 5 \times 10^7\yr$. For comparison, Fig. \[fig\_KBOs\_3\] also plots data for actual KBOs that do not reside in any strong mean-motion resonance. These objects, taken from C06, comprise both low-$e$ classical and high-$e$ scattered KBOs as classified by the Deep Ecliptic Survey (Elliot et al. 2005). The simulated test particles have their eccentricities and inclinations excited up to $\sim$0.1, values that match actual classical KBOs. But the simulated particles fail to embody the extreme degree of dynamical heating exhibited by scattered KBOs. Marauding oligarchs in this simulation stir planetesimals at 40–45 AU too briefly.
Since simulations with $N_{\rm olig}=3$ more efficiently generate solar-system-like planetary spacings than do simulations with $N_{\rm olig}=5$, we can more thoroughly map out the possible extents to which KBOs are stirred for $N_{\rm olig}=3$. Figs. \[fig\_RvsT\_best\] and \[fig\_KBOs\_best\] document one simulation, representative of several percent of the solar-system-like realizations generated using $N_{\rm olig}=3$, in which KBOs are stirred considerably. Even here, however, the proportion of simulated KBOs that simultaneously attain inclinations $i \gtrsim 10^{\circ}$ and perihelion distances $r_{\rm p} \gtrsim 35$ AU is less than observed. The proportion of simulated KBOs having eccentricities $e \gtrsim 0.3$ also seems under-represented.
More Compact Initial Conditions {#sec_compact}
-------------------------------
As described in §§\[sec\_require\]–\[sec\_two\], the simulations that begin with $N_{\rm olig} = 4$ or 5 oligarchs often do yield 4-planet systems, but the orbital spacings of the resultant systems do not match those of the solar system. Multiple ejections displace Jupiter too far inward and displace Saturn, Uranus, and Neptune too far outward. The problem of excessive spreading is exacerbated by the need to have Neptune conclude its orbital evolution by migrating smoothly and slowly outward from $a\approx 23$ to 30 AU to produce the population of resonant KBOs (Murray-Clay & Chiang 2006, and references therein; but see Levison et al. 2006 for an alternative theory for the origin of resonant KBOs). This last constraint implies that our simulations should place Neptune on a final orbit near $a\approx 23$ AU.
In this sub-section we attempt to remedy the problem of excessive spreading by adjusting our initial conditions. In anticipation of Jupiter’s inward displacement, we locate that planet initially at $a_{\rm J} = 5.7\AU$. In anticipation of Saturn’s outward displacement, we set $a_{\rm S} = 8\AU$ initially. The innermost oligarch is also shifted inwards, to $a_1 = 12 \AU$. Initial semi-major axes for remaining oligarchs are still given by eqn. (\[eqn\_define\_a\]): $a_2$ through $a_5$ equal 13.7, 15.5, 17.7, and 20.1 AU. Finally, so that all oligarchs lie initially inside the disk, we extend the inner edge of the disk inward to 10 AU. All remaining parameters remain unchanged from their values in §\[sec\_init\].
The distribution of final semi-major axes for resultant 4-planet systems is given by Fig. \[fig\_FinalA\_compact\], constructed in similar fashion to Fig. \[fig\_finalA\]. The more compact initial configuration produces reasonably close matches to current orbital spacings in the solar system for $N_{\rm olig} = 4$. Results for $N_{\rm olig}=3$ are also acceptable if we allow for the subsequent outward migration of Neptune that is seemingly demanded by resonant KBOs (see first paragraph of this sub-section). The case $N_{\rm olig} = 5$ still suffers, however, from excessive spreading.
Figs. \[fig\_RvsT\_compact\] and \[fig\_KBOs\_compact\] sample one simulation using the revised compact configuration for $N_{\rm olig}=4$. We highlight this simulation because it reproduces solar system properties, insofar as (1) Uranus and Neptune have final semi-major axes less than 30 AU, and (2) the Kuiper belt at 40–45 AU is significantly stirred. Though outcome (1) is not infrequent—occurring in, e.g., 16 out of 100 runs with $\sigma = 0.04\gm/{\rm cm}^2$ and $N_{\rm olig} = 4$—outcome (2) is less probable, characterizing only several percent of runs already culled to satisfy (1). Most runs that satisfy (1) stir KBOs to eccentricities and inclinations of just a few percent. By contrast, the simulation showcased in Fig. \[fig\_KBOs\_compact\] excites large eccentricities and inclinations similar to those sported by actual KBOs. Nevertheless, the most extreme of scattered KBOs, having perihelion distances $r_{\rm p} \gtrsim 40$ AU and inclinations $i > 20^{\circ}$, are still under-represented. In short, our revised compact configuration stirs KBOs to about the same degree as our original configuration.
SUMMARY AND OUTLOOK {#sec_sum}
===================
In §\[sec\_answers\], we answer the three questions posed in §\[sec\_intro\]. In §\[sec\_commentary\], we place our work in a broader context and mention some directions for future work.
Answers to Questions Posed in §\[sec\_intro\] {#sec_answers}
---------------------------------------------
1. Of all our simulations that initially place $N_{\rm olig}=5$ Neptune-mass planets between 15 and 25 AU, and that have disk surface densities $\sigma \approx 0.1 \gm/{\rm cm}^2 \approx 0.07 \Sigma$ (where $\Sigma \sim 1.5 \gm/{\rm cm}^2$ is the original surface density in oligarchs), 50% result in the ejection of a single oligarch (Fig. \[fig\_branching\]). For runs that begin with $N_{\rm olig}=4$ and 3 oligarchs, we achieve similar outcomes for $\sigma/\Sigma \sim 0.05$ and $\sim$$0.03$, respectively. Jupiter is responsible for the vast majority of ejections, which occur within $\sim$$10^7\yr$. The likelihood of a single ejection remains as high as 20% if the above $\sigma$’s are increased by factors of 2–3. Roughly speaking then, we find that instability and ejection require $\sigma/\Sigma \lesssim 0.1$. By comparison, GLS04 estimate that $\sigma/\Sigma \lesssim 1$ for instability. The difference arises partly because nearest-neighboring oligarchs in our simulations are separated by 5 Hill sphere radii, whereas their analysis of shear-dominated oligarchy assumes the separation is closer to $\sim$1 Hill sphere radius. Our choice of $5 R_{\rm H}$ is motivated by the half-width of an oligarch’s annular feeding zone in a shear-dominated disk. This half-width spans $2.5R_{\rm H}$ (Greenberg et al. 1991). Because oligarchs separated by $5R_{\rm H}$ viscously stir each other more slowly (Fig. \[fig\_stir\], phase 2) than do oligarchs separated by $1R_{\rm H}$, we find a threshold value for $\sigma$ lower than what GLS04 estimate.
2. For certain choices of $\sigma$ and initial semi-major axes, systems starting with $N_{\rm olig}=3$ or 4 oligarchs frequently end with 2 surviving oligarchs on nearly circular and co-planar orbits inside 30 AU (Figs. \[fig\_finalA\] and \[fig\_FinalA\_compact\]). For example, of all runs that (a) use our revised set of initial semi-major axes (§\[sec\_compact\]), (b) begin with $N_{\rm olig}=4$ oligarchs, and (c) have $\sigma = 0.04\gm/{\rm cm}^2 \approx 0.02 \Sigma$, 44% end with solar-system-like configurations in which the outermost surviving oligarch orbits inside 30 AU. This percentage decreases with increasing $N_{\rm olig}$. This is because surviving oligarchs spread outward, well beyond the current orbit of Neptune, as they scatter more oligarchs inward for eventual ejection by Jupiter. To eject efficiently more than one oligarch requires that $\sigma$ be reduced considerably below the previously mentioned threshold of $\sim$$0.1\Sigma$. For example, we find that for $N_{\rm olig}=4$ and our original set of initial semi-major axes (§\[sec\_init\_con\]), setting $\sigma \sim 0.02 \gm/{\rm cm}^2 \sim 0.01\Sigma$ maximizes the likelihood of 2 ejections at $\sim$50%.
3. In a small fraction of runs that successfully place Jupiter, Saturn, and 2 oligarchs on solar-system-like orbits inside 30 AU, test particles (KBOs) located initially in a dynamically cold ring at 40–45 AU have their eccentricities and inclinations considerably excited by velocity-unstable oligarchs. We observe maximum eccentricities of $\sim$0.8 and maximum inclinations of $\sim$$20^{\circ}$ (Figs. \[fig\_KBOs\_best\] and \[fig\_KBOs\_compact\]). In runs characterized by the greatest degrees of excitation, orbits of simulated KBOs resemble those of observed classical KBOs and some observed scattered KBOs. However, no run reproduces the large proportion of observed scattered KBOs having inclinations $\gtrsim 20^{\circ}$. There may also be a problem in generating enough KBOs having eccentricities $\gtrsim 0.3$, given the observational selection bias against finding such objects.
These results are complementary to those of Levison & Morbidelli (2007), who concentrate on the limit $\sigma \gg \Sigma$ and find that they cannot produce solar-system-like outcomes. In comparison, we study the case $\sigma \lesssim \Sigma$ and find positive results.
Commentary {#sec_commentary}
----------
### The Compactness of Our Preferred Initial Conditions
We find that a shear-dominated oligarchy can readily produce a solar-system-like outcome if it contains just a few excess oligarchs—about 1 or 2 extra—and if the oligarchs initially reside inside 20 AU. We are driven to these parameters because to scatter excess oligarchs inward (toward Jupiter for eventual ejection), surviving oligarchs must be scattered outward. The right amount of outward spreading is achieved for suitably compact initial configurations and not too many ejections.
Our favored initial conditions are about as compact as those of Tsiganis et al. (2005), who place their outermost ice giant initially at 17 AU. By comparison, in our revised set of initial conditions for $N_{\rm olig}=4$, the outermost oligarch is located at 17.7 AU. But we stress that our study differs from theirs in that we base our initial conditions on considerations of shear-dominated oligarchic accretion. An ice giant cannot form at 17 AU within the gas photoevaporation time of a few $\times$ $10^7\yr$ without strong gravitational focussing (§\[sec\_intro\]; Levison & Stewart 2001; GLS04). This need for gravitational focussing can be met by a highly dissipative disk of planetesimals (GLS04; Rafikov 2004). It is this disk, and the multiple ($>2$) ice giants that it spawns, that we have modeled.
Why such a disk would not form Neptune-mass oligarchs outside 20 AU is an open question. For some ideas on what limits the sizes of planetary systems, see Youdin & Shu (2002) and Youdin & Chiang (2004).
### Reducing the Disk Surface Density (“Clean-Up” and Migration) {#sec_reduce}
Velocity instability and the ejection of a single oligarch require $\sigma/\Sigma \lesssim 0.1$. Ejecting more than 1 oligarch requires still lower values of $\sigma/\Sigma \sim 0.01$. How can $\sigma$ reach such low values? Accretion and/or ejection of planetesimals by oligarchs are natural possibilities. The rate at which $\sigma$ decreases, compared to the rate at which oligarchs stir each other, determines whether more than 1 oligarch escapes. If the rate of depletion of $\sigma$ is sufficiently slow, then after the first oligarch escapes, surviving planets may occupy orbits so spread apart that they remain stable even as $\sigma$ decreases further. On the other hand, if the rate of depletion is fast, then conditions required to eject more than 1 oligarch can be met. We leave investigation of the time dependence of $\sigma$ for future work.
Given an initial surface density in oligarchs of $\Sigma \sim 1 \gm/{\rm
cm}^2$, the disk surface densities relevant for instability and ejection range from $\sigma \approx 0.1$ to $0.01 \gm/{\rm cm}^2$. These $\sigma$’s are still higher than surface densities characterizing the Kuiper belt today, which are on the order of $0.001 \gm/{\rm cm}^2$ (integrated over all KBO sizes). “Cleaning up” the disk mass remains an unsolved problem (GLS04). Again, the mass could either be accreted or ejected by surviving planets. Planetesimal ejection drives planetary migration. For Neptune to expand its orbit from $\sim$23 to 30 AU, as seemingly demanded by the large observed population of resonant KBOs (C06), the planet must scatter at least $\sim$$7/30 \sim 25$% of its own mass in planetesimals, or about $4 M_{\earth}$. Spreading such a mass over a disk of radius 30 AU yields a surface density on the order of $0.03 \gm/{\rm cm}^2$, which falls within the range of $\sigma$’s that we find characterize instability. Clean-up and migration go hand-in-hand.
The disk masses that we found to be relevant for instability are smaller than the masses in the planets. This raises concern about the validity of our approximation that disk properties remain fixed throughout the simulation. One way of testing this approximation is to compare the change in the total angular momentum of all planets (brought about by dynamical friction) to the angular momentum available in the disk. If the former were much larger than the latter, then our neglect of back-reaction upon the disk would be a poor assumption. For the simulations displayed in Figs. \[fig\_RvsT\_best\] and \[fig\_RvsT\_compact\], we find that the change in the $z$-component of angular momentum of all planets (including ejected ones) is nearly identical to the angular momentum available in the disk, indicating that our assumption of a fixed disk may be only marginally valid. (An analogous test for the energy would be inconclusive since the disk is supposed to be a sink of energy by virtue of dissipative collisions). For ideas on how to treat planet-disk interactions self-consistently, see Lithwick & Chiang (2007) and Levison & Morbidelli (2007).
### Disk Optical Depths and Comparison to Debris Disks
A planetesimal disk of surface density $\sigma$ has a geometric vertical optical depth $\tau_p \sim 4 \times 10^{-6} [\sigma/(0.01 \gm/{\rm cm}^{2})] (10 \m / p)$, where $p$ is the assumed planetesimal radius. Collisions between planetesimals, which occur over a timescale $\sim$$1/(\Omega\tau_p) \sim 10^7 \yr$ at 30 AU, generate smaller dust particles whose optical depth is orders of magnitude higher. For example, if the dust size distribution obeys a Dohnanyi (1969) spectrum, then the geometric, vertical optical depth in $s$-sized grains would be $\tau_d \sim \tau_p (p/s)^{1/2} \sim 0.01 [\sigma/(0.01 \gm/{\rm cm}^2)] (10\m
/ p)^{1/2} (\mu{\rm m}/s)^{1/2}$. This is comparable to the vertical optical depths of some of the brightest extra-solar debris disks observed, e.g., $\beta$ Pic (Artymowicz 1997), HR 4796A (Li & Lunine 2003), and AU Mic (Strubbe & Chiang 2006), systems that are all $\sim$$10^7\yr$ old. The observed paucity of stars with optically thicker disks implies that large populations of planetesimals having sizes $p < 10 \m$ at stellocentric distances of $\sim$30 AU cannot be maintained for longer than $\sim$$10^7\yr$. The surface density in such collisional objects must be reduced by at least 2 orders of magnitude below planet-forming values of $\sim$$1 \gm/{\rm cm}^2$ within this timescale. In other words, planetesimals that are both collisional and planet-forming, like the kind espoused by GLS04, must be cleaned up fairly quickly.
### Instability in Our Solar System and Others {#sec_instab_ss}
We have demonstrated that more than 2 ice giants may once have orbited the Sun. The current architecture of the outer solar system may well have resulted from a prior era of dynamical instability during which Uranus, Neptune, and 1 or 2 other ice giants crossed paths.
We expect similar instabilities to afflict all nascent planetary systems. Perhaps planet-planet instabilities are reflected in the large orbital eccentricities exhibited by most extra-solar gas giants (Marzari & Weidenschilling 2002; Ford et al. 2003; but for an alternative view see Goldreich & Sari 2003). The case of Upsilon Andromedae fits this picture (Ford et al. 2005). The difference between our solar system and systems like Ups And might be the surface density of the parent disk at the time of the last planet-planet scattering (Ford 2006). The time of last scattering will vary widely because of the chaotic nature of multi-planet systems. In the case of the solar system, the disk surface density must have been large enough at the time of last scattering for dynamical friction to damp the eccentricities and inclinations of surviving planets back down.
Viscous stirring rates vary with the semi-major axis separation between oligarchs. At least in the case without gas giants, the time for oligarchs to undergo close encounters increases by several orders of magnitude as their semi-major axis spacing is increased from 3 to 7 mutual Hill radii (Chambers et al. 1996). The disk surface density required for instability depends directly on this time, i.e., $\sigma_{\rm crit} \propto 1/t_{\rm unstable}$ in our eqn. (\[eqn\_semi\_empirical\]). Whether $t_{\rm unstable}$ varies as strongly with oligarch spacing when perturbations by Jupiter and Saturn are included is not known, but preliminary experiments by us suggest that it does not. When we change the oligarch spacing from our standard 5 Hill radii to 3 Hill radii in runs that include Jupiter, Saturn, and $N_{\rm olig}=5$ oligarchs, we find that the probability of 1 ejection still peaks at $\sim$50% for $\sigma/\Sigma \approx 0.1$ (accounting for the factor of 2 increase in $\Sigma$ due to the shorter spacing).
### Evidence for the Velocity Instability in the Kuiper Belt
Did velocity-unstable ice giants excite the large eccentricities and inclinations of the scattered Kuiper belt, as proposed by C06? Our provisional answer is no, as we were unable to reproduce the large inclinations of scattered KBOs. For runs with disk surface densities down to $\sim$$0.01 \gm /{\rm cm}^2$, oligarchs spend too little time, less than $\sim$$10^7\yr$, passing through the Kuiper belt. Moreover, oligarchs in our simulations have orbital inclinations that rarely exceed $10^{\circ}$.
To remedy the situation, we might appeal to still lower disk surface densities, on the order of $\sim$$0.001 \gm/{\rm cm}^2$, for which dynamical friction cooling times for embedded planets would be as long as $\sim$$10^8\yr$. We found this region of parameter space difficult to explore. Out of 100 simulations starting with (a) $N_{\rm olig}=4$, (b) our compact set of initial conditions, and (c) $\sigma = 0.002 \gm/{\rm cm}^2$, only 6 yielded systems each with 2 surviving oligarchs at the end of the integrations at $t = 10^8\yr$. Unfortunately, 5 of these 6 systems had not stabilized, and it was unclear whether more oligarchs would be ejected were the integrations to continue. Furthermore, over these long timescales, effects resulting from time variations in $\sigma$ (see §\[sec\_reduce\]) might be expected to be important, and we have not modeled these.
Cooling times for planets and, by extension, KBO heating times might also be prolonged in more realistic treatments of dynamical friction that incorporate non-axisymmetries and the clearing of gaps in the disk. Ways of numerically simulating the response of planetesimal disks to planets can be found in Levison & Morbidelli (2007) and Lithwick & Chiang (2007).
We thank Ben Collins, Yoram Lithwick, Ruth Murray-Clay, Re’em Sari, and an anonymous referee for helpful exchanges. We also thank Harold Levison and Alessandro Morbidelli for generously sharing some of their ideas about planet-disk interactions and suggesting that we examine the angular momentum and energy budgets of our simulations. Support for E.B.F. was provided by a Miller Research Fellowship and by NASA through Hubble Fellowship grant HST-HF-01195.01A 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. E.C. acknowledges grants from the National Science Foundation (AST-0507805), NASA (JPL-1264475), and the Alfred P. Sloan Foundation.
Artymowicz, P. 1997, Annual Review of Earth and Planetary Sciences, 25, 175\
Binney, J., & Tremaine, S. 1987, [*Galactic Dynamics*]{} (Princeton: Princeton University Press), 424\
Butler, R.P., et al. 2006, , 646, 505\
Chambers, J.E. 2006, Icarus, 180, 496\
Chambers, J.E. 1999, , 304, 793\
Chambers, J.E., Wetherill, G.W., & Boss, A.P. 1996, Icarus, 119, 261\
Chiang, E.I., & Lithwick, Y. 2005, , 628, 520\
Chiang, E.I., et al. 2006, in [*Protostars and Planets V*]{}, eds. B. Reipurth, D. Jewitt, & K. Keil (Tucson: Univ. Arizona Press), in press (astro-ph/0601654) (C06)\
Collins, B.F., & Sari, R. 2006, , 132, 1316\
Collins, B.F., Schlichting, H.E., & Sari, R. 2007, , in press\
Dohnanyi, J.W. 1969, JGR, 74, 2531\
Elliot, J.L., et al. 2005, , 129, 1117\
Ford, E.B. 2006, in [*New Horizons in Astronomy: Frank N. Bash Symposium*]{}, ASP Conference Series Vol. 352 (San Francisco: Astronomical Society of the Pacific), 15\
Ford, E.B., Lystad, V., & Rasio, F.A. 2005, Nature, 434, 873\
Ford, E.B., Rasio, F.A., & Yu, K. 2003, in [*Scientific Frontiers in Research on Extrasolar Planets*]{}, ASP Conference Series Vol. 294, eds. D. Deming & S. Seager (San Francisco: Astronomical Society of the Pacific), 181\
Goldreich, P., Lithwick, Y., & Sari, R. 2004, , 42, 549 (GLS04)\
Goldreich, P., & Sari, R. 2003, , 585, 1024\
Goldreich, P., & Tremaine, S. 1982, , 20, 249\
Greenberg, R., et al. 1991, Icarus, 94, 98\
Ida, S., & Makino, J. 1993, Icarus, 106, 210\
Kenyon, S.J., & Luu, J.X. 1999, 118, 1101\
Levison, H.F., & Stewart, G.R. 2001, Icarus, 153, 224\
Levison, H.F., et al. 2006, in [*Protostars and Planets V*]{}, eds. B. Reipurth, D. Jewitt, & K. Keil (Tucson: Univ. Arizona Press), in press\
Levison, H.F., & Morbidelli, A. 2007, Icarus, in press\
Li, A., & Lunine, J. I. 2003, , 590, 368\
Lissauer, J.J., Pollack, J.B., Wetherill, G.W., & Stevenson, D.J. 1995, in [*Neptune and Triton*]{}, ed. D. Cruikshank (Tucson: Univ. Arizona Press), 37\
Lithwick, Y., & Chiang, E. 2007, , in press (astro-ph:0607241)\
Marzari, F., & Weidenschilling, S.J. 2002, Icarus, 156, 570\
Matsuyama, I., Johnstone, D., & Hartmann, L. 2003, , 582, 893\
Murray-Clay, R.A., & Chiang, E.I. 2006, , 651, 1194\
Rafikov, R.R. 2004, , 128, 1348\
Shu, F.H., Johnstone, D., Hollenbach, D. 1993, Icarus, 106, 92\
Stewart, G.R., & Ida, S. 2000, Icarus, 143, 28\
Strubbe, L.E., & Chiang, E.I. 2006, , 648, 652\
Thommes, E.W., Duncan, M.J., & Levison, H.F. 1999, Nature, 402, 635\
Thommes, E.W., Duncan, M.J., & Levison, H.F. 2002, , 123, 2862\
Tsiganis, K., et al. 2005, Nature, 435, 459\
Wisdom, J., & Holman, M. 1991, , 102, 1528\
Youdin, A.N., & Chiang, E.I. 2004, , 601, 1109\
Youdin, A.N., & Shu, F.H. 2002, , 580, 494\
[^1]: The problem does not disappear by merely raising the disk mass above the minimum-mass value, since the gravitationally unfocussed growth rate scales only linearly with the disk surface density.
[^2]: Chambers et al. (1996) provide fitting formulae for $t_{\rm unstable}$ as functions of oligarch mass and orbital spacing, but only for the case without the gas giants Jupiter and Saturn.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Object pose increases interclass object variance which makes object recognition from 2D images harder. To render a classifier robust to pose variations, most deep neural networks try to eliminate the influence of pose by using large datasets with many poses for each class. Here, we propose a different approach: a class-agnostic object pose transformation network (OPT-Net) can transform an image along 3D yaw and pitch axes to synthesize additional poses continuously. Synthesized images lead to better training of an object classifier. We design a novel eliminate-add structure to explicitly disentangle pose from object identity: first ‘eliminate’ pose information of the input image and then ‘add’ target pose information (regularized as continuous variables) to synthesize any target pose. We trained OPT-Net on images of toy vehicles shot on a turntable from the iLab-20M dataset. After training on unbalanced discrete poses (5 classes with 6 poses per object instance, plus 5 classes with only 2 poses), we show that OPT-Net can synthesize balanced continuous new poses along yaw and pitch axes with high quality. Training a ResNet-18 classifier with original plus synthesized poses improves mAP accuracy by 9% over training on original poses only. Further, the pre-trained OPT-Net can generalize to new object classes, which we demonstrate on both iLab-20M and RGB-D. We also show that the learned features can generalize to ImageNet.'
author:
- Anonymous ECCV submission
- Yunhao Ge
- Jiaping Zhao
- Laurent Itti
bibliography:
- 'OPT-Net.bib'
title: 'Pose Augmentation: Class-agnostic Object Pose Transformation for Object Recognition'
---
![Object pose transformation with OPT-Net. The first column shows input images from the test dataset, and the remaining columns show target pose images transformed by OPT-Net. Integer poses (1, 2, 3, 4, 5, 6 in red) are defined in the training dataset, while decimal poses (1.5, 2.5, 3.5, 4.5, 5.5 in green) are new poses, which shows OPT-Net can achieve continuous pose transformation. []{data-label="fig1"}](figure1.png){height="2in"}
Introduction and related work
=============================
In object recognition from 2D images, object pose has a significant influence on performance. An image depends on geometry (shape), photometry (illumination and material properties of objects) and dynamics (as objects move) of the scene. Thus, every image is a mixture of instance-specific information and nuisance factors [@ma2012invitation], such as 3D viewpoint, illumination, occlusions, shadows, etc. Nuisance factors often depend on the task itself. Specifically, in object recognition from 2D images, we care for instance-specific information like shape, while the dynamics of pose is a nuisance that often degrades classification accuracy [@ma2012invitation].
Deep convolution neural networks (CNNs) have achieved great success in object recognition [@krizhevsky2012imagenet; @simonyan2014very; @he2016deep; @szegedy2017inception; @huang2017densely] and many other tasks, such as object detection [@girshick2014rich; @ren2015faster; @redmon2016you; @fang2017rmpe], image segmentation [@ronneberger2015u; @milletari2016v; @he2017mask], etc. Most research tries to discount pose, by eliminating pose information or improving pose robustness of a classifier. Typical CNN architectures, such as LeNet [@lecun2015lenet] AlexNet [@krizhevsky2012imagenet] and VGG [@simonyan2014very] use convolution layers and pooling layers to make the high-level feature representations invariant to object pose over some limited range [@zhao2017learning]. In contrast, recent results have shown that explicitly modeling pose information can help an object recognition task [@zhao2017learning; @bengio2013representation; @wohlhart2015learning; @bakry2014untangling]. Some approaches use multi-task learning where pose information can be an auxiliary task or regularization to improve the main object recognition task [@zhang2014facial; @zhang2014improving; @huang2013multi; @su2015render]. These neural networks have the potential to disentangle content from their instantiation attributes [@ranzato2007unsupervised; @zhao2015stacked; @goroshin2015learning]. Training on multiple views of the object can improve recognition accuracy [@su2015multi]. A common method is collecting all poses of the object and creating a pose-balanced dataset, with the hope that pose variations will average out. However, collecting pose-balanced datasets is hard and expensive. One notable such dataset is iLab-20M which comprises 22 million images of 704 toy vehicles captured by 11 cameras while rotating on a turntable [@borji2016ilab]. Here, we use a subset of this data to learn about pose transformations, then transferring this knowledge to new datasets (RGB-D [@lai2011large], ImageNet [@deng2009imagenet]).
2D images can be seen as samples of 3D poses along yaw and pitch axes (Fig.2(a)). We want our OPT-Net to imitate the 3D pose transformation along these two axes. Thus given any single pose image, we can ’rotate’ the object along yaw and pitch axes to any target pose. Instead of directly training a transformation model to continuously ’rotate’ images, we start with a discrete transform, which is easier to constrain. Then we can make the pose representation continuous and normalize the continuous transform process. Here, we use sampled discrete poses along yaw and pitch as our predefined poses (Fig.2(b), 6 poses along the yaw axis and 3 poses along pitch axis). We treat different object poses as different domains so that discrete pose transformation can be seen as an image-to-image translation task, where a generative model can be used to synthesize any target pose given any input pose. Recently, Generative Adversarial Networks (GAN) [@goodfellow2014generative] have shown a significant advantage in transforming images from one modality into another modality [@mirza2014conditional; @isola2017image; @sangkloy2017scribbler; @zhu2017unpaired; @kim2017learning]. GANs show great performance in various tasks, such as style transfer [@chang2018pairedcyclegan; @karras2019style], domain adaptation [@hoffman2017cycada; @tzeng2017adversarial], etc. However, there is a high cost in our task, because we should train specific GANs for all pairs of poses [@bhattacharjee2018posix]. StarGAN [@choi2018stargan] and CollaGAN [@lee2019collagan] proposed a method for multi-domain mapping with one generator and showed great results in appearance changes such as hair color, age, and emotion transform. However, pose transform creates a large, nonlinear spatial change between input and output images. The traditional structure of the generators (Unet [@ronneberger2015u], Vnet [@milletari2016v]) has few shared structures which satisfy all randomly paired pose transformation. It makes StarGAN training hard to converge (see Exp 4.1).
Learning a better representation could also reduce variance due to pose. [@zhu2014multi] tried to learn better representation features to disentangle identity rotation and view features. InfoGAN [@chen2016infogan] learns disentangled representations in an unsupervised manner. [@kan2016multi] seeks a view-invariant representation shared by views.
To combine the idea of better representation and multi-domain image transformation, we propose a class-agnostic object pose transformation neural network (OPT-Net), which first transforms the input image into a canonical space with pose-invariant representation and then transform it to the target domain. We design a novel eliminate-add structure of the OPT-Net and explicitly disentangle pose from object identity: OPT-Net first ‘eliminates’ the pose information of the input image and then ‘adds’ target pose information to synthesize any target pose. Convolutional regularization is first used to implicitly regularize the representation to keep only the key identification information that may be useful to any target pose. Then, our proposed pose-eliminate module can explicitly eliminate the pose information contained in the canonical representation by adversarial learning. We also add a discriminator leveraging pose classification and image quality classification to supervise the optimization of transforming.
Overall our contributions are multifold: (1) developed OPT-Net, a novel class-agnostic object pose transformation network with an eliminate-add structure generator that learns the class-agnostic transformation among object poses by turning the input into a pose-invariant canonical representation. (2) design a continuous representation of 3D object pose and achieve continuous pose transforming in 3D, which can be learned from limited discrete sampled poses and adversarial regularization. (3) demonstrated the generative OPT-Net significantly boosts the performance of discriminative object recognition models. (4) showed OPT-Net learns class-agnostic pose transformations, generalizes to out-of-class categories and transfers well to other datasets like RGB-D and ImageNet.
![(a) Discrete predefined pose images sample . (b) Predefined sample poses and pose change along pitch and yaw axes. (c) Given any pose (first column), OPT-Net can transform it along pitch and yaw axes to target poses (remaining columns)[]{data-label="fig:example"}](figure2.png){height="3cm"}
Object Pose Transforming Network
================================
As shown in Fig. 3, the proposed OPT-Net has an eliminate-add structure generator, a discriminator and a pose-eliminate module.
Eliminate-add structure of the generator
----------------------------------------
![Flow of OPT-Net, consisting of three modules: eliminate-add structure generator $G$, discriminator $D$, and pose-eliminate module. (a) Pose transformation sketch (b) Origin to target pose transformation. In the pose ‘eliminate’ part, $G$ takes in the original pose image and first uses both implicit regularization and the explicit pose-eliminate module to eliminate pose information of the input, yielding a pose-invariant canonical representation. Then, in the pose ‘add’ part, the representation features are concatenated with a target pose mask and the target pose image is synthesized. $D$ learns to distinguish between real and fake images and to classify real images to their correct pose. (c) Training OPT-Net: $G$ first maps the original pose image to target pose and synthesizes a fake image, then $G$ tries to reconstruct the original pose image from the fake image given the original pose information.[]{data-label="fig:example"}](figure3.png){width="\linewidth"}
The generator ($G$) of OPT-Net transforms an input object pose image x into a target object pose y conditioned on the target pose label $c$, $G(x,c)\to$y$ $. Different from the hair color, gender and age transform, which have more appearance transfer with smaller shape changes, object pose transformation creates large shape differences. Our eliminate-add structure generator (Fig.3 (a)) first turns the input pose image into a pose-invariant canonical representation by ‘eliminating’ pose information, and then ‘adds’ target pose information to turn the representation into the target pose. As shown in Fig. 3(b), given an input image, we randomly select the target pose domain. We do not input target pose along with the input image. Instead, in the ‘eliminate’ part, the first several convolution layers with stride $s>2$ are used to implicitly regularize the preserved representation features. This implicit regularization makes the representation features contain only key information for the transformation (appearance, color, shape), and eliminates useless information which may hinder transformation (pose). At the same time (Fig. 3(b)), the ‘pose-eliminate module’ ($P_{elim}$) explicitly forces the representation to contain as little pose information as possible, by predicting equal probability for every pose. After both implicit and explicit elimination of pose information, the input image is turned to a pose-invariant canonical representation space. We then ‘add’ the target pose information by concatenating it with the representation feature map. The remaining layers in the generative model transform the concatenated features into the target pose image. This eliminate-add structure is shared and can be used for any pose transformation. This shared structure makes the generator easy to converge. To control the translation direction, as shown in Fig. 3(b), we use an auxiliary classifier as discriminator $D$ to guide the image quality and pose transform. Given one image, the discriminator has two outputs, the probability that the input image is real, which represents the quality of the synthesized image, and the output pose, which should match the desired target pose, $D:x\to\{ D_{src}(x), D_{cls} (x)\} $
Pose-eliminate module
---------------------
The pose-eliminate module ($P_{elim}$) takes the preserved representation feature $x_r$ as input and outputs pose classification $\{ P_{elim}(x_r) \}$. $P_{elim}$ can be treated as a discriminator which forms an adversarial learning framework with the ‘eliminate’ part of the generator ($G_{elim}$). The canonical representation features of real images with pose labels are used to train $P_{elim}$. We use Cross-Entropy loss to make $P_{elim}$ predict the correct pose from the pose-invariant feature after $G_{elim}$. Different from traditional adversarial training, when using $P_{elim}$ to train $G_{elim}$, we want the generator to eliminate all pose information in the pose-invariant feature, which makes $P_{elim}$ produce equal output probability for every pose. We use the uniform probability ($1/N$) as the ground truth label to compute the pose-eliminate loss, which is used to optimize the $G_{elim}$.
Continuous pose transforming training
-------------------------------------
We design a 2-dimension linear space to represent pitch and yaw values, in which we could interpolate and achieve continuous pose representation (Fig.1) . The yaw and pitch values can be duplicated as a matrix that has the same dimension as the canonical representation features, which is easy to be concatenated. The mask vector has N (we use 3) dimensions along the channel dimension which can be adjusted depending on the canonical representation features channel. We start the training on discrete sampled poses (which can be represented as integer in linear space). After the network has converged, we randomly sample decimal poses as target poses and use a style consistency loss to regularize the synthesized images, which keeps pose representation consistent along yaw and pitch axes.
Loss Function
-------------
Our goal is to train a generator $G$ that learns object pose transformations along yaw and pitch axes. The overall loss is formed by adversarial loss, domain classification loss, reconstruction loss, pose-eliminate loss and style consistency loss.
[**Adversarial Loss.**]{} The adversarial loss is used to make the synthesized image indistinguishable from real images. $$\label{Eq.1}
L_{adv}=E_{x}[logD_{src}(x)] + E_{x,c}[log(1-D_{src}(G(x,c)))]$$ $ D_{src}(x) $ represent the probability that input x belongs to the real images given by $D$. The generator $G$ tries to minimize the loss, while the discriminator $D$ tries to maximize it.
[**Pose Classification Loss.**]{} The pose classification loss is used to guide the pose transformation which makes the synthesized image $y$ belong to the target pose $c$. This pose classification loss is used to optimize both $D$ and $G$. The pose classification loss of $D$ is defined as $$\label{Eq.2}
L_{cls}^{r}=E_{x,c'}[-logD_{cls}(c'|x)]$$ The loss for $D$ is similar to a traditional Cross-Entropy loss for classification, where$ D_{cls}(c'|x) $ means the predicted probability of real image x belongs to the ground truth pose label $c' $ . The pose classification loss of $G$ is defined as $$\label{Eq.3}
L_{cls}^{f}=E_{x,c}[-logD_{cls}(c|G(x,c))]$$ $G$ tries to minimize this loss to make the synthesized fake image $G(x,c) $ be classified as the target pose $c$.
[**Reconstruction Loss.**]{} To make the synthesized image preserve the content information and change only the object pose, as shown in fig3(c), we use the cycle consistency loss [@zhu2017unpaired] to optimize $G$. $$\label{Eq.4}
L_{rec}=E_{x,c,c'}[\|x-G(G(x,c)|c')\|_{1}]$$ where $G$ can reconstruct the original image $x$ by transforming the synthesized fake target pose image $G(x,c) $ back to the original pose $c'$ . $L1$ norm is used as reconstruction loss.
[**Pose-eliminate Loss.**]{} In the eliminate-add structure of $G$, to eliminate the pose information in preserved canonical representation features, we designed pose-eliminate loss to optimize the pose eliminate module ( $P_{elim}$) and the eliminate part of $G$,($G_{elim}$ ). The pose eliminate loss is $$\label{Eq.5}
L_{pose}^{P}=E_{x,c'}[-logP_{elim}(c'|G_{elim}(x))]$$ where $P_{elim}(c'|G_{elim}(x)) $ means the predicted probability of the canonical representation features of a real image belongs to the ground truth pose label $c' $ . The pose eliminate loss for $G_{elim}$ is defined as $$\label{Eq.6}
L_{pose}^{G}=-E_{x}{\sum_{c_{i}=1}^{N}1/N\cdot log(P_{elim}(c_{i}|G_{elim}(x)))}$$ where $N$ is the number of pose classes we defined, $c_{i} $ represent the pose label, $c_{i}\in [0, N) $ , $P_{elim}(c_{i}|G_{elim}(x))$ represent the probability of the synthesized canonical representation belongs to the $c_{i} $ pose. In ideal situations, the $P_{elim}$ can hardly predict the correct pose from canonical representation features and output equal probability for every pose, which means the pose information is eliminated in preserved canonical features. We use equal prediction of every pose to optimize $G_{elim}$ instead of minimizing the pose classification accuracy of to avoid a ‘cheated optimize’ that $P_{elim}$ tries to predict all input to a fixed pose class.
[**Style consistency Loss.**]{} After the converge of the previous loss, we randomly sample decimal target pose instead of all integers to make continuous pose transforming, the style consistency loss can regularize the synthesized images. The equation of style consistency loss is same as adversarial loss above, but the target pose is randomly sampled decimal value along yaw and pitch axes.
[**Full Loss Function.**]{} Finally, we optimize: $$\label{Eq.8}
L_{G}=L_{adv} + \lambda_{cls}L_{cls}^{f} + \lambda_{rec}L_{rec} + \lambda_{pose}L_{pose}^{G}$$ $$\label{Eq.9}
L_{D}=-L_{adv} + \lambda_{cls}L_{cls}^{r}$$ $$\label{Eq.10}
L_{P_{elim}}=L_{pose}^{P}$$ where $ \lambda_{cls}$, $ \lambda_{rec} $ and $ \lambda_{pose} $ are hyper-parameters that control the relative importance of classification, reconstruction, and pose-eliminate losses.
Experimental Methods
====================
Datasets
--------
[**iLab-20M dataset [@borji2016ilab].**]{} The iLab-20M dataset is a controlled, parametric dataset collected by shooting images of toy vehicles placed on a turntable using 11 cameras at different viewing points. There are in total 15 object categories with each object having 25 160 instances. Each object instance was shot on more than 14 backgrounds (printed satellite images), in a relevant context (e.g., cars on roads, trains on rail tracks, boats on water). In total, 1,320 images were captured for each instance and background combinations: 11 azimuth angles (from the 11 cameras), 8 turntable rotation angles, 5 lighting conditions, and 3 focus values (-3, 0, and +3 from the default focus value of each camera). The complete dataset consists of 704 object instances, with 1,320 images per object-instance/background combination, almost 22M images (18 times of ImageNet).
[**RGB-D dataset.**]{} The RGB-D Object Dataset consists of 300 common household objects organized into 51 categories. This dataset was recorded using a Kinect style 3D camera. Each object was placed on a turntable and video sequences were captured for one whole rotation. For each object, there are 3 video sequences, each recorded with the camera mounted at a different height so that the object is shot from different viewpoints.
Network Implementation
----------------------
Our OPT-Net consists of two parts, pose ‘eliminate’ part $G_{elim}$ and pose ‘add’ part $G_{add}$. As shown in Fig. 3(b), the $G_{elim}$ first has two convolution layers with stride size of two to downsample the input image. Then, three Resnet blocks [@he2016deep] form the backbone of $G_{elim}$. The output $G_{elim}(x)$ is the pose-invariant canonical representation feature. The canonical representation feature is copied to different streams, one concatenates with the target pose mask, forming the input of $G_{add}$ to synthesize the target pose image. The other one is treated as the input of $P_{elim}$ to predict the pose class. $G_{add}$ first has 6 blocks of Resnet as a backbone and ends with two up-sampling convolution layers to transform the canonical representation features to a target pose image, given a target pose information mask. For the discriminator $D$, we adopt the PatchGAN [@isola2017image] network.
$P_{elim}$ has a traditional classification network structure, which has the first two convolution layers with stride size of 2 to downsample the input features, followed with 3 Resnet blocks as backbone and ending with one down sampling convolution layers. In the end, the feature map is resized to N-dimensional output vectors and we use Softmax to obtain the prediction of pose class.
We use Wasserstein GAN objective with a gradient penalty \[1, 4\] to stabilize the training process. We adjust the$ \lambda_{pose} $ during training the generator, at the beginning epochs of training, improving the value of $ \lambda_{pose} $ can accelerate the convergence of generator, which makes the synthesized fake pose image have meaningful corresponding spacial structure. We gradually reduce the value of $ \lambda_{pose} $. At the last ending part of the training, $ \lambda_{pose} $ can be very small to make the optimization concentrate on improving the image quality.
Experiments and Results
=======================
We have five main experiments: in Section 4.1 on object pose transformation task, we compare OPT-Net with baseline StarGAN [@choi2018stargan] and CycleGAN [@zhu2017unpaired] by quantitatively and qualitatively comparing the synthesized object pose image quality. In Section 4.2, we use the OPT-Net as a generative model to help the training of a discriminative model for object recognition, by synthesizing missing poses and balancing a pose bias in the training dataset. In Section 4.3, we further show the class-agnostic transformation property of OPT-Net by generalizing the pretrained OPT-Net to new datasets. In Section 4.4, we study the influence of object pose information for objects which are mainly distinguishable by shape, as opposed to other features like color. Finally, in Section 4.5, we further demonstrate how the learned pose features in OPT-Net and object recognition model with the iLab-20M dataset can generalize to other datasets like ImageNet.
Object Pose Transformation Experiments
--------------------------------------
Because the baseline models can only do discrete pose transform, we fix the pitch value and use 6 different yaw viewpoints among the 88 different views of iLab-20M as our predefined pose to implement our OPT-Net. As is shown in Fig. 2, the selected 6 viewpoints have big spatial variance which can better represent the general object pose transformation task. In training set, each pose has nearly 26k images with 10 vehicle classes (Table 2). Each class contains 20$\sim$80 different instances. The test set has the same 10 vehicle categories, but different instances than the training set. Both training and test datasets are 256x256 RGB images.
![Object pose transform comparation between StarGAN, CycleGAN and OPT-Net.[]{data-label="fig:example"}](figure4.png){width="\linewidth"}
The training dataset is used to train our OPT-Net and the baseline models, StarGAN, and CycleGAN. Our OPT-Net has one generator, one discriminator and one pose-eliminate module; StarGAN has one generator and one discriminator, while CycleGAN needs 30 generators and 30 discriminators to achieve 6 pose transformations.
**Qualitative evaluation**. The experiment results are shown in Fig. 4. Compared with StarGAN, which struggles with large pose variations, the synthesized target pose images by OPT-Net are high quality with enough details, and the colors and appearances are similar to the real images. One possible reason is that our eliminate-add structure eliminates the conflicts between different directions on pose transformation and the pose-invariant representation breaks the barriers in making big spacial changes. Moreover, CycleGAN leads to inferior synthesized images in generalization on the test dataset. Thus our eliminate-add generator structure learns more generalized knowledge with less possible of over-fitting which can better generalized on new instances. Fig.1 shows more transformation results of OPT-Net.
**Quantitative evaluation**. The real target pose image of input is used as the ground truth label. To reduce background influence, we segment the foreground vehicle with the Graph-Based Image Segmentation method and only compute mean squared error (MSE) and peak signal to noise ratio (PSNR) of foreground between the synthesized target object pose and ground truth. (Table 1)
The result is the mean MSE and PSNR computed by 200 different instances, the MSE and PSNR for each instance is the average of 6 synthesized fake pose images. Table 1 demonstrates that the quality of synthesized target pose images by our OPT-Net is better than CycleGAN and StarGAN.
---------- -------------- ------------- ----------------------------------------------
CycleGAN StarGAN OPT-Net
Mean MSE 479.42 502.51 **[374.76]{}\
Mean PSNR & 21.94 & 21.95 & **[23.04]{}\
****
---------- -------------- ------------- ----------------------------------------------
: Average Mean squared error (MSE; lower is better) and peak-signal-to-noise ratio (PSNR; higher is better) for different methods
Object Recognition Experiment
-----------------------------
We design an object recognition experiment to explore the performance of OPT-Net as a generative model to help the training of a discriminative model. Two different training datasets are tailored from iLab-20M, pose-unbalanced (P-UB) and pose-balanced (P-B). In P-UB (Table 2), 5 classes of vehicles (boat, car, semi, tank, and van) have all 6 pose images (same poses as 4.1), while the other 5 classes (bus, military car, monster, pickup, and train) have only two poses (pose2 and pose5), which has significant pose bias. In P-B, each category among 10 classes of vehicles has all 6 pose images (no pose bias). The test dataset is a pose-balanced dataset which contains different instances of the 10 classes of vehicles that were not in either training dataset (P-UB and P-B). The classification neural network we used is Resnet-18 [@he2016deep] (no pre-training).
--------- --------- --------- --------- --------- --------- -------
Pose1 Pose2 Pose3 Pose4 Pose5 Pose6
boat
bus
car
mil
monster
pickup
semi
tank
train
van
--------- --------- --------- --------- --------- --------- -------
: Poses used in the pose-unbalanced (P-UB) training dataset to train OPT-Net
We first train the classification model on P-UB and P-B, calculating the test accuracy of each class of vehicles on the test dataset. To evaluate the performance of OPT-Net, we first train it on P-UB to learn the object transformation ability. After training, for each category in P-UB which have only pose2 and pose5 (bus, military car, monster, pickup, and train), we use the trained OPT-Net to synthesize the missing 4 poses (pose1, pose3, pose4, pose6). We combine the synthesized images with P-UB and form a synthesized-pose-balanced (S-P-B) training dataset. To show continuous transforms, we also interpolate pose values and synthesize 5 new poses beyond the predefined ones, and form a synthesized-additional-pose-balanced (SA-P-B) training dataset. S-P-B and SA-P-B were used to train the same resnet-18 classification model from scratch and to calculate test accuracy of each class of vehicles in the test dataset. We also use common data augmentation methods (random crop, horizontal flip, scale resize, etc) to augment the P-UB dataset to the same number of images as P-B, called A-P-UB (Table 3).
------------ --------- ----------- ----------------- --------------- --------------- ------
Dataset P-UB P-B S-P-B SA-P-B A-P-UB Test
Source real real synthesized synthesized augmented real
Number 25166 37423 37423 66041 37423 4137
------------ --------- ----------- ----------------- --------------- --------------- ------
: Different training and testing datasets for object recognition
The test accuracy of each class is shown in Table 4. From P-UB to S-P-B, the overall accuracy improved from 52.26% to 59.15%, which shows the synthesized missing pose images by OPT-Net can improve the performance of object recognition. It is also shown that OPT-Net, as a generative model, can help the discriminative model. Specifically, the vacant pose categories show significant improvement in accuracy: military improved by 11.68%, monster improved by 14.97%, pickup and train improved by 8.74% and 16.12% respectively. The comparison of S-P-B and A-P-UB shows that synthesized images by OPT-Net are better than traditional augmented images in helping object recognition. Because of the continuous pose transformation ability, our OPT-Net can synthesize additional poses different from the 6 poses in P-B. With these additional poses, SA-P-B (61.23%) performs even better than the P-U (59.20%), achieve 9% improvement compared with P-UB.
------------- ---------- ---------- ----------- ---------- --------
Category P-UB P-B S-P-B SA-P-B A-P-UB
boat 54.0 61.6 65.4 57.7 51.3
**bus** 35.2 42.5 38.1 47.8 37.2
car 85.1 76.3 79.8 64.0 78.9
**mil** 73.8 84.2 85.4 86.4 70.7
**monster** 45.3 67.4 60.2 66.0 52.9
**pickup** 17.8 26.7 26.6 36.5 18.7
semi 83.9 79.8 79.0 83.5 86.1
tank 78.1 69.4 78.6 77.0 72.5
**train** 41.1 65.1 57.2 58.1 43.1
van 23.6 18.6 24.2 20.7 21.0
**overall** 52.3 59.2 59.2 **61.2** 52.3
------------- ---------- ---------- ----------- ---------- --------
: Testing object recognition accuracy (%) of each class after trained on different training dataset. Comparing S-P-B and SA-P-B with P-UB shows how much classification improves thanks to adding synthesized images for missing poses in the training set, reaching or surpassing the level of when all real poses are available (P-B). Our synthesized poses yield better learning than traditional data augmentation (A-P-UB)
Class-agnostic Object Transformation Experiment
-----------------------------------------------
Our proposed OPT-Net can simultaneously make pose transformation on different classes of vehicles, which demonstrate that the learned object pose transformation has not fixed with object classes, it is a class-agnostic object pose transformation. To further explore the class-agnostic property of OPT-Net, we design experiments that generalize OPT-Net’s ability for object pose transformation from one dataset to other datasets.
15 categories of objects from RGB-D are used. They are both common household objects with big spatial variance between different object poses. Similar poses of objects in RGB-D are selected and defined as the same pose as iLab-20M. For each pose, RGB-D contains only about 100 images which cannot train our OPT-Net from scratch, thus we use RGB-D to finetune OPT-Net pre-trained on iLab-20M. We can see (Fig. 5) that our pre-trained OPT-Net can generalize well to other datasets, which demonstrates that OPT-Net is a class-agnostic object pose transformation framework.
![Generalization results of OPT-Net on RGB-D dataset pretrained on iLab-20M.[]{data-label="fig:example"}](figure5.png){width="\linewidth"}
To further explore the performance of OPT-Net as a generative model to help a discriminative model of object recognition, we split RGB-D into a pose-unbalanced (P-UB) training dataset, where each category randomly takes 3 poses among all 6 poses; pose-balanced (P-B), and test dataset similar to 4.2.
We first use P-UB to finetune the pretrained OPT-Net, and then use the trained OPT-Net to synthesize missing poses of household objects in RGB-D. The synthesized images and the original pose-unbalanced images form the synthesized pose balanced (S-P-B) training dataset. Similarly, to eliminate the influence of the number of training images, we created A-P-UB using common data augmentation methods. We trained Alexnet [@krizhevsky2012imagenet] on the 4 training datasets separately, and showed the test accuracy for each category in Table 5.
------------- --------- ---------- ---------- --------
Dataset P-UB P-B S-P-B A-P-UB
Accuracy(%) 99.1 99.9 **99.7** 99.2
------------- --------- ---------- ---------- --------
: Overall object recognition accuracy for different training dataset in RGB-D
The (small) accuracy improvement in S-P-B compared with P-UB demonstrates that our pretrained OPT-Net can be generalized to different datasets after finetune, which can help the discriminative model in object recognition. While the overall improvement is small, below we show that this is not the case uniformly across all object categories.
Object Pose Significance on Different Object Recognition Tasks
--------------------------------------------------------------
Because the accuracy improvement in RGB-D is smaller than in iLab-20M, we tested whether this was the case across all object categories, or whether those which look more alike would benefit more from synthesized images from OPT-Net. Indeed, maybe classifying a black keyboard vs. a blue stapler can easily be achieved by size or color even without pose-dependent shape analysis. To verify our hypothesis, we use the confusion matrix of classification to select categories which are more confused by classifier: marker, comb, toothbrush, stapler, lightbulb, and sponge. We then assign different fixed poses to each category to improve overall pose variance and form P-UB-1 (randomly fix 1 pose for each category), P-UB-2 (randomly fix 2 poses for each category), and P-UB-3 (randomly fix 3 poses for each category) pose-unbalanced datasets (suppl. Table 1).
Similarly, we create 3 other training datasets using the same method as in 4.2 and 4.3: (S-P-B: use pretrained OPT-Net to synthesize the missing poses; P-B, and A-P-UB for each unbalanced datasets), and report the object recognition performance on the test dataset in Table 6.
-------------- ----------- ------------- ------------ ----------- ------------- ----------
Dataset P-UB-1 A-P-UB-1 S-P-B-1 P-UB-2 A-P-UB-2 S-P-B-2
Accuracy(%) 75.1 77.6 **83.2** 90.4 91.2 **94.2**
Dataset P-UB-3 A-P-UB-3 S-P-B-3 P-B
Accuracy(%) 99.3 99.2 **99.4** 99.8
-------------- ----------- ------------- ------------ ----------- ------------- ----------
: Object recognition overall accuracy for different datasets
The results in Table 6 demonstrate that object pose information has different degrees of impact on the object recognition task. Compared with the results in 4.3, where the improvement between P-UB and S-P-B is less than 1%, here, when the class variance is small, OPT-Net can improve more accuracy after synthesizing the missing poses in the unbalanced dataset. The accuracy improvement in experiment group 1 (P-UB-1 and S-P-B-1) is 8.1%. This result verified our hypothesis that pose balance is more important in small between-class variance object cognition tasks.
Meanwhile, comparing the different accuracy improvements in different experimental groups, group 2 (P-UB-2 and S-P-B-2) is 3.8%, while group 3 (P-UB-3 and S-P-B-3) is 0.1%. This demonstrates that when class-variance is fixed, the more pose bias we have, the more accuracy improvement we will get with the help of our OPT-Net pose transformation.
------------------------- ----------- ----------- ----------- -----------
Pretrain dataset boat car mil tank
No pretrain 0 0 0 0
S-P-B (StarGAN) 16.33 41.84 17.20 6.21
S-P-B (OPT-Net) **31.14** **76.20** **30.47** **12.86**
------------------------- ----------- ----------- ----------- -----------
: Generalization to ImageNet (percent correct)
Generalization to Imagenet
--------------------------
We directly use the pretrained OPT-Net on iLab-20M to synthesize images of different poses on ImageNet (suppl. Fig. 1). Results are not as good and might be improved using domain adaptation in future work. However, the discriminator of OPT-Net makes decent prediction of image poses: Fig. 6 shows the top 8 ImageNet images for each of our 6 poses.
![Top 8 ImageNet images for each pose predicted by discriminator in OPT-Net without finetune.[]{data-label="fig:example"}](figure6.png){height="3cm"}
To test object recognition in ImageNet, we replace real images by OPT-Net synthesized images in S-P-B (4.2) and form a S-P-B (OPT-Net) dataset (all synthesized images). Similarly, we use StarGAN synthesized images form S-P-B (StarGAN). We use a resnet18 10-class vehicles classifier pretrained with this two synthesized datasets and predict 4 classes of vehicles in ImageNet which have similar meanings as iLab-20M, with good results on some classes like car (Table 7). To further explore generalization, we pretrian an AlexNet on S-P-B which synthesized pose images by StarGAN and OPT-Net respectively and then finetune it on ImageNet. Table 8 shows significantly better accuracy compared to training from scratch when using only a small number of images per class, demonstrating generalization from iLab-20M to ImageNet.
------------------------------------------------------- ----------- ----------- -----------
Number of images per class 10 20 40
Alexnet (scratch) 4.32 17.01 25.35
Alexnet (pretrain on S-P-B (StarGAN)) 7.57 18.91 27.14
Alexnet (pretrain on S-P-B (OPT-Net)) **10.94** **19.75** **27.82**
------------------------------------------------------- ----------- ----------- -----------
: Top-5 recognition accuracy (%) on ILSVRC-2010 test set
Conclusions
===========
We proposed OPT-Net, a class-agnostic object pose transformation network (OPT-Net) to synthesize any target poses continuously given a single pose image. The proposed eliminate-add structure generator can first eliminate pose information and turn the input to a pose-invariant canonical representation, then adding the target pose information to synthesize the target pose image. OPT-Net also gives a more common framework to solve big variance continuous transformation problems. OPT-Net generated images have higher visual quality compared to existing methods. We also demonstrate that the OPT-Net, as a generative model can help the discriminative model in the object recognition task, which achieve a 9% accuracy improvement. We design experiments to demonstrate that pose balance is more important in small between-class variance object cognition tasks. Finally, we demonstrate the learned pose features in OPT-Net with the iLab-20M dataset can better generalize to other datasets like ImageNet.
Acknowledgement
===============
This work was supported by the National Science Foundation (grant number CCF-1317433), C-BRIC (one of six centers in JUMP, a Semiconductor Research Corporation (SRC) program sponsored by DARPA), and the Intel and CISCO Corporations. The authors affirm that the views expressed herein are solely their own, and do not represent the views of the United States government or any agency thereof.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the non-collinear exchange coupling across the trilayer magnetic junction composed of an intermediate layer with Rashba interaction and two sandwiching ferromagnetic ones. To compute the mediated exchange coupling, one needs to go beyond the single-particle argument and integrate over the contributions from the whole Fermi surface. Surprisingly, we find that the topology of the Fermi surface plays a crucial role in determining whether the oscillatory RKKY or the spiral interactions would dominate. At the end, we discuss the connection of our numerical results to experiments and potential applications.'
author:
- 'Wen-Min Huang'
- 'Cheng-Hung Chang'
- 'Hsiu-Hau Lin'
title: 'Non-collinear Exchange Coupling in Trilayer Magnetic Junction and its Connection to Fermi Surface Topology'
---
The central theme of spintronics is to manipulate the extra spin degrees of freedom in condensed matter systems[@Wolf01; @Zutic04; @MacDonald05; @Sun04b; @Sun06], as compared with the traditional electronic devices where only the charge part was utilized. One of the classic examples, which merges charge and spin sectors together in a single device is the spin field effect transistor (SFET) proposed by Datta and Das[@Datta89] more than a decade ago. It was suggested that the Rashba interaction, whose strength is controlled by the gate voltage, causes the spins of the itinerant carriers to spiral and can be used to modulate the transport currents. Another more recent proposal, now under the name of spin Hall effect[@Murakami03; @Sinova04], explores the possibility to generate spin currents (or spin accumulations) by electric gates via spin-orbital interactions and has received some primitive verifications in experiments.
On the other hand, it is less explored how the spin-orbital interaction will reshape our understanding in the more conventional magnetic junctions[@Sun04a]. In this Letter, we study the non-collinear exchange coupling across the F/N/F trilayer magnetic junction (TMJ) as shown in Fig. \[TMJ\], where the intermediate layer consists of semiconductors (such as GaAs) with significant Rashba interaction. Since the TMJ resembles the Datta-Das SFET, one may naively guess that the carrier-mediated exchange coupling can be explained following similar arguments. Within the single-particle picture, Datta and Das had demonstrated the spins of the itinerant carriers undergo precessions due to the Rashba interaction and, therefore, lead to the non-collinear coupling between the ferromagnetic layers.
![\[TMJ\] (Color online) Schematic plot for the trilayer magnetic junction where $\phi_x(r)$ is the non-collinear angle between the ferromagnets. ](fig1){width="7cm"}
However, there is a sharp difference between SFET and TMJ. In the SFET, the time-reversal symmetry is broken (by the source-drain bais which drives the current) while, in the TMJ, the symmetry is preserved and gives rise to Kramers degeneracy. In fact, this non-trivial degeneracy has a profound influence upon the effective exchange coupling across the junction. It turns out that the single-particle picture (with specific momentum) employed in Datta-Das’ original work[@Datta89] fails to explain the magnetic behavior because the inclusion of the [*whole*]{} Fermi surface is crucially important. For instance, the quantum interferences between the Kramers-degenerate patches of the Fermi surfaces give rise to the oscillatory RKKY interaction. To explore the subtle competition between the spiral and the RKKY interactions, one needs to integrate over the whole Fermi surfaces by numerical approach.
The outcome is rather surprising – the dominance of either spiral or RKKY interactions depends on the topology of the Fermi surface. When the Rashba interaction is weak (compared with the Fermi energy), the Fermi surface consists of two cocentered circles with opposite chiralities. In this regime, the RKKY interaction dominates over the spiral. However, as one gradually increases the strength of the Rashba interaction, the inner Fermi circle shrinks to zero first, then reappears again but with the same chirality as the outer circle (as shown in Fig. \[topology\]). The topological change of the Fermi surface magically alters the dominant coupling from the of RKKY to the spiral. Therefore, in the strong Rashba limit, the non-collinear exchange coupling mainly comes from the spiral interaction with minor quantum corrections. It is rather amazing that the transition between different magnetic behaviors coincides with the change of the Fermi surface topology.
![\[topology\] (Color online) Band structure of the Rashba Hamiltonian and the Fermi surface topology. In weak Rashba regime, the Fermi surface consists of two concentric circles with opposite chiralities while there is only one chirality present in strong Rashba regime. ](fig2){width="7.5cm"}
In the following, we present the analytic arguments and numerical results which support the claims we made in above. First of all, we model the intermediate layer by 2D electron gas with Rashba interaction, $$\label{TH}
H = \int d^2r\: \Psi^{\dag} \left[\frac{k^2}{2m^*}\textbf{1}
+\gamma_{R} (k_y\sigma^x- k_x\sigma^y)\right] \Psi,$$ where $\gamma_{R}$ is the strength of the Rashba interaction and $\Psi^{\dag}, \Psi$ are the two-component spinors of the creation/annhilation operators for itinerant carriers. The Rashba Hamiltonian can be brought into its eigenbasis in momentum space, $$\varphi_{k\lambda}(\vec{r}) = e^{i\vec{k}\cdot\vec{r}} u_{\lambda}(\phi)
= \frac{e^{i\vec{k}\cdot\vec{r}}}{\sqrt{2}}\left(
\begin{array}{c}
-i\lambda e^{-i\theta_{k}}\\
1
\end{array}\right)$$ with $\theta_{k} = \tan^{-1} (k_{y}/k_{x})$. Due to the spin-orbital interaction, spin is no longer the good quantum number but replaced by the chirality instead, $$\begin{aligned}
\lambda = (\hat{\bm{k}} \times \hat{\bm{s}})\cdot \hat{\bm{z}} = \pm 1,\end{aligned}$$ where the hat denotes the unit vector. It is important to remind the readers that, under the time reversal transformation, both momentum and spin reverse their directions and make the chirality invariant.
After integrating out the itinerant carriers[@Sun04a], the exchange coupling between the ferromagnetic layers is described by an effective Heisenberg Hamiltonian, $H_{\rm eff} = \sum_{ij} J_{ij} S^{i}_{L}S^{j}_{R}$. Within the linear response theory, the mediated exchange coupling is proportionally to the spin susceptibility tensor, $$\label{sus1}
\chi_{ij}(\textbf{r})=\int_{0}^{\infty}dt\left\langle\left\langle i
\left[\sigma^{i}(\vec{r},t),\sigma^{j}(0,0)\right]\right\rangle\right\rangle
e^{-\eta t}.$$ Here $\sigma^{i}(\vec{r},t)=\sum_{\alpha\beta}\psi^{\dag}_{\alpha}
(\vec{r},t)\sigma^{i}_{\alpha\beta}\psi_{\beta}(\vec{r},t)$ is the spin density operator for the itinerant carriers. Besides, $\langle\langle...\rangle\rangle \equiv \textrm{tr} [e^{-\beta H} ...]$ represents the thermal average at finite temperature and $\eta$ is the spin relaxation rate. Transforming into the eigenbasis, the susceptibility tensor can be expressed as summations of the product of a weight function and the particle-hole propagator over all possible quantum numbers, $$\label{susr}
\chi_{ij}(\vec{r}) = \sum_{k_1 \lambda_{1}}\sum_{k_2 \lambda_{2}}
W_{ij}(\vec{r}) \:
\left[\frac{f(\epsilon_{k_{1}\lambda_{1}})-f(\epsilon_{k_{2}\lambda_{2}})}{\epsilon_{k_{2}\lambda_{2}}-\epsilon_{k_{1}\lambda_{1}}-i\eta}
\right],$$ where $\epsilon_{k\lambda} = k^2/2m^*- \lambda k\gamma_R$ is the dispersion for the particle with momentum $k$ and chirality $\lambda$. The weight function inside the summations is $$W_{ij}(\vec{r}) =
e^{i(\vec{k_2}-\vec{k_1}) \cdot \vec{r}}
( u^{\dag}_{\lambda_1} \sigma^i u^{}_{\lambda_2}) ( u^{\dag}_{\lambda_2} \sigma^j u^{}_{\lambda_1}).$$
While the derivation of the spin susceptibility tensor is straightforward, it still requires rather involved numerics. However, the numerical task can be greatly reduced by various symmetry arguments. Let’s take the component $\chi_{xy}(\vec{r}) = \chi_{xy}(r,\theta)$ as a working example. Since the operators $\sigma_x, \sigma_y$ carry the 2D angular momentum $m = \pm 1$, making use of the SO(2) rotational invariance, the corresponding susceptibility $\chi_{xy}(r,\theta)$ contains linear combinations of $m=0, \pm 2$ by tensor analysis. That is to say, $\chi_{xy}(r,\theta) = f_0(r) + f_2(r) \cos 2\theta + g_2(r) \sin 2\theta$, where $f_0(r), f_2(r), g_2(r)$ are some functions without angular dependence. Furthermore, applying the parity symmetry in $y$ direction, it requires $\chi_{xy}(r,\theta) = -\chi_{xy}(r,-\theta)$ and enforces the functions $f_0(r), f_2(r)$ to vanish. Finally, the Onsager relation from the time-reversal symmetry indicates $\chi_{yx}(r,\theta) = \chi_{xy}(r,\theta+\pi) = g_2(r) \sin 2\theta$. Utilizing the rotational SO(2), parity $P_y$ (or equivalently $P_x$), and time reversal symmetries, one can work out the remaining components of the susceptibility tensor, $$\begin{aligned}
\chi_{ij}(r,\theta)=\left[
\begin{array}{ccc}
g_0+g_2\cos2\theta & g_2\sin2\theta & g_1\cos\theta \\
g_2\sin2\theta & g_0-g_2\cos2\theta & g_1\sin\theta \\
-g_1\cos\theta & -g_1\sin\theta & h_0 \end{array} \right].\end{aligned}$$ It is rather remarkable that the symmetry arguments make the angular dependence explicit and reduce the numerical task down to evaluation of four real scalar functions, $g_0(r)$, $g_1(r)$, $g_2(r)$ and $h_0(r)$. The Rashba Hamiltonian we study here further constrains $h_0(r) = g_0(r)+g_2(r)$, which reduces the number down to three.
![\[rkky\] (Color online) The spiral angle in weak Rashba regime with $k_R/k_F = 0.042$ at $T=30$ K. The smooth minus $\pi$-jumps originate from the opposite tendency of angle evolution of the Rashba spiral and RKKY effect. ](fig3){width="6cm"}
Suppose the ferromagnet on the left of the TMJ is aligned along the $z$-axis, we are interested in the mediated non-collinear exchange coupling proportional to $\chi_{iz}(r,\theta=0)$, where $r$ is the width of intermediate layer. Since $\chi_{yz}(r,0)=0$, the induced moment is captured by the spiral angle (shown in Fig. \[TMJ\]), $$\begin{aligned}
\phi_x(r) = \tan^{-1} \left[ \frac{\chi_{zz}(r,0)}{\chi_{xz}(r,0)} \right]
= \tan^{-1} \left[ \frac{g_0(r)+g_2(r)}{g_1(r)} \right].\end{aligned}$$ For realistic materials[@Luo; @Das; @Nitta; @Engle; @Heida98], we choose the spin splitting $\Delta_{R} \equiv 2 k_F \gamma_R = 5$ meV and the Fermi energy $\epsilon_F = 60$ meV. Or equivalently, it corresponds to the Rashba coupling $\gamma_R=8.91 \times 10^{-12}$ eV m and the carrier density $n_{2D} = 1.25 \times 10^{12}$ cm$^{-2}$. Thus, it falls into the weak Rashba regime, characterized by the dimensionless parameter $k_R/k_F = 0.042 \ll 1$, where $k_R=2 m^* \gamma_R$ and $k_F$ is the Fermi momentum in the absence of Rashba splitting. From Fig. \[rkky\], it is clear that the non-collinear angle between the ferromagnets $\phi_x(r)$ shows RKKY-like oscillations with the gradual upswing trend due to the Rashba interaction. The numerical results, drastically different from the spin-precession argument in Datta-Das SFET, demonstrate the importance of the quantum interferences from all patches of the Fermi surface.
The RKKY oscillation with upward trend can be understood in a simple picture. Taking the asymptotic limit $k_F r \gg 1$, the reduced spin susceptibility along the radial direction $\chi_{ab}(r)$, where $a,b = x,z$, can be well approximated as 1D Rashba system. Applying a local gauge transformation, $U(r)=e^{-i k_R r \sigma^y/2}$[@Alei; @Ima], the Rashba Hamiltonian can be mapped into the 1D free electron gas with the well-known RKKY spin susceptibility. Since the local gauge transformation is nothing but the local rotation about the $y$-axis with the spiral angle $\phi(r) = k_R r$, the reduced susceptibility is approximately the usual RKKY oscillation twisted by a local spiral transformation, $$\chi_{ab}(r) \approx \left[
\begin{array}{cc}
\cos k_R r & -\sin k_R r \\
\sin k_R r & \cos k_R r
\end{array} \right]_{ac} \chi^{RKKY}_{cb}(r),$$ where the summation over the repeated index $c=x,z$ is implied. The gauge argument explains why our numerical results resemble the RKKY oscillation but with a gradual spiral background.
![\[rashba\] (Color online) The spiral angle in the strong Rashba regime where $k_R/k_F = 2.6$ and $T=30$ K. It is clear that the spiral exchange dominates with minor ripples from the RKKY interaction.](fig4){width="6cm"}
Note that the Fermi surface composed of the concentric circles with opposite chiralities is topologically equivalent to that for the 1D electron gas with opposite spins. It is exactly because of this topological equivalence so that the local gauge transformation is possible. Therefore, we were motivated to investigate what happens when the topology of the Fermi surface changes in the strong Rashba regime where only one chirality is present. By increasing Rashba coupling to $k_R/k_F =2.6$, our numerical results, shown in Fig. \[rashba\], demonstrate robust spiral structure with minor oscillatory ripples. It is rather amazing that the change of Fermi surface topology swings the magnetic property from the RKKY dominated to the spiral. Another route to enter the strong Rashba regime is by reducing the carrier density. Since the transition occurs at $k_R/k_F =1$, for the Rashba coupling $\gamma_R=8.91 \times 10^{-12}$ eV m, one needs to reduce the density below $n_{2D} \sim 2.18 \times 10^9$ cm$^{-2}$, which may be achieved by applying gate voltage (see discussions below).
It is worth emphasizing that one should not confuse the robust spiral exchange here with the spin precession from single-particle argument. Since the chirality is even under time reversal transformation, the Kramers degeneracy connects opposite patches of the same circle. Therefore, the quantum interferences leading to RKKY oscillations are still present, as manifest in our numerical results. The puzzle is why the spiral interaction, when only one chirality is present, always takes the leading role, rendering the RKKY into minor ripples on the spiral backbone. At the point of writing, we do not have a simple physical interpretation for the interesting transition driven by the change of Fermi surface topology. But, as expected, the minor oscillatory ripples get further suppressed when the ratio $k_R/k_F$ increases.
In addition to its connection to Fermi surface topology, the mediated non-collinear coupling opens up potential applications in many magnetic devices by electrical manipulations. For instance, it was demonstrated in some experiments[@Nitta; @Engle; @Heida98] that the Rashba coupling can be controlled by the gate voltage. On phenomenology ground, the Rashba coupling reacts to the external electric field linearly, $\gamma_R = b \langle E\rangle$, where the coefficient $b$ is inversely proportional to the energy band gap and the effective mass[@Lommer]. Meanwhile, one can also use the external gate to manipulate the density of itinerant carriers. It comes into notice that the density concentration can be enhance up to 70% (corresponding to 30 % increase in Fermi energy) by electric means[@Nitta; @Heida98]. Since the carrier density responds to external electric field more sensitively than the Rashba coupling, we numerically compute the non-collinear angle $\phi_x(r)$ at different Fermi energies but keeping the Rashba coupling $\gamma_R=8.91 \times 10^{-12}$ eV m fixed. As shown in Fig. \[gate\], by changing the gate voltage, it is possible to induce sudden reversal of magnetic moments because of the RKKY oscillations. However, due to the presence of Rashba interaction, the angular jump deviates from $\pm \pi/2$ with upswing spiral background. Note that, before applying the idea of carrier-mediated non-collinear exchange coupling to realistic materials, one must keep in mind that the simple model we studied here does not include strain effects, disorders and surface roughness. In particular, the randomness of the interface can ruin our predictions and more careful ensemble average must be included.
While we mainly concentrate on the magnetic aspect of TMJ in previous paragraphs, the transport aspect is as important. In a recent paper[@Bauer03], Bauer [*et al.*]{} showed the interesting universal angular magnetoresistance and also the spin torque in ferromagnetic/normal-metal heterostructures. Besides, Myers [*et al.*]{}[@Myers] demonstrated the possibility to reverse the domain orientation by the spin torque from the injected currents. Compared with the spin torque effect in magnetic tunneling junction[@Tulapurkar05], the presence of the non-collinear exchange should deliver even richer phenomena. It is interesting to explore how the mediated non-collinear coupling and the spin torque compete and reshape our understanding in magnetic junctions.
![\[gate\]The spiral angle versus the Fermi energy for the junction of the width 90 nm at $T = 30$ K. The slightly upward tendency and the non-symmetry location of each region is caused by the Rashba spiral coupling.](fig5){width="6cm"}
In conclusion, we found the carrier-mediated non-collinear exchange coupling across the trilayer magnetic junction can not be explained by the simple spin-precession argument within single-particle picture. As the strength of the Rashba coupling increases, the mediated exchange goes from the oscillatory RKKY dominated to the robust spiral. Surprisingly, the change of magnetic behavior coincides with the transition of Fermi surface topology. As the nanotechnology advances in recent years, we believe that the clean and sharp interfaces can be realized in experiments and the effects we studied here would become important and measurable in experiments.
We acknowledges the grant supports from the National Science Council in Taiwan through NSC 94-2112-M-007-031(HHL), NSC 93-2112-M007-005 (HHL) and NSC 94-2112-M-009-025 (CHC). Financial supports for HHL through Ta-You Wu Fellow from National Center for Theoretical Sciences is also greatly appreciated.
[99]{} S. A. Wolf [*et al.*]{}, Science [**294**]{}, 1488 (2001).
I. Zutic, J. Fabian and S. Das Sarma, Re. Mod. Phys. [**76**]{}, 323 (2004).
A. H. MacDonald, P. Schiffer and N. Samarth, Nature Mat. [**4**]{}, 195 (2005).
S.-J. Sun and H.-H. Lin, Phys. Lett. A [**327**]{}, 73 (2004).
S.-J. Sun and H.-H. Lin, Eur. Phys. J. B [**49**]{}, 403 (2006).
S. Datta and B. Das, Appl. Phys. Lett. [**56**]{}, 665 (1989).
S. Murakami, N. Nagaosa and S.-C. Zhang, Science [**301**]{}, 1348 (2003).
J. Sinova [*et al.*]{}, Phys. Rev. Lett. [**92**]{}, 126603 (2004).
S.-J. Sun, S.-S. Cheng and H.-H. Lin, Appl. Phys. Lett. [**84**]{}, 2862 (2004).
J. Luo, H. Munekata, F.F. Fang and P.J. Stiles, Phys. Rev. B. [**38**]{}, R10142 (1988).
B. Das, D.C. Miller, S. Datta, R. Reifenberger, W.P. Hong, P.K. Bhattacharya, J. Singh and M. Jaffe, Phys. Rev. B. [**39**]{}, 1411 (1989).
J. Nitta, T. Akazaki, H. Takayanagi and T. Enoki, Phys. Rev. Lett. [**78**]{}, 1335 (1997).
G. Engels, J. Lange, Th. Schäpers and H. Lüth, Phys. Rev. B. [**55**]{}, R1958 (1997).
J.P. Heida, B.J. van Wees, J.J. Kuipers, T.M. Klapwijk and G. Borghs, Phys. Rev. B [**57**]{}, R11911(1998).
I.L. Aleiner and V.I. Fal’ko, Phys. Rev. Lett. [**87**]{}, 256801 (2001).
H. Imamura, P. Bruno and Y. Utsumi, Phys. Rev. B [**69**]{}, 121303(R) (2004).
G. Lommer, F. Malcher and U. Rössler, Phys. Rev. Lett. [**60**]{}, 728 (1988).
G. E. W. Bauer, Y. Tserkovnyak, D. Huertas-Hernando and A. Brataas, Phys. Rev. B [**67**]{}, 094421 (2003).
E.B. Myers [*et al.*]{}, Science [**285**]{},867(1999).
A. A. Tulapurkar [*et al.*]{}, Nature [**438**]{}, 339 (2005).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The is a system of cosmic microwave background (CMB) polarimeters, each similar to the experiment. In this paper we report results from the 2012 and 2013 observing seasons, during which the consisted of five receivers all operating in the same (150 GHz) frequency band and observing field as . We again find an excess of power over the lensed- expectation of $>5\sigma$ in the range $30<\ell<150$ and confirm that this is not due to systematics using jackknife tests and simulations based on detailed calibration measurements. In map difference and spectral difference tests these new data are shown to be consistent with . Finally, we combine the maps from the two experiments to produce final $Q$ and $U$ maps which have a depth of 57 nK$\,$deg (3.4 $\mu$K$\,$arcmin) over an effective area of 400 deg$^2$ for an equivalent survey weight of 250,000 $\mu$K$^{-2}$. The final $BB$ band powers have noise uncertainty a factor of 2.3 times better than the previous results, and a significance of detection of excess power of $>6\sigma$.'
author:
- ' and Collaborations: P. A. R. Ade'
- 'Z. Ahmed'
- 'R. W. Aikin'
- 'K. D. Alexander'
- 'D. Barkats'
- 'S. J. Benton'
- 'C. A. Bischoff'
- 'J. J. Bock'
- 'J. A. Brevik'
- 'I. Buder'
- 'E. Bullock'
- 'V. Buza'
- 'J. Connors'
- 'B. P. Crill'
- 'C. D. Dowell'
- 'C. Dvorkin'
- 'L. Duband'
- 'J. P. Filippini'
- 'S. Fliescher'
- 'S. R. Golwala'
- 'M. Halpern'
- 'M. Hasselfield'
- 'S. R. Hildebrandt'
- 'G. C. Hilton'
- 'V. V. Hristov'
- 'H. Hui'
- 'K. D. Irwin'
- 'K. S. Karkare'
- 'J. P. Kaufman'
- 'B. G. Keating'
- 'S. Kefeli'
- 'S. A. Kernasovskiy'
- 'J. M. Kovac'
- 'C. L. Kuo'
- 'E. M. Leitch'
- 'M. Lueker'
- 'P. Mason'
- 'K. G. Megerian'
- 'C. B. Netterfield'
- 'H. T. Nguyen'
- 'R. O’Brient'
- 'R. W. Ogburn IV'
- 'A. Orlando'
- 'C. Pryke'
- 'C. D. Reintsema'
- 'S. Richter'
- 'R. Schwarz'
- 'C. D. Sheehy'
- 'Z. K. Staniszewski'
- 'R. V. Sudiwala'
- 'G. P. Teply'
- 'K. L. Thompson'
- 'J. E. Tolan'
- 'A. D. Turner'
- 'A. G. Vieregg'
- 'A. C. Weber'
- 'J. Willmert'
- 'C. L. Wong'
- 'K. W. Yoon'
bibliography:
- 'ms.bib'
title: ' / V: Measurements of Polarization at Degree Angular Scales and 150 GHz by the '
---
Introduction
============
Precision polarimetry of the cosmic microwave background (CMB) has become a mainstay of observational cosmology. The model predicts a polarization of the CMB at the level of a few $\mu$K, with a characteristic pattern. The $EE$ power spectrum has been detected over a wide range of angular scales by many experiments, including [@kovac02], [@barkats04; @bischoff08], [@readhead04; @sievers07], 03 [@montroy06], [@page07; @bennett13], [@wu07], [@pryke09; @brown09], [@chiang10; @barkats14], [@quiet11; @quiet12], [@polarbear14], [@b2respap14], [@naess14], [@crites14], and [@planckI]. These measurements have been in broad agreement with theoretical expectations and other cosmological data sets. Improved $EE$ power spectrum data are important because they may eventually constrain the model parameters better than cosmic variance limited CMB temperature data [@rocha04; @galli14].
Of greater interest is the component of the polarization pattern. Though the $EE$ power spectrum is higher, the $BB$ power spectrum is more sensitive to new physics because the linear density perturbations at the surface of last scattering, which are the main source of $TT$ and $EE$ power, cannot generate power. On small angular scales, $BB$ power instead arises from the gravitational lensing of power by the large scale structure of the universe [@zaldarriaga98]. The lensing $BB$ power thus cleanly traces the growth of structure, complementary to other methods, providing information about possible extensions to such as neutrino mass or a nontrivial dark energy equation of state. Measurements by [@hanson13] [@polarbear6645; @polarbear6646; @polarbear14], and [@b2respap14] have provided the first evidence of $BB$ power from gravitational lensing.
On large angular scales, lensing contributes only a small amount of $BB$ power. However, inflationary gravitational waves (IGW) may be a source of $BB$ power on these scales [@polnarev85; @seljak97a; @kamionkowski97; @seljak97b]. The recent detection by of power on degree angular scales in excess above the lensing expectation is especially exciting because it could be evidence of primordial gravitational waves and cosmic inflation [@b2respap14]. The contribution of foregrounds to the observed signal is uncertain, and preliminary data from have suggested that polarized dust in the field may be brighter than models had predicted [@planckiXXX]. Regardless, a confirmation of the signal, whether cosmological or galactic in origin, is a top priority of observational cosmology today [@dodelson14; @caligiuri14].
The telescope is a microwave polarimeter at the South Pole designed to follow up the observations. The quickly deployed a large number of detectors at 150 GHz by installing five receivers of a design very similar to with minimal changes. All five receivers were installed in time for the 2012 observing season and continued, with upgrades, to observe at 150 GHz through the end of 2013. The modular, multi-receiver structure of the also allows individual receivers to be tuned to different frequencies. Two of the receivers began observing at 95 GHz in 2014, which will help to discriminate the signal from foregrounds.
In this paper, we present the results of the 150 GHz observations by the of the field during the 2012 and 2013 seasons. We begin with sections describing the instrument, calibrations, and analysis methods. We proceed with the maps and angular power spectra obtained from this data set and perform extensive internal consistency checks. The confirms the signal at $>5\sigma$. We then proceed to test for consistency between the and data, and finally combine the two sets of maps to a final result.
This paper is the latest in a series of publications by the and collaborations. [@b2respap14] (hereafter the Results Paper) is directly analogous to this paper. [@b2instpap14] (hereafter the Instrument Paper) presented the full details of the instrument—differences are summarized in §\[sec:instrument\] of this paper. [@b2systpap14] (hereafter the Systematics Paper) presents a detailed analysis of instrumental systematics, which are treated for the 2012–13 data in §\[sec:systematics\] of the current paper. Two additional papers, [@b2beams14] (the Beams Paper) and [@dets2014] (the Detectors Paper), describe the beam characterization and the detectors for both and the .
The instrument {#sec:instrument}
===============
The instrument shares much of its design with , details of which are presented in the Instrument Paper. In this section, we describe the main features common to both instruments and the substantive changes and upgrades unique to the . Figure \[fig:keckinstr\] shows the receiver design for the .
Cryostat and cryogenic system {#sec:cryo}
-----------------------------
The comprises five independent cryostats [@sheehy10] built by Atlas Technologies[^1]. Inside each cryostat is a closed-cycle, three-stage ($^4$He/$^3$He/$^3$He) sorption refrigerator [@duband99] that cools the focal plane unit (FPU) to approximately 270 mK. Other optical elements are held at cryogenic temperatures to minimize the thermal load on the FPUs.
The main difference between and is the bulk refrigeration system. While used a bath of liquid helium, the uses a set of Cryomech[^2] PT-410 pulse tube refrigerators. Each cryostat has its own pulse tube refrigerator aligned along the optical axis. The helium gas is pulsed at a common frequency of 1.2 Hz, and the pressure in each system is optimized to achieve the lowest base temperature. After optimization, the pulse tubes’ copper mounting surfaces typically reach 40 K and 3 K with comparable performance in all five cryostats. These surfaces are thermally connected to the telescope insert by stacks of ultra high purity aluminum foil.
Optics {#sec:optics}
------
The optics use an on-axis, refractive design which was originally demonstrated in the telescope [@takahashi10]. The entire optics chain is essentially unchanged from [@aikin10]. The two-lens design was chosen to accommodate the flat telecentric focal plane, with the image of the primary at infinity as viewed from the focal plane.
The lenses are made from high density polyethylene (HDPE) and cooled to 4 K. In order to sufficiently reduce the infrared loading on the cooling stages, there is a 3 mm nylon filter and two polytetrafluoroethylene (PTFE) filters of thickness 12.7 mm and 34.3 mm in the optics path, cooled to an intermediate stage of 50 K. A second nylon filter of 5.2 mm is placed between the objective and eyepiece lens, heat sunk to 4 K. A metal mesh low-pass filter [@ade06] with a cutoff of 8.3 cm$^{-1}$ (225 GHz) is placed above the final nylon filter to prevent any stray radiation that was not absorbed by the plastic filters from thermalizing in the detectors.
All surfaces surrounding the optical path are blackened with Eccosorb HR10[^3] cut in half and epoxied with Stycast 2850 loaded with carbon. The Stycast covers the HR10 to prevent particulate shedding during cryogenic cycling. The lining is designed such that stray light terminates on cold surfaces. Later configurations installed after the 2013 observing season included baffling on the inside of the telescope tube to further reduce reflections.
The aperture stop is similarly made from a ring of 1.9 cm thick Eccosorb AN-74, beveled at 40$\deg$ with an inner diameter of 26.4 cm. Approximately 20% of the total throughput is absorbed by the aperture stop. The aperture stop is placed on the lower surface of the objective lens.
Focal plane unit {#sec:fpu}
----------------
The detectors, developed at Jet Propulsion Laboratory (JPL) for joint use in the , SPIDER and experiments, consist of planar antennas, each an array of slot sub-antennas combined in phase, and feeding into dual series Ti and Al Transition-Edge Superconducting (TES) bolometers. A detailed description is given in the Detectors Paper. Each pixel consists of two interleaved phased antenna arrays with orthogonal polarization directions. The signal is bandpass filtered by a lithographed filter in stripline after the antenna and terminated on a thermally isolated island which also contains the series TES. A single Si tile contains 64 detector pairs and a focal plane unit has 4 tiles. In each focal plane, 8 detector pairs are left “dark”. Dark detectors consist of the complete TES island structure, but are not connected to their corresponding antennas.
The focal plane units in the were slightly modified versus those described in the Instrument Paper. The spacing between the tiles was increased in order to reduce the electromagnetic coupling between the pixels near the edge of the tile and the copper plate. The feed network of the antennas was also redesigned to reduce the dipole beam mismatch between the polarized pairs, significantly improving the performance versus .
The TES detectors are voltage biased and the current is inductively coupled to time-domain multiplexing SQUIDs [@dekorte03]. The uses the NIST developed MUX09s, which have a gradiometric design that reduces the sensitivity of the SQUIDs to uniform magnetic fields by three orders of magnitude in comparison to the MUX07a design used in [@stiehl11].
After the deployment of , we discovered that the aliased noise from the multiplexing system was affecting the overall sensitivity of the instrument. One way to mitigate the aliased noise is to increase the Nyquist inductors that limit the bandwidth of the detectors. The choice of inductance is balanced with the need for the L/R time constant to be fast enough for the detectors to be in stable negative electrothermal feedback. The first focal plane produced for the has Nyquist chips with an inductance of 1.35 $\mu$H consistent with , and all subsequent focal planes have an increased inductance of 2 $\mu$H. This limits the bandwidth to $\le18$ kHz.
Readout {#sec:readout}
-------
The configuration of the room temperature electronics that interface with is similar to that described in the Instrument Paper. A Multi-Channel Electronics (MCE) crate provided by the University of British Columbia mounts directly to the outside of each cryostat to interface with the SQUIDs, and supply the detector bias [@mce08]. The MCE crates use lower power SQUID series array readout cards compared to earlier designs in order to stay compatible with a development program to improve operability on balloon-borne telescopes.
Housekeeping thermometry is read through “backpacks” attached to the cryostats similar to , and the signals are collated and digitized in a common BLASTbus2 crate provided by the University of Toronto [@benton14]. Both the housekeeping and detector electronics are connected to a set of Linux-based computers and recorded to disk using the control software at a sample rate of 20 Hz [@story12].
Mount {#sec:mount}
-----
All five receivers are attached to a common telescope mount located at the Martin A. Pomerantz Observatory (MAPO) at the Amundsen-Scott South Pole Station. This mount was previously used for the [@leitch02] and [@hinderks09] experiments. A new front end “drum” for the cryostats was installed in 2010. The platform was leveled at that time to account for gradual shifts of the building on the snow relative to the horizon.
The mount has three axes: elevation, azimuth, and boresight. The rotation around the boresight is referred to as “deck rotation” and allows for cancellation of some systematic effects and/or tests for their presence.
Characterization {#sec:char}
================
The has been extensively characterized in laboratory tests and with [*in situ*]{} calibration measurements. The characterization program was very similar to the one described in Sections 10 and 11 of the 2 Instrument Paper. This section summarizes these measurements, focusing particularly on detector properties that have been reoptimized since the fabrication of the 2 detectors. The spectral band, optical efficiency, and bolometer thermal conductance have a strong effect on the ultimate sensitivity of the instrument, and are tuned to minimize noise while allowing stable operation under typical South Pole atmospheric loading conditions. These detector properties, summarized in Table \[tab:detparams\], are described in Sections \[sec:fts\], \[sec:oe\], and \[sec:detparam\]. We have also extensively measured the far-field beams with *in situ* observations of a mast-mounted source. The beam mapping measurement and its results are summarized here in Section \[sec:beams\] and described more fully in the Beams Paper.
[lc]{}
------------------------------------------------------------------------
Detector Parameter & Median Value\
Optical efficiency, $\eta$ & 24%\
Band center, $\langle \nu \rangle$ & 151 GHz\
Spectral bandwidth, $\Delta\nu$ &42 GHz\
Normal resistance, $R_N$ & 62 m$\Omega$\
Operating resistance, $R_\mathrm{op}$ & 0.68 $R_N$\
Saturation power, $P_\mathrm{sat}$ & 9.9 pW\
Optical loading, $P_\mathrm{opt}$ &3.1 pW\
Thermal conductance, $G_c$ & 90 pW/K\
Transition temperature, $T_{c}$ & 520 mK\
Thermal conductance exponent, $\beta$ & 2.5\
Spectral response {#sec:fts}
-----------------
The frequency response of the antennas and lumped element filters was tuned to give a fractional bandwidth of 25%. The $\sim 150$ GHz observing band is bracketed by the 118.8 GHz oxygen line on the low side and the 183.3 GHz water line on the high side.
The frequency response $S(\nu)$ of the 150 GHz detectors was characterized using a Martin-Puplett Fourier transform spectrometer (FTS) [@karkare14]. From these spectra, the band center and the bandwidth are calculated. The band center is defined to be $$\label{eqn:ftsbc}
\langle \nu\rangle = \frac{\int \nu I(\nu) S(\nu)\upd \nu}{\int I(\nu) S(\nu) \upd \nu}$$ and the bandwidth $$\label{eqn:ftsbw}
\Delta \nu = \frac{(\int I(\nu) S(\nu)\upd \nu)^2}{\int I^2(\nu) S^2(\nu)\upd \nu},$$ where $I(\nu)$ is the source spectrum relative to a Rayleigh-Jeans spectrum. For a Rayleigh-Jeans spectrum, the 150 GHz detectors were measured to have a band center of 151.5$\pm$1.9 GHz and a bandwidth of 41.8$\pm$1.4 GHz. The effective band center shifts to 150.6, 152.8 and 148.6 GHz respectively for a source spectrum of CMB, dust, and synchrotron radiation using the current best-fit models [@planckiXXX; @bennett13] The standard deviations are dominated by variation from tile to tile, with smaller variation from detector to detector within a tile.
Optical efficiency {#sec:oe}
------------------
The end-to-end optical efficiency is defined as the fraction of incident light absorbed by the detectors. This is dependent on the losses in the optics, the antennas, and the bandpass filters. A higher optical efficiency increases the sensitivity of the detectors. Because it also increases the optical loading and photon noise, it must be taken into account when optimizing the thermal conductivity of the detector.
The optical efficiency is measured in the lab using a beam-filling, microwave absorbing cone of AN-72. The power change on the detector for a source in the Rayleigh-Jeans limit ($h\nu\ll kT$) is: $$P_\mathrm{opt} = kT\eta \Delta \nu$$ where $\eta$ is the optical efficiency and $\Delta \nu$ is the bandwidth as defined in §\[sec:fts\]. The detector loading was measured with the cone at both room temperature and liquid nitrogen temperature and converted into optical efficiency with the measured bandpass of 42 GHz.
The median measured end-to-end optical efficiency of the receivers was 24%. The optical efficiency of the early detectors was lower than the 38% optical efficiency of the detectors. Detector testing after initial deployment suggested that the optical efficiency was being reduced by microscopic stress-induced cracks in the niobium microstrips connecting the antenna networks to the TES bolometers. In later generations of detectors the Nb film stress was decreased from $\sim 1000$ MPa to $<300$ MPa. The optical efficiency was observed to increase as described in more detail in the Detectors Paper.
Thermal conductance {#sec:detparam}
-------------------
As is discussed in the Instrument Paper, the detector parameters can be tuned during fabrication in order to optimize the noise performance. In particular, the thermal conductance can be tuned to minimize the phonon noise while maintaining a margin of safety ensuring operability under normal loading conditions.
The phonon noise is the thermal fluctuations from the substrate to the detector island through the SiN isolation legs. The noise-equivalent-power (NEP) is dependent on the thermal conductance $G$ across the legs (see e.g. @hiltonirwin2005) as $$\label{eqn:phonon}
\mathrm{NEP}=\sqrt{4k_BGT_c^2F}.$$ where $F$ is a numerical factor describing the non-linearity of the thermal conductance between the substrate temperature and the bath temperature (typically 0.5 for these detectors).
The saturation power of the detectors is dependent on the thermal conductance as: $$\label{eqn:satpow}
P_\mathrm{sat}=G_0 T_0 \frac{(T_c/T_0)^{\beta+1}-1}{\beta+1}$$ where the exponent $\beta$ is roughly 2.5 for these detectors. For the the loading from the optics and the sky was modeled to be $\sim$22 . The optical efficiency is used to convert the loading temperature to a power deposited on the detector. For the median optical efficiency described in §\[sec:oe\], this corresponds to $P_\mathrm{opt}=$3.1 pW of loading under normal observing conditions. Assuming a safety factor of 2, the optimal $G_c$ is then 67 pW/K.
The thermal conductance $G_c$ was measured using detector load curves with the substrate held at different temperatures. This method used “dark” detectors that were purposefully disconnected from their antennas to avoid the optical loading effects. The detectors used in had higher $G_c$, with two tiles centered at 100 pW/K and two centered at 140 pW/K. The tiles fabricated in later runs for the had lower thermal conductances, with a median $G_c$ of 90 pW/K. Several tiles had a much lower $G_c$ of 30–50 pW/K, expected to give lower phonon noise but a smaller margin of safety against saturation.
Finally, the margin of safety can be verified by measuring the electrical power $P_\mathrm{J}$ required to keep the detector in transition during standard observation. The standard observing schedule includes load curves (bolometer current-voltage measurements) taken once per hour. These have been used retrospectively to assess the safety margin under actual atmospheric conditions. With the telescope pointed at 55$\deg$ in elevation, the detectors were found to have a median margin of safety of 6.8 pW, corresponding to a safety factor of 3.2.
Beams {#sec:beams}
-----
The beam shapes were measured [*in situ*]{} at the South Pole by scanning on a large thermal noise source mounted $\sim 200$ m away, in the optical far field. All receivers in the have a beam width of 0.22 degrees, with very low levels of ellipticity. As in the dominant differential beam imperfection for the is differential pointing. The beam mapping campaign, extracted beam parameters, and residual beam features are described in detail in the Beams Paper. In this paper we use the high-fidelity per-detector beam maps as a convolution kernel for simulations to place a limit on the false $B$-mode signal from beam imperfections. The simulations and results are described in §\[sec:systematics\].
Observations and data set {#sec:obs}
=========================
Observations {#sec:obsstrat}
------------
The observation strategy of the in the 2012 and 2013 seasons was very similar to that used by , as described in the Results and Instrument Papers. The same field as (and ) was observed—a region centered at RA 0h, Dec. $-57.5\deg$. As viewed from the South Pole, the observing field remains at constant elevation and rotates in azimuth once per day. For fifty minutes periods the telescope scanned in azimuth at a fixed elevation, forming a “scanset” with 102 half-scans. Between scansets, the azimuth was updated by approximately 12.5$\deg$ to account for the sky rotation, and stepped in elevation by $0.25\deg$. Before and after each scanset, an elevation nod was performed to calibrate the relative gain of the detectors. The scan rate in azimuth was $2.8\deg/s$.
A group of ten scansets over successive azimuth ranges (and stepping in elevation) is called a “phase”. Table \[tab:phase\] shows the phases for the . The elevation ranges were switched between phases after each full cycle of schedules. Since the briefest sub-kelvin hold time among the set of five helium sorption refrigerators was $\sim$48 hours, the standard observing schedule consisted of four CMB phases and one galaxy phase between fridge cycles.
[lccll]{}
------------------------------------------------------------------------
Phase & LST Time & Field & Elevation \[deg\] & Azimuth \[deg\]\
A & Day 0 23:00 & Cryo service & &\
B & Day 1 05:30 & CMB & 55.00–57.25 & 120–300\
C & Day 1 14:30 & CMB & 57.50–59.75 & -10–170\
D & Day 1 23:00 & Galaxy & 55.00–56.50 & 130–270\
E & Day 2 05:30 & CMB & 57.50–59.75 & 120–300\
F & Day 2 14:30 & CMB & 55.00–57.25 & -10–170\
As for , the mount allows for rotation of the whole apparatus around the line of sight—referred to as “deck rotation”. This rotation was performed between each two day schedule. For and the in 2012, four deck angles were used: 68, 113, 248, and 293. These four angles provide coverage in $Q$ and $U$ and allow for cancellation of systematic effects whose sign reverses under 180$\deg$ rotation. In 2013, the started observing at eight deck angles: 23, 68, 113, 158, 203, 248, 293, and 338$\deg$. This allows for a more complete cancellation of beam systematics—see the Systematics Paper.
Data selection {#sec:cuts}
--------------
As described in the Results and Instrument Papers, data cuts are applied in three distinct stages. A few cuts remove half-scans from the scansets, while a larger number cut entire scansets from the final map. The final cut stage is the channel selection cut which is applied during the final coaddition stage. These three stages provide the necessary flexibility and granularity. The cuts are summarized in Table \[tab:cuts\] for the 2012–2013 data set (Cf. Table 7 of the Instrument Paper).
[lccc]{}
------------------------------------------------------------------------
Cut & Total time \[10$^6$ s\] & Integration \[$10^9$ det$\cdot$s\] & Fraction cut \[%\]\
Before cuts & 18.3 & 30.4 &\
Channel cuts & 18.3 & 27.5 & 9.5\
Synchronization & 18.1 & 27.2 & 1.2\
Deglitching & 17.8 & 23.6 & 11.8\
Passing channels 1 & 17.6 & 23.3 & 0.58\
Elnod calibration & 17.1 & 19.9 & 11.2\
TES fractional resistance & 17.1 & 19.6 & 0.88\
Time stream skewness & 17.1 & 17.8 & 6.2\
Time stream variance & 17.0 & 17.4 & 1.2\
Noise stationarity & 16.9 & 17.0 & 1.4\
FPU Temperature & 16.7 & 16.7 & 0.84\
Telescope pointing & 15.7 & 15.8 & 3.0\
Passing data & 15.4 & 15.6 & 0.55\
2012–2013 data set and sensitivity {#sec:sens}
----------------------------------
The telescopes continuously took data through the South Pole winter. Each of 2012 and 2013 produced nearly 4500 scansets of data for a total of $18\times10^6$ s of data. After the data selection cuts, an average of 50% of the data remain, for a total of 15.6$\times 10^9$ dets$\cdot$s. This is shown graphically in the top panel of Figure \[fig:livetime\] (Cf. Figure 23 of the Instrument Paper).
The instantaneous sensitivity of the instruments is measured using two methods: by taking the average of the time stream noise spectra between 0.1–1 Hz and by measuring the standard deviation of the noise-only maps weighted by the square root of integration time [@kernasovskiy12]. Both methods yielded a noise-equivalent-temperature (NET) of 11.5 for 2012 and 9.5 for 2013. The middle panel of Figure \[fig:livetime\] shows the instantaneous sensitivity calculated with the time stream based method for the 2012–2013 seasons. Using the same method as described in the Results and Instrument Papers the map depth for the 2012–2013 150 GHz data is 74 nK in nominal square-degree pixels (4.4 $\mu$K$\,$arcmin) over an effective area of 390 square degrees for a total sensitivity of 2.6 nK. An equivalent way of expressing the sensitivity of the data set is the survey weight $W=1/s^2=$150,000 $\mu$K$^{-2}$, where $s$ is the total sensitivity. This expression is useful because it scales linearly with integration time, number of detectors, and statistical sensitivity to $r$.
Low level data reduction, map making and simulations
====================================================
Analysis pipeline {#sec:pipeline}
-----------------
The and data analysis uses the same code and proceeds in parallel, enabling cross checks between these independent data sets. The process used here is exactly the same as described in the Results Paper—a summary follows.
Low level reduction {#sec:lowreduc}
-------------------
As detailed in the Results Paper, the first step in the low level reduction is to deconvolve the temporal response of the instrument. The TES detectors have a fast response of $\sim$1 ms that can be ignored. Both the MCE and the apply low pass filters to the data which must be accounted for in the deconvolution.
At this stage, relative calibration is accomplished by dividing the time streams by the individual detector gains derived from elevation nods. The data are then multiplied by the median gain across the array in order to remove dependence on atmospheric variation.
Pairmaps
--------
The sum and difference of each detector pair is taken. Each half-scan is subjected to third order polynomial filtering to remove atmospheric variation. In addition, the mean of the half-scans over the scanset is subtracted to remove any scan-synchronous contamination. The pointing of each detector pair is reconstructed from the telescope pointing model and per-pair offset angles refined by regressing per-channel maps against a 5 template. At this point the time stream data of each pair is binned into a rectangular grid of pixels forming per scanset “pairmaps”. We also sample and bin the 143 GHz temperature map and its derivatives to use in the deprojection of beam systematics. For further details see the Results Paper and the Systematics Paper.
Full maps
---------
Finally, the pairmaps are coadded into final full maps, and also into various pairs of jackknife splits. The deprojection templates are fit and removed during this process. Absolute calibration of the maps is performed by comparing the power spectrum of the temperature map with the 143 GHz map as described in the Instrument Paper.
Figure \[fig:tqu\_maps\] shows the resulting temperature and Stokes $Q$ and $U$ polarization maps for the 150 GHz data from the 2012–2013 observing seasons. The left column shows the final maps and the right hand column shows a difference (jackknife) map which is consistent with noise alone. The vertical/horizontal stripes in the $Q$ maps and the diagonal stripes in the $U$ maps are characteristic of the polarization signal, which dominates the maps.
Simulations {#sec:sims}
-----------
We create signal and noise simulations exactly as described in the Results Paper. We generate realizations of noise by randomly flipping the signs of the pairmaps when co-adding to full maps. Several kinds of signal simulations are made by resampling input maps from the program (part of the [^4] package [@healpix]). The simulated data are then binned into pairmaps and combined to full maps exactly in parallel with the treatment of the real data.
Results
=======
Power spectra {#sec:powspec}
-------------
The maps are converted into angular power spectra exactly as described in the Results Paper. The matrix based purification of the $Q$ and $U$ maps is performed prior to inversion to form s to avoid $E$ to $B$ mixing due to the sky-cut and filtering. A variant of the procedure [@hivon02] is used to noise debias the auto spectra and correct for the effects of the time stream filtering.
The resulting power spectra for the using the 150 GHz data from the 2012–2013 observing seasons is shown in Figure \[fig:powspecres\], along with a temporal jackknife. The power spectrum is inconsistent with cosmology without foregrounds at $5.0\sigma$ (over the first 5 band powers). Although the noise is lower than for , the first two band power values are also lower so the significance is somewhat smaller.
An overall rotation is applied to the maps to minimize the high-$\ell$ $TB$ and $EB$ spectra [@kaufman14]. For the 2012+2013 data this adjustment is $\approx -0.5\deg$. This rotation makes no practical difference to the power spectrum.
$E$ and $B$ maps
----------------
and maps can be made by performing an inverse Fourier transform as shown in Figure \[fig:ebmaps\]. The maps created are inherently apodized, as the and components are generated from apodized Q and U maps. These are compared to a lensed-+noise simulation.
Internal consistency tests {#sec:jacks}
--------------------------
The data was split in 16 different ways to test for internal consistency. The motivations behind these splits are described in the Results Paper and the Systematics Paper. If a contaminating signal exists in only one half of the data split, then it should show up with as much significance in the jackknife as in the signal map. However, some jackknives are more sensitive to certain systematics than the signal map because of inherent cancellation effects which operate in the full map. Each of the jackknife categories is summarized below.
The first set of jackknives probes for systematics which differ between different subsets of channels. This includes division in the multiplexing system, as well as divisions in the focal plane layout: tile, focal plane inner/outer, tile inner/outer, mux row, mux column. As is documented in the Beams Paper, there are systematics that are highly dependent on the position of the detector in the focal plane. For instance, the ellipticity of the beams is greater in the detectors near the outside of the focal plane than the inside.
The next set of jackknives is temporal. This includes both the longest time scale of 2012 data versus 2013 data, and the shortest timescale of left going scans versus right going scans. Owing to the changes between the 2012 and 2013 observing seasons, an early/late season jackknife acts as an alternative temporal split. The first is sensitive to the effects of different observing schedules and detectors changed between seasons, while the second is sensitive to detector time constants.
Another set of jackknives is based on external contamination. This includes the azimuth jack, which divides the data based on the direction the telescope is pointed with respect to the ground (see Table \[tab:phase\]). One half of this jackknife is data taken in the direction of the main South Pole Station and associated operations, while the other half points into the desolate Antarctic plateau.
A set of jackknives that particularly amplifies the differential beam properties is the deck rotation jackknives. As is described in the Systematics Paper, a 180$^\circ$ deck rotation cancels out differential pointing. The deck jackknife, which differences the 180$^\circ$ rotations, amplifies the leakage by an order of magnitude in comparison to that present in the fully coadded data. The also started taking data at 90$^\circ$ compliment deck rotations in the 2013 observing season, and this jackknife is sensitive to differential gain or differential beam width leakage. The alternative deck jackknife is defined to be the difference of the 90$^\circ$ rotations for 2013. In this special case, the statistics for the 2012 and 2013 data are separate.
Maps are made from each half of the data split and then differenced. The differenced maps are divided by a factor of two in order to keep the noise amplitudes equivalent to the signal map. The consistency with lensed-+noise simulations is calculated with a simple $\chi^2$ statistic: $$\chi^2 = \left( \mathbf{d} - \langle\mathbf{m}\rangle\right)^\mathrm{T} \mathbf{D}^{-1}
\left(\mathbf{d} - \langle\mathbf{m}\rangle\right)
\label{eqn:chisq}$$ where $\mathbf{d}$ is the vector of observed band power values, $\langle\mathbf{m}\rangle$ is the mean of the lensed-+noise simulations (except where alternative signal models are considered), and $\mathbf{D}$ is the band power covariance matrix as evaluated from those simulations.
A $\chi$ statistic is also considered to probe for sets of band powers which are systematically above or below the expectation. This is defined as: $$\chi = \sum_i \frac{d_i - \langle m_i \rangle}{\sigma_{m_i}}$$ where the $d_i$ are the observed band power values and $\langle m_i \rangle$ and $\sigma_{m_i}$ are the mean and standard deviation of the lensed-+noise simulations.
For each of these statistics, we calculate the probability to exceed (PTE) the observed value by comparing to the values obtained in the 500 lensed-+noise simulations. Since the distribution of the band powers of the auto spectra is approximately $\chi^2$ distribution, there is some non-Gaussianity to the statistics. In particular, the lowest band power only has 9 effective modes which will increase the tails of the distribution. However, by calculating the PTE against the simulations, any non-Gaussianity is fully reflected in the PTE value. The PTE for the $\chi$ and $\chi^2$ using band powers 1–5 and 1–9 is given in Table \[tab:ptes\]. Note that these statistics are correlated (especially along each row of the table). The distribution of the PTE values is shown in Figure \[fig:ptedist\].
[l c c c c ]{} Jackknife & Band powers & Band powers & Band powers & Band powers\
& 1–5 $\chi^2$ & 1–9 $\chi^2$ & 1–5 $\chi$ & 1–9 $\chi$\
Systematics {#sec:systematics}
===========
Experimental systematics can create false polarization and must be shown to be tightly controlled. The systematics in were fully explored in the Systematics Paper and were shown to be below the level equivalent to $r$=0.003–0.006. These limits were derived from forward simulations of the measured instrumental properties. If a given property did not have a measured level, appropriate upper limits were used.
The beam systematics in the are expected to be below those of because of the larger number of detectors (increased averaging down of incoherent effects), and the increased number of receiver orientations (both instantaneously due to the “clocking” of the five receivers at $72^\circ$ increments around the boresight, and, in 2013, the increased number of deck angles of observation).
As described in the Results and Systematics Papers we produce simulated time streams by convolving an input temperature map with high precision per channel measurements of the actual beam shapes. We then pass these simulated time streams through the mapping process, including all filtering and deprojection, to assess the residual contamination due to beam non-ideality. The results are shown in Figure \[fig:beamsim\]. The beam maps for the do not provide as uniform and redundant coverage of all detectors as those for and additional analysis is required to construct composite beam maps that have consistently high signal-to-noise and are free of artifacts from the beam mapping measurement. For the purposes of the current paper we use the preliminary beam map results to set an upper limit on the residual contamination, as indicated by the down arrows in the figure.
Other forms of systematic contamination were considered, such as electromagnetic interference (EMI) contamination, magnetic pickup, thermal pickup, and detector pointing. Each of these systematics was quantified to be below the level when averaged over the entire array, and thus safely ignorable.
Consistency with {#sec:b2consis}
=================
The resulting power spectra of and the can be compared to assess the compatibility of the two sets of results. Although there is much that the two experiments share in terms of hardware, design, and location, there are also potential systematics that are different: the bulk refrigeration system, the ground shield, and the time at which the observations occurred. Comparing the results is a powerful additional systematics check. The power spectra for both the , and the cross between the two are shown in the upper panel of Figure \[fig:specjack\].
A rigorous comparison can be done in two ways: directly comparing the maps and comparing the auto and cross power spectra. The latter can be a more powerful comparison if the maps have different noise levels—since and the have comparable noise levels, all four methods (map and the three combinations of auto and cross spectra) have approximately equal statistical power.
To test the compatibility of the resulting band powers with null we compare them to the differences of signal+noise simulations which share common input skies. In such tests it is necessary that the simulations contain power roughly equal to the real sky as the cross terms between signal and noise increase the fluctuation of the differences even for perfectly common sky coverage. For example, the (un-debiased) auto spectrum of a map $M$ composed of a signal $S$ and noise $N$ can be written $$M \!\! \times \!\! M = (S+N) \! \times \! (S+N)=S \!\! \times \!\! S + 2(S \!\! \times \!\! N) +N \!\! \times \!\! N.$$ The difference of such auto spectra between experiments with common sky coverage is then $$M_1 \!\! \times \!\! M_1 - M_2 \!\! \times \!\! M_2 =
2(S \!\! \times \!\! N_1 - S \!\! \times \!\! N_2)
+ N_1 \!\! \times \!\! N_1
- N_2 \!\! \times \!\! N_2$$ where $M_1$ and $M_2$ refer to the first and second experiment. The signal auto spectrum $S \!\! \times \!\! S$ cancels out. However, the cross terms between the signal and noise $2(S \!\! \times \!\! N_1 - S \!\! \times \!\! N_2)$ do not cancel, and they increase the fluctuations between the two experiments over the noise-only case in proportion to the common signal. To account for this extra variance, we use signal simulations with additional power that matches the amplitude of the observed signal in excess of in band powers 1–5. (The origin of the extra signal over is not important here—only its approximate amplitude.) The results are shown in Figure \[fig:specjack\].
We then proceed to calculate the PTE of the $\chi$ and $\chi^2$ statistics versus the simulated distributions using the same spectra and band power ranges as in §\[sec:jacks\], and give the results in Table \[tab:specptes\]. In both the figure and the table we note the effect of the two band powers at $\ell \approx 220$ that are high with respect to lensed- in B2xB2 (as noted in the Results Paper) but not in KxK and B2xK—as expected these also show up in the map difference. Again note that the PTE values are correlated (both along and between rows of the table) so overinterpretation should be avoided. Our conclusion is that the and data are consistent—especially in the lowest five band powers where an IGW contribution would be strongest.
[l c c c c]{} Jackknife & Band powers & Band powers & Band powers & Band powers\
& 1–5 $\chi^2$ & 1–9 $\chi^2$ & 1–5 $\chi$ & 1–9 $\chi$\
Combination with {#sec:comb}
=================
Having shown that the results are consistent with we now proceed to combine the maps by adding the accumulation quantities (equivalent to a noise weighted combination of the maps). This results in $Q$ and $U$ maps which have a depth of 57 nK$\,$deg (3.4 $\mu$K$\,$arcmin) over an effective area of 400 square degrees. Following §\[sec:sens\], the map depth and effective area are combined for a total sensitivity of 2.0 nK and a total survey weight of 250,000 $\mu$K$^{-2}$.
The observation regions and strategies are sufficiently similar that it is found empirically using simulations that the purification matrix of either experiment delivers adequate purity when applied to the combined map (with contamination equivalent to $r<10^{-3}$).
The final $BB$ spectrum is shown in Figure \[fig:b2kcomb\] and is inconsistent with the lensed- expectation at $>6\sigma$ (for either band powers 1–5 or 1–9). The lensed-+noise error bars as plotted are approximately a factor two smaller than those of the previous only results—saturation on the (small) sample variance of the lensing component is occurring—the noise component is a factor 2.3 times smaller. All the spectra (including $TT$, $EE$ etc.) are available for download at <http://bicepkeck.org/> together with the ancillary data, noise information etc., required to use them.
Conclusions {#sec:conc}
===========
We have presented the instrument and the 2012–2013 (150 GHz) data set. The instantaneous instrumental sensitivity of 9.5 is the best reported to date. The same area of sky as previously observed by was mapped to a depth in $Q$ and $U$ of 74 nK$\,$deg (4.4 $\mu$K$\,$arcmin). The resulting power spectra are consistent with lensed-except for an excess at degree angular scales in $BB$ which has a significance of $5.0\sigma$. Extensive jackknife tests argue against a systematic origin for the signal, and further statistical tests indicate that the maps and spectra are consistent with the previous results. Finally the two sets of maps are combined to produce maps with noise of 57 nK$\,$deg (3.4 $\mu$K$\,$arcmin) over an effective area of 400 deg$^2$ for a survey weight of 250,000 $\mu$K$^{-2}$. The final $BB$ spectrum is inconsistent with lensed- at a significance of $>6\sigma$. The combined map results (for all spectra) are available for download. There does not appear to be any reason to consider the results as more reliable than the results or vice versa. We therefore emphasize that we regard the combined results as the best available data set at this time.
The origin of the excess power shown in Figure \[fig:b2kcomb\], and previously reported in the Results Paper, has been extensively debated in the literature [@mortonson14; @flauger14; @fuskeland14]. Recently, concrete information on the strength of polarized dust emission at high galactic latitude has become available in @planckiXXX. It appears that dust emission is a significant contribution to the signal observed by and the . Therefore, in an upcoming paper, the and combined maps are cross correlated with maps of the same region to constrain the dust contribution to the observed signal.
During the 2014 season two of the receivers operated at 95 GHz and a future analysis will use this data to further constrain the dust contribution. In the 2015 season 3 will provide increased sensitivity at 95 GHz and operation of receivers at 220 GHz is also planned.
The project has been made possible through support from the National Science Foundation under Grants ANT-1145172 (Harvard), ANT-1145143 (Minnesota) & ANT-1145248 (Stanford), and from the Keck Foundation (Caltech). The development of antenna-coupled detector technology was supported by the JPL Research and Technology Development Fund and Grants No.06-ARPA206-0040 and 10-SAT10-0017 from the NASA APRA and SAT programs. The development and testing of focal planes were supported by the Gordon and Betty Moore Foundation at Caltech. Readout electronics were supported by a Canada Foundation for Innovation grant to UBC. The computations in this paper were run on the Odyssey cluster supported by the FAS Science Division Research Computing Group at Harvard University. The analysis effort at Stanford and SLAC is partially supported by the U.S. Department of Energy Office of Science. We thank the staff of the U.S. Antarctic Program and in particular the South Pole Station without whose help this research would not have been possible. Most special thanks go to our heroic winter-overs Robert Schwarz and Steffen Richter. We thank all those who have contributed past efforts to the –series of experiments, including the team.
[^1]: <http://www.atlasuhv.com/>
[^2]: <http://www.cryomech.com/>
[^3]: Emerson & Cuming Microwave Products, Randolph, MA 02368. Phone: 781-961-9600. Web: <http://eccosorb.com/>
[^4]: <http://healpix.sourceforge.net/>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In a recent paper “Head-on collision of electron acoustic solitary waves in a plasma with nonextensive hot electrons" \[Astrophys. Space Sci.**338**, 271–278 (2012)\] Eslami, Mottaghizadeh and Pakzad deal with the problem of the head-on collisions between two weakly nonlinear electron-acoustic solitary waves. Unfortunately, their treatment is deficient and leads to erroneous conclusions.'
author:
- Frank Verheest
title: 'Comment on “Head-on collision of electron acoustic solitary waves in a plasma with nonextensive hot electrons"'
---
In a recent paper, @Headon2012 deal with the problem of the head-on collisions between two weakly nonlinear electron-acoustic solitary waves in a two-electron plasma with hot nonextensive and cold components, in the presence of a neutralizing ion background. Unfortunately, their treatment is deficient and leads to erroneous conclusions.
For clarity in what follows, equations from the paper by @Headon2012 will be denoted as (EMP.1) and higher, and equations in this Comment as (1) and higher. Let us start the discussion from the Poisson equation (EMP.3), which is repeated here for ease of exposition: - n\_c - \[1+(q-1)\]\^ + ( 1 + ) = 0. The cold electron density $n_c$ has to be determined from the relevant fluid equations (EMP.1) and (EMP.2). Using their expansions (EMP.6) to lowest orders gives from (EMP.3) or that n\^1 &=& - \^1, \[pois1\]\
n\^2 &=& - \^2 + (\^1)\^2. \[pois2\] This part is straightforward, since terms with derivatives only occur to third or higher order. The order of the expansion in $\eps$ is denoted by superscripts, in their notation, and their expansions (EMP.6) start with terms in $\eps$ outside equilibrium.
In (EMP.10) a separable form is proposed for $\phi^1$, \^1 = \_1\^1(,) + \_2\^1(,), which gives (EMP.11) or n\^1 = - , and thus fulfils . However, @Headon2012 claim that $n^2$ and $\phi^2$ have separability properties analogous to the first order ones, cfr.(EMP.13) and (EMP.14), hence \^2 &=& \_1\^2(,) + \_2\^2(,),\
n\^2 &=& - . It is immediately clear that this cannot hold, since there would remain from that \[\_1\^1(,) + \_2\^1(,)\]\^2 = 0, and except for $q=3$ one would have to conclude that $\phi_1^1(\xi,\tau)=0$ and $\phi_2^1(\eta,\tau)=0$, wiping out all first order terms.
The problem is worse, because the equations of continuity (EMP.1) and of motion (EMP.2) also contain nonlinear contributions to second order. One can try to eliminate e.g. $u^2$ to arrive at another relation between $n^2$ and $\phi^2$, but disentangling the information is bedevilled by the fact that derivatives with respect to $\xi$ and $\eta$ occur together, preventing an immediate integration.
A way forward would be to propose a general decomposition \^2 = \_1\^2(,) + \_2\^2(,) + \_3\^2(,,). This includes a mixed term $\phi_3^2(\xi,\eta,\tau)$, which cannot be separated into parts not depending either on $\eta$ or on $\xi$, as these would be in $\phi_1^2(\xi,\tau)$ or $\phi_2^2(\eta,\tau)$, respectively. Using a similar decomposition for $n^2$ (and also for $u^2$, but let us concentrate on the densities and the electrostatic potential) allows to determine first from (EMP.1) and (EMP.2) that n\_1\^2 &=& - \_1\^2(,) + \^2, && \[nn21\]\
n\_2\^2 &=& - \_2\^2(,) + \^2. && \[nn22\] When this is substituted into the terms in $n_1^2(\xi,\tau)$, $n_2^2(\eta,\tau)$, $\phi_1^2(\xi,\tau)$ and $\phi_2^2(\eta,\tau)$ cancel, because of the linear dispersion properties, so that for the terms only in $(\xi,\tau)$ or $(\eta,\tau)$ there remains that \^2 = 0, \[n21\]\
\^2 = 0, \[n22\] after having divided out common nonzero factors. In addition, when one combines the mixed contributions and eliminates $n_3^2(\xi,\eta,\tau)$ and $u_3^2(\xi,\eta,\tau)$, there is a differential equation for $\phi_3^2(\xi,\eta,\tau)$ to fulfil, \_3\^2(,,) &=& && \_1\^1(,) \_2\^1(,). Now the choice is clear.
A first and generic possibility is that $q$ does not annul the coefficients in and , but then all first order variables vanish and indicates that $\phi_3^2(\xi,\eta,\tau)=0$. Hence, the second order is indeed separable as claimed by @Headon2012, but there is no first order left, $\phi_1^1(\xi,\tau)=0$ and $\phi_2^1(\eta,\tau)=0$, and from here on the remainder of the paper is automatically null and void.
The other choice is that $q$ is special, $q=(3(1+\alpha)/(1-3\alpha)$, so that the first order variables remain in the loop, but for the second order quantities besides and , one has to find a solution for $\phi_3^2(\xi,\eta,\tau)$, which is far from trivial but certainly nonzero, as becomes \_3\^2(,,) &=& ( + - ) && \_1\^1(,) \_2\^1(,). About this part of the discussion the authors are completely silent, and now the second order variables certainly are not given by (EMP.13)–(EMP.15), so that also here the remainder of the paper presents no valid information.
Analogous criticisms invalidate the results in an earlier paper by the same authors [@Headon2011], dealing with ion rather than electron acoustic modes, with obvious notational differences but having a similar structure.
Some of the papers in the literature start the expansion with terms in $\eps^2$ (outside equilibrium), thereby *implicitly assuming* (apparently without checking!) that the model is simple enough so that the coefficients corresponding to those in and never vanish. Then the terms in $\eps^3$ of the expansion do not contribute, and to the next order the relevant KdV equations and phase shifts are obtained.
This is certainly the case when a simple plasma model is considered with cold ions and Boltzmann electrons, without additional species [@Hilmi], but not immediately for many other models treated in the literature.
However, the plasma model investigated by @Headon2012 is rich enough to admit critical values for the parameters, and so they were right to start their expansions with terms in $\eps$, but did not work that out as it should have been, with the unfortunate consequence that their paper is incorrect and incomplete, either way.
There are other, but far less important, blemishes in the paper by @Headon2012. One is that in (EMP.8) a factor $-\,\alpha$ is missing in front of the last two terms, as can immediately be seen by referring to the original equation of motion (EMP.2).
Another is that $\lambda$ is used in two different meanings, once in (EMP.12) and (EMP.15) where it really should be the velocity $c$ mentioned in the stretching (EMP.5), whereas the other $\lambda$, defined towards the bottom of the left hand column on page 2, is essentially $1/c^2$. Furthermore, there is an evident typo in the nonlinear term in (EMP.18).
It is interesting to remark that the value of $q$ which annuls the coefficient in (12) and (13) also annuls the coefficient $A$ of the nonlinear term in the KdV equations (EMP.17) and (EMP.18). Given the way the nonlinearities work, this should not come as a surprise.
To conclude, the paper by @Headon2012 is marred by an erroneous algebra (for generic $q+1>0$) or by a deficient discussion (when $q$ takes on a critical value), leaving the paper without validity.
Demiray, H.: Interactions of nonlinear ion-acoustic waves in a collisionless plasma. J. Comput. Appl. Math. **206**, 826–831 (2007).
Eslami, P., Mottaghizadeh, M., Pakzad, H. R.: Head-on collision of ion-acoustic solitary waves in a plasma with a $q$-nonextensive electron velocity distribution. Physica Scripta **84**, 015504 (2011).
Eslami, P., Mottaghizadeh, M., Pakzad, H. R.: Head-on collision of electron acoustic solitary waves in a plasma with nonextensive hot electrons. Astrophys. Space Sci. **338**, 271–278 (2012).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We look for a Brans-Dicke type of generalization of Horava-Lifshitz gravity. It is shown that such a generalization is possible within the detailed balance condition. The resulting theory reduces in the IR limit to the usual Brans-Dicke theory with a negative cosmological constant for certain values of parameters. We then consider homogeneous and isotropic cosmological situation in the context of this generalized theory, and find some interesting features of the Brans-Dicke scalar field in determining the behavior of the universe.'
author:
- Joohan Lee
- Tae Hoon Lee
- Phillial Oh
title: 'Coupling of Brans-Dicke scalar field with Horava-Lifshitz Gravity'
---
Recently, a new theory of gravity has been proposed by Horava[@ph1; @ph2; @ph3]. This theory, being based on anisotropic scaling of space and time, breaks the spacetime symmetry. It has a much better UV behavior than the theories with spacetime diffeomorphism symmetry such as general relativity, but reduces to Einstein’s gravity in the infrared limit, thereby recovering spacetime diffeomorphism symmetry. Physical constants such as the speed of light, Newton’s constant, and cosmological constants all emerge from the relevant deformation of the non-relativistic theory at short distance. These interesting features as well as related findings have received a great deal of attention[@HLG].
On the other hand, even if we consider only the low energy limit of the gravity there are many alternative theories and extensions of the Einstein theory. In particular, in the context of cosmology various models with a scalar field have been considered and a possible role of the scalar field in explaining the behavior of the universe in the early inflationary stage as well as the late stage has been investigated[@QE; @KE]. Therefore, it would be interesting to see if Horava’s theory can be extended in such a way that in the infrared limit it reduces to those alternative theories. In this regard, of particular interest for us is the one with a non-minimally coupled scalar field because minimally coupled scalar source had already been investigated[@ph1; @MCS]. Typical examples would be the Brans-Dicke theory[@BD] and the the gravity with a dilaton field[@DG] arising for instance in the string theory.
In this paper, we extend the Horava-Lifshitz gravity to include the Brans-Dicke field as a concrete example of the non-minimally coupled scalar field. It turns out that such an extension is possible within the context of the detailed balance condition and in the IR limit reduces to the four-dimensional Brans-Dicke theory with negative cosmological constant when parameters of the theory satisfy certain conditions.
We then study cosmological implication of the theory assuming homogeneity and isotropy. Without the negative cosmological constant and the dark radiation term[@gc] the equations are those of the Brans-Dicke theory. So, we concentrate only on their effects on cosmology. Still, we find several interesting features. In the early universe limit there exists a solution where the scale factor $a(t)$ grows like $t^{1/2}$, which corresponds to the behavior of the universe in the presence of the normal radiation. Furthermore, in the large universe limit we find a solution which increases exponentially in spite of the existence of a negative cosmological term. This is contrary to the usual expectation that the exponential solution is possible only for a positive cosmological constant[@uk; @ac]. Both these aspects are possible because of the Brans-Dicke field.
Let us consider the four-dimensional Brans-Dicke theory[@BD], where the action is given by S=d\^4x(R-\^[-1]{}g\^\_\_). Decomposition of this action into $(3+1)$ form, including the speed of light, $c$, yields[^1]R && N([R]{}+c\^[-2]{}(K\_[ab]{}K\^[ab]{}-K\^2)) -2Nc\^[-2]{}K-2ND\^2,\
-\^[-1]{} g\^\_\_&=&N\^[-1]{}c\^[-2]{}\^2 -N\^[-1]{}D\^aD\_a, where the four metric $g$ is decomposed into the lapse function $N$, the shift vector $N^a$ and the three metric $q_{ab}$, and the corresponding three-dimensional covariant derivative and its scalar curvature are denoted respectively by $D_a$, ${\cal R}$. The Brans-Dicke parameter is assumed positive, $\omega>0$. In the first equation an irrelevant total divergence term was dropped. The time derivatives of the three-metric and the scalar field are encoded in the following quantities; K\_[ab]{}&&[12N]{}([g]{}\_[ab]{}-D\_aN\_b-D\_bN\_a),\
&&[1N]{}(-N\^a\_a).Using the above result the Brans-Dicke action can be split into the two parts $S_{BD}=S_{BD}^K+S_{BD}^V$, where the kinetic and potential parts are S\_[BD]{}\^K&=&c\^[-1]{}dtd\^3x N( (K\_[ab]{}K\^[ab]{}-K\^2)-2K+\^[-1]{}\^2),\
S\_[BD]{}\^V&=&cdtd\^3x N(-2D\^2-\^[-1]{}D\^aD\_a).Re-scaling the scalar field $\phi$ and the corresponding field $\pi$, we find S\_[BD]{}\^K&=&dtd\^3x N( (K\_[ab]{}K\^[ab]{}-K\^2)-2K+\^[-1]{}\^2),\
S\_[BD]{}\^V&=&c\^2dtd\^3x N(-2D\^2-\^[-1]{}D\^aD\_a).Note that the factor of $c^2$ appears in front of the potential term. For the later purpose regarding the detailed balance it is important to express the kinetic part in the following matrix form; S\_[BD]{}\^K=dtd\^3x N (
[cc]{} K\_[ab]{} &\
) (
[cc]{} G\^[abcd]{} & -q\^[ab]{}\
-q\^[cd]{} & \^[-1]{}\
) (
[c]{} K\_[cd]{}\
\
), where G\^[abcd]{}= (q\^[ac]{}q\^[bd]{}+q\^[ad]{}q\^[bc]{})-q\^[ab]{}q\^[cd]{}.\[sm\] Note that the matrix in the middle of the kinetic part of the action can be regarded as the supermetric on the space of $(q_{ab},\phi)$, naturally extending the DeWitt metric on the space of three-metrics.
We intend to construct a Brans-Dicke type extension of Horava-Lifshitz gravity with the detailed balance condition. So, we choose the action of the form, $S_{HLBD}=S_{HLBD}^K+S_{HLBD}^V$, where the kinetic part is S\_[HLBD]{}\^K=dt d\^3x N (
[cc]{} K\_[ab]{} &\
) (
[cc]{} G\^[abcd]{}() & -q\^[ab]{}\
-q\^[cd]{} & \^[-1]{}\
) (
[c]{} K\_[cd]{}\
\
)and the potential part is of the form S\_[HLBD]{}\^V=-dt d\^3x N (
[cc]{} &\
)
(
[cc]{} G\^[abcd]{}() & -q\^[ab]{}\
-q\^[cd]{} & \^[-1]{}\
)
\^[-1]{}(
[c]{}\
\
)for some suitable choice of function $W(q,\phi)$. The supermetric $G^{abcd}(\lambda)$ was slightly deformed compared to the Eq. (\[sm\]) to include the parameter $\lambda$ as usual, G\^[abcd]{}()(q\^[ac]{}q\^[bd]{}+q\^[ad]{}q\^[bc]{})-q\^[ab]{}q\^[cd]{}.Such a choice of the action is a natural generalization of the Horava-Lifshitz gravity in the context of the detailed balance condition. The factor of two was inserted in front of the variation of $W$ with respect to $\phi$ to compensate for different normalization in time derivatives in Eqs.(4) and (5). It is a straightforward matter to calculate the inverse supermetric. It comes out to be of form (
[cc]{} \^[-1]{} [G]{}\_[abcd]{} & -Aq\_[ab]{}\
-Aq\_[cd]{} & B\
),where \_[abcd]{}= (q\_[ac]{}q\_[bd]{}+q\_[ad]{}q\_[bc]{})-|q\_[ab]{}q\_[cd]{}, with A&=&,\
B&=&,\
|&=&.Note that this inverse supermetric is well-defined even for $\lambda=1/3$ contrary to the pure gravity case and becomes singular instead when $\lambda=1$ and $\omega=-3/2$ ($\omega>0$ is assumed in this work and it is nonsingular if $\lambda>1/3$.). The singualr case corresponds to the conformal scalar. If we take the limit of $\omega\rightarrow\infty$, $A$ and $B$ vanish and $\lambda=1/(3\lambda-1)$, reproducing the pure gravity case.
We choose W=c\_1d\^3x ([R]{}-2\_b)-c\_2d\^3x\^[-1]{}D\^aD\_a.In general all possible marginal and relevant terms can be included. The above choice of $W$ corresponds to keeping only terms important in the infrared limit. Then, from &=&-c\_1\_bq\^[ab]{}+Q\^[ab]{},\
&=& -c\_1\_b +Q,where Q\^[ab]{}&& c\_1(-G\^[ab]{}+D\^aD\^b-q\^[ab]{}D\^2),\
Q&& c\_1-c\_2(-\^[-1]{}D\^2+\^[-2]{}D\^aD\_a),with $G_{ab}$ being the Einstein tensor constructed with the three-dimensional metric, we find after a straightforward calculation that S\_[BDHL]{}\^V&=&dtd\^3xN{+(-\^[-1]{}D\^aD\_a) +(-2D\^2)}\
&-&dtd\^3xN(Q\^[ab]{}\^[-1]{}[G]{}\_[abcd]{}Q\^[cd]{}-2AQ\^[ab]{}q\_[ab]{}Q+BQ\^2),\[quad\]where &=&(c\_1\_b)\^2\
&=&-(c\_1)\^2\_b\
&=&-(c\_1)\^2\_b.The second line of Eq. (\[quad\]) collects all the quadratic terms.
When $c_1=c_2$ and $\lambda=1$, the theory recovers four-dimensional diffeomorphism symmetry, as one can see from the fact that in the infrared limit the potential part of the action becomes S\_[BDHL]{}\^V\_[IR]{}= -(c\_1)\^2\_bdtd\^3x N(([R]{}-2)-2D\^2-\^[-1]{}D\^aD\_a),where =\_b.This expression coincides with that of the Brans-Dicke theory except that the cosmological constant term is present. Comparison with the kinetic part yields the speed of light c\^2=-(c\_1)\^2\_b.As in the case of the Horava gravity the constant $\Lambda_b$ must be negative, consequently allowing only negative cosmological constant $\Lambda$. The Newton constant is related to the expectation value of the scalar field $<\phi>$ as follows,G\_N=.
Now, we consider the homogeneous, isotropic cosmology. We will restrict ourself to the case of $\lambda=1$, $c_1=c_2$, and set the speed of light to unity, i.e., $c=1$. We choose vanishing shift vector $N^a=0$, the three-metric to be the usual maximally symmetric ones with curvature constant $k=-1,0,+1,$ ds\^2= a\^2(t)(+r\^2(d\^2+\^2d\^2)).In this case the higher derivative terms become greatly simplified due to homogeneity and isotropy, Q\^[ab]{}=kc\_1 q\^[ab]{}, Q=6kc\_1.Substituting this result into the action, (\[quad\]), yields the following mini-superspace action,S\_[BDHL]{}=dt a\^3,where the lapse field is set $N=1$ after deriving the field equation. The field equations (including the matter) become 3 ()\^2+3- \^[-1]{}\^2+(-)+ (kc\_1)\^2&=&\_m,\
-2-()\^2- -2- \^[-1]{}\^2-(-)+(kc\_1)\^2&=&p\_m,\
-\^[-1]{}\^2+3-3-3()\^2-(-)-(kc\_1)\^2&=&0,where the matter is assumed, for consistency, to satisfy the usual form of the continuity equation;+3(\_m+p\_m)=0.Only two equations are independent and can be chosen to be 3 H\^2+3H- ()\^2&=&\^[-1]{}\_m-(-)- (),\
(2+3)( +3H)&=&\^[-1]{}(\_m-3p\_m)+2+,\[fried\]where $H\equiv(\dot{a}/{a})$ is the Hubble constant and B\^2=(kc\_1)\^2= .The first in Eq. (\[fried\]) is the Friedmann equation of the Brans-Dicke theory with a negative cosmological term and the dark radiation term included. In the absence of those two terms the equations simply become those of the usual Brans-Dicke theory[@HKim]. Therefore, we restrict our attention to the new effects resulting from those two terms.
Consider two limiting cases for vacuum, $\rho_m=p_m=0$. First, for small $a$ the dark radiation term dominates, so 3 H\^2+3H- ()\^2&=&- (),\
(2+3)( +3H)&=& .These equations can be solved by H&=&,\
&=&,with the two constants $h$ and $g$ satisfying 3h\^2+3hg-g\^2&=&-B\^2,\
(2+3)g(g+h)&=&B\^2.Eliminating the dark energy terms from the equations we find two possibilities $h=-(1/2)g$, or $3h=-(\omega+3)g$. Only the first one gives rise to a solution when $B^2$ is positive, so we get h=-g=.The two signs represent contracting and expanding phases. Solving for $a$ and $\phi$ reads a\^2(t)&=&2ht,\
(t)&=&,where $\phi_0$ is the integration constant. This early universe behavior, $a(t)\sim t^{1/2}$, is the one corresponding to the normal (not dark) radiation source. This is a rather unexpected result due to the scalar field.
Secondly, in the large $a$ limit cosmological term dominates over the curvature and dark radiation terms. We get 3 H\^2+3H- ()\^2&=& ,\
(2+3)(()\^+()\^2 +3H)&=&2.Similarly to the previous case, these two equations are solved by H&=&h,\
&=&g,with $h$ and $g$ satisfying 3h\^2+3hg-g\^2&=&,\
(2+3)g(g+h)&=&.In this case, we get either $h=-\frac{1}{2}g$ or $h=(\omega+1)g$. Only the first is allowed for negative $\Lambda$, and we find h=-=.For the positive sign the solution represents the universe exponentially expanding. It is interesting to note that such a solution exists even for a negative cosmological constant. This is in sharp contrast to the Horava gravity case. Note that the effective cosmological constant $3h^2$ is suppressed by the factor of $\omega$.
To summarize, we have condtructed a Brans-Dicke extension of the Horava-Lifshitz gravity with the detailed balance condition satisfied. We have investigated its IR limit and shown that the resulting IR theory is the Brans-Dicke theory with a negative cosmological constant and a dark radiation term. By studying its cosmological solutions we have shown that exponentially expanding solution at late time and power law expanding solution at early time can exist. This is in contrast with the pure gravity of Horava. Although we focused on the Brans-Dicke theory in this paper the analysis can be generalized to other non-minimally coupled scalar field gravity theory. It would be interesting to investigate further cosmological aspects of the resulting theories.
THL was supported by the Soongsil University Research Fund. PO was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MEST) through the Center for Quantum Spacetime(CQUeST) of Sogang University with grant number 2005-0049409.
[99]{}
P. Horava, [*JHEP*]{}**0903**(2009) 020, arXiv:0812.4287. P. Horava, [*Phys. Rev.*]{} **D79**(2009) 084008, arXiv:0901.3775. P. Horava, [*Phys. Rev. Lett.*]{} **102** (2009) 161301, arXiv:0902.3657. G. Calcagni, [*JHEP*]{} [**0909**]{}, 112 (2009); H. Lu, J. Mei, and C. N. Pope, [*Phys. Rev. Lett.*]{} [**103**]{}, 091301 (2009); E. Kiritsis and G. Kofinas, [*Nucl. Phys.*]{} [**B821**]{}, 467 (2009); T. Takahashi and J. Soda, [*Phys. Rev. Lett.*]{} [**102**]{}, 231301 (2009); R. Brandenberger, [*Phys. Rev.*]{} [**D80**]{}, 043516 (2009); C. Charmousis, G. Niz, A. Padilla, and P. M. Saffin, [*JHEP*]{} [**0908**]{}, 070 (2009); S. Mukohyama, [*JCAP*]{} [**0906**]{}, 001 (2009); R.-G. Cai, L.-M. Cao, and N. Ohta, [*Phys. Rev.*]{} [**D80**]{}, 024003 (2009); H. Nastase, arXiv:0904.3604; A. Kehagias and K. Sfetsos, [*Phys. Lett.*]{} [**B678**]{}, 123 (2009); X. Gao, arXiv:0904.4187; M. Li and Y. Pang, [*JHEP*]{} [**0908**]{}, 015 (2009); T. P. Sotiriou, M. Visser, and S. Weinfurtner, [*JHEP*]{} [**0910**]{}, 033 (2009); R.-G. Cai, Y. Liu, and Y.-W. Sun, [*JHEP*]{} [**0906**]{}, 010 (2009); Y.-S. Piao, [*Phys. Lett.*]{} [**B681**]{}, 1 (2009); T. P. Sotiriou, M. Visser, and S. Weinfurtner, [*Phys. Rev. Lett*]{}. [**102**]{}, 251601 (2009); R.-G. Cai, B. Hu, and H.-B. Zhang, [*Phys. Rev.*]{} [**D80**]{}, 041501 (2009); R.-G. Cai, L.-M. Cao, and N. Ohta, [*Phys. Lett.*]{} [**B679**]{}, 504 (2009); E. O Colgain and H. Yavartanoo, [*JHEP*]{} [**0908**]{}, 021 (2009); Y. S. Myung and Y.-W. Kim, arXiv:0905.0179; M. Visser, [*Phys. Rev.*]{} [**D80**]{}, 025011 (2009); D. Orlando and S. Reffert, [*Class. Quant. Grav.*]{} [**26**]{}, 155021 (2009); A. Ghodsi and E. Hatefi, arXiv:0906.1237. B. Ratra and P.J.E. Peebles, [*Phys. Rev.*]{} **D37** (1988) 3406; R.R Caldwell, R. Dave and P.J. Steinhardt, [*Phys. Rev. Lett.*]{} **80**, 1582 (1998). C. Armendariz-Picon, V. Mukhanov and P.J. Steinhardt, [*Phys. Rev. Lett.*]{} **85**, 4438 (2000); T. Okabe and M. Yamaguchi, [*Phys. Rev.*]{} **D62**, 023511 (2000); M. Malquarti, E.J. Copland, A.R. Liddle and M. Trodden, [*Phys. Rev.*]{} **D67**, 123503 (2003). S. Mukohyama, [*JCAP*]{} [**0906**]{}, 001 (2009); E. Kiritsis and G. Kofinas, [*Nucl. Phys.*]{} [**B821**]{}, 467 (2009); B. Chen and Qing-Guo Huang, [*Phys. Lett.*]{} **B683**, 108 (2010). C. Brans and R. H. Dicke, [*Phys. Rev.*]{} **124**, 925 (1961). C.G. Callan, D. Friedan, E.J. Martinec and M.J. Perry, [*Nucl. Phys.*]{} **B262**, 597 (1985). G. Calcagni, [*JHEP*]{}**09**, 112 (2009), arXiv:0904.0829. K. Uehara, C.W. Kim, [*Phys. Rev.*]{} **D26**, 2575 (1982). M. Arik, M. C. Calik, [*JCAP*]{} **0501**, 013 (2005), gr-qc/0403108. B. Kelleher, [*Class. Quantum Grav.*]{} **21**, 483 (2004). Hongsu Kim, [*Mon. Not. Roy. Astron. Soc.*]{} **364**, 813 (2005), astro-ph/0408577.
[^1]: This result was considered in the context of conformal gravity in Ref. [@bk].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The aim of this paper is to show that fermion and boson random point processes naturally appear from representations of CAR and CCR which correspond to gauge invariant generalized free states (also called quasi-free states). We consider particle density operators $\rho(x)$, $x\in{{\Bbb R}}^d$, in the representation of CAR describing an infinite free Fermi gas of finite density at both zero and finite temperature [@AW], and in the representation of CCR describing an infinite free Bose gas at finite temperature [@ArWoods]. We prove that the spectral measure of the smeared operators $\rho(f)=\int dx\, f(x)\rho(x)$ (i.e., the measure $\mu$ which allows to realize the $\rho(f)$’s as multiplication operators by ${\langle}\cdot,f{\rangle}$ in $L^2(d\mu)$) is a well-known fermion, resp. boson process on the space of all locally finite configurations in ${{\Bbb R}}^d$.'
author:
- EUGENE LYTVYNOV
---
\[section\] \[section\] \[section\] \[section\] \[section\] \[section\]
addtoreset[equation]{}[section]{}
[**Fermion and boson random point processes as particle distributions of infinite free Fermi and Bose gases of finite density** ]{}
[*Institut für Angewandte Mathematik, Universität Bonn, Wegelerstr. 6, D-53115 Bonn, Germany; SFB 256, Univ. Bonn, Germany; BiBoS, Univ. Bielefeld, Germany*]{}
Introduction: representations of current algebras; boson and fermion processes
==============================================================================
The nonrelativistic quantum mechanics of many identical particles may be described by means of a field $\psi(x)$, $x\in{{\Bbb R}}^d$, satisfying either canonical commutation relations (CCR) and describing bosons: $$\begin{gathered}
[\psi(x),\psi(y)]_- =
[\psi^*(x),\psi^*(y)]_-=\pmb0 ,\notag\\ [\psi^*(x),\psi(y)]_-
=\delta(x-y)\pmb1\label{z7eawr76},\end{gathered}$$ or satisfying canonical anticommutation relations (CAR) and describing fermions: $$\begin{gathered}
[\psi(x),\psi(y)]_+ =
[\psi^*(x),\psi^*(y)]_+ =\pmb0,\notag\\ [\psi^*(x),\psi(y)]_+
=\delta(x-y)\pmb1.\label{tzew45}\end{gathered}$$ Here, $[A,B]_\mp=AB\mp BA$ is the commutator, resp. anticommutator. The statistics of the system is thus determined by the algebra which is to be represented.
In the formulation of nonrelativistic quantum mechanics in terms of particle densities and currents, one defines $$\begin{aligned}
\rho(x)&{:=}\psi^*(x)\psi(x),\notag\\
J(x)&{:=}(2i)^{-1}\big( \psi^*(x)\nabla
\psi(x)-(\nabla\psi^*(x))\psi(x)\big).\label{u8zearg6}\end{aligned}$$ Using CCR or CAR, one can formally compute the commutation relations satisfied by the smeared operators $\rho(f){:=}\int_{{{\Bbb R}}^d}dx\, f(x)\rho(x)$ and $J(v){:=}\int_{{{\Bbb R}}^d}dx\, v(x)\cdot J(x)$. These turn out to be $$\begin{gathered}
_-=\pmb0,\notag\\
[\rho(f),J(v)]_-=i\rho(v\cdot\nabla f),\notag\\
[J(v_1),J(v_2)]_-=-iJ(v_1\cdot\nabla v_2-v_2\cdot \nabla
v_1),\label{zue6ed}\end{gathered}$$ independently of whether one starts with CCR or CAR.
Thus, in a nonrelativistic current theory, the particle statistics is not determined by a choice of an equal-time algebra, but instead may be determined by the choice of a representation of the algebra, see e.g. [@DaSh; @Go; @GoMeSha] and the references therein. Since the operators $\rho(f)$ and $J(v)$ are generally speaking unbounded, one usually starts with study of the group $G$ obtained by exponentiating the algebra $\frak g$ generated by the commutation relations . More precisely, considering $\rho(f)$ and $J(v)$ to be self-adjoint, the corresponding one-parameter groups are $${\cal
U}(tf)=\exp\big[it\rho(f)\big],\qquad {\cal
V}(\phi_t^v)=\exp\big[it J(v))\big],\label{zftft}$$ where $\phi_t^v$ is the one-parameter group of diffeomorphisms (or flows) on ${{\Bbb R}}^d$ generated by the vector field $v$: $$\frac{\partial \phi_t^v}{\partial t}=v(\phi_t^v),\qquad
\phi_{t=0}^v(x)=x.$$ From equation , the operators satisfy the group law $${\cal U}(f_1){\cal
V}(\psi_1){\cal U}(f_2){\cal V}(\psi_2)={\cal U}(f_1+f_2\circ
\psi_1){\cal V}(\psi_2\circ\psi_1).$$ Hence, the group $G$ is the semidirect product $S({{\Bbb R}}^d)\wedge \operatorname{Diff}({{\Bbb R}}^d)$, where $S({{\Bbb R}}^d)$ is the Schwartz space of rapidly decreasing functions on ${{\Bbb R}}^d$ and $\operatorname{Diff}({{\Bbb R}}^d)$ is a certain group of diffeomorphisms of ${{\Bbb R}}^d$ of Schwartz’s type (and thus containing all diffeomorphisms with compact support, i.e., which are identical outside a compact set). Due to physical interpretation, the currents $\rho(f)$ and $J(v)$ (and the group $G$) can be taken as fundamental structures of quantum mechanics. It should be noted that, given a representation of the group $G$ and its current algebra $\frak g$, corresponding operators $\psi(x), \psi^*(x)$ may, in general, not exist.
Let ${\cal U}(f)$ be a continuous unitary cyclic representation of $S({{\Bbb R}}^d)$ in a Hilbert space $\cal H$ with cyclic vector $\Omega$. The functional $L(f){:=}({\cal U}(f)\Omega,\Omega)$ satisfies the conditions of the Bochner–Minlos theorem, and hence is the Fourier transform of a probability measure $\mu$ on $S'({{\Bbb R}}^d)$, the dual of $S({{\Bbb R}}^d)$: $$L(f)=\int_{S'({{\Bbb R}}^d)}\exp\big[i{\langle}\omega,f{\rangle}\big]\,
\mu(d\omega).$$ Therefore, ${\cal H}$ can be realized as $L^2(S'({{\Bbb R}}^d);\mu)$, $\Omega= 1$, and ${\cal U}(f)$ as the multiplication operator by $\exp\big[i{\langle}\cdot,f{\rangle}\big]$ in $L^2(S'({{\Bbb R}}^d);\mu)$.
Let now ${\cal U}(f){\cal V}(\psi)$ be a continuous unitary cyclic representation of the group $S({{\Bbb R}}^d)\wedge
\operatorname{Diff}({{\Bbb R}}^d)$ in $\cal H$. Suppose additionally that the cyclic vector $\Omega$ is also cyclic for (the smaller family of unitary operators) ${\cal U}(f)$. From some physical reasons, one believes that, in the spinless case, the latter condition is always fulfilled as long as one does not deal with parastatistics. Then, realizing the Hilbert space ${\cal H}$ as $L^2(S'({{\Bbb R}}^d);\mu)$ just as above, by [@Go], one has that the measure $\mu$ is quasi-invariant for $\operatorname{Diff}({{\Bbb R}}^d)$ and the operators ${\cal V}(\psi)$ become $$\label{zgufgtgf12} ({\cal V}(\psi)F)(\omega)=
\chi_\psi(\omega)F(\psi^*\omega)\bigg(\frac{d\mu(\psi^*\omega)}{d\mu(\omega)}\bigg)^{1/2},
\qquad \text{$\mu$-a.e.\ $\omega\in S'({{\Bbb R}}^d)$},$$ for all $F\in L^2(S'({{\Bbb R}}^d);\mu)$. Here, $d\mu(\psi^*\omega)/d\mu(\omega)$ is the Radon–Nikodym derivative of the transformed measure with respect to the original measure, and $\chi_\psi(\omega)$ is a cocycle, i.e., $\chi_\psi(\cdot)$ is a complex-valued function of modulus one, depending on $\psi$, defined $\mu$-a.e., and satisfying, for each $\psi_1,\psi_2\in\operatorname{Diff}({{\Bbb R}}^d)$, $\chi_{\psi_2}(\omega)\chi_{\psi_1}(\psi_2^*\omega)=\chi_{\psi_1\circ\psi_2}(\omega)$ $\mu$-a.e.
A powerful method of construction of continuous unitary cyclic representations of $G$ has been the method of generating functional. A continuous complex-valued function $E$ on $G$ is called a generating functional on $G$ if the following conditions are fulfilled: 1) $E(e)=1$, where $e$ is the identity element of $G$; 2) $\sum_{i,j=1}^N\bar\lambda_i\lambda_j E(g_i^{-1}g_j)\ge0$ for all $\lambda_i\in{{\Bbb C}}$, $g_i\in G$, $i=1,\dots,N$, $N\in{{\Bbb N}}$. By Araki’s theorem [@Arrr], $E$ is a generating functional on $G$ if and only if there exists a continuous unitary cyclic representation $\pi$ of $G$ in $\cal H$ with cyclic vector $\Omega$ such that $E(g)=(\Omega,E(g)\Omega)$, $g\in G$. Thus, one may implicitly construct unitary representations of $G$ by finding generating functionals.
In [@GoGrPoSha] (see also [@AKR1; @AKR2; @VGG]), the case of an infinite free Bose gas at zero temperature with average particle density $\rho>0$ was studied in the formalism of local current algebras. Goldin [*et al*]{}. started with considering a system of $N$ bosons in a box of volume $V$. The physical Hilbert space $\cal H$ is now $L^2_{\mathrm s}(V^N)$, the subspace of $L^2(V^N)$ consisting of all symmetric functions (we also used the letter $V$ to denote the box itself). The Hamiltonian for $N$ boson particles is given by $H_{N,V}=-\frac12\sum_{i=1}^N\Delta_i$ with periodic boundary conditions, and the normalized ground state wave function $\Omega_{N,V}=V^{-N/2}$. The representation of the group $G$ in the box $V$ is given by $$\begin{aligned}
{\cal
U}_{N,V}(f)F(x_1,\dots,x_N&)=\exp\bigg(i\sum_{i=1}^N
f(x_i)\bigg)F(x_1,\dots,x_N),\notag\\ {\cal
V}_{N,V}(\psi)F(x_1,\dots,x_N)&=F(\psi(x_1),\dots,\psi(x_N))\prod_{i=1}^N
\sqrt{J_\psi(x_i)},\label{574}\end{aligned}$$ where $f$ and $\psi$ have support inside the box $V$ and $J_\psi(x)=\det(\partial
\psi^k(x)/\partial x_l)_{k,l=1}^d$ is the Jacobian of the flow. Thus, one can write down the generating functional $$\label{gzdrtdtds} E_{N,V}(f,\psi)=
(\Omega_{N,V},{\cal U}_{N,V}(f){\cal
V}_{N,V}(\psi)\Omega_{N,V})$$ of this representation and take the so-called $N/V$-limit, i.e., the limit as $N,V\to\infty$, $N/V\to\rho$. The limiting functional then has the form $$E(f,\psi)=\exp\bigg(\rho\int_{{{\Bbb R}}^d}\big(e^{if(x)}\sqrt{J_\psi(x)}-1\big)\,dx\bigg).$$ The authors then showed that $\Omega$ is cyclic for ${\cal U}(f)$, $f\in S({{\Bbb R}}^d)$, and hence this representation can be realized on the space $L^2(S'({{\Bbb R}}^d);\mu)$, where the Fourier transform of the measure $\mu$ is equal to $$\label{zugqzgu}\int_{S'({{\Bbb R}}^d)}\exp\big(i{\langle}\omega,
f{\rangle}\big)\,\mu(d\omega)=\exp\bigg(\rho\int_{{{\Bbb R}}^d}\big(e^{if(x)}-1\big)\,dx\bigg).$$ Thus, $\mu=\pi_\rho$ is the Poisson measure with intensity $\rho\,
dx$. This measure is concentrated on the space $\Gamma_{{{\Bbb R}}^d}$ of all locally finite configurations in ${{\Bbb R}}^d$. As for the operators ${\cal V}(\psi)$ in this representation, the general formula now takes the following form: all the cocycles are identically equal to one and the Radon–Nykodym derivative is given by $$\frac
{d\pi_\rho(\psi^*\gamma)}{d\pi_\rho(\gamma)}=\prod_{x\in\gamma}J_\psi(x),\qquad
\text{$\pi_\rho$-a.e.\ $\gamma\in\Gamma_{{{\Bbb R}}^d }.$}$$ One may also derive an explicit formula for the action of the operators $J(v)$ in $L^2(\Gamma_{{{\Bbb R}}^d};\pi_\rho)$, which particularly shows that these are certain differential operators on the configuration space (see [@AKR1] for details).
Furthermore, it was shown in [@GoGrPoSha] that the representation of $G$ defined by is unitarily equivalent to the representation in the symmetric Fock space ${{\cal F}}_{\mathrm s}(L^2({{\Bbb R}}^d))$ in which the operators $\rho(x)$, $J(x)$ are defined by formula with $$\label{uilsagfz} \psi(x)=\psi_{\mathrm
F}(x)+\sqrt\rho,\qquad \psi^*(x)=\psi^*_{\mathrm F}(x)+\sqrt\rho$$ and $\Omega$ is the vacuum vector in ${{\cal F}}_{\mathrm s}(L^2({{\Bbb R}}^d))$. In , $\psi_{\mathrm F}(x)$, $\psi^*_{\mathrm
F}(x)$ are the standard annihilation and creation operators in the Fock space, respectively. In fact, this unitary equivalence has played a crucial role in the study of the representation defined by .
On the other hand, the obtained unitary $I:{\cal F}_{\mathrm
s}(L^2({{\Bbb R}}^d))\to L^2(\Gamma_{{{\Bbb R}}^d};\pi_\rho)$ coincides with the well-known chaos decomposition for the Poisson measure (e.g.[@Sur]). The operators $$\rho(x)=\psi^*(x)\psi(x)=\psi_{\mathrm F}^*(x)\psi_{\mathrm
F}(x)+\sqrt\rho\,\psi^*_{\mathrm F}(x)+\sqrt{\rho}\,\psi_{\mathrm
F}(x)+\rho$$ are known in quantum probability as quantum Poisson white noise (e.g. [@HuPa; @mey]). The $\pi_\rho$ can also be thought of as the spectral measure of the family $(\rho(f))_{f\in S({{\Bbb R}}^d) }$ (cf.[@BeKo Ch. 3] and [@Ber; @Ly]).
In fact, it was in Araki and Woods’ paper [@ArWoods] dealing with representations of CCR that the operators first appeared in the description of an infinite free Bose gas at zero temperature.
In [@M1; @M2] (see also [@Girard]), a unitary cyclic representation of the group $G$ describing an infinite free Fermi gas at zero temperature was, in particular, studied. Menikoff started with a system of $N$ free Fermi particles in a cubic box $V$ in ${{\Bbb R}}^3$ with edges of length $L$. If $k_f$ is the Fermi momentum of the system, the number of the particles in the box $V$ is equal to $$\label{guzftzf} N=\# \big\{k_n =(2\pi/L)(n_1,n_2,n_3): \, n_1,n_2,n_3\in{{\Bbb Z}},\,
|k_n|\le k_f\big\}.$$ The Hilbert space of the system is then ${\cal H}=L^2_{\mathrm a}
(V^N)$, the subspace of $L^2(V^N)$ consisting of all antisymmetric functions, and the Hamiltonian of the system $H_{N,V}=-\frac12\sum_{i=1}^N\Delta_i$. The normalized ground state of $H_{N,V}$ is $$\Omega_{N,V}(x_1,\dots,x_N)=(V^{-N}/N!)^{1/2}\,\det\big(\exp(ik_n\cdot
x_m)\big)_{n,m=1}^N,$$ where the $k_n$’s are as in . The representation of the group $G$ in the box $V$ is given by the same formulas but with $F\in
L^2_{\mathrm s}(V^N)$. By , we get for the average particle density $$N/V= N/(L^3)\to \rho=(4/3)
\pi(k_f/2\pi)^3\quad\text{as }L\to\infty.$$ Taking the limit $N,V\to\infty$, $N/V\to \rho$ of the generating functional $E_{N,V}(f,\psi)$ again given by , one gets the following results. Let $${\varkappa}(x){:=}(2\pi)^{-3}\int_{\{|\lambda|< k_f\} }e^{i\lambda
x}\,d\lambda=3\rho(\sin z-z\cos z)/(z^3)\big|_{z=k_f|x|},\qquad
x\in{{\Bbb R}}^3,$$ and let $$R_n(y_1,\dots,y_n;x_1,\dots,x_n){:=}\det
({\varkappa}(x_i-y_j))_{i,j=1}^n,\qquad n\in{{\Bbb N}}.$$ Then, the limiting generating functional is given by $$\begin{gathered}
\label{56}
E(f,\psi)=1+\sum_{n=1}^\infty \int dx_1\int dy_1\dotsm \int
dx_n\int dy_n
\prod_{i=1}^n\big[\delta(x_i-y_i)(e^{if(x_i)}T_{x_i}(\psi)-1)\big]\\
\times R_n(y_1,\dots,y_n;x_1,\dots,x_n),\end{gathered}$$ where $ T_x(\psi)
g(x){:=}g(\psi(x))\sqrt{J_\psi(x)}$. In particular, $$\label{jhifztfd} L(f)=1+\sum_{n=1}^\infty
\int_{({{\Bbb R}}^3)^n}\bigg(\prod_{i=1}^n (e^{if(x_i)}-1)\bigg)
\det({\varkappa}(x_i-x_j))_{i,j=1}^n\, dx_1\dotsm dx_n,$$ which is the Fourier transform of a measure $\mu_{\mathrm a}$ on $S'({{\Bbb R}}^3)$. Furthermore, it follows from that the measure $\mu_{\mathrm a}$ has correlation functions $$\label{tzdds} k_{\mu_{\mathrm a}}^{(n)}(x_1,\dots,x_n)=\det
({\varkappa}(x_i-x_j))_{i,j=1}^n.$$ Menikoff mentioned in [@M2] that from the existence of correlation functions it should follow that the measure $\mu_{\mathrm a}$ is concentrated on $\Gamma_{{{\Bbb R}}^d}$, however he could not prove it.
Two important problems remained open after [@M1; @M2]: 1) Is there any connection with the representation of CAR for an infinite free Fermi gas at zero temperature? Does corresponding $\psi(x),\, \psi^*(x)$ operators exist? 2) Is $\Omega$—the cyclic vector of the representation defined by —also cyclic for the ${\cal U}(f)$’s, and if it is so, what is the form of the Radon–Nikodym derivative and the cocycles in the formula in this case?
When studying statistical properties of a chaotic beam of fermions by using wavepacket formalism, Benard and Macchi [@BeMa] arrived at measures on the configuration space over a bounded volume whose correlation functions are given by a formula of type . In [@Ma1; @Ma2], Macchi called a measure on the configuration space a fermion process if the respective correlation functions are given by in which ${\varkappa}(x-y)$ is a non-negative definite function. She gave sufficient conditions for the existence of such a measure. Fermion process (also called determinantal random point fields) are often met with in random matrix theory, probability theory, representation theory, and ergodic theory. We refer to the paper [@Sosh] containing an exposition of recent as well as sufficiently old results on the subject. Scaling limits of fermion point processes are proved in [@Scale1; @Scale2; @Spohn]. We also refer to the recent papers [@Shi; @ShiYoo] for a discussion of different problems connected with fermion processes on the configuration space over the lattice ${{\Bbb Z}}^d$.
In a parallel way, a boson process was defined and studied in [@BeMa; @Ma0; @Ma1]. This process is defined as a probability measure $\mu$ on $\Gamma_{{{\Bbb R}}^d}$ whose correlation functions are given by $$k_\mu^{(n)}={\operatorname{per}}({\varkappa}(x_i-x_j))_{i,j=1}^n,$$ where ${\varkappa}(x-y)$ is again a non-negative definite function and the permanent ${\operatorname{per}}A$ of a matrix $A$ contains the same terms as the corresponding determinant $\det A$ but with constant positive signs for each product of matrix elements in place of the alternating positive and negative signs of the determinant. It should be mentioned that any boson process is a Cox process, i.e., a Poisson process with a random intensity measure, see [@DV Sec. 8.5].
In [@F; @FF] (see also [@FF2]), point processes were constructed and studied which correspond to locally normal states of a boson system, and which can be interpreted as the position distribution of the state. More precisely, let ${\cal L}({\cal F
}_{\mathrm s}(L^2({{\Bbb R}}^d)))$ be the von Neumann algebra of all bounded linear operators in the symmetric Fock space ${\cal F
}_{\mathrm s}(L^2({{\Bbb R}}^d))$, and let $\cal A$ be its $C^*$-subalgebra obtained as the uniform closure of all local von Neumann algebras in ${\cal L}({\cal F }_{\mathrm s}(L^2({{\Bbb R}}^d)))$. Let $\omega$ be a locally normal state on $\cal A$ (cf.[@BraRo]). As well known, the symmetric Fock space ${\cal F
}_{\mathrm s}(L^2({{\Bbb R}}^d))$ may be isomorphically realized as the $L^2$-space $L^2(\Gamma_{{{\Bbb R}}^d}^{\mathrm fin};\lambda)$, where $\Gamma_{{{\Bbb R}}^d}^{\mathrm fin}$ is the space of all finite configurations in ${{\Bbb R}}^d$ and $\lambda$ is the Lebesgue–Poisson measure on $\Gamma_{{{\Bbb R}}^d}^{\mathrm fin}$. (Notice the evident inclusion $\Gamma_{{{\Bbb R}}^d}^{\mathrm fin}\subset \Gamma_{{{\Bbb R}}^d}$.) Then, every bounded function $F$ on $\Gamma_{{{\Bbb R}}^d}^{\mathrm fin}$ determines a bounded operator ${\cal M}_F$ of multiplication by $F$. A function $F$ on $\Gamma_{{{\Bbb R}}^d}$ is called local if there exists a compact set $\Lambda\subset {{\Bbb R}}^d$ such that $F(\gamma)=F(\gamma\cap\Lambda)$ for all $\gamma\in\Gamma_{{{\Bbb R}}^d}$. The restriction of $F$ to $\Gamma_{{{\Bbb R}}^d}^{\mathrm fin}$ is a local function on $\Gamma_{{{\Bbb R}}^d}^{\mathrm fin}$, for which we preserve the notation $F$. In [@FF], it was proved that there exists a unique probability measure $\mu_\omega$ on $\Gamma_{{{\Bbb R}}^d}$ such that, for all bounded local functions $F$ on $\Gamma_{{{\Bbb R}}^d}$, $$\int_{\Gamma_{{{\Bbb R}}^d}}F(\gamma)\,\mu_\omega(d\gamma)=\omega({\cal
M }_F).$$ Some properties of such point processes were also studied in [@FF]. In [@FF2], it was shown that, if the reduced density matrices $\rho^{(n)}_\omega(x_1,\dots,x_n;y_1\dots,y_n)$ of the state $\omega$ exist and are continuous, then the correlation functions of the measure $\mu_\omega$ are given by $$k_{\mu_\omega}^{(n)}(x_1,\dots,x_n)=\rho_\omega^{(n)}(x_
1,\dots,x_n;x_1,\dots,x_n).\label{zggtfft}$$ Furthermore, the special case corresponding to the ideal Bose gas (cf. [@BraRo]) was studied in detail in [@F]. By , $\mu_\omega$ is now the boson measure with $${\varkappa}(x)=\sum_{n=1}^\infty \frac{z^n}{(4\pi \beta
n)^{d/2}}\,\exp\big[-|x|^2/(4n\beta)\big],$$ where $\beta$ is the inverse temperature and $z$ the activity. This boson measure was proved to be an infinite divisible point process. It should also be noted that the local normality of the states we discuss below, in Section 2, was established in e.g. [@BraRo].
The aim of this paper is to show a connection between representations of CAR, resp. CCR describing infinite free Fermi, resp. Bose gases of finite density and the fermion, resp.boson random point processes.
In Section 2, we recall Araki and Wyss’ representations of CAR in the the antisymmetric Fock space that describes infinite free Fermi gases at both finite and zero temperature [@AW], and Araki and Woods’ representation of CCR in the symmetric Fock space that describes an infinite free Bose gas at finite temperature [@ArWoods]. The results of this section are essentially known (with the only exception that the corresponding “annihilation” and “creation” operators $\psi(x),\psi^*(x)$ has not been treated without smearing).
In Section 3, we prove that the corresponding particle density operators are well-defined and form a family $(\rho(f))_{f\in
S({{\Bbb R}}^d)}$ of commuting selfadjoint operators. Then, we introduce the space ${\frak H}_{\sharp}$ ($\sharp=$a in the fermionic case and $\sharp=$s in the bosonic case) as the closed linear span of the vectors of the form $\rho(f_1)\dotsm\rho(f_n)\Omega$. Restricted to this space, $(\rho(f))_{f\in S({{\Bbb R}}^d)}$ evidently becomes a cyclic family. Using the spectral theory of cyclic families of commuting selfadjoint operators [@BeKo; @Sam], we then show that there exist a unique probability measure $\mu_\sharp$ on $S'({{\Bbb R}}^d)$—the Schwartz space of tempered distributions—and a unitary operator $I_\sharp:{\frak
H}_\sharp\to L^2(S'({{\Bbb R}}^d);\mu_\sharp)$ such that $I_\sharp\Omega=1$ and $I_\sharp\rho(f)I_\sharp^{-1}={\langle}\cdot,f{\rangle}\cdot$ for each $f\in
S({{\Bbb R}}^d)$.
Next, we introduce an operator field ${{:}\,}\rho(x_1)\dotsm\rho(x_n){\,{:}}$ via a recurrence relation, and prove that $${{:}\,}\rho(x_1)\dotsm\rho(x_n){\,{:}}=\psi^*(x_n)\dots
\psi^*(x_1)\psi(x_1)\dotsm\psi(x_n).$$ Thus, ${{:}\,}\rho(x_1)\dotsm\rho(x_n){\,{:}}$ is nothing but a normal (Wick) product of $\rho(x_1),\dots,\rho(x_n)$. Using this and results of [@BKKL], we explicitly calculate the correlation functions of $\mu_\sharp$. This enables us, first, to show that $\mu_\sharp$ is concentrated on the configuration space $\Gamma_{{{\Bbb R}}^ d}$, and second, to identify $\mu_{\mathrm a}$ as a fermion process, and $\mu_{\mathrm s}$ as a boson process. In particular, starting from the representation of CAR [@AW] corresponding to an infinite free Fermi gas at zero temperature, we arrive at the the same probability measure $\mu_{\mathrm a}$ as Menikoff did in [@M2].
Thus, the main results of the paper are as follows: 1) We introduced the particle density operators (quantum white noise) $\rho(x)$, $x\in{{\Bbb R}}^d$, corresponding to an infinite free Fermi gas (of finite density) at zero and at finite temperature, resp. an infinite free Bose gas at finite temperature, proved the well-definedness of $\rho(x)$ and the essential self-adjointness of the corresponding field $(\rho(f))_{f\in S({{\Bbb R}}^d)}$. 2) We proved a one-to-one correspondence between the operator field $(\rho(f))_{f\in S({{\Bbb R}}^d)}$ and a fermion, resp. boson point process $\mu_\sharp$ on $\Gamma_{{{\Bbb R}}^d}$. Furthermore, the constructed unitary isomorphism between the spaces ${\frak
H}_{\sharp}$ and $L^2(\Gamma_{{{\Bbb R}}^d};\mu_{\sharp})$ can be thought of as a kind of a chaos decomposition of $L^2(\Gamma_{{{\Bbb R}}^d};\mu_{\sharp})$ (compare with the Poisson case).
We also note that, though the very existence of a fermion process under a slightly stronger condition on the function ${\varkappa}$ in terms of its Fourier transform has been known before (cf.[@BO Proposition 4.1]), as a by-product of our results we get a new proof of the existence of fermion (as well as boson) processes.
It is still an open problem to show that also the operators $J(v)$ may be realized on $L^2(\Gamma_{{{\Bbb R}}^d};\mu_\sharp)$, but we hope that the obtained unitary operator between the latter space and the corresponding subspace of the Fock space may be of some help to tackle this problem.
Infinite free Fermi and Bose gases of finite density {#gzdtese}
====================================================
We first recall the construction of a cyclic representation of CAR whose state (constructed with respect to the cyclic vector) is a gauge invariant generalized free state. This representation is due to Araki and Wyss [@AW]. Generalized free states were first defined and studied by Shale and Stinespring [@ShSt]. Since that generalized free states (also called quasi-free states) have been studied by several authors, see e.g. [@Araki; @BV; @BraRo; @DA; @MV; @PS; @R] and the references therein.
Let $H$ be a separable real Hilbert space and let $H_{{{\Bbb C}}}$ denote its complexification. We suppose that the scalar product in $H_{{{\Bbb C}}}$, denoted by $(\cdot,\cdot)_{H_{{{\Bbb C}}}}$, is antilinear in the first dot and linear in the second one. Let $${{\cal F}}_{\mathrm a}
(H)\big(={{\cal F}}_{\mathrm a}(H_{{\Bbb C}})\big){:=}\bigoplus_{n=0}^\infty
{{\cal F}}_{\mathrm a}^{(n)} (H)$$ denote the antisymmetric Fock space over $H$. Here, ${{\cal F}}_{\mathrm a}^{(0)}(H){:=}{{\Bbb C}}$ and, for $n\in{{\Bbb N}}$, ${{\cal F}}_{\mathrm a}^{(n)}(H){:=}H_{{{\Bbb C}}}^{\wedge n}$, $\wedge$ standing for antisymmetric tensor product. By ${{\cal F}}_{\mathrm a,\, fin}(H)$ we denote the subset of ${{\cal F}}_{\mathrm
a}(H)$ consisting of all elements $f=(f^{(n)})_{n=0}^\infty\in
{{\cal F}}_{\mathrm a}(H)$ for which $f^{(n)}=0$, $n\ge N$, for some $N\in {{\Bbb N}}$. We endow ${{\cal F}}_{\mathrm a,\, fin}(H)$ with the topology of the topological direct sum of the spaces ${{\cal F}}_{\mathrm a}^{(n)}
(H)$. Thus, the convergence in ${{\cal F}}_{\mathrm a,\, fin}(H)$ means uniform finiteness and coordinate-wise convergence.
For $f\in H_{{\Bbb C}}$, we denote by $a(f)$ and $a^*(f)$ the standard annihilation and creation operators on ${{\cal F}}_{\mathrm a}(H)$. They are defined on the domain ${{\cal F}}_{\mathrm a,\, fin}(H)$ through the formula $$\begin{aligned}
a(f) h_1\wedge\dotsm\wedge h_n &{:=}\frac1{\sqrt
n}\, \sum_{i=1}^n (-1)^{i+1} (f,h_i)_{H_{{{\Bbb C}}}} h_1\wedge \dotsm
\wedge h_{i-1}\wedge \check h_i\wedge h_{i+1}\dotsm \wedge
h_n,\notag \\ a^*(f) h_1\wedge\dotsm\wedge h_n &{:=}\sqrt{n+1}\,
f\wedge h_1\wedge\dotsm\wedge h_n,\label{ftwsea}\end{aligned}$$ where $h_1,\dotsm,h_n\in H_{{{\Bbb C}}}$ and $\check h_i$ denotes the absence of $h_i$. The operator $a^*(f)$ is the restriction to ${{\cal F}}_{\mathrm a,\, fin}(H)$ of the adjoint of $a(f)$ in ${{\cal F}}_{\mathrm a}(H)$, and both $a(f)$ and $a^*(f)$ act continuously on ${{\cal F}}_{\mathrm a,\, fin}(H)$. The annihilation and creation operators satisfy CAR: $$\begin{gathered}
_+=\pmb 0,\\ [a(g),a^*(f)]_+
=(g,f)_{H_{{{\Bbb C}}}} \pmb 1 \end{gathered}$$ for all $f,g\in H_{{{\Bbb C}}}$.
Let $K$ be a linear operator in $H_{{\Bbb C}}$ such that $\pmb 0\le K\le
\pmb 1$. We take the direct sum $H\oplus H$ of two copies of the Hilbert space $H$, and construct the antisymmetric Fock space ${{\cal F}}_{\mathrm a}(H\oplus H)$. For $f\in H_{{\Bbb C}}$, we denote $a_1(f){:=}a(f,0)$, $a_2(f){:=}a(0,f)$ and analogously $a_i^*(f)$, $i=1,2$. Let also $K_1{:=}K^{1/2}$, $K_2{:=}(\pmb 1-K)^{1/2}$. We then set, for $f\in H_{{\Bbb C}}$, $$\psi(f){:=} a_2 (K_2f)+a^*_1(JK_1f),\qquad \psi^*(f){:=}
a_2^* (K_2f)+a_1(JK_1f),\label{gztde45}$$ where $J:H_{{\Bbb C}}\to H_{{\Bbb C}}$ is the operator of complex conjugation: $Jf{:=}\overline f$. As easily seen, the operators $\{\psi(f),\psi^*(f)\mid f\in H_{{\Bbb C}}\}$ again satisfy CAR. Let $H_i$ denote the closure of $\operatorname{Im}K_i$ in $H_{{\Bbb C}}$, $i=1,2$. Then, restricted to the subspace ${{\cal F}}_{\mathrm a}(H_1\oplus H_2)$, the operators $\{\psi(f),\psi^*(f)\mid f\in H_{{\Bbb C}}\}$ form a cyclic representation of CAR with cyclic vector $\Omega{:=}(1,0,0,\dots)$—the vacuum in ${{\cal F}}_{\mathrm a}(H\oplus H )$.
Let ${\frak A}_{\mathrm a}(H_{{\Bbb C}})$ denote the $C^*$-algebra generated by the operators $\psi(f)$, $f\in H_{{\Bbb C}}$, and let $\omega_{\mathrm a}$ be the state on ${\frak A}_{\mathrm a}(H_{{\Bbb C}})$ defined by $\omega_{\mathrm
a}(\Psi){:=}(\Psi\Omega,\Omega)_{{{\cal F}}_{\mathrm a}(H_1\oplus H_2)}$, $\Psi\in{\frak A}_{\mathrm a}(H_{{\Bbb C}})$. The $n$-point functions of $\omega_{\mathrm a}$ are given by the formula $$\label{32rsfd} \omega_{\mathrm a}(\psi^*(f_n)\dotsm \psi^*(f_1)\psi(g_1)\dotsm
\psi(g_m))=\delta_{n,m} \det ((f_i,Kg_j)_{H_{{\Bbb C}}})$$ for all $f_1,\dots,f_n,g_1,\dots, g_m\in H_{{\Bbb C}}$. Thus, $\omega_{\mathrm
a}$ is a gauge invariant generalized free state corresponding to the operator $K$.
An analogous representation of CCR was constructed by Araki and Woods [@ArWoods] (historically it preceded the representation of CAR [@AW]). Let us outline it. In the symmetric Fock space over $H$, denoted by ${{\cal F}}_{\mathrm s}(H)$, we construct the standard annihilation and creation operators, $b(f)$ and $b^*(f)$, which satisfy CCR: $$\begin{gathered}
_- =\pmb0,\\ [b(g),b^*(f)]_- =(g,f)_{H_{{\Bbb C}}}\pmb 1\end{gathered}$$ for all $f,g\in H_{{\Bbb C}}$. Let now ${\cal K}$ be a bounded linear operator in $H_{{\Bbb C}}$ such that ${\cal
K}\ge\pmb0$. We set ${\cal K}_1{:=}{\cal K}^{1/2}$, ${\cal
K}_2{:=}(\pmb1+{\cal K})^{1/2}$. Analogously to , we define the following operators in ${{\cal F}}_{\mathrm s}(H\oplus H)$: $$\varphi(f){:=} b_2
({\cal K}_2f)+b_1^*(J{\cal K}_1f),\qquad \varphi^*(f){:=}b_2^*
({\cal K }_2 f)+b_1(J{\cal K}_1 f)\label{gztfdrde45}$$ for $f\in H_{{\Bbb C}}$. These operators again satisfy CCR and form a cyclic representation of CCR in the Hilbert space ${{\cal F}}_{\mathrm s
}({\cal H}_1\oplus {\cal H}_2)$, where ${\cal H}_i$ is the closure of $\operatorname{Im}{\cal K}_i$ in $H_{{\Bbb C}}$, $i=1,2$. Let ${\frak
A}_{\mathrm s}(H_{{\Bbb C}})$ denote the $C^*$-algebra generated by the operators $\varphi(f)$, $f\in H_{{\Bbb C}}$, and let $\omega_{\mathrm s}$ be the state on ${\frak A}_{\mathrm s}(H_{{\Bbb C}})$ defined by $
\omega_{\mathrm s}(\Psi){:=}(\Psi\Omega,\Omega)_{{{\cal F}}_{\mathrm
s}(H\oplus H)}$, $\Psi \in {\frak A}_{\mathrm s}(H_{{\Bbb C}})$. The $n$-point functions of $\omega_{\mathrm s}$ are given by $$\label{gtfr56e54}
\omega_{\mathrm s}(\varphi^*(f_n)\dotsm \varphi^*(f_1)\varphi(g_1)\dotsm
\varphi(g_m))=\delta_{n,m} {\operatorname{per}}((f_i,Kg_j)_{H_{{\Bbb C}}}).$$
We now proceed to consider an infinite free Fermi gas of finite density, which is a special case of representation . Let $H{:=}L^2({{\Bbb R}}^d;dx)=L^2({{\Bbb R}}^d)$, and so $H_{\mathrm{{\Bbb C}}}=L^2_{{{\Bbb C}}}({{\Bbb R}}^d)=L^2({{\Bbb R}}^d\to{{\Bbb C}};dx)$. To fix notations, we define the Fourier transform of a function $f\in
L_{{{\Bbb C}}}^1({{\Bbb R}}^d)$ by $${{\cal F}}f(\lambda){:=}\hat f(\lambda) {:=}
(2\pi)^{-d/2}\int_{{{\Bbb R}}^d} e^{-i x\cdot \lambda} f(x)\,dx,\qquad
\lambda \in{{\Bbb R}}^d,$$ and the inverse Fourier transform by $${{\cal F}}^{-1}f(x){:=}\check f(x){:=}(2\pi)^{-d/2} \int_{{{\Bbb R}}^d}
e^{i\lambda\cdot x}f(\lambda)\,d\lambda,\qquad x\in{{\Bbb R}}^d,$$ so that ${{\cal F}}$ can be extended by continuity from $L_{{{\Bbb C}}}^2({{\Bbb R}}^d)\cap
L_{{{\Bbb C}}}^1({{\Bbb R}}^d)$ to a unitary operator on $L_{{{\Bbb C}}}^2({{\Bbb R}}^d)$, and ${{\cal F}}^{-1}$ is the inverse operator of ${{\cal F}}$.
Let $k$ be the inverse Fourier transform of a function $\hat k$ satisfying the following conditions: $$\label{rde4w3}0\le\hat k\le1,\quad \hat k\in
L^1({{\Bbb R}}^d).$$ We define $K{:=}{{\cal F}}^{-1}\hat k\cdot{{\cal F}}$, where $f\cdot$ denotes the operator of multiplication by a function $f$. Using this $K$, we construct the operators $\psi(f),\psi^*(f)$ defined in the Fock space ${{\cal F}}_{\mathrm
a}(H\oplus H)$ by formula . Notice that now $$H_{1}={{\cal F}}^{-1}L_{{{\Bbb C}}}^2(\operatorname{supp}\hat k;dx),\qquad
H_{2}={{\cal F}}^{-1}L_{{{\Bbb C}}}^2(\operatorname{supp}(1-\hat k);dx).$$
This representation of CAR describes an infinite free Fermi gas with density distribution $\hat k(\cdot)$ in “momentum space” [@AW], see also [@DA]. In particular, if $\beta$ is the inverse temperature, $\mu$ the chemical potential, and $m$ the mass of a particle, the corresponding infinite free Fermi gas is described by $$\label{gzter}\hat
k(\lambda)=\frac{\exp(\beta\mu-\beta\,\frac{|\lambda|^2}{2m})}{1+\exp(\beta\mu-\beta\,\frac{|\lambda|^2}{2m})},\qquad
\lambda\in{{\Bbb R}}^d.$$ For the limit $\beta\to\infty$ of zero temperature, we obtain $$\label{gzftrsres}\hat
k(\lambda)=\pmb1_{B(\sqrt{2m\mu})}(\lambda),\qquad
\lambda\in{{\Bbb R}}^d,$$ where $B(r)$ denotes the ball in ${{\Bbb R}}^d$ of radius $r>0$ centered at the origin, and $\pmb1_X(\cdot)$ denotes the indicator of a set $X$. Notice that, in the case of , we have $H_{1}=H_{2}=H_{{{\Bbb C}}}$, while in the case of $H_{{{\Bbb C}}}=H_{1}\oplus H_{2}$.
We will now need a rigging of ${{\cal F}}_{\mathrm a}(H)$. Let ${\cal
D}({{\Bbb R}}^d)$ denote the space of all real-valued infinite differentiable functions on ${{\Bbb R}}^d$ with compact support. For $p\in{{\Bbb N}}$, we define a weighted Sobolev space $S_p({{\Bbb R}}^d)$ as the closure of ${\cal D }({{\Bbb R}}^d)$ with respect to the Hilbert norm $$\|f\|_p^2{:=}\int_{{{\Bbb R}}^d}A^p f(x)f(x)\,dx,\qquad f\in{\cal
D}({{\Bbb R}}^d),$$ where $$\label{z76t45}Af(x){:=}-\Delta
f(x)+(|x|^2+1)f(x),\qquad x\in{{\Bbb R}}^d,$$ is the harmonic oscillator. We identify $S_0({{\Bbb R}}^d)=L^2({{\Bbb R}}^d)$ with its dual and obtain $$S({{\Bbb R}}^d){:=}\projlim_{p\to\infty}S_p({{\Bbb R}}^d)\subset
L^2({{\Bbb R}}^d)\subset {\operatornamewithlimits{ind\,lim}}_{p\to\infty}S_{-p}({{\Bbb R}}^d){=:} S'({{\Bbb R}}^d).$$ We recall that the Fourier transform ${{\cal F}}$ is a continuous bijection of $S_{{{\Bbb C}}}({{\Bbb R}}^d)$ onto $S_{{{\Bbb C}}}({{\Bbb R}}^d)$, and, extended by continuity, it is a continuous bijection of $S_{{{\Bbb C}}}'({{\Bbb R}}^d)$ onto $S_{{{\Bbb C}}}'({{\Bbb R}}^d)$. Here, $S_{{\Bbb C}}({{\Bbb R}}^d)$ and $S'_{{\Bbb C}}({{\Bbb R}}^d)$ denote the complexification of $S({{\Bbb R}}^d)$ and $S'({{\Bbb R}}^d)$, respectively.
Denoting $\Phi{:=}S_{{{\Bbb C}}}({{\Bbb R}}^d)\oplus S_{{{\Bbb C}}}({{\Bbb R}}^d) $, $\Phi_p{:=}S_{p,{{\Bbb C}}}({{\Bbb R}}^d)\oplus S_{p,{{\Bbb C}}}({{\Bbb R}}^d)$, and $\Phi'{:=}S'_{{{\Bbb C}}}({{\Bbb R}}^d)\oplus S'_{{{\Bbb C}}}({{\Bbb R}}^d)$, we get $
\Phi=\projlim_{p\to\infty}\Phi_p$ and $\Phi'=
{\operatornamewithlimits{ind\,lim}}_{p\to\infty}\Phi_{-p}$. We set, for $n\in{{\Bbb Z}}_+$, $${{\cal F}}_{\mathrm a}^{(n)}(\Phi){:=}\projlim_{p\to\infty}{{\cal F}}_{\mathrm
a}^{(n)}(\Phi_p),\qquad {{\cal F}}_{\mathrm
a}^{(n)}(\Phi'){:=}{\operatornamewithlimits{ind\,lim}}_{p\to\infty}{{\cal F}}^{(n)}_{\mathrm
a}(\Phi_{-p}).$$ Let ${{\cal F}}_{\mathrm a,\, fin} (\Phi)$ denote the topological direct sum of the spaces ${{\cal F}}_{\mathrm
a}^{(n)}(\Phi)$, $n\in{{\Bbb Z}}_+$. The dual of ${{\cal F}}_{\mathrm a,\, fin (\Phi)}$ with respect to the zero space ${{\cal F}}_{\mathrm a}(H\oplus H )$ is ${{\cal F}}_{\mathrm a,\,
fin}^*(\Phi)=\times_{n=0}^\infty {{\cal F}}^{(n)}_{\mathrm a}(\Phi')$, the topological product of the spaces ${{\cal F}}^{(n)}_{\mathrm
a}(\Phi')$. It consists of all sequences of the form $F=(F^{(0)},F^{(1)},F^{(2)},\dots)$ such that $F^{(n)}\in
{{\cal F}}^{(n)}_{\mathrm a}(\Phi')$, and convergence in ${{\cal F}}_{\mathrm
a,\, fin}^*(\Phi)$ means coordinatewise convergence. Thus, we have constructed the nuclear triple $${{\cal F}}_{\mathrm a,\,
fin}(\Phi)\subset {{\cal F}}_{\mathrm a}(H \oplus H)\subset {{\cal F}}_{\mathrm
a,\, fin}^*(\Phi).$$
Noting that $\hat k^{1/2} \in L^2({{\Bbb R}}^d)$ and $(1-\hat k)^{1/2}\in
S'({{\Bbb R}}^d)$, we define $${\varkappa}_1{:=} (2\pi)^{-d/2}{{\cal F}}^{-1}\hat
k^{1/2}\in L_{{\Bbb C}}^2({{\Bbb R}}^d),\qquad
{\varkappa}_2{:=}(2\pi)^{-d/2}{{\cal F}}^{-1}(1-\hat k)^{1/2}\in
S_{{\Bbb C}}'({{\Bbb R}}^d).$$ Then, for any $f\in L_{{\Bbb C}}^2({{\Bbb R}}^d)$, $$K_1f(x)={\varkappa}_1*f(x)=\int_{{{\Bbb R}}^d}{\varkappa}_1(x-y) f(y)\,dy,\qquad
\text{a.e.\ $x\in{{\Bbb R}}^d$},$$ and for any $f\in S_{{\Bbb C}}({{\Bbb R}}^d)$, $K_2f={\varkappa}_2* f$, where the convolution of a generalized function with a test one is defined in the usual way. For each $x\in {{\Bbb R}}^d$, we define ${\varkappa}_{1,x}\in L_{{\Bbb C}}^2({{\Bbb R}}^d)$ and ${\varkappa}_{2,x}\in S_{{\Bbb C}}'({{\Bbb R}}^d)$ by $$\label{new}{\langle}{\varkappa}_{i,x},f{\rangle}={\langle}{\varkappa}_i,f(x+\cdot){\rangle},\qquad f\in
S_{{\Bbb C}}({{\Bbb R}}^d),\ i=1,2,$$ where ${\langle}\cdot,\cdot{\rangle}$ denotes the dual pairing (generated by the scalar product in $H_{{\Bbb C}}$). Then, for any $f\in S_{{\Bbb C}}({{\Bbb R}}^d)$, $$K_i f(x)={\langle}{\varkappa}_{i,x},f{\rangle},\qquad\text{a.e.\
$x\in{{\Bbb R}}^d$},\ i=1,2.\label{sw5e45w}$$
Using formulas , we can easily define, for each $(f_1,f_2)\in\Phi'$, an annihilation operator $a(f_1,f_2)$ acting continuously on ${{\cal F}}_{\mathrm a,\,fin}(\Phi) $, and a creation operator $a^*(f_1,f_2)$ acting continuously on ${{\cal F}}^*_{\mathrm
a,\,fin}(\Phi) $. Analogously to the above, we then introduce operators $a_i(f)$ and $a^*_i(f)$, $i=1,2$, for each $f\in
S_{{\Bbb C}}'({{\Bbb R}}^d)$.
Taking to notice and , we now set, for each $x\in{{\Bbb R}}^d$, $$\psi(x){:=}a_2({\varkappa}_{2,x})+a^*_1(\overline
{\varkappa}_{1,x}),\qquad
\psi^*(x){:=}a^*_2({\varkappa}_{2,x})+a_1(\overline{\varkappa}_{1,x}).$$ These operators act continuously from ${{\cal F}}_{\mathrm a,\,fin}(\Phi)
$ into ${{\cal F}}^*_{\mathrm a,\,fin}(\Phi) $, and we have the following integral representation: for each (real-valued) $f\in S({{\Bbb R}}^d)$ $$\label{3425w4278} \psi(f)=\int_{{{\Bbb R}}^d} dx\, f(x)\psi(x),\qquad
\psi^*(f)=\int_{{{\Bbb R}}^d} dx\, f(x)\psi^*(x).$$ The integration in and below is to be understood in the following sense: for example, the first equality in means: ${\langle}\psi(f)G_1,G_2{\rangle}=\int_{{{\Bbb R}}^d}
f(x){\langle}\psi(x)G_1,G_2{\rangle}\, dx$ for any $G_1,G_2\in {{\cal F}}_{\mathrm
a,\,fin}(\Phi) $. The operators $\psi(x),\psi^*(x)$ satisfy the CAR , the formulas making sense after integration with test functions.
Now, let us briefly consider the bosonic case. Let $H{:=}L^2({{\Bbb R}}^d)$ and let $k$ be the inverse Fourier transform of a function $\hat k$ satisfying the following conditions: $$\label{763457z} 0\le \hat k\le C\quad\text{for some
}C\in(0,\infty),\quad \hat k\in L^1({{\Bbb R}}^d).$$ We define ${\cal K}{:=}{{\cal F}}^{-1}\hat k\cdot{{\cal F}}$, and using this $\cal K$, we construct the operators $\varphi(f),\varphi^*(f)$ defined on the symmetric Fock space ${{\cal F}}_{\mathrm s}(H\oplus H)$ by formula . If we additionally suppose that $ \hat k(x)>0$ a.e. $x\in {{\Bbb R}}^d$, then the obtained representation of CCR describes an infinite free Bose gas at finite temperature and with density distribution $\hat k$ in “momentum space” [@ArWoods], see also [@DA].
Analogously to the above, we construct the triple $${{\cal F}}_{\mathrm
s,\, fin}(\Phi)\subset {{\cal F}}_{\mathrm s}(H \oplus H)\subset
{{\cal F}}_{\mathrm s,\, fin}^*(\Phi),$$ and using it, we make sense of the operators $\varphi(x)$, $\varphi^*(x)$, $x\in{{\Bbb R}}^d$. These satisfy the CCR with $\psi$ replaced by $\varphi$.
Particle density operators and their spectral measure
=====================================================
We will again consider the fermionic case in detail, and then outline the bosonic case.
Fermionic case
--------------
We suppose that holds. For each $x\in{{\Bbb R}}^d$, we define a particle density operator $${\rho_{\mathrm a}}(x){:=}\psi^*(x)\psi(x).$$ Since $\overline{\varkappa}_{1,x}\in
L^2({{\Bbb R}}^d)$, the operator $\psi(x)$ acts continuously from ${{\cal F}}_{\mathrm fin}(\Phi)$ into ${{\cal F}}_{\mathrm fin}(H\oplus H)$, and $\psi^*(x)$ acts continuously from ${{\cal F}}_{\mathrm fin}(H\oplus H)$ into ${{\cal F}}^*_{\mathrm fin}(\Phi)$. Therefore, $\rho_{\mathrm
a}(x)$ is a well-defined, continuous operator from ${{\cal F}}_{\mathrm
fin}(\Phi)$ into ${{\cal F}}^*_{\mathrm fin}(\Phi)$.
We then define $${\rho_{\mathrm a}}(f){:=}\int_{{{\Bbb R}}^d}dx\, f(x){\rho_{\mathrm a}}(x),\qquad
f\in S({{\Bbb R}}^d).$$
\[74dzt\] For each $f\in S({{\Bbb R}}^d)$, the operator ${\rho_{\mathrm a}}(f)$ is well-defined and continuous on${{\cal F}}_{\mathrm
a,\,fin}(H\oplus H) $.
[*Proof*]{}. 1. We first prove the statement for $\rho_{\mathrm a,\,1}(f){:=}\int_{{{\Bbb R}}^d}dx\,
f(x)a^*_2({\varkappa}_{2,x})a^*_1(\overline{\varkappa}_{1,x})$. As easily seen, it suffices to show that $$\label{fze5e}\int_{{{\Bbb R}}^d}dx\,
f(x){\varkappa}_{2,x}\otimes \overline{\varkappa}_{1,x}\in H^{\otimes2}
,$$ where $\otimes$ denotes the usual tensor product. For any $g,h\in S_{{\Bbb C}}({{\Bbb R}}^d)$, we have $$\begin{aligned}
& \left{\langle}\int_{{{\Bbb R}}^d}dx\, f(x){\varkappa}_{2,x}\otimes
\overline{\varkappa}_{1,x},g\otimes h\right{\rangle}\notag\\ &\qquad{:=}
\int _{{{\Bbb R}}^d} f(x){\langle}{\varkappa}_{2,x},g{\rangle}\,{\langle}\overline{\varkappa}_{1,x},h{\rangle}\, dx\notag\\ &\qquad=
\int_{{{\Bbb R}}^d}f(x)(K_2g)(x)(JK_1Jh)(x)\, dx\notag
\\ &\qquad =\int_{{{\Bbb R}}^d}\overline {f(x)(K_1Jh)(x)}\, (K_2g)(x) \,dx\notag\\
& \qquad =\int_{{{\Bbb R}}^d}\overline{({{\cal F}}(f\cdot K_1J h ))(\lambda)}
\,({{\cal F}}K_2g)(\lambda) \,d\lambda\notag\\ &\qquad =\int_{{{\Bbb R}}^d}
(2\pi)^{-d/2} \overline{\hat f * (\hat k^{1/2}\cdot\widehat
Jh)(\lambda)}(1-\hat k)^{1/2}(\lambda)\hat
g(\lambda)\,d\lambda.\notag\\ &\qquad
=\int_{{{\Bbb R}}^d}(2\pi)^{-d/2}\int_{{{\Bbb R}}^d}\overline{\hat
k^{1/2}(\xi)\widehat{Jh}(\xi)\hat f(\lambda-\xi)}\,d\xi\, (1-\hat
k )^{1/2}(\lambda)\hat g(\lambda)\,d\lambda\notag\\
&\qquad=\int_{{{\Bbb R}}^{2d}}(2\pi)^{-d/2} \overline{\hat f(\lambda+\xi)}
(1-\hat k)^{1/2}(\lambda)\hat k^{1/2}(-\xi) \hat g(\lambda)\hat h
(\xi)\,d\lambda\,d\xi.\label{weqew3}\end{aligned}$$ Since $|1-\hat k|\le 1$ and $\hat k^{1/2},\hat f\in L^2({{\Bbb R}}^d)$, the function $$G_f(\lambda,\xi){:=} (2\pi)^{-d/2} \hat
f(\lambda+\xi) (1-\hat k)^{1/2}(\lambda)\hat k^{1/2}(-\xi),\qquad
\xi,\lambda\in{{\Bbb R}}^d,$$ belongs to $L^2({{\Bbb R}}^{2d})$. Therefore, by , $$\left{\langle}\int_{{{\Bbb R}}^d}dx\, f(x){\varkappa}_{2,x}\otimes
\overline{\varkappa}_{1,x},g\otimes h\right{\rangle}= {\langle}{{\cal F}}_{2d}^{-1}(G_f),g\otimes h{\rangle},\qquad g,h\in S({{\Bbb R}}^d),$$ where ${{\cal F}}_{2d}$ denotes the Fourier transform on $L_{{{\Bbb C}}}^2({{\Bbb R}}^{2d})$. By linearity and continuity, this implies $$\int_{{{\Bbb R}}^d}dx\,
f(x){\varkappa}_{2,x}\otimes
\overline{\varkappa}_{1,x}={{\cal F}}_{2d}^{-1}(G_f)\in L^2({{\Bbb R}}^d)^{\otimes 2}
.$$
2\. We now prove the statement for $$\rho_{\mathrm a,\,
2}(f){:=}\int_{{{\Bbb R}}^d}dx\,
f(x)a^*_2({\varkappa}_{2,x})a_2({\varkappa}_{2,x}).$$ For any $g_i,h_i\in
S_{{{\Bbb C}}}({{\Bbb R}}^d)$, $i=1,2$, we have $$\begin{aligned}
&\left{\langle}\int_{{{\Bbb R}}^d} dx\, f(x)a^*_2({\varkappa}_{2,x})
a_2({\varkappa}_{2,x})(g_1,g_2),(h_1,h_2)\right{\rangle}\\ &\qquad=
\int_{{{\Bbb R}}^d}f(x)\overline{K_2g_2(x)}\, K_2h_2(x)= {\langle}K_2(f\cdot
K_2g_2),h_2{\rangle},\end{aligned}$$ and therefore $$\int_{{{\Bbb R}}^d}
dx\,a^*_2({\varkappa}_{2,x})
a_2({\varkappa}_{2,x})\restriction{{\cal F}}^{(1)}(\Phi)=\pmb0\oplus
\big(K_2(f\cdot K_2)\big){=:}{\cal A}_{2,f}.$$ Evidently ${\cal
A}_{2,f} $ is continuous on $H_{{{\Bbb C}}}\oplus H_{{{\Bbb C}}}$.
For any linear continuous operator $\cal A$ on $H_{{{\Bbb C}}}\oplus
H_{{{\Bbb C}}}$, we define the second quantization of $\cal A$, denoted by $d\Gamma({\cal A })$, as the operator in ${{\cal F}}_{\mathrm a}(H\oplus
H )$ with domain $D(d\Gamma({\cal A})){:=}{{\cal F}}_{\mathrm a,\,
fin}(H\oplus H)$, given by $$d\Gamma({\cal A})\restriction
{{\cal F}}_{\mathrm a}^{(n)}(H\oplus H){:=}{\cal
A}\otimes\pmb1\otimes\dots\otimes\pmb1+\pmb1\otimes{\cal
A}\otimes\pmb1\otimes\dots\otimes\pmb1+\pmb1\otimes \dots\otimes
\pmb1\otimes{\cal A}.$$ $d\Gamma({\cal A})$ acts continuously on ${{\cal F}}_{\mathrm a,\, fin}(H\oplus H)$. Then, an easy calculation shows that $\rho_{\mathrm a,\,2 }(f)=d\Gamma({\cal A}_{2,f})$.
3\. Analogously, we get $$\begin{aligned}
\rho_{\mathrm a,\, 3
}(f){:=}&\int_{{{\Bbb R}}^d}dx\, f(x)
a_1(\overline{\varkappa}_{1,x})a^*_1(\overline{\varkappa}_{1,x})\\ =&
\int_{{{\Bbb R}}^d} f(x){\langle}\overline{\varkappa}_{1,x},\overline{\varkappa}_{1,x}{\rangle}\, dx\, \pmb1-d\Gamma({\cal A}_{1,f})\\=&\int_{{{\Bbb R}}^d}f(x)\,
dx\,(2\pi)^{-d}\int_{{{\Bbb R}}^d}\hat k(\lambda)\,d\lambda\,
\pmb1-d\Gamma({\cal A}_{1,f}),\end{aligned}$$ where ${\cal
A}_{1,f}{:=}\big(JK_1J(f\cdot JK_1J)\big)\oplus\pmb0$.
4\. Finally, since $\rho_{\mathrm a,1}(f)$ is a continuous operator from ${{\cal F}}_{\mathrm a}^{(n)}(H\oplus H)$ into ${{\cal F}}_{\mathrm
a}^{(n+2)}(H\oplus H)$ for each $n\in{{\Bbb Z}}_+$, and since $$\rho_{\mathrm a,\, 4}(f){:=}\int_{{{\Bbb R}}^d}dx\,
f(x)a_1(\overline{\varkappa}_{1,x})a_2({\varkappa}_{2,x})$$ is its adjoint, we have that $\rho_{\mathrm a,\, 4}(f)$ acts continuously from ${{\cal F}}_{\mathrm a}^{(n+2)}(H\oplus H)$ into ${{\cal F}}_{\mathrm
a}^{(n)}(H\oplus H)$, and hence, continuously on ${{\cal F}}_{\mathrm
a,\, fin}(H\oplus H)$.$\blacksquare$
Using anticommutation relations , we easily prove the following
For each $f_1,f_2\in S({{\Bbb R}}^d)$, we have on ${{\cal F}}_{\mathrm a,\, fin}(H\oplus
H)$: $$\label{hjfcz}{\rho_{\mathrm a}}(f_1){\rho_{\mathrm a}}(f_2)={\rho_{\mathrm a}}(f_2){\rho_{\mathrm a}}(f_1).$$
We define a Hilbert space ${\frak H}_{\mathrm a}$ as the closure of the linear span of the set $$\big\{\, \Omega, {\rho_{\mathrm a}}(f_1)\dotsm{\rho_{\mathrm a}}(f_n)\Omega\mid
f_1,\dots,f_n\in S({{\Bbb R}}^d),\ n\in{{\Bbb N}}\,\}$$ in ${{\cal F}}_{\mathrm
a}(H\oplus H)$.
[[It is not hard to see that ${\frak H}_{\mathrm a}$ is a subspace of the space $$\label{guhdesrw}\bigoplus_{n=0}^\infty
P_{{\mathrm a},\, 2n}\big(H_{1,\,{{\Bbb C}}}^{\otimes n }\otimes
H_{2,\,{{\Bbb C}}}^{\otimes n}\big),$$ where $P_{{\mathrm
a},\, 2n}:(H_{{{\Bbb C}}}\oplus H_{{{\Bbb C}}})^{\otimes 2n}\to (H_{{{\Bbb C}}}\oplus
H_{{{\Bbb C}}})^{\wedge 2n}$ is the antisymmetrization operator. Evidently, is a subspace of ${{\cal F}}_{\mathrm
a}(H_1\oplus H_2)$. Whether ${\frak H}_{\mathrm a}$ coincides with or it is a proper subspace of it, is an open problem (see also Remark \[zurftt\] below). ]{}]{}
We next define ${\frak H}_{\mathrm a,\,fin}{:=}{\frak H}_{\mathrm
a}\cap {{\cal F}}_{\mathrm a,\,fin}(H\oplus H)$, ${\frak H}_{\mathrm
a,\,fin}$ being dense in ${\frak H}_{\mathrm a}$. Let us consider the ${\rho_{\mathrm a}}(f)$’s as operators in ${\frak H}_{\mathrm a}$ with domain ${\frak H}_{\mathrm a,\, fin}$.
\[ugtder\] The operators ${\rho_{\mathrm a}}(f)$, $f\in
S({{\Bbb R}}^d)$, are essentially selfadjoint in ${\frak H}_{\mathrm a
}$.
[*Proof*]{}. The operators are evidently symmetric. The proof of essential selfadjointness is quite standard (see e.g.[@BeKo Ch. 3, subsec. 3.8] and [@Ly Lemma 4.1]), so we only outline it.
As easily seen from the proof of Lemma \[74dzt\], we have for any $g^{(n)}\in {{\cal F}}_{\mathrm a}^{(n)}(H\oplus H)$ $${\rho_{\mathrm a}}(f)g^{(n)}=\sum_{j=1}^4 \rho_{{\mathrm a},\, j}(f)g^{(n)}\in
\bigoplus_{i=n-2,\,n,\,n+2}{{\cal F}}_{\mathrm a}^{(i)}(H\oplus H),$$ and moreover $$\label{zue53398}\|\rho_{{\mathrm a},\,j}(f)g^{(n)}\|_{{{\cal F}}_{\mathrm a}(H\oplus
H)}\le C_1
\max\{\|f\|_{L^2({{\Bbb R}}^d)},\|f\|_{L^1({{\Bbb R}}^d)},\|f\|_{L^\infty({{\Bbb R}}^d)}\}\,\|g^{(n)}\|_{{{\cal F}}_{\mathrm
a }^{(n)}(H\oplus H)}$$ for $j=1,\dots,4$ and some $C_1>0$. From here, it is not hard to show that, for every $g^{(n)}\in {{\cal F}}_{\mathrm a}^{(n)}(H\oplus H)$, the series $$\sum_{m=0}^\infty \frac {\|\rho_{{\mathrm a}}(f)^m
g^{(n)}\|_{{{\cal F}}_{\mathrm a}(H\oplus H)}}{m!}\, t^m$$ converges for $$0<t<\big(4C_1
\max\{\|f\|_{L^2({{\Bbb R}}^d)},\|f\|_{L^1({{\Bbb R}}^d)},\|f\|_{L^\infty({{\Bbb R}}^d)}\}\big)^{-1}.$$ Therefore, any vector from ${\frak H}_{\mathrm a,\,fin}$ is analytical for ${\rho_{\mathrm a}}(f)$. By Nelson’s analytic vector criterium (e.g. [@RS2 Th. X.39]), the lemma follows.$\blacksquare$
We denote by ${\rho_{\mathrm a}}^\sim(f)$ the closure of ${\rho_{\mathrm a}}(f)$ in ${\frak
H}_{\mathrm a}$, which is a selfadjoint operator by Lemma \[ugtder\].
\[he54w\] For any $f_1,f_2\in S({{\Bbb R}}^d)$, the operators $\rho_{\mathrm a}^\sim(f_1)$ and $\rho_{\mathrm a
}^\sim(f_2)$ commute in the sense of their resolutions of the identity.
[*Proof*]{}. Since ${\rho_{\mathrm a}}(f_1)$ is essentially selfadjoint, the set $({\rho_{\mathrm a}}(f_1)+i\pmb1){\frak
H}_{\mathrm a,\, fin}$ is dense in ${\frak H}_{\mathrm a}$. Furthermore, $({\rho_{\mathrm a}}(f_1)+i\pmb1){\frak H}_{\mathrm a,\,
fin}\subset {\frak H}_{\mathrm a,\, fin}$. Thus, by the proof of Lemma \[ugtder\], the operator ${\rho_{\mathrm a}}(f_1)\restriction({\rho_{\mathrm a}}(f_1)+i\pmb1){\frak H}_{\mathrm a,\,
fin}$ has a dense set of analytical vectors. Hence, the lemma follows from [@BeKo Ch. 5, Th. 1.15].$\blacksquare$
\[e5q2343\] Let $k$ be the inverse Fourier transform of a function $\hat k$ satisfying . Let the Hilbert space ${\frak H}_{\mathrm a }$ and the operators ${\rho_{\mathrm a}}^\sim(f)$, $f\in S({{\Bbb R}}^d)$, be defined as above. Then, there exist a unique probability measure $\mu_{\mathrm
a}$ on $(S'({{\Bbb R}}^d),{\cal B}(S'({{\Bbb R}}^d)))$ (${\cal B}(S'({{\Bbb R}}^d))$ denoting the Borel $\sigma$-algebra on $S'({{\Bbb R}}^d)$) and a unique unitary operator $I_{\mathrm a}:{\frak H}_{\mathrm a}\to
L^2(S'({{\Bbb R}}^d),{\cal B}(S'({{\Bbb R}}^d));\mu_{\mathrm a})$ such that $I_{\mathrm a}\Omega=1$ and the following formula holds $$\label{ttre} I_{\mathrm a}\,{\rho_{\mathrm a}}^\sim(f) I_{\mathrm a}^{-1}={\langle}\cdot,
f{\rangle}\cdot,\qquad f\in S({{\Bbb R}}^d).$$
[[In terms of the spectral theory of commuting selfadjoint operators (e.g. [@BeKo; @Sam]), Theorem \[e5q2343\] states that the family $({\rho_{\mathrm a}}^\sim(f))_{f\in S({{\Bbb R}}^d)}$ has a spectral measure $\mu_{\mathrm a}$ on $(S'({{\Bbb R}}^d),{\cal B}(S'({{\Bbb R}}^d)))$. Furthermore, since the operators $\rho_{\mathrm a}(f)$ have a Jacobi type form in ${\cal F}_{\mathrm a}(H_1\oplus H_2)$, this result is close in spirit to [@bere; @Ly].]{}]{}
[*Proof of Theorem*]{} \[e5q2343\]. Let $(h_k)_{k=0}^\infty$ be the sequence of Hermite functions forming an orthonormal basis in $L^2({{\Bbb R}}^d)$ and let $a_k>0$ be the eigenvalue of the operator $A$ (defined by ) belonging to the eigenvector $h_k$, $k\in{{\Bbb Z}}_+$.
We denote by ${{\Bbb R}}^\infty{:=}{{\Bbb R}}^{{{\Bbb Z}}_+}$ the space of all sequences of the form ${\bf x}=(x_0,x_1,x_2,\dots)$, $x_k\in{{\Bbb R}}$, $k\in{{\Bbb Z}}_+$, and we endow ${{\Bbb R}}^\infty$ with the product topology. The Borel $\sigma$-algebra ${\cal B}({{\Bbb R}}^\infty)$ coincides with the cylinder $\sigma$-algebra ${\cal C}_\sigma({{\Bbb R}}^\infty)$.
\[rtw342\] There exist a unique probability measure $\tilde\mu_{\mathrm a}$ on $({{\Bbb R}}^\infty,{\cal
B}({{\Bbb R}}^\infty))$ and a unique unitary operator $\tilde I_{\mathrm a
}: {\frak H}_{\mathrm a}\to L^2({{\Bbb R}}^\infty,{\cal B
}({{\Bbb R}}^\infty);\tilde\mu_{\mathrm a})$ such that $\tilde I_{\mathrm
a }\Omega=1$ and, for each $k\in{{\Bbb Z}}_+$, $\tilde I_{\mathrm
a }\,{\rho_{\mathrm a}}^{\sim}(h_k)\tilde I^{-1}_{\mathrm a}=x_k\cdot$, where $x_k\cdot$ denotes the operator of multiplication by $x_k$.
[*Proof*]{}. For $f\in S({{\Bbb R}}^d)$, we have $f=\sum_{k=0}^\infty \langle f,h_k\rangle h_k$, where the series converges in each space $S_p({{\Bbb R}}^d)$, $p\in{{\Bbb N}}$, and hence in $S({{\Bbb R}}^d)$. Next, it follows from that, for each fixed $G\in{\frak H}_{\mathrm a,\, fin}$, the mapping $$\label{4896743w} S({{\Bbb R}}^d)\ni f\mapsto {\rho_{\mathrm a}}(f)G\in{\frak
H}_{\mathrm a}$$ is continuous. Therefore, $\Omega$ is a cyclic vector for the family $({\rho_{\mathrm a}}^\sim(h_k))_{k=0}^\infty$. Thus, $({\rho_{\mathrm a}}^\sim(h_k))_{k=0}^\infty$ is a countable family of commuting selfadjoint operators having a cyclic vector, and hence the lemma follows from [@Sam Ch. 1, Th. 4].$\blacksquare$
For each $p\in{{\Bbb N}}$, we define the following measurable function on ${{\Bbb R}}^\infty$: $${{\Bbb R}}^\infty\ni{\bf x}=(x_k)_{k=0}^\infty\mapsto
\|{\bf x}\|_{-p}^2{:=}\sum_{k=0}^\infty
x_k^2a_k^{-p}\in{{\Bbb R}}_+\cup\{+\infty\}.$$ Let $${\cal
S}_{-p}{:=}\{{\bf x}\in{{\Bbb R}}^\infty:\|x\|_{-p}^2<\infty\},\quad
p\in{{\Bbb N}},\qquad {\cal S}'{:=}\bigcup_{p\in{{\Bbb N}}}{\cal S}_{-p}.$$ Evidently, ${\cal S}_{-p},\,{\cal S}'\in{\cal B}({{\Bbb R}}^\infty)$. By using the monotone convergence theorem and Lemma \[rtw342\], we get $$\begin{aligned}
\int_{{{\Bbb R}}^\infty}\|{\bf
x}\|_{-p}^2\,d\tilde\mu_{\mathrm a
}(x)&=\int_{{{\Bbb R}}^\infty}\sum_{k=0}^\infty x_k^2 a_k^{-p}\,
d\tilde\mu_{\mathrm a}(x) \notag\\ &=\sum_{k=0}^\infty
a_k^{-p}\int_{{{\Bbb R}}^\infty} x_k^2\,d\tilde\mu_{\mathrm a}(x)=\sum_{k=0} ^\infty a_k^{-p}\|\psi(h_k)\Omega\|_{{\frak
H}_{\mathrm a}}^2.\label{zur6z}\end{aligned}$$ For some $C_2>0$, $$\label{rtw348975}\max\{\|f\|_{L^2({{\Bbb R}}^d)},\|f\|_{L^1({{\Bbb R}}^d)},\|f\|_{L^\infty({{\Bbb R}}^d)}\}\le
C_2\|f\|_{S_d({{\Bbb R}}^d)},\qquad f\in S({{\Bbb R}}^d),$$ and since the inclusion $S_d({{\Bbb R}}^d)\hookrightarrow L^2({{\Bbb R}}^d)$ is of Hilbert–Schmidt type, $$\label{urt67}\sum_{k=0}^\infty a_k^{-d}<\infty.$$ By , –, $$\int_{{{\Bbb R}}^\infty}\|x\|_{-2d}^2\, d\tilde\mu_{\mathrm a}(x)\le C_1^2
C_2^2\sum_{k=0}^\infty a_k^{-d}<\infty.$$ This yields that $$\label{zr564}\tilde\mu_{\mathrm a}({\cal S}_{-2d})=\tilde\mu_{\mathrm a}({\cal S}')=1.$$ Let ${\cal B}({\cal S}')$ denote the trace $\sigma$-algebra of ${\cal B}({{\Bbb R}}^\infty)$ on ${\cal S}'$. By , we can consider $\tilde\mu_{\mathrm a}$ as a probability measure on $({\cal S}',{\cal B}({\cal S}'))$.
Noticing that the mapping $${\cal S}'\ni{\bf
x}=(x_0,x_1,x_2,\dots)\mapsto {\cal E}{\bf x}{:=}\sum_{k=0}^\infty
x_kh_k \in S'({{\Bbb R}}^d)$$ is a measurable bijection, we define a probability measure $\mu_{\mathrm a}$ on $(S'({{\Bbb R}}^d),{\cal
B}(S'({{\Bbb R}}^d)))$ by $\mu_{\mathrm a}{:=}\tilde \mu_{\mathrm
a}\circ{\cal E}^{-1}$, and a unitary operator ${\cal U}:L^2({\cal
S}',{\cal B}({\cal S}');\tilde\mu_{\mathrm a})\to
L^2(S'({{\Bbb R}}^d),{\cal B}(S'({{\Bbb R}}^d));\mu_{\mathrm a})$ by $${\cal
U}F(\omega){:=}F({\cal E}^{-1}\omega),\qquad \omega\in S'({{\Bbb R}}^d).$$ Setting $I_{\mathrm a}{:=}{\cal U}\tilde I_{\mathrm a}$, we get a unitary operator acting from ${\frak H}_{\mathrm a}$ onto $L^2(S'({{\Bbb R}}^d);\mu_{\mathrm a})$ such that $I_{\mathrm a}\Omega=1$ and $$\label{hjre54w}I_{\mathrm a}\,{\rho_{\mathrm a}}^\sim(h_k)
I_{\mathrm a}^{-1}={\langle}\cdot,h_k{\rangle}\cdot,\qquad
k\in{{\Bbb Z}}_+.$$ Furthermore, using the continuity of mapping , we easily conclude from that holds. Thus, the theorem is proved.$\blacksquare$
The configuration space $\Gamma_{{{\Bbb R}}^d}$ over ${{\Bbb R}}^d$ is defined as the set of all locally finite subsets (configurations) in ${{\Bbb R}}^d$: $$\Gamma_{{{\Bbb R}}^d}{:=}\big\{\,\gamma\subset{{\Bbb R}}^d\mid
|\gamma\cap\Lambda|<\infty\text{ for each compact
$\Lambda\subset{{\Bbb R}}^d$}\,\big\}.$$ Here, $|\Lambda|$ denotes the cardinality of a set $\Lambda$. We can identify any $\gamma\in\Gamma_{{{\Bbb R}}^d}$ with the positive Radon measure $\sum_{x\in\gamma}\delta_x\in{\cal M}({{\Bbb R}}^d)$, where $\delta_x$ is the Dirac measure with mass at $x$, $\sum_{x\in\varnothing}\delta_x{:=}$zero measure, and ${\cal
M}({{\Bbb R}}^d)$ stands for the set of all positive Radon measures on ${\cal B}({{\Bbb R}}^d)$. The space $\Gamma_ {{{\Bbb R}}^d}$ is endowed with the relative topology as a subset of the space ${\cal M}(X)$ with the vague topology. We denote by ${\cal B}(\Gamma_X)$ the Borel $\sigma$-algebra on $\Gamma_{{{\Bbb R}}^d}$.
We endow ${\cal D}({{\Bbb R}}^d)$ with its natural projective limit topology and denote by ${\cal D}'({{\Bbb R}}^d)$ the dual space of ${\cal
D}({{\Bbb R}}^d)$. One can show that $\Gamma_{{{\Bbb R}}^d}$ belongs to the cylinder $\sigma$-algebra ${\cal C}_\sigma({\cal D}'({{\Bbb R}}^d))$, and furthermore, the trace $\sigma$-algebra of ${\cal C}_\sigma({\cal
D}'({{\Bbb R}}^d))$ on $\Gamma_{{{\Bbb R}}^d}$, resp. $S'({{\Bbb R}}^d)$, coincides with ${\cal B}(\Gamma_X)$, resp. ${\cal B}(S'({{\Bbb R}}^d))$. Thus, any probability measure $\nu$ on $(S'({{\Bbb R}}^d),{\cal B}(S'({{\Bbb R}}^d)))$ can be considered as a measure on $({\cal D}'({{\Bbb R}}^d),{\cal
C}_\sigma({\cal D}'({{\Bbb R}}^d)))$, and if additionally $\nu(\Gamma_{{{\Bbb R}}^d})=1$, $\nu$ can be considered as probability measure on $(\Gamma_{{{\Bbb R}}^d},{\cal B}(\Gamma_{{{\Bbb R}}^d}))$ as well.
Our next aim is to show that $\mu_{\mathrm a}$ is supported by $\Gamma_{{{\Bbb R}}^d}$. To this end, let us recall the notion of correlation functions of a probability measure $\nu$ on $(\Gamma_{{{\Bbb R}}^d},{\cal B}(\Gamma_{{{\Bbb R}}^d}))$.
Let ${\hat\otimes}$ stand for the symmetric tensor product. For any $g^{(n)}\in{\cal D}({{\Bbb R}}^d)^{{\hat\otimes}n}$($=$the space of all smooth, symmetric, compactly supported functions on $({{\Bbb R}}^d)^n$), we define a function $\Gamma_{{{\Bbb R}}^d}\ni\gamma\mapsto{\langle}{:}\gamma^{\otimes
n}{:},g^{(n)}{\rangle}\in{{\Bbb R}}$ by $$\label{zudsfu}
{\langle}{:}\gamma^{\otimes n}{:},g^{(n)}{\rangle}{=}
\sum_{x_1\in\gamma}\,\sum_{x_2\in\gamma,\, x_2\ne x_1}\dots
\sum_{x_n\in\gamma,\, x_n\ne x_1,\dots,x_n\ne
x_{n-1}}g^{(n)}(x_1,\dots,x_n)$$ (the number of the non-zero summands on the right hand side of is finite). The functions $(k_\nu^{(n)})_{n=1}^\infty$ with $k_\nu^{(n)}:({{\Bbb R}}^d)^n\to{{\Bbb R}}$ being measurable and symmetric, are called correlation functions of the measure $\nu$ if, for each $g^{(n)}\in{\cal D}({{\Bbb R}}^d)^{{\hat\otimes}n}$, $n\in{{\Bbb N}}$, $$\label{hgfztdzzttz}
\int_{\Gamma_{{{\Bbb R}}^d}} {\langle}{:}\gamma^{\otimes
n}{:},g^{(n)}{\rangle}\,\nu(d\gamma)=\int_{({{\Bbb R}}^d)^n}
g^{(n)}(x_1,\dots,x_n) k_{\nu}^{(n)}(x_1,\dots,x_n)\, dx_1\dotsm
dx_n$$ (if the measure $\nu$ has correlation functions, then these are a.s. uniquely defined).
As easily seen from , the kernels ${:}\gamma^{\otimes n}{:}\in{\cal D}'({{\Bbb R}}^d)^{{\hat\otimes}n}$ satisfy the recursion relation $$\begin{gathered}
{:}\gamma^{\otimes 1}(x){:}= \gamma(x),\notag\\ {:}\gamma^{\otimes
(n+1)}(x_1,\dots,x_{n+1}){:}=\big(\gamma(x_{n+1})\,{:}\gamma^{\otimes
n}(x_1,\dots,x_n){:}\notag\\\text{} -\sum_{i=1}^n
\delta(x_{n+1}-x_i)\,{:}\gamma^{\otimes
n}(x_1,\dots,x_n){:}\big)^\sim,\qquad n\in{{\Bbb N}},\label{hfztfdzw}\end{gathered}$$ where $(\cdot)^\sim$ denotes symmetrization of a function. Replacing $\gamma\in\Gamma_{{{\Bbb R}}^d}$ with an arbitrary $\omega\in{\cal D}'({{\Bbb R}}^d)$, we may now define ${:}\omega^{\otimes n}{:}\in{\cal D}'({{\Bbb R}}^d)^{\hat\otimes n}$ and introduce, analogously to , the notion of correlation functions $(k_\nu^{(n)})_{n=1}^\infty$ for any probability measure $\nu$ on ${\cal D}'({{\Bbb R}}^d)$. (We, however, remark that the introduction of correlation functions for a measure on ${\cal D}'({{\Bbb R}}^ d)$ is only of “technical” nature, since one always expects that a measure having correlation functions is supported by $\Gamma_{{{\Bbb R}}^d}$, see the arguments below).
Following , we introduce operators $$\begin{gathered}
{{:}\,}{\rho_{\mathrm a}}(x){\,{:}}={\rho_{\mathrm a}}(x),\notag\\
{{:}\,}{\rho_{\mathrm a}}(x_{n+1}){\rho_{\mathrm a}}(x_n)\dotsm
{\rho_{\mathrm a}}(x_1){\,{:}}=\big({\rho_{\mathrm a}}(x_{n+1})\,{{:}\,}{\rho_{\mathrm a}}(x_1)\dotsm{\rho_{\mathrm a}}(x_n){\,{:}}\notag\\ \text{}
-\sum_{i=1}^{n}\delta(x_{n+1}-x_i)\,{{:}\,}{\rho_{\mathrm a}}(x_1)\dotsm
{\rho_{\mathrm a}}(x_n){\,{:}}\big)^\sim,\label{zur65e}\end{gathered}$$ which make sense after integration with test functions.
The following proposition shows that ${{:}\,}{\rho_{\mathrm a}}(x_n)\dotsm{\rho_{\mathrm a}}(x_1){\,{:}}$ is the “normal product” of the the operators ${\rho_{\mathrm a}}(x_1),\dots,{\rho_{\mathrm a}}(x_n)$ (compare with [@MSh subsecs. 2.B and 2.C]).
\[higztf\] For each $n\in{{\Bbb N}}$ and $f_1,\dots,f_n\in S({{\Bbb R}}^d)$, we have on ${\cal F}_{\mathrm a,\, fin}(H\oplus H)$: $$\begin{gathered}
\label{teeszr6} \int_{({{\Bbb R}}^d)^n}dx_1\dotsm dx_n\,
f_1(x_1)\dotsm f_n(x_n) \,{:}{\rho_{\mathrm a}}(x_1)\dotsm{\rho_{\mathrm a}}(x_n){\,{:}}\\
=\int_{({{\Bbb R}}^d)^n} dx_1\dotsm
dx_n\,f_1(x_1)\dotsm f_n(x_n)\,\psi^*(x_n)\dotsm\psi^*(x_1)\psi(x_1)\dotsm\psi(x_n).\end{gathered}$$
[*Proof*]{}. We first note that, for $n\ge2$, the well-definedness of the operator on the right hand side of on ${\cal F}_{\mathrm a,\, fin}(H\oplus H)$ may be proved by using arguments analogous to those as in the proof of Lemma \[74dzt\]. We prove the proposition by induction. For $n=1$, is trivially satisfied. Suppose that holds for some $n\in{{\Bbb N}}$. Then, by the induction hypothesis, , and , we have $$\begin{gathered}
{{:}\,}{\rho_{\mathrm a}}(x_{n+1})\dotsm{\rho_{\mathrm a}}(x_1){\,{:}}=\big(\psi^*(x_{n+1})\psi(x_{n+1})\psi^*(x_n)\dotsm\psi^*(x_1)\psi(x_1)\dotsm\psi(x_n)\\
\text{} -\sum_{i=1}^n
\delta(x_{n+1}-x_i)\psi^*(x_n)\dotsm\psi^*(x_1)\psi(x_1)\dotsm\psi(x_n)\big)^\sim\\
=\big(-\psi^*(x_{n+1})\psi^*(x_n)\psi(x_{n+1})\psi^*(x_{n-1})\dotsm
\psi^*(x_1)\psi(x_1)\dotsm \psi(x_n)\\ \text{}
-\sum_{i=1}^{n-1}\delta(x_{n+1}-x_i)\psi^*(x_n)\dotsm\psi^*(x_1)\psi(x_1)\dotsm\psi(x_n)\big)^\sim\\
=\big(\psi^*(x_{n+1})\psi^*(x_n)\psi^*(x_{n-1})\psi(x_{n+1})\psi^*(x_{n-2})\dotsm\psi^*(x_1)\psi(x_1)\dots\psi(x_n)\\
\text{}-\delta(x_{n+1}-x_{n-1})\psi(x_{n-1})^*\psi(x_n)^*\psi(x_{n-2})^*\dotsm
\psi^*(x_1)\psi(x_1)\dotsm \psi(x_n)\\ \text{}
-\sum_{i=1}^{n-1}\delta(x_{n+1}-x_i)\psi^*(x_n)\dotsm\psi^*(x_1)\psi(x_1)\dotsm\psi(x_n)\big)^\sim\\
=\big(\psi^*(x_{n+1})\psi^*(x_n)\psi^*(x_{n-1})\psi(x_{n+1})\psi^*(x_{n-2})\dotsm\psi^*(x_1)\psi(x_1)\dots\psi(x_n)\\
-\sum_{i=1}^{n-2}\delta(x_{n+1}-x_i)\psi^*(x_n)\dotsm\psi^*(x_1)\psi(x_1)\dotsm\psi(x_n)\big)^\sim\\
=\dots=\big((-1)^n\psi^*(x_{n+1})\dotsm\psi^*(x_1)\psi(x_{n+1})\psi(x_1)\dotsm\psi(x_n)\big)^\sim\\
=\psi^*(x_{n+1})\dotsm\psi^*(x_1)\psi(x_1)\dotsm\psi(x_{n+1}),\end{gathered}$$ the formulas above making sense after integration with test functions. $\blacksquare$
\[jkgztd\] For any $f_1,\dots,f_n\in S({{\Bbb R}}^d)$, $n\in{{\Bbb N}}$, $$\begin{gathered}
\label{z8t54}\bigg(
\int_{({{\Bbb R}}^d)^n}dx_1\dots dx_n\, f_1(x_1)\dotsm f_n(x_n) {{:}\,}{\rho_{\mathrm a}}(x_1)\dotsm{\rho_{\mathrm a}}(x_n){\,{:}}\,\Omega,\Omega\bigg)_{{\frak
H}_{\mathrm a}}\\=\int_{({{\Bbb R}}^d)^n}(f_1{\hat\otimes}\dotsm{\hat\otimes}f_n)(x_1,\dots,x_n)\det({\varkappa}(x_i-x_j))_{i,j=1}^n\,dx_1\dotsm
dx_n,\end{gathered}$$where ${\varkappa}(x){:=}(2\pi)^{-d/2} k(x)$, $x\in{{\Bbb R}}^d$.
[*Proof*]{}. By Proposition \[higztf\], $$\begin{gathered}
\bigg( \int_{({{\Bbb R}}^d)^n}dx_1\dotsm dx_n\,
f_1(x_1)\dotsm f_n(x_n) {{:}\,}{\rho_{\mathrm a}}(x_1)\dotsm{\rho_{\mathrm a}}(x_n){\,{:}}\,\Omega,\Omega\bigg)_{{\frak
H}_{\mathrm a}}\notag\\ =\int_{({{\Bbb R}}^d)^n}(f_1{\hat\otimes}\dotsm{\hat\otimes}f_n)(x_1,\dots,x_n)\big( a_1(\overline\varkappa_{1,x_n})\dotsm
a_1(\overline\varkappa_{1,x_1})a_1^*(\overline\varkappa_{1,x_1})\dotsm
a^*_1(\overline\varkappa_{1,x_n})\Omega,\Omega\big)_{{\frak
H}_{\mathrm a}}\, dx_1\dotsm dx_n \notag\\ =
\int_{({{\Bbb R}}^d)^n}(f_1{\hat\otimes}\dotsm{\hat\otimes}f_n)(x_1,\dots,x_n)\sqrt{n!}\,\big(
a_1(\overline\varkappa_{1,x_n})\dotsm
a_1(\overline\varkappa_{1,x_1})
\overline{\varkappa}_{1,x_1}\wedge\dotsm\wedge
{\varkappa}_{1,x_n},\Omega\big)_{{\frak H}_{\mathrm a}}\, dx_1\dotsm
dx_n \notag\\ = \int_{({{\Bbb R}}^d)^n}(f_1{\hat\otimes}\dotsm{\hat\otimes}f_n)(x_1,\dots,x_n)\,n!\,\big( \overline{\varkappa}_{1,x_1}\otimes
\dots \otimes\overline{\varkappa}_{1,x_n},
\overline{\varkappa}_{1,x_1}\wedge\dotsm\wedge\overline{\varkappa}_{1,x_n}\big)_{{\frak
H}_{\mathrm a}}\, dx_1\dotsm dx_n \notag\\ =
\int_{({{\Bbb R}}^d)^n}(f_1{\hat\otimes}\dotsm{\hat\otimes}f_n)(x_1,\dots,x_n) \det\big((
\overline{\varkappa}_{1,x_i},\overline{\varkappa}_{1,x_j})_{H_{{\Bbb C}}}\big)_{i,j=1}^n\,
dx_1\dotsm dx_n .\label{56rtzhb}\end{gathered}$$ Next, for any $f,g\in
S_{{\Bbb C}}({{\Bbb R}}^d)$, $$\begin{gathered}
\int_{({{\Bbb R}}^d)^2}(
\overline{\varkappa}_{1,x},\overline{\varkappa}_{1,y})_{H_{{\Bbb C}}}\,
f(x)\overline g(y)\, dx\, dy =
\int_{({{\Bbb R}}^d)^3}{\varkappa}_1(x-z)\overline{\varkappa}_1(y-z)f(x)\overline
g(y)\, dx\, dy\, dz\\ = (K_1g,K_1f)_{H_{{\Bbb C}}}=(Kg,f)_{H_{{\Bbb C}}}
=\int_{{{\Bbb R}}^d}\int_{{{\Bbb R}}^d}(2\pi)^{-d/2}\overline{k(x-y)g(y)}\, dy\,
f(x)\, dx\\ = \int_{({{\Bbb R}}^d)^2}{\varkappa}(y-x)f(x)\overline g(y)\, dx\,
dy.\end{gathered}$$ Hence, $$\label{jiuszutr}(
\overline{\varkappa}_{1,x},\overline{\varkappa}_{1,y})_{H}={\varkappa}(y-x)\qquad
\text{a.e.\ $(x,y)\in ({{\Bbb R}}^d)^2 $}.$$ Furthermore, for each $x\in{{\Bbb R}}^d$, $$\begin{aligned}
(
\overline{\varkappa}_{1,x},\overline{\varkappa}_{1,x})_{H_{{\Bbb C}}}&=\int_{{{\Bbb R}}^d}|{\varkappa}_1(y-x)|^2\,
dy=\int_{{{\Bbb R}}^d}|{\varkappa}_1(y)|^2\,dy=\int_{{{\Bbb R}}^d}(2\pi)^{-d}\hat
k(\lambda)\,d\lambda\notag\\ &=(2\pi)^{-d/2}({{\cal F}}^{-1}\hat
k)(0)={\varkappa}(0).\label{hjas}\end{aligned}$$ Evidently, for any $(x_1,\dots,x_n)\in({{\Bbb R}}^d)^n$, $$\label{hghg}\det({\varkappa}(x_i-x_j))_{i,j=1}^n=\det({\varkappa}(x_j-x_i))_{i,j=1}^n.$$ Thus, – imply .$\blacksquare$
\[zure54e\] The measure $\mu_{\mathrm a}$ has correlation functions, which are given by $$\label{zutedrrd}
k_{\mu_{\mathrm a}}^{(n)}(x_1,\dots,x_n)=\det
({\varkappa}(x_i-x_j))_{i,j=1}^n,\qquad\text{{{\rm $(x_1,\dots,x_n)\in({{\Bbb R}}^d)^n$,
$n\in{{\Bbb N}}$}}.}$$
[*Proof*]{}. By Theorem \[e5q2343\], we have, for any $f_1,\dots,f_n\in S({{\Bbb R}}^d)$, $$I_{\mathrm a}^{-1} \big( {\langle}{{:}\,}\cdot^{\otimes n}{\,{:}},f_1{\hat\otimes}\dotsm{\hat\otimes}f_n{\rangle}\big) =
\int_{({{\Bbb R}}^d)^n}(f_1{\hat\otimes}\dotsm{\hat\otimes}f_n)(x_1,\dots,x_n) {{:}\,}{\rho_{\mathrm a}}(x_1)\dotsm {\rho_{\mathrm a}}(x_n){\,{:}}\,\Omega .$$ From here and Proposition \[jkgztd\] the statement easily follows. $\blacksquare$
\[zfsders\] Let the conditions of Theorem [[\[e5q2343\]]{}]{} be fulfilled. Then, $\mu_{\mathrm a}(\Gamma_{{{\Bbb R}}^d})=1$, the correlation functions of $\mu_{\mathrm a}$ are given by , and the Fourier transform of $\mu_{\mathrm
a}$ is calculated as follows: for each $f\in S({{\Bbb R}}^d)$ $$\label{zhgders98} \int
e^{i{\langle}\omega,f{\rangle}}\,\mu_{\mathrm
a}(d\omega)=\sum_{n=0}^\infty\frac1{n!}\int_{({{\Bbb R}}^d)^n}(e^{if(x_1)}-1)\dotsm
(e^{if(x_n)}-1)\det({\varkappa}(x_i-x_j))_{i,j=1}^n\, dx_1\dotsm
dx_n.$$
[*Proof*]{}. We evidently have $$|{\varkappa}(x)|\le(2\pi)^{-d}\|\hat k\|_{L_1({{\Bbb R}}^d)}\qquad \forall
x\in{{\Bbb R}}^d.$$ Hence, by [@M2 Corollary 3] and Corollary \[zure54e\] $$\label{hiwegfuw} |k_ {\mu_{\mathrm a}}^{(n)}(x_1,\dots,x_n)|\le
\big((2\pi)^{-d}\|\hat k\|_{L^1({{\Bbb R}}^d)}\big)^n n^{n/2}\qquad
\forall (x_1,\dots,x_n)\in ({{\Bbb R}}^d)^n,\ n\in{{\Bbb N}}.$$ By [@BKKL Theorem 2] (see also [@KK Theorem 6.5]), the bound implies that the measure $\mu_{\mathrm a}$ is concentrated on $\Gamma_{{{\Bbb R}}^d}$. Finally, formula follows in a standard way from Lemma \[zure54e\] and bound (see e.g. [@BKKL Remark 2]).$\blacksquare$
By Theorem \[zfsders\], $\mu_{\mathrm a}$ is a fermion process [@DV; @Ma1], or a determinantal random point field in terms of [@Sosh].
\[zurftt\][[Let us suppose that, in addition to condition , the function $\hat k$ satisfies $$\int_{{{\Bbb R}}^d}\hat
k(\lambda)|\lambda|^n\,d\lambda<\infty,\qquad \forall n\in{{\Bbb N}}.$$ Then, using formula , for each $v\in {\cal
V}_0({{\Bbb R}}^d)$, one can construct $J(v)$ as a selfadjoint operator in ${\cal F}(H_1\oplus H_2)$. Here, ${\cal V}_0({{\Bbb R}}^d)$ denotes the set of all smooth, compactly supported vector fields on ${{\Bbb R}}^d$. Thus, one gets a representation of the full algebra $\frak g$ (see Introduction). However, it is still an open problem, whether ${\frak H}_{\mathrm a}$ is an invariant subspace for the operators $J(v)$. If it were so, we could evidently construct a representation of the algebra $\frak g$, as well as the group $G$ in the space $L^2(\Gamma_{{{\Bbb R}}^d};\mu_{\mathrm a})$. ]{}]{}
Bosonic case
------------
We now suppose that holds. For each $x\in{{\Bbb R}}^d$, we introduce a particle density operator $\rho_{\mathrm
s}(x){:=}\varphi^*(x)\varphi(x)$, acting continuously from ${{\cal F}}_{\mathrm s}(\Phi)$ into ${{\cal F}}_{\mathrm s}^*(\Phi)$. Then, the operators $$\rho_{\mathrm s}(f){:=} \int_{{{\Bbb R}}^d}dx\,
f(x)\rho_{\mathrm s}(x),\qquad f\in S({{\Bbb R}}^d),$$ act continuously on ${{\cal F}}_{\mathrm fin}(H\oplus H)$. Using and an estimates of type , we show that $\rho_{\mathrm s
}(f)$ are essentially selfadjoint and their closures $\rho^\sim_{\mathrm s}(f)$ constitute a cyclic family of commuting selfadjoint operators in the Hilbert space ${\frak H}_{\mathrm
s}$—the closure of the linear span of the vectors $$\big\{\,\Omega,\, \rho_{\mathrm s }(f_1)\dotsm\rho_{\mathrm
s}(f_n)\Omega\mid f_1,\dots,f_n\in S({{\Bbb R}}^d),\ n\in{{\Bbb N}}\,\big\}$$ in ${{\cal F}}_{\mathrm s}(H\oplus H)$.
We then construct the spectral measure $\mu_{\mathrm s}$ of the operator family $(\rho^\sim_{\mathrm s}(f))_{f\in S({{\Bbb R}}^d)}$ as a probability measure on $(S'({{\Bbb R}}^d),{\cal B}(S'({{\Bbb R}}^d)))$. Furthermore, with the help of formula we show that $\mu_{\mathrm s}$ has correlation functions, which are given by the following formula: $$\label{zudrrd}
k_{\mu_{\mathrm s}}^{(n)}(x_1,\dots,x_n)={\operatorname{per}}({\varkappa}(x_i-x_j))_{i,j=1}^n,\qquad\text{{{\rm $(x_1,\dots,x_n)\in({{\Bbb R}}^d)^n$,
$n\in{{\Bbb N}}$}},}$$ where ${\varkappa}(x){:=}(2\pi)^{-d/2}k(x)$. Next, for every bounded $\Lambda\in{\cal B}({{\Bbb R}}^d)$, we evidently have the following estimate: $$\label{zure64e} \frac1{n!}\int_{\Lambda^n}|k_{\mu_{\mathrm s}}^{(n)}(x_1,\dots,x_n)|
\, dx_1\dotsm dx_n\le \big(|\Lambda|\,C_3\big)^n,$$ where $|\Lambda|$ denotes the volume of $\Lambda$ and $C_3{:=}\sup_{x\in{{\Bbb R}}^d}|{\varkappa}(x)|<\infty$. Hence, by and [@BKKL Th. 2], we get $\mu_{\mathrm s
}(\Gamma_{{{\Bbb R}}^d})=1$. Thus, we get the following
Let $k$ be the inverse Fourier transform of a function $\hat k$ satisfying . Let the Hilbert space ${\frak H}_{\mathrm s}$ and the operators $\rho^\sim_{\mathrm
s}(f)$, $f\in S({{\Bbb R}}^d)$, be defined as above. Then, there exist a unique probability measure $\mu_{\mathrm s}$ on $(S'({{\Bbb R}}^d),{\cal B }(S'({{\Bbb R}}^d)))$ and a unique unitary operator $I_{\mathrm s}:{\frak H}_{\mathrm s}\to L^2(S'({{\Bbb R}}^d);\mu_{\mathrm
s})$ such that $I_{\mathrm s}\Omega=1$ and $$I_{\mathrm
s}\,\rho^\sim_{\mathrm s}(f)I_{\mathrm s}^{-1}={\langle}\cdot,f{\rangle}\cdot,\qquad f\in S({{\Bbb R}}^d).$$ Furthermore, $\mu_{\mathrm s}(\Gamma_{{{\Bbb R}}^d})=1$ and the correlation functions $(k_{\mu_{\mathrm s}}^{(n)})_{n=0}^\infty$ of the measure $\mu_{\mathrm s}$ are given by formula .
By [@Ma0; @Ma1] (see also [@DV]), $\mu_{\mathrm s}$ is a boson process.
[**Acknowledgements**]{}
I am grateful to Yu. Kondratiev for drawing my attention to the fermion processes and for his permanent interest in this work. I would like to thank S. Albeverio, G. Goldin and Yu. Samoilenko for useful discussions. I am also grateful to the referees of the paper for many suggestions on improvement of the first version of the paper. The financial support of SFB 256, DFG Research Projects 436 RUS 113/593, and BMBF Research Project UKR-004-99 is gratefully acknowledged.
[99]{}
S. Albeverio, Yu. G. Kondratiev and M. Röckner, “Analysis and geometry on configuration spaces”, [*J. Func. Anal.*]{} [**154**]{} (1998) 444–500.
S. Albeverio, Yu. G. Kondratiev and M. Röckner, “Diffeomorphism groups and current algebras: configuration space analysis in quantum theory”, [*Rev.Math. Phys.*]{} [**11**]{} (1999) 1–23.
H. Araki, “Factorizable representation of current algebra. Non commutative extension of the L' evy-Kinchin formula and cohomology of a solvable group with values in a Hilbert space”, [*Publ. RIMS Kyoto Univ.*]{} [**5**]{} (1969/70) 361–422.
H. Araki, “On quasifree states of CAR and Bogoliubov automorphisms”, [*Publ. RIMS Kyoto Univ.*]{} [**6**]{} (1970/71) 385–442.
H. Araki and E. Woods, “Representations of the C.C.R. for a nonrelativistic infinite free Bose gas”, [*J. Math. Phys.*]{} [**4**]{} (1963) 637–662.
H. Araki and W. Wyss, “Representations of canonical anticommutation relation”, [*Helv. Phys. Acta*]{} [**37**]{} (1964) 136–159.
E. Balslev and A. Verbeure, “States on Clifford algebras”, [*Commun. Math. Phys.*]{} [**7**]{} (1968) 55–76.
Ch. Benard and O. Macchi, “Detection and “emission” processes of quantum particles in a “chaotic state””, [*J. Math. Phys.*]{} [**14**]{} (1973) 155–167.
Yu. M. Berezansky, “Commutative Jacobi fields in Fock space”, [*Integral Equations Operator Theory*]{} [**30**]{} (1998) 163–190.
Yu. M. Berezansky, “Poisson measure as the spectral measure of Jacobi field”, [*Infin. Dimens. Anal. Quantum Probab. Relat. Top.*]{} [**3**]{} (2000) 121–139.
, [*Spectral Methods in Infinite Dimensional Analysis*]{}, Kluwer, Dordrecht, 1994.
Yu. M. Berezansky, Yu. G. Kondratiev, T. Kuna and E. Lytvynov, “On a spectral representation for correlation measures in configuration space analysis”, [*Meth. Func.Anal. and Topol.*]{} [**5**]{} (1999), no. 4, 87–100.
A. Borodin and G. Olshanski, “Point processes and the infinite symmetric group. Part III: Fermion point processes”, Preprint, 1998, available viahttp://xxx.lanl.gov/abs/math.RT/9804088.
O. Bratteli and W. D. Robinson, [*Operator Algebras and Quantum-Statistical Mechanics I, II*]{}, Springer-Verlag, New York/Berlin, 1979, 1981.
D. J. Daley and D. Vere-Jones, [*An Introduction to the Theory of Point Processes*]{}, Springer-Verlag, New York, 1988.
R. Dashen and D. H. Sharp, “Currents as coordinates for hadrons”, [*Phys. Rev.*]{} [**165**]{} (1968) 1857–1867.
G. F. Dell’Antonio, “Structure of the algebra of some free systems”, [*Comm. Math. Phys.*]{} [**9**]{} (1968) 81–117.
K.-H. Fichtner, “On the position distribution of the ideal Bose gas”, [*Math. Nachr.*]{} [**151**]{} (1991) 59–67.
K.-H. Fichtner and W. Freudenberg, “Point processes and states of infinite boson systems”, Preprint NTZ Leipzig, 1986.
K.-H. Fichtner and W. Freudenberg, “Point processes and the position distribution of infinite boson systems”, [*J. Statist. Phys.*]{} [**47**]{} (1987) 959–978.
A. Girard, “Current algebras of free systems at finite temperature”, [*J. Math. Phys.*]{} [**14**]{} (1973) 353–365.
G. A. Goldin, “Nonrelativistic current algebras as unitary representations of groups”, [*J. Math. Phys.*]{} [**12**]{} (1971) 462–487.
G. A. Goldin, J. Grodnik, R. T. Powers and D. H. Sharp, “Nonrelativistic current algebra in the $N/V$ limit”, [*J. Math. Phys.*]{} [**15**]{} (1974) 88–100.
G. A. Goldin, R. Menikoff and D. H. Sharp, “Particle statistics from induced representations of a local current group”, [*J. Math. Phys.*]{} [**21**]{} (1980) 650–664.
R. L. Hudson and K. R. Parthasarathy, “Quantum Ito’s formula and stochastic evolutions”, [*Commun. Math. Phys.*]{} (1984) 301–323.
Yu. G. Kondratiev and T. Kuna, “Harmonic analysis on configuration spaces I. General theory”, Preprint no. 626, Bonn University, 1999, to appear in [*Infinite Dimens. Anal. Quantum Prob. Related Topics*]{}.
, “Multiple Wiener integrals and non-Gaussian white noises: a Jacobi field approach”, [*Meth. Func. Anal.and Topol.*]{} [**1**]{} (1995) 61–85.
O. Macchi, “Distribution statistique des instants d’émission des photoélectrons d’une lumière thermique”, [*C. R. Acad. Sci. Paris Ser. A*]{} [**272**]{} (1971) 437–440.
O. Macchi, “The coincidence approach to stochastic point processes”, [*Adv. Appl. Prob.*]{} [**7**]{} (1975) 83–122.
O. Macchi, “The [F]{}ermion process—a model of stochastic point process with repulsive points”, pp. 391–398 in [*Transactions of the Seventh Prague Conference on Information Theory, Statistical Decision Functions, Random Processes and of the Eighth European Meeting of Statisticians (Tech. Univ. Prague, Prague, 1974), Vol. A*]{}, Reidel, Dordrecht, 1977.
J. Manuceau and A. Verbeure, “Representations of anticommutation relations and Bogolioubov transformations”, [*Comm. Math. Phys.*]{} [**8**]{} (1968) 315–326.
P. A. Meyer, [*Quantum Probability for Probabilists*]{}, Lect. Notes in Math., Vol. 1538, Springer-Verlag, Berlin/New York, 1993.
R. Menikoff, “The Hamiltonian and generating functional for a nonrelativistic local current algebra”, [*J. Math. Phys.*]{} [**15**]{} (1974) 1138–1152.
R. Menikoff, “Generating functionals determining representations of a nonrelativistic local current algebra in the $N/V$ limit”, [*J. Math. Phys.*]{} [**15**]{} (1974) 1394–1408.
R. Menikoff and D. H. Sharp, “Representations of a local current algebra: Their dynamical determination”, [*J. Math. Phys.*]{} [**16**]{} (1975) 2341–2352.
R. T. Powers and Strømer, “Free states of the canonical abticommutation relations”, [*Comm. Math. Phys.*]{} (1970) 1–33.
M. Reed and B. Simon, [*Methods of Modern Mathematical Physics. II. Fourier Analysis, Self-Adjointness*]{}, Academic Press, New York, 1975.
G. Rideau, “On some representations of the anticommutation relations”, [*Comm. Math. Phys.*]{} [**9**]{} (1968) 229–241.
Y. S. Samoilenko, [*Spectral Theory of Families of Self-Adjoint Operators*]{}, Kluwer, Dordrecht, 1987.
D. Shale and W. F. Stinespring, “States on the Clifford algebra”, [*Ann. Math.*]{} [**80**]{} (1964) 365–381.
T. Shirai and Y. Takahashi, “Random point field associated with certain Feedholm determinant II: fermion shift and its ergodic and Gibbs properties”, Preprint, 2001, available via http://neptune.math.titech.ac.jp/ shirai/preprint.html.
T. Shirai and H. J. Yoo, “Glauber dynamics for fermion point processes”, Preprint, 2001, available via http://neptune.math.titech.ac.jp/ shirai/preprint.html.
A. Soshnikov, “Determinantal random point fields”, [*Russian Math.Surveys*]{} [**55**]{} (2000) 923–975.
A. Soshnikov, “Gaussian fluctuation for the number of particles in Airy, Bessel, sine, and other determinantal random point fields”, [*J. Statist. Phys.*]{} [**100**]{} (2000) 491–522.
A. Soshnikov, “Gaussian limit for determinantal random point fields”, Preprint, 2001, available via http://xxx.lanl.gov/abs/math.PR/0006037.
H. Spohn, Interacting Brownian particles: A study of Dyson’s model, pp. 151–179 in [*Hydrodynamic Behavior and Interacting Partcle Systems*]{}, G. Papanicolau (ed.), Springer-Verlag, New York, 1987.
D. Surgailis, On multiple Poisson stochastic integrals and associated Markov semigroups, [*Probab. Math.Statist.*]{} [**3**]{} (1984), 217–239.
A. M. Vershik, I. M. Gelfand, and M. I. Graev, “Representations of the group of diffeomorphisms”, [*Russian Math. Surveys*]{} (1975), no. 6, 1–50.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $(M,g)$ be a closed oriented negatively curved surface. A unitary connection on a Hermitian vector bundle over $M$ is said to be transparent if its parallel transport along the closed geodesics of $g$ is the identity. We study the space of such connections modulo gauge and we prove a classification result in terms of the solutions of certain PDE that arises naturally in the problem. We also show a local uniqueness result for the trivial connection and that there is a transparent $SU(2)$-connection associated to each meromorphic function on $M$.'
address: ' Department of Pure Mathematics and Mathematical Statistics, University of Cambridge, Cambridge CB3 0WB, UK'
author:
- 'Gabriel P. Paternain'
title: Transparent connections over negatively curved surfaces
---
Introduction
============
Let $(M,g)$ be a closed Riemannian manifold and let $E\to M$ be a Hermitian vector bundle of rank $n$ over $M$. A unitary connection $\nabla$ on $E$ is said to be [*transparent*]{} if its parallel transport along every closed geodesic of $g$ is the identity. These connections are “ghosts" or “invisible" from the point of view of the closed geodesics of $g$. Clearly, if $\nabla$ is transparent any other connection gauge equivalent to it will also be transparent. The goal of the present paper is the study of transparent connections modulo gauge transformations when $(M,g)$ is a closed oriented negatively curved surface.
The motivation for studying this problem comes from several a priori unrelated quarters. Transparent connections on $S^2$ (and $\re \mathbb P^2$) arise in a natural way in the theory of integrable systems and solitons when studying the Bogomolny equations $D\Phi=\star F$ in $(2+1)$-dimensional Minkowski space [@Wa1; @Wa2]. Here $\Phi$ is the Higgs field, $F$ is the curvature of the connection, $\star$ is the Hodge star operator of the metric and $D$ is the induced connection on the endomorphism bundle. The condition of having trivial holonomy along the closed geodesics of a compactified space-like plane picks up finite dimensional families of solutions and enables the use of methods from twistor theory over a compact twistor space. [@A]. In fact, using a more refined twistor correspondence, L. Mason has recently classified all transparent connections on $S^2$ and $\re \mathbb P^2$ with the standard round metric [@Ma1]. For the case of $S^2$, his results say that the space of transparent connections modulo gauge is in 1-1 correspondence with holomorphic vector bundles $W\to (\mathbb C\mathbb P^2)^*$ and positive definite Hermitian metrics on $W$ restricted to the real slice $(\mathbb R\mathbb P^2)^*$. Similar results are obtained for anti-self-dual Yang-Mills connections over $S^2\times S^2$ with split signature, see [@Ma].
The problem of determining a connection from its parallel transport along geodesics is a natural integral-geometry problem that can be considered also in the case of manifolds with boundary or $\re^d$ with appropriate decay conditions at infinity. It arises for example when one considers the wave equation associated to the Schrödinger equation with external Yang-Mills potential $A$ and the inverse problem of determining the potential $A$ from the Dirichlet-to-Neumann map $\Lambda_{A}$. There are various results known for the integral-geometry problem. Local uniqueness theorems under various assumptions on the connection or its curvature are proved by V.A. Sharafutdinov [@Sha], R. Novikov [@No] and D. Finch and G. Uhlmann [@FU]. A global uniqueness result for connections with compact support is proved by G. Eskin in [@E]. In the case of $\re^2$, Novikov shows (building on the work of Ward previously mentioned) that global uniqueness may fail and in fact, his construction gives non-trivial transparent connections over $\re \mathbb P^2$. He also shows global uniqueness (with reconstruction) for $d\geq 3$ without assuming compact support.
As we mentioned before, in the present paper we will discuss transparent connections when the metric is negatively curved, or more generally, when its geodesic flow is Anosov. While our main focus here will be in the non-abelian case, we should mention that the abelian case $n=1$ is also of interest, but it can be reduced to known results (see Theorem \[thm:abelian\]) to obtain that transparent connections, when they exist, are unique up to gauge equivalence. The abelian case arises also when discussing positivity of entropy production in dissipative geodesic flows or thermostats [@DP], thus showing that the problem of understanding transparent connections also pops up naturally in dynamical systems and non-equilibrium statistical mechanics.
Our first result (Theorem \[thm:livtop\]) asserts that not all bundles over a surface of genus `g` carry transparent connections. In fact we show that a complex vector bundle $E$ over $M$ admits a transparent connection if and only if $2-2\texttt{g}$ divides its first Chern class $c_{1}(E)$. This result, and subsequent ones, are based on the classical Livsic theorem for non-abelian cocycles which is recalled in Section \[sec:livsic\].
One of the obvious differences with the abelian case is the appearance of the following ghosts. Let $K$ be the canonical line bundle and $K^s$ with $s\in \Z$ be its tensor powers (if $s=0$ we get the trivial bundle). The powers $K^s$ for $s\neq 0$ carry the Levi-Civita connection which we denote by $\nabla_{\ell}^{s}$. If $s=0$ we understand that this is the trivial connection. Note that the Levi-Civita connection on $TM$ ($=K^{-1}$) is transparent, since the parallel transport along a closed geodesic $\gamma$ must fix $\dot{\gamma}(0)$ and consequently any vector orthogonal to it since the parallel transport is an isometry and the surface is orientable. Thus any $\nabla_{\ell}^{s}$ is transparent. Given an $n$-tuple of integers $S:=(s_{1},\dots,s_{n})$, the connection $$\nabla_{\ell}^{S}:=\nabla_{\ell}^{s_{1}}\oplus\cdots \oplus\nabla_{\ell}^{s_{n}}$$ defines a transparent unitary connection on the bundle $E_{S}:=K^{s_{1}}\oplus\cdots\oplus K^{s_{n}}$. Clearly $c_{1}(E_{S})=(2\texttt{g}-2)(s_{1}+\dots+s_{n})$ and any complex vector bundle $E$ supporting a transparent connection is isomorphic to $E_S$ for $S$ such that $c_{1}(E)=c_{1}(E_{S})$.
Now let $E$ be a Hermitian vector bundle and consider a unitary isomorphism $\tau:E\to E_{S}$, where $S$ is such that $c_{1}(E)=c_{1}(E_{S})$. The unitary connection $\tau^*\nabla_{\ell}^S$ is a transparent connection on $E$ and its gauge equivalence class, denoted by $[S]$, is independent of $\tau$. Note that $[S_1]=[S_2]$ if and only if $S_1$ and $S_2$ coincide up to a permutation. However, as we shall see below, these will not be the only ghosts.
Given two transparent connections $\nabla^1$ and $\nabla^2$ write $\nabla^2=\nabla^1+A$ where $A\in \Omega^{1}(M,\mbox{\rm ad}\,E)$. Let $\pi: SM\to M$ be the unit circle bundle and $X$ the vector field of the geodesic flow of the metric. The Livsic theorem will provide solutions $u\in \Omega^{0}(SM,\mbox{\rm Aut}\,\pi^*E)$ of $D_{X}u+Au=0$, where $D$ is the connection on the bundle of endomorphisms of $\pi^*E$ induced by $\nabla^1$. Inspired by the methods in [@GK] we will show that these solutions must have a [*finite*]{} Fourier expansion (cf. Theorem \[thm:deg\]), in other words, $u$ must be a polynomial in the velocities. The degree of this polynomial will allow us to define a distance function on the set of transparent connections modulo gauge. As a consequence of these results we will derive the following local uniqueness statement for the trivial connection on the trivial bundle:
[**Theorem A**]{}. [*Consider the trivial bundle and let $\nabla$ be a transparent connection with curvature $F_{\nabla}$. Let $K<0$ be the Gaussian curvature of the surface and suppose that the Hermitian matrix $\pm i\star F_{\nabla}(x)-K(x)\,\mbox{\rm Id}$ is positive definite for all $x\in M$. Then $\nabla$ is gauge equivalent to the trivial connection.*]{}
Thus, if a transparent connection has small enough curvature, it must be gauge equivalent to the trivial connection. Note that Theorem A is sharp, since a ghost $\nabla\in [S]$ with $\sum_{k=1}^{n} s_{k}=0$, $s_{k}\in \{0,\pm 1\}$ and $S\neq 0$ has $\pm i\star F_{\nabla}(x)-K(x)\,\mbox{\rm Id}$ positive semi-definite for all $x\in M$.
Our second main result is a classification of the set $\T$ of transparent connections on $E$ modulo gauge in terms of the solutions of certain non-linear PDE that arises naturally in the problem. In order to describe this PDE, recall that the unit sphere bundle $SM$ of $M$ has a canonical frame $\{X,H,V\}$ where $X$ is the geodesic vector field, $V$ is the vertical vector field and $H=[V,X]$ is the horizontal vector field. Let $f:SM\to \mathfrak{u}(n)$ be a smooth function, where $\mathfrak{u}(n)$ denotes the set of all $n\times n$ skew-Hermitian matrices. Consider the PDE: $$H(f)+VX(f)=[X(f),f].
\label{eq:keypde}$$ Observe that the set $\mathcal H$ of solutions to (\[eq:keypde\]) is invariant under the action of $U(n)$ given by $f\mapsto q^{-1}fq$, where $q\in U(n)$.
We shall say that two functions $f,h:SM\to \mathfrak u(n)$ are [*$V$-cohomologous*]{} if there exists a smooth function $u:SM\to U(n)$ such that $f=u^{-1}V(u)+u^{-1}h u$.
Given a constant matrix $c\in \mathfrak u(n)$ with $e^{2\pi c}=\mbox{\rm Id}$ we consider the $U(n)$-invariant subset $\mathcal H_{c}\subset \mathcal H$ given by those solutions $f$ which are $V$-cohomologous to $c$. The set $\mathcal H_{c}$ only depends on $\mbox{\rm tr}(c)$ (see Lemma \[lemma:traces\]).
[**Theorem B**]{}. [*Let $E$ be a Hermitian vector bundle over a closed oriented surface of genus*]{} $\texttt{g}$ [*whose geodesic flow is Anosov. Suppose that*]{} $2-2\texttt{g}$ [*divides $c_{1}(E)$ and let $c\in \mathfrak u(n)$ be a constant matrix with $e^{2\pi c}=\mbox{\rm Id}$ and*]{} $c_{1}(E)=(2\texttt{g}-2)\,\mbox{\rm tr}(ic)$. [*Then, there is a 1-1 correspondence between $\T$ and $\mathcal H_{c}/U(n)$.*]{}
In the abelian case $n=1$, it is not hard to see that the only solutions to (\[eq:keypde\]) are the constants provided $K<0$. This can be shown using the energy estimates method (the Pestov identity) which also gives some information about (\[eq:keypde\]) for $n\geq 2$. This is discussed at the end of the paper, Subsection \[sub:last\].
For $n\geq 2$, the constant solutions in $\mathcal H_{c}$ correspond precisely to the Levi-Civita ghosts $[S]$, but as we mentioned before, there are other ghosts besides $[S]$ and these have to come from non-constants solutions to (\[eq:keypde\]). To see that this is the case we consider functions $f$ which only depend on the base point $x$. Under such assumption, it is easy to see that (\[eq:keypde\]) turns into $2\star df=[df,f]$, which only depends on the conformal class of the metric $g$. We discuss this equation in Subsection \[sub:sol\] for the case of $SU(2)$ and we show that its non-zero solutions correspond precisely with the set of holomorphic maps $f:M\to \mathbb C\mathbb P^1$. We also show that all these maps are $V$-cohomologous to the zero matrix. In this way, via Theorem B, we virtually obtain as many $SU(2)$-transparent connections on the trivial bundle (modulo gauge) as meromorphic functions on $M$; these are all the transparent connections at distance one from the trivial connection (cf. Corollary \[cor:last\] for the precise statement).
There are several questions which are worthy of further consideration. In particular, it would be interesting to exhibit elements in $\mathcal H_c$ which have dependence on the velocities. It seems that one can deal with this issue using an appropriate Bäcklund transformation, but we leave it as the subject of a future paper. The inclusion of a Higgs field $\Phi$ and the problem of understanding transparent pairs $(\nabla,\Phi)$ is also of interest, but it will also be discussed elsewhere.
[*Acknowledgements:*]{} I am very grateful to N. Dairbekov, M. Dunajski, L. Mason, R. Novikov, V. Sharafutdinov, R. Spatzier and G. Uhlmann for several useful comments and discussions related to this paper. I am also greateful to the referee for several comments and corrections that improved the presentation.
The Livsic theorem for non-abelian cocycles {#sec:livsic}
===========================================
Let $X$ be a closed manifold and $\phi_t:X\to X$ a smooth transitive Anosov flow. Recall that $\phi_t$ is Anosov if there is a continuous splitting $TX=E^0\oplus E^{u}\oplus E^{s}$, where $E^0$ is the flow direction, and there are constants $C>0$ and $0<\rho<1<\eta$ such that for all $t>0$ we have $$\|d\phi_{-t}|_{E^{u}}\|\leq C\,\eta^{-t}\;\;\;\;\mbox{\rm
and}\;\;\;\|d\phi_{t}|_{E^{s}}\|\leq C\,\rho^{t}.$$ It is very well known that the geodesic flow of a closed negatively curved Riemannian manifold is a transitive Anosov flow.
Let $G$ be a compact Lie group; for the purposes of this paper it is enough to think of $G$ as $U(n)$.
A $G$-valued cocycle over the flow $\phi_t$ is a map $C:X\times\re\to G$ that satisfies $$C(x,t+s)=C(\phi_{t}x,s)\,C(x,t)$$ for all $x\in X$ and $s,t\in\re$. \[def1\]
In this paper the cocycles will always be smooth. In this case $C$ is determined by its infinitesimal generator $A:X\to\mathfrak g$ given by $$A(x):=\left.\frac{d}{dt}\right|_{t=0}C(x,t).$$ The cocycle can be recovered from $A$ as the unique solution to $$\frac{d}{dt}C(x,t)=dR_{C(x,t)}(A(\phi_{t}x)),\;\;\;C(x,0)=\mbox{\rm Id},$$ where $R_{g}$ is right translation by $g\in G$.
Let $C$ be a smooth cocycle such that $C(x,T)=\mbox{\rm Id}$ whenever $\phi_{T}x=x$. Then, there exists a smooth function $u:X\to G$ such that $$C(x,t)=u(\phi_{t}x)u(x)^{-1}.$$ \[livsic\]
The existence of a Hölder continuous function $u$ (assuming $A$ is Hölder) was proved by Livsic [@L1; @L2]. Smoothness of $u$ was proved by Niţică and Török [@NT].
In our applications we will need to consider non trivial vector bundles. Suppose $E$ is a rank $n$ Hermitian vector bundle over $X$ and $\phi_t:X\to X$ is as above, a smooth transitive Anosov flow.
A cocycle over $\phi_t$ is an action of $\re$ by bundle automorphisms which covers $\phi_t$. In other words, for each $(x,t)\in X\times \re$, we have a unitary map $C(x,t):E_{x}\to E_{\phi_{t}x}$ such that $C(x,t+s)=C(\phi_{t}x,s)\,C(x,t)$.
If $E$ admits a unitary trivialization $f:E\to X\times \C^n$, then $$f\,C(x,t)\,f^{-1}(x,a)=(\phi_{t}x,D(x,t)a),$$ where $D:X\times\re\to U(n)$ is a cocycle as in Definition \[def1\].
Let $E^*$ denote the dual vector bundle to $E$. If $E$ carries a Hermitian metric $h$, we have a conjugate isomorphism $\ell_h: E\to E^*$, which induces a Hermitian metric $h^*$ on $E^*$. Given a cocycle $C$ on $E$, $C^*:=\ell_{h}\,C\,\ell_{h}^{-1}$ is a cocycle on $(E^*,h^*)$.
Let $E$ be a Hermitian vector bundle over $X$ such that $E\oplus E^*$ is a trivial vector bundle. Let $C$ be a smooth cocycle on $E$ such that $C(x,T)=\mbox{\rm Id}$ whenever $\phi_T x=x$. Then $E$ is a trivial vector bundle. \[trivial\]
As explained above, the cocycle $C$ on $E$ induces a cocycle $C^*$ on $E^*$. On the trivial vector bundle $E\oplus E^*$ we consider the cocycle $C\oplus C^*$. Clearly $C\oplus C^*(x,T)=\mbox{\rm Id}$ everytime that $\phi_T x=x$. Choose a unitary trivialization $f:E\oplus E^*\to X\times \C^{2n}$ and write $$f\,C\oplus C^*(x,t)\,f^{-1}(x,a)=(\phi_{t}x,D(x,t)a).$$ By Theorem \[livsic\], there exists a smooth function $u:X\to U(2n)$ such that $D(x,t)=u(\phi_{t}x)u^{-1}(x)$. Since $\phi_t$ is a transitive flow, we may choose $x_0\in X$ with a dense orbit and without loss of generality we may suppose that $u(x_0)=\mbox{\rm Id}$. Let $$\{e_{1}(x_0),\dots,e_{n}(x_{0})\}$$ be a unitary frame at $E_{x_{0}}$. Write $f(x_0,e_{i}(x_{0}))=(x_{0},a_{i})$, where $a_{i}\in \C^{2n}$ and let $$e_{i}(x):=f^{-1}(x,u(x)a_{i}).$$ Clearly at every $x\in X$, $\{e_{1}(x),\dots,e_{n}(x)\}$ is a smooth unitary $n$-frame of $E_{x}\oplus E_{x}^{*}$. We claim that in fact $e_{i}(x)\in E_x$ for all $x\in X$. This, of course, implies the triviality of $E$. Note that $$e_{i}(\phi_{t}x_{0})=f^{-1}(\phi_{t}x_{0},u(\phi_{t}x_{0})a_{i})=
f^{-1}(\phi_{t}x_{0},D(x_{0},t)a_{i})=C\oplus C^{*}(x_{0},t)e_{i}(x_{0}).$$ But $e_{i}(x_{0})\in E_{x_{0}}$, thus $e_{i}(\phi_{t}x_{0})\in E_{\phi_{t}x_{0}}$. It follows that $e_{i}(x)\in E_{x}$ for a dense set of points in $X$. By continuity of $e_{i}$, $e_{i}(x)\in E_{x}$ for all $x\in X$.
The hypothesis of $E\oplus E^*$ being trivial is not needed in Proposition \[trivial\]. Ralf Spatzier has informed the author that it is possible to adapt the proof of the usual Livsic periodic data theorem to show directly that $E$ is trivial. However, this weaker version is all that we will need in this paper.
A proof of the measurable Livsic theorem for bundles (which we do not use here) may be found in [@GS].
Transparent connections and the Livsic theorem {#tliv}
==============================================
Let $M^d$ be a closed orientable Riemannian manifold whose geodesic flow $\phi_t$ is Anosov. The geodesic flow acts on the unit sphere bundle $SM$ and we let $\pi:SM\to M$ be the footpoint projection.
Let $E\to M$ be a Hermitian vector bundle of rank $n$ and let $\nabla$ be a unitary connection on $E$. Given a geodesic $\gamma:\re\to M$ with initial conditions $(x,v)\in SM$, we can consider the parallel transport of $\nabla$ along $\gamma$. The parallel transport $P_{x,\gamma(t)}:E_{x}\to E_{\gamma(t)}$ is an isometry and defines a smooth cocycle $C$ over the geodesic flow on the Hermitian vector bundle over $SM$ given by the pull-back bundle $\pi^* E$. The connection $\nabla$ is transparent if and only if $C(x,v,T)=\mbox{\rm Id}$ every time that $\phi_{T}(x,v)=(x,v)$.
Arbitrary bundles over an Anosov surface
----------------------------------------
Suppose $d=2$. In this case, complex vector bundles $E$ over $M$ are classified topologically by the first Chern class $c_{1}(E)\in H^{2}(M,\Z)=\Z$. Since $c_{1}(E^*)=-c_{1}(E)$ and $c_1$ is additive with respect to direct sums, we see that $E\oplus E^*$ is the trivial bundle and therefore we will be able to apply Proposition \[trivial\]. In fact we will show:
Let $M$ be a closed orientable surface of genus $\mbox{\tt g}$ whose geodesic flow is Anosov. A complex vector bundle $E$ over $M$ admits a transparent connection if and only if $2-2\tt{g}$ divides $c_{1}(E)$. \[thm:livtop\]
Suppose $E$ admits a transparent connection. As explained above we may apply Proposition \[trivial\] to deduce that $\pi^*E$ is a trivial bundle and since $c_{1}(\pi^*E)=\pi^* c_{1}(E)$ we conclude that $\pi^*c_{1}(E)=0$. Consider now the Gysin sequence of the unit circle bundle $\pi:SM\to M$, $$0\to H^{1}(M,\Z)\stackrel{\pi^*}{\to}H^{1}(SM,\Z)\stackrel{0}{\to}
H^{0}(M,\Z)\stackrel{\times(2-2\texttt{g})}{\longrightarrow} H^{2}(M,\Z)\stackrel{\pi^*}{\to}
H^{2}(SM,\Z)\to\cdots.$$ We see that $\pi^*c_{1}(E)=0$ if and only if $c_1(E)$ is in the image of the map $H^0(M,\Z)\to H^{2}(M,\Z)$ given by cup product with the Euler class of the unit circle bundle. Equivalently, $2-2\texttt{g}$ must divide $c_1(E)$.
Let $K$ be the canonical line bundle of $M$. We can think of $K$ as the cotangent bundle to $M$; it has $c_{1}(K)=2\texttt{g}-2$. The tensor powers $K^s$ of $K$ (positive and negative) generate all possible line bundles with first Chern class divisible by $2-2\texttt{g}$ and they all carry the unitary connection induced by the Levi-Civita connection of the Riemannian metric on $M$. All these connections are clearly transparent. Topologically, all complex vector bundles over $M$ whose whose first Chern class is divisible by $2-2\texttt{g}$ are of the form $K^s\oplus \varepsilon$, where $\varepsilon$ is the trivial vector bundle. Since the trivial connection on the trivial bundle is obviously transparent, it follows that every complex vector bundle whose first Chern class is divisible by $2-2\texttt{g}$ admits a transparent connection.
Arbitrary bundles over an Anosov 3-manifold
-------------------------------------------
Suppose $M$ is a closed 3-manifold whose geodesic flow is Anosov. Complex vector bundles $E$ over $M$ are also classified topologically by $c_{1}(E)\in H^{2}(M,\Z)$, hence as in the two dimensional case, $E\oplus E^{*}$ is the trivial bundle. Thus if $E$ admits a transparent connection, Proposition \[trivial\] implies that $\pi^{*}E$ is trivial. However now the argument with the Gysin sequence that we explained in the proof of Theorem \[thm:livtop\] shows that $\pi^{*}:H^{2}(M,\Z)\to H^{2}(SM,\Z)$ is injective and thus $c_{1}(E)=0$. Therefore if $E$ admits a transparent connection, it must be trivial. This shows that the problem for 3-manifolds is in some sense simpler than the problem for surfaces, at least, there are no obvious transparent connections besides the trivial one.
The abelian case
----------------
The goal of this subsection is to show the following result.
Let $M$ be a closed orientable Riemannian manifold whose geodesic flow is Anosov and let $E$ be a Hermitian line bundle over $M$. Then, any two transparent connections are gauge equivalent. \[thm:abelian\]
Let $\nabla^1$ and $\nabla^2$ be transparent connections. We may write $\nabla^2=\nabla^1+A$, where $A\in \Omega^{1}(M,\mbox{\rm ad}\,E)$. Since $E$ is a line bundle, $A=i\theta$, where $\theta$ is a real-valued 1-form in $M$. Since $\nabla^1$ and $\nabla^2$ are transparent, $$\int_{\gamma}\theta\in 2\pi\,\Z
\label{trans}$$ for every closed geodesic $\gamma$. Consider the cocycle over $\phi_t$, $C:SM\times\re\to S^1$ defined as follows. Given $(x,v)\in SM$, let $\gamma:\re\to M$ be the unique geodesic with initial conditions $(x,v)$. Set $$C(x,v,t):=\exp\left(i\int_{0}^t\theta_{\gamma(s)}(\dot{\gamma}(s))\,ds \right).$$ By (\[trans\]), the cocycle $C$ has the property that $C(x,v,T)=1$, every time that $\phi_T(x,v)=(x,v)$, thus by Theorem \[livsic\], there is a smooth function $u:SM\to S^1$ such that $C(x,v,t)=u(\phi_{t}(x,v))u^{-1}(x,v)$. If we differentiate this equality with respect to $t$ and set $t=0$ we obtain $$ui\theta=du(X),
\label{ec}$$ where $X$ is the geodesic vector field. The function $u$ gives rise to a real-valued closed 1-form in $SM$ given by $\varphi:=\frac{du}{iu}$. Since $\pi^*:H^1(M,\re)\to H^{1}(SM,\re)$ is an isomorphism (this follows easily from the Gysin sequence, since $M$ cannot be the 2-torus), there exists a closed 1-form $\omega$ in $M$ and a smooth function $f:SM\to\re$ such that $$\varphi=\pi^*\omega+df.$$ When this equality is applied to $X$ and combined with (\[ec\]) one obtains $$\theta_{x}(v)-\omega_{x}(v)=df(X(x,v))$$ for all $(x,v)\in SM$. This cohomological equation is actually equivalent -via the classic Livsic theorem for $\re$-values cocycles- to saying that $$\int_{\gamma}\theta-\omega=0$$ for every closed geodesic $\gamma$. It is known that this implies that $\theta-\omega$ is exact. This was proved by V. Guillemin and D. Kazhdan [@GK] for surfaces of negative curvature, by C. Croke and Sharafutdinov [@CS] for arbitrary manifolds of negative curvature and by N.S. Dairbekov and Sharafutdinov [@DS] for manifolds whose geodesic flows is Anosov.
If $\theta-\omega$ is exact, $\theta$ must be closed and by (\[trans\]), $[\theta]/2\pi\in H^{1}(M,\Z)$. Thus there exists a smooth function $g:M\to S^1$ such that $\theta=dg/ig$. This is precisely the statement that $\nabla^1$ and $\nabla^2$ are gauge equivalent.
Setting up the Fourier analysis
===============================
Let $E$ be a complex Hermitian vector bundle over $M$ and let $\nabla$ be a unitary connection on $E$. If $\pi:SM\to M$ denotes the canonical projection, then $\nabla$ induces a unitary connection on the pull-back bundle $\pi^*E$ that we denote by $\pi^*\nabla$. This pull-back connection induces in turn a unitary connection on the bundle $\mbox{\rm End}\,\pi^*E$ of complex endomorphisms of $\pi^*E$, which we denote by $D$. Note that $\mbox{\rm End}\,\pi^*E$ naturally inherits a Hermitian metric determined by the trace $\mbox{\rm tr}(u\,w^*)$ where $u,w\in \Omega^{0}(SM,\mbox{\rm End}\,\pi^*E)$. This Hermitian metric together with the Liouville measure $\mu$ of $SM$ combine to give an $L^2$-inner product of sections $$\langle u,w^*\rangle=\int_{SM} \mbox{\rm tr}(u\,w^*)\,d\mu.$$
Let $F_{\nabla}\in \Omega^2(M,\mbox{\rm ad}\,E)$ be the curvature of $\nabla$. Then $F_{\pi^*\nabla}\in \Omega^2(SM,\mbox{\rm ad}\,\pi^*E)$ is given by $F_{\pi^*\nabla}=\pi^*F_{\nabla}$ and $F_{D}\in \Omega^2(SM,\mbox{\rm ad}\,\mbox{\rm End}\,\pi^*E)$ is given by $F_{D}=[F_{\pi*\nabla},\,\cdot]$. Note that if $\star$ denotes the Hodge star operator of the metric, then $\star F_{\nabla}\in \Omega^{0}(M,\mbox{\rm ad}\,E)$.
The vertical vector field $V$ and the connection $D$ induce a first order differential operator $$D_{V}:\Omega^{0}(SM,\mbox{\rm End}\,\pi^*E)\to
\Omega^{0}(SM,\mbox{\rm End}\,\pi^*E)$$ which in fact is independent of $\nabla$, for if $\nabla'$ is another connection and we write $\nabla'=\nabla+A$, then $D'=D+[\pi^{*}A,\cdot]$ and $D'_{V}=D_{V}$ since $\pi^*A(V)=0$.
Note that $-iD_{V}$ is self-adjoint, since $V$ preserves the Lioville measure $\mu$. Indeed, observe that the compatibility of $D$ with the Hermitian metric implies $V\langle u,w\rangle=\langle D_{V}u,w\rangle+\langle u,D_{V}w\rangle$. Since the integral of $V\langle u,w\rangle$ with respect to $\mu$ vanishes, $(D_{V})^*=-D_{V}$. We also have an orthogonal decomposition $$L^{2}(SM,\mbox{\rm End}\,\pi^*E)=\bigoplus _{n\in\Z}H_{n}$$ such that $-iD_{V}$ acts as $n\,\mbox{\rm Id}$ on $H_n$. To see this, triangulate $M$ in such a way that both $SM\to M$ and $E\to M$ are trivial over each face $M_r$ of the triangulation. Since $L^{2}(SM,\mbox{\rm End}\,\pi^*E)$ is isomorphic to $\oplus_{r}L^{2}(SM_{r},\mbox{\rm End}\,\pi^*E)$ we are reduced to the case of both bundles being trivial in which case the claim is clear because $D_{V}u=V(u)$, where $u$ is a matrix valued function on $M_{r}\times S^1$.
Following Guillemin and Kazhdan in [@GK] we introduce the following first order differential operators $$\eta_{+},\eta_{-}:\Omega^{0}(SM,\mbox{\rm End}\,\pi^*E)\to
\Omega^{0}(SM,\mbox{\rm End}\,\pi^*E)$$ given by
$$\eta_{+}:=\frac{D_{X}-i\,D_{H}}{2}$$ $$\eta_{-}:=\frac{D_{X}+i\,D_{H}}{2},$$ where $H=[V,X]$. We recall the other two structure equations of the Riemannian metric: $X=-[V,H]$ and $[X,H]=KV$, where $K$ is the Gaussian curvature of the surface.
The next lemma describes the commutation relations between these operators.
We have $$\begin{aligned}
[-iD_{V},\eta_{+}]&=\eta_{+},\\
[-iD_{V},\eta_{-}]&=-\eta_{-},\\
[\eta_{+},\eta_{-}]&=\frac{i}{2}\left(K\,D_{V}+[\star F_{\nabla},\,\cdot]\right).\end{aligned}$$ \[commeta\]
In order to prove the lemma we need the following general preliminary observation: if $U$ and $W$ are vector fields in $SM$ then $$F_{D}(U,W)=[D_{U},D_{W}]-D_{[U,W]}.
\label{curva}$$ As noted before $$F_{D}=[\pi^*F_{\nabla},\,\cdot].
\label{culift}$$ Thus for any vector field $U$, $F_{D}(V,U)=0$ ($V$ is vertical) and hence $[D_{V},D_{U}]=D_{[V,U]}$. If we now take $U=X,H$ and we use the commutation relations $[V,X]=H$ and $[V,H]=-X$ we obtain $[D_{V},D_{X}]=D_{H}$ and $[D_{V},D_{H}]=-D_{X}$. The first two commutation relations in the lemma follow easily from this. To prove the third relation note that $2[\eta_{+},\eta_{-}]=i[D_{X},D_{H}]$. Using (\[curva\]) and (\[culift\]) together with $[X,H]=K\,V$ we see that $[D_{X},D_{H}]=F_{D}(X,H)+K\,D_{V}=
[\star F_{\nabla},\cdot]+K\,D_{V}$ and the third commutation relation follows.
Let us set $\Omega_{n}:=H_{n}\cap \Omega^{0}(SM,\mbox{\rm End}\,\pi^*E) $. The first two commutation relations in the lemma imply that $\eta_{+}:\Omega_{n}\to \Omega_{n+1}$ and $\eta_{-}:\Omega_{n}\to \Omega_{n-1}$. It also follows easily from the fact that $X$ and $H$ preserve $\mu$ and the definitions that $\eta_{+}^{*}=-\eta_{-}$ and $\eta_{-}^{*}=-\eta_{+}$. Indeed, like $D_{V}$, $D_{X}$ and $D_{H}$ are skew-Hermitian since both $X$ and $H$ preserve the Liouville measure $\mu$.
Modified operators
------------------
We will now modify the operators $\eta_{+}$ and $\eta_{-}$ to suit our purposes. Consider a second unitary connection $\nabla^{0}$ and write $\nabla^0=\nabla+A$, where $A\in \Omega^{1}(M,\mbox{\rm ad}\,E)$. We may regard $A$ and $\star A$ as elements of $\Omega^{0}(SM,\mbox{\rm ad}\,\pi^*E)$ and if we do so, then $D_{V}A=\star A$ since $(D_{V}A)(x,v)=A(x,iv)=(\star A)(x,v)$. This certainly implies that $D_{V}^{2}A=-A$. Having this in mind, we decompose $A$ as $A=A_{-1}+A_{1}$ where $$A_{1}:=\frac{A-iD_{V}A}{2}\in \Omega_{1},\;\;\;A_{-1}:=\frac{A+iD_{V}A}{2}\in \Omega_{-1}.$$ Observe that this decomposition corresponds precisely with the usual decomposition of 1-forms on a surface: $$\Omega^{1}(M,\mbox{\rm ad}\,E)\otimes \C=\Omega^{1,0}(M,\mbox{\rm ad}\,E)\oplus \Omega^{0,1}(M,\mbox{\rm ad}\,E),$$ given by the eigenvalues $\pm i$ of the Hodge star operator.
We now set $\mu_{+}:=\eta_{+}+A_{1}$ and $\mu_{-}:=\eta_{-}+A_{-1}$. It is straightforward to check that $\mu_{+}:\Omega_{n}\to \Omega_{n+1}$ and $\mu_{-}:\Omega_{n}\to \Omega_{n-1}$. It is also easy to check that $\mu_{+}^{*}=-\mu_{-}$ and $\mu_{-}^{*}=-\mu_{+}$.
We will need the following auxiliary lemma.
The following relation holds $$\frac{i}{2}\star (\nabla A+A\wedge A)=\eta_{+}(A_{-1})-\eta_{-}(A_{1})
+A_{1}A_{-1}-A_{-1}A_{1}.$$
\[auxiliar\]
Using the definitions we derive $$A_{1}A_{-1}-A_{-1}A_{1}=\frac{i}{2}(AD_{V}(A)-D_{V}(A)A),$$ $$\eta_{+}(A_{-1})-\eta_{-}(A_{1})=\frac{i}{2}(D_{X}D_{V}A-D_{H}A).$$ But it is easy to check that $$\star(A\wedge A)=AD_{V}(A)-D_{V}(A)A,$$ and since $$\begin{aligned}
\star(\nabla A)&=D_{X}(\pi^*A)(X,H)\\
&=D_{X}(\pi^*A(H))-D_{H}(\pi^*A(X))-\pi^*A([X,H])\\
&=D_{X}D_{V}A-D_{H}A,\end{aligned}$$ the lemma is proved.
The next lemma describes the commutator $[\mu_{+},\mu_{-}]$.
Given $u\in \Omega^{0}(SM,\mbox{\rm End}\,\pi^*E)$ we have $$[\mu_{+},\mu_{-}]u=\frac{i}{2}\left(K\,D_{V}u+(\star F_{\nabla^{0}})\,u-u\,
(\star F_{\nabla})\right).$$
Using the definitions we compute $$[\mu_{+},\mu_{-}]u=[\eta_{+},\eta_{-}]u+(\eta_{+}(A_{-1})-\eta_{-}(A_{1}))u
+(A_{1}A_{-1}-A_{-1}A_{1})u.$$ Using Lemmas \[commeta\] and \[auxiliar\] we obtain $$[\mu_{+},\mu_{-}]u=\frac{i}{2}\left(K\,D_{V}u+[\star F_{\nabla},u]\right)
+\frac{i}{2}\star(\nabla A+A\wedge A)u.$$ The lemma is now a consequence of the fact that $F_{\nabla^{0}}=F_{\nabla}+\nabla A +A\wedge A$.
Let $|\cdot|$ stand for the $L^2$-norm in $\Omega^{0}(SM,\mbox{\rm End}\,\pi^*E)$.
Given $u\in \Omega^{0}(SM,\mbox{\rm End}\,\pi^*E)$ we have $$|\mu_{+}u|^2=|\mu_{-}u|^2+\frac{i}{2}\left(\langle K\,D_{V}u,u\rangle
+\langle (\star F_{\nabla^{0}})u,u\rangle-\langle u(\star F_{\nabla}),u\rangle\right).$$ \[cor:1\]
Using that $\mu_{+}^{*}=-\mu_{-}$ and $\mu_{-}^{*}=-\mu_{+}$ we derive $$\begin{aligned}
|\mu_{+}u|^2&=\langle \mu_{+}u,\mu_{+}u\rangle\\
&=\langle (\mu_{+})^*\mu_{+}u,u\rangle\\
&=\langle -\mu_{-}\mu_{+}u,u\rangle\\
&=\langle -\mu_{+}\mu_{-}u,u\rangle+
\langle [\mu_{+},\mu_{-}]u,u\rangle\\
&=\langle (\mu_{-})^*\mu_{-}u,u\rangle+\langle [\mu_{+},\mu_{-}]u,u\rangle\\
&=|\mu_{-}u|^2+\langle [\mu_{+},\mu_{-}]u,u\rangle\end{aligned}$$ and the corollary follows from the previous lemma.
A distance between transparent connections
==========================================
Let $\nabla^1$ and $\nabla^2$ be two unitary connections. We may write $\nabla^2=\nabla^1+A$, where $A\in \Omega^{1}(M,\mbox{\rm ad}\,E)$.
If $\pi:SM\to M$ denotes the canonical projection, we obtain unitary connections on the pull-back bundle $\pi^{*}E$ which are related by $$\pi^*\nabla^{2}=\pi^*\nabla^{1}+\pi^* A,$$ where $\pi^*A\in \Omega^{1}(SM,\mbox{\rm ad}\,\pi^*E)$. These connections on $\pi^*E$ induce in turn connections $D^1$ and $D^2$ on the bundle $\mbox{\rm End}\,\pi^*E$ of complex endomorphisms of $\pi^*E$ and are related by $$D^{2}=D^{1}+[\pi^* A,\,\cdot].$$
Suppose now that both $\nabla^1$ and $\nabla^2$ are transparent. As explained in Section \[tliv\], they induce smooth cocycles $C_1$ and $C_2$ on $\pi^*E$. By Proposition \[trivial\], $\pi^*E$ is trivial and via a unitary trivilization we may use the Livsic theorem \[livsic\] to conclude that there exists a smooth $u\in \Omega^{0}(SM,\mbox{\rm Aut}\,\pi^*E)$ such that $$C_{2}(x,v,t)=u(\phi_t(x,v))C_{1}(x,v,t)u^{-1}(x,v).$$ Take $\xi\in E_{x}$. Since $C_{1}(x,v,t)\xi$ (resp. $C_{2}(x,v,t)\xi$) is $\nabla^1$-parallel (resp. $\nabla^2$-parallel) along the geodesic determined by $(x,v)$, if we apply $\nabla^1$ to the previous equality and set $t=0$ we obtain at $(x,v)$: $$-A\xi=(D_{X}u)u^{-1}\xi$$ where $D:=D^1$, and since this holds for all $\xi$ we derive $$D_{X}u+Au=0.
\label{keyec}$$
The main result that we will prove about the solutions $u$ of (\[keyec\]) is that they have a [*finite*]{} Fourier expansion.
Given an element $u\in \Omega^{0}(SM, \mbox{\rm End}\,\pi^*E)$, we write $u=\sum_{m\in\Z}u_{m}$, where $u_m\in \Omega_m$. We will say that $u$ has degree $N$, if $N$ is the smallest non-negative integer such that $u_{m}=0$ for all $m$ with $|m|\geq N+1$.
Let $u\in \Omega^{0}(SM, \mbox{\rm End}\,\pi^*E)$ be a smooth solution to (\[keyec\]). Then $u$ has degree $N<\infty$. Moreover $N\leq l-1$ where $l$ is the smallest positive integer such that the Hermitian operators $$\mbox{\rm End}\,E_{x}\ni \alpha\mapsto -l\,K(x)\alpha
+(i\star F_{\nabla^{2}}(x))\alpha-\alpha(i\star F_{\nabla^1}(x)),$$ $$\mbox{\rm End}\, E_{x}\ni \alpha\mapsto -l\,K(x)\alpha
-(i\star F_{\nabla^{2}}(x))\alpha+\alpha(i\star F_{\nabla^1}(x))$$ are positive definite for all $x\in M$. \[thm:deg\]
Since $D_{X}=\eta_{+}+\eta_{-}$, equation (\[keyec\]) may be rewritten as $$\mu_{+}(u)+\mu_{-}(u)=0.$$ Projecting onto $\Omega_m$-components we obtain $$\mu_{+}(u_{m-1})+\mu_{-}(u_{m+1})=0
\label{eq:mu}$$ for all $m\in \Z$. Since $K<0$, there exists a positive integer $l$ such that the Hermitian operators $$u\mapsto -l\,K u
+(i\star F_{\nabla^{2}})u-u(i\star F_{\nabla^1}),$$ $$u\mapsto -l\,K u
-(i\star F_{\nabla^{2}})u+u(i\star F_{\nabla^1})$$ are positive definite for all $x\in M$. Using Corollary \[cor:1\], we can find a constant $c>0$ such that $$|\mu_{+}(u_{m})|^2\geq |\mu_{-}(u_{m})|^2+c|u_{m}|^2
\label{ineq:1}$$ for all $m\geq l$. There is also a constant $d>0$ such that $$|\mu_{-}(u_{m})|^2\geq |\mu_{+}(u_{m})|^2+d|u_{m}|^2
\label{ineq:2}$$ for all $m\leq -l$. Combining (\[eq:mu\]) and (\[ineq:1\]) we obtain $$|\mu_{+}(u_{m+1})|\geq |\mu_{+}(u_{m-1})|
\label{ineq:3}$$ for all $m\geq l-1$. Similarly, it follows from (\[eq:mu\]) and (\[ineq:2\]) that $$|\mu_{-}(u_{m-1})|\geq |\mu_{-}(u_{m+1})|
\label{ineq:4}$$ for all $m\geq -l+1$. Since the function $u$ is smooth, $\mu_{+}(u_{m})$ must tend to zero in the $L^2$-topology as $m\to\infty$. It follow from (\[ineq:3\]) that $\mu_{+}(u_{m})=0$ for $m\geq l-2$. However, (\[ineq:1\]) implies that $\mu_{+}$ is injective for $m\geq l$ and thus $u_{m}=0$ for $m\geq l$. Similarly, using (\[ineq:2\]) and (\[ineq:4\]) we deduce that $u_{m}=0$ for $m\leq -l$. This shows that $u$ has finite degree $N\leq l-1$.
Let $\T$ denote the space of transparent connections modulo gauge transformations. Using the previous theorem we can introduce a distance function on $\T$ as follows. Given $[\nabla^1],[\nabla^{2}]\in\T$ we set $\d([\nabla^1],[\nabla^2]):=N$, where $N$ is the smallest degree of $u\in \Omega^{0}(SM,\mbox{\rm Aut}\,\pi^*E)$ which solves $D_{X}u+Au=0$. It is easy to check that this definition does not depend on the chosen representatives as we now explain. Let $\varphi, \psi \in \Omega^{0}(M,\mbox{\rm Aut}\,E)$ and write $\varphi^{*}\nabla^2=\psi^*\nabla^1+\tilde{A}$. One checks that $$\tilde{A}=\varphi^{-1}A\varphi+\varphi^{-1}\nabla^{1}\varphi-\psi^{-1}\nabla^{1}\psi$$ and using this one also checks that $\varphi^{-1}u\psi$ solves $(\tilde{D}_{X}+\tilde{A})(\cdot)=0$, where $\tilde{D}$ is the connection induced by $\psi^{*}\nabla^{1}$. Since $\varphi^{-1}u\psi$ has the same degree as $u$, $\d$ is well defined.
$\d([\nabla^1],[\nabla^2])$ defines a distance function on $\T$.
Suppose $\d([\nabla^1],[\nabla^2])=0$. This means that there exists $u\in \Omega^{0}(SM,\mbox{\rm Aut}\,\pi^*E)$ which solves $D^1_{X}u+Au=0$ and $D^1_{V}u=0$. But this last equality means that $u(x,v)$ is independent of $v$. Indeed, consider a unitary trivialization of $E$ over a neighbourhood $V$ of $x$ and write $\nabla^1=d+C$, where $C$ is a $\mathfrak{u}(n)$-valued 1-form on $V$. Then $$0=D^1_{V}u=V(u)+[\pi^{*}C(V),u]=V(u).$$ This implies that we may write $u=w\circ\pi$, where $w\in \Omega^{0}(M,\mbox{\rm Aut}\,E)$. But $$(D^1_{X}u)(x,v)=(D^1_{X}w\circ\pi)(x,v)=(\nabla^{1}_{d\pi(X)}w)(x)=\nabla^{1}_{v}w.$$ Thus $\nabla^{1}w+Aw=0$, which combined with $\nabla^{2}w=\nabla^{1}w+[A,w]$, implies that $w^{*}\nabla^2=\nabla^1$. Hence $[\nabla^{1}]=[\nabla^{2}]$ as desired.
To show that $\d$ is symmetric, it suffices to note that if $u$ solves $D_{X}^{1}u+Au=0$, then $u^{*}$ solves $D^{2}_{X}u^{*}-Au^{*}=0$ and that $u$ and $u^{*}$ have the same degree.
In order to prove the triangle inequality, consider $\nabla^2=\nabla^1+A$ with $D^{1}_{X}u+Au=0$, and $\nabla^{3}=\nabla^{1}+B$ with $D^{1}_{X}w+Bw=0$. Obviously $\nabla^3=\nabla^2+(B-A)$. Using that $D^2=D^1+[\pi^*A,\,\cdot]$ and that $D^{1}_{X}u^{*}=u^{*}A$, we compute $$\begin{aligned}
D_{X}^{2}(wu^{*})&=(D^{2}_{X}w)u^{*}+w(D_{X}^{2}u^{*})\\
&=(D^{1}_{X}w+Aw-wA)u^{*}+w(D^{1}_{X}u^{*}+Au^{*}-u^{*}A)\\
&=(A-B)wu^{*}\end{aligned}$$ and since $\deg(wu^{*})\leq \deg(u)+\deg(w)$ it follows that $\d([\nabla^{3}],[\nabla^{2}])\leq \d([\nabla^{3}],[\nabla^{1}])+\d([\nabla^{2}],[\nabla^{1}])$.
[*Proof of Theorem A.*]{} Let us apply Theorem \[thm:deg\] when $\nabla=\nabla^2$ and $\nabla^1$ is the trivial connection $d$. The hypothesis of $\pm i\star F_{\nabla}(x)-K(x)\,\mbox{\rm Id}$ being positive definite for all $x\in M$ implies that $\d([\nabla],[d])=0$. Thus $\nabla$ is gauge equivalent to the trivial connection.
Proof of Theorem B
==================
Levi-Civita ghosts
------------------
As in the introduction, let $K$ be the canonical line bundle and $K^s$ with $s\in \Z$ be its tensor powers (if $s=0$ we get the trivial bundle). The powers $K^s$ for $s\neq 0$ carry the Levi-Civita connection which we denote by $\nabla_{\ell}^{s}$. If $s=0$ we understand that this is the trivial connection. Given an $n$-tuple of integers $S:=(s_{1},\dots,s_{n})$, the connection $$\nabla_{\ell}^{S}:=\nabla_{\ell}^{s_{1}}\oplus\cdots \oplus\nabla_{\ell}^{s_{n}}$$ defines a transparent unitary connection on the bundle $E_{S}:=K^{s_{1}}\oplus\cdots\oplus K^{s_{n}}$. Clearly $c_{1}(E_{S})=(2\texttt{g}-2)(s_{1}+\dots+s_{n})$ and any complex vector bundle $E$ supporting a transparent connection is isomorphic to $E_S$ for $S$ such that $c_{1}(E)=c_{1}(E_{S})$.
Now let $E$ be a Hermitian vector bundle and consider a unitary isomorphism $\tau:E\to E_{S}$, where $S$ is such that $c_{1}(E)=c_{1}(E_{S})$. The unitary connection $\tau^*\nabla_{\ell}^S$ is a transparent connection on $E$ and its gauge equivalence class, denoted by $[S]$, is independent of $\tau$. Note that $[S_1]=[S_2]$ if and only if $S_1$ and $S_2$ coincide up to a permutation.
The next lemma will allows us to see these ghosts in a different form, more appropriate for our purposes.
Let $L$ be a $\mathfrak{u}(n)$-valued 1-form on $SM$. It defines a unitary connection $d_{L}:=d+L$ on the trivial bundle $SM\times \C^n$.
Suppose $L(X)=L(H)=0$ and $L(V)=c$, where $c\in \mathfrak{u}(n)$ is a constant matrix such that $e^{2\pi\,c}=\mbox{\rm Id}$. Let $is_{1},\dots,is_{n}$ be the eigenvalues of $-c$, where $s_{k}\in \Z$. Set $S=(s_{1},\dots,s_{n})$. Then, there exists a unitary trivialization $\psi:\pi^*E_{S}\to SM\times \C^n$ such that $\psi^{*}(d_{L})=\pi^{*}\nabla_{\ell}^S$. \[lemma:ghosts\]
Since $c\in \mathfrak{u}(n)$, we can find a matrix $a\in U(n)$ such that $-a^{-1}ca$ is a diagonal matrix with entries $is_{1},\dots,is_{n}$. Hence we might as well assume that $-c$ has already this diagonal form. It is now clear that it suffices to prove the lemma for the case $n=1$ and we let $s:=s_{1}\in \Z$. The case $s=0$ is obvious and we will prove the lemma for $s<0$ (the case $s>0$ is similar). If $s<0$, then $K^s$ is just $(TM)^{\otimes m}$, where $m=-s>0$. Given $v\in T_{x}M$, we let $v^{m}\in (TM)^{\otimes m} $ be the tensor product of $v$ with itself $m$ times.
We define a unitary trivialization $\psi:\pi^{*}(TM)^{\otimes m}\to SM\times \C$ as follows. Given $(x,v)\in SM$ and $w\in (T_{x}M)^{\otimes m}$ we set $\psi(x,v,w)=(x,v,\lambda)$, where $\lambda\in\C$ is the unique number such that $w=\lambda v^{m}$. Note that the Riemannian metric on $M$ determines the unitary structure on $(T_{x}M)^{\otimes m}$. The real 2-dimensional tangent space $T_{x}M$ carries the complex structure $iv$ that rotates a vector $v\in T_{x}M$ by $\pi/2$ according to the orientation of the surface. We will show that $(\psi^{-1})^*(\pi^{*}\nabla_{\ell}^{s})=d_{L}$.
Let $\xi \in\Omega^{0}(SM,\pi^*(T_{x}M)^{\otimes m})$ be the section given by $\xi(x,v)=v^m$. Consider a smooth function $f:SM\to \C$ and note that $(\psi^{-1})^{*}f=f\xi$. By the definition of the Levi-Civita connection $$(\pi^{*}\nabla_{\ell}^{s})_{X}(\xi)=(\pi^{*}\nabla_{\ell}^{s})_{H}(\xi)=0\,$$ and thus $$(\pi^{*}\nabla_{\ell}^{s})_{X}(f\xi)=X(f)\xi=(\psi^{-1})^{*}(d_{L,X}(f)),$$ $$(\pi^{*}\nabla_{\ell}^{s})_{H}(f\xi)=H(f)\xi=(\psi^{-1})^{*}(d_{L,H}(f)).$$ We finally check what happens on $V$. Note that for any affine connection $\nabla$ on $TM$ we have $\pi^*\nabla_{V}(v)=iv$. Using the definition of the induced connection on a tensor product we deduce $$(\pi^{*}\nabla_{\ell}^{s})_{V}(\xi)=mi\xi,$$ hence $$(\pi^{*}\nabla_{\ell}^s)_{V}(f\xi)=V(f)\xi+fmi\xi.$$ On the other hand $$d_{L,V}(f)=V(f)+imf$$ and the lemma follows.
The next lemma, like the previous one, does not need any curvature assumption; only that we are working on a surface which is not a torus. The relation of being $V$-cohomologous is an equivalence relation and given $f:SM\to \mathfrak{u}(n)$, let $[f]_{V}$ denote the class of $f$.
Let $c_{1},c_{2}\in \mathfrak{u}(n)$ be two matrices such that $e^{2\pi c_{k}}=\mbox{\rm Id}$ for $k=1,2$. Then $[c_{1}]_{V}=[c_{2}]_{V}$ if and only if $\mbox{\rm tr}(c_1)=\mbox{\rm tr}(c_2)$. \[lemma:traces\]
Suppose first that $\mbox{\rm tr}(c_1)=\mbox{\rm tr}(c_2)$. The matrix $c_k$ determines a bundle $E_{S_{k}}$ and let $\psi_{k}:\pi^*E_{S_{k}}\to SM\times \C^n$ be the unitary trivialization given by the previous lemma. By hypothesis, we may take a unitary isomorphism $\phi: E_{S_{1}}\to E_{S_{2}}$ and let $\rho: \pi^*E_{S_{1}}\to \pi^*E_{S_{2}}$ be the induced isomorphism, $\rho(x,v,\xi)=(x,v,\phi_{x}(\xi))$. Let us write $\varphi:=\psi_2\circ\rho\circ\psi_{1}^{-1}(x,v,a)=(x,v,w(x,v)a)$ where $w:SM\to U(n)$ and $a\in\C^n$. Let $G$ be the unique $\mathfrak{u}(n)$-valued 1-form on $SM$ such that $\psi_{1}^*(d_{G})=\rho^*\pi^*\nabla^{S_{2}}_{\ell}=\pi^*\phi^*\nabla^{S_{2}}_{\ell}$.
Write $\phi^*\nabla^{S_{2}}_{\ell}=\nabla_{\ell}^{S_{1}}+A$. Since $\pi^*\nabla_{\ell}^{S_{1}}=\psi_{1}^*(d_{L_{1}})$ we must have $G=L_{1}+\psi_{1}\,\pi^*A\,\psi_{1}^{-1}$ which gives $G(V)=L_{1}(V)=c_{1}$. But $\varphi^*(d_{L_{2}})=d_{G}$, that is, $G=w^{-1}dw+w^{-1}L_{2}w$. Applying the last equality to $V$ we derive $c_1=w^{-1}V(w)+w^{-1}c_{2}w$, i.e., $[c_{1}]_{V}=[c_{2}]_{V}$.
Suppose now that there is $w:SM\to U(n)$ such that $c_1=w^{-1}V(w)+w^{-1}c_{2}w$. Taking traces $$\mbox{\rm tr}(c_1)-\mbox{\rm tr}(c_2)=h^{-1}V(h),$$ where $h:=\det w:SM\to S^1$. Arguing as in the proof of Theorem \[thm:abelian\], the function $h$ gives rise to a real-valued closed 1-form in $SM$ given by $-ih^{-1}dh$. Since $\pi^*:H^1(M,\re)\to H^{1}(SM,\re)$ is an isomorphism (this follows easily from the Gysin sequence, since $M$ is not the 2-torus), there exists a closed 1-form $\omega$ in $M$ and a smooth function $f:SM\to\re$ such that $$-ih^{-1}dh=\pi^*\omega+df.$$ Applying this equality to $V$ we derive $$iV(f)=\mbox{\rm tr}(c_1)-\mbox{\rm tr}(c_2)$$ which clearly implies $\mbox{\rm tr}(c_1)=\mbox{\rm tr}(c_2)$.
The next lemma is not needed in what follows, but it illustrates the distance in $\mathcal T$.
Suppose $\sum_{k=1}^{n}s_k=0$, so that $\nabla^{S}_{\ell}$ induces a connection on the trivial bundle. Then $\d([S],[d])=\max|s_k|$, where $d$ is the trivial connection. \[lemma:dist\]
Let $\phi:E_{S}\to M\times \C^n$ be a unitary trivialization and let $\rho:\pi^{*}E_{S}\to SM\times \C^n$ be the induced unitary trivialization, $\rho(x,v,\xi)=(x,v,\phi_{x}(\xi))$. Let $A$ be the unique $\mathfrak{u}(n)$-valued 1-form on $M$ given by $\phi^*(d_{A})=\nabla_{\ell}^{S}$, where $d_{A}=d+A$. Of course, we also have $\rho^{*}(d_{\pi^{*}A})=\pi^{*}\nabla^{S}_{\ell}$. By Lemma \[lemma:ghosts\] there is a unitary trivialization $\psi: \pi^{*}E_{S}\to SM\times \C^n$ such that $\psi^{*}(d_{L})=\pi^{*}\nabla_{\ell}^{S}$. Hence $(\rho\psi^{-1})^{*}(d_{\pi^{*}A})=d_{L}$. In other words, if we write $\rho\psi^{-1}(x,v,\xi)=(x,v,u(x,v)\xi)$, where $u:SM\to U(n)$, then $L=u^{-1}du+u^{-1}\,\pi^{*}A\,u$. Since $L(X)=0$, $u$ solves $X(u)+Au=0$. An inspection of the construction of $\psi$ in Lemma \[lemma:ghosts\] reveals that $\psi$ has polynomial dependence on the velocities with degree given by $\max|s_{k}|$. It follows that $\deg(u)=\max|s_{k}|$ and thus $\d([S],[d])\leq \max|s_k|$. Finally note that equality must hold since if $w$ is another solution of $X(u)+Au=0$, then $u^{*}w$ must be constant. Indeed, a simple calculation shows that $X(u^{*}w)=0$ and the claim follows from the transitivity of the geodesic flow of $X$.
Proof of Theorem B
------------------
(Forward direction.) The matrix $c$ determines a bundle $E_S$ and by considering a unitary isomorphism $\tau:E\to E_S$ we may suppose $E=E_{S}$. Let $\nabla$ be a transparent connection on $E_{S}$ and let $C$ be its associated cocycle in $\pi^*E_{S}$. Let $\psi:\pi^*E_{S}\to SM\times \C^n$ be the unitary trivialization given by Lemma \[lemma:ghosts\].
Write $$\psi\,C(x,v,t)\,\psi^{-1}(x,v,a)=(\phi_{t}(x,v),D(x,v,t)a),$$ where $D:SM\times\re\to U(n)$ is a cocycle as in Definition \[def1\]. By the Livsic theorem \[livsic\] there exists a smooth function $u:SM\to U(n)$ such that $D(x,v,t)=u(\phi_{t}(x,v))u^{-1}(x,v)$. Let $\Gamma:\re \to SM$ be $\Gamma(t)=\phi_{t}(x,v)$. By the definition of $C$, $\Gamma^{*}\pi^*\nabla(t\mapsto C(x,v,t)\xi)=0$ for any $\xi\in E_{S}(x)$. Now let $G$ be the unique $\mathfrak{u}(n)$-valued 1-form on $SM$ such that $\psi^{*}(d_{G})=\pi^*\nabla$, where $d_{G}=d+G$. Then $\Gamma^{*}d_G(t\mapsto D(x,v,t)a)=0$ for all $a\in \C^n$. Equivalently $$\frac{dD}{dt}+G(X)D=0$$ and setting $t=0$, we obtain: $X(u)+G(X)u=0$.
As in the proof of Lemma \[lemma:traces\], write $\nabla=\nabla_{\ell}^S+A$. Since $\pi^*\nabla_{\ell}^S=\psi^*(d_{L})$ we must have $G=L+\psi\,\pi^*A\,\psi^{-1}$ which gives $G(V)=L(V)=c$.
Now let us set $B:=u^{-1}du+u^{-1}Gu$. Then $d_{G}$ and $d_{B}$ are gauge equivalent, but $B(X)=0$.
Since $F_{\pi^*\nabla}(\cdot\,,V)=0$, we must also have $F_{B}(\cdot\,,V)=0$. Using that $F_{B}=dB+B\wedge B$ and $B(X)=0$ we compute $$F_{B}(X,V)=dB(X,V)+[B(X),B(V)]=dB(X,V).$$ But $$dB(X,V)=XB(V)-VB(X)-B([X,V])=XB(V)+B(H),$$ hence $$B(H)=-XB(V).
\label{eq:cuXV}$$ We also compute $$F_{B}(H,V)=dB(H,V)+[B(H),B(V)],$$ and $$dB(H,V)=HB(V)-VB(H)-B([H,V])=HB(V)-VB(H),$$ hence $$HB(V)-VB(H)+[B(H),B(V)]=0.
\label{eq:cuHV}$$ Combining (\[eq:cuXV\]) and (\[eq:cuHV\]) we derive the following non-linear PDE for $f:=B(V)$ $$H(f)+VX(f)-[X(f),f]=0.
\label{eq:nlpde}$$ This is precisely equation (\[eq:keypde\]) in the Introduction and since $f=B(V)=u^{-1}V(u)+u^{-1}cu$ it follows that $f\in \mathcal H_{c}$. Note that $f$ is defined exclusively in terms of $u$ and $c$ and $u$ must solve $X(u)+G(X)u=0$. However, up to right multiplication by an element $q\in U(n)$, there is only one such solution. Indeed, if $w$ is another solution, then $X(u^*w)=0$ and by transitivity of the geodesic flow there is $q\in U(n)$ such that $w=uq$. This implies that $f$ is uniquely defined in $\mathcal H_{c}/U(n)$. To complete the correspondence in the forward direction, we must check that if we consider a connection gauge equivalent to $\nabla$ we obtain the same $f$. A connection $\nabla^1$ gauge equivalent to $\nabla$ determines a connection $d_{G_1}$ in $SM$ gauge equivalent to $d_{G}$. In other words, there is a smooth function $r:SM\to U(n)$ such that $G_1=r^{*}dr+r^{*}Gr$. But if $u$ solves $X(u)+G(X)u=0$, then $r^{*}u$ solves $X(w)+G_1(X)w=0$ (unique up to multiplication on the right by an element in $U(n)$). Next observe that $$G_1(V)=c=r^*V(r)+r^*cr$$ and $$f_{1}=u^*rV(r^*u)+u^*rcr^*u=u^*V(u)+u^*(rV(r^*)+rcr^*)u=f$$ thus obtaining a well defined map $\T\mapsto \mathcal H_{c}/U(n)$.
(Backward direction.) Suppose now that we have a solution $f$ of (\[eq:nlpde\]) such that there is $u:SM\to U(n)$ with $f=u^*V(u)+u^*cu$. Define a $\mathfrak u(n)$-valued 1-form $G$ on $SM$ by setting: $$\begin{aligned}
G(X)&=-X(u)u^*,\\
G(H)&=-uX(f)u^*-H(u)u^*,\\
G(V)&=c,\end{aligned}$$ and define an element $\A\in \Omega^{1}(SM,\mbox{\rm ad}\,\pi^*E_{S})$ by $$\A:=\psi^{-1}(G-L)\psi.$$ Clearly $\A(V)=0$ and we wish to show that there exists $A\in \Omega^{1}(M,\mbox{\rm ad}\,E_{S})$ such that $\A=\pi^*A$. For this, it suffices to show that $D_{V}\A(X)=\A(H)$ and $D_{V}\A(H)=-\A(X)$, where $D$ here stands for the connection induced by $\nabla_{\ell}^S$. Equivalently, using the unitary isomorphism $\psi$, we are required to show that $D_{V}^LG(X)=G(H)$ and $D_{V}G(H)=-G(X)$, where $D^L$ is induced by $d_{L}$. Explicitly this means $V(G(X))+[c,G(X)]=G(H)$ and $V(G(H))+[c,G(H)]=-G(X)$. Using the definition of $G(X)$, the structure equations of the metric and $uf=V(u)+cu$ we compute: $$\begin{aligned}
V(G(X))&=-VX(u)u^*-X(u)V(u^*)\\
&=-XV(u)u^*-H(u)u^*-X(u)V(u^*)\\
&=-X(uf-cu)u^*-H(u)u^*-X(u)(u^*c-fu^*)\\
&=-uX(f)u^*-H(u)u^*+[c,X(u)u^*]\\
&=G(H)-[c,G(X)].\end{aligned}$$ Similarly we compute (we omit some of the details) $$\begin{aligned}
V(G(H))&=-V(u)X(f)u^*-uVX(f)u^*-uX(f)V(u^*)-VH(u)u^*-H(u)V(u^*)\\
&=X(u)u^*+u([X(f),f]-VX(f)-H(f))u^*+[c,uX(f)u^*]+[c,H(u)u^*]\\
&=-G(X)-[c,G(H)]+u([X(f),f]-VX(f)-H(f))u^*.\\\end{aligned}$$ Thus, if $f$ satisfies equation (\[eq:nlpde\]) we have $V(G(H))+[c,G(H)]=-G(X)$ as desired. Since $\A=\pi^*A$, $\nabla:=\nabla_{\ell}^S+A$ defines a transparent connection on $E_S$. To complete the backward correspondence we need to discuss what happens when we have another solution $w$ to the equation $f=u^*V(u)+u^*cu$. In this case $w$ determines a connection $d_{G_{1}}$ on $SM$ and it is straightforward to check that if we let $r:=uw^*$, then $G_1=r^*dr+r^*Gr$. In other words, $d_{G}$ and $d_{G_{1}}$ are gauge equivalent in $SM$. It follows that $\pi^*\nabla$ and $\pi^*\nabla^1$ are gauge equivalent. But this implies that $\nabla$ and $\nabla^1$ are gauge equivalent. Indeed, suppose $\pi^*\nabla$ and $\pi^*\nabla^1$ are gauge equivalent via $\varphi\in \Omega^0(SM,\mbox{\rm Aut}\,\pi^*E_S)$, i.e. $\pi^{*}\nabla^1=\varphi^{-1}D\varphi+\pi^*\nabla$. Apply this to $V$ to obtain $D_{V}\varphi=0$, so $\varphi$ only depends on the base point $x$. Applying the equality to $X$ we deduce that there is $\varphi\in\Omega^0(M,\mbox{\rm Aut}\,E_S)$ such that $\nabla^1=\varphi^{-1}\nabla\varphi+\nabla$ and thus $\nabla^1$ and $\nabla$ are gauge equivalent. The proof of Theorem B is now complete.
[**Addendum to Theorem B.**]{} We claim that $\mbox{\rm tr}(f)=\mbox{\rm tr}(c)$ and thus $\mbox{\rm tr}(f)$ is constant and determined by the topology of $E$. Consider the transparent connections induced by $\nabla$ and $\nabla_{\ell}^S$ on the line bundle $\det E_{S}$. By Theorem \[thm:abelian\] they must be gauge equivalent; in other words there is a smooth function $g:M\to S^1$ such that $\mbox{\rm tr}(A)=dg/g$. Recall that $X(u)+G(X)u=0$ and $G(X)=\psi A_{x}(v)\psi^{-1}$. Hence $\mbox{\rm tr}(G(X))=\mbox{\rm tr}(A)=X(g)/g$. Since $X(\det u)=\det u\,\mbox{\rm tr}(u^{*}X(u))$ we derive $X(\det u)=\det u(-X(g)/g)$ and thus $X(g\det u)=0$. By transitivity of the geodesic flow $g\det u$ is a constant and hence $V(\det u)=0$. But this is equivalent to $\mbox{\rm tr}(u^*V(u))=0$. Since $f=u^*V(u)+u^* c u$, $\mbox{\rm tr}(f)=\mbox{\rm tr}(c)$ as desired.
[One can also compute the curvature $F_{B}(X,H)$ of the connection $d_B$ from the theorem. Using that $B(X)=0$ we derive: $$F_{B}(X,H)=dB(X,H)= XB(H)-B([X,H])=-X^{2}(f)-Kf.
\label{eq:ode}$$ Note that $F_{B}(X,H)$ is conjugate to $\star F_{\nabla}\circ\pi$ via a unitary trivialization. ]{} \[rem:curva\]
Transparent connections at distance one from the trivial connection
-------------------------------------------------------------------
Let $\nabla$ be a transparent connection on the trivial bundle with $\d([\nabla],[d])=1$. If we follow the proof of Theorem B, we see that in this case $\psi$ is the identity and if we write $\nabla=d+A$, then $G=\pi^*A$. Since $\d([\nabla],[d])=1$, there exists a smooth function $u:SM\to U(n)$ such that $u=u_{-1}+u_{0}+u_{1}$ and $X(u)+Au=0$. Also, $B=u^{-1}du+u^{-1}\,\pi^*A\, u$ and $f=B(V)=u^{*}V(u)=-V(u^{*})u$.
$f\in \Omega_0$.
By separating $X(u)+Au=0$ into even and odd parts we deduce $$X(u_0)+Au_{0}=0,$$ $$X(u_{-1}+u_{1})+A(u_{-1}+u_{1})=0.$$ These two equations yield $$X(u^{*}_{0}(u_{-1}+u_{1}))=u_{0}^{*}A(u_{-1}+u_{1})+u_{0}^{*}(-A)(u_{-1}+u_{1})=
0,$$ and since the geodesic flow is transitive $u^{*}_{0}(u_{-1}+u_{1})$ must be constant and thus $$u_{0}^{*}u_{1}=u_{0}^{*}u_{-1}=0.
\label{eq:u0u1}$$ Using the special form of $u$ we derive $$\begin{aligned}
f&=(u_{-1}^{*}+u_{0}^{*}+u_{1}^{*})(-iu_{-1}+iu_{1})\\
&=i(u^{*}_{-1}u_{1}-u_{-1}^{*}u_{-1}+u_{0}^{*}u_{1}-u_{0}^{*}u_{-1}
+u_{1}^{*}u_{1}-u_{1}^{*}u_{-1}).\\\end{aligned}$$ Using that $u^{*}u=\mbox{\rm Id}$ we see that the terms of degree $\pm 2$, $u_{-1}^{*}u_{1}\in \Omega_{2}$ and $u_{1}^{*}u_{-1}\in\Omega_{-2}$ must vanish. Using (\[eq:u0u1\]) we obtain $$f=i(u_{1}^{*}u_{1}-u_{-1}^{*}u_{-1})\in\Omega_{0}.$$
Suppose the Hermitian matrix $\pm i\star F_{\nabla}(x)-2K(x)\,\mbox{\rm Id}$ is positive definite for all $x\in M$. Then $f\in \Omega_0$.
This follows right away from the last lemma and Theorem \[thm:deg\] which implies that $\d([d],[\nabla])\leq 1$.
Since $f\in\Omega_0$ we can think of $f$ as a function which depends only on the base point $x$. Thus $X(f)(x,v)=df_{x}(v)$ and $H(f)(x,v)=df_{x}(iv)$. Since $VX(f)=XV(f)+H(f)=H(f)$, equation (\[eq:nlpde\]) gives $2H(f)=[X(f),f]$ and we can rewrite this in terms of matrix valued 1-forms as $$2\star df=[df,f].
\label{eq:pdeonm}$$
We discuss this equation in the next subsection. Note that if we wish $f$ to be $V$-cohomologous to a matrix $c$ as in Theorem B we must have $e^{2\pi f(x)}=\mbox{\rm Id}$.
Solutions to $2\star df=[df,f]$. {#sub:sol}
--------------------------------
If we let $A:=\frac{1}{2}\star df$, then $2\star df=[df,f]$ may be rewritten as $d_{A}f=0$, so $f$ is covariant constant relative to the connection $d_{A}$. This implies that $f$ only hits one adjoint orbit of the adjoint action of $U(n)$ on $\mathfrak{u}(n)$. To see that this is the case observe first that $d\mbox{\rm tr}(f^m)=\mbox{\rm tr}(d_{A}f^{m})=0$ for any $m$ and thus the eigenvalues of $f(x)$ must be constant (and belong to $i\Z$ if $e^{2\pi f(x)}=\mbox{\rm Id}$). Also, the multiplicities of the eigenvalues do not change with $x$. Indeed, let $\xi\in \C^n$ be an eigenvector of $f(x)$ with eigenvalue $\lambda$ and let $\gamma:[0,1]\to M$ be a curve connecting $x$ to $y$. Let $\xi(t)$ be the parallel transport of $\xi$ along $\gamma$. Since $f(\gamma(t))\xi(t)$ is also parallel ($d_{A}f=0$), it must equal $\lambda\xi(t)$ and thus $f(y)\xi(1)=\lambda \xi(1)$ which shows that parallel transport preserves the eigenspaces of $\lambda$.
Suppose now $f:M\to \mathfrak{su}(2)$. The discussion above implies that $f^{2}=-\lambda^2\mbox{\rm Id}$ for some constant $\lambda$. This implies that $df\,f=-f\,df$, so we rewrite $2\star df=[df,f]$ as $\star df=df\,f$. Applying $\star$ we derive $df=\lambda^2df$. Hence if $\lambda^2\neq 1$, $f$ must be constant. Let us suppose then that $f^{2}=-\mbox{\rm Id}$, so $f$ hits the adjoint orbit of $$\left(\begin{array}{cc}
i&0\\
0&-i\\
\end{array}\right)$$ which we denote by $\bf S$ and we identify with the 2-sphere. For $g\in \bf S$ and $X\in T_{g}\bf S$, let $J_{g}(X):=Xg$. Clearly $J_{g}^2=-\mbox{\rm Id}$, so $J_g$ is a complex structure in $\bf S$ and the equation $\star df=df\,f$ simply says $df_{x}(iv)=J_{f(x)}(df_{x}(v))$, i.e. $f:M\to\bf S$ is a holomorphic map.
We now wish to show that given such a map $f:M\to\bf S$, then $f$ is $V$-cohomologous to the zero matrix, that is, there exists $u:SM\to SU(2)$ such that $f=u^*V(u)$. This would show that $\mathcal H_{0}\cap\Omega_0$ can be identified with the set of holomorphic maps $f:M\to \mathbb \C\mathbb P^1$ as claimed in the introduction.
Consider a map $f:M\to \bf S$ and let $L(x)$ (resp. $U(x)$) be the eigenspace corresponding to the eigenvalue $i$ (resp. $-i$) of $f(x)$. We have an orthogonal decomposition $\C^{2}=L(x)\oplus U(x)$ for every $x\in M$. Consider sections $\alpha\in \Omega^{1,0}(M,\C)$ and $\beta\in \Omega^{1,0}(M,\mbox{\rm Hom}(L,U))=\Omega^{1,0}(M,L^*U)$ such that $|\alpha|^2+|\beta|^2=1$. Such pair of sections always exists; for example, we can choose a section $\tilde{\beta}$ with a finite number of isolated zeros and then choose $\tilde{\alpha}$ such that it does not vanish on the zeros of $\tilde{\beta}$. Then we set $\alpha:=\tilde{\alpha}/(|\tilde{\alpha}|^2+|\tilde{\beta}|^2)^{1/2}$ and $\beta:=\tilde{\beta}/(|\tilde{\alpha}|^2+|\tilde{\beta}|^2)^{1/2}$.
Note that $\bar{\alpha}\in \Omega^{0,1}(M,\C)$ and $\beta^*\in \Omega^{0,1}(M,\mbox{\rm Hom}(U,L))=\Omega^{0,1}(M,U^*L)$. Using the orthogonal decomposition we define $u:SM\to SU(2)$ by $$u(x,v)=\left(\begin{array}{cc}
\alpha(x,v)&\beta^*(x,v)\\
-\beta(x,v)&\bar{\alpha}(x,v)\\
\end{array}\right).$$ Clearly $u=u_{-1}+u_{1}$, where $$u_{1}=\left(\begin{array}{cc}
\alpha&0\\
-\beta&0\\
\end{array}\right)$$ and $$u_{-1}=\left(\begin{array}{cc}
0&\beta^*\\
0&\bar{\alpha}\\
\end{array}\right).$$ It is straightforward to check that $uf=V(u)$.
Combining the discussion above with Theorem B (and its addendum) we derive:
The set of transparent $U(2)$-connections modulo gauge transformations at distance one from the trivial connection is in 1-1 correspondence with holomorphic maps $f:M\to \C\mathbb P^1$ up to composition with an orientation preserving isometry of $\C\mathbb P^1$. \[cor:last\]
We can actually compute the distance $\d([A],[B])$ where $A$ and $B$ define transparent connections at distance one from the trivial connection. Let $u=u_{-1}+u_{1}$ solve $X(u)+Au=0$ and let $w=w_{-1}+w_{1}$ solve $X(w)+Bw=0$. Then $r=wu^*$ solves $X(r)+Br-rA=0$ or equivalently $D_{X}^{A}r+(B-A)r=0$. In fact it is easy to check using arguments already used before that $wqu^*$, where $q\in SU(2)$ is a constant matrix, are all the solutions of $X(r)+Br-rA=0$. Now observe that $wqu^*=w_{-1}qu^{*}_{-1}+w_{1}qu_{1}^*+w_{-1}qu_{1}^*+w_{1}qu^{*}_{-1}$. Thus $wqu^*$ has terms only of degree zero or $\pm 2$. It follows that $\d([A],[B])=2$ unless $[A]=[B]$. Hence the distance induced via Corollary \[cor:last\] on the space of holomorphic maps $f:M\to\C\mathbb P^1$ (modulo $SU(2)$) is just the discrete distance.
The energy estimates method {#sub:last}
---------------------------
In order to deal with equation (\[eq:nlpde\]) one may try to use the energy estimates method (the Pestov identity) in the case of matrix valued functions as done by L.B. Vertgeim [@V], Sharafutdinov [@Sha] and Finch and Uhlmann [@FU]. However in order to control the non-linear term given by the bracket in (\[eq:nlpde\]) one ends up requiring some assumption of smallness on the connection or its curvature.
In our case the relevant integral identity takes virtually the same form as in the case of complex valued functions; we omit its proof here which is a straightforward generalization of the case $n=1$, which may be found in the form below in [@SU Lemma 2.1]. Let $f:SM\to \mathbb M_{n}(\C)$ be a smooth function, where $\mathbb M_{n}(\C)$ denotes the set of $n\times n$ complex matrices. Then
$$2\int_{SM}\langle H(f),VX(f)\rangle\,d\mu=\int_{SM}|H(f)|^2\,d\mu+\int_{SM}|X(f)|^2\,d\mu-\int_{SM}K|V(f)|^2\,d\mu,$$ where $\langle A,B\rangle=\Re\,\mbox{\rm tr}(AB^*)$ for $A,B\in \mathbb M_{n}$. If $f$ satisfies equation (\[eq:nlpde\]), then $$2\int_{SM}\langle H(f),[X(f),f]\rangle\,d\mu=3\int_{SM}|H(f)|^2\,d\mu+\int_{SM}|X(f)|^2\,d\mu-\int_{SM}K|V(f)|^2\,d\mu.
\label{eq:ee}$$ The last equality gives right away that $f$ is constant if $n=1$. Indeed, if $K<0$ the right hand side of the equality is $\geq 0$ and the left hand side vanishes since the bracket must vanish. This implies $H(f)=X(f)=V(f)=0$ and thus $f$ is constant. For $n\geq 2$ it is not clear how to deal with the term in the left hand side for arbitrary $f$. Here is an attempt in the spirit of [@FU].
Using the Cauchy-Schwartz inequality we can estimate the left hand side of (\[eq:ee\]) by $$2\int_{SM}\langle H(f),[X(f),f]\rangle\,d\mu\leq 2 \max||f||\int_{SM}(|X(f)|^2+|H(f)|^2)\,d\mu,$$ where $||f||$ is the operator norm of $f$. Hence if $2\max ||f||\leq 1$, (\[eq:ee\]) gives when $K<0$ that $H(f)=V(f)=0$ and again $f$ must be constant ($X=-[V,H]$). One can now try to estimate $\max||f||$ using the curvature of $\nabla$ and Remark \[rem:curva\]. If for example $K=-1$ we can solve the ODE (\[eq:ode\]) explicitly as $$2f(x,v)=\int_{0}^{\infty}e^{-s}F_{B}(X,H)(\phi_{s}(x,v))\,ds+\int_{-\infty}^{0}e^{s}F_{B}(X,H)(\phi_{s}(x,v))\,ds.$$ Hence if the operator norm of $\star F_{\nabla}$ is everywhere $\leq 1/2$, so is $F_{B}(X,H)$, and then $\max ||f||\leq 1/2$. By the argument above $f$ must be constant. However, this seems to give a weaker result than Theorem A.
Finally we note that (\[eq:ee\]) shows that if $f$ is a solution of (\[eq:nlpde\]) which is also [*odd*]{} (i.e. $f(x,-v)=-f(x,v)$), then it must be identically zero. Indeed, in this case $H(f)$, $X(f)$ and $VX(f)$ are even functions, but $[X(f),f]$ is odd. It follows that $[X(f),f]=0$ and by (\[eq:ee\]), $f$ must be a constant, and thus identically zero.
[aa]{}
C. Anand, [*Ward’s solitons,*]{} Geom. Topol. [**1**]{} (1997) 9–20.
C.B. Croke, V.A. Sharafutdinov, [*Spectral rigidity of a negatively curved manifold,*]{} Topology [**37**]{} (1998) 1265–1273.
N.S. Dairbekov, G.P. Paternain, [*Entropy production in Gaussian thermostats,*]{} Comm. Math. Phys. [**269**]{} (2007) 533–543.
N.S. Dairbekov, V.A. Sharafutdinov, [*Some problems of integral geometry on Anosov manifolds,*]{} Ergodic Theory Dynam. Systems [**23**]{} (2003) 59–74.
G. Eskin, [*On non-abelian Radon transform,*]{} Russ. J. Math. Phys. [**11**]{} (2004) 391–408.
D. Finch, G. Uhlmann, [*The X-ray transform for a non-abelian connection in two dimensions,*]{} Inverse Problems [**17**]{} (2001) 695–701.
E.R. Goetze, R.J. Spatzier, [*On Livšic’s theorem, superrigidity, and Anosov actions of semisimple Lie groups,*]{} Duke Math. J. [**88**]{} (1997) 1–27.
V. Guillemin, D. Kazhdan, [*Some inverse spectral results for negatively curved 2-manifolds,*]{} Topology [**19**]{} (1980) 301–312.
A.N. Livsic, [*Certain properties of the homology of $Y$-systems,*]{} Mat. Zametki [**10**]{} (1971) 555–564.
A.N. Livsic, [*Cohomology of dynamical systems,* ]{} Izv. Akad. Nauk SSSR Ser. Mat. [**36**]{} (1972) 1296–1320.
L.J. Mason, [*Global anti-self dual Yang-Mills fields in split signature and their scattering,*]{} J. Reine Angew. Math. [**597**]{} (2006) 105–133.
L.J. Mason, [*Lecture at Cambridge, November 2007*]{}.
V. Niţică, A. Török, [*Regularity of the transfer map for cohomologous cocycles,*]{} Ergodic Theory Dynam. Systems [**18**]{} (1998) 1187–1209.
R. Novikov, [*On determination of a gauge field on $\re^d$ from its non-abelian Radon transform along oriented straight lines*]{}, J. Inst. Math. Jussieu [**1**]{} (2002) 559–629.
V.A. Sharafutdinov, [*On an inverse problem of determining a connection on a vector bundle,*]{} J. Inverse and Ill-Posed Problems [**8**]{} (2000) 51–88.
V.A. Sharafutdinov, G. Uhlmann, [*On deformation boundary rigidity and spectral rigidity of Riemannian surfaces with no focal points,*]{} J. Differential Geom. [**56**]{} (2000) 93–110.
R.S. Ward, [*Soliton solutions in an integrable chiral model in $2+1$ dimensions,*]{} J. Math. Phys. [**29**]{} (1988) 386–389.
R.S. Ward, [*Twistors, geometry, and integrable systems,*]{} The geometric universe (Oxford, 1996) 99–108 Oxford Univ. Press, Oxford, 1998.
L.B. Vertgeim, [*Integral geometry with a matrix weight, and a nonlinear problem of recovering matrices,*]{} Sov. Math.-Dokl. [**44**]{} (1992) 132–135.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Stellar Yields for Galactic Modeling Applications ([`SYGMA`]{}) code is an open-source module that models the chemical ejecta and feedback of simple stellar populations (SSPs). It is intended for use in hydrodynamical simulations and semi-analytic models of galactic chemical evolution. The module includes the enrichment from asymptotic giant branch (AGB) stars, massive stars, Type-Ia supernovae (SNe Ia) and compact binary mergers. An extensive and extendable stellar yields library includes the NuGrid yields with all elements and many isotopes up to Bi. Stellar feedback from mechanic and frequency-dependent radiative luminosities are computed based on NuGrid stellar models and their synthetic spectra. The module further allows for customizable initial-mass functions and SN Ia delay-time distributions to calculate time-dependent ejecta based on stellar yield input. A variety of r-process sites can be included. A comparison of SSP ejecta based on NuGrid yields with those from Portinari et al. (1998) and Marigo (2001) reveals up to a factor of 3.5 and 4.8 less C and N enrichment from AGB stars at low metallicity, a result we attribute to NuGrid’s modeling of hot-bottom burning. Different core-collapse supernova explosion and fallback prescriptions may lead to substantial variations for the accumulated ejecta of C, O and Si in the first $10^7\, \mathrm{yr}$ at $Z=0.001$. An online interface of the open-source [`SYGMA`]{} module enables interactive simulations, analysis and data extraction of the evolution of all species formed by the evolution of simple stellar populations.'
author:
- Christian Ritter
- Benoit Côté
- Falk Herwig
- 'Julio F. Navarro'
- 'Chris L. Fryer'
bibliography:
- 'apj-jour.bib'
- 'astro.bib'
title: 'SYGMA: Stellar Yields for Galactic Modeling Applications'
---
Code verification {#s.appendixveri}
=================
We compare [`SYGMA`]{} results calculated with the same yields as W09 (M01P98) with the W09 results. This serves two goals. The first is simply code verification. The second is to provide some estimate on the kinds of uncertainties that are introduced by small code design and implementation differences, which are in addition to the uncertainties in the yield input data. We choose for this comparison the widely used W09 work, but would expect to find similar outcomes when comparing with other SSP codes. As this appendix shows, the differences are small but not entirely insiginficant. Our response to these differences is to make our code public so that all code design and implementation details can be scrutinized, and changed if deemed appropriate.
We apply the same massive star model factors of 0.5, 2 and 0.5 as in W09, initial abundances from Table 1 of W09, and – as much as possible – the same chemical evolution parameters ([Table\[tab:ce\_parameter\]]{}). The initial mass range for which yields were ejected is not given W09 and we choose the range from ${\ensuremath{M_{\rm ZAMS}}}=0.8{\ensuremath{\, M_\odot}}$ to ${\ensuremath{M_{\rm ZAMS}}}=100{\ensuremath{\, M_\odot}}$ to match best Fig. 2 in W09.
{width="1.\textwidth"}
Another important parameter is the transition initial mass $M_{\rm
mass}$ that delineates white dwarf and supernova outcomes. It is not given in W09, but must be between ${\ensuremath{M_{\rm ZAMS}}}=7$ and $9{\ensuremath{\, M_\odot}}$ in the M01P98 set. The actual value of $M_{\rm mass}$ is still a matter of some debate [@poelarends:08; @doherty:15; @doherty:17; @jones:16]. For this section we have adopted a value ([Table\[tab:ce\_parameter\]]{}) that agrees best with the results shown in Fig. 2 of W09.
![Evolution of fraction of total ejecta for transition masses of ${\ensuremath{M_{\rm ZAMS}}}=7.5{\ensuremath{\, M_\odot}}$, ${\ensuremath{M_{\rm ZAMS}}}=8{\ensuremath{\, M_\odot}}$ and ${\ensuremath{M_{\rm ZAMS}}}=8.5{\ensuremath{\, M_\odot}}$. []{data-label="fig:ce_comparisons"}](fig_10.pdf){width="95.00000%"}
Overall the [`SYGMA`]{} and W09 SSP models agree well, but differences can be seen as low- and intermediate mass stars start to contributed – especially for C (up to $10\%$ differnce) and Fe (Figure \[fig:wiersma\_fig2\]). The choice of the transition mass $M_{\rm mass}$ determines the appearance of the AGB star ejecta and the drop in total massive star ejecta (Figure \[fig:ce\_comparisons\]). The N yields increase smoothly with initial mass which leads to a smooth increase of the SSP ejecta similar to W09. The differences in the C and N evolution could be due to different yield interpolation methods used in the initial mass transition region from AGB to massive stars.
Online availability {#s.appendixA}
===================
The [`SYGMA`]{} web interface allows to simulate, analyse, and extract SSP ejecta which includes all stable elements and many isotopes up to Bi. We introduce the yield sets and parameters which are available within the web interface. Yields for AGB stars and massive stars can be selected from the NuGrid sets NuGrid$_{\rm d/r/m}$ (Table \[tab:yieldcompilations\]) SNIa yields are from [@thielemann:03] and Pop III yields are from [@heger:10]. The available metallicities are $Z=0.02$, $0.01$, $0.006$, $0.001$ and $0.0001$, $0$. Yields are applied in the initial mass range from $1{\ensuremath{\, M_\odot}}$ to $30{\ensuremath{\, M_\odot}}$. Chemical evolution parameters such as IMF and SNIa DTD can be set.
SSP ejecta can be extracted in the form of tables which contain for each time step the fraction of elements and isotopes of choice. As an example parts of a table which contains the normalized mass of elements ejected over $10^{10}\, \mathrm{yr}$ by a SSP of $1{\ensuremath{\, M_\odot}}$ at $Z=0.02$ is presented in Table \[tab:abu\_table\].
[lccccccc]{} 1.000E+07 &3.921E-04 &1.391E-04 &2.118E-03 &4.179E-05 &1.268E-08 &7.086E-10 &3.273E-02\
\
1.000E+08 &1.208E-03 &8.412E-04 &5.497E-03 &1.025E-03 &3.713E-08 &3.192E-09 &1.642E-01\
\
1.000E+09 &2.629E-03 &1.212E-03 &7.025E-03 &1.301E-03 &5.625E-08 &7.224E-09 &2.793E-01\
\
1.000E+10 &3.257E-03 &1.408E-03 &8.052E-03 &1.655E-03 &6.211E-08 &9.051E-09 &3.691E-01\
The [`SYGMA`]{} code and the yield library can be accessed via [<http://nugrid.github.io/NuPyCEE>](http://nugrid.github.io/NuPyCEE/). We provide an online documentation based on SPHINX[^1], guides and teaching material in form of Jupyter notebooks. [`SYGMA`]{} web interface is accessible through the NuPyCEE web page and hosted on NuGrid’s Web Exploration of NuGrid Datasets: Interactive (WENDI) platform at <http://wendi.nugridstars.org>. WENDI is a [`Cyberhubs`]{} service [@2018ApJS..236....2H]. Access to the figures of this work are provided through WENDI. NuGrid’s stellar and nucleosynthesis data sets are available at <http://nugridstars.org/data-and-software/yields/set-1> and can be analyzed with WENDI.
[^1]: http://www.sphinx-doc.org
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we present in a topological way the construction of the orientable surface with only one end and infinite genus, called *The Infinite Loch Ness Monster*. In fact, we introduce a flat and hyperbolic construction of this surface. We discuss how the name of this surface has evolved and how it has been historically understood.'
address:
- 'Fundación Universitaria Konrad Lorenz. CP. 110231, Bogotá, Colombia.'
- 'Fundación Universitaria Konrad Lorenz. CP. 110231, Bogotá, Colombia.'
author:
- 'John A. Arredondo'
- Camilo Ramírez Maluendas
title: On the Infinite Loch Ness monster
---
Introduction {#Introduction}
============
The term Loch Ness Monster is well known around the world, specially in The Great Glen in the Scottish highlands, a rift valley which contains three important lochs for the region, called Lochy, Oich and Ness. The last one, people believe that a monster lives and lurks, baptized with the name of the loch. The existence of the monster is not farfetched, people say, taking into account that the Loch Ness is deeper than the North Sea and is very long, very narrow and has never been known to freeze (see Figure \[real-mons\]).
\[real-mons\]
*Image by xKirinARTZx, taken from devianart.com*
The earliest report of such a monster appeared in the Fifth century, and from that time different versions about the monster passed from generation to generation [@Ste]. A kind of modern interest in the monster was sparked by 1933 when George Spicer and his wife saw the monster crossing the road in front of their car. After that sighting, hundreds of different reports about the monster have been collected, including photos, portrayals and other descriptions. In spite of this evidence, without a body, a fossil or the monster in person, The Loch Ness Monster is only part of the folklore.
In a different context, in mathematics, the term *Loch Ness Monster* is well known, and not in the folklore. In number theory there is a family of functions called exponential sums, which in general take the form $$s(n)=\sum_{n=1}^N e^{2\pi i f(n)},$$ and for the special case in which $$f(n)=(ln (n))^4$$ the graph of the curve associated to that function is called *Loch Ness Monster*, dubbed to the curve by J. H. Loxton [@Lox], [@Lox1].
\[curve-mons\]
From view of the Kerékjártó’s theorem of classification of noncompact surfaces (*e.g.*, [@Ker], [@Ian]), the *Infinite Loch Ness Monster* is the name of the orientable surface which has infinite genus and only one end [@Val]. Simply, É. Ghys (see [@Ghy]) describes it as the orientable surface obtained from the Euclidean plane which is attached to an infinity of handles (see Figure \[Figure3\]). Or alternatively, from a geometric viewpoint one can think that the Infinite Loch Ness monster is the only orientable surface having infinitely many handles and only one way to go to infinity.
![*The Infinite Loch Ness monster.*[]{data-label="Figure3"}](LNM.pdf "fig:")\
In the seventies, the interest by several authors (*e.g.*, [@Sow], [@Ni], [@Can]) on the qualitative study in the noncompact leaves in foliations of closed manifolds had grown. Ongoing in this line of research considering closed 3-manifolds foliated by surfaces, A. Phillips and D. Sullivan proved that the quasi-isometry types of the surfaces well known as the *Jacob’s ladder*[^1], the *Infinite jail cell windows* [@Spiv p.24], and the *Infinite jangle gym* (see Figure \[Figure4\]) cannot occur in foliations of $S^3$, or in fact in orientable foliation of any manifold with second Betti number zero. Nevertheless, all these surfaces are diffeomorphic to the *Infinite Loch Ness monster* (see [@PSul]). Roughly speaking from the historical point of view, this nomenclature to this topological surface appeared published on *Leaves with Isolated ends in Foliated 3-Manifolds* ([@Can2 1977]), however the authors wedge the term *Infinite Loch Ness monster* to preliminary manuscript of [@PSul], which was published the following year. Under these evident, one can consider to A. Phillips and D. Sullivan as the *Infinite Loch Ness monster*’s parents.
-------------------------------------------------------------------------------------------------- -- ------------------------------------------------------------------------------------------------
![*Surfaces having only one end and infinite genus.*[]{data-label="Figure4"}](jacob.pdf "fig:") ![*Surfaces having only one end and infinite genus.*[]{data-label="Figure4"}](jail.pdf "fig:")
![*Surfaces having only one end and infinite genus.*[]{data-label="Figure4"}](jungle.pdf "fig:")
-------------------------------------------------------------------------------------------------- -- ------------------------------------------------------------------------------------------------
Perhaps the reader has found on the literature other names for this surface with infinite genus and only one end, for example, the *infinite-holed torus* (see [@Spiv p.23]). Figure \[Figure5\].
![*The infinite-holed torus.*[]{data-label="Figure5"}](lochness.pdf "fig:")\
The Infinite Loch Ness monster has also appeared in the area of Combinatory. Its arrival was in 1929 when J. P. Petrie told H. S. M. Coxeter that had found two new infinite regular polyhedra. As soon as J. P. Petrie begun to describe them and H. S. M. Coxeter understood this, the second pointed out a third possibility. Later they wrote a paper calling this mathematical objets the *skew polyhedra* [@Cox1], or also known today as the Coxeter-Petrie polyhedra. Indeed, they are topologically equivalent to the Infinite Loch Ness monster as shown in [@ARV]. Given that from a combinatory view one can think that skew polyhedra are multiple covers of the first three Platonic solids, J. H. Conway and *et. al.* [@Con p.333] called them the *multiplied tetrahedron*, the *multiplied cube*, and the *multiplied octahedron*, and denoted them $\mu T$, $\mu C$, and $\mu O$, respectively. See Figure \[Polyhedra\].
---------------------------------------------------------------------------------------- --------------------------------------------------------------------------------- -- -- --
 
---------------------------------------------------------------------------------------- --------------------------------------------------------------------------------- -- -- --
*Images by Tom Ruen, distributed under CC BY-SA 4.0.* \[Polyhedra\]
In billiards, an interesting area of Dynamical Systems, during 1936 the mathematicians *R. H. Fox* and *R. B. Kershner* [@Fox] (later, used it by A. B. Katok and A. N. Zemljakov [@KZ]) associated to each *billiard* $\phi_P$ coming from an Euclidian compact polygon $P\subseteq \mathbb{E}^2$ a surface $S_{P}$ with structure of translation, which they called *Ueberlagerungsfläche* and means *covered surface*, and a projection map $\pi_p: S_p\to \phi_P$ mapping each geodesic of $S_P$ onto a *billiard trajectory* of $\phi_P$ (see Table \[tabla1\] and Figure \[billar\]). Later, F. Valdez published a paper [@Val], which proved that the surface *Ueberlagerungsfläche* $S_P$ associated to the billiard $\phi_P$, being $P\subseteq \mathbb{E}^2$ a polygon with almost an interior vertex of the form $\lambda \pi$ such that $\lambda$ is a irrational number, is the Infinite Loch Ness monster.
$$\xymatrix{
& *+[F]{ P\subseteq \mathbb{E}^2 }\ar@(u,u)[dr]^{\text{ \,\, \emph{Ueberlagerungsfl\"ache}}} \ar@(u,u)[dl]_{Billiard}& \\
\phi_P & & S_P\ar[ll]_{\pi_P}
}$$
![*Billiard associated to a rectangle triangle.*[]{data-label="billar"}](octagono.pdf "fig:")\
Building the Infinite Loch Ness Monster {#Building}
=======================================
A tame Infinite Loch Ness Monster {#tame}
---------------------------------
An easy and simple way to get an Infinite Loch Ness monster from the Euclidean plane is using the operation well-known as the gluing straight segments. Actually, it consists of drawing two disjoint straight segments $l$ and $l^{'}$ of the same lengths on the Euclidean plane $\mathbb{E}^2$, then we cut along to $l$ and $l^{'}$ turns $\mathbb{E}^2$ into a surface with a boundary consisting of four straight segments (see Figure \[gluemarks\]).
![*Two straight segments on $\mathbb{E}^2$.*[]{data-label="gluemarks"}](pegarmarcas.pdf)
Finally, we glue this segments using translations to obtain a new surface $S,$ which is homeomorphic to the torus pictured by only one point (see Figure \[genus\]). The operation described above is called *gluing the straight segments* $l$ and $l^{'}$ [@RaVa].
----------------------------------------------------------------------------------- --------------------------------------------------------------------------------------
![*Gluing straight segments.*[]{data-label="genus"}](plano_con_un_asa.pdf "fig:") ![*Gluing straight segments.*[]{data-label="genus"}](torus_without_point.pdf "fig:")
*Gluing the two straight segments on $\mathbb{E}^2$.* *Torus pictured by only one point.*
----------------------------------------------------------------------------------- --------------------------------------------------------------------------------------
Note that to build a Loch Ness monster from the Euclidian plane using the gluing straight segments is necessary to draw on it a countable family of straight segment and suitable glue them. It means, we consider $\mathbb{E}^2$ a copy of the Euclidean plane equipped with a fixed origin $\overline{0}$ and an orthogonal basis $\beta= \{e_{1},e_{2}\}$. On $\mathbb{E}^2$ we draw[^2] the countable family of straight segments following: $$\mathcal{L}:= \{l_{i} =((4i-1)e_{1}, \, 4ie_{1}) : \forall i \in \mathbb{N}\} \text{ (see Figure \ref{glue})}.$$
![*Countable family of straight segments $\mathcal{L}$.*[]{data-label="glue"}](marka.pdf)
Now, we cut $\mathbb{E}^2$ along $l_{i}$, for each $i\in \mathbb{N}$, which turns $\mathbb{E}^2$ into a surface with boundary consisting of infinite straight segments. Then, we glue the straight segments $l_{2i-1}$ and $l_{2i}$ as above (see Figure \[Figure3\]). Hence, the surface $S$ comes from the Euclidean plane attached to an infinitely many handles, which appear gluing the countable disjoint straight segments belonged to the family $\mathcal{L}$. In other words, the mathematical object $S$ is the Infinite Loch Ness monster.
From view of differential geometric, the surface $S$ is conformed by two kind of points. The set of *flat points* conformed by all points in $S$ except the ends of the straight segments $l_i$, for every $i\in \mathbb{N}$. To each one of this elements there exist an open isometric to some neighborhood of the Euclidean plane. Since the curvature is invariant under isometries then the curvature in the flat points is equal to zero. The other ones, are called *singular points*, in this case they are the end points of the straight segments $l_i$, for each $i\in\mathbb{N}$. Their respective neighborhood is isometric to cyclic branched covering $2:1$ of the disk in the the Euclidean plane, *i.e.*, they are *cone angle singularity of angle* $4\pi$ (see Figure \[cone\]). The surfaces having this kind of structure are known as *tame translation surfaces* (see *e.g.*, [@PSV]).
![*Cone angle singularity of angle $4\pi$.*[]{data-label="cone"}](punto_conico.pdf)
Hyperbolic Infinite Loch Ness Monster {#hyperbolic}
-------------------------------------
An application of the Uniformization Theorem (see *e.g.*, [@Abi], [@Muc]) ensures the existence of a subgroup $\Gamma$ of the isometries group of the hyperbolic plane $Isom(\mathbb{H})$ acting on the hyperbolic plane $\mathbb{H}$ performing the quotient space $\mathbb{H} / \Gamma$ in a hyperbolic surface homeomorphic to the Infinite Loch Ness monster. In other words, there exist a hyperbolic polygon $P\subseteq \mathbb{H}$, which is suitable identifying its sides by hyperbolic isometries to get the Infinite Loch Ness monster. An easy way to define the polygon $P$ is as follows[^3].
First, we consider the countable family conformed by the disjoint half-circles $\mathcal{C}=\{C_{4n}: n\in\mathbb{Z}\}$ having $C_{4n}$ center in $4n$ and radius equal to one, for every $n\in \mathbb{Z}$. See Figure \[circles\]. In other words, $C_{4n}:=\{z\in \mathbb{H}: |z- 4n|=1\}$
![*Family of half-circles $\mathcal{C}$.*[]{data-label="circles"}](circulos.pdf)
Removing the half-circle $C_{4n}$ of the hyperbolic plane $\mathbb{H}$ we get two connected component, which are called the *inside* of $C_{4n}$ and the *outside* of $C_{4n}$, respectively (see Figure\[inside\]). They are denoted as $\check{C}_{4n}$ and $\hat{C}_{4n}$, respectively.
![*Inside and outside.*[]{data-label="inside"}](inner_outner.pdf)
Hence, our connected hyperbolic polygon $P\subseteq\mathbb{H}$ is the closure of the intersection of the outsides following (see Figure \[poligono\])
$$\label{eq:4}
P:=\overline{\bigcap\limits_{n\in\mathbb{Z}} \hat{C}_{4n}}=\bigcap\limits_{n\in\mathbb{Z}} \{z\in\mathbb{H}: |z-4n|\geq 1\}.$$
![*Family of half-circles $\mathcal{C}$ and hyperbolic polygon $P$.*[]{data-label="poligono"}](poligono.pdf)
The boundary of $P$ is conformed by the half-circle belonged to the family $\mathcal{C}$. Then for every $m\in\mathbb{Z}$ the hyperbolic geodesics $C_{4(4m)}$ and $C_{4(4m+2)}$ are identified as it is shown in Figure \[identificacion\] by some of the following Möbius transformations:
$$\label{eq:5}
\begin{array}{rl}
f_{m}(z) & :=\dfrac{(16m+8)z-(1+16m(16m+8))}{z-16m}\\
f_{m}^{-1}(z) & := \dfrac{-16mz+(1+16m(16m+8))}{-z+(16m+8)}.
\end{array}$$
![*Gluing the side of the hyperbolic polygon $P$.*[]{data-label="identificacion"}](Identificacion.pdf)
Analogously, the hyperbolic geodesics $C_{4(4m+1)}$ and $C_{4(4m+3)}$ are identified as it is shown in Figure \[identificacion\] by the Möbius transformations: $$\label{eq:6}
\begin{array}{rl}
g_{m}(z) & :=\dfrac{(16m+8)z-(1+(16m+4)(16m+8))}{z-(16m+4)}, \\
g_{m}^{-1}(z) & := \dfrac{-(16m+4)z+(1+(16m+4)(16m+8))}{-z+(16m+8)}.
\end{array}$$
Through the Möbius transformations above, the inside of the half-circle $C_{4(4m)}$ (the half-circle $C_{4(4m+1)}$, respectively) is send by the map $f_m(z)$ (the map $g_m(z)$, respectively) into the outside of the half-circle $C_{4(4m+2)}$ (the half-circle $C_{4(4m+3)}$, respectively). Furthermore, the outside of the half-circle $C_{4(4m)}$ (the half-circle $C_{4(4m+1)}$, respectively) is send by $f_{m}(z)$ (the map $g_m(z)$, respectively) into the inside of the half-circle $C_{4(4m+2)}$ (the half-circle $C_{4(4m+3)}$, respectively).
Hence, the hyperbolic surface $S$ get glued the side of the polygon $P$ is the Infinite Loch Ness Monster. From the polygon $P$ we deduce that noncompact quotient space $S$ comes whit a hyperbolic structure having infinite area. Fortunately, the identification defined above takes the pairwise disjoint straight segment in the boundary of $P$ performing into the only one end of the surface $S$.
![*Subregion $P_m$.*[]{data-label="subregion"}](subregion.pdf)
Furthermore, for each integer number $n\in\mathbb{Z}$ we consider the subregion $P_m\subseteq P$, which is gotten by the intersection of $P$ and the strip $\{z\in \mathbb{H}:4(4m)-2 <Re(z)< 4(4m+3)+2\}$ (See Figure \[subregion\]), then restricting to $P_m$ the identification defined above it is turned into a torus with one hole $S_m$ (see Figure \[subsurface\]), which is a subsurface of $S$. Given the elements of the countable family $\{S_m:m\in\mathbb{Z}\}$ are pair disjoint subsurfaces of $S$ then it performs infinite genus in the hyperbolic surface $S$. In other words, $S$ is the Infinite Loch Ness monster.
![*Topological subregion $P_m$ and torus with one hole $S_m$.*[]{data-label="subsurface"}](toro.pdf)
From the analytic point of view, we have built a Fuchsian subgroup $\Gamma$ of $PSL(2,\mathbb{Z})$, where $\Gamma$ is infinitely generated by the set of Möbius transformations $\{f_m(z), g_m(z), f^{-1}_m(z), g^{-1}_m(z): \text{ for all } m\in \mathbb{Z}\}$ (see equations \[eq:5\] and \[eq:6\]), having the subset $P\subseteq \mathbb{H}$ as fundamental domain[^4]. Then $\Gamma$ acts on the hyperbolic plane $\mathbb{H}$. Defining the subset $K\subseteq \mathbb{H}$ as follows, $$\label{eq:7}
K:=\{w\in\mathbb{H}: f(w)=w \text{ for any } f\in \Gamma-\{Id\}\}\subseteq\mathbb{H},$$ the Fuchsian group $\Gamma$ acts freely and properly discontinuously on the open subset $\mathbb{H}-K$. Hence, the quotient space $$\label{eq:8}
S:= (\mathbb{H}-K)/\Gamma$$ is a well-defined hyperbolic surface homeomorphic to the Infinite Loch Ness monster. Moreover, it follows from an application of the Uniformization Theorem that the fundamental group $\pi_1(S)$ of the Infinite Loch Ness monster is isomorphic to $\Gamma$.
[00]{}
[^1]: E. Ghys calls Jacob’s ladder to the surface with two ends and each ends having infinite genus (see [@Ghy]). However, M. Spivak calls this surface the doubly infinite-holed torus (see [@Spiv p.24])
[^2]: Straight segments are given by their ends points.
[^3]: The reader can also found in [@AR] a great variety of hyperbolic polygons that perform hyperbolic surfaces having infinite genus.
[^4]: To deepen in these topics we suggest to reader [@MB], [@KS].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The moduli spaces of trigonal curves of odd genus $g\geq5$ are proven to be rational.'
address: 'Graduate School of Mathematical Sciences, the University of Tokyo, 3-8-1 Komaba, Meguro-ku, Tokyo 153-8914, Japan'
author:
- Shouhei Ma
title: The rationality of the moduli spaces of trigonal curves of odd genus
---
Introduction {#sec: intro}
============
The object of this article is to prove the following.
\[main\] The moduli space $\mathcal{T}_g$ of trigonal curves of genus $g=2n+1$ with $n\geq2$ is rational.
By a *trigonal curve* we mean an irreducible smooth projective curve which admits a degree $3$ morphism to ${{{\mathbb P}}}^1$. A trigonal curve of genus $g\geq5$ has a unique $g_3^1$, so that the space $\mathcal{T}_g$ to be studied is regarded as a sublocus of $\mathcal{M}_g$, the moduli space of curves of genus $g$. Shepherd-Barron [@SB] proved the rationality of $\mathcal{T}_g$ for $g=4n+2$ with $n\geq1$. Hence the space $\mathcal{T}_g$ is rational possibly except when the genus $g$ is divisible by $4$. For the one lower gonality, Katsylo and Bogomolov [@Ka1], [@B-K] established the rationality of the moduli spaces of hyperelliptic curves.
The proof of Theorem \[main\] is based on the classical relation between trigonal curves and the Hirzebruch surfaces ${{\mathbb{F}}}_N={{{\mathbb P}}}({{\mathcal{O}_{{\mathbb P}^{1}}}}\oplus{{\mathcal{O}_{{\mathbb P}^{1}}}}(N))$. Recall that a canonically embedded trigonal curve $C\subset{{{\mathbb P}}}^{g-1}$ of genus $g\geq5$ lies on a unique rational normal scroll $S$. The scroll $S$ may obtained either as the intersection of quadrics containing $C$, or as the scroll swept out by the lines spanned by the fibers of the trigonal map. The surface $S$ is the image of a Hirzebruch surface ${{\mathbb{F}}}_N$ by a linear system $|{{\mathcal{O}_{\pi}}}(1)\otimes\pi^{\ast}{{\mathcal{O}_{{\mathbb P}^{1}}}}(a)|$, $a>0$, where $\pi:{{\mathbb{F}}}_N\to{{{\mathbb P}}}^1$ is the natural projection. The trigonal map of $C$ is the restriction of $\pi$. When $C$ is general in the moduli $\mathcal{T}_g$, we have $N=0$ or $1$ depending on whether $g$ is even or odd. Thus, if $L_{3,b}$ denotes the line bundle ${{\mathcal{O}_{\pi}}}(3)\otimes\pi^{\ast}{{\mathcal{O}_{{\mathbb P}^{1}}}}(b)$ on ${{\mathbb{F}}}_1$ with $g=2b+1$, we have the birational equivalence $$\mathcal{T}_g \sim |L_{3,b}|/{\operatorname{Aut}}({{\mathbb{F}}}_1).$$ Here $|L_{3,b}|/{\operatorname{Aut}}({{\mathbb{F}}}_1)$ stands for a rational quotient of the linear system $|L_{3,b}|$ by the algebraic group ${\operatorname{Aut}}({{\mathbb{F}}}_1)$. Then Theorem \[main\] is equivalent to the following assertion in invariant theory.
\[main’\] For the line bundle $L_{3,b}$ on the Hirzebruch surface ${{\mathbb{F}}}_1$ the quotient $|L_{3,b}|/{\operatorname{Aut}}({{\mathbb{F}}}_1)$ is rational for $b\geq2$.
The rest of this article is devoted to the proof of this theorem. In Section \[sec: symmetric product\] we construct an ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-equivariant map from $|L_{3,b}|$ to $S^b{{\mathbb{F}}}_1$, the symmetric product of ${{\mathbb{F}}}_1$, which plays crucial role in the proof. In Section \[sec: g>7\] the rationality for $g\geq9$ is established by using the rational normal curves. In Section \[sec: g<9\] the rationality of $\mathcal{T}_7$ and $\mathcal{T}_5$ is proved,
Throughout this article we work over the field of complex numbers. We denote by $\pi:{{\mathbb{F}}}_1\to{{{\mathbb P}}}^1$ the natural projection. The $(-1)$-curve on ${{\mathbb{F}}}_1$ is denoted by $\Sigma$. The line bundle ${{\mathcal{O}_{\pi}}}(a)\otimes\pi^{\ast}{{\mathcal{O}_{{\mathbb P}^{1}}}}(b)$ on ${{\mathbb{F}}}_1$ will be written as $L_{a,b}$. The bundle ${{\mathcal{O}_{\pi}}}(1)$ is the pullback of ${{\mathcal{O}_{{\mathbb P}^{2}}}}(1)$ by the blow-down ${{\mathbb{F}}}_1\to{{{\mathbb P}}}^2$.
Symmetric product of the Hirzebruch surface {#sec: symmetric product}
===========================================
Let ${{{\mathbb P}}}\mathcal{E}$ be the projective space bundle ${{{\mathbb P}}}\pi_{\ast}{{\mathcal{O}_{\pi}}}(2)$ on ${{{\mathbb P}}}^1$. The variety ${{{\mathbb P}}}\mathcal{E}$ parametrizes unordered pairs $q_++q_-$ of two points of ${{\mathbb{F}}}_1$ which lie on the same $\pi$-fiber. We have a rational map $$\label{map 1}
\varphi_1 : |L_{3,b}| \dashrightarrow S^b({{{\mathbb P}}}\mathcal{E}), \quad C\mapsto \sum_{i=1}^b(q_{i+}+q_{i-})$$ defined as follows. If $C|_{\Sigma}=p_1+\cdots+p_b$ and $F_i$ is the $\pi$-fiber passing $p_i$, we set $q_{i+}+q_{i-}=C|_{F_i}-p_i$. The map $\varphi_1$ is clearly ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-equivariant. Next we define a rational map $$\label{map 2}
\varphi_2 : S^b({{{\mathbb P}}}\mathcal{E}) \dashrightarrow S^b{{\mathbb{F}}}_1, \quad \sum_{i=1}^b(q_{i+}+q_{i-})\mapsto \sum_{i=1}^{b}q_i$$ as follows. If $F_i$ is the $\pi$-fiber passing $\{ q_{i+}, q_{i-} \}$ and $p_i=F_i\cap\Sigma$, there exists a unique involution $\iota_i$ of $F_i\simeq{{{\mathbb P}}}^1$ which fixes $p_i$ and interchanges $q_{i+}$ and $q_{i-}$. Then we let $q_i\in F_i$ be the fixed point of $\iota_i$ other than $p_i$. By the uniqueness of $\iota_i$ the map $\varphi_2$ is ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-equivariant. We study the composition map $$\label{map 3}
\varphi = \varphi_2 \circ \varphi_1 : |L_{3,b}| \dashrightarrow S^b{{\mathbb{F}}}_1.$$
\[lemma 2.1\] The map $\varphi$ is dominant with a general fiber being an open set of a linear subspace of $|L_{3,b}|$.
For a general point $q_1+\cdots+q_b \in S^b{{\mathbb{F}}}_1$ let $F_i$ be the $\pi$-fiber passing $q_i$ and let $p_i=F_i\cap\Sigma$. We take an inhomogeneous coordinate $x_i$ of $F_i\simeq{{{\mathbb P}}}^1$ in which $p_i$ is $\{ x_i=0 \}$ and $q_i$ is $\{ x_i=\infty \}$. The involution of $F_i$ fixing $p_i$ and $q_i$ is given by $x_i\mapsto -x_i$. A smooth curve $C\in |L_{3,b}|$ is contained in $\varphi^{-1}(q_1+\cdots+q_b)$ if and only if $C|_{F_i}$ has the equation $x_i(\alpha_ix_i^2+\beta_i)=0$ for each $i=1,\cdots, b$. Since these are $2b$ linear conditions on $|L_{3,b}|$, namely the vanishing of the coefficient of $x_i^2$ and the constant term for $C|_{F_i}$, the second assertion is proved. The dominancy of $\varphi$ is a consequence of the dimension counting ${\dim}|L_{3,b}|=4b+9>2b$.
\[lemma 2.2\] The group ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$ acts on $S^b{{\mathbb{F}}}_1$ almost freely if $b\geq4$.
First we treat the case $b\geq5$. If a general point $p_1+\cdots+p_b \in S^b{{\mathbb{F}}}_1$ is fixed by a $g\in{{{\operatorname{Aut}}(\mathbb{F}_1)}}$, then $g$ stabilizes a general $b\geq5$ point set of the $(-1)$-curve $\Sigma$ so that $g$ acts trivially on $\Sigma$. Hence $g$ fixes each $p_i$. As ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$ acts almost freely on $({{\mathbb{F}}}_1)^b$, it follows that $g={\rm id}$.
Next we study the case $b=4$. Let $f:{{\mathbb{F}}}_1\to{{{\mathbb P}}}^2$ be the blow-down. For a general $p_1+\cdots+p_4\in S^4{{\mathbb{F}}}_1$ there exists a unique smooth conic $Q$ passing $f(\Sigma)$ and $f(p_1), \cdots, f(p_4)$. Any $g\in{{{\operatorname{Aut}}(\mathbb{F}_1)}}$ fixing $p_1+\cdots+p_4$, regarded as an element of ${{{\rm PGL}}}_3$, preserves $Q$ and the five point set $f(\Sigma), f(p_1), \cdots, f(p_4)$ on it. Hence $g$ acts trivially on $Q$, which implies that $g={\rm id}$.
We shall apply the no-name lemma (see [@Do], and also [@C-G-R] for non-reductive groups) to the map $\varphi$ when $b\geq4$. For that we note the following.
\[linearization\] Every line bundle on ${{\mathbb{F}}}_1$ admits an ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-linearization.
We have canonical ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-linearizations on the bundles $K_{{{\mathbb{F}}}_1}=L_{-2,-1}$, $\pi^{\ast}K_{{{{\mathbb P}}}^1}=L_{0,-2}$, and $f^{\ast}K_{{{{\mathbb P}}}^2}=L_{-3,0}$ where $f:{{\mathbb{F}}}_1\to{{{\mathbb P}}}^2$ is the blow-down of $\Sigma$. These induce ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-linearizations on $L_{1,0}$ and $L_{0,1}$. Since ${\rm Pic}({{\mathbb{F}}}_1)$ is freely generated by $L_{1,0}$ and $L_{0,1}$, the lemma is proved.
By Lemma \[linearization\] the ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-action on $|L_{3,b}|$ is induced by an ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-representation on $H^0(L_{3,b})$. Then Lemma \[lemma 2.1\] shows that $|L_{3,b}|$ is ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-birational to the projectivization of an ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-linearized vector bundle on an open set of $S^b{{\mathbb{F}}}_1$. By Lemma \[lemma 2.2\] we may apply the no-name lemma to see the
\[prop 2.4\] For $b\geq4$ we have a birational equivalence $$|L_{3,b}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}} \sim {{{\mathbb P}}}^{2b+9}\times(S^b{{\mathbb{F}}}_1/{{{\operatorname{Aut}}(\mathbb{F}_1)}}).$$
Thus the rationality of $|L_{3,b}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}}$ for $b\geq4$ is reduced to a stable rationality of $S^b{{\mathbb{F}}}_1/{{{\operatorname{Aut}}(\mathbb{F}_1)}}$.
Projection of rational normal curve {#sec: g>7}
===================================
In this section we prove a stable rationality of the quotient $S^b{{\mathbb{F}}}_1/{{{\operatorname{Aut}}(\mathbb{F}_1)}}$ to derive Theorem \[main’\] for $b\geq4$. For an integer $d\geq0$ we consider the universal curve $f:\mathcal{H}_d \to |L_{1,d}|$ over the linear system $|L_{1,d}|$. The variety $\mathcal{H}_d$ is defined as a divisor on ${{\mathbb{F}}}_1\times|L_{1,d}|$, and $f$ is the restriction of the second projection ${{\mathbb{F}}}_1\times|L_{1,d}| \to |L_{1,d}|$. The bundle $L_{0,1}$ on ${{\mathbb{F}}}_1$ induces a relative hyperplane bundle for $f$ which we denote by $\mathcal{O}_f(1)$. Let $$\mathcal{H}_{d,b}={{{\mathbb P}}}f_{\ast}\mathcal{O}_f(b).$$ An open set of $\mathcal{H}_{d,b}$ parametrizes pairs $(H, q_1+\cdots+q_b)$ where $H\in|L_{1,d}|$ is smooth and $q_1, \cdots, q_b$ are $b$ points on $H$. Note that a smooth $H\in|L_{1,d}|$ is a section of $\pi$.
\[lemma 3.1\] For $4\leq b\leq2d+2$ we have a birational equivalence $$\label{eqn 3.1}
\mathcal{H}_{d,b}/{{{\operatorname{Aut}}(\mathbb{F}_1)}} \sim {{{\mathbb P}}}^{2d+2-b}\times(S^b{{\mathbb{F}}}_1/{{{\operatorname{Aut}}(\mathbb{F}_1)}}).$$
Consider the evaluation map $$\psi : \mathcal{H}_{d,b} \dashrightarrow S^b{{\mathbb{F}}}_1, \quad (H, q_1+\cdots+q_b) \mapsto q_1+\cdots+q_b.$$ The fiber $\psi^{-1}(q_1+\cdots+q_b)$ over a general $q_1+\cdots+q_b$ is an open set of the sub linear system of $|L_{1,d}|$ of curves passing $q_1, \cdots, q_b$. Since ${\dim}|L_{1,d}|=2d+2\geq b$, $\psi^{-1}(q_1+\cdots+q_b)$ is non-empty and of dimension $2d+2-b$. In particular, $\psi$ is dominant. Then we may apply the no-name lemma for $\psi$ as like the proof of Proposition \[prop 2.4\] to deduce the equivalence .
By a comparison of Proposition \[prop 2.4\] and Lemma \[lemma 3.1\], it suffices for the proof of Theorem \[main’\] for $b\geq4$ to show the rationality of $\mathcal{H}_{d,b}/{{{\operatorname{Aut}}(\mathbb{F}_1)}}$ for *one* $d$ in the range $b\leq 2d+2\leq 3b+9$. We begin with the
\[lemma 3.2\] For $d\geq5$ we have a birational equivalence $$\mathcal{H}_{d,b}/{{{\operatorname{Aut}}(\mathbb{F}_1)}} \sim {{{\mathbb P}}}^b\times(|L_{1,d}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}}).$$
This lemma is an application of the no-name method for the fibration $\mathcal{H}_{d,b} \to |L_{1,d}|$. Since the bundle $L_{0,1}$ on ${{\mathbb{F}}}_1$ admits an ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-linearization, so is the bundle $\mathcal{O}_f(1)$ on the universal curve $\mathcal{H}_d$. Hence the sheaf $f_{\ast}\mathcal{O}_f(b)$ on $|L_{1,d}|$ is ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-linearized. It remains to check the almost freeness of the ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-action on $|L_{1,d}|$ for $d\geq5$. For a general $H\in|L_{1,d}|$ the intersection $H\cap\Sigma$ is a general $d$ point set of $H\simeq {{{\mathbb P}}}^1$. If a $g\in {{{\operatorname{Aut}}(\mathbb{F}_1)}}$ stabilizes $H$, then we have $g(H\cap\Sigma)=H\cap\Sigma$ so that $g$ acts trivially on $H$. This is enough for concluding that $g={\rm id}$.
Blowing-down ${{\mathbb{F}}}_1$ to ${{{\mathbb P}}}^2$, we see that the quotient $|L_{1,d}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}}$ is birational to the ${{{\rm PGL}}}_3$-quotient of the space $\mathcal{X}_d$ of rational plane curves of degree $d+1$ having an ordinary $d$-fold point. Let $\widetilde{\mathcal{X}}_d$ be the space of morphisms $\phi:{{{\mathbb P}}}^1\to{{{\mathbb P}}}^2$ such that $\phi^{\ast}{{\mathcal{O}_{{\mathbb P}^{2}}}}(1)\simeq{{\mathcal{O}_{{\mathbb P}^{1}}}}(d+1)$ and $\phi({{{\mathbb P}}}^1)\in \mathcal{X}_d$. We have $$|L_{1,d}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}} \sim {{{\rm PGL}}}_2\backslash\widetilde{\mathcal{X}}_d/{{{\rm PGL}}}_3.$$
Let ${{{\mathbb P}}}V_{d+1}=|{{\mathcal{O}_{{\mathbb P}^{1}}}}(d+1)|^{\vee}$ and $\Gamma_{d+1}\subset {{{\mathbb P}}}V_{d+1}$ be the rational normal curve $\phi_0({{{\mathbb P}}}^1)$ where $\phi_0$ is the embedding associated to ${{\mathcal{O}_{{\mathbb P}^{1}}}}(d+1)$. Recall that every morphism $\phi:{{{\mathbb P}}}^1\to{{{\mathbb P}}}^2$ with $\phi^{\ast}{{\mathcal{O}_{{\mathbb P}^{2}}}}(1)\simeq{{\mathcal{O}_{{\mathbb P}^{1}}}}(d+1)$ is the composition of $(1)$ the isomorphism $\phi_0:{{{\mathbb P}}}^1 \to \Gamma_{d+1}$, $(2)$ the projection $\Gamma_{d+1}\to {{{\mathbb P}}}(V_{d+1}/W)$ from a $(d-2)$-plane ${{{\mathbb P}}}W\subset {{{\mathbb P}}}V_{d+1}$ which is disjoint from $\Gamma_{d+1}$, and $(3)$ an isomorphism ${{{\mathbb P}}}(V_{d+1}/W)\to {{{\mathbb P}}}^2$. The group ${{{\rm PGL}}}_3$ acts on $\widetilde{\mathcal{X}}_d$ by transformation of an isomorphism ${{{\mathbb P}}}(V_{d+1}/W)\to {{{\mathbb P}}}^2$. Hence the quotient $\widetilde{\mathcal{X}}_d/{{{\rm PGL}}}_3$ is naturally birational to the locus $\mathcal{Y}_d$ in the Grassmannian $\mathbf{G}(d-2, {{{\mathbb P}}}V_{d+1})$ consisting of $(d-2)$-planes ${{{\mathbb P}}}W$ such that (i) ${{{\mathbb P}}}W\cap\Gamma_{d+1}=\emptyset$ and (ii) there exists a $(d-1)$-plane ${{{\mathbb P}}}U$ containing ${{{\mathbb P}}}W$ with ${{{\mathbb P}}}U\cap\Gamma_{d+1}$ being a $d$ point set. For such a ${{{\mathbb P}}}W$, the $(d-1)$-plane ${{{\mathbb P}}}U$ is spanned by the point set ${{{\mathbb P}}}U\cap\Gamma_{d+1}$ because of the fact that any distinct $d$ points on a rational normal curve in ${{{\mathbb P}}}^{d+1}$ are linearly independent. Also ${{{\mathbb P}}}U$ is uniquely determined by ${{{\mathbb P}}}W$ for an irreducible plane curve of degree $d+1$ has at most one singularity of multiplicity $d$. These two facts imply that $\mathcal{Y}_d$ is identified with an open set of the locus $$\mathcal{Z}_d \subset \mathbf{G}(d-2, {{{\mathbb P}}}V_{d+1})\times|{{\mathcal{O}_{{\mathbb P}^{1}}}}(d)|$$ of pairs $({{{\mathbb P}}}W, \mathbf{p})$ such that $\mathbf{p}=p_1+\cdots+p_d$ is a distinct $d$ point set on ${{{\mathbb P}}}^1$ and ${{{\mathbb P}}}W$ is a hyperplane of the $(d-1)$-plane ${{{\mathbb P}}}U_{\mathbf{p}}=\langle \phi_0(p_1), \cdots, \phi_0(p_d)\rangle$. We arrived at the birational equivalence $$\mathcal{X}_d/{{{\rm PGL}}}_3 \sim \mathcal{Z}_d/{{{\rm PGL}}}_2.$$ Now we prove the
\[prop 3.3\] If $d\geq5$ is odd, the ${{{\rm PGL}}}_2$-quotient of $\mathcal{Z}_d$ is rational. Hence $|L_{1,d}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}}$ is rational too.
The morphism $$\label{eqn 3.5}
\mathcal{Z}_d \to |{{\mathcal{O}_{{\mathbb P}^{1}}}}(d)|, \quad ({{{\mathbb P}}}W, \mathbf{p})\mapsto \mathbf{p}$$ is dominant with the fiber over a general $\mathbf{p}$ being ${{{\mathbb P}}}U_{\mathbf{p}}^{\vee}$. The vector space $U_{\mathbf{p}}$ is a subspace of $V_{d+1}=H^0({{\mathcal{O}_{{\mathbb P}^{1}}}}(d+1))^{\vee}$. Since $d+1$ is even, the bundle ${{\mathcal{O}_{{\mathbb P}^{1}}}}(d+1)$ is ${{{\rm PGL}}}_2$-linearized so that the ${{{\rm PGL}}}_2$-action on ${{{\mathbb P}}}V_{d+1}$ is induced by a ${{{\rm PGL}}}_2$-representation on $V_{d+1}$. Therefore $\mathcal{Z}_d$ is ${{{\rm PGL}}}_2$-isomorphic to the projectivization of a ${{{\rm PGL}}}_2$-linearized vector bundle on an open set of $|{{\mathcal{O}_{{\mathbb P}^{1}}}}(d)|$. As ${{{\rm PGL}}}_2$ acts almost freely on $|{{\mathcal{O}_{{\mathbb P}^{1}}}}(d)|$, the no-name method applied to the fibration shows that $$\mathcal{Z}_d/{{{\rm PGL}}}_2 \sim {{{\mathbb P}}}^{d-1}\times(|{{\mathcal{O}_{{\mathbb P}^{1}}}}(d)|/{{{\rm PGL}}}_2).$$ The quotient $|{{\mathcal{O}_{{\mathbb P}^{1}}}}(d)|/{{{\rm PGL}}}_2$ is rational by Katsylo [@Ka1].
*Proof of Theorem \[main’\] for* $b\geq4$. We may take an odd $d\geq5$ in the range $b \leq 2d+2 \leq 3b+9$. By Proposition \[prop 2.4\], Lemma \[lemma 3.1\], and Lemma \[lemma 3.2\] we have $$|L_{3,b}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}} \sim {{{\mathbb P}}}^{4b+7-2d}\times(|L_{1,d}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}}).$$ Then $|L_{1,d}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}}$ is rational by Proposition \[prop 3.3\].
The case $g\leq7$ {#sec: g<9}
=================
The rationality of $\mathcal{T}_7$ {#ssec: g=7}
----------------------------------
We consider the ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-equivariant map $\varphi:|L_{3,3}|\dashrightarrow S^3{{\mathbb{F}}}_1$ defined in . The group ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$ acts almost transitively on $S^3{{\mathbb{F}}}_1$, with the stabilizer $G$ of a general point $q_1+q_2+q_3$ being isomorphic to $\frak{S}_3$ by the permutation action on the set $\{ q_1, q_2, q_3\}$. As proved in Lemma \[lemma 2.1\], the fiber $\varphi^{-1}(q_1+q_2+q_3)$ is an open set of a sub linear system ${{{\mathbb P}}}V\subset|L_{3,3}|$. Then by the slice method (see [@Do]) we have the birational equivalence $$|L_{3,3}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}} \sim {{{\mathbb P}}}V/G.$$ The $G$-action on ${{{\mathbb P}}}V$ is induced by a $G$-representation on $V$ because the bundle $L_{3,3}$ admits an ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-linearization. It is well-known that for any linear representation $V'$ of $\frak{S}_3$ the quotient ${{{\mathbb P}}}V'/\frak{S}_3$ is rational (apply the no-name method for the irreducible decomposition). Hence the quotient ${{{\mathbb P}}}V/G$ is rational, and Theorem \[main’\] is proved for $b=3$.
The rationality of $\mathcal{T}_5$ {#ssec: g=5}
----------------------------------
We consider the ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$-equivariant map $\varphi_1 : |L_{3,2}| \dashrightarrow S^2({{{\mathbb P}}}\mathcal{E})$ defined in .
\[lemma 4.1\] The group ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$ acts almost transitively on $S^2({{{\mathbb P}}}\mathcal{E})$ with the stabilizer $G$ of a general point $\mathbf{q}=(q_{1+}+q_{1-})+(q_{2+}+q_{2-})$ being isomorphic to $\frak{S}_2\ltimes(\frak{S}_2\times\frak{S}_2)$.
Since ${{{\operatorname{Aut}}(\mathbb{F}_1)}}$ and $S^2({{{\mathbb P}}}\mathcal{E})$ have the same dimention, it suffices to calculate the stabilizer $G$. If $p_{i\pm}\in{{{\mathbb P}}}^2$ is the image of $q_{i\pm}$ by the blow-down ${{\mathbb{F}}}_1\to{{{\mathbb P}}}^2$, the group $G$ is identified with the group of those $g\in{{{\rm PGL}}}_3$ such that for each $i=1, 2$ we have $g(\{ p_{i+}, p_{i-}\}) = \{ p_{j+}, p_{j-}\}$ for some $1\leq j\leq2$.
Let $F_i$ be the $\pi$-fiber passing $q_{i\pm}$ and let $p_i=F_i\cap\Sigma$. The fiber $\varphi_1^{-1}(\mathbf{q})$ is an open set of the sub linear system ${{{\mathbb P}}}V\subset|L_{3,2}|$ of curves passing $q_{1+}, \cdots, q_{2-}$ and $p_1, p_2$. Similarly as Section \[ssec: g=7\], the slice method applied to the map $\varphi_1$ implies that $$|L_{3,2}|/{{{\operatorname{Aut}}(\mathbb{F}_1)}} \sim {{{\mathbb P}}}V/G,$$ where the $G$-action on ${{{\mathbb P}}}V$ is induced by a $G$-representation on $V$. Let ${{{\mathbb P}}}W\subset{{{\mathbb P}}}V$ be the sub linear system defined by $${{{\mathbb P}}}W = 2F_1 + 2F_2 + 2\Sigma + |L_{1,0}|.$$ Since the group $G$ preserves the curves $F_1+F_2$ and $\Sigma$, the subspace ${{{\mathbb P}}}W$ is invariant under the $G$-action. Since $G$ is finite, we have a $G$-decomposition $V=W\oplus W^{\perp}$ where $W^{\perp}$ is a $G$-invariant subspace. The group $G$ acts almost freely on the linear system $|L_{1,0}|$. Hence we may apply the no-name lemma for the projection ${{{\mathbb P}}}V\dashrightarrow {{{\mathbb P}}}W$ from ${{{\mathbb P}}}W^{\perp}$ to see that $${{{\mathbb P}}}V/G \sim {{\mathbb{C}}}^9\times({{{\mathbb P}}}W/G).$$ The quotient ${{{\mathbb P}}}W/G$, being of dimension $2$, is rational by Castelnuovo’s theorem. This completes the proof of rationality of $\mathcal{T}_5$.
[99]{}
Bogomolov, F. A.; Katsylo, P. I. *Rationality of some quotient varieties.* Mat. Sb. (N.S.) **126**(**168**) (1985), 584–589.
Chernousov, V.; Gille, P.; Reichstein, Z. *Resolving $G$-torsors by abelian base extensions.* J. Algebra **296** (2006), 561–581.
Dolgachev, I. V. *Rationality of fields of invariants.* Algebraic geometry, Bowdoin, 1985, 3–16, Proc. Symp. Pure Math., **46**, Part 2, Amer. Math. Soc., Providence, 1987.
Katsylo, P. I. *Rationality of the moduli spaces of hyperelliptic curves.* Izv. Akad. Nauk SSSR. **48** (1984), 705–710.
Shepherd-Barron, N. I. *The rationality of certain spaces associated to trigonal curves.* Algebraic geometry, Bowdoin, 1985 (Brunswick, Maine, 1985), 165–171, Proc. Symp. Pure Math., 46, Part 1, Amer. Math. Soc., Providence, RI, 1987.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'M. Castro'
- 'T. Duarte'
- 'G. Pace'
- 'J. -D. do Nascimento Jr.'
date: 'Received : Accepted : '
title: 'Mass effect on the lithium abundance evolution of open clusters: Hyades, NGC 752, and M67'
---
Introduction {#sec:Intro}
============
Lithium ($^7$Li) is a fragile element that is destroyed at temperatures above $\sim$2.5 $\times~10^6$ K. The lithium depletion in stars depends on several factors such as mass, age, metallicity, rotation, magnetic fields, mass loss, and mixing mechanisms [e.g. @dm84; @dp97; @ventura98; @charbonneltalon05]. Abundance measurements of this element in open cluster stars allow us to investigate mixing in stellar interior and to put constraints on the non-convective mixing processes as a function of mass, age, rotation, and metallicity for coeval stars. To understand the mixing mechanisms and to explain the complex evolution of stellar chemical elements, we need to constrain the effect of mass. As a part of this complexity, several unexpected results related to the depletion of lithium in stars have been found and reveal our limited understanding of the physics acting in the stars’ interiors. One of these unexpected results concerns the measurements of lithium abundance in late-F and early G-type stars, showing empirical evidence that is in contradiction with the standard model predictions, which only include convection. For stars in these spectral classes, authors have shown lithium depletion during the main sequence [@BT86; @boesgaard87; @ph88]. However, convective zones of low-mass main sequence (MS) stars do not reach regions in the stellar interior, where the temperature is sufficiently high to drive lithium destruction [see e.g. @randich06 and references therein]. This depletion is mass-age dependent [@donascimento09; @melendez10; @pace12]. For stars of one solar mass, including the Sun and those stars, which are spectroscopically indistinguishable from the Sun called solar twins [@cayrel96], the extremely low lithium abundance of about one hundred times lower than its original value as measured in meteorites, represents a long-standing puzzle.
During the pre-main-sequence (PMS), studies showed that Li destruction in stars with mass that is equal or larger to the solar one should be weak [@martin93; @jones97]. @randich97 determined the lithium abundance in 28 stars of the 28 Myr old cluster IC 2602. They compared the lithium abundance distribution as a function of stellar mass of this sample with those of the $\alpha$ Persei (50 Myr old) and the Pleiades (70 Myr old) clusters, and showed that only latest-type stars present a significant lithium depletion during the PMS. Early-K and G-type stars have a slightly lower lithium abundance in the oldest clusters, and F-type stars do not present any significant lithium depletion.
According to the standard stellar models, lithium depletion should be a unique function of mass, age, and chemical composition. Stars with close effective temperature in a given cluster should have undergone the same amount of lithium depletion. In this context, lithium abundances have been studied extensively in open clusters of different ages and revealed a complex behavior with an important scattering [see e.g. @pinsonneault97 and references therein]. The star-to-star scatter in lithium abundance exists among solar-type stars of the $\sim$2 Gyr old cluster NGC 752, as well as in the solar age, solar metallicity open cluster M67 [e.g. @pasquini97; @jones99]. From the study of the open cluster NGC 752 by @hobbs86a, @ph88, and @balachandran95, and from the 4 Gyr old open cluster M67 studied by @hobbs86b, @spite87, and @garcialopez88, authors conclude about a significant depletion relative to younger systems that increases with age. Despite decades of observations in clusters and field stars, the mechanism that drives the evolution of the surface lithium abundance and its scattering in main-sequence field stars and stars belonging to open clusters is far from being fully understood, and the lithium depletion expected from theoretical models needs to be adjusted to reproduce the mass and age-dependence. As for the open cluster M67 [@pace12], we explore this mass-age dependence and the star-to-star scatter at the same mass and different ages, with particular attention to the quality of the spectroscopy behind the lithium abundance determination. In this work, we continue to homogenize previous studies on open clusters by obtaining a consistent set of temperatures and masses, and by applying corrections to the lithium abundances when needed, due to revised effective temperatures, by taking into account non-LTE effects on the lithium abundances. We analyze stars from the main-sequence to the red giant branch, comparing our models to the observed behavior of lithium abundance as a function of mass and age.
This paper is organized as follows. In Sect. \[sec:Obs\], we describe the working sample of the three clusters. In Sect. \[sec:Models\], we present details of our evolutionary stellar models, and in Sect. \[sec:cmd\] we construct the color-magnitude diagrams of the open clusters. In Sect. \[sec:predictions\], we discuss the evolution of the lithium abundance predicted from our models. In Sect. \[sec:comparisons\], we compare the models with the observations from the studied clusters. Finally, in Sects. \[sec:discussion\] and \[sec:conclusions\], we present our discussion and conclusions.
Working sample {#sec:Obs}
==============
Our working sample is composed of lithium non-LTE abundance measurements for 67 F-G dwarfs in the Hyades from @takeda13. @takeda13 re-studied 49 stars with lithium abundance measurements that are available in the literature [@randich07; @thorburn93; @soderblom90; @boe88; @BT86; @duncan83; @rebolo88], and added 19 more stars to their sample. This sample is currently the largest for lithium abundances of Hyades stars. We decided not to use the star HD 285252 of the original sample because its lithium abundance is very low and uncertain ($A({\rm Li}) = -0.29$). @takeda13 do not directly provide uncertainty in their lithium abundance measurements, but they do provide all necessary data in electronic format to compute it. The uncertainty on lithium abundance reported in Table \[tab:interphyades\] was obtained by adding its three components quadraticall , i.e. that originated from the error of temperature, gravity, and microturbolence, respectively. Each term was defined as $1/2(A(\mathrm{Li},P-\Delta P) - A(\mathrm{Li},P+\Delta P))$, where $P$ is the parameter considered (temperature, gravity, or microturbulence), $\Delta P$ its uncertainty according to @takeda13, which assumes $\Delta T = 100 \ K$, and the two other parameters were kept constant. The temperature is the only significant source of error for the lithium abundance.
Masses and stellar parameters of the 67 sample stars were computed for each star by taking the closest isochrone point to the stellar data-point on the color-magnitude diagram (see Sect. \[sec:cmd\]). The isochrone was chosen by imposing \[Fe/H\] = 0.13 [@paulson03] and selecting the best fitting age, which was found to be 0.79 Gyr. B-V colours and absolute V magnitude are taken from @takeda13, who in turn used photometry and parallaxes for individual Hyades members from the Hipparcos catalogue [@debruijne; @esa97]. Lithium abundances for the individual Hyades stars are directly derived by @takeda13, and do not take into account this adjustment on the stellar parameters, which was applied by projection on the isochrone.
Takeda’s lithium abundance measurements are already accurate and homogeneous. Revisiting them, as we did for NGC752 stars, would not change the conclusions of this paper, so we did not. The calculation of stellar parameters for Hyades stars was mainly intended to derive stellar masses, which we chose as an independent variable to examine the trend of lithium abundances, as we did for the other clusters.\
Equivalent width (EW) measurements of the lithium line at 6708 [Å]{} for NGC 752 members were taken from the following sources: @sestito04 [@hobbs86a; @ph88]. @psh88 give lithium abundances (mostly upper limits) for 11 giants, which were transformed into (upper limits of) EWs by reverting the LTE analysis using the @lind09 code. Since @psh88 do not mention what metallicty they adopted, we assumed a solar metallicity for the stars. There are three stars in common between the older works and that of @sestito04 In these cases, the value from the latter work is adopted. The mean difference between the EWs that is published in the two works for these three stars in common is used as an estimation of the errors on the EW for the other measurements given in @hobbs86a, because they are not provided by the authors. The stars in common are PLA701, PLA859, and PLA921 (H146, H207, and H229, respectively, in the numbering system by @heinemann26), and the couple of measurements are, respectively, 42 and 45, 26.2 and 7, and 54 and 63. The values are in m$\AA$, the result from @sestito04 being the first of each pair. The average difference, after rounding, is 10 m$\AA$, which we adopted as error on EW for all stars for which only the old measurements are available. The EWs collected were used to compute abundances with the code by K. Lind, employing the full non-LTE analysis and the stellar parameters computed, as described below.
Errors on lithium abundances were calculated by adding quadratically its two main components, i.e. that due to the equivalent width and that due to the temperature. For the former, we simply computed: $dA(\mathrm{Li})_{EW}=1/2(A(\mathrm{Li}, EW+\Delta_{EW})-A(\mathrm{Li}, EW-\Delta_{EW}))$, where temperature, gravity, and metallicity were kept constant. To estimate the error due to the uncertainty in the temperature, we computed each stellar temperature in two different ways. The most accurate value was drawn from the isochrone in the CMD, as described in more detail below. We will refer to it as $T_{\mathrm{iso}}$. The other value, $T_{\mathrm{phot}}$, was obtained from the photometric calibration of @casagrande10, or @kucinskas05 for the giants, and the photometry in the reference papers. We then compute, for each dwarf, the following: $dA(\mathrm{Li})_{T}=A(\mathrm{Li}, T_{\mathrm{iso}})-A(\mathrm{Li}, T_{\mathrm{phot}})$.
The mean cluster iron content estimates range from $[\mathrm{Fe/H}] = -0.09$ dex [@ht92 based on 8 dwarfs] to $[\mathrm{Fe/H}] = 0.08 \pm 0.04$ . The other two studies (@sestito04, based on 18 dwarfs, and @reddy, based on 4 giants) favour a substantially solar metallicity ($[\mathrm{Fe/H}] = 0.01 \pm 0.04$ and $[\mathrm{Fe/H}] = -0.02 \pm 0.05$, respectively). In conclusion, the range of possible values for the cluster metallicity is aproximately $[\mathrm{Fe/H}] = 0.0 \pm 0.1$.
As we did for the Hyades, we computed stellar parameters for each star (effective temperature and mass) from the isochrone point closest to the stellar data point on the color-magnitude diagram. We adopted the best fitting isochrone, with an age of 1.59 Gyr for $[\mathrm{Fe/H}] = 0.0$ (see Sect. \[sec:cmd\]).\
The stars of both samples of the Hyades and NGC 752 clusters were all supposed to be single stars when included in the reference studies. When we checked the data using the study of @vanleeuwen07, queried through WEBDA, it turned out that these samples were contaminated by binaries. The contamination is about 20% for the Hyades cluster (14 out of 67) and for NGC 752 (10 out of 46). All known binaries in both samples are identified in Tables \[tab:interphyades\] and \[tab:interpngc752\] and in Figs. \[fig:cmd\_hyades\], \[fig:cmd\_ngc752\], \[fig:li-mass\_hyades\], and \[fig:li-mass\_ngc752\].\
The data relative to lithium abundance for the M67 cluster, corrected for non-LTE effects using the grid of @lind09, were analyzed in @pace12, to whom we redirect the reader.
Stellar evolutionary models {#sec:Models}
===========================
We present here the models computed using the Toulouse-Geneva stellar evolution code (TGEC) [@huibonhoa08; @donascimento09]. We used the same input physics as in @pace12. In particular, the transport of chemical elements and angular momentum is driven by different processes, which we briefly describe hereafter.
In the outer layers, convection is modeled according to the @bohm58 formalism of the mixing-length theory. The mixing-length parameter $\alpha = l/H_p$, where $l$ is the characteristic mixing length and $H_p$ is the pressure scale height, is a free parameter in our models. Below the convective zone, we introduce a convective undershooting with a depth of 0.09 $H_p$ so that, for the Sun, the combined mixing reaches the depth deduced by helioseismology .
In the radiative zone, microscopic diffusion, which is the process of element segregation by gravitational and thermal diffusion [@eddington16; @chapman17], is treated by computing convergent series of a maxwellian distribution function, which is the solution at equilibrium of the Bolzmann equation for dilute collision-dominated plasmas, with sucessive approximations . For collisons between charged ions, we use the @paquette86 method, which introduces a screened coulomb potential. Microscopic diffusion is essential in stellar models to account for abundances anomalies in Ap and Am stars [@michaud70] and to improve the consistency with helioseismology [@cox89; @bahcall95; @cd96; @richard96; @turcotte98; @schlattl02].
When a star rotates, centrifugal effects modify the gravity equipotencials, which are no longer spherical. It induces a macroscopic circulation of matter in the radiative zone between polar and equadorial regions, the so-called meridional circulation [@eddington26; @sweet50]. @zahn92 suggested that the meridional flow induces a transport of angular momentum, creating shears which become unstable in the horizontal direction, while the vertical shears are stabilized by the density gradient. The coupling between the meridional advection and the horizontal turbulence leads to an anisotropic mixing of the chemical species, parameterized as an effective diffusion coefficient. showed that the nuclear-induced $\mu$-gradients slowly stabilize the circulation and expel it from the core toward the external layers. @vt03 show that this feedback effect, which is due to the $\mu$-gradients, strongly modifies the meridional circulation and meanwhile reduces the efficiency of diffusion. Following @zahn92, , and , the rotational mixing is computed by a diffusion coefficient that we write as
$$D_{turb} = \alpha_{turb} r | U_{\rm r} |$$
with
$$\alpha_{\rm turb} = C_{\rm v} + \frac{1}{30 C_{\rm h}}$$
where $C_{\rm v}$ and $C_{\rm h}$ are unknown parameters relating to the vertical and horizontal part of the shear-induced anisotropic turbulent coefficient that is due to the transport of angular momentum, respectively. According to the assumption of strong anisotropic turbulence, they must satisfy the condition: $C_{\rm v} << C_{\rm h}$. $\mathbf{U_{\rm r}}$ is the vertical velocity amplitude which includes a classical meridional circulation term, which is directly linked to the rotation velocity $\Omega$, a $\mu$-gradient-induced term, a term related to the time variations of the differential rotation, and a term that appears in the case of large horizontal turbulence [see @tv03b for more details].
At the bottom of the convective zone, the transition layer between the differential rotation of the convective zone and the radiative interior rotating as a solid body is called the tachocline. This shear layer undergoes a strong anisotropic turbulence, with much stronger viscous transport in the horizontal than in the vertical direction, reducing the differential rotation and inhibiting its spread deep inside the radiative interior . We assume that the absolute size of the tachocline is constant with time. @brun98 showed it can be modeled by an exponential diffusion coefficient added to the turbulent diffusion coefficient. Thus, we use an effective diffusion coefficient:
$$D_{eff} = D_{turb} + D_{tacho}$$
with
$$D_{tacho} = D_{bcz} \exp \left( \ln 2 . \frac{r - r_{bcz}}{\Delta} \right)$$
where $D_{\rm bcz}$ and $r_{\rm bcz}$ are, respectively, the diffusion coefficient and the radius at the base of the convective zone, and $\Delta$ is the half width of the tachocline [@richard04]. This is the effective diffusion coefficient that appears in the equation of the chemical transport for the mean concentration of the different species.\
We calibrated our models with the Sun as in @richard04. There is no a priori reason why the calibration of the parameters for the model of 1.00 [$\mathrm{M_{\odot}}$]{} should also hold for different masses and evolutionary times but, given that no other standard calibrators are available, our approach is the best possible one for now. The calibration procedure of our models is the same as described in Sect. 4.2 in @pace12. The mixing-length parameter $\alpha$ and the initial helium abundance $Y_{\rm ini}$ enable us to calibrate the present evolutionary status of the Sun. By triangulation, we found the best couple $(\alpha,Y_{\rm ini})$ that allows us to reproduce the solar radius and the solar luminosity at the solar age [$L_{\odot} = 3.8515 \pm 0.0055 \times 10^{33}$ erg.s$^{-1}$ and $R_{\odot} = 6.95749 \pm 0.00241 \times 10^{10}$ cm at $t_{\odot} = 4.57 \pm 0.02$ Gyr @richard04]. For the best-fit solar model, we obtained $L = 3.8501 \times 10^{33}$ erg.s$^{-1}$ and $R = 6.95524 \times 10^{10}$ cm at an age $t = 4.57$ Gyr with $\alpha = 1.69$ and $Y_{\rm ini} = 0.268$. The rotation-induced mixing efficiency (meridional circulation and turbulent motions) is calibrated using the parameters $C_{\rm h}$ and $\alpha_{\rm turb}$. $C_{\rm v}$ is determined from these two parameters through Eq. 2. As explained in @richard04, the feedback currents, which are due to $\mu$-gradients, are strongly dependent on $C_{\rm h}
$, which is directly related to the horizontal turbulence. A strong horizontal turbulence tends to homogenize the horizontal layers and thereby smooths the horizontal $\mu$-gradients. In this case, the small induced $\mu$-currents may not be able to compensate for the currents because of classical meridional circulation and the mixing may remain efficient during a long timescale. In particular, we intend to put a mixing that is efficient and deep enough to smooth the diffusion-induced helium gradient, which lies below the surface convective zone, thus improving the agreement between the model and seismic sound-speed profiles. On the other hand, weak horizontal turbulence leads to important $\mu$-currents, which strongly reduce the mixing. In our models, this had to be weak and shallow enough to avoid the destruction of beryllium. We used the values $C_{\rm h}$ = 9000 and $\alpha_{\rm turb}$ = 1, and obtained an excellent agreement with helioseismology, which was more accurate than 1% for most of the stars in terms of sound velocity, except in the deep interior, where the discrepancy reaches 1.5%. We also obtained a slight destruction of beryllium by a factor of 1.17 with respect to the meteoritic value, which is well within the error in the determination of the solar beryllium abundance following @gs98. Beryllium destruction in low-mass stars will be subject of a further study. Finally, the tachocline is calibrated using the parameters $D_{\rm bcz}$ and $\Delta$ to obtain the solar lithium abundance at the solar age. Helioseismic constraints indicate that the solar tachocline thickness, defined as the region where the effective diffusion coefficient $D_{tacho}$ increases from 0.08 to 0.92 of its maximum value $D_{bcz}$, should be lower than 0.04 $R_{\rm \odot}$ [@charbonneau99], leading to a $\Delta$ lower than 0.01135 $R_{\rm \odot}$ [@richard04]. Using $D_{\rm bzc} = 2.02 \times 10^{5}$ m$^{2}$.s$^{-1}$ and $\Delta = 0.60 \times 10^{9}$ cm = 0.0086 $R_{\rm \odot}$, we obtained a surface lithium content relative to the initial one ${\rm Li}/{\rm Li}_0 = 0.60 \times 10^{-2}$. To obtain the lithium abundance $A({\rm Li})$, the question of the initial abundance on the ZAMS is important. Stars with different masses in the range 0.8 - 2.0 [$\mathrm{M_{\odot}}$]{}do not suffer the same lithium depletion during the PMS. In the Sun, this depletion should have been low and, therefore, we kept following the calibration procedure from @richard96 [@richard04], which neglects the lithium destruction during the PMS in solar models. Consequently, we adopted the meteoritic lithium abundance from @asplund09 $A({\rm Li})_0 = 3.26$ for the initial abundance, and obtained at the solar age $A({\rm Li}) = 1.04$, which agrees well with solar lithium abundance measurements from @asplund09 ($A({\rm Li}) = 1.05 \pm 0.10$).
The three steps of the calibration are not completely independent as the variations in the rotation-induced mixing parameters slightly modify the luminosity and radius of the model. Some iterations are necessary to obtain the best-fit solar model.\
![Color-magnitude diagram of the Hyades open cluster. Open circles are the whole sample of Hyades stars with magnitudes from the Hipparcos catalogue. Filled black circles are the sample of stars form @takeda13 for which we have a lithium abundance estimation. Blue diamonds represent the known binaries of this sample. The continuous line is the isochrone of 0.79 Gyr, calculated from our TGEC models with extra mixing, and the dotted line is the isochrone of 0.69 Gyr, calculated from standard models.[]{data-label="fig:cmd_hyades"}](./fig1.eps){width="9" height="9cm"}
We computed two grids of evolutionary models for each cluster, one grid of rotation-induced mixing, and another grid of standard models (with microscopic diffusion), to check the influence of the rotation-induced mixing on the lithium abundance depletion. For the Hyades, we computed models with a range of masses from 0.80 to 2.25 [$\mathrm{M_{\odot}}$]{} with a step of 0.01 [$\mathrm{M_{\odot}}$]{} and with a metallicity of ${\rm [Fe/H]} = 0.13$. For NGC 752, we computed models with masses from 0.80 to 1.77 [$\mathrm{M_{\odot}}$]{} with a step of 0.01 [$\mathrm{M_{\odot}}$]{} with a metallicity of ${\rm [Fe/H]} = 0.0$. Finally, for M67, our models have been computed with masses in the range 0.90 to 1.34 [$\mathrm{M_{\odot}}$]{}, with a step of 0.01 [$\mathrm{M_{\odot}}$]{}, and with a metallicity of ${\rm [Fe/H]} = 0.01$, which is a simple average of different estimates for the metallicity of M67 [see @pasquini08]. We ran the models from the zero age main sequence (ZAMS) to the top of the red giant branch (RGB) for the most massive stars. The input parameters for all the models are the same as for the 1.00 [$\mathrm{M_{\odot}}$]{} model. In particular, we used the same initial lithium abundance for all masses. For solar mass and larger, we can neglect lithium depletion during the PMS, and the choice of the meteoritic lithium abundance since initial abundance of the models is straightforward. For lower masses than the solar one, initial lithium abundance should be lower since the lithium depletion during the PMS increases with decreasing mass. However, it is a tricky task to choose a value for the lithium abundance at the ZAMS as a function of stellar mass. @dantona94 determined that, depending on the physics input of convection treatment and opacities, a model with 0.8 [$\mathrm{M_{\odot}}$]{} presents a lithium depleted of between 1.7 and 3.3 dex, at the age of $\alpha$ Persei. The smallest destruction ($\sim 1.7$ dex) is the most consistent with lithium abundance observations of $\alpha$ Per and corresponds to models with the @kurucz91 opacities and convection treatment. In this case, we chose to adopt the meteoritic lithium abundance as initial abundance as well, keeping in mind that, for these masses, it should be lower.
![Color-magnitude diagram of the open cluster NGC 752. Open circles are the whole sample of NGC 752 stars from @daniel94. Filled black circles are our working sample of stars for which we have a lithium abundance estimation. Blue diamonds represent the known binaries of this sample. The continuous line is the isochrone of 1.59 Gyr calculated from our TGEC models with extra mixing and the dotted line is the isochrone of 1.51 Gyr calculated from standard models.[]{data-label="fig:cmd_ngc752"}](./fig2.eps){width="9" height="9cm"}
Color-magnitude diagrams {#sec:cmd}
========================
Accurate determinations of cluster ages and the masses of their members are an essential requirement when comparing our models to the cluster observations and to study the evolution with stellar mass of lithium abundance in each cluster. The classic way to obtain both is to compare model isochrones with data points on the color-magnitude diagram, as in @pace12.
To construct the color-magnitude diagram of the open cluster Hyades presented in Figure \[fig:cmd\_hyades\], we used a reddening $E(B - V) = 0.001$ mag [@taylor06], a distance modulus $(m - M)_0 = 3.35$ mag [@mcArthur11; @perryman98], and a metallicity $[\mathrm{Fe/H}] = +0.13$ [@paulson03]. We plot the whole sample of Hyades stars with magnitudes from the Hipparcos catalogue along with our working sample of stars form @takeda13 for which we have lithium abundance estimation. Blue diamonds represent the known binaries of this sample. For the isochrones, we computed two grids of models, as presented in Sect. \[sec:Models\]. Matching the position of the turn-off hook of the isochrones with the left envelope of the observations, we estimated an age of 0.69 Gyr with the grid of standard models, and 0.79 Gyr with the grid of models with extra mixing, which falls right in between the classical estimation of 625 Myr [@perryman98] and that by @brandtHuang15 of 950 Myr. The data points that are found significantly far away from the isochrone could be due to photometric errors, non-members, or binaries. We can see that the known binaries are mostly off the isochrone. Also, a non-detected companion in a binary system would make the star appear more luminous and colder than it is, therefore a bias could exist in the CMD that moves some points up and/or to the left. However, in the case of the Hyades, this effect should be very small as the cluster has been very well studied.
The color-magnitude diagram for NGC 752 is presented in Figure \[fig:cmd\_ngc752\]. We built it using the grids of models exposed in Sect. \[sec:Models\]. Then, we shifted the resulting isochrone according to a reddening of $E(B - V) = 0.035$ mag and using a distance modulus of $(m - M)_0 = 8.25$ mag [@daniel94]. From the position of the turn-off hook of the isochrones, we estimated an age of 1.51 Gyr with the grid of standard models, and 1.59 Gyr with the grid of models with extra mixing. The positions of the turn-off hook and the giant branch of the isochrone depend on the input physics in the models such as metallicity, diffusion, and overshooting at the convective core. In particular, the cut-off of the diffusion when the model enters into the RGB or a bad calibration of the overshooting, could explain why the giants are off the isochrone.
For the open cluster, M67, see the paper @pace12, where model calibration and mass determination, a color-magnitude diagram, and an estimation of the cluster age of 3.87 Gyr have already been presented. In the present work, we added an isochrone that is made up of standard models to the color-magnitude diagram of M67 and found an age of 3.60 Gyr. The new plot, including this isochrone, is presented in Fig. \[fig:cmd\_m67\].
![Color-magnitude diagram of the open cluster M67. The black filled circles represent our work sample, i.e., stars with lithium abundance measurements. Stars with photometry available from [@yadav08], but with no lithium abundances, are displayed as open circles. The continuous isochrone from TGEC models with extra mixing corresponds to an age of 3.87 Gyr. The dotted line is the isochrone of 3.60 Gyr calculated from standard models.[]{data-label="fig:cmd_m67"}](./fig3.eps){width="9" height="9cm"}
Predicted lithium evolution as a function of mass and age {#sec:predictions}
=========================================================
In this section, we study the evolution of lithium abundance predicted by the models, as a function of mass in each cluster, and as a function of age, and we compare the evolution between the three clusters. In Fig. \[fig:li-mass\_models\], we present the lithium abundance as a function of mass for our model isochrones with ages that are estimated previously for each cluster. Continuous lines are the isochrones from models with extra mixing, whereas dotted lines are the isochrones from standard models. We can identify several interesting facts in this figure. Models with mass lower than $\sim1.10$ [$\mathrm{M_{\odot}}$]{}, including rotation-induced mixing, seem to destroy their lithium early in the MS, but depletion continues throughout the main sequence, with a lower rate after Hyades age though. For 1.00 [$\mathrm{M_{\odot}}$]{}, the lithium abundance difference between models of 1.59 Gyr for NGC 752 and models of 3.87 Gyr for M67 is about 0.5 dex. In this mass range, which is characterized by the presence of an extensive outer convective zone, lithium destruction is strongly dependent on stellar mass. We notice that the lithium depletion predicted at the Hyades age is greater than at the age of NGC 752 owing to the difference of metallicity between the models of the Hyades and the models of NGC 752. Indeed, the Hyades members are much more metallic than the NGC 752 members, and thus the models of the former have a deeper convective zone than members of the latter (see Fig. \[fig:rcz-mass\_models\]), leading to a larger and quicker lithium depletion. The mixing that occurs below that convective zone is mainly driven by rotation during the PMS and the beginning of the MS. However, surface lithium destruction during the PMS is significant only for the latest-type stars, and stars with masses around the solar mass present a slight lithium depletion at the ZAMS @randich97. In our models, PMS is not taken into account, but we calibrated our solar model to obtain the lithium abundance at the solar age, adopting the meteoritic lithium abundance of @asplund09 as initial abundance. This implies that we probably used a stronger mixing than would have been necessary with a lower initial abundance owing to the slight depletion during the PMS. For lower masses, which should present a lower lithium abundance at the ZAMS, caused by the lithium destruction during the PMS, we expect lower lithium abundance than calculated by our models at the ages of the three clusters, and thus a steeper isochrone between 0.80 and 1.10 [$\mathrm{M_{\odot}}$]{}.
![Lithium abundance as a function of mass for isochrones with ages estimated for each cluster. Contiunuous isochrones are from models with extra mixing and dotted isohrones from standard models. Magenta lines represent the isochrones for the Hyades with $[\mathrm{Fe/H}] = +0.13$ and ages of 0.79 Gyr and 0.69 Gyr, respectively; blue lines represent the isochrones for NGC 752 with $[\mathrm{Fe/H}] = +0.00$ and ages of 1.59 Gyr and 1.51 Gyr, respectively; and green lines represent the isochrones for M67 with $[\mathrm{Fe/H}] = +0.01$ and ages of 3.87 Gyr and 3.60 Gyr, respectively.[]{data-label="fig:li-mass_models"}](./fig4.eps){width="9" height="9"}
For models with extra mixing with masses between 1.10 and 1.30 [$\mathrm{M_{\odot}}$]{}, we observe the presence of a lithium plateau where the element is barely destroyed during evolution in the MS and independently of stellar mass. In these stars, the outer convective zone is very shallow and the turbulence is weak. The rotation-induced mixing below the convective zone does not sink deep enough to bring the lithium to the destruction layers. For these masses, lithium destruction during the PMS is negligible and it is consistent to use the same initial lithium abundance as for the solar model. Furthermore, a decrease in the mixing efficiency should not have any significant effect. The lithium abundance on this plateau seems to slightly decrease with time, at a rate of about 0.2 dex between Hyades age and M67 age.
When the mass increases, lithium abundance decreases with mass and age. This is due to the gravitational diffusion process below the convective zone. Lithium particles continuously sink below the convective zone owing to gravity and the amount of lithium in the convective zone decreases with time. For larger masses, the process is more efficient.
![Relative inner convective zone radius as a function of stellar mass for isochrones with ages estimated for each cluster. The line types and colors are the same as in Fig. \[fig:li-mass\_models\].[]{data-label="fig:rcz-mass_models"}](./fig5.eps){width="9" height="9"}
For the highest masses, we can see the dramatic decrease of lithium abundance with age due to the entry of the models in the red giant branch. When the star leaves the main sequence, the outer convective zone deepens during the first dredge-up and the lithium is diluted in lithium-free layers. In M67, this occurs in stars with masses around 1.30 [$\mathrm{M_{\odot}}$]{}, although in NGC 752 this process can be observed for masses larger than 1.70 [$\mathrm{M_{\odot}}$]{}.
Standard models have a totaly different behavior. For masses lower than about 1.25 [$\mathrm{M_{\odot}}$]{}, depletion of lithium abundance is very weak and is due to microscopic diffusion below the convective zone. When the mass increases, the outer convective zone shrinks and the diffusion is more efficient with time. For larger masses, the convective zone is very thin (see Fig. \[fig:rcz-mass\_models\]) and microscopic diffusion takes place on a timescale much smaller than the stellar life time on the main sequence. Surface lithium abundance is depleted very quickly. From comparison with models with extra mixing, we can deduce that the rotation-induced mixing prevents the lithium depeletion from microscopic diffusion in this range of masses, until the first dredge-up. These statements are long-standing results [see @schatzman69].
Figure \[fig:rcz-mass\_models\] shows the relative radius at the base of the convective zone along our three isochrones for the three open clusters and for the two sets of models as a function of stellar mass. The depth of the convection zone for the two sets of models is very similar, in contradiction with the completely different behaviour of the lithium depletion. However, in the case of the models with meridional circulation, the extra mixing lies below the base of the convective zone, increasing greatly the depth of the chemical transport. For low-mass stars, the outer convective zone becomes shallower when the mass increases. For mass lower than 1.10 [$\mathrm{M_{\odot}}$]{}, this explains the mass-dependence of the lithium depletion. The convective zone of stars in the lithium plateau, i.e. between 1.10 and 1.30 [$\mathrm{M_{\odot}}$]{}, is not deep enough for mixing mechanisms to reach the layers in which lithium could be destroyed. For larger masses, in the cases of the Hyades (M $>$ 1.60 [$\mathrm{M_{\odot}}$]{}) and NGC 752 (1.40 [$\mathrm{M_{\odot}}$]{}$<$ M $<$ 1.70 [$\mathrm{M_{\odot}}$]{}), we can see that the slight slope in lithium depletion as a function of mass is not due to the depth of the convective zone, which remains nearly constant in this range of masses. As we explained, this lithium destruction is due to microscopic diffusion, which increases with mass and age. For larger masses, the dramatic decrease of lithium abundances corresponds to the strong deepening of the convective zone during the first dredge-up, at the beginning of the RGB. In the case of standard models, as explained above, the base of the convective zone is not deep enough to reach the lithium destruction layers, and the lithium depletion is controlled by the microscopic diffusion process, which is efficient for masses larger than about 1.25 [$\mathrm{M_{\odot}}$]{}.
Comparison of the predicted lithium abundance and observations {#sec:comparisons}
==============================================================
For the **Hyades**, we estimated the mass for each star of our sample by picking the closest isochrone point to its data point on the color-magnitude diagram, and interpolating the value of the mass from the values for which the models were actually run. We used the isochrone calculated from models with extra mixing. Results of this interpolation are given in Table \[tab:interphyades\]. In Figure \[fig:li-mass\_hyades\], we present the lithium abundances of our sample as a function of the inferred stellar mass. We also plot the corresponding isochrones from both set of models. For stars with masses lower than 1.10 [$\mathrm{M_{\odot}}$]{}, lithium destruction occurs as soon as the evolution starts. Stars with these masses present low lithium abundance at the age of the Hyades. Our models with extra mixing reproduce the depletion as observed in Fig. \[fig:li-mass\_hyades\]. The models present a more severe depletion for this mass range. Standard models do not reproduce to any extent the lithium abundances of these stars.
For this cluster, the lithium plateau at A(Li) $\sim 3.0$ for stars with masses larger than 1.10 [$\mathrm{M_{\odot}}$]{} is also well reproduced. However, the Li-dip for stars with masses between 1.30 and 1.45 [$\mathrm{M_{\odot}}$]{} is not reproduced by our models, thus suggesting that the mixing mechanism responsible for the Li-dip is not related with meridional circulation. However, it is interesting to note that the strong decrease of the lithium abundance in standard models corresponds to the Li-dip. For stars with masses larger than 1.60 [$\mathrm{M_{\odot}}$]{}, there is a lack of lithium abundances estimation and we cannot give any formal comparison for the Hyades age. Known binaries have the same behavior as single stars.\
![Lithium abundances of our sample as a function of the inferred stellar mass for the Hyades. Red filled circles represent stars with the lithium abundance determined. Green triangles represent the lithium abundance upper limits. Blue diamond represent the known binaries. Continuous line corresponds to the lithium abundance predicted by models with rotation-induced mixing at the age of the Hyades, whereas the dotted line corresponds to the standard models.[]{data-label="fig:li-mass_hyades"}](./fig6.eps){width="9" height="9cm"}
For **NGC 752**, results of the interpolation are given in Table \[tab:interpngc752\]. In Figure \[fig:li-mass\_ngc752\], we plot the lithium abundances of our sample as a function of the inferred stellar mass for this cluster. We also present the corresponding isochrones from our two sets of models, with the ages deduced from the color-magnitude diagram. As in the Hyades, we observe a good agreement of the models with rotation-induced mixing for the stars of NGC 752 with masses that are lower than 1.10 [$\mathrm{M_{\odot}}$]{}. However, for stars in the plateau, observed lithium abundances are lower by $\sim$0.3 dex in relation to to the models’ predictions. The Li-dip is still not reproduced. The models with masses larger than 1.60 [$\mathrm{M_{\odot}}$]{} seem to deplete too much lithium. Again, standard models do not reproduce the lithium distribution of the NGC 752 cluster, and the decrease of lithium abundance for more massive stars corresponds to the Li-dip. The lithium abundances of binaries have no pecularity compared to single stars. This implies that the binarity has no influence on lithium depletion of these stars and thus, and that there is no tidal effect in the observed systems. It suggests that there are no close binaries, and that an analysis of the spectra of the two components of each system should be done in the future to know if they can be separated or not.\
![Same as Fig. \[fig:li-mass\_hyades\] but for the open cluster NGC 752.[]{data-label="fig:li-mass_ngc752"}](./fig7.eps){width="9" height="9cm"}
Figure \[fig:li-mass\_m67\] is the **M67** counterpart of Figs. \[fig:li-mass\_hyades\] and \[fig:li-mass\_ngc752\]. It is published by @pace12 and shown here for better clarity and discussion. The results of the interpolation on the isochrone are also available in the online version of @pace12. In this paper, the authors observed a good agreement between models and observations for stars with masses lower than 1.10 [$\mathrm{M_{\odot}}$]{}. The difference between the observed and the predicted plateaus is about 0.45 dex, showing an increase of this difference with age. The Li-dip, which lies around 1.34 [$\mathrm{M_{\odot}}$]{} [@balachandran95], is not entirely reproduced in depth, the decrease of lithium abundance in the models is only due to the deepening of the outer convective zone, when the star enters in the RGB. Extra mixing mechanisms responsible for the Li-dip are not yet identified.
![Same as Figs. \[fig:li-mass\_hyades\] and \[fig:li-mass\_ngc752\], but for the open cluster M67. The open squares represents the deviant stars as discussed in the Sect. 5.3 of @pace12.[]{data-label="fig:li-mass_m67"}](./fig8.eps){height="9cm" width="9cm"}
Discussion {#sec:discussion}
==========
Based on previous results with M67 [@pace12] and the new analysis for the Hyades and NGC 752, the calibration of evolution models on the atmospheric parameters of the Sun enables us to reproduce the qualitative distribution of lithium abundance as a function of the stellar mass for different ages. Lithium abundance is a crescent function of mass for $M < 1.10$ [$\mathrm{M_{\odot}}$]{}, where the depth of the outer convective zone plays a crucial role in the efficiency of the meridional circulation to bring lithium from the base of the convective zone to the destruction layers. As expected, solar-calibrated mixing is suitable for solar-type stars at all stages of the evolution. For lower masses, our mixing seems to deplete too much lithium, too soon, and this discrepancy would be even larger if lithium depletion during the PMS had been taken into account.
For higher masses, we observe a plateau between 1.10 and 1.30 [$\mathrm{M_{\odot}}$]{}in the three clusters where lithium remains constant as a function of stellar mass. This plateau is reproduced by our models. For these masses the outer convective zone is too shallow to allow the meridional circulation to bring lithium particles to the destruction layers. The lithium abundance in this plateau decreases with time, due to microscopic diffusion below the convective zone. Indeed, in Fig. \[fig:li-mass\_models\], we can see that standard models present nearly the same lithium abundances as models with rotation-induced mixing in this range of masses. However, our models fail to reproduce quantitatively this decrease and the difference between the observed and the predicted plateaus increases with age. Since the turbulence, which is due to meridional circulation, prevents lithium to diffuse below the convective zone, it can indicate that our parametrization leads to a too strong efficiency of the meridional circulation for theses masses.
In the three clusters, we observe a Li-dip, which lies around 1.34 [$\mathrm{M_{\odot}}$]{}. Our models with meridional circulation calibrated on the Sun do not reproduce this Li-dip. These results can be interpreted in different ways. As already stated, there is no a priori reason why the calibration of the parameters of the extra mixing for the solar model should also hold for different masses, and at different ages. It has been shown and discussed by @tv03b and @pace12 that it is possible to reproduce the Li-dip by calibrating the meridional circulation for each mass. We chose not to use this artificial parametric way and we focused on the study of the discrepancy in the solar mixing as a function of mass and its evolution. Furthermore, our results indicate that lithium depletion in these stars likely involves another mechanism besides meridional circulation. This other mechanism can be linked with angular momentum transport, like internal gravity waves . Internal gravity waves (IGW) are produced and excited by stochastic movement at the base of the convective zone and they transport mainly negative angular momentum through the radiative zone. They are mostly efficient in low-mass stars. If we introduce IGW in a solar model, part of the angular momentum should be transported by these waves, reducing the efficiency of the meridional circulation. Thus, we would have to increase the turbulent diffusion coefficients to reproduce the solar lithium depletion. This increase could be enough to destroy lithium in greater masses, where IGW are much less efficient, and reproduce the decrease of the lithium abundance in the plateau with time and the Li-dip. These new physics need to be tested for different masses. Moreover, it cannot be a coincidence that, for the three clusters, the Li-dip corresponds to the masses for which microscopic diffusion becomes very efficient. A possible explanation could be that, for some reason, the rotation-induced turbulence, which prevents microscopic diffusion being efficient in this range of masses, becomes frozen.
We are also able to observe that for lower masses, the dispersion of lithium abundance seems to increase with age. The same trend is found for clusters studied by @randich08 and @randich09 for stars with a similar range of effective temperature. Figure \[fig:hist\_Li-mass\] shows the cumulative distribution of lithium abundance for a shallow range of mass around the solar value (between 0.95 and 1.05 [$\mathrm{M_{\odot}}$]{}) in both open clusters with solar metallicity, NGC 752 and M67. Since the Hyades is overmetallic, it would produce an additional bias in comparison. Indeed, a higher metallicity implies a deeper convective zone, and thus a larger efficiency of the transport mechanisms of lithium particles below the convective zone. In this figure, M67 presents a dispersion in A(Li) from 0.6 to 2.2 dex for masses between 0.95 and 1.05 [$\mathrm{M_{\odot}}$]{}. NGC 752 presents a lower dispersion from 1.6 to 2.3 dex in the same range of masses. The dispersion seems to increase with age between 1.59 Gyr and 3.87 Gyr. However, the low number of stars with masses between 0.95 and 1.05 [$\mathrm{M_{\odot}}$]{} for NGC 752 is not enough for a definitive conclusion about the evolution of the lithium abundance dispersion with age for solar-type stars. In @duarte15 [ in prep.], the authors show that the same dispersion exists for solar twin field stars. The authors use a sample of 88 solar twin stars from @ramirez14 and do a comparative study of lithium abundance, rotation period, and magnetic activity as a function of stellar mass. Lithium abundance in solar twin field stars span a wide range of values, which cannot be explained by the small differences in mass. This range of dispersion for stars with estimated ages of between 2.0 and 5.0 Gyr is equivalent to the one we can see in the MS region in M67 ($M \leqslant 1.10$ [$\mathrm{M_{\odot}}$]{}).
![Cumulative distribution for lithium abundance of NGC 752 (dotted line) stars and M67 (continuous blue line) stars with a stellar mass between 0.95 and 1.05 [$\mathrm{M_{\odot}}$]{}.[]{data-label="fig:hist_Li-mass"}](./fig9.eps){width="9cm" height="9cm"}
It will be important to understand this point in a near future, because it could be closely connected with the internal differential rotation evolution. Some studies put forward the idea that the dispersion in lithium abundances in solar-type stars is due to the different rotation history [see @charbonneltalon05]. The way that the magnetic field of the protostar is coupled with the accretion disk determines the magnetic brake that the star undergoes during the PMS, and the initial rotation rate of the star at the ZAMS . Furthermore, stellar winds have a strong influence on rotation in the MS, removing angular momentum from the surface convective zone of the star. This increases the differential rotation between the convective zone and the radiative interior, and thus the efficency of the tachocline, which plays a major role on lithium destruction. As a result, the resulting decrease in the rotation rate causes significant changes in the wind strength [@johnstone15].
This result indicates that lithium abundance is not such a good age indicator for solar-type stars, when used alone. Rotation period through gyrochronology [@barnes10] and asteroseismology must be taken into account to infer stellar age.
Conclusions {#sec:conclusions}
===========
In this work, we have focused our study on the evolution of the atmospheric lithium abundance for open cluster, low-mass star members as a function of mass and age. We used three open clusters of different ages: the Hyades, with an estimated age of 0.79 Gyr, NGC 752, with an age of 1.59 Gyr, and M67, with 3.87 Gyr. For each cluster, we show how strong the dependence of lithium abundance is with mass. Models that include meridional circulation reproduce most of the shape of lithium depletion. However, they fail to reproduce the Li-dip. Comparisons between clusters show that microscopic diffusion plays a role in F-type stars, where meridional circulation is inefficient in bringing lithium from the bottom of the shallow convective zone to the destruction layers. This study shows that there are two main reasons why using lithium abundance is tricky as an indicator of age. First, the lithium abundance shows a strong dependence on mass of a same age. The second reason is its dependency with transport mechanisms, which can be a function of rotation between others parameters. Comparisons of the rotation period distributions for solar-type stars in open clusters with different ages show that core and envelope rotations have to be coupled with a characteristic timescale of angular momentum transfer between them from a few Myr for the fastest rotators to nearly a hundred Myr for the slowest rotators [@denissenkov10; @irwin07]. This angular momentum transfer seems to be mostly due to magnetic braking, but the interaction between meridional circulation and the differential rotation during this braking causes large turbulence which leads to a diffusion process of lithium from the base of the convective zone to the destruction layers. However, our models do not calculate this magnetic braking and the destruction of lithium is only driven by turbulence owing to classical meridional circulation. In the future, we intend to include the angular momentum redistribution in the stellar interior by magnetic braking in our models. Furthermore, in field stars one has to consider the intrinsic difficulty of getting accurate masses. This generates a greater dispersion and complicates the interpretation of the lithium abundance evolution as a function of age. However, for stars with very accurate mass determination, it is possible to use lithium abundance to diffentiate a young star from an old one. At the moment, we need observations of solar analog stars (with an effective temperature close to the solar one) of clusters with intermediate masses and solar metallicity to clarify the evolution of the dispersion of the lithium abundance in stars with solar mass.
This research made use of the SIMBAD database, operated at CDS, Strasbourg, France, and of WEBDA, an open cluster database developed and maintained by Jean-Claude Mermilliod. G.P. is supported by grant SFRH/BPD/39254/2007 and by the project PTDC/CTE-AST/098528/2008, funded by Fundação para a Ciência e a Tecnologia (FCT), Portugal. Research activities of the Stellar Board at the Federal University of Rio Grande do Norte are supported by continuous grants from CNPq and FAPERN Brazilian Agencies. J.D.N. and M.C. would like to acknowledge support from CNPq ([*Bolsa de Produtividade*]{}).
Asplund, M., Grevesse, N., Sauval, A. J., & Scott, P. 2009, , 47, 481 Bahcall, J. N., Pinsonneault, M. H., & Wasserburg, G. J. 1995, RvMP, 67, 781 Balachandran, S. 1995, , 446, 203 Barnes, S. A. 2010, , 722, 222 Basu, S., & Antia, H. M. 1997, , 287, 189 Boesgaard, A. M., & Tripicco, M. J. 1986, , 302, L49 Boesgaard, A. M. 1987, , 99, 1067 Boesgaard, A. M., & Budge, K. G. 1988, , 332, 410 Böhm-Vitense, E. 1958, , 46, 108 Brandt, T. D., & Huang, C. X. 2015, , 807, 58 Brun, A. S., Turck-Chièze, S., Zahn, J.-P. 1998, in *Structure and Dynamics of the Interior of the Sun and Sun-like Stars*, ESA Publications Division, SP-418, 439 Canuto, V. M., & Mazzitelli, I. 1990, , 370, 295 Carrera, R., & Pancino, E. 2011, , 535, A30 Casagrande, L., Ramírez, I., Meléndez, J., Bessell, M., & Asplund, M. 2010, , 512, A54 Cayrel de Strobel, G. 1996, A&ARv, 7, 243 Chapman, S. 1917, , 77, 540 Chapman, S., & Cowling, T. G. 1970, in *The Mathematical Theory of Non-Uniform Gases*, Cambridge University Press, 3rd ed. Charbonneau, P., Christensen-Dalsgaard, J., Henning, R. et al. 1999, , 527, 445 Charbonnel, C., & Talon, S. 2005, Science, 309, 2189 Christensen-Dalsgaard, J., Dappen, W., Ajukov., S. V, et al. 1996, Science, 272, 1286 Cox, A. N., Guzik, J. A., & Kidman, R. B. 1989, , 342, 1187 Daniel, S. A., Latham, D. W., Mathieu, R. D., & Twarog, B. A. 1994, , 106, 281 D’Antona, F., & Mazzitelli, I. 1984, , 138, 431 D’Antona, F., & Mazzitelli, I. 1994, , 90, 467 de Bruijne, J. H. J., Hoogerwerf, R., & de Zeeuw, P. T. 2001, , 367, 111 Deliyannis, C. P. & Pinsonneault, M. H. 1997, 488, 836 Denissenkov, P. A., Pinsonneault, M., Terndrup, D. M., & Newsham, G. 2010, , 716, 1269 do Nascimento Jr., J.-D., Castro, M., Meléndez, J., et al. 2009, , 501, 687 Duarte, T., Soares, J., Castro, M., & do Nascimento Jr., J.-D. 2015, in prep. Duncan, D. K., & Jones, B. F. 1983, , 271, 663 Eddington, A. S. 1916, , 77, 16 Eddington, A. S. 1926, in *The Internal Constitution of the Stars* (1959 New York: Dover) Garcia Lopez, R. J., Rebolo, R., & Beckman, J. E. 1988, , 100, 1489 Grevesse, N., & Sauval, A. J. 1998, , 85, 161 Heinemann, K. 1926, Astronomische Nachrichten, 227, 193 Hobbs, L. M., & Pilachowski, C. 1986a, , 309, L17 Hobbs, L. M., & Pilachowski, C. 1986b, , 311, L37 Hobbs, L. M., & Thorburn, J. A. 1992, , 104, 669 Hui-Bon-Hoa, A. 2008, Ap&SS, 316, 55 Irwin, J., Hodgkin, S., Aigrain, S., et al. 2007, , 377, 741 Johnstone, C. P., Güdel, M., Brott, I., & Lüftinger, T. 2015, , 577, A28 Jones, B. F., Fischer, D., Shetrone, M., & Soderblom, D. R. 1997, , 114, 352 Jones, B. F., Fischer, D., & Soderblom, D. R. 1999, , 117, 330 Ku[č]{}inskas, A., Hauschildt, P. H., Ludwig, H.-G., et al. 2005, , 442, 281 Kurucz, R. L. 1991, in *Stellar Atmospheres: Beyond the Classical Models*, ed. L. Crivellari, I. Hubeny, & D. G. Hummer (NATO ASI Ser.; Dordrecht; Kluwer), 441 Lind, K., Asplund, M., & Barklem, P. S. 2009, , 503, 541 Maeder, A., & Zahn, J.-P. 1998, , 334, 1000 Martin, E. L., Rebolo, R., Garcia Lopez, R. J., Magazzú, A., & Pavlenko, Y. V. 1993, in IAU Colloquim 137, *Inside the Stars*, ed. A. Baglin & W. Weiss (ASP Conf. Series; San Francisco: ASP), 40, 165 Matt, S. P., & Pudritz, R. E. 2005, , 632, 135 McArthur, B. E., Benedict, G. F., Harrison, T. E., & van Altena, W. 2011, , 141, 172 Meléndez, J., Ramírez, I., Casagrande, L., et al. 2010, , 328, 193 Mestel, L., & Moss, D. L. 1986, , 221, 25 Michaud, G. 1970, , 160, 641 Pace, G., Castro, M., Meléndez, J., Théado, S., & do Nascimento Jr., J.-D. 2012, , 541, A150 Paquette, C., Pelletier, C., Fontaine, G., & Michaud, G. 1986, , 61, 177 Pasquini, L., Randich, S., & Pallavicini, R. 1997, , 325, 535 Pasquini, L., Biazzo, K., Bonifacio, P., Randich, S., & Bedin, L. R. 2008, , 489, 677 Paulson, D. B., Sneden, C., & Cochran, W. D. 2003, , 125, 3185 Perryman, M. A. C. 1997, Proceedings of the ESA Symposium ‘Hipparcos - Venice ’97’, 13-16 May, Venice, Italy, ESA SP-402 (July 1997), p. 1-4 Perryman, M. A. C., Brown, A. G. A., Lebreton, Y., et al. 1998, , 331, 81 Pilachowski, C. A., & Hobbs, L. M. 1988, , 100, 336 Pilachowski, C., Saha, A., & Hobbs, L. M. 1988, , 100, 474 Pinsonneault, M. H. 1997, , 35, 557 Ramírez, I., Meléndez, J., Bean, J., et al. 2014, , 572, A48 Randich, S., Aharpour, N., Pallavicini, R., Prosser, C. F., & Stauffer, J. R. 1997, , 323, 86 Randich, S., Sestito, P., Primas, F., Pallavicini, R., & Pasquini, L. 2006, , 450, 557 Randich, S., Primas, F., Pasquini, L., Sestito, P., & Pallavicini, R. 2007, , 469, 163 Randich, S. 2008, MmSAI, 79, 516 Randich, S. 2009, in *The Ages of Stars*, IAU Symp. 258, eds. E. E. Mamajek, D. R. Soderblom, & R. F. G. Wyse, 133 Rebolo, R., & Beckman, J. E. 1988, , 201, 267 Reddy, A. B. S., Giridhar, S., & Lambert, D. L. 2012, , 419, 1350 Richard, O., Vauclair, S., Charbonnel, C., & Dziembowski, W.A. 1996, , 312, 1000 Richard, O., Théado, S., & Vauclair, S. 2004, SoPh, 220, 243 Schatzman, E. 1969, , 3, 331 Schlattl, H. 2002, , 395, 85 Sestito, P., Randich, S., & Pallavicini, R. 2004, , 426, 809 Soderblom, D. R., Oey, M. S., Johnson, D. R. H., & Stone, R. P. S. 1990, , 99, 595 Spiegel, E. A., & Zahn, J.-P. 1992, , 265, 106 Spite, F., Spite, M., Peterson, R. C., & Chaffee, F. H., Jr. 1987, , 171, L8 Sweet, P. A. 1950, , 110, 548 Takeda, Y., Honda, S., Ohnishi, T., et al. 2013, , 65, 53 Talon, S., & Charbonnel, C. 2005, , 440, 981 Talon, S., & Zahn, J.-P. 1997, , 317, 749 Taylor, B. J. 2006, , 132, 2453 Théado, S., & Vauclair, S. 2003, , 587, 795 Thorburn, J. A., Hobbs, L. M., Deliyannis, C. P., & Pinsonneault, M. H. 1993, , 415, 150 Turcotte, S. Richer, J., Michaud, G., Iglesias, C. A., & Rogers, F. J. 1998, , 504, 539 van Leeuwen, F. 2007, , 474, 653 Vauclair, S., & Théado, S. 2003, , 587, 777 Ventura, P., Zeppieri, A., Mazzitelli, I., & D’Antona, F. 1998, , 331, 1011 Yadav, R. K. S., Bedin, L. R., Piotto, G., et al. 2008, , 484, 609 Zahn, J.-P. 1992, , 265, 115
[ C C C C D C C C D ]{}
\
ID & [$T_{\rm eff}$]{}& $T_{\rm iso}$ & $\log g$ & $(\log g)_{\rm iso}$ & Mass & A(Li) & dA(Li) & Binarity\
& (K) & (K) & & & [$\mathrm{M_{\odot}}$]{}& (dex) & (dex) &
[[** – continued from previous page**]{}]{}\
ID & [$T_{\rm eff}$]{}& $T_{\rm iso}$ & $\log g$ & $(\log g)_{\rm iso}$ & Mass & A(Li) & dA(Li) & Binarity\
& (K) & (K) & & & [$\mathrm{M_{\odot}}$]{}& (dex) & (dex) &\
\
\
24357 & 6934 & 7000 & 4.30 & 4.30 & 1.52 & 3.19 & 0.07 & n\
26462 & 6971 & 7014 & 4.30 & 4.19 & 1.55 & 3.25 & 0.07 & n\
26015 & 6795 & 7079 & 4.32 & 4.18 & 1.58 & 2.51 & 0.07 & y\
26911 & 6783 & 6823 & 4.32 & 4.24 & 1.46 & 3.24 & 0.07 & y\
27561 & 6728 & 6823 & 4.33 & 4.24 & 1.46 & <1.56 & & n\
18404 & 6714 & 6668 & 4.33 & 4.26 & 1.40 & <1.55 & & n\
25102 & 6705 & 6606 & 4.33 & 4.28 & 1.37 & <1.95 & & n\
28736 & 6693 & 6714 & 4.33 & 4.25 & 1.42 & 1.53 & 0.07 & n\
26345 & 6660 & 6591 & 4.34 & 4.29 & 1.36 & <1.51 & & n\
28568 & 6656 & 6591 & 4.34 & 4.29 & 1.36 & <1.91 & & n\
28911 & 6651 & 6606 & 4.34 & 4.28 & 1.37 & 1.89 & 0.07 & n\
27534 & 6598 & 6668 & 4.34 & 4.26 & 1.40 & <1.95 & & n\
29225 & 6593 & 6561 & 4.35 & 4.29 & 1.35 & <1.95 & & n\
27848 & 6558 & 6652 & 4.35 & 4.27 & 1.39 & 1.85 & 0.07 & n\
31845 & 6558 & 6501 & 4.35 & 4.31 & 1.32 & 2.17 & 0.07 & n\
28406 & 6554 & 6501 & 4.35 & 4.31 & 1.32 & 2.44 & 0.07 & n\
27483 & 6533 & 6934 & 4.36 & 4.20 & 1.52 & <1.12 & & y\
27731 & 6507 & 6426 & 4.36 & 4.33 & 1.29 & 2.05 & 0.07 & n\
28483 & 6474 & 6501 & 4.37 & 4.31 & 1.32 & 1.98 & 0.07 & n\
28608 & 6465 & 6353 & 4.37 & 4.35 & 1.26 & 2.28 & 0.07 & n\
30869 & 6339 & 6714 & 4.39 & 4.25 & 1.42 & 1.79 & 0.08 & y\
27383 & 6310 & 6441 & 4.40 & 4.33 & 1.30 & 2.37 & 0.08 & y\
27991 & 6310 & 6792 & 4.40 & 4.24 & 1.45 & 2.80 & 0.08 & y\
26784 & 6291 & 6397 & 4.40 & 4.34 & 1.28 & 3.02 & 0.08 & n\
27808 & 6275 & 6223 & 4.41 & 4.39 & 1.21 & 3.06 & 0.08 & n\
28394 & 6242 & 6353 & 4.41 & 4.35 & 1.26 & 2.17 & 0.08 & y\
30809 & 6239 & 6223 & 4.41 & 4.39 & 1.21 & 1.60 & 0.08 & n\
28363 & 6202 & 6745 & 4.42 & 4.25 & 1.43 & 2.63 & 0.08 & y\
30738 & 6202 & 6397 & 4.42 & 4.34 & 1.28 & 3.12 & 0.08 & y\
28205 & 6199 & 6194 & 4.42 & 4.40 & 1.20 & 3.06 & 0.08 & n\
28635 & 6186 & 6123 & 4.42 & 4.42 & 1.17 & 3.01 & 0.08 & n\
30810 & 6174 & 6652 & 4.43 & 4.27 & 1.39 & 2.82 & 0.08 & y\
35768 & 6122 & 6353 & 4.44 & 4.35 & 1.26 & 2.69 & 0.08 & n\
28033 & 6118 & 6223 & 4.44 & 4.39 & 1.21 & 3.14 & 0.08 & y\
27406 & 6107 & 6137 & 4.44 & 4.41 & 1.18 & 2.93 & 0.08 & n\
28237 & 6107 & 6194 & 4.44 & 4.40 & 1.20 & 2.81 & 0.08 & n\
20430 & 6079 & 6309 & 4.45 & 4.37 & 1.24 & 2.95 & 0.09 & y\
14127 & 6079 & 5984 & 4.45 & 4.46 & 1.12 & 2.72 & 0.09 & n\
29419 & 6045 & 6095 & 4.45 & 4.43 & 1.16 & 2.86 & 0.09 & n\
30589 & 6037 & 6137 & 4.45 & 4.41 & 1.18 & 2.87 & 0.09 & n\
25825 & 5983 & 5984 & 4.46 & 4.46 & 1.12 & 2.74 & 0.09 & n\
27859 & 5961 & 6039 & 4.47 & 4.44 & 1.14 & 2.73 & 0.09 & n\
28344 & 5924 & 6011 & 4.47 & 4.45 & 1.13 & 2.76 & 0.09 & n\
20439 & 5894 & 5984 & 4.48 & 4.46 & 1.12 & 2.76 & 0.09 & n\
28992 & 5844 & 5834 & 4.49 & 4.49 & 1.07 & 2.64 & 0.09 & n\
26767 & 5812 & 5794 & 4.50 & 4.50 & 1.06 & 2.61 & 0.09 & n\
26736 & 5757 & 5834 & 4.51 & 4.49 & 1.07 & 2.45 & 0.09 & n\
28099 & 5735 & 5834 & 4.51 & 4.49 & 1.07 & 2.38 & 0.09 & n\
26756 & 5640 & 5571 & 4.53 & 4.55 & 0.99 & 2.08 & 0.10 & n\
27282 & 5553 & 5610 & 4.54 & 4.54 & 1.00 & 1.78 & 0.10 & n\
240648& 5527 & 5533 & 4.54 & 4.55 & 0.98 & 1.85 & 0.10 & n\
19902 & 5522 & 5636 & 4.55 & 4.53 & 1.01 & 1.59 & 0.10 & n\
28593 & 5516 & 5508 & 4.55 & 4.56 & 0.97 & 1.36 & 0.10 & n\
31609 & 5508 & 5432 & 4.55 & 4.57 & 0.95 & 1.61 & 0.10 & n\
27250 & 5485 & 5357 & 4.55 & 4.58 & 0.93 & 1.49 & 0.10 & n\
27732 & 5449 & 5395 & 4.55 & 4.58 & 0.94 & 1.31 & 0.11 & y\
32347 & 5429 & 5470 & 4.56 & 4.57 & 0.96 & 1.49 & 0.11 & n\
242780& 5429 & 5432 & 4.56 & 4.57 & 0.95 & 1.40 & 0.11 & n\
283704& 5426 & 5432 & 4.56 & 4.57 & 0.95 & 1.36 & 0.11 & n\
284574& 5303 & 5395 & 4.57 & 4.58 & 0.94 & 1.04 & 0.11 & n\
284253& 5297 & 5272 & 4.57 & 4.59 & 0.91 & 0.75 & 0.11 & n\
285773& 5254 & 5116 & 4.58 & 4.62 & 0.87 & 0.75 & 0.11 & n\
30505 & 5249 & 5236 & 4.58 & 4.60 & 0.90 & 0.57 & 0.11 & y\
28258 & 5235 & 5236 & 4.58 & 4.60 & 0.90 & 0.29 & 0.11 & n\
27771 & 5196 & 5164 & 4.58 & 4.61 & 0.88 & 0.50 & 0.12 & n\
28462 & 5172 & 4988 & 4.58 & 4.63 & 0.84 & 0.82 & 0.12 & n\
285367& 5114 & 5164 & 4.59 & 4.61 & 0.88 & 0.10 & 0.12 & n\
[ D C D C C D C C C D ]{}
\
ID & $T_{\rm iso}$ & $(\log g)_{\rm iso}$ & $T_{\rm phot}$ & Mass & EW$_{Li}$ & dEW$_{Li}$ & A(Li) & dA(Li) & Binarity\
& (K) & & (K) & [$\mathrm{M_{\odot}}$]{}& (m$\AA$) &(m$\AA$) & (dex) & (dex) &\
[[** – continued from previous page**]{}]{}\
ID & $T_{\rm iso}$ & $(\log g)_{\rm iso}$ & $T_{\rm phot}$ & Mass & EW$_{Li}$ & dEW$_{Li6708\AA}$ & A(Li) & dA(Li) & Binarity\
& (K) & & (K) & [$\mathrm{M_{\odot}}$]{}& (m$\AA$) &(m$\AA$) & (dex) & (dex) &\
\
\
H10 & 6854 & 3.82 & 6463 & 1.68 & 35 & 10 & 2.69 & 0.30 & y\
H88 & 6501 & 4.29 & 6195 & 1.29 & <10 & &<2.00 & & n\
H94 & 5308 & 4.59 & 5333 & 0.89 & <5 & &<0.78 & & n\
H123 & 6668 & 4.14 & 6523 & 1.44 & <34 & &<2.67 & & n\
H139 & 6501 & 4.29 & 6294 & 1.29 & <10 & &<2.00 & & n\
H185 & 6266 & 4.36 & 6030 & 1.19 & 42 & 10 & 2.31 & 0.21 & n\
H189 & 6668 & 4.17 & 6416 & 1.41 & 65 & 10 & 3.00 & 0.18 & n\
H222 & 6652 & 4.03 & 6555 & 1.55 & 45 & 10 & 2.68 & 0.17 & n\
H227 & 5116 & 4.62 & 4917 & 0.85 & <12 & &<0.99 & & n\
H235 & 6606 & 4.21 & 6203 & 1.36 & <3 & &<1.53 & & y\
H237 & 6223 & 4.38 & 5797 & 1.17 & 43 & 10 & 2.29 & 0.34 & y\
H254 & 6668 & 4.05 & 6617 & 1.53 & 49 & 10 & 2.74 & 0.17 & n\
H259 & 6637 & 4.19 & 6432 & 1.39 & <17 & &<2.33 & & n\
H265 & 5688 & 4.53 & 5614 & 0.99 & <15 & &<1.60 & & n\
H266 & 6668 & 4.15 & 6412 & 1.44 & 42 & 10 & 2.64 & 0.21 & y\
H293 & 6412 & 4.31 & 6188 & 1.25 & <15 & &<1.32 & & n\
H302 & 6606 & 4.22 & 6440 & 1.36 & <5 & &<1.76 & & n\
H304 & 6441 & 4.30 & 6199 & 1.26 & <3 & &<1.43 & & n\
PLA475 & 5929 & 4.47 & 5722 & 1.07 & 59 & 10 & 2.45 & 0.18 & n\
PLA520 & 5929 & 4.47 & 5893 & 1.07 & 67 & 12 & 2.52 & 0.15 & n\
PLA552 & 5888 & 4.48 & 5857 & 1.06 & 61 & 7 & 2.44 & 0.13 & y\
PLA648 & 6324 & 4.35 & 5844 & 1.21 & 24 & 8 & 2.28 & 0.37 & n\
PLA699 & 5847 & 4.50 & 5713 & 1.04 & 48 & 10 & 2.28 & 0.16 & y\
PLA701 & 5794 & 4.50 & 5526 & 1.03 & 42 & 12 & 2.17 & 0.26 & n\
PLA786 & 5727 & 4.52 & 5414 & 1.01 & 20 & 7 & 1.77 & 0.31 & n\
PLA859 & 5714 & 4.52 & 5584 & 1.00 & 26 & 7 & 1.88 & 0.18 & n\
PLA864 & 5902 & 4.47 & 5867 & 1.06 & 75 & 9 & 2.56 & 0.14 & n\
PLA889 & 5970 & 4.46 & 5956 & 1.08 & 78 & 11 & 2.63 & 0.14 & n\
PLA921 & 6053 & 4.44 & 5949 & 1.11 & 54 & 9 & 2.49 & 0.15 & n\
PLA964 & 5888 & 4.48 & 5854 & 1.06 & 64 & 9 & 2.46 & 0.14 & n\
PLA983 & 5767 & 4.51 & 5765 & 1.02 & 45 & 7 & 2.19 & 0.14 & n\
PLA993 & 5482 & 4.57 & 5475 & 0.94 & 20 & 6 & 1.57 & 0.19 & n\
PLA1012 & 6165 & 4.40 & 5996 & 1.15 & 74 & 10 & 2.74 & 0.14 & n\
PLA1107 & 5445 & 4.58 & 5614 & 0.93 & 24 & 5 & 1.62 & 0.17 & n\
PLA1284 & 5902 & 4.48 & 5564 & 1.06 & 42 & 9 & 2.26 & 0.29 & y\
PLA1365 & 5662 & 4.54 & 5470 & 0.99 & 48 & 8 & 2.14 & 0.18 & n\
H77 & 4764 & 2.93 & 4580 & 1.75 & 59\* & 10 & 1.41 & 0.25 & n\
H311 & 4677 & 2.75 & 4718 & 1.76 & <12\* & &<0.52 & & n\
H137 & 4634 & 2.68 & 4773 & 1.76 & <10\* & &<0.39 & & n\
H208 & 4623 & 2.65 & 4682 & 1.76 & 70\* & 10 & 1.35 & 0.14 & y\
H213 & 4666 & 2.73 & 4812 & 1.76 & <12\* & &<0.51 & & y\
H1 & 4764 & 2.93 & 4934 & 1.75 & <10\* & &<0.55 & & n\
H295 & 4731 & 2.87 & 4891 & 1.76 & <10\* & &<0.52 & & n\
H186 & 4931 & 3.30 & 4891 & 1.74 & <5\* & &<0.39 & & n\
H110 & 4645 & 2.70 & 5153 & 1.76 & <5\* & &<0.07 & & y\
H24 & 4645 & 2.69 & 4812 & 1.76 & <13\* & &<0.53 & & n\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We demonstrate that the ($s$-wave) geometric spectrum of the Efimov energy levels in the unitary limit is generated by the radial motion of a primitive periodic orbit (and its harmonics) of the corresponding classical system. The action of the primitive orbit depends logarithmically on the energy. It is shown to be consistent with an inverse-squared radial potential with a lower cut-off radius. The lowest-order WKB quantization, including the Langer correction, is shown to reproduce the geometric scaling of the energy spectrum. The (WKB) mean-squared radii of the Efimov states scale geometrically like the inverse of their energies. The WKB wavefunctions, regularized near the classical turning point by Langer’s generalized connection formula, are practically indistinguishable from the exact wave functions even for the lowest ($n=0$) state, apart from a tiny shift of its zeros that remains constant for large $n$.
PACS: 67.85.-d, 03.65.Sq, 31.15.xg
author:
- 'Rajat K. Bhaduri'
- Matthias Brack
- 'M. V. N. Murthy'
title: A semiclassical analysis of the Efimov energy spectrum in the unitary limit
---
Introduction
============
Two particles that are just shy of binding may develop an infinite number of shallow bound states when a third particle is added. This was predicted by Efimov [@efimov] forty years back, and has only been recently verified experimentally with an ultra-cold gas of optically trapped $^{133}$Cs atoms [@kraemer]. Subsequently, Barontini [*et al.*]{} [@baront] have have found evidence for two kinds of Efimov trimers with $^{41}$K and $^{87}$Rb atoms. Efimov considered three identical bosons interacting pairwise with an interaction whose range $r_0$ is much smaller than the interatomic scattering length $a$. Using the hyperspherical coordinates for the three-body problem, he showed that the effective potential in the hyperradial coordinate $R$ between the length scales of $r_0$ and $|a|$ is of an inverse-square type. In the symmetric $L=0$ three-body state, this effective interaction is sufficiently attractive to give rise to the Efimov spectrum for the trimers. A signature of the Efimov spectrum is its geometric scaling, with the ratio of the adjacent energy levels being constant. This was predicted by Efimov [@efimov] and has been verified by recent experiments [@zacca].
Although in experiments the Efimov spectrum can only be measured for large but finite scattering lengths $a$, our semiclassical analysis of the geometric spectrum in Sec. II is done in the unitary limit ($|a|\to\infty$). The geometric scaling of the spectrum then holds right up to the accumulation point at $E=0$, and the effective potential is of inverse-square type for all distances $R>r_0$. This limit itself has interesting properties and given rise to various theoretical research [@rajeev; @bra; @grif]. Our contribution to this research here is a semiclassical description of the unitary Efimov system.
In the periodic orbit theory (POT) [@gutz], there is an intimate connection between classical periodic orbits and the quantum energy spectrum of a system through so-called trace formulae (cf. [@book] for an introduction and applications of the POT). In Sec. II, we derive an exact trace formula from the Efimov spectrum and show that the action of the classical periodic orbit generating the quantum spectrum depends logarithmically on the energy. The corresponding average level density leads, via an inverse Abel transform, to an inverse-squared radial one-body potential that creates the Efimov spectrum in the limit $E\to 0$.
In Sec. \[secinv\], we first re-derive the quantum-mechanical solutions of an inverse-squared potential with a lower cut-off radius $R_c$. The exact quantum wave function is a modified Bessel function of imaginary order [@rajeev]. Next we calculate the first-order WKB wave function (including the Langer correction [@lang]). We show that its leading term in the classically allowed region is identical with the leading term of the exact wave function for high-lying states ($n\gg1$), responsible for the geometric scaling of the spectrum. In the classically forbidden region, it decays exponentially like the exact one. The WKB eigenvalues are obtained by quantizing the classical action of the inverse-squared potential, whose leading term is the action appearing in the semiclassical trace formula discussed in Sec. \[secgeo\]. We find that the WKB spectrum, although shlightly phase shifted with respect to the exact one, reproduces the geometric scaling of the quantum-mechanical energies and mean-squared radii with the same scaling factor. Remarkably, the WKB wave functions, when regularized near the classical turning point by Langer’s generalized connection formula [@lang], are – apart from their slightly shifted zeros – practically indistinguishable from the quantum-mechanical wave functions even for the lowest state ($n=0$).
Geometric spectrum {#secgeo}
==================
In the unitary limit, three identical bosons in the symmetric $s$ state ($L=0)$, have the energy spectrum given by E\_n=E\_0 (-2n/s\_0), n=0,1,2,…,\[spectrum\] so that $E_{n+1}/E_n$ is constant and independent of $n$ and $E_0$. This is called the geometric spectrum. $E_0<0$ and $s_0>0$ are constants that depend on the system. For three equal-mass bosons, one has (/s\_0)=22.694 s\_0=1.00624. \[space\] The energy $E_0$, which corresponds to the lowest quantum state of the system $(n=0)$, introduces a length scale whose origin will be made clear soon. Our objective here is to derive an exact semiclassical trace formula for the density of states corresponding to the spectrum given by Eq. (\[spectrum\]). This will enable us to identify the action of a single periodic orbit that generates the above spectrum. For an energy spectrum governed by only one quantum number, there is a rather simple way of deriving a trace formula. Following [@book] (Chapter 3.2.2), we write $E_n=f(n)$, with degeneracy $D(n)=1$. The function $f(n)$ is monotonic with a differentiable inverse, $f^{-1}(x)=F(x)$. The exact density of states, defined by g(E)=\_n(E-E\_n), \[dos\] can then be rewritten, using Poisson resummation, as g(E)=|F’(E)|\_[n=0]{}\^ (n-F(E))=|F’(E)|. \[trace\] This result, which is exact, can be split into two parts: a Thomas-Fermi (TF) term $\widetilde{g}(E)=|F'(E)|$ which gives its average behaviour, and an oscillating term which we denote by $\delta g(E)$.
In the semiclassical POT [@gutz], the oscillating part of the exact density of states of a quantum Hamiltonian is expressed as a sum over the periodic orbits of the corresponding classical Hamiltonian: g(E)=\_ \_[k=1]{}\^ [A]{}\_[,k]{} \[osc\] The sum is over primitive periodic orbits ($k=1$), denoted by $\Gamma$, and their repetitions (harmonics) $k>1$. The amplitude ${\cal A}_{\Gamma\!,k}$ of a periodic orbit (assumed here to be isolated in phase space) depends on its primitive period and on its stability matrix. The phase factor $\sigma_{\Gamma\!,k}$ is called the Maslov index of the periodic orbit. Comparing the simple trace formula (\[trace\]) with the general form (\[osc\]), we see that it has only one primitive periodic orbit $\Gamma$ with action $S(E)=2\pi\hbar F(E)$, and an amplitude $2F'(E)$ which is proportional to its period. The Maslov index is zero. Since the spectrum is for $s$ states only, the action is that of the [*radial motion*]{} in a central potential. For the geometric spectrum, inverting , we have n(E)=(E\_0/E)=F(E). \[nofE\] Therefore the action $S(E)$ of the primitive orbit is given by S(E)=s\_0 (E\_0/E). \[active\] Substituting $F(E)$ in Eq. (\[trace\]), we obtain g(E)= {1+2\_[k=1]{}\^ }. \[trace2\] This is an exact trace formula, representing a Fourier decomposition of . When summed over all harmonics, it reproduces the quantized spectrum (\[spectrum\]).
The smooth part of the density of states $\widetilde{g}(E)$ is given by the first term on the r.h.s. of (\[trace2\]). The total number of Efimov states between the energies $-|E_0|$ and $-|E|$ is given by its integration over this interval, which yields Eq. .
Efimov trimers are formed when the two-body scattering length $|a|\gg r_0$, where the latter is of the order of the range of the intermolecular potential. One may then take the shallowest state to be $E_a\simeq \hbar^2\!/ma^2$, and the deepest state to have energy $E_0\simeq \hbar^2\!/mr_0^2$. As $a\rightarrow
\pm \infty$, there is an accumulation of states near zero energy. Substituting these in , we find the total number of Efimov states to be N= (|a|/r\_0). \[numb2\] The same result was obtained by Efimov following a different route.
In order to obtain a hint to the kind of one-body potential that can generate a spectrum of the form , we apply an idea [@wusp] that exploits the properties of the Abel transform [@abel]. The TF level density obtained above, interpreted as the $s$-state level density of a radially-symmetric potential $V(r)$, can be written [@book; @rajat] as (E) = - = \_[r\_1]{}\^[r\_2]{}, \[gtf\] where $r_1$ and $r_2$ are the lower and upper turning points with $V(r_1) = E_0$ and $V(r_2)=E$, respectively, and the potential is assumed to be monotonously increasing from its minimum value $E_0$ to the energy $E$. Using the substitution $dr=y(V)\,dV$, so that $y(V)=1/V'(r)$, we can rewrite as (E) = \_[E\_0]{}\^E . \[gev\] The r.h.s. above represents an Abel transform [@abel] of the function $y(V)$ which, by the inverse transform, is given by y(V)=. \[yV\] Using $\widetilde{g}{\,'}(E)=s_0/2\pi E^2$, the integral above is elementary and leads to y(V) = . \[yVef\] Assuming that $V(r)\to 0$ for $E\to 0$, the second term will be leading. We thus get asymptotically y(V) (-V\_[as]{})\^[-3/2]{} = E0. \[yVefas\] (Note that both $E$ and $V$ are always negative.) The potential $V_{as}(r)$ which solves the r.h.s. above is given by V\_[as]{}(r) = -. \[Vasr\] An inverse-squared potential of the type is, indeed, shown in the following section to be responsible for the asymptotic quantum-mechanical $s$-state spectrum of the Efimov three-body system in the limit $E\to 0$.
The inverse-squared potential {#secinv}
=============================
The three-body problem, after eliminating the center-of-mass coordinates, contains six degrees of freedom. In the hyperspherical formalism, these are described by a hyperradius R= and five hyperangular coordinates [@bra]. In the adiabatic approximation, for fixed $R$, the Schrödinger equation for the angular coordinates is solved to obtain a complete set of adiabatic eigenstates and the corresponding eigenvalues $\epsilon(R)$. In the unitary limit $a\rightarrow \pm\infty$, the angular variables decouple and one gets an effective inverse-squared potential in the coordinate $R$. The uncoupled hyperradial wavefunction in the $L=0$ state is $\Psi(R)=R^{-5/2}u(R)$. The reduced wavefunction $u(R)$ obeys the Schrödinger equation (cf. also a pedagogical review of the Efimov effect [@rajat]) u(R)=u(R), \[scheq\] where $m$ is the mass of the atom and $s_0$ the constant given in . In order to regularize the inverse-squared potential V\_0(R)=-, \[inverse\] we introduce a lower cut-off radius $R_c\simeq r_0$ as shown schematically in Fig. 1.
{width="10cm"} \[fig1\]
For future convenience we introduce a dimensionless scaled variable $x$ by x=R/R\_+, x\_c = R\_c/R\_+, \[scales\] where $R_+$ is the classical turning point at the energy $E$ E=V(R\_+)=-\^2 s\_0\^2/2mR\_+\^2, \[R+\] (see Sec. \[secwkb\] below), and the lowest energy $E_c$ of the classical particle is given by E\_c=V(R\_c)=-\^2 s\_0\^2/2mR\_c\^2, so that x\_c\^2 = E/E\_c. \[xcE\]
Quantum-mechanical solutions {#secqm}
----------------------------
We now solve the Schrödinger equation with the lower boundary condition $u(R_c)=0$. The second boundary condition comes from the requirement that the wave functions vanish at infinity. We introduce the following transformation u(R)= w(R) and obtain the equation for the function $w(R)$ as ++w =- w. Using the scaled variable $x=R/R_+ = R\sqrt{-2mE/\hbar^2 s_0^2}$, the above equation reduces to the standard form for the modified Bessel function +-w =w. \[Bes\] Thus, the energy scales away, which is a unique feature of the inverse-squared potential whose $R$ dependence is the same as that of the kinetic energy operator. The solution to the equation , given in [@rajeev], is a modified Bessel function with imaginary index $is_0$ w(x)=K\_[is\_0]{}(y), y = s\_0 x, which for real $y$ and $s_0$ is a real function that vanishes exponentially as $y\to\infty$. The eigenstates are found from its zeros as explained below.
To compute the function $K_{i\nu}(y)$, we use a power series expansion given in [@da]: K\_[i]{}(y)= - \_[k=0]{}\^, \[Kx\] where the phase $\phi_{\nu,k}$ is given by \_[,k]{}=(1+k+i)=\_[,0]{}+\_[s=1]{}\^k (). $\phi_{\nu,0}$ and a convergent series for its calculation [@abro] are given by \_[,0]{}=(1+i) =-+ \_[s=0]{}\^(-), where $\gamma=0.577215664\dots$ is Euler’s constant. Numerically we find \_[s\_0,0]{}=-0.30103393…\[phi0\]
As stated in [@da], the function $K_{i\nu}(y)$ has an infinite sequence of non-degenerate zeros $y_n$ ($n=1,2,\dots$) with $0 < \dots < y_{n+1}<y_n<y_{n-1}\dots<y_1<\nu$, and no zeros for $y\geq \nu$. The solution to the Schrödinger equation (\[scheq\]) is therefore given by u(x) = CK\_[i s\_0]{}(y), (y=s\_0 x) \[ux\] where $C$ is a normalisation factor. The eigenspectrum is obtained from the zeros $y_n$ via the boundary condition $K_{i\nu}(y_n)=0$ ($n=1,2,\dots$).
{width="16.5cm"} \[fig2\]
It is an important feature of this system [@grif] that all eigenfunctions $u_n(x)$ are given in terms of the universal solution simply by letting the variable $y$ in start at the ($n$+1)-th zero $y_{n+1}$. Fig. 2 shows this universal function in a doubly logarithmic plot, exhibiting the first 20 zeros. For $x>1$, i.e. $\ln(x)>0$, we notice the beginning of the exponential tail.
We can now associate the zeros with eigenvalues of the scaled cut-off $x_c$ and write the eigenfunctions as u\_n(x) = C\_nK\_[i s\_0]{}(s\_0 x), x (x\_c)\_n=y\_[n+1]{}/s\_0, n=0,1,…\[unx\] where $C_n$ is the normalization constant of the $n$-th state. Note that $n$ here is the number of zeros $x_{n,j}>(x_c)_n$ ($j=0,1,\dots,n$) of the functions $u_n(x)$. We found numerically that the $C_n$ are practically identical for all $n\geq 2$ (see Tab. \[tab1\] below), which is due to the fact that all wave functions are peaked around $x=1$ and the regions below the first three zeros, $y<y_3$, give only exponentially small contributions to the norms.
In the region of the maximum of $u(x)$ and the exponential tail for $y\simg1$, we had to include contributions to the sum in up to $k_{max}\simg 25$. For $y\simg 11$, we did not, in fact, obtain convergence of the $k$ sum. However, we found numerically that the leading term of an asymptotic form of $K_{i s_0}(y)$ for large $y$, given in [@da] u\_n(x) \~ C\_n(-y), (y 1) \[qtail\] becomes sufficiently accurate for $y \simg 9$.
For $y \siml 0.3$, the terms with $k>0$ of the series in become numerically insignificant. Since the largest zero is found at $y_1=0.0653423$ \[cf. $\ln(x_c)_0=-2.73434$ in Tab. \[tab1\] below\], this means that all zeros of $K_{i\nu}(y)$ are given by the leading term with $k=0$, yielding the asymptotic solution given in [@bra] K\_[i]{}(y) . (n1) \[Kasy\] Its zeros give the geometrical spectrum E\_n/E\_c = (x\_c)\_n\^2 \~ ()\^[2]{} (-2n/s\_0-2/s\_0+2\_[s\_0,0]{}/s\_0) (n1) \[eratas\] which, apart from constants, is the same as that discussed in Sec. \[secgeo\]; in particular, the constant ratio $E_{n+1}/E_n$ remains the same. From we get (x\_c)\_n \~ -n/s\_0 + \_0, (n1) \[specas\] where $\alpha_0$ is, using the actual constants in and , given by \_0 = -/s\_0 + \_[s\_0,0]{}/s\_0 + (2/s\_0) = -2.73434. \[alpha\]
In column 1 of Tab. \[tab1\], we give some selected eigenvalues $\ln(x_c)_n$ obtained numerically from the exact solutions above. Although the expression is mathematically correct only asymptotically for large $n$, we find that the exact numerical eigenvalues agree with the r.h.s. of and within five digits – which corresponds to our numerical accuracy – even for $n=0$.
$n$ $\ln(x_c)_n$ [(QM)]{} $\ln(x_c)_n$ [(WKB)]{} $\delta\ln(x_c)_n$ $1/C_n$
----- ----------------------- ------------------------ -------------------- -----------
0 -2.73434 -2.64717 0.08717 0.3649953
1 -5.85644 -5.77053 0.08591 0.3651889
2 -8.97854 -8.89263 0.08591 0.3651892
10 -33.95535 -33.86943 0.08592 0.3651892
11 -37.07745 -36.99154 0.08591 0.3651892
20 -65.17635 -65.09044 0.08591 0.3651892
63 -199.42668 -199.34078 0.08590 0.3651892
100 -314.94440 -314.85850 0.08590 0.3651892
It is quite remarkable that, with the value $s_0\simeq 1$ given for the present system, the geometrical nature of the Efimov spectrum is preserved with a high numerical accuracy all the way down to $n=0$ in the inverse-squared potential. For appreciably larger values of $s_0$ this would not be the case.
We emphasize once more that the geometric nature of the asymptotic Efimov spectrum, given by equations and , is a direct consequence of the inverse-square behavior of the potential , with its strength determined by the parameter $s_0$. The ground-state energy $E_0$ in , and with it the parameter $\phi_{s_0,0}$ in , is determined by the particular choice of the regularization of the potential at small distances. In the present case, this was done by the hard-wall cut-off at $R_c$ and the corresponding boundary condition. Other forms of regularization would lead to the same geometric scaling as in but to different ground-state energies $E_0$, such as e.g. in [@mofri] where the potential was regularized by the addition of a repulsive $1/R^4$ potential at short distances.
WKB solutions {#secwkb}
-------------
We now present our results using the WKB approximation. Since the WKB method is well known, we need not present it here and refer for details to the text book by Migdal [@migdal], who also discusses explicitly the Langer correction for radially symmetric potentials [@lang].
In the scaled variable $x$, the classical orbits are bounded inside the classically allowed region $x_c \leq x \leq 1$. Applying the Langer correction with the potential , the effective classical potential becomes V(R)=-, (RR\_c) \[VR\] which is identical with the asymptotic potential found from the inverse Abel transform of the TF level density in Sect. II. The classical momentum then is P(R)=, \[pr\] which, in the scaled variable $x$, becomes P(R)==, p\_0=. (x\_cx<1) \[Px\]
The standard WKB wave function in the classically allowed region has the form u\_[in]{}\^[WKB]{}(x) = , (x\_cx < 1) \[WKBin\] where $S_{in}(x)$ is the action integral along the orbit from $x$ to 1: S\_[in]{}(x)=\_[x]{}\^1 p(x)dx=s\_0\_[x]{}\^1 dx. (x\_cx < 1) This integral can be found analytically [@gr] and becomes S\_[in]{}(x) = s\_0 . (x\_cx < 1) \[sin\] The boundary condition that the wave function vanishes at the lower turning point $x=x_c$ leads, like in the quantum-mechanical case, to the quantization of the eigenenergies as shown below.
Outside the classically allowed region ($x>1$), the exponentially decaying WKB wave function has the form u\_[out]{}\^[WKB]{}(x)=, (x)=, (x>1) \[WKBout\] with the action $S_{out}(x)$ given by S\_[out]{}(x) = \_1\^x (x)dx = s\_0 . (x>1) \[sout\] For large $x\gg 1$, the function has – apart from the normalization – the same exponential tail as the exact function : u\_[out]{}\^[WKB]{}(x) \~ (-s\_0 x). (s\_0 x 1) \[wkbtail\] Equating the two asymptotic forms and , we can determine the constant $B$ as B\_n = C\_ne\^[-/2]{}, \[tailmatch\] so that both wave functions agree exactly in the asymptotic tail region for each state $|n\rangle$.
Since both WKB functions and diverge at the upper classical turning point $x=1$, where $p(1)=\kappa(1)=0$, one must regularize these functions. This is done in the standard “connection” [@lang; @migdal] by a linear approximation to the potential $V(R)$ in the neighborhood of $R=R_+$ and matching the corresponding Airy function solution to the above WKB wave functions at some distances on either side of $x=1$. The requirement that the wavefunction be continuous and continuously differentiable at both matching points leads to the relation $A_n=2B_n$ for the normalization constants and the phase $-\pi/4$ in .
We now have to fulfil the lower boundary condition at $x=x_c$: u\_[in]{}\^[WKB]{}(x\_c) = 0 = 0, \[zeros\] where $S_0(E)$ is the action of the primitive periodic orbit going from $x_c$ to $x=1$ and back: S\_0(E)=2\_[x\_c]{}\^1 p(x)dx, which, using and becomes S\_0(E)=s\_0. \[S0\] For $|E|\ll|E_c|$, i.e. for the the shallow states ($n\gg 1$), the second and third terms above become negligible and the leading term reproduces the action (with $E_0=E_c$) obtained from the periodic orbit theory applied to the geometric spectrum, as discussed in Sec. \[secgeo\]. Eq. has the solutions S\_0(E)- = (n+12), n=0,1,2,…which yields the WKB quantization condition S\_0(E\_n\^[WKB]{}) = P(R)dR = 2(n+3/4), n=0,1,2,…\[WKBquant\] Note that the constant 3/4 in differs from the usual value 1/2, which one obtains for smooth potentials, due to the hard-wall reflection at the lower turning point $x_c$.
Using only the asymptotically leading logarithmic term in yields the geometric spectrum E\_n\^[WKB]{} \~ E\_c, (n1) which corresponds to (x\_c)\_n \~ -n/s\_0 -3/4s\_0 = -n/s\_0 - 2.34158. (n1) This is the same as the quantum-mechanical result , apart from a different shift (denoted by $\alpha_0$ there). In column 2 of Tab. \[tab1\], we give the WKB eigenvalues obtained from the quantization condition , using the full action , in terms of the scaled values $\ln(x_c)_n$. We see that they come very close to the exact quantum values. In fact, there remains a slight shift in $\ln(x_c)_n$ that becomes constant for $n\geq 1$ (cf. the third column in Tab. \[tab1\]). A similar (but different) shift between the asymptotic exact and WKB spectra has been obtained for an attractive $1/R^2$ potential regularized differently [@mofri] (see [@fried] for its interpretation).
We now have to regularize the WKB functions , near the turning point $(x=1)$ where they diverge. While the standard connection to the Airy solution of the locally linearized potential leads to the WKB quantization condition , as described above, we found that it does not yield satisfactory wave functions. The reason is that the asymptotic Airy solutions do not come sufficiently close to the WKB solutions on either side of the turning point. We were, however, successful when using Langer’s generalized “connection formula” for one isolated classical turning point \[Ref. [@lang], Eqs. (11a) and (11b) with $\eta=0$\]. Expressable in terms of a single Airy function Ai$(\xi)$, it reads u\_n\^[uni]{}(x) = D\_n[Ai]{}(),x (x\_c)\_n \[uuni\] where $D_n$ is a normalization constant, and $$\begin{aligned}
\xi & = & \left[\frac32\,\frac{1}{\hbar}\,S_{out}(x)\right]^{2/3},\quad
~{\cal S}(x) = \frac{S_{out}(x)}{\hbar \kappa(x)}\qquad\!\hbox{for}
\quad x\geq 1\,,\;\;\xi \geq 0\,,\label{xiout}\\
\xi & = & -\left[\frac32\,\frac{1}{\hbar}\,S_{in}(x)\right]^{2/3},\quad
{\cal S}(x) = \frac{S_{in}(x)}{\hbar p(x)}\qquad\hbox{for}
\quad x\leq 1\,,\;\;\xi \leq 0\,.\label{xiin}\end{aligned}$$ The superscript “uni” in stands for “uniform”, because it turns out that is a global uniform approximation that can be used not only in the vicinity of the classical turning point, but throughout the whole domain $(x_c)_n \leq x < \infty$. By construction [@lang], it yields the asymptotic WKB solutions , sufficiently far from the turning point $x=1$.
{width="15.3cm"}
\[fig3\]
The normalized uniform WKB wavefunction for $n=5$ is shown in Fig. 3, along with the exact one, in a doubly logarithmic plot. Apart from the slightly shifted zeros, there is a surprisingly good agreement. Fig. 4 shows the ground-state wave function for $n=0$, both exactly (solid line), the “raw” WBK approximation (dotted line), and in the uniform approximation (dashed line). Even for this lowest state, the quantum-mechanical and the uniform WKB solutions are practically indistinguishable.
[*Note added after publication of this paper:*]{}
The above global uniform approximation is also discusse in:\
C. M. Bender and S. A. Orszag: [*Advanced Mathematical Methods for Scientists and Engineers*]{} (Springer-Verlag New York 1991), Ch. 10.4, pp. 510 ff.
{width="15.3cm"}
\[fig4\]
Mean squared radii
------------------
The size of our system in the $n$-th state is given by the mean squared hyperradius defined as R\^2 = = R\_+\^2(E\_n)\_[x\_c]{}\^[1]{} dxx\^2|u\_n(x)|\^2, where Eq. and the normalized wavefunctions $u_n(x)$ have been used. Since $R_+^2$ scales like $1/E_n$, see , the ratio of mean squared radii of two successive states is $E_n/E_{n+1}$, if the expectation values $\langle n|x^2|n\rangle$ remain independent of $n$.
That this is, indeed, the case for all $n\geq 1$, is demonstrated in Tab. \[tab2\], both for the quantum-mechanical and the uniform WKB solutions. Thus the mean squared radius increases by the same geometric scaling factor as the energy decreases, as discussed in [@bra], also in the WKB approximation.
$n$ $\langle n|x^2|n\rangle$ [(QM)]{} $\langle n|x^2|n\rangle$ [(WKB)]{}
----- ----------------------------------- ------------------------------------
0 1.3265 1.3408
1 1.3251 1.3392
2 1.3251 1.3392
5 1.3251 1.3392
10 1.3251 1.3392
Summary
=======
To summarize, we have shown that the ($s$-wave) geometric spectrum of the Efimov energy levels in the unitary limit is semiclassically generated by a single periodic orbit whose action depends logarithmically on the energy. The smooth part of the $s$-state level density, obtained by the periodic orbit theory, is consistent with an attractive inverse-squared radial one-body potential. We have re-derived the quantum spectrum of the inverse-squared potential with a lower cut-off and shown, for $s_0\simeq 1$, that it reproduces the geometric Efimov spectrum not only for shallow states, but yields the same constant ratio $E_{n+1}/E_n$ all the way down to the ground state with $n=0$. We have given an analytical expression for the zeros of the eigenfunctions. The WKB quantization of the classical system (including the Langer correction) yields the same spectrum, although slightly phase shifted, which preserves the same constant ratio $E_{n+1}/E_n$ down to $n=0$. The action of the classical system has as its leading term precisely the action obtained from the periodic orbit theory. The WKB wave functions, when regularized around the classical turning point using Langer’s generalized connection formula, reproduce the exact ones surprisingly well, apart from the slightly shifted zeros – even for the ground state ($n=0$). Both the quantum and the WKB solutions reproduce the geometric scaling of the mean squared radii of the Efimov states which is inverse to that of their eigenenergies.
As mentioned in the introduction, the Efimov spectrum is experimentally not measured in the unitary limit, but for finite and large scattering lengths $|a|\gg r_0 $. Strictly speaking, therefore, the $1/R^2$ nature of the effective potential is only guaranteed in the range $R_c \leq R \siml |a|$. Then, the geometric scaling of the spectrum would only hold within the corresponding energy range, the approximate number of Efimov bound states being given by Eq. . Also, the derivation of the inverse-squared potential given in Sec. II and the analytical exact wavefunctions given in Sec. III A would no longer hold. However, once we assume an inverse-squared potential in the range $R_c \leq R \leq |a|$, our WKB calculations still go through provided we restrict the outer turning point to $R_+ \leq |a| $.
R.K.B. and M.B. are grateful to the IMSc, Chennai, for its hospitality, excellent working conditions and financial support. We are also grateful to an anonymous referee for valuable suggestions. M.B. acknowledges enlightening discussions with H. Friedrich, C. Eltschka and D. L. Sprung.
[99]{}
V. Efimov, Phys. Lett. B [**33**]{}, 563 (1970); Sov. J. Nucl. Phys. [**12**]{}, 589.
T. Kraemer [*et al.*]{}, Nature [440]{}, 315 (2006).
G. Barontini [*et al.*]{}, Phys. Rev. Lett. [**103**]{}, 043201 (2009).
M. Zaccanti [*et al.*]{}, Nature Phys. [**5**]{}, 586 (2009);\
N. Gross [*et al.*]{} Phys. Rev. Lett. [**103**]{}, 163202 (2009);\
S. E. Pollack, D. Dries, and R. G. Hulet, Science [**326**]{}, 1683 (2009).
K. S. Gupta and S. G. Rajeev, Phys. Rev. D [**48**]{}, 5940 (1993).
E. Braaten and H.-W. Hammer, Phys. Rep. [**428**]{}, 259 (2006).
A. M. Essin and D. G. Griffiths, Am. J. Phys. [**74**]{}, 109 (2006).
M. C. Gutzwiller, J. Math. Phys. [**12**]{}, 343 (1971);\
R. Balian and C. Bloch, Ann. Phys. (N.Y.) [**69**]{}, 76 (1972);\
M. V. Berry and M. Tabor, Proc. R. Soc. Lond. A [**349**]{}, 101 (1976).
M. Brack and R. K. Bhaduri: [*Semiclassical Physics*]{} (Bolder, Westview Press, 2003).
R. E. Langer, Phys. Rev. [**51**]{}, 669 (1937).
Hua Wu and D. W. L. Sprung, Phys. Rev. A [**48**]{}, 2595 (1993).
G. Arfken: [*Mathematical Methods for Physicists*]{}, 2nd edition (Academic Press, 1970).
R. K. Bhaduri, A. Chatterjee, and B. P. van Zyl, Am. J. Phys. [**79**]{}, 274 (2011). Note that the prefactor in brackets in Eq. (A.4) should rooted: $(m/2\hbar^2)^{1/2}$.
T. M. Dunster, SIAM J. Math. Anal. [**21**]{}, 995 (1990).
M. Abramowitz and I. A. Stegun: [*Handbook of Mathematical Functions*]{}, 9th printing (New York, Dover, 1972).
M. Moritz, C. Eltschka, and H. Friedrich, Phys. Rev. A [**64**]{}, 042102 (2001); see also\
H. Friedrich in: [*Proceedings of the fourth international conference on dynamical systems and differential equations*]{} (held in Wilmington, USA, 2002), Eds. W. Feng, S. Hu, and X. Lu (AIMS Press, USA, 2003), p. 288.
A. B. Migdal: [*Qualitative methods in quantum theory*]{} (Reading, Benjamin, 1977)
I. S. Gradshteyn and I. M. Ryzhik: [*Table of Integrals, Series, and Products*]{} (New York, Academic Press, 5th edition, 1994).
H. Friedrich: [*Theoretical Atomic Physics*]{} (Springer, Berlin, 1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We are re-examining the problem of stellar migration in disc galaxies from a diffusion perspective. We use for the first time the formulation of the diffusion rates introduced by @1979PhR....52..263C, applied to both energy $E$ and angular momentum [$L_\mathrm{z}$]{} in self-consistent N$-$body experiments. We limit our study to the evolution of stellar discs well after the formation of the bar, in a regime of adiabatic evolution. We show that distribution functions of Chirikov diffusion rates have similar shapes regardless the simulations, but different slopes for energy and angular momentum. Distribution functions of derived diffusion time scales $T_D$ have also the same form for all simulations, but are different for $T_D(E)$ and $T_D(L_\mathrm{z})$. Diffusion time scales are strongly dependent on [$L_\mathrm{z}$]{}. $T_D(E) \lesssim 1$ Gyr in a [$L_\mathrm{z}$]{} range roughly delimited by the set of stellar bar resonances (between the Ultra Harmonic Resonance and the Outer Lindblad Resonance). Only particles with low [$L_\mathrm{z}$]{} have $T_D(L_\mathrm{z}) \lesssim 10$ Gyr, i.e. the simulation length. In terms of mass fraction, 35 to 42% turn out to diffuse energy in a characteristic time scale shorter than 10 Gyr, i.e. simulations length, while 60 to 64% undergo the diffusion of the angular momentum on the same time scale. Both the diffusion of [$L_\mathrm{z}$]{} and $E$ are important in order to grasp the full characterisation of the radial migration process, and we showed that depending on the spatial region considered, one or the other of the two diffusions dominates.'
author:
- Hervé Wozniak
bibliography:
- 'diffusion.bib'
title: Stellar migration in galaxy discs using the Chirikov diffusion rate
---
Introduction {#sec:introduction}
============
Stellar migration of the galactic disc stars has been invoked as a dynamical mechanism to explain the dispersion of stellar metallicity observed in the solar neighbourhood. The age$-$metallicity relation (AMR) shows that the dispersion of stellar metallicity increases with the age of the stars . Another relation, the age$-$velocity dispersion relation (AVR) suggests the existence of a heating mechanism of the stellar disc. Stellar migration could then also be a cause, although other mechanisms have been proposed.
However, stellar migration does not take place in a hypothetical perfectly axisymmetric disc, made of stars rotating in circular orbits. One or more gravitational perturbations are at the origin of any deviation from this hypothetical perfection. These perturbations can be intrinsic (density waves such as bars or spiral arms, two-body relaxation, ...) or extrinsic (galaxy satellites, encounters, mergers, gas accretion...). Also, the amplitude of the perturbations, and thus their ability to reproduce observations, depends on the underlying physical mechanism invoked.
Two classes of models have recently been proposed. They have been the focus of attention since then. All of them distinguish between the effects of blurring (i.e. the radial migration of a star is due to epicyclic motion around a fixed guiding radius), and churning (i.e. the radial migration is due to a change in this guiding radius).
@2002MNRAS.336..785S have shown that spiral waves, possibly transient, have the ability to modify the angular momentum of stars without changing the distribution function, so that the disc does not heat up as a result of these changes. These angular momentum changes essentially result in a variation in the mean radius of stellar orbits over time while keeping their low eccentricity. The dominant mechanism is thus churning. These spiral waves have their own pattern speeds with which stars may resonate. @2002MNRAS.336..785S, confirmed by @2012MNRAS.426.2089R, have shown that angular momentum exchanges take place mainly at corotation. Therefore, the corotation scattering mechanism might be responsible for stellar migration.
For @2010ApJ...722..112M, resonances are also responsible for stellar migration, but their mechanism differs somewhat from @2012MNRAS.426.2089R. Indeed, @2010ApJ...722..112M, confirmed by , consider the interactions between a stellar bar and a spiral structure. In this case, at least two patterns exist and the resonances may overlap. Resonance overlap introduces additional chaos by increasing the efficiency of orbit scattering, which also modifies the angular momentum of stars. Indeed, motion in chaotic regions can be diffusive
We are therefore faced with two mechanisms, which are not irreconcilable from the point of view of galactic dynamics, but which have different observational consequences on the AMR and AVR. Indeed, a related important question for the AVR is whether or not these phenomena cause just the right amount of chaos in the disc to explain increases in the velocity dispersion over time. Whether we deal with stellar migration, disc heating, or stochasticity, stellar motions can be studied from different viewpoints. Indeed, each analysis uses a different methodological framework but the fundamental observational fact is that stars do not stay at their birth site. Therefore, the fundamental question, which is still under debate, is not ultimately to know which dynamical process is solely responsible for the radial migration of stars, but rather what are the relative intensities of each of these processes, whether they contribute together to the same phenomenon, or whether they are ultimately only different points of view of the same phenomenon whose root cause should still be determined.
We have decided to tackle the problem with tools of non-linear physics. This article is only a preliminary step towards answering the fundamental questions mentioned above. Our approach here is to reanalyse the [*diffusion*]{} of quantities such as energy and angular momentum. We have measured diffusion by applying for the first time the Chirikov diffusion coefficient to galactic N-body simulations. have already addressed the issue but in the general context of Fokker-Planck diffusion.
After some fundamental considerations on the dynamics of a rotating disc subjected to perturbations (Section \[sec:theory\]), we introduce the Chirikov diffusion rate in Section \[sec:chirikov\], then N$-$body simulations on which we have applied this tool (Section \[sec:nbody\]). Sections \[sec:chirikov\_results\] and \[sec:timescales\] are dedicated to the analysis of the results that will be discussed in Section \[sec:discussion\]. Section \[sec:axi\] focuses on an axisymmetric case for the sake of comparison.
Angular momentum and energy variations {#sec:theory}
======================================
The dynamics of a rotating stellar disc forced by a spiral or a bar (or any other driving force) is well-known and reminded by @2002MNRAS.336..785S in the context of stellar migration. For any galactic dynamical system subjected to one perturbative frequency, the only integral of motion is Jacobi’s integral defined as: $$\label{eq:one}
E_\mathrm{J} = E - \Omega_\mathrm{p} \, L_\mathrm{z}$$ where $E$ is the classical energy in the non-rotating inertial frame, [$L_\mathrm{z}$]{} is the angular momentum component on the z-axis ${\mathbf{e_z}}$ chosen to be conveniently the rotation axis, and assuming that ${\mathbf{\Omega}} =
\Omega_p\,{\mathbf{e_z}}$, the frequency of the perturbation. $E$ is the sum of the kinetic and potential energy in a closed system. Therefore, since $\Delta E_\mathrm{J} \equiv 0$, any variation of $E$ is linearly related to [$L_\mathrm{z}$]{} and vice-versa as $\Delta E =
\Omega_\mathrm{p}\,\Delta L_\mathrm{z}$.
From the point of view of Hamiltonian dynamics, the existence of two pattern speeds in a galactic disc is similar to a physical system with motions on two different time scales, the so-called slow-fast systems. The fast system could be the bar which has the greater pattern speed whereas a spiral structure could be the slow one. The same situation occurs with the phenomenon of the bar-in-the-bar . For such systems, adiabatic invariants are important dynamical quantities as approximate integral of motion: on the one hand the motion over long time ranges is almost regular if several such adiabatic invariants exist. On the other hand, dissolution of these invariants is one possible mechanism for onset of chaotic dynamics. Indeed, resonant phenomena in fast motion lead generally to dynamical chaos and transport in large regions of the phase space as they destroy adiabatic invariance.
![Mass distribution in the plane $\Delta E$ vs [$\Delta L_\mathrm{z}$]{} for [`RunC`]{} (cf. Section \[sec:nbody\]) between times $t\!=\!3.16$ and $t\!=\!3.27$ Gyr. Colorbar is scaled in log([M$_{\sun}$]{}) per bin. The dashed line is the location of $\Delta E =
\Omega_p\,\Delta L_\mathrm{z}$, where $\Omega_p$ is the bar pattern speed.[]{data-label="fig:dedlz_pfaa"}](PFAA3100-PFAA3000_dE_dLz.pdf){width="\hsize"}
If several patterns coexist in the disc, each one being possibly variable, the $\Delta E$ vs $\Delta L_\mathrm{z}$ plane must exhibit several coexisting slopes. This is what simulations of Section \[sec:nbody\] would suggest. Indeed, Figure \[fig:dedlz\_pfaa\] shows that other patterns exist. Therefore, not all mass transfers have the stellar bar as the responsible party. This figure shows also that there is a continuum of $\Delta E$ and $\Delta L_\mathrm{z}$ values for which mass transfers occurs.
In fact, working with the ratio $\Delta E / \Delta L_\mathrm{z}$ (as @2002MNRAS.336..785S and @2012MNRAS.426.2089R did) masks too much the complexity of the redistribution in $E$ and [$L_\mathrm{z}$]{}, as this technique focuses on the dominant patterns, whether they are due to the bar or the spiral structures. Alternatively, the study of variations in $E$ and [$L_\mathrm{z}$]{} can be based on the difference of these quantities between two times. These times can be distant: this is the case between final and initial values, the notion of “final” being understood here as representative of a typical state of the galaxy. This approach was adopted by all stellar migration studies since @2002MNRAS.336..785S, but only for [$L_\mathrm{z}$]{}, with $\Delta
L_\mathrm{z} = L_\mathrm{z} (t) - L_\mathrm{z}(0)$.
$\Delta L_\mathrm{z}(\Delta t) =
L_\mathrm{z}(t_2)-L_\mathrm{z}(t_1)$ is another approach used by, for instance, @2012MNRAS.426.2089R. It has the merit of focusing on the relative variations with respect to an earlier time, whatever the meaning to be given at that time. The two times can be close, and as close as we want, so that the difference tends towards a differential. If the study is limited to the consequences of the development of certain structures (stellar bar, spiral arms for example), the initial state can be chosen wisely in order to isolate the perturbation created by these structures. Finally, pushed at infinitely small time intervals, this formulation expresses an instantaneous variation of the angular momentum $\dot{L_\mathrm{z}}$ which is related to the net torque acting on the system of particles.
It should be recalled that, at the level of individual particles, taking a particular time as a reference situation is not necessarily more correct than taking the initial time. Indeed, because of the combined effect of relaxation and bar formation, instantaneous individual angular momenta may not be representative of time-averaged angular momenta. For instance, if a particle is able to move [ *alternatively*]{} outward and inward in radius while preserving the circularity of its orbit (what is typical of an epicyclic orbit with low $\kappa$ frequency), it contributes to the instantaneous $\Delta
L_\mathrm{z}$ taken at any particular time. But averaged on several rotations, the time-averaged $\overline{L_\mathrm{z}}$ is only representative of the mean radius, and thus $\Delta
\overline{L_\mathrm{z}} \sim 0$. On the contrary, if that particle moves adiabatically outward or inward, exclusively, then $\Delta
\overline{L_\mathrm{z}} \neq 0$. The particle then moves to a nearby region of the phase space, its mean radius and/or rotational velocity having been modified. There is diffusion.
It is therefore necessary to average the measurements in one way or another, both on the angular momentum and the energy of individual particles. The averaging is intended to cancel the influence of bounded energy/angular momentum oscillations and emphasise the accumulating changes, related to the diffusion process.
Introduction to Chirikov diffusion rate {#sec:chirikov}
=======================================
The diffusion of $E$ means that the energy of the body, as a result of the accumulation of small random variations, can take larger, as well as smaller values, as compared to the unperturbed energy. Similarly, diffusion of [$L_\mathrm{z}$]{} means that even if the trajectory of unperturbed motion of the body was close to circular, a perturbation may bring about trajectories with high eccentricity. If $E$ increases at constant rotational velocity, it will necessarily generate an increase in [$L_\mathrm{z}$]{} because the radius increases. But a similar effect can be achieved by increasing the velocity. Since positions and velocities vary together, the real diffusion rate depends on both $E$ and [$L_\mathrm{z}$]{}.
In this context, the [*Chirikov diffusion rate*]{} [@1979PhR....52..263C eq. 4.6] appears to be a natural choice. Applied on $E$, for individual particles, it is defined as : $$\label{eq:chirikov}
D_n(E) = \overline{(\Delta\overline{E})^2/\Delta t}.$$ Although the original definition deals only with $E$, we may extend the definition of Equation (\[eq:chirikov\]) to compute $D_n(L_\mathrm{z})$. In Equation (\[eq:chirikov\]), $\overline{E}$ is the value of energy averaged over a period of $\Delta t_n=10^n$ (in time unit of the system). In our case, it is convenient to choose $\Delta
t_n \equiv \Delta t_2$ as the minimum time that separates two snapshots, i.e. 100 code units (105.49 Myr). Indeed, the whole simulation (10.54 Gyr) is naturally segmented at regular intervals. $\overline{E}$ and $\overline{L_\mathrm{z}}$ are thus computed on the fly for each particles and stored with snapshots.
$\Delta\overline{E}$ is then the difference between two intervals (snapshots), $\Delta t$ being the time difference between the snapshots. These snapshots are not necessarily consecutive because the second averaging concerns all possible pair combinations. This procedure, initiated by @1979PhR....52..263C, ensures that all time scales are represented by the definition of $D_n$.
Several other definitions of $D_n$ exist and an abundant literature concerns the interpretation to be given to the evolution of $D_n$ with the [*strength*]{} of the perturbation(s), and its asymptotic behaviour when $n$ increases [@1992rcd..book.....L]. As a general rule, $D_n$ sharply increases above a certain threshold of perturbation strength meaning that the motion is moving from regular to chaotic. We are not looking here for any critical value of the perturbation strength, as the definition of this strength can be the subject of much debate. Indeed, each region of a galactic disc is subjected to perturbations of different intensities while all these regions remain connected through gravitation. It is therefore very difficult to highlight particular threshold values of the perturbation intensity in a large N$-$body system, considered as a whole. Our objective is rather to qualify the different types of particle populations with noticeable $D_n$ values, or range of $D_n$, possibly different from one region to another. This allows to determine whether some regions are more stochastic than others and, if so, whether $E$ or [$L_\mathrm{z}$]{} is the more diffusive quantity.
N$-$body experiments {#sec:nbody}
====================
Several simulations have been performed to check the dependence of our results on certain quantities, such as the total mass, or the initial scale of the stellar disc. Rather than starting from a cosmological situation, including all kinds of effects that are difficult to control (accretion of dwarf galaxies, cold gas flows, star formation, etc.), we preferred to start with an idealised situation.
Initial stellar populations are set up to reproduce such idealised, but typical, disc galaxies. Positions and velocities for $N_\mathrm{s}$ particles are drawn from a superposition of two axisymmetrical @mn75 discs of mass $M_1$ and $M_2$ (cf. Table \[tab:simul\]), of scale lengths $l_1$ and $l_2$ kpc and a scale height $h_z$. Initial positions have been truncated to $R=30$ kpc for [`RunA`]{} and [`RunB`]{}, $R=40$ kpc for the more massive [`RunC`]{}. Scale lengths and scale heights have been chosen such as the superposition of the two axisymmetric distributions shapes the initial spatial configuration as disc galaxy with a small but significant bulge.
Initial velocity dispersions are computed solving numerically the Jeans equations according to the @1993ApJS...86..389H method. The initial velocity dispersion was chosen to be anisotropic with $\sigma_r = \sigma_z$ and $\sigma_{\theta}^2 = \sigma_r^2 \kappa^2 /
(4\Omega^2)$, where $\sigma_r$, $\sigma_{\theta}$ and $\sigma_z$ are three components of the velocity dispersion along respectively the radial, azimuthal and vertical directions and $\kappa$ and $\Omega$ are respectively the radial and angular epicyclic frequencies. They are related by $\kappa^2 = 4\Omega^2+rd\Omega^2/dr$
![Initial $Q$ Toomre parameter as a function of radius for the three simulations.[]{data-label="fig:qq"}](fig_qq.pdf){width="0.75\hsize"}
As the Toomre parameter ($Q=\sigma_r\,\kappa/(3.36 G \mu)$ where $\mu$ is the mass surface density and $G$ the gravitational constant) has not been explicitly constrained, the resulting values are displayed in Figure \[fig:qq\]. The three simulations are unstable ($Q
< 1$) in their central region, i.e. at roughly one scale length around the centre. This is typically the region where the bar is formed.
[@llllllll@]{} Model & $N_\mathrm{s}$ & $M_1$ & $M_2$ &$l_1$ &$l_2$&$h_z$ &$\times 10^{7}$&$\times 10^{10}$ M$_{\sun}$&$\times 10^{10}$ M$_{\sun}$&(kpc)&(kpc)&(kpc) [`RunA`]{}& 4. & 0.63 & 3.57 &0.57 &2.0 & 0.5 [`RunB`]{}& 4.4 & 1.1 & 11.0 &1.0 &3.5 & 0.5 [`RunC`]{}& 4. & 3.0 & 17.0 &1.14 &4.0 & 1.0
[`RunC`]{} and [`RunA`]{} use similar initial parameters than respectively and . [`RunB`]{} has similar initial conditions than the run named “SimS” in but is made exclusively of stellar particles for the same total mass. All runs are computed until 10.54 Gyr.
![Initial (solid lines) and final (dotted lines) distribution functions (DF) for the three simulations. DFs have been normalised to DF(0) which is also the maximum. The bump in the final DF, around $L_\mathrm{z}\approx$ 1100, 2500 and 3500 [kpckms$^{-1}$]{} for, resp., [`RunA`]{}, [`RunB`]{}, and [`RunC`]{}, is typical of the ‘hot’ particle population which is able to cross the corotation and explore both the bar and the disc.[]{data-label="fig:fdlzinit"}](fig_fdlz.pdf "fig:"){width="0.75\hsize"} ![Initial (solid lines) and final (dotted lines) distribution functions (DF) for the three simulations. DFs have been normalised to DF(0) which is also the maximum. The bump in the final DF, around $L_\mathrm{z}\approx$ 1100, 2500 and 3500 [kpckms$^{-1}$]{} for, resp., [`RunA`]{}, [`RunB`]{}, and [`RunC`]{}, is typical of the ‘hot’ particle population which is able to cross the corotation and explore both the bar and the disc.[]{data-label="fig:fdlzinit"}](fig_fde.pdf "fig:"){width="0.75\hsize"}
The evolution is computed with a particle–mesh N$-$body code, derived from the original version of the Geneva group . The broad outline of the code is the following: the gravitational forces are computed with a particle–mesh method using a 3D log–polar grid with $(N_R, N_\phi,
N_Z)=(60,64,312)$ active cells. The smallest radial cell in the central region is 36 pc large and the vertical sampling is 50 pc. The extent of the mesh is 100 kpc in radius and $\pm 7.8$ kpc in height. Since we used a polar grid and we need an accurate determination of the forces in the central region, we have improved the pre-computation of self-forces by subdividing each cell in $(n_r, n_\phi,
n_z)=(32,6,6)$ subcells. Self-forces are then linearly interpolated before being subtracted from the gravitational forces.
In a perfectly collisionless simulation of a stable equilibrium model, each particle would conserve its specific energy. The combination of a particle-mesh code, an initial relaxed distribution and a large number of particles ensures that the sources of numerical diffusion are minimised. However, we have also performed a control run ([`RunC^\mathtt{axi}`]{}) which will be detailed in Section \[sec:axi\].
For convenience, the units in which the discussion will be conducted have been chosen to avoid the power of 10. Thus, the specific angular momenta will be in [kpckms$^{-1}$]{} while the specific total energies will be in [kpc$^2$Myr$^{-2}$]{}. In addition, since all particles have the same mass, all particle number distributions can also be read as mass fractions. The initial and final distribution functions (DF, cf. Figure \[fig:fdlzinit\]) are typical of such type of simulation. Expressed as a function of [$L_\mathrm{z}$]{}, these DF display similar trends in regard to the DF obtained by 3D $N$-body simulations, e.g. those of @1978ApJ...226..521Z, @1987MNRAS.225..653S or . The shape of these DFs has been explained by a superposition of various families of orbits . Orbits of the bump (around $L_\mathrm{z}\approx$ 1100, 2500 and 3500 [kpckms$^{-1}$]{} for, resp., [`RunA`]{}, [`RunB`]{}, and [`RunC`]{}) are mostly disc orbits which also populate the corotation region of the bar. These orbits spend most of their time outside the bar and sometimes enter inside the bar from the $L_{1,2}$ Lagrangian points. This last kind of orbits as well as Lagrangian orbits form the ‘hot’ population described first by @1987MNRAS.225..653S. This ‘hot’ population may contribute up to 30% of the total mass.
![Evolution of $\overline{L_\mathrm{z}}$ (top) and $\overline{E}$ (bottom), for 3847 particles selected for [`RunC`]{} (cf. Sect \[sec:nbody\] for details) at $t\!=\!3.27$ Gyr with $\overline{L_\mathrm{z}} = 3000\pm0.3$ [kpckms$^{-1}$]{}. Colorbar is scaled in log([M$_{\sun}$]{}) per bin.[]{data-label="fig:evol"}](Chirikov_part_3000_9999_xdr.pdf){width="\hsize"}
Figure \[fig:evol\] displays the evolution of $\overline{L_\mathrm{z}}$ and $\overline{E}$, for a group of particles selected for [`RunC`]{} at $t\!=\!3.27$ Gyr with $\overline{L_\mathrm{z}}
= 3000\pm0.3$ [kpckms$^{-1}$]{}, typical of the ‘hot’ population. This selection represents 3847 particles. Although these particles are selected over a narrow interval in $\overline{L_\mathrm{z}}$, the values of $\overline{E}$ show initially a larger amplitude, the maximum being around $-0.0075$ [kpc$^2$Myr$^{-2}$]{}. As the group evolves, the amplitude of $\overline{L_\mathrm{z}}$ increases rapidly, until it reaches a range of values from $\approx 0$ (or even negative for some particles) to $\approx 4500$. The $\overline{L_\mathrm{z}}$ distribution mode increases until it reaches a value of $\approx 3500$ at $t\!=\!10.54$ Gyr. Values of $\overline{E}$ also vary over the same time interval. However, its distribution gradually spreads only on the negative side. Let us recall here that these are average values over an interval of $\Delta t_2 \!=\!105$ Myr, i.e. between an half and a quarter of the bar rotation period. Averaging over a larger $\Delta t_n$ does not change the result.
{width="0.33\hsize"} {width="0.33\hsize"} {width="0.33\hsize"}
By looking for potential differences between these simulations, we can focus on power spectra of the $m=2$ frequencies between $t=2.11$ and $10.54$ Gyr for [`RunA`]{} and [`RunB`]{} and between $t=3.16$ and $t=10.54$ Gyr for [`RunC`]{} (cf Sect. \[sec:chirikov\_results\] for explanation of these time ranges). In Fig. \[fig:omega\_init\], the bar frequency largely dominates. Several other patterns exist and give rise to overlaps of resonances. These overlaps are usually temporary because the resonance system linked to the bar slides outwards during the evolution of the disc and the slowing down of the bar. Moreover, since the time window is wide, only long-lived structures appear in this figure. Transient structures, often with a lifetime of less than one orbital period, are erased. However, their role is essential. In a future article, I will analyze more closely their connections with more permanent structures.
Chirikov diffusion rate in N$-$body experiments {#sec:chirikov_results}
===============================================
{width="0.33\hsize"} {width="0.33\hsize"} {width="0.33\hsize"}
As [$D_2(E)$]{} and [$D_2(L_\mathrm{z})$]{} sum up all fluctuations occurring in the disc, we have intentionally restricted the time interval to the epoch well after the bar formation. Therefore, the computation of $D_2(E)$ was performed between $t=3.16$ and $t=10.54$ Gyr for [`RunC`]{} (i.e.70 snapshots) and between $t=2.11$ and $10.54$ Gyr for [`RunA`]{} and [`RunB`]{} (i.e. 80 snapshots), spaced by $\Delta t_2$, i.e. on resp. 2415 and 3160 unique pairs. The starting times have been chosen in order to avoid the strong perturbations caused by the formation of the bar, which are not of interest to us here. This rules out the strong redistribution in $E$ and [$L_\mathrm{z}$]{} made by the formation of the bar. Doing so, we can examine the impact of driving forces in a quieter phase of the galaxy.
Although particles escaping the grid are tracked throughout their trajectory by a ballistic approximation, we chose to exclude them from our analyses as soon as they came out of even one time step. This drastic procedure ensures that we limit numerical errors to their lowest values.
Figure \[fig:chirikov\] shows the distributions for [$D_2(E)$]{} and [$D_2(L_\mathrm{z})$]{} for the three simulations. A first lesson that can be drawn from these figures is the universality of the distributions shape when $D_2$ is normalised to its maximum. Approximatively, $$\log n(D_2)\propto \gamma D_2 / \max(D_2)$$ where $\gamma$ is different for [$D_2(E)$]{} and [$D_2(L_\mathrm{z})$]{}, and $n(D_2)$ is the fraction of particles number or, equivalently, the mass fraction. $\gamma$ is close to $-5.0$ for [$D_2(E)$]{} and between $-6.28$ and $-6.79$ for [$D_2(L_\mathrm{z})$]{}. The shape is represented by a linear regression valid over a wider range of $D_2/\max(D_2)$ for [$L_\mathrm{z}$]{} than for $E$.
Deviations from a linear fit are also instructive. For [$D_2(E)$]{}, two regions deserve to be commented on. The three distributions show a dip for $D_2(E) / \max(D_2(E)) < 0.04-0.06$. It accounts for a maximum of 30% of the total mass. The second region is at the opposite: the distribution drops when [$D_2(E)$]{} reach $\approx 80$ % of the maximum.
However, the normalisation by $\max(D_2)$, which allows to compare the profiles between them, masks an important element. Indeed, these maxima are different from one simulation to another, in a sensitive way because they approximately scale with the square of the total energy or angular momentum. Table \[tab:max\] gives the values of these maxima.
[@lrll@]{} Model & $\max(D_2(E))$ & $\max(D_2(L_\mathrm{z}))$ & $M_\mathrm{tot}$ & [(kpc$^4$Myr$^{-4}$)Myr$^{-1}$]{} & [(kpc$^2$km$^2$s$^{-2}$)Myr$^{-1}$]{} & [M$_{\sun}$]{} [`RunA`]{}& $3.93\times 10^{-8}$ & $ 276.7$ & $4.2\times10^{10}$ [`RunB`]{}& $1.43\times 10^{-7}$ & $ 859.2$ & $1.2\times10^{11}$ [`RunC`]{}& $3.32\times 10^{-7}$ & $2292.8$ & $2.0\times10^{11}$ [`RunC^\mathtt{axi}`]{} & $4.46\times 10^{-7}$ & $ 145.6$ & $2.0\times10^{11}$
Diffusion time scales {#sec:timescales}
=====================
{width="0.33\hsize"} {width="0.33\hsize"} {width="0.33\hsize"}
{width="0.33\hsize"} {width="0.33\hsize"} {width="0.33\hsize"}
The interpretation of [$D_2(E)$]{} may seem complicated because this quantity mixes information on the quadratic evolution of the $E$ fluctuations at different time scales. Large fluctuations of $E$ over long times can contribute as much as small fluctuations over very short times. Formally, we can also estimate a diffusion time scale $T_D$ by renormalizing $D_n$ by $E^2$ and $L_\mathrm{z}^2$ respectively. As Chirikov diffusion rate takes care of all sources of perturbation, such as particle-wave interactions, it can be seen as a generalisation of several diffusion time definitions, such as @1942psd..book.....C’s one on the two-body relaxation times of stellar systems.
The diffusion time scale, defined as: $$T_D(E) = \overline{E}^2 / D_2(E)$$ for each individual body, may thus seem more intuitive. The same definition holds with [$L_\mathrm{z}$]{} to compute $T_D(L_\mathrm{z})$. $E^2(t\!=\!0)$ or $L^2_\mathrm{z}(t\!=\!0)$ can be used instead of respectively $\overline{E}^2$ or $\overline{L_\mathrm{z}}^2$ without any significant change. The results for all three simulations are displayed in Figure \[fig:timescales\] and Figure \[fig:timescalesLz\], where the frequency distribution of particles (or mass fraction since all particles have the same individual mass) is plotted against $T_D(E)$ and $T_D(L_\mathrm{z})$. For the sake of clarity, we have restricted these figures to the range $10^{-4} - 90$ Gyr, but $T_D$ can reach much higher values for a few particles.
An obvious outcome is the similarity between the distributions for the three simulations. This form of universality is primarily linked to the similarities of [$D_2(E)$]{} and [$D_2(L_\mathrm{z})$]{} distributions for the three simulations. It is also due to the shape of the distribution functions DF($E$) and DF([$L_\mathrm{z}$]{}) (Figure \[fig:fdlzinit\]) which, although they differ in detail, share the same form. Moreover, it should be stressed here that the time scale chosen is absolute, in Gyr, and not normalised to a maximum as we have done in Figure \[fig:chirikov\]. Time scales are thus quantitatively comparable in terms of values.
$E$ diffusion time scale
------------------------
Dealing first with $T_D(E)$ distribution only, a first region appears between 0.1 Myr (the minimum time step observed during numerical integration) and a local minimum located at $\approx 5$ Gyr ([`RunA`]{}), $\approx 7$ Gyr ([`RunB`]{}), and $\approx 4$ Gyr ([`RunC`]{}). This time range covers most of the dynamical time scales present in the galaxy’s disc, from its central part to its outermost border. The decrease in mass fraction as a function of $T_D(E)$ is slower than exponential. This region represents roughly 33% ([`RunA`]{}), 39% ([`RunB`]{}) and 27% ([`RunC`]{}) of the total mass. It is noteworthy that particles with $T_D(E) \lesssim 0.1$ Myr represent a negligible mass, but $\approx
35-42$% of the total mass lie in the range $T_D(E) < 10.54$ Gyr. Apart from the fact that $T_D(E)$ is calculated on 1 Gyr more for [`RunA`]{} and [`RunB`]{} than for [`RunC`]{}, we did not find any other simple possible cause that would explain these differences in mass fraction. For instance, we do not see any scaling with the total mass or the initial disc scale length. Differences in the evolution of these three simulations, notably the formation of the bar, the emergence of the spiral arms, etc., are possibly at the origin of these differences in mass fraction.
A second feature, a bump centred at $\approx 10$ Gyr for [`RunA`]{} and [`RunB`]{}, and $\approx 8$ Gyr for [`RunC`]{}, might be the footprint of the limited time length of the simulations. On the contrary, no signature due to sampling is detected (i.e. 100 Myr for the calculation of $D_2$).
Finally, $\approx 58-65$% of particles have $T_D(E) >
10.54$ Gyr. This means that most of the mass undergoes energy fluctuations that only become significant over times longer than the simulations length, and therefore, in practice, over times that might be greater than the age of the Universe.
![Distribution function DF([$L_\mathrm{z}$]{}) for [`RunC`]{} at $t=0$. Dashed line: all particles ($N_\mathrm{s}$). Solid line: only particles that neither escaped from the grid and used for $T_D(E)$ computations. Red line: particles with $T_D(E) > 10.54$ Gyr. Green line: $4 < T_D(E) < 10.54$ Gyr. Blue line: $T_D(E) < 4.0$ Gyr. All DFs have been normalised to DF(0) with all particles, which is also the maximum.[]{data-label="fig:tde_fdlz"}](PFAA_fig_tde_fdlz.pdf){width="0.75\hsize"}
In order to understand the properties of the particle populations that contribute to the different time scales, we have plotted in Figure \[fig:tde\_fdlz\] DF([$L_\mathrm{z}$]{}) for various selection of particles made on $T_D(E)$ for [`RunC`]{} (cf. appendix for other simulations). The reference time is the origin of the simulation ($t=0$). Particles with $T_D(E) > 10.54$ Gyr come essentially from populations with $L_\mathrm{z} < 1500$ [kpckms$^{-1}$]{}(red curve in Figure \[fig:tde\_fdlz\]) that are typically well inside the innermost resonances of the bar. For the sake of comparison, an hypothetical circular orbit at the bar Ultra-Harmonic resonance (UHR) at $t=3.16$ Gyr has $L_\mathrm{z}
\approx 2200$ [kpckms$^{-1}$]{}. Although the bar is a major gravitational perturbation, which has the ability to cause significant mass redistribution, the fact that resonances isolate the central region from the rest of the galaxy possibly limits the onset of energy diffusion. Therefore, the diffusion time scales in $E$ are longer than simulation length in the innermost region.
Particles with $4 < T_D(E) < 10.54$ Gyr (green curve in Figure \[fig:tde\_fdlz\]) come from a fraction of the bar population which is close to the corotation barrier. This region contains many bifurcations of orbit families by period doubling [@1983ApJ...275..511C]. An infinite cascade of this type of bifurcation then forms a sequence that leads to stochasticity.
Finally, particles with $T_D(E) < 4$ Gyr come massively from both the ‘hot’ population and the disc. Their diffusion time scale is comparable to or shorter than typical dynamical time scales in the disc.
We can therefore summarise the global trend of $T_D(E)$ by saying that it decreases from the centre to the most external regions. This trend will be further discussed in Sect. \[sec:discussion\] and refined.
[$L_\mathrm{z}$]{} diffusion time scale
---------------------------------------
Dealing now with $T_D(L_\mathrm{z})$, a noteworthy observation is that two slopes appear for $0.001 \lesssim T_D(L_\mathrm{z}) \lesssim 0.3$ Gyr, and $3\lesssim T_D(L_\mathrm{z}) \lesssim 10$ Gyr in a $\log-\log$ diagram. These ranges are those on which a line has been fitted by a standard algorithm of linear regression. Therefore, for $0.001 \lesssim
T_D(L_\mathrm{z}) \lesssim 0.3$ Gyr, $$\label{eq:ntd}
n(T_D) \propto T_D^{\beta_\mathrm{L}}$$ where $\beta_\mathrm{L} \approx -0.42$ in most cases and $n(T_D)$ is the distribution of particle frequencies (or mass fraction). The second slope with index $\beta_\mathrm{H}$ seems to depend on the simulation parameters.
![Same as Figure \[fig:tde\_fdlz\] but for $T_D(L_\mathrm{z})$. The noticeable time scale is now 0.3 Gyr instead of 4 Gyr.[]{data-label="fig:tdlz_fdlz"}](PFAA_fig_tdlz_fdlz.pdf){width="0.75\hsize"}
Below the time scale of 0.3 Gyr, which is also the typical mean bar rotation period for all three simulations, mass is made of particles with low [$L_\mathrm{z}$]{} (Figure \[fig:tdlz\_fdlz\] for the case of [`RunC`]{}). These particles represent only a small mass fraction, between 8.6 and 9.5%.
On the other side of the distribution, 36-40% of the mass has $T_D(L_\mathrm{z}) \ge 10.54$ Gyr. All kinds of orbits contribute to this population, but it should be noted that, in the case of [`RunC`]{}, all particles with $L_\mathrm{z} > 3300$ [kpckms$^{-1}$]{}, i.e. a large fraction of the ‘hot’ population and all disc particles, have very long $T_D(L_\mathrm{z})$.
Finally, about half of the mass (51-54%) has intermediate diffusion times, between 0.3 and 10.54 Gyr. The particles inside the bar form the largest part of this population responsible for the diffusion of the angular momentum. Probably a small fraction of the ‘hot’ population also belongs to this category but it is difficult to quantify its contribution more precisely without a detailed orbit analysis that is postponed to a future paper.
In comparison, the global trend of $T_D(L_\mathrm{z})$ seems to be opposite to that of $T_D(E)$: the diffusion time scale increases with the radius.
The axisymmetric case {#sec:axi}
=====================
An instructive element of comparison is to look at what happens to $D_2(E)$, $D_2(L_\mathrm{z})$, $T_D(E)$, and $T_D(L_\mathrm{z})$ in case a simulation is forced to remain axisymmetric. Both $E$ and [$L_\mathrm{z}$]{} are now isolating integral of motion. Diffusion rates would be zero if the gravitational potential were due to an infinite number of particles. The potential would then be smooth and stationary. The individual energy of the particles would then be perfectly preserved. Poissonian shot noise due to potential discreteness, forces accuracy and the finite number of particles is however unavoidable. We thus need reference values.
![Particle number (or mass fraction) as a function of [$D_2(E)$]{} and [$D_2(L_\mathrm{z})$]{} for [`RunC^\mathtt{axi}`]{} when the mass density is forced to remain axisymmetric. As for Figure \[fig:chirikov\], the number of particles is normalised to $N_\mathrm{s}$ and [$D_2(E)$]{} and [$D_2(L_\mathrm{z})$]{} to their respective maxima (see text for values).[]{data-label="fig:chirikov_axi"}](PFAB_Chirikov4_3000_9999_xdr_0.pdf){width="\hsize"}
[`RunC`]{} was recalculated by forcing the axisymmetrisation of the mass density at each time step, any other parameter being similar to [`RunC`]{}. Let us call it [`RunC^\mathtt{axi}`]{}. The gravitational potential therefore remains close to axisymmetric, no bar or spiral structure can develop. Only axisymmetric waves can propagate in the first Gyr of the simulation, carrying initial angular momentum towards the external regions. For comparison purposes, all rates and time scales were calculated in the same way as [`RunC`]{}, i.e. between $t=3.16$ and $t=10.54$ Gyr.
Regarding $D_2(E)$ (Figure \[fig:chirikov\_axi\]), even if its maximum ($4.46\times10^{-7}$) is close but a little greater than that of [`RunC`]{} (cf. Table \[tab:max\]), the distribution shape is significantly different. Indeed, a large mass fraction has low values of $D_2(E)$ (i.e. less than 10% of the maximum). Beyond that, the distribution is flatter than for [`RunC`]{}, which results in a higher $\gamma$ slope ($\approx -3.8$ instead of $\approx -5$).
For $D_2(L_\mathrm{z})$, not only $\max(D_2(L_\mathrm{z}))\approx 146$ is much lower than for [`RunC`]{} ($\approx 2300$), and this for the same total mass, but the shape of the distribution is no longer close to a linear relation between the mass fraction and $D_2/\max(D_2)$. The scale of $D_2(L_\mathrm{z})$ has been reduced by a factor of 16. This can be easily understood because a large angular momentum diffusion is not expected in an axisymmetric simulation as [$L_\mathrm{z}$]{} is an integral of motion.
![Same as Figure \[fig:timescales\] (top) and Figure \[fig:timescalesLz\] (bottom) for [`RunC^\mathtt{axi}`]{} when the mass density is forced to remain axisymmetric. []{data-label="fig:timescales_axi"}](PFAB_Chirikov_td_e.pdf "fig:"){width="0.75\hsize"} ![Same as Figure \[fig:timescales\] (top) and Figure \[fig:timescalesLz\] (bottom) for [`RunC^\mathtt{axi}`]{} when the mass density is forced to remain axisymmetric. []{data-label="fig:timescales_axi"}](PFAB_Chirikov_td_lz.pdf "fig:"){width="0.75\hsize"}
Unexpectedly, the shape of the $T_D(E)$ distribution (Figure \[fig:timescales\_axi\] top row) between 0.1 Myr and 4 Gyr is significantly different from those shown in Figure \[fig:timescales\]. It looks like the $T_D(L_\mathrm{z})$ distribution for non-axisymmetric simulations. The two slopes $\beta_\mathrm{L}$ and $\beta_\mathrm{H}$ are moreover similar to those displayed in Figure \[fig:timescalesLz\]. If we select the particles of [`RunC^\mathtt{axi}`]{} with $T_D(E) < 2$ Gyr, it appears (Figure \[fig:td\_fdlz\_axi\]) that their [$L_\mathrm{z}$]{} corresponds to the so-called ‘hot’ population, although formally this population cannot exist here because the bar and associated resonances are absent. These particles are therefore the ones likely to be most affected by a perturbation, their diffusion time being already the shortest in the axisymmetric case. Particles with $ 2 < T_D(E) < 10.54$ Gyr would then be orbits located in the region inside the corotation, where the families of orbits undergo bifurcations. Here again, in the axisymmetric case, and therefore in absence of any pattern frequency, there is no specific resonances.
![Distribution function DF([$L_\mathrm{z}$]{}) for [`RunC^\mathtt{axi}`]{} at $t=0$. Dashed lines: all particles ($N_\mathrm{s}$). Solid lines: only particles that neither escaped from the grid and used for $T_D(E)$ and $T_D(L_\mathrm{z})$ computations. Red lines: particles with $T_D(E)$ or $T_D(L_\mathrm{z}) > 10.54$ Gyr. Green lines: $2 < T_D(E) < 10.54$ (top) or $0.3 < T_D(L_\mathrm{z}) <
10.54$ Gyr (bottom). Blue lines: $T_D(E) < 2.0$ Gyr (top) or $T_D(L_\mathrm{z}) < 0.3$ Gyr (bottom). All DFs have been normalised to DF(0) with all particles, which is also the maximum.[]{data-label="fig:td_fdlz_axi"}](PFAB_fig_tde_fdlz.pdf "fig:"){width="0.75\hsize"} ![Distribution function DF([$L_\mathrm{z}$]{}) for [`RunC^\mathtt{axi}`]{} at $t=0$. Dashed lines: all particles ($N_\mathrm{s}$). Solid lines: only particles that neither escaped from the grid and used for $T_D(E)$ and $T_D(L_\mathrm{z})$ computations. Red lines: particles with $T_D(E)$ or $T_D(L_\mathrm{z}) > 10.54$ Gyr. Green lines: $2 < T_D(E) < 10.54$ (top) or $0.3 < T_D(L_\mathrm{z}) <
10.54$ Gyr (bottom). Blue lines: $T_D(E) < 2.0$ Gyr (top) or $T_D(L_\mathrm{z}) < 0.3$ Gyr (bottom). All DFs have been normalised to DF(0) with all particles, which is also the maximum.[]{data-label="fig:td_fdlz_axi"}](PFAB_fig_tdlz_fdlz.pdf "fig:"){width="0.75\hsize"}
The $T_D(L_\mathrm{z})$ distribution of [`RunC^\mathtt{axi}`]{} (Figure \[fig:timescales\_axi\] bottom row) displays now a unique slope with $\beta \approx -0.53$ that extends from 0.01 to $\approx 10$ Gyr. But particles in this range account for only 7% of the total mass compared to more than 50% for the other three non-axisymmetric simulations. Only orbits with very low [$L_\mathrm{z}$]{} (and thus close to the centre) contribute to this region. The rest of the mass (93%) has $T_D(L_\mathrm{z}) \gg 10$ Gyr. This trend is much expected as [$L_\mathrm{z}$]{} is an integral of motion in axisymmetric discs. This illustrate that when a bar, spiral structure, and any other pattern appear in the [`RunC`]{} simulation, these collective oscillations are solely responsible for the diffusion of the angular momentum. As soon as axisymmetry is broken, $T_D(L_\mathrm{z})$ starts decreasing, leading to $T_D(L_\mathrm{z}) < 10.54$ Gyr for circular orbits with $L_\mathrm{z} < L_\mathrm{z}(\mathrm{corotation})$.
A final observation worth mentioning: the number of escaped particles, defined as those having gone outside the grid even one time step, is only 1.5% for [`RunC^\mathtt{axi}`]{}, compared to 14% for [`RunC`]{}.
Discussion {#sec:discussion}
==========
Role of resonances
------------------
At resonances, particles can undergo two types of phenomena that are both important for the galactic dynamics. The first one is capture (or trapping) into the resonance. Any particle follows closely an adiabatic trajectory until it crosses a resonant surface. There, the particle trajectory starts following this surface rather than the adiabatic trajectory, causing strong adiabatic invariant (hereinafter referred to as $I$) changes along the exact motion. As the particle may escape from the resonance (after a finite but unpredictable time), it starts following a different adiabatic trajectory with value of $I$ completely different from the initial one. Initial conditions of particles to be captured and those to cross the resonance without capture are entangled. At the coarse-grained level, this leads to capture probabilities. These capture probabilities are of order of the perturbation strength so that we do not expect a large part of particles to be captured. However, due to the disc rotation, phase trajectories of the averaged system are closed, allowing particles to cross the resonant surface again and again. This significantly increases the probability to be captured on the long term, making resonant capture important for the galactic disc dynamics.
The second phenomenon, scattering, takes place for particles that cross the resonance without capture. The particle trajectory follows closely the adiabatic trajectory, but at the resonant surface it shifts by a very small amount rather than being captured. After crossing the resonance, the particle follows a new adiabatic trajectory, which has moved from a distance of order of the perturbation strength from the original one. The amount by which $I$ changes depends on the initial conditions. As for capture, multiple scattering is possible and lead to diffusion of the adiabatic invariant on the long term. Therefore, particles passing through resonance change their energy even in absence of trapping. However, if the resonant system contains a separatrix, the mean energy change due to scattering is finite. Multiple scatterings lead to either acceleration or deceleration of particles. In absence of such a separatrix, the energy change over multiple scatterings is diffusive.
In the context of epicyclic approximation, localising resonances requires the computation of the circular orbit frequency $\Omega$ and the radial epicyclic frequency $\kappa$. Strictly speaking, these frequencies predict the oscillation frequencies of the orbits in the axisymmetrical case limits only. They do not provide any indication of whether families of periodic orbits do follow such oscillations when the bar growth breaks the axisymmetry. However, a number of previous orbital studies suggest that the epicyclic approximation could lead to an acceptable estimation of the resonance locations, in particular if we are mainly interested in their evolution rather than their accurate absolute position. For instance, using a careful integration of orbits to compute $\Omega$ and $\kappa$, found that the error on the corotation radius remains within 10%.
In the case of barred galaxies, the resonant area is a large region around the idealised corotation radius, in the sense of a radial solution of the equation $\Omega(R_\mathrm{cor})=\Omega_p$ as defined by the linear theory of resonances. Indeed, Lagrangian points are defined as being the points of equilibrium between centrifugal and centripetal forces along the main axis of the bar perturbation. The radii of these points converge towards the corotation circle when the bar perturbation vanishes. [$L_{1,2}$]{} and [$L_{4,5}$]{} are the radii of the Lagrangian points along respectively the major-axis and intermediate axis (minor-axis in a face-on projection) of the bar perturbation. have shown that the relative amplitude of the difference between [$L_{1,2}$]{} and [$L_{4,5}$]{} rarely exceeds 15% even in very strong bar phases. A standard value for a slowly evolving bar seems to be in the range 5 to 10%. Moreover, the amplitude of the difference between [$L_{1,2}$]{} and [$L_{4,5}$]{} is roughly proportional to the bar strength. When the bar gets stronger, the difference between [$L_{1,2}$]{} and [$L_{4,5}$]{} increases, [$L_{4,5}$]{} being always smaller than [$L_{1,2}$]{}. Therefore, the corotation radius always lies between the Lagrangian points radii but is closer to [$L_{1,2}$]{} than [$L_{4,5}$]{}. Thus, when the corotation is mentioned, the region concerned is an oval ring whose width depends on the strength of the bar. @2007MNRAS.379.1155C focused on the action of the bar in the redistribution of [$L_\mathrm{z}$]{} and, in particular, the capture of particles by corotation. Their work clearly shows that the scope of the action of the bar goes well beyond corotation. In particular, they show how strong the variations of [$L_\mathrm{z}$]{} of the particles trapped around stable Lagrange points are.
It is moreover well established that a rotating stellar bar transports angular momentum outwards, resulting in a decrease in [$\Omega_\mathrm{p}$]{}. This decrease is almost linear with time in absence of a dissipative component (gas) and any star formation. For instance, this is the case of the three N$-$body simulations described in Sect. \[sec:nbody\]. As a result, the corotation radius increases over time, as do other resonance radii. It is the whole resonance system that moves, whether it is the one generated by the bar or by any other structure likely to lose/gain angular momentum.
{width="0.24\hsize"} {width="0.24\hsize"} {width="0.24\hsize"} {width="0.24\hsize"}
In Figure \[fig:td\_lz\] we have displayed diffusion time scales $T_D(E)$ and $T_D(L_\mathrm{z})$ averaged over sets of particles sampled by ${\overline{L_\mathrm{z}}}$ ranges, for the case of [`RunC`]{}. ${\overline{L_\mathrm{z}}}$ is now time-averaged over $\approx
7$ Gyr. It is here a proxy for the mean radial position of particles. In this $\langle T_D \rangle - {\overline{L_\mathrm{z}}}$ plot, we can overlay the approximate position of bar resonances determined in the linear epicyclic approximation. Since the entire resonance system moves outwards during the evolution of the galaxy, we have plotted the position of the UHR at $t=3.16$ Gyr and the OLR at $t=10.54$ Gyr. This delimits the range in [$L_\mathrm{z}$]{} occupied by all bar resonances during the evolution. The specific range covered by the corotation is approximately represented by the shaded area but can be more extended when [$L_{1,2}$]{} and [$L_{4,5}$]{} positions are considered.
This averaged view of time scales and angular momentum confirms the statements made in Section \[sec:timescales\]. In average, $\langle
T_D(E)\rangle$ decreases from the centre to the outermost regions. In the range of ${\overline{L_\mathrm{z}}}$ delimited by resonances positions, $\langle T_D(E)\rangle$ remains below 1 Gyr, and goes down to 0.1 Gyr. Typical $\langle T_D(E)\rangle$ values outside OLR remain of the order of the Gyr or below.
In contrast to $\langle T_D(E)\rangle$, $\langle T_D(Lz)\rangle$ increases from the centre outwards. In the bar resonance region, it reaches values much higher than 10 Gyr. The diffusion of [$L_\mathrm{z}$]{} is therefore more effective well inside the UHR. In order to avoid any misunderstanding, we remind here that we only study the phase after the formation of the bar ($t> 3$ Gyr), in a regime that can be considered as quiet.
There have been some debates on the action of the OLR as a barrier to stellar migration . In the case of our simulations, $\langle
T_D(E)\rangle$ does indeed show a bump just after the OLR but the characteristic time scale remains of the Gyr order. $\langle
T_D(Lz)\rangle$ continues to increase well beyond the bar OLR, even for [`RunC^\mathtt{axi}`]{}. Therefore, we can not confirm a specific signature of a barrier due to the bar OLR.
![Projection of mass distribution on $x-y$ plane for particles selected according to $T_D(E)$ (top) and $T_D(L_\mathrm{z})$ (bottom) for [`RunC`]{} at $t=10.54$ Gyr. The $\log$ greyscale is identical for all four figures. Contour labels are spaced by 0.25 dex. The white or black circles show the position for the innermost UHR (dot-dashed), the corotation (full line), and the outermost OLR (long-dashed). Spatial scale is in kpc.[]{data-label="fig:td_xy"}](PFAA_fig_tdxy.pdf){width="\hsize"}
Formally, [$L_\mathrm{z}$]{} is a good proxy for the radius only for orbits close to circular. In order to verify the true spatial distribution, the mass distribution obtained for short and long diffusion times (arbitrarily defined as shorter or longer than the simulation length) can be projected. Figure \[fig:td\_xy\] shows the mass projected in the x$-$y plane for $T_D$ less than or greater than 10.54 Gyr. It globally confirms the analysis of Sect. \[sec:timescales\] but suggests that a more detailed analysis must be performed. A few features deserve to be mentioned. Particles with $T_D(E) > 10.54$ Gyr includes a population that might be trapped around $L_{4,5}$ Lagrangian points. They also included stable $x_1$ orbits inside the UHR, where their shape is purely elliptical . Beyond the UHR bifurcation, $x_1$ orbits become rectangular-like, develop loops, and can become unstable. Their contribution is visible in the mass distribution for $T_D(E) < 10.54$ Gyr. A much more detailed study of the $T_D$ spatial distribution, and its relation to orbits families, is postponed to a dedicated future paper.
Stochastic diffusion
--------------------
Results similar to have been reached: the role of the ‘hot’ population is highlighted in both studies and the diffusion time scale depends on the radial position. However, a quantitative comparison with is not straightforward. The definition of their diffusion coefficient is different and they expressed it as a function of time and radius. Their study is based on Fourier’s law of heat conduction. Heat conduction is a non-equilibrium phenomenon. A coarse-grained description of the phenomenon with a clear separation between microscopic and macroscopic scales can be assumed. At the microscopic scale, heat carriers which are molecules and atoms in gas and liquids, phonons in solids, evolve as a result of a deterministic Hamiltonian description, whereas at macroscopic scale phenomenological Fourier’s law implies a diffusive transfer of energy. However a rigorous derivation of this law starting from a microscopic Hamiltonian description is still lacking [@2008AdPhy..57..457D].
Microscopically we have to think about heat carriers colliding randomly and the result is a heat diffusion. However, in a pure stellar N$-$body system, hard collisions are rare. Encounters are the dominant process, especially weak ones, which makes the Fokker-Planck equation the traditional tool for the study of stellar systems through the frictional and diffusion coefficients [@1992rcd..book.....L; @2008gady.book.....B].
It can be shown that a DF whose evolution over time is governed by the Fokker-Planck equation also follows a diffusion equation of Fourier’s form, with the same diffusion coefficient, provided that a relationship with the friction coefficient is respected [cf @1992rcd..book.....L]. Therefore, the formalism used by may be similar to that of Fokker-Planck.
For simple Hamiltonians, a generalised Fokker-Planck equation can be derived to include the energy drift due to scattering and fast transport in phase space due to trapping/escape. This derivation goes beyond the purpose of this paper.
The Chirikov coefficient implicitly includes all effects due to resonances, and resonances overlaps due to several forcing patterns, as well as effects due to a noisy potential.
Limitations
-----------
For our first paper on this topic, we have decided to restrict the exploration of $E$ and [$L_\mathrm{z}$]{} variations to the simplest type of simulations, the pure N$-$body case. Indeed, the absence of a gaseous component is a major main limitation that has several clear consequences. Without gas, there is no possibility to form a new population whose kinematics might cool down the disc . Another missing fluid is dark matter. The main effect of a live dark halo (except to flatten the rotation curve of the disc at a large distance) is to permit the exchange of angular momentum with the stellar disc. The rate and the amplitude of these exchanges depend on the velocity dispersion of both the disc and the halo, and on the relative halo mass. Depending on the rate at which the stellar disc losses its angular momentum, the bar grows quite differently. Considering @2006ApJ...637..214M simulations as representative, roughly 2/3 of the angular momentum loses by the bar-unstable part of the stellar disc is absorbed by the halo, the rest going to the outer disc. Most of these exchanges happen during the buckling of the bar, which, in the case of our simulations, occurs for $t < 3$ Gyr.
A final limitation comes from the genuine nature of galaxies, which are much more complex than these idealised simulations. Much of this complexity comes from perturbations by random sources. These sources can be intrinsic (such as molecular clouds or Poissonian shot noise) or extrinsic (satellites accretion, globular clusters, etc.). All these perturbations could contribute to reduce diffusion times, but this remains to be demonstrated in the specific case at hand.
Conclusions
===========
We have computed @1979PhR....52..263C diffusion rates ([$D_2(E)$]{} and [$D_2(L_\mathrm{z})$]{}) and related diffusion time scales ($T_D(E)$ and $T_D(L_\mathrm{z})$) for energy ($E$) and angular momentum ([$L_\mathrm{z}$]{}) in pure N$-$body simulations of disc galaxies developing bars and spiral structures. These quantities were only calculated once the bar was perfectly settled in order to reflect the evolution of the disc under the effect of its presence.
We can summarise our results as follows:
1. Both $E$ and [$L_\mathrm{z}$]{} diffuse during the evolution of a stellar disc, under the effect of intrinsic perturbations caused by the bar and spiral structures. In particular, bars and spiral structures are responsible for shortening diffusion time scales.
2. Diffusion time scales are shorter than the simulations length (i.e. $\approx 10$ Gyr) for different particle populations depending on whether the diffusion of $E$ or [$L_\mathrm{z}$]{} is considered. Consequently, the regions affected by the diffusion differ according to the quantity that diffuses.
3. The distribution function of Chirikov diffusion rates $D_2$ has the same shape regardless the simulation considered. It can be approximate by the equation $ \log n(D_2)\propto \gamma D_2 /
\max(D_2) $ where $\gamma\approx -5.0$ for [$D_2(E)$]{} and $\gamma$ is in the range $[-6.61 ; -6.21]$ for [$D_2(L_\mathrm{z})$]{}.
4. At first order, values of [$D_2(E)$]{} remain within the same range in axisymmetric and non-axisymmetric simulations unlike for [$D_2(L_\mathrm{z})$]{}.
5. $T_D(E)$ is shorter than simulation length for particles belonging to the ‘hot’ population, the disc, and families of orbits lying between the bar UHR and corotation. It is minimal (and shorter than 1 Gyr) in the region delimited by the set of bar resonances (between UHR and OLR).
6. $T_D(L_\mathrm{z})$ is shorter than simulation length mainly for particles inside the bar region (i.e. inside the UHR).
7. On average, $T_D(L_\mathrm{z})$ increases with radius while $T_D(E)$ tends to decrease from the centre to the most external regions.
8. The so-called ’hot’ population, which navigates between the bar and the disc, plays only an important role in diffusion of $E$.
This article is limited to a first exploration of the results obtained with the Chirikov diffusion rate. Next articles will explore in greater depth the phenomena of migration, diffusion and resonance, particularly in terms of orbital structure.
{width="0.45\hsize"} {width="0.45\hsize"}
{width="0.45\hsize"} {width="0.45\hsize"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the distributions of $\epsilon$-drawdowns and $\epsilon$-drawups of the most liquid futures financial contracts of the world at time scales of $30$ seconds. The $\epsilon$-drawdowns (resp. $\epsilon$-drawups) generalise the notion of runs of negative (resp. positive) returns so as to capture the risks to which investors are arguably the most concerned with. Similarly to the distribution of returns, we find that the distributions of $\epsilon$-drawdowns and $\epsilon$-drawups exhibit power law tails, albeit with exponents significantly larger than those for the return distributions. This paradoxical result can be attributed to (i) the existence of significant transient dependence between returns and (ii) the presence of large outliers (dragon-kings) characterizing the extreme tail of the drawdown/drawup distributions deviating from the power law. The study of the tail dependence between the sizes, speeds and durations of drawdown/drawup indicates a clear relationship between size and speed but none between size and duration. This implies that the most extreme drawdown/drawup tend to occur fast and are dominated by a few very large returns. We discuss both the endogenous and exogenous origins of these extreme events.'
address:
- 'Dept. of Management, Technology and Economics, ETH Zürich, Zürich, Switzerland'
- 'Swiss Finance Institute, c/o University of Geneva'
author:
- Vladimir Filimonov
- Didier Sornette
title: |
Power law scaling and “Dragon-Kings”\
in distributions of intraday financial drawdowns
---
Extreme events ,drawdowns ,power law distribution ,tail dependence ,“Dragon-King” events ,financial markets ,high-frequency data
Introduction
============
Traditionally, market risk is proxied by the distribution of asset log-returns $r_t^{(\Delta t)}$ on different scales $\Delta t$. Such distribution for most assets in various classes is well-known to have a power law tail $\mathrm{Pr}[r_t^{(\Delta t)} >x] \sim x^{-\mu}$ with the “tail exponent” $\mu$ in the range from 3 to 5 both for daily [@deVries1994; @Pagan1996; @Gopikrishnan1998; @Cont2001; @Sornette2005] and intraday returns [@Chakraborti2011_1]. Though the second and possibly third order moments of the distribution exist, the traditional volatility measures based on variance of returns are not sufficient for quantifying the risk associated with extreme events.
Much better metrics to capture systematic events are the so-called *drawdowns* (and their complements, the *drawups*), which are traditionally defined as a persistent decrease (respectively increase) in the price over consecutive $\Delta t$-time intervals [@JohansenSornette2001JofRisk]. In other words, a drawdown is the cumulative loss from the last maximum to the next minimum of the price, and a drawup is the the price change between a local minimum and the following maximum. By definition, drawups and drawdowns alternate: a drawdown follows a drawup and vice versa.
In contrast to simple returns, drawdowns are much more flexible measures of risk as they also capture the transient time-dependence of consecutive price changes. Drawdowns quantify the worst-case scenario of an investor buying at the local high and selling at the next minimum (similarly drawups quantifies the upside potential of buying at the lowest price and selling at the highest one). The duration of drawdowns is not fixed as well: some drawdowns can end in one drop of duration $\Delta t$, when others may last for tens to hundreds $\Delta t$’s. The distribution of drawdowns contains information that is quite different from the distribution of returns over a fixed time scale. In particular, a drawdown reflects a transient “memory” of the market by accounting for possible inter-dependence during series of losses [@SornetteJohansen2000; @JohansenSornette2001JofRisk]. During crashes, positive feedback mechanisms are activated so that previous losses lead to further selling, strengthening the downward trend, as for instance as a result of the implementation of so-called portfolio insurance strategies [@Lelandrub]. The resulting drawdowns will capture these transient positive feedbacks, much more than the returns or even the two-point correlation functions between returns or between volatilities. In contrast to autocorrelation measures, which quantify the average (or global) serial linear dependence between returns over a generally large chosen time period, drawdowns are local measures, i.e. they account for rather instantaneous dependences between returns that are specific to a given event. Statistically, drawdowns are related to the notion of “runs” that is often used in econometrics [@Campbell1996].
This paper presents an analysis of the statistical properties of intraday drawdowns and drawups. Our tests are performed on the most liquid Futures Contracts of the world. Our results are thus of general relevance and are offered as novel “stylized facts” of the price dynamics. We discuss and quantify the distribution of intra-day extreme events, and compare distributional characteristics of drawdowns with those of individual returns. In so doing, we discover that the generally accepted description of the tail of the distribution of returns by a power law distribution is incorrect: we find highly statistically significant upward deviations from the power law by the most extreme events. These deviations are associated with well-known events, such as the “flash-crash” of May 6, 2010. Statistical tests designed to detect such deviations confirm their high significance, implying that these events belong to a special class of so-called *“Dragon-Kings”* [@Sornette2009; @SorouillonDK12]: these events are generated with different amplifying mechanisms than the rest of the population. We show that some of these events can be attributed to an internal mutual-excitation between market participants, while others are pure response to external news. As for extreme drawdowns, there are in principle two end-member generating mechanisms for them: (i) one return in the run of losses is an extreme loss and, by itself alone, makes the drawdown extreme; (ii) rare transient dependences between negative returns make some runs especially large. We document that most of the extreme drawdowns are generated by the second mechanism, that is, by emerging spontaneous correlation patterns, rather than by the domination of one or a few extreme individual returns.
The paper is organized as follows. Section \[sec:data\] discusses the high-frequency data and cleaning procedures. Section \[sec:dd\] presents the detection method of the so-called *$\epsilon$-drawdowns* that we use as a proxy of transient directional price movements. Section \[sec:descript\_stat\] provides descriptive statistics of the detected events. Section \[sec:distrib\_dd\] focuses on the properties of the distributions of drawdowns and quantify their tails as belonging essentially to a power law regime. In section \[sec:dragonkings\], we present a generalised Dragon-King test (DK-test), derived and improved from [@PisarenkoSornette2012], which allows us to quantify the statistical significance of the deviations of extreme drawdowns from the power law distribution calibrated on the rest of the distribution. Section \[sec:distrib\_aggragated\] describes the aggregated distributions over all tickers and validates that our findings hold both at individual and global levels. For this, we use the generalised DK-test as well as the parametric U-test also introduced by [@PisarenkoSornette2012] and study their respective complementary merits. Section \[sec:dependence\] examines the interdependence of the speed and durations of extreme drawdowns with respect to their size. Section \[sec:conclusion\] concludes.
The data {#sec:data}
========
We use tick data for the most actively traded Futures Contracts on the World Indices (see Table \[tb:contracts\]) from January 1, 2005 to December 30, 2011. For Futures on the OMX Stockholm 30 Index (OMXS), our dataset starts from September 1, 2005; and for Futures on Hong Kong indexes (HSI and HCEI), our datasets are limited to the period before April 1, 2011. For Futures on the BOVESPA index (BOVESPA), we restrict our analysis to the period after January 1, 2009, ignoring the relatively inactive trading in 2005–2008.
Many of the contracts presented in Table \[tb:contracts\] are traded almost continuously (e.g. E-mini S&P 500 futures contracts are traded every business day from Monday to Friday with only two trading halts: from 16:15 to 17:00 CDT and from 15:15 to 15:30 CDT). Though it is being progressively changing [@FilimonovSornette2013_apparent], most of the daily volume is traded within so-called Regular Trading Hours (RTH, in case of E-mini contracts: 8:30–15:15 CDT). For Asian exchanges, the activity within Regular Trading Hours is also non-uniform. For the analysis, we have limited ourselves only to the part of RTH where the trading is the most active (in terms of volume), which we refer to as an Active Trading Hours (ATH) in Table \[tb:contracts\].
Region Codename Underlying index ATH / Exchange
--------------- ---------- --------------------------- ------------------------------
AEX AEX (Netherlands) 09:00 – 17:30 CET / Euronext
CAC CAC40 (France) 09:00 – 17:30 CET / MONEP
DAX DAX (Germany) 09:00 – 17:30 CET / Eurex
FTSE FTSE (UK) 08:00 – 17:30 GMT / LIFFE
MIB FTSE MIB (Italy) 09:00 – 17:30 CET / MIL
IBEX IBEX (Spain) 09:00 – 17:30 CET / MEFF
STOXX Euro STOXX (Europe) 09:00 – 17:30 CET / Eurex
OMXS OMX Stockholm 30 (Sweden) 09:00 – 17:25 CET / OMX
SMI SMI (Switzerland) 09:00 – 17:25 CET / Eurex
ES S&P 500, E-mini (US) 08:30 – 15:15 CDT / CME
DJ Dow Jones, E-mini (US) 09:00 – 15:15 CDT / CBOT
NQ NASDAQ, E-mini (US) 08:30 – 15:15 CDT / CME
HSI Hang Seng (Hong Kong) 09:45 – 12:30 HKT / HKFE
HCEI HCEI (Hong Kong) 09:45 – 12:30 HKT / HKFE
TAMSCI TAMSCI (Taiwan) 08:45 – 13:45 SGT / SEDT
NIFTY NIFTY (India) 10:00 – 15:30 IST / NSEI
NIKKEI Nikkei 225 (Japan) 12:30 – 15:10 JST / OSA
TOPIX TOPIX (Japan) 12:30 – 15:10 JST / OSA
Australia ASX S&P/ASX 200 (Australia) 09:50 – 16:30 AEDT / SFE
South America BOVESPA BOVESPA (Brazil) 09:05 – 17:15 BRT / SPCFE
: Description of the Futures Contracts used for analysis.[]{data-label="tb:contracts"}
All Futures Contracts presented of Table \[tb:contracts\] are traded in different cycles with different expiration date. Moreover, for each Futures at every moment, 5–6 contracts with different maturities are traded simultaneously. Typically, for all Index Futures contracts, most of the trading activity is going at the so-called “Front Month” contract with the nearest maturity: in order to avoid settlement, financial investors “roll over” contracts to the next maturity typically one week before the expiration. At the rollover, the liquidity (measured in volume) of the contract that is going to expire is switched to the contract that will expire at the following quarter.
![For each analyzed contract (Table \[tb:contracts\]), daily closing price (upper panel, black lines, right scale) and intraday volatility (upper panel, blue lines, left scale) are presented. Green up triangles and red down triangles denote respectively the 5 largest drawups and the 5 largest drawdowns for $\Delta t=30$ sec and $\epsilon_0=1$ (see Section \[sec:dd\]). The lower panel of each asset presents the number of “largest” drawdowns per month (drawdowns with normalized size larger than $\hat x_m$, see Section \[sec:distrib\_dd\] and Table \[tb:fit\_dd\_norm\_returns\]).[]{data-label="fig:price_vol"}](fig1_price_vol.pdf){width="\textwidth"}
In order to construct a continuous time series from Futures Contracts with different maturities, we have “rolled” the previous contract to the next one on the second Thursdays of the expiration months. This approach is different from the traditional one “rolling” the Front Month contract at expiration date. For any given date, our approach amounts to selecting the contract with the largest daily volume. Note, however, that this “rolling” procedure, being well-suited for Futures contracts on financial indices, does not apply directly to other futures contracts such as commodity futures, where the physical delivery plays an important role (see for instance [@Masteika2012] for a discussion of various methods of “rolling” procedures). Figure \[fig:price\_vol\] presents the dynamics of daily closing prices of the analyzed Front Month contracts together with their intraday volatility estimated using [@GarmanKlass1980]’s method.
Before our analysis, we have cleaned the high-frequency data according to standard rules described in [@Falkenberry2002; @Brownlees2006; @BarndorffNielsen2009]. Namely, we have (i) ignored all trades and quotes outside active trading hours (see Table \[tb:descr\]); (ii) ignored all recordings with transaction price, bid or ask equal to zero; (iii) deleted all quotes entries for which bid-ask spread is negative; (iv) deleted entries with the spread larger than 20 times the median spread of the day; (v) delete entries with corrected trades; (vi) deleted entries with prices above ‘ask’ plus the bid-ask spread and entries with prices below ‘bid’ minus bid-ask spread. Moreover, we ignored all trading days with gaps in data (due to issues on recording or low trading activity) longer than 5 minutes.
Detection of drawdowns {#sec:dd}
======================
Traditional measure of drawdowns (drawups) as sums of consecutive negative (positive) returns is very rigid and sensitive to noise: even tiny deviation of the price in the opposite direction will break large drawdowns into parts. Consider as an example a total peak-to-valley drop of $40\%$ of total duration of 10 days, with daily returns $$\{+0.8\%, -5\%. -3\%, -10\%, -2\%, +0.01\%, -8\%, -13\%, -3\%, -4\%, -2.01\%, +1.2\%\}~.
\label{rtjhetyngb}$$ The second return $-5\%$ following the first positive return $+0.8\%$ determines the start of a drawdown. The last return $+1.2\%$ determines the end of the last drawdown. In the standard definition, this series is characterised by two drawdowns of $-20\%$, characterised by the followed return series, $$\begin{aligned}
&&\{+0.8\%, -5\%. -3\%, -10\%, -2\%, +0.01\%\}, \nonumber \\
&&\{+0.01\%, -8\%, -13\%, -3\%, -4\%, -2.01\%, +1.2\%\}~,
\label{rtjhewrthttyngb}\end{aligned}$$ where we include the two positive returns at the boundaries to make clear the start and end of the drawdowns. Does this makes sense? Following [@JohSorepsidd10], we argue that investors gauging the market dynamics will be financially and psychology hurt by the total loss of $-40\%$, and will be insensitive to the tiny positive return $+0.01\%$ that technically defines two separate drawdowns. As a risk measure, it is intuitive that the event to register is the total loss of $-40\%$. This motivates the introduction of robust measures of drawdowns and we use the so-called *epsilon-drawdown* (*$\epsilon$-drawdown*) measure introduced by [@JohSorepsidd10]. An $\epsilon$-drawdown is defined as a standard drawdown, except for the fact that positive returns smaller than some defined threshold controlled by the parameter $\epsilon >0$ are considered as “noise” and do not end a drawdown run. Only when a positive return occurs, which is larger than the threshold, is the drawdown deemed to end. In the above example, taking a threshold of, say, $0.5\%$ leads to characterise the series (\[rtjhetyngb\]) as a single $\epsilon$-drawdown of amplitude $-40\%$, which is a faithful embodiment of the realised losses of investors. The two pure drawdowns of $20\%$, when analysed statistically for instance via the distribution of their sizes that loses all information about their closeness, paint a much milder picture of the true loss. In practice, $\epsilon$ can be either a pre-defined constant or time-dependent. The second option is preferable to account for the clustering and memory effects of the volatility [@Cont2001]. For large-scale analyses and in order to compare different market regimes, the dynamics of volatility should be taken into account for the choice of $\epsilon$. The most transparent way is to choose $\epsilon=\epsilon_0\sigma$, where $\sigma$ is a measure of the realized volatility estimated over a preceding time period as discussed later and $\epsilon_0$ is a constant. This time-adaptive choice for $\epsilon$ allows for a scaling of the tolerance in the definition of $\epsilon$-drawdown that takes into account the recent level of “noise”. For $\epsilon=0$, one recovers the classical definition of a drawdown (respectively, drawup) as a sequence of consecutive strictly negative (respectively, positive) returns.
Technically, we define the sequence of drawups and drawdowns as follows. Consider the total time interval $[t_1, t_2]$. We first discretize it in $N=[(t_2-t_1)/\Delta t]$ periods of length $\Delta t$, where the square brackets denote the floor function. This allows us to construct the discrete returns at time scale $\Delta t$ from the log-price series $p(t)=\log P(t)$ as $$\label{eq:ret}
r_k=\log P(t_1+k\Delta t)-\log P(t_1+(k-1)\Delta t),\quad k=1,2, \dots, N~.$$ The time $k=k_0=1$ is defined as a beginning of a drawup if $r_1>0$ and a drawdown if $r_1<0$. Then, for each $k>k_0$, we calculate the cumulative sum $$p_{k_0, k}=\sum_{i=k_0}^k r_i$$ and test the largest deviation $\delta_{k_0,k}$ of the price trajectory from a previous extremum: $$\label{eq:delta}
\delta_{k_0,k}=\begin{cases}
\displaystyle
p_{k_0, k}-\min_{k_0\leq i\leq k} p_{k_0,i} \qquad\text{for drawdowns},\\
\displaystyle
\max_{k_0\leq i\leq k} p_{k_0,i}-p_{k_0, k} \qquad\text{for drawups}.\\
\end{cases}$$ We stop the procedure when $\delta_{k_0,k}$ becomes larger than $\epsilon$: $$\label{eq:cond}
\delta_{k_0,k}>\epsilon=\epsilon_0\sigma,$$ where $\sigma$ is a measure of the volatility in the recent past and $\epsilon_0$ is a constant.
For a drawdown, the lowest price $k_1=\arg\min_{k_0\leq i\leq k} p_{k_0,i}$ is defined as its end and $k+1$ is defined as starting the following drawup. Respectively, for a drawup, the highest price $k_1=\arg\max_{k_0\leq i\leq k} p_{k_0,i}$ is defined as its end and $k+1$ as the beginning of the following drawdown. The procedure restarts from the time $k_1$, from which we compute the cumulative sum $p_{k_1, k}$ and the maximum deviation $\delta_{k_1,k}$, looking for the next value of $k$ that will satisfy $\delta_{k_1,k}>\epsilon$. Then, the next extreme value of $\delta_{k_1,k}$ within $[k_1, k]$ provides us $k_2$, and so on.
![Illustration to the definition of $\epsilon$-drawups and $\epsilon$-drawdowns.[]{data-label="fig:drawdown_illustration"}](fig0_dd.pdf){width="65.00000%"}
Figure \[fig:drawdown\_illustration\] illustrates the procedure described above and the fact that drawdowns are always followed by drawups and vice versa, by definition. Drawdowns always start and end with negative returns, and correspondingly drawups always start and end with positive returns. It is important to notice that, similarly to other measures of trends, neither strict drawdowns nor $\epsilon$-drawdowns are causal, in the sense that at time $k$ it is impossible to say if the current drawdown (or drawup) is over or not. The classical definition ($\epsilon=0$) requires one-step look-ahead in order to conclude about the existence of the change of the trend and the end of the drawdown. The required look-ahead for $\epsilon$-drawdowns increases with $\epsilon$, and also exhibits some time variability through the realized volatility at the time of observation.
For the present analysis, we aggregate tick-by-tick data within each day into 30-seconds bars ($\Delta t=30$ sec). The chosen size of $\Delta t$ results from a tradeoff. On the one hand, $\Delta t$ should be long enough to reduce the microstructure noise (e.g. resulting from the bid-ask bounce) and to capture only systemic events while being insensitive to price drops and jumps due to “fat-finger trades” or the transient sparsity of the order book. On the another hand, $\Delta t$ should not be too large so as to miss fast events that may happen on small time scales. For example, during one of the most dramatic intraday price fluctuations — the so-called “flash crash” of May 6, 2010, that started on E-mini S&P 500 futures contracts [@FlashCrash2010_report], and then almost instantly propagated to constituting stocks — the price of E-mini Futures plunged by 5.7% in 4 minutes and then recovered back by 4% over the next 3 minutes. For instance, taking $\Delta t=15$ minutes would prevent us from detecting any anomaly and would make use blind to this fast drawdown followed by an equally fast drawup. As this flash crash is associated with the largest intraday swing of the Dow Jones index (998.5 points or approximately 9% in 5 minutes), we must select parameters that ensure a suitable capture of the most salient developments of financial markets.
In addition to $\Delta t$, the definition of $\epsilon$-drawdowns requires the specification of $\epsilon$ defined by expression in terms of the recent volatility $\sigma$. Traditionally, the realized volatility at time $t$ in a running window of size $\tau$ is estimated as the standard deviation of the returns in the interval $[t-\tau, t)$. For an intraday estimation in a time interval $[t_1, t_2]$, this estimation cannot be directly applied due to the impact of the trading day opening, which is characterised by a very large transient volatility. This leads to either truncate the window at its beginning in order to avoid or minimise the influence of the opening period, which results in a smaller number of samples and thus higher variance of the initial estimates, or to include the overnight trading period prior to $t_1$, which amounts to mixing different regimes in one estimate. In order to be consistent within the trading day, for the $\sigma$ in expression , we use the volatility calculated on returns of the previous trading day. Specifically, $\sigma$ is defined as the standard deviation of the log-returns $r_k$ defined by equation : $$\label{eq:vol}
\sigma^2=\frac 1N \sum_{k=0}^N r_k^2=
\frac 1N \sum_{k=0}^N
\Big(\log P(t_1+k\Delta t)-\log P(t_1+(k-1)\Delta t)\Big)^2,$$ where $N= [(t_2-t_1)/\Delta t]$. Note that expression is slightly different from the traditional definition of volatility $\sigma_{\Delta t}$, which is normalized by the time scale $\Delta t$: $\sigma_{\Delta t}=\sigma/\sqrt{\Delta t}$.
The intraday volatility is subjected to the so-called “signature plot” effect [@AndersenBollerslev2000; @Muzy2013hawkes; @SaiSorepps14]: the volatility $\sigma_{\Delta t}$ is a decreasing function of $\Delta t$. Moreover, in the presence of the bid-ask bounce, decreasing $\Delta t$ tends to increase the negative autocorrelation of returns at the first lag. The choice of the parameter $\epsilon_0$ should thus be adapted to the time-scale $\Delta t$. For example, for $\Delta t$ of the order of few seconds, when most of the volatility $\sigma$ can be attributed to the bid-ask bounce, $\epsilon_0$ should be taken larger than for $\Delta t$ of the orders of minutes, in order to achieve the same level of aggregation. On the other hand, $\epsilon_0$ regulates the minimum (and also typical) size of the detected drawdowns, and plays a role of an effective “scaling parameter”.
For the analysis presented below, we have selected $\Delta t=30$ seconds and $\epsilon_0=1$. We have tested that our results are robust with respect to other values of these parameters to within at least a factor $2$.
Descriptive statistics of intraday drawdowns {#sec:descript_stat}
============================================
We characterize drawdowns (drawups) by the following properties:
- *duration*: $\tau=(k_{end}-k_{start})\Delta t$, where $k_{start}$ and $k_{end}$ are time index of the beginning and end of a drawdown (drawup). Since we work with discrete times, durations are always multiple of $\Delta t$;
- *size*: $\Delta P=|P(t_1+k_{end}\Delta t)-P(t_1+k_{start}\Delta t)|$ (this is a multiple of a tick size);
- *return*: $r=|\log P(t_1+k_{end}\Delta t)-\log P(t_1+k_{start}\Delta t)|$;
- *normalized return*: $r^{norm}=r/\sigma$, where $\sigma$ is the volatility of the previous trading day;
- *speed*: $v=r/\tau$;
- *normalized speed*: $v^{norm}=r^{norm}/\tau=v/\sigma$.
Over the 7 years of the period 2005–2011, all the studied markets have passed through different regimes (see Figure \[fig:price\_vol\]): low-volatile period of 2005–2007, increase of uncertainty in 2008 and volatility burst during the peak of the sub-prime crisis in October 2008. After the relaxation to the pre-crisis level by the second half of 2009–early 2010, the volatility spiked again in the mid-2010 and second half of 2010 due to the Eurozone debt crisis. For example, for E-mini S&P 500 Futures Contracts, the highest intraday volatility over the whole period ($0.083$ on October 10, 2008) is almost 50 times larger than the lowest ($0.0015$ on February 15, 2007). Thus, normalizing drawdowns by the instantaneous volatility is essential to obtain meaningful results aggregated over such a long period and especially across multiple assets.
\(a) Drawdowns (b) Drawups
Table \[tb:descr\] presents descriptive statistics of durations, sizes, normalized returns and speeds of drawups and drawdowns detected for $\Delta t=30$ sec and $\epsilon_0=1$ on the entire interval of 2005–2011 for different contracts (for BOVESPA, we considered the period 2009–2011). Changes in $\Delta t$ and $\epsilon_0$ modify the descriptive statistics as follows. Increasing $\Delta t$ or $\epsilon_0$ decreases proportionally the number of drawdowns and drawups, and magnifies their durations $\tau$ together with their sizes $\Delta P$ and their normalized returns $r^{norm}$. For instance, for $\Delta t=2$ min and $\epsilon_0=2$, the typical duration $\tau$ is roughly 14–15 times larger than for $\Delta t=30$ sec and $\epsilon_0=0.5$. While changes of $\Delta P$ are proportional to changes in $\Delta t$ and $\epsilon_0$, the normalized returns $r^{norm}$ only gradually increase with $\Delta t$ and, at the same time, are much more sensitive to $\epsilon_0$. Interestingly, the normalized speed $v^{norm}$ is almost insensitive to changes in $\epsilon_0$, but decreases with increase of $\Delta t$ (i.e. larger drawdowns are typically slower than shorter drawdowns). This decrease is almost proportional for most of the distribution of normalized speeds, except in its tail where a different scaling holds. Such non-trivial scaling can be interpreted as due to the multifractal properties of the price dynamics [@ArneodoMuzySornette1998; @Calvet2002review; @Muzy2005] resulting from the interplay between long-term memory in the system together with nonlinear amplification [@FilimonovSornette2011_SEMF].
Table \[tb:descr\] shows that the statistics for drawdowns and drawups are almost identical for each asset and for all analyzed parameters, except the statistics of extreme (maximal) values. Typically (for developed European and US markets), extreme drawdowns are larger than extreme drawups in terms of normalized returns $r^{norm}$, which concurs with the empirical evidence of gain-loss asymmetry [@Cont2001; @Jensen2003]. However, for TAMSCI (Taiwan), the extreme drawup of 5% at the opening of the market on September 10, 2009 ($r^{norm}=127.4$) is more than twice larger than the maximum observed drawdown ($r^{norm}=57.7$). For NIKKEI, TOPIX (Japan) and BOVESPA (Brazil), the largest drawup is only slightly larger than the most extreme drawdown.
We observe that both median and 90%-quantiles (as well as mean not reported in the table) of the durations $\tau$ of drawdowns and drawups are almost equal to each other for all Futures contracts, except NIKKEI and TOPIX. Though the price of contracts vary in a wide range and so does the sizes $\Delta P$ of drawdowns, the statistics (mean, median and quantiles) of normalized returns $r^{norm}$ are very similar across contracts. At the same time, maximal values of the normalized returns differ by a large factor: the smallest extreme value ($r^{norm}=44.3$ for HCEI) is 4.8 times smaller than the largest one ($r^{norm}=215.0$ for CAC). Finally, mean and quantile statistics for normalized speeds $v^{norm}$ agree almost perfectly across all contracts.
We also note the exceptional properties of Japanese markets (contracts NIKKEI and TOPIX), which exhibit the shortest (minimal values of $\tau$) and the smallest (minimal values of $r^{norm}$) drawdowns and drawups among all analyzed contracts. Moreover, while the normalized speed $v^{norm}$ of drawdowns and drawups are typically close to each other across different contracts, the normalized speed of the fastest drawdown on TOPIX Futures Contracts is at least twice smaller than that of any other analyzed contract.
Distribution of normalized size of drawdowns {#sec:distrib_dd}
============================================
Figure \[fig:ccdf\] shows the complementary cumulative distribution functions (ccdf) of the normalized returns $r^{norm}$ of the drawdowns and drawups defined at the top of the previous section \[sec:descript\_stat\], for $\Delta t=30$ sec and $\epsilon_0=1$, together with the distributions of negative and positive price log-returns for the value $\Delta t=30$. Together with the empirical distributions, we present the distributions of returns of drawdowns for the null model constructed by reshuffling log-returns within the active trading hours of each trading day independently. This reshuffling destroys all temporal correlations while keeping unchanged the marginal distributions of log-returns as well as the secular dynamics of the intraday volatility (see Figure \[fig:price\_vol\]).
![Complementary cumulative distribution function (ccdf) for the normalized returns of drawdowns (red down triangles) and drawups (green up triangles) for $\Delta t=30$ sec and $\epsilon_0=1$. Blue and magenta lines depict the distributions of drawdowns and drawups of the null model constructed by reshuffled returns. Black straight lines correspond to power law fits of the tails of the distributions of drawdowns (see Table \[tb:fit\_dd\_norm\_returns\]). Red and green dots lines show the distributions of normalized returns at $\Delta t=30$ sec. Dashed black lines correspond to power law fits of the tails of these distributions of returns (see Table \[tb:fit\_dd\_norm\_returns\]). []{data-label="fig:ccdf"}](fig2_ccdf_normalized_returns.pdf){width="\textwidth"}
One can observe in Figure \[fig:ccdf\] that the tails of the distributions of drawdowns and drawups deviate from those obtained for the null model. While the drawdowns in the null model follow approximately a Weibull distribution with shape parameter estimated around $0.9$, the real data is much better characterised by a fatter power law tail. This is evidence that time dependence between returns play an important role in the directional price movements and especially in the occurrence of large drawdowns and drawups. Interestingly, for some of the analyzed contracts (CAC, FTSE, ES and DJ), one can observe significant deviations in the null model distribution that are associated with a few extreme events. They correspond to drawdowns whose size is essentially controlled by a single return of enormous magnitude: as seen from Figure \[fig:ccdf\], the sizes of these drawdowns are identical to the sizes of the return outliers.
In order to quantify the power law approximation of the tails, $$\label{eq:pl_ccdf}
F(x)=\mathrm{Pr}\left[r^{norm} >x\right] = \left(\frac{x_{m}}{x}\right)^{\alpha},\quad x\ge x_m$$ we have employed the framework proposed in [@Clauset2009]. Assuming a known value of $x_m$, the (Hill) maximum likelihood estimator (MLE) yields the well-known closed form expression for the exponent $\hat \alpha$: $$\label{eq:mle}
\hat \alpha=N\cdot\left[\sum_{i=1}^N\ln\frac{x_i}{x_m}\right]^{-1},$$ where $N$ is the number of observations in the tail (i.e. such that $x_i\ge x_m$). The standard error on $\hat\alpha$ is derived from the width of the likelihood maximum [@SorknoRO96; @Newman2005; @Clauset2009]: $$\label{eq:mle_error}
\sigma_{\hat\alpha} \simeq \frac{\hat\alpha}{\sqrt{N}}$$ Following @Clauset2009, we have scanned different values of $x_m$, fitting the distribution and calculating the Kolmogorov-Smirnov distance [@deGrootSchervish2011_Prob_n_Stats] between the cumulative distribution function (cdf) of the data and fitted model: $$\label{eq:KS_dist}
D = \max_{x\ge x_m}|S(x)-F(x)|,$$ where $S(x)$ is the empirical cdf for the observed data and $F(x)$ is given by . As an alternative test, we have used the Anderson-Darling distance [@deGrootSchervish2011_Prob_n_Stats], which is more sensitive to the deviations in the tails of the distributions $$\label{eq:A2_dist}
A^2 = N\int_{x_m}^\infty \frac{(S(x)-F(x))^2}{F(x)(1-F(x))}dF(x)
= -N -\sum_{i=1}^N\frac{2i-1}{N}\Big[
\ln F(x_i)+\ln (1-F(x_{N+1-i}))
\Big],$$ where the input data is assumed to be ordered ($x_m\leq x_1<x_2<\dots<x_N$). We then used the value of $\hat x_m$ that minimizes $D$ (or $A^2$) as an estimate of the lower bound $x_m$. An illustration of the power law fits of the tails of the distributions of drawups and drawdowns is presented in Figure \[fig:ccdf\].
Table \[tb:fit\_dd\_norm\_returns\] summarises the results obtained by fitting the power law distribution to the empirical ccdf of the normalized returns ($r^{norm}$) of drawdowns and drawups using the Kolmogorov-Smirnov distance for the lower bound detection. The Anderson-Darling distance provides values similar to those reported in Table \[tb:fit\_dd\_norm\_returns\] except for the contracts ES, NQ, HCEI, NIKKEI, TOPIX and BOVESPA for which the more conservative $A^2$ distance cuts off too much of the data by estimating a lower bound $\hat x_m\gtrsim 20$. The resulting dataset contains less than 100 points, which prevents a robust estimation. This behavior resulting from the use of the $A^2$ distance was previously discussed in @Clauset2009. Let us mention that alternative methods have recently been introduced that improve on or complement the maximum likelihood (Hill) estimation of exponent $\alpha$ together with the selection of the lower bound $x_m$ using the Kolmogorov-Smirnov or Anderson-Darling distances. These methods include [@Carpenter2013; @Deluca2013; @Wager2014]), each of which present advantages over the Hill estimator but have also their own limitations.
\(a) Drawdowns and negative returns
--------- ------- ----------------- --------------------- ------ ----------------- ---------------------
$N_{x\ge \hat x_m}$ $N_{x\ge \hat x_m}$
AEX 14.83 4.25 $\pm$ 0.13 1068 3.72 4.18 $\pm$ 0.05 6342
CAC 14.38 4.34 $\pm$ 0.13 1076 3.06 3.90 $\pm$ 0.03 14311
DAX 13.32 4.00 $\pm$ 0.10 1481 3.45 3.53 $\pm$ 0.03 10226
FTSE 15.38 4.56 $\pm$ 0.17 752 3.23 4.33 $\pm$ 0.04 11505
MIB 18.43 4.71 $\pm$ 0.26 337 3.49 4.16 $\pm$ 0.05 6844
IBEX 12.70 4.51 $\pm$ 0.12 1493 3.68 4.20 $\pm$ 0.05 6138
STOXX 11.60 4.34 $\pm$ 0.12 1326 3.09 4.19 $\pm$ 0.04 9590
OMXS 14.71 5.00 $\pm$ 0.20 655 3.78 4.47 $\pm$ 0.09 2724
SMI 16.25 4.71 $\pm$ 0.19 605 4.12 4.48 $\pm$ 0.08 3386
ES 11.73 4.85 $\pm$ 0.15 1041 2.56 4.24 $\pm$ 0.03 17897
DJ 12.40 4.19 $\pm$ 0.13 1042 3.86 4.23 $\pm$ 0.07 4127
NQ 16.18 5.01 $\pm$ 0.27 358 3.76 4.36 $\pm$ 0.06 4633
HSI 11.07 4.03 $\pm$ 0.16 662 2.89 4.00 $\pm$ 0.06 4561
HCEI 18.18 5.54 $\pm$ 0.56 99 3.40 4.80 $\pm$ 0.10 2332
TAMSCI 14.27 4.50 $\pm$ 0.19 591 2.37 3.72 $\pm$ 0.03 15115
NIFTY 15.85 4.88 $\pm$ 0.23 456 3.84 4.21 $\pm$ 0.07 3850
NIKKEI 5.56 4.18 $\pm$ 0.10 1841 5.14 5.00 $\pm$ 0.59 71
TOPIX 11.67 4.81 $\pm$ 0.27 321 3.69 5.93 $\pm$ 0.18 1125
ASX 12.42 5.67 $\pm$ 0.18 972 5.17 5.62 $\pm$ 0.25 490
BOVESPA 15.92 5.37 $\pm$ 0.36 221 2.69 4.01 $\pm$ 0.05 6758
--------- ------- ----------------- --------------------- ------ ----------------- ---------------------
: Estimates of the lower boundary $\hat x_m$ and of the exponent $\hat \alpha$ of the power law fits of the distributions of normalized returns $r^{norm}$ of the drawdown and drawup for $\Delta t=30$ sec and $\epsilon_0=1$ and of the normalized positive and negative price log-returns for $\Delta t=30$ sec. The exponents $\hat \alpha$ are presented together with the estimation of its standard error ; the number of observations in the power law tail ($N_{x\ge \hat x_m}$) is also reported. []{data-label="tb:fit_dd_norm_returns"}
\(b) Drawups and positive returns
--------- ------- ----------------- --------------------- ------ ----------------- ---------------------
$N_{x\ge \hat x_m}$ $N_{x\ge \hat x_m}$
AEX 16.74 5.10 $\pm$ 0.23 486 3.37 4.27 $\pm$ 0.05 8065
CAC 14.05 4.82 $\pm$ 0.16 950 3.17 4.09 $\pm$ 0.04 11218
DAX 13.50 4.77 $\pm$ 0.15 1059 3.37 3.89 $\pm$ 0.04 9585
FTSE 12.58 4.99 $\pm$ 0.13 1481 3.05 4.40 $\pm$ 0.04 12992
MIB 13.70 5.04 $\pm$ 0.17 876 3.18 4.42 $\pm$ 0.05 8339
IBEX 12.56 4.96 $\pm$ 0.13 1363 3.47 4.34 $\pm$ 0.05 6814
STOXX 10.61 4.74 $\pm$ 0.12 1584 2.56 4.21 $\pm$ 0.03 18461
OMXS 12.86 4.82 $\pm$ 0.15 1013 3.00 4.51 $\pm$ 0.06 6721
SMI 11.54 4.72 $\pm$ 0.10 2073 3.42 4.60 $\pm$ 0.06 6587
ES 10.41 4.75 $\pm$ 0.12 1525 2.54 4.33 $\pm$ 0.03 18327
DJ 11.16 4.52 $\pm$ 0.12 1431 3.73 4.46 $\pm$ 0.07 4368
NQ 14.63 5.55 $\pm$ 0.25 487 3.07 4.35 $\pm$ 0.04 10172
HSI 13.91 4.48 $\pm$ 0.28 259 3.08 4.13 $\pm$ 0.07 3456
HCEI 15.14 5.43 $\pm$ 0.39 191 3.39 4.65 $\pm$ 0.10 2279
TAMSCI 14.31 4.61 $\pm$ 0.22 457 2.95 4.02 $\pm$ 0.05 6551
NIFTY 11.41 4.61 $\pm$ 0.12 1406 3.55 4.47 $\pm$ 0.07 3891
NIKKEI 5.88 4.49 $\pm$ 0.12 1425 3.91 5.42 $\pm$ 0.31 313
TOPIX 13.36 5.22 $\pm$ 0.39 180 3.63 5.65 $\pm$ 0.17 1051
ASX 11.94 5.45 $\pm$ 0.17 1029 2.82 4.98 $\pm$ 0.05 9786
BOVESPA 12.62 4.93 $\pm$ 0.23 460 2.78 4.17 $\pm$ 0.06 5669
--------- ------- ----------------- --------------------- ------ ----------------- ---------------------
: Estimates of the lower boundary $\hat x_m$ and of the exponent $\hat \alpha$ of the power law fits of the distributions of normalized returns $r^{norm}$ of the drawdown and drawup for $\Delta t=30$ sec and $\epsilon_0=1$ and of the normalized positive and negative price log-returns for $\Delta t=30$ sec. The exponents $\hat \alpha$ are presented together with the estimation of its standard error ; the number of observations in the power law tail ($N_{x\ge \hat x_m}$) is also reported. []{data-label="tb:fit_dd_norm_returns"}
Table \[tb:fit\_dd\_norm\_returns\] summarizes the values of the lower boundary $\hat x_m$ and exponent $\hat \alpha$ of the power law fits of the distributions of the normalized returns ($r^{norm}$) for drawdowns and drawups for $\Delta t=30$ sec and $\epsilon_0=1$. Our tests show that, when increasing both $\Delta t$ and $\epsilon_0$, the exponent $\hat\alpha$ slightly increases. However, overall, the exponent estimates are consistent for all values of $\Delta t$ and $\epsilon_0$ and coincide with a relatively good precision between drawups and drawdowns.
Figure \[fig:boxplot\] shows that the distributions of both drawups and drawdowns have lighter tails than the distribution of returns. On average, the exponent $\hat\alpha_{du}$ of the power law tail of the distribution of drawups is larger than the corresponding exponent $\hat\alpha_{pos}$ for positive returns by $0.43$. Similarly, the exponent $\hat\alpha_{dd}$ of the power law tail of the distribution of drawdowns is larger than corresponding exponent $\hat\alpha_{neg}$ for negative returns by $0.29$. Overall, the estimated exponents for drawups and drawdowns have values in the interval $4<\hat\alpha<5$ and for log-returns in the interval $3.5<\hat\alpha<4.5$.
![Box plot of estimated exponents $\hat\alpha$ (Table \[tb:fit\_dd\_norm\_returns\]); and exponents for drawdowns (red down triangles), drawups (green up triangles), positive (black squares) and negative (black circles) returns for individual contracts.[]{data-label="fig:boxplot"}](fig3_boxplot_alpha.pdf){width="\textwidth"}
These results on the exponents presented in figure \[fig:boxplot\] are paradoxical, since drawdowns cannot, by construction, be smaller than single period returns, as seen from Figure \[fig:ccdf\] in which one can observe that the distribution of returns exhibits a first-order stochastic dominance with respect to the distribution of drawdowns. As a consequence, the distribution of drawdowns should embody better the extreme loss occurrences. Taking at face value that the tail exponents of return distributions are smaller than the tail exponents of drawdown distributions, as documented in Table \[tb:fit\_dd\_norm\_returns\], would imply that their extrapolations will intersect at some point. This would imply that there can be such values of losses that are reached with higher probability in a single step rather than in a sequence of steps constituting drawdown (and similarly for drawups), which is impossible logically. A first comment is that this mathematical paradox can never be reached in real life, as this “intersection” would occur at unrealistically large values of $x\sim10^6$ (i.e. for returns that are $\sim10^6$ times larger than the volatility of the previous trading day). There is an additional argument to resolve this paradox, which is that extrapolating the power law tails assumes that they hold and would hold firmly further for larger events. But already in our data, we observe significant deviations from the power law tails for the most extreme events, as shown in figure \[fig:ccdf\], so that a power law approximation becomes highly questionable.
The fact that the distributions of individual returns $r_t^{(\Delta t)}$ have fatter tails than the distribution of normalized returns $r^{norm}$ for drawdowns made of these returns is a signature of the temporal correlation structure between returns. Moreover, it suggests that drawdowns in the tail of their marginal distribution do not result from the largest individual returns but are constructed from a sequence of smaller returns that aggregate transiently either due to external or internal market forces.
![For each analyzed contract, fraction of drawdowns (left) and drawups (right) belonging to the power law tail of their marginal distributions that contain $N=0,1,2,3,4,5$ or $6\le N\le 10$ (color-coded) large returns.[]{data-label="fig:fraction"}](fig7_fraction.pdf){width="\textwidth"}
To test this hypothesis, we calculated the number $N$ of large log-returns (in the power law tail) in each drawdown belonging to the power law tail of its normalized return $r^{norm}$ larger than the corresponding $\hat{x}_m$ given in Table \[tb:fit\_dd\_norm\_returns\]. For all analyzed drawdowns and drawups, we found that the maximal value of $N$ is 10. Then, for each $N=0,1,\dots,10$, we calculated the number of drawdowns (drawups) that contained $N$ large log-returns. Figure \[fig:fraction\] shows that the majority of drawdowns and drawups (more than 50%) do not contain any log-returns from the tail of the distribution ($N=0$). In particular, the large drawdowns and drawups for NIKKEI and ASX are constructed only from the returns that are smaller than $\hat{x}_m$, i.e. from the body of the distributions.
![Histogram of the relative contribution ($C$) of large individual log-returns (from the tail of distribution) to the normalized return of large drawdowns (red) and drawups (green) for ES Futures Contract. Vertical line denotes $C=1$ that corresponds to the situation when the drawdown is constructed only from large individual log-returns. []{data-label="fig:contribution"}](fig8_relative_contribution.pdf){width="70.00000%"}
Moreover, when they are present, the contribution of these large log-returns to the size of the drawdowns (drawups) is not special. In order to illustrate this, for each analyzed contract and each large drawdown/drawup that contain at least one large log-return (i.e. $N\ge1$), we have calculated the relative contribution $C$ of the total sum of the large individual normalized log-returns $\sum r_t^{(\Delta t)}$ to the total normalized return $r^{norm}$ of the event ($C=\sum r_t^{(\Delta t)}/r^{norm}$), which are shown in Figure \[fig:contribution\] for the E-mini S&P 500 Futures Contract (codename ES). Other contracts give similar qualitative results. One can observe that the distribution of $C$ is skewed towards small values. The mean and median values of $C$ for different contracts vary in the range $0.35-0.45$, indicating that, on average, large individual returns contribute no more than 50% of the total return of a drawdown/drawup. Very rarely, one can observe $C\ge1$, which corresponds to the situation when the whole trend is constructed only from individual large returns, and smaller returns of negative sign only provide transient corrections (see the discussion of the drawdown detection method in Section \[sec:dd\] and Figure \[fig:drawdown\_illustration\]).
Quantification of the extreme drawdowns {#sec:dragonkings}
=======================================
In the previous section, we have analyzed the tails of the distributions of normalized returns for individual Futures Contracts, which have been found well approximated by power laws over several orders of magnitude. However, some extreme events deviate substantially from the power law tail approximation (see Figure \[fig:ccdf\]). The important question discussed in the present section is whether this deviation is statistically significant or not. Statistical significance here is not a mere quantification of goodness-of-fit of distributional characteristic, but addresses the question whether these extreme events are “outliers” in the sense of generating mechanisms. If present, we will refer to these special events as “Dragon Kings”, the notion coined by one of us [@Sornette2009]. To understand what the term means, we first need to explain what are the implications of power law distributions taken as a reference point.
Recall that power law distributions embody the notion that extreme events are not exceptional events because they belong to the same distribution, which exhibits the remarkable property of scale invariance: the ratio of the frequencies of two event sizes is proportional to the ratio of the sizes and independent of the absolute values of the sizes. Only power laws have this property. In this sense, extreme events are just scaled-up versions of their smaller siblings. This is usually interpreted as evidence of the same mechanism underlying the generation of the whole population, from small to extreme sizes. As a consequence, because of the common mechanism, these tail events are intrinsically unpredictable because nothing special takes place before their occurrence that could hint of some special organisation that would put extreme events apart. In this sense, extreme events in the far right tail of power laws have been suggested as illustrating the concept of “black swans” [@Taleb2007]. This colourful term is actually a layman version of the concept of “unknown unknowns” introduced by Knight in 1930 that there exists events that we could not conceive before they happen. Therefore, strictly in the sense of Knightian uncertainty, extreme events in the far right tail of power laws are not black swans because we can conceive of their occurrence since we can quantify their frequencies. Only their specific occurrence time is not predictable.
In contrast, “Dragon King” events reveal the existence of mechanisms of self-organization that are not apparent from the (power law) distribution of their smaller siblings. Dragon Kings are often associated with a neighborhood of a phase transition, a bifurcation or a tipping point. This distinctive feature (of the approach towards a tipping point) is crucial to learn how to diagnose in advance the symptoms associated with a coming Dragon King (for a more elaborated discussion, we refer to [@Sornette2009; @SorouillonDK12]).
There have already been a number of empirical examples of Dragon Kings documented in the literature in natural and socio-economic systems, identified statistically as the outliers that occur more frequently that predicted by the power law distributions calibrated on the rest of the population: stock market crashes [@JohansenSornette2001JofRisk; @JohansenSornette1998]; some capitals in the distributions of agglomeration sizes [@PisarenkoSornette2012]; extreme events in the distribution of hydrodynamic turbulent velocity fluctuations, acoustic emissions associated with material failure or epileptic seizures occurring in the strong coupling regime [@Sornette2009]. Recently, similar “runaway” phenomena, that correspond to “negative” Dragon Kings, were documented in the distribution of citations [@GolosovskySolomon2012].
In order to detect Dragon Kings in the distributions, we employ a modified version of the *DK-test* (“Dragon-King test”) proposed by @PisarenkoSornette2012. The DK-test is based on a statistics that allows one to test quantitatively if the largest $r\ge1$ events in the tail of the empirical distribution belongs to the same distribution as the rest of the sample assumed to be a power law . We start by transforming with the following nonlinear mapping: $$\label{eq:pareto_to_exp}
y=\ln\frac{x}{x_m}.$$ If $x$ follows the Pareto distribution , then the random variable $y$ is distributed according to the exponential law: $$\label{eq:exp_ccdf}
F(y)=\mathrm{Pr}\left[Y>x\right] = 1-\exp(-\alpha y),\quad y\ge 0.$$ Given independent observations $x_i$ (and thus $y_i)$, we construct a statistical test for the following hypothesis:
> $H_0$: all observations of the sample are independently generated by the same exponential distribution .
versus its alternative:
> $H_1$: $r$ observations with the first $r$ ranks ($y_1\ge \dots\ge y_{r-1}\ge y_r$) are generated by a different distribution.
We consider the differences $\delta y_i=y_{i}-y_{i-1}$ and construct the following auxiliary variables: $$\label{eq:exp_ccdf2}
z_i=i\cdot \delta y_i=\begin{cases}
i\cdot(y_{i}-y_{i-1}), \quad &i=1,\dots,N-1;\\
N\cdot y_N,\quad &i=N.
\end{cases}$$ According to @PisarenkoSornette2012, the test statistics $$\label{eq:T}
T=\frac{z_1+\dots+z_r}{z_{r+1}+\dots+z_N}\cdot\frac{N-r}{r}$$ is distributed according to the $f$-distribution with $(2r,2N-2r)$ degrees of freedom. The corresponding p-value for the hypothesis $H_0$ is given by: $$\label{eq:pvalue}
p(r;y_1,\dots,y_N)=1-F(T, 2r, 2N-2r),$$ where $F(t, a, b)$ is the cumulative distribution function (cdf) of the $f$-distribution with $(a,b)$ degrees of freedom. Following @PisarenkoSornette2012, we will use threshold the $p_0=0.1$ for rejecting the null hypothesis (for $p<0.1$) and declaring the $r$ largest events as “outliers”.
We note the existence of a limitation of the original DK-test of @PisarenkoSornette2012, which appears when the value of the observation with the first rank (or the first few ranks) is extremely large. In this case, this single observation will contribute heavily to the sum in the numerator of expression , which will remain large even for large values of $r$. For example, in synthetic samples of 100 variables distributed according to a pure exponential, adding a single outlier with a size $2\cdot y_1$ (i.e. twice larger than the maximal observed value) is sufficient to distort the statistic (\[eq:T\]) and make it declare that the 10 to 20 largest observations are“outliers”, when using the original formulation of DK-test [@PisarenkoSornette2012]. The original DK-test thus does not fail to correctly identify the presence of Dragon Kings, but it may over-estimate their numbers due to the contamination from the existence of a super-large one.
The natural solution to this problem when testing if the $r$-th rank is an outlier is to remove the $r-1$ largest observations from the set before calculating the DK statistics, In other words, when testing if rank $r$ is an outlier, we remove the $r-1$ larger values and make the rank $r$-varialble the new rank 1. For instance, for $r=2$, we remove the largest observation $r_1$ from the set and then test the null hypothesis versus its alternative that $y_2$ is generated by a different distribution than the rest of the observations ($y_3\ge y_4\ge\dots\ge y_N$). We employ this procedure iteratively starting with $r=1$ in order to find the minimal value of $r$ for which the null hypothesis $H_0$ can no more be rejected.
However, this procedure does not address the lack of power of the DK-test when several large outliers are present, as occurs for instance in the distributions of drawdowns for the CAC and FTSE contracts shown in Figure \[fig:ccdf\]. As an illustration, consider the synthetic sample of 100 exponentially distributed variables $y_1,\dots, y_{100}$, with two additional introduced outliers of sizes respectively $2\cdot y_1$ and $2.2\cdot y_1$. The application of the iterative procedure described above fails to detect the largest outlier of size $2.2\cdot y_1$. The reason is similar to the one mentioned before: in this case, $y_2$ contributes substantially to the sum in the denominator of expression and the overall value of $T$ is not large enough to reject $H_0$. In other words, the value $2.2\cdot y_1$ is not detected as an outlier in the presence of the value $2\cdot y_1$ together with the other 100 exponentially distributed variables. Removing $2.2\cdot y_1$ and applying the DK-test on the variable $2\cdot y_1$ compared with the other remaining 100 exponentially distributed variables $y_1,\dots, y_{100}$ as suggested in our iterative procedure does diagnose $2\cdot y_1$ as an outlier. This creates a paradox, as the largest value is not an outlier but the second largest one would be.
To overcome all these issues, we propose the following modification of the original DK-test of @PisarenkoSornette2012. The $r$ largest observations $y_1,\dots, y_r$ are diagnosed as being “Dragon-Kings” or “outliers” if and only if the remaining observations $y_{r+1},\dots,y_N$ contains no outliers and if each of the variable $y_1,\dots, y_r$ when introduced individually in the remaining set individually can be qualifier as an outlier using the F-test . In other words, we will require that the following inequalities are simultaneously satisfied: $$\label{eq:modified_DK}
\begin{cases}
p_1=p(1;y_1,y_{r+1},\dots,y_N)<p_0;\\
p_2=p(1;y_2,y_{r+1},\dots,y_N)<p_0;\\
\dots\\
p_r=p(1;y_{r},y_{r+1},\dots,y_N)<p_0;\\
p_{r+1}=p(1;y_{r+1},y_{r+2},\dots,y_N)\ge p_0.\\
\end{cases}$$ This set of conditions (\[eq:modified\_DK\]) captures the logic that, for the first $r$ ranks to be outliers, the distribution of the other $N-r$ variables is an unperturbed exponential law when removing these $r$ outliers, and each of the $r$ outliers should be individually diagnosed as aberrant.
\(a) Drawdowns
\(b) Drawups
Table \[tb:dk\_dd\] presents the results of the modified DK-test for the distribution of normalized returns $r^{norm}$ of drawdowns and drawups for different contracts. For each test , we have sampled the $N=200$ largest events of the distribution tail. The largest number of “Dragon-Kings” were found in drawdowns of the CAC contract (3 events), then 2 “Dragon-Kings” in the drawdowns of the DAX, and FTSE and in the drawups of the NQ contracts; and 1 “Dragon-King” was detected in the drawdowns of the IBEX, ES, DJ and NQ and in the drawups of the CAC and TAMSCI contracts.
Table \[tb:dk\_dd\] documents that the most famous intraday price swing, the so-called “flash crash” of May 6, 2010 (13:41:30 EST), can be diagnosed as a “Dragon King” event for all three US E-mini futures contracts: ES, DJ and NQ. Over the time of 4 minutes (5 for NQ), the price of these index futures contracts dropped by $r^{norm}=132.62$, $165.04$ and $149.19$ respectively, which means that the price drop was more than a 130-sigma event (i.e., 132–165 times larger than the volatility of 30-second returns). As shown by @FilimonovSornette2012_Reflexivity, the dynamics of the high-frequency mid-quote price during this “flash-crash” exhibited a unique pattern, indicating an extreme degree of self-excitation during that event. As reported, slightly before and during the price drop, the system became critical, being essentially driven by the internal feedback mechanisms rather than the external information flow.
Two consecutive drawups occurring after the “flash crash” of NQ (13:46:00 and 13:49:30 EST with $r^{norm}=96.24$ and $80.39$) can be also claimed to be “outlier” of the respective probability distribution. The “flash crash” started on E-mini S&P 500 futures contracts [@FlashCrash2010_report], and then almost instantly propagated to the constituting stocks of the index as a result of the arbitrage between ETF and Futures and between ETF and the underlying assets [@BenDavid2011]. The “Flash crash” of May 6, 2010 was attributed to the activity of high-frequency traders according to the joint SEC and CFTC report [@FlashCrash2010_report]. These high-frequency traders did not trigger the crash but contributed significantly to the market volatility and extraordinarily amplified the initial price drop.
While being the best known, the “flash crash” of May 6, 2010 is not the largest one in relative values in the data analysed here. A remarkable drawdown of $r^{norm}=214.99$ (“215-sigma event”) was experienced in the futures on the CAC index on December 27, 2010 at 09:03:00 CET. This “mini flash crash” at the opening of that trading day initially started on the DAX index and quickly propagated to other other European markets through cross-market arbitrage, many of which experienced large drawdowns. However, only for the CAC and DAX futures contracts can this event be quantified as a “Dragon King” according to our modified DK-test (\[eq:modified\_DK\]). The CAC futures contract recovered from the “mini flash crash” via an extreme drawup of size $r^{norm}=157.99$, which is also qualified as a dragon-king (the hypothesis $H_0$ that this event belongs to the overall distribution can be rejected).
These two “flash crashes” in US markets on May 6 and in Europe on December 27, 2010 represent endogenous events, where were not generated by external news but by a self-exciting activity of the market participants. But not all dragon-king events are generated internally, as illustrated by the plunge of European markets in response to the 7 July 2005 London bombings, when a series of coordinated suicide attacks targeted the civilian public transportation system in central London during the morning rush hour. At that time, almost all analyzed contracts experienced an outstanding drop and, for the CAC, DAX, FTSE and IBEX contracts, these price fluctuations can be quantified as “Dragon Kings” of their respective drawdown distributions.
In general, the Asian, Australian and South American contracts do not exhibit any outstanding price fluctuations that can be quantified as “Dragon King”, with a single exception. On the opening of the Taiwanese markets on September 10, 2009, the price of TAMSCI futures contracts experienced a drawup with $r^{norm}=127.45$. After the drawup of the CAC contract following the “mini flash crash” described above, this is the largest normalized return over all observed drawups for all analyzed contracts. This extraordinary drawup of the market resulted from the announcements of an agreement between Taiwan and mainland China to allow mainland investors to buy stocks in Taiwan.
\(a) Normalized velocity, $v^{norm}$
\(b) Duration, $\tau$
Table \[tb:dk\_dd\] shows that some of the reported extreme (in terms of normalized return) drawdowns and drawups are also the fastest. This is the case for the drawdowns observed for CAC, SMI, ES, DJ and NQ and for the drawups for FTSE and NQ. However, the normalized speeds of these events are not exceptional. Moreover, from the point of view of the speed statistics, these events cannot be quantified as “Dragon Kings”. Table \[tb:dk\_dd\_dur\_speed\] (a) reports the “outliers” that our modified DK-test detects in the distributions of the normalized speeds. None of the reported “Dragon Kings” of the normalized returns (Table \[tb:dk\_dd\]) is also exceptional in terms of speed. The only common event in these two tables — the drawup for the FTSE contract on 2005-07-07 10:18:30 GMT — cannot be quantified as a “Dragon King” with respect to its normalized return $r^{norm}$ (this is vividly illustrated in Figure \[fig:ccdf\]). It is also interesting to observe that, in general, the power law fit of the distribution of normalized speeds is more robust than the distribution of normalized returns (Figure \[fig:ccdf\]).
Finally, Table \[tb:dk\_dd\_dur\_speed\] (b) reports the “Dragon Kings” detected in the distributions of duration $\tau$ and shows that all these events are not even close to be the largest or the fastest. This raises naturally the issue of the dependence between these different measures. For this, we will discuss the so-called tail dependence (dependence of extreme values) later in Section \[sec:dependence\].
Aggregated distributions {#sec:distrib_aggragated}
========================
In the previous section, we have analyzed the distribution of normalized returns for individual contracts. As discussed above, we are considering normalized characteristics of drawups and drawdowns (such as returns $r^{norm}$ and speeds $v^{norm}$), which allows us to compare directly these values for events of different time periods and different assets. Moreover, we can aggregate all values from different contracts, to check if our conclusions are robust with respect to sampling.
![Complementary cumulative distribution function (ccdf) for the (i) aggregated normalized returns, (ii) aggregated normalized speeds and (iii) aggregated durations of drawdowns (red down triangles) and drawups (green up triangles) for $\epsilon_0=1$ and $\Delta t=30$ sec. Black straight lines correspond to power law fits of the tails of distributions of drawdowns (see Table \[tb:fit\_aggregated\]). Red and green dots on the plot (i) correspond to distributions of the aggregated normalized log-returns over the time scale $\Delta t=30$ sec. The dashed black lines depict the power law fits of the tails of these distributions (see Table \[tb:fit\_aggregated\]).[]{data-label="fig:ccdf_aggregated"}](fig4_ccdf_aggregated.pdf){width="\textwidth"}
Figure \[fig:ccdf\_aggregated\] presents the corresponding aggregated empirical distribution functions for the normalized returns, speeds and durations, and the power law approximations of the tails of these distributions. Table \[tb:fit\_aggregated\] lists the parameters of the power law fits .
Characteristic Event type $\hat x_m$ $\hat\alpha$ $N_{x\ge \hat x_m}$
---------------- ------------ ------------ ----------------- --------------------- -- --
DD 17.62 4.64 $\pm$ 0.06 5197
DU 16.03 5.01 $\pm$ 0.07 5943
RNeg 3.55 4.12 $\pm$ 0.01 101512
RPos 3.03 4.25 $\pm$ 0.01 172093
DD 0.10 4.14 $\pm$ 0.02 39677
DU 0.10 4.32 $\pm$ 0.02 35911
DD 960.00 4.46 $\pm$ 0.04 12322
DU 1020.00 4.46 $\pm$ 0.04 10871
: Estimates of the lower boundary $\hat x_m$ and exponent $\hat \alpha$ of the power law fits of the distributions of (i) normalized returns $r^{norm}$, (ii) normalized speeds $v^{norm}$ and (iii) durations $\tau$ for normalized drawdown (DD) and drawup (DU), as well as for normalized log-returns : positive (RPos) and negative (Rneg). The number of observations qualified in the power law tail ($N_{x\ge \hat x_m}$), log-likelihood ratio ($\mathcal{R}$) and p-values for the significance of likelihood ratio test are also given.[]{data-label="tb:fit_aggregated"}
These results validate our previous findings. First, fits of the distribution of drawdowns/drawups are much more robust than for individual log-returns. Second, the reported exponents of the power law tails for the aggregated distributions lie in the same range of values that we reported for individual contracts (see Figure \[fig:boxplot\]). Third, the estimated exponent $\hat\alpha$ for drawdowns is significantly larger than the exponent $\hat\alpha$ for drawups (the difference is larger than 6 standard errors).
Figure \[fig:ccdf\_aggregated\] and Table \[tb:fit\_aggregated\] show that the power law approximation of the normalized speeds of drawdowns and drawups is almost perfect and holds for more than 5 orders of magnitude in the vertical axis. In contrast, the fits of the tails of the distributions of durations are relatively poor. In particular, the hypothesis that the distribution of drawup durations is a power law can be rejected in favor of the stretched exponential distribution family using the nested Wilks’ test [@Sornette2005].
An important observation from Figure \[fig:ccdf\_aggregated\] is that the extreme events of individual distributions (Figure \[fig:ccdf\]) also branch off the aggregated distribution. One can clearly see that up to ten of the largest drawdowns and drawups deviate substantially from the power law fit of the tail. Interestingly, the original [@PisarenkoSornette2012] and the modified (Section \[sec:dragonkings\]) DK-tests give contradictory and confusing results. However, “the absence of evidence is not the evidence of absence”, which summarises the fallacy of the argumentum ad ignorantiam. In other words, we argue that the failure to diagnostic the largest drawdowns as outliers reflects the lack of power of these tests. As discussed above, the 15–20 extremes events in the tail contribute substantially to the numerator in expression and lead to a spurious identification of up to 400 outliers in the tail (the $H_0$ hypothesis is rejected for values of ranks $r$ up to $r=400$). On the another hand, the modified DK-test sins at the other extreme by being too conservative and fails to reject $H_0$ for any $r$, because it requires the simultaneous rejection of $H_0$ for $y_{r},y_{r+1},\dots,y_N$ and the acceptance of $H_0$ for $y_{r+1},\dots,y_N$. This typically does not occur when outliers are not a few far-standing events, but are organised with a continuous and smooth deviation of the tail as in Figure \[fig:ccdf\_aggregated\].
To address this issue and test for the presence of a change of regime in the tail of the distribution, we employ the parametric U-test [@PisarenkoSornette2012], which tests deviation of the tail with respect to the fitted power law distribution rather than with respect to the rest of sample as in the DK-test. We present a slightly modified description of the U-test from that found in [@PisarenkoSornette2012] and provide a closed-form solution of the maximum likelihood estimation of the power law exponent.
We select the lower threshold $x_m$ for the calibration of the power law in the distribution tail and apply the same nonlinear transformation as for the DK-test. Then, by visual inspection, we determine a candidate for the rank $r$ such that observations smaller than $y_{r+1}$ (i.e., of rank larger than $r$) are distributed according to the exponential distribution and the total number of outliers is not larger than $r$.
The exponent $\alpha$ of the exponential distribution can be estimated with the Maximum Likelihood method applied to the subsample $y_{r+1},\dots,y_N$, where the likelihood with right-censored observations is given by $$\label{eq:lik}
\mathcal{L}(\alpha|y_{r+1},\dots,y_N)=\big[1-F(y_{r+1})\big]^r\prod_{k=r+1}^Nf(y_k),$$ where $F(y)$ is the cdf of the exponential distribution and $f(x)=\alpha\exp(-\alpha x)$ is the corresponding pdf. The exponent $\alpha$ can then be estimated by maximizing $\log\mathcal{L}$. In the case of an exponential distribution , this yields the closed form expression $$\label{eq:alpha_exp}
\hat\alpha=(N-r)\cdot\left[ry_{r+1}+\sum_{k=r+1}^Ny_k\right]^{-1}$$ The p-values that the $k$ smallest ranks deviate from the null hypothesis of the exponential distribution can be then obtained from the following equation (see derivations in @PisarenkoSornette2012): $$\label{eq:pvalue_U}
p_k=1-\mathcal{B}\big(F(y_k);n-k+1, k\big),\quad k=1,\dots, r,$$ where $\mathcal{B}(y; a, b)$ is the normalized incomplete beta-function, and the exponent $\alpha$ for the probability distribution is taken to be equal to the MLE (\[eq:alpha\_exp\]). The event $y_k$ for which $p_k<p_0=0.1$ can be then diagnosed as an outlier with respect to the fitted exponential distribution of the tail. This corresponds to the original event $x_k$ being a “Dragon-King” with respect to the fitted power law distribution.
Table \[tb:dk\_aggregated\] presents results of the application of the U-test to the aggregated distribution of normalized returns $r^{norm}$ of drawdowns, where the threshold $x_m$ is selected by using the Kolmogorov-Smirnov test (Table \[tb:fit\_aggregated\]). The first important observation is that all “Dragon-King” events that were detected for individual contracts (Table \[tb:dk\_dd\]) are also quantified as belonging to a different regime than the power law for the aggregated distribution. This supports the findings of Section \[sec:dragonkings\].
However, not all extremes (events with rank 1) of the individual contracts are present n the upper tail of the aggregated distribution. For drawdowns, the extreme tail of the aggregated distribution contains mostly events from European and US markets. For example, the extreme drawdowns for OMXS, HCEI, TAMSCI, NIKKEI, TOPIX, ASX and BOVESPA are not qualified as an outliers (neither they were reported as “Dragon-Kings” at the individual level). For drawups, the number of detected outliers is much smaller than for drawdowns.
On the contrary, several events classified as outliers at the aggregate level were not reported as individual “Dragon-Kings” using the (conservative) modified DK-test. For example, we were unable to reject the null hypothesis $H_0$ for the two largest events occurring in the SMI futures contracts (Figure \[fig:ccdf\]). For $r=2$, we report the following p-values: $p_1=0.03$, $p_2=0.13$ and $p_3=0.47$; for $r=3$, we obtain p-values: $p_1=0.01$, $p_2=0.05$, $p_3=0.47$ and $p_4=0.99$. In both cases, one inequality of the system does not hold. This results from the fact that the second event $x_2$ is not sufficiently larger than $x_3$, which leads to the absence of rejection of the null (no dragon-kings) for $r=2$ ($p_2=0.13>0.1$). And the third event $x_3$ only slightly deviates from the tail ($p_3=0.47$ for $r=3$), which leads to the absence of rejection of the null for larger $r$ values.
Finally, all events detected as outliers of the aggregate distribution (Table \[tb:dk\_aggregated\]) can be detected in their corresponding individual distributions using the U-test. However, being dependent on the calibration of the exponent of the power law, the U-test is subjected to estimation errors that need to be accounted properly. The nonparametric DK-test is free from this drawback, at the cost of having more limited power. In general, as with any statistical testing problem, it is always a good practice to consider several different tests to confirm the conclusions.
Tail dependence characteristics of extreme drawdowns {#sec:dependence}
====================================================
Are extreme drawdowns (drawups) associated with the largest speed and/or the longer durations? Clarifying the interdependence between size, speed and duration of extreme drawdown (drawups) is important to better understand their generating mechanism. In previous sections, we have already commented that events that are extreme with respect to one characteristic may not be extreme with respect to another (see Tables \[tb:dk\_dd\] and \[tb:dk\_dd\_dur\_speed\]). The largest (with respect to $r^{norm}$) drawdowns and drawups are often not the fastest, and by far not the longest events in the population. The occurrence of an extreme normalized speed $v^{norm}$ does not ensure that the event will have an extreme size. Moreover, the longest drawdowns and drawups typically have relatively small returns.
Our goal here is to quantify the mutual interdependence between size, speed and duration. Generally, the complete information about the dependence between two random variables $X$ and $Y$ is contained in their copula structure. Here, we consider a simpler metric, the tail dependence, which is defined at the probability of observing a very large value of one variable conditional on the occurrence of a very large value of the other variable [@Sornette_Risks2005]: $$\label{eq:tail_dependence}
\lambda=\lim_{u\to1}\lambda_u=
\lim_{u\to1}\mbox{Pr}[X>F_X^{-1}(u)|Y>F_Y^{-1}(u)],$$ where $F_X(\cdot)$ and $F_Y(\cdot)$ are the marginal cumulative distributional functions of $X$ and $Y$. In practice, it is difficult to work with the asymptotic tail dependence $\lambda$, which is defined in the empirically unattainable limit $u \to 1$. We will thus consider $\lambda_u$ for fixed value of probability $u\lesssim1$ and document the behaviour of $\lambda_u$ as $u$ approaches $1$ from below.
![Tail dependence coefficients $\lambda_u$ for normalized returns $r^{norm}$ and normalized speeds $v^{norm}$ (solid lines) and for normalized returns $r^{norm}$ and durations $\tau$ (dashed lines) of the aggregated drawdowns (red) and drawups (green) for different contracts and different probabilities $u$.[]{data-label="fig:td_aggregated"}](fig6_td_aggregated.pdf){width="70.00000%"}
Figure \[fig:td\_aggregated\] shows the non-asymptotic tail dependence coefficients $\lambda_u$ of (i) $X=r^{norm}$ and $Y=v^{norm}$ and (ii) $X=r^{norm}$ and $Y=\tau$ for the aggregated probability distributions $F(r^{norm}, v^{norm})$ and $F(r^{norm}, \tau)$ (marginal distributions are presented in Figure \[fig:ccdf\_aggregated\]). In other words, figure \[fig:td\_aggregated\] quantifies the probability that the observed drawdown is large, conditional on it being (i) fast or (ii) long. One can observe that $\lambda_u$ of the normalized returns conditional on the durations decreases monotonously with $u$ and converges to zero as $u\to1$. This indicates an absence of dependence of the extreme values of size and durations. In other words, the longest drawdowns and drawups do not belong to the highest quantiles in term of sizes.
On the contrary, the tail dependence $\lambda_u$ between returns and speed is significant and tends to increase for $u \to 1$, except very close to $1$ ($u>0.9997$) due to the finite size of the data sample. The estimated value of the tail dependence $\lambda_u$ between returns and speed is approximatively in the range $0.09-0.12$ for drawdowns and $0.075-0.12$ for drawups, i.e., conditional on a very large speed, there is about a 10% probability that the corresponding drawdown (drawup) is extreme in normalised return. Figure \[fig:tail\_dependence\], which presents the tail dependence coefficients $\lambda_u$ at three probability levels $u=0.990, 0.995$ and $0.999$ for individual contracts, supports our previous findings at the aggregate level. With the exception of the OMXS, HCEI and ASX contracts that are characterised by monotonously decaying $\lambda_u$, all other analyzed future contracts exhibit clear signatures of non-zero tail dependence with $\lambda_u$ varying in the range $0.03-0.035$ (for NIKKEI, TOPIX) and $0.1-0.12$ (for CAC, DAX, AEX, STOXX, DJ, NIFTY).
![Tail dependence coefficient $\lambda_u$ between normalized returns $r^{norm}$ and normalized speeds $v^{norm}$ of drawdowns (red bars) and drawups (green bars, inverted x-axis) for different contracts and different probability levels $u$. The last row corresponds to the aggregated distributions. []{data-label="fig:tail_dependence"}](fig5_tail_dependence.pdf){width="\textwidth"}
Conclusion {#sec:conclusion}
==========
We have investigated the distributions of $\epsilon$-drawdowns and $\epsilon$-drawups of the most liquid futures financial contracts of the world at time scales of $30$ seconds. The $\epsilon$-drawdowns and $\epsilon$-drawups defined by expressions (\[eq:delta\]) with (\[eq:cond\]) are proposed as robust measures of the risks to which investors are arguably the most concerned with. The time scale of $30$ seconds for the time steps used to defined the drawdown and drawups is chosen as a compromise between robustness with respect to microstructure effects and reactivity to regime changes in the time dynamics. Similarly to the distribution of returns, we find that the distributions of $\epsilon$-drawdowns and $\epsilon$-drawups exhibit power law tails, albeit with exponents significantly larger than those for the return distribution. This paradoxical result can be attributed to (i) the existence of significant transient dependence between returns and (ii) the presence of large outliers (termed dragon-kings [@Sornette2009; @SorouillonDK12]) characterizing the extreme tail of the drawdown/drawup distributions deviating from the power law. We present the generalised non-parametric DK-test together with a novel implementation of the parametric U-test for the diagnostic of the dragon-kings. Studying both the distributions of $\epsilon$-drawdowns and $\epsilon$-drawups of individual future contracts and of their aggregation confirm the robustness and generality of our results. The study of the tail dependence between drawdown/drawup sizes, speeds and durations indicates a clear relationship between size and speed but none between size and duration. This implies that the most extreme drawdown/drawup tend to occur fast and are dominated by a few very large returns. These insights generalise and extend previous studies on outliers of drawdown/drawup performed at the daily scale [@JohansenSornette2001JofRisk; @JohSorepsidd10].
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank Professor Frédéric Abergel and the Chair of Quantitative Finance of l’École Centrale de Paris (<http://fiquant.mas.ecp.fr/liquidity-watch/>) for the access to the high-frequency data used in the present analysis. We are very grateful to Professor Yannick Malevergne for many fruitful discussions while preparing this manuscript.
The statistical analysis of the data was performed using open source software: Python 2.7 (<http://www.python.org>) and libraries: Pandas [@McKinney_Pandas2012], NumPy (<http://www.numpy.org/>), SciPy (<http://www.scipy.org/>), IPython [@Perez_IPython2007] and Matplotlib [@Hunter_Matplotlib2007].
[45]{} natexlab\#1[\#1]{}url \#1[`#1`]{}urlprefix
Andersen, T. G., Bollerslev, T., Diebold, F. X., Labys, P., Mar. 2000. [Great realisations]{}. Risk, 105–108.
Arneodo, A., Muzy, J.-F., Sornette, D., 1998. [“Direct” causal cascade in the stock market]{}. The European Physical Journal B - Condensed Matter and Complex Systems 2 (2), 277–282.
Bacry, E., Delattre, S., Hoffmann, M., Muzy, J.-F., 2013. [Modeling microstructure noise with mutually exciting point processes]{}. Quantitative Finance 13 (1), 65–77.
Barndorff-Nielsen, O. E., Hansen, P. R., Lunde, A., Shephard, N., Nov. 2009. [Realized kernels in practice: trades and quotes]{}. The Econometrics Journal 12 (3), C1–C32.
Ben-David, I., Franzoni, F. A., Moussawi, R., 2011. [ETFs, Arbitrage, and Shock Propagation]{}.
Borland, L., Bouchaud, J.-P., Muzy, J.-F., Zumbach, G., Mar. 2005. [The Dynamics of Financial Markets: Mandelbrot’s multifractal cascades, and beyond]{}. Wilmott Magazine.
Brownlees, C. T., Gallo, G. M., Dec. 2006. [Financial econometric analysis at ultra-high frequency: Data handling concerns]{}. Physica A: Statistical Mechanics and its Applications 51 (4), 2232–2245.
Calvet, L. E., Fisher, A. J., 2002. [Multifractality in asset returns: theory and evidence]{}. Review of Economics and Statistics 84 (3), 381–406.
Campbell, J. Y., Lo, A. W., MacKinlay, A. C., 1996. [The Econometrics of Financial Markets]{}. Princeton University Press.
Carpentier, A., Kim, A. K. H., Dec. 2013. [Honest and adaptive confidence interval for the tail coefficient in the Pareto model]{}.
, [SEC]{}, Oct. 2010. [Findings Regarding the Market Events of May 6, 2010: Report of the Staffs of the CFTC and SEC to the Joint Advisory Committee on Emerging Regulatory Issues]{}. Tech. rep.
Chakraborti, A., Toke, I. M., Patriarca, M., Abergel, F., Jul. 2011. [Econophysics review: I. Empirical facts]{}. Quantitative Finance 11 (7), 991–1012.
Clauset, A., Shalizi, C. R., Newman, M. E. J., 2009. [Power-Law Distributions in Empirical Data]{}. SIAM Review 51 (4), 661.
Cont, R., 2001. [Empirical properties of asset returns: stylized facts and statistical issues]{}. Quantitative Finance 1, 223–236.
de Vries, C. G., Leuven, K. U., 1994. [Stylized facts of nominal exchange rate returns]{}. In: van der Ploeg, F. (Ed.), The Handbook of International Macroeconomics. Blackwell, Oxford, pp. 348–389.
DeGroot, M. H., Schervish, M. J., 2011. [Probability and Statistics]{}, 4th Edition. Pearson.
Deluca, A., Corral, [Á]{}., Sep. 2013. [Fitting and goodness-of-fit test of non-truncated and truncated power-law distributions]{}. Acta Geophysica 61 (6), 1351–1394.
Falkenberry, T. N., Sep. 2002. [High Frequency Data Filtering]{}. Tech. rep., Tick Data, Inc.
Filimonov, V., Sornette, D., 2011. [Self-excited multifractal dynamics]{}. Europhysics Letters 94 (4), 46003.
Filimonov, V., Sornette, D., 2012. [Quantifying reflexivity in financial markets: Toward a prediction of flash crashes]{}. Physical Review E 85 (5), 056108.
Filimonov, V., Sornette, D., Aug. 2013. [Apparent criticality and calibration issues in the Hawkes self-excited point process model: application to high-frequency financial data]{}. Swiss Finance Institute Research Paper No. 13-60.
Garman, M. B., Klass, M. J., 1980. [On the Estimation of Security Price Volatilities from Historical Data]{}. The Journal of Business 53 (1), 67–78.
Golosovsky, M., Solomon, S., May 2012. [Runaway events dominate the heavy tail of citation distributions]{}. The European Physical Journal Special Topics 205 (1), 303–311.
Gopikrishnan, P., Meyer, M., Amaral, L. A. N., Stanley, H. E., Jul. 1998. [Inverse cubic law for the distribution of stock price variations]{}. The European Physical Journal B - Condensed Matter and Complex Systems 3 (2), 139–140.
Hunter, J. D., 2007. [Matplotlib: A 2D Graphics Environment]{}. Computing in Science and Engineering 9 (3), 90–95.
Jensen, M. H., Johansen, A., Simonsen, I., 2003. [Inverse statistics in economics: the gain-loss asymmetry]{}. Physica A: Statistical Mechanics and its Applications 324 (1-2), 338–343.
Johansen, A., Ledoit, O., Sornette, D., 2000. [Crashes as Critical Points]{}. International Journal of Theoretical and Applied Finance 3 (2), 219–255.
Johansen, A., Sornette, D., Feb. 1998. [Stock market crashes are outliers]{}. The European Physical Journal B - Condensed Matter and Complex Systems 1, 141–143.
Johansen, A., Sornette, D., 2001. [Large stock market price drawdowns are outliers]{}. Journal of Risk 4 (2), 69–110.
Johansen, A., Sornette, D., 2010. [Shocks, Crashes and Bubbles in Financial Markets]{}. Brussels Economic Review (Cahiers economiques de Bruxelles) 53 (2), 201–253.
Malevergne, Y., Pisarenko, V., Sornette, D., Sep. 2005. [Empirical distributions of stock returns: between the stretched exponential and the power law?]{} Quantitative Finance 5 (4), 379–401.
Malevergne, Y., Sornette, D., Dec. 2005. [Extreme Financial Risks: From Dependence to Risk Management]{}. Springer.
Masteika, S., Rutkauskas, A. V., Alexander, J. A., 2012. [Continuous futures data series for back testing and technical analysis]{}. In: 2012 International Conference on Economics, Business and Marketing Management. pp. 265–269.
McKinney, W., 2012. [Python for Data Analysis]{}. O’Reilly Media.
Newman, M., Sep. 2005. [Power laws, Pareto distributions and Zipf’s law]{}. Contemporary Physics 46 (5), 323–351.
Pagan, A., May 1996. [The econometrics of financial markets]{}. Journal of Empirical Finance 3 (1), 15–102.
P[é]{}rez, F., Granger, B. E., May 2007. [IPython: A System for Interactive Scientific Computing]{}. Computing in Science and Engineering 9 (3), 21–29.
Pisarenko, V., Sornette, D., May 2012. [Robust statistical tests of Dragon-Kings beyond power law distributions]{}. The European Physical Journal Special Topics 205 (1), 95–115.
Rubinstein, M., 1988. [Portfolio insurance and the market crash]{}. Financial Analysts Journal Jan-Feb, 38–47.
Saichev, A., Sornette, D., 2014. [A simple microstructure return model explaining microstructure noise and Epps effects]{}. International Journal of Modern Physics C 25 (6), 1450012 (36 pages).
Sornette, D., Dec. 2009. [Dragon-Kings, Black Swans and the Prediction of Crises]{}. International Journal of Terraspace Science and Engineering 2 (1).
Sornette, D., Knopoff, L., Kagan, Y., Vanneste, C., 1996. [Rank-ordering statistics of extreme events: application to the distribution of large earthquakes]{}. J.Geophys.Res. 101, 13883–13893.
Sornette, D., Ouillon, G., 2012. [Dragon-kings: mechanisms, statistical methods and empirical evidence]{}. Eur. Phys. J. Special Topics 205, 1–26.
Taleb, N. N., 2007. [The Black Swan: The Impact of the Highly Improbable]{}. Random House.
Wager, S., Sep. 2014. [Subsampling extremes: From block maxima to smooth tail estimation]{}. Journal of Multivariate Analysis 130, 335–353.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The structures and distributions of light nuclei are investigated within a microscopic correlation model. Two particle correlations are responsible for the scattering of model particles either to low momentum- or to high momentum-states. The low momentum states form the model space while the high momentum states are used to calculate the G-matrix. The three and higher order particle correlations do not play a role in the latter calculation especially if the correlations induced by the scattering operator are of sufficient short range. They modify however, via the long tail of the nuclear potential, the Slater determinant of the (A) particles by generating excited Slater’s determinants.'
author:
- TOMASELLI
- 'K[Ü]{}HL'
- URSESCU
- FRITZSCHE
date: 'Received: date / Accepted: date'
title: Correlation Effects on the Charge Radii of Exotic Nuclei
---
[example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore
Introduction {#intro}
============
Correlation effects in nuclei have first been introduced in nuclei by Villars [@vil63], who proposed the unitary-model operator (UMO) to construct effective operators. The method was implemented by Shakin [@sha66; @sha68] for the calculation of the G-matrix from hard-core interactions.
The UMO is based on the separation of the two body potential in a short and a long components. Within this separation the effective n-body Hamiltonian contains only the long component. The short-range component is considered up to the two body correlation and produces no energy shift in the pair state.
Non perturbative approximations of the UMO have been recently applied to even nuclei in Ref. \[4,5\] which is treated here in more detail. The basics formulas of the Boson Dynamic Correlation Model (BDCM) presented in the above quoted paper have been obtained by solving the n-body problem in terms of the long range component of the two-body force. This component has the effect of generating a new correlated model space (effective space) which departs from the originally adopted one (shell model). The amplitudes of the model wave functions are calculated in terms of non linear equation of motions (EoM).
By linearizing the systems of commutator equations, which characterize the EoM, we derive the eigenvalue equations for our model space. Within this correlated formalism we generate a model that includes not only the ladder diagrams of Ref. \[6\] but also the folded diagrams of Kuo [@kuo01].
The n-body matrix elements which define the eigenvalue equations are calculated exactly via the Cluster Factorization Theory (CFT) [@tom06].
In this paper the BDCM model is applied to calculate the influence of the correlations on the charge distributions of the lithium isotopes and on the charge distribution of $^6$He.
The value obtained for the charge radius of the correlated $^6$He is slightly bigger than the radius calculated in other theories \[9-13\] and that derived within the isotopic-shift IS theory [@wan04]. A charge radius which agrees with the radii calculated in the Refs. \[9-13\] and those calculated in the cluster models of Refs. \[15-17\] is on the other hand obtained by considering only two protons in the $s_{\frac{1}{2}}$. This non correlated radius agrees also with the radius derived at Argonne within the IS theory [@wan04]. Correlations have therefore the property to increase the charge radius of $^6$He as observed for the isotopes of Lithium.
The calculations performed in Ref. \[18\] for the charge radii of the lithium isotopes, although in good agreement with those measured at GSI-TRIUMF \[19-21\] and analyzed with the help of Ref. \[22\], are always slightly larger than those measured. For the stable isotope $^6$Li the calculated radius agrees with the value obtained with the electron scattering experiments of Ref. \[23\]. However, the charge radii calculated in the IS theory could also depend on the nuclear correlations. The calculations performed in Ref. \[24\] for the field shift (FS) of $^7$Li show that the departure from a point nuclear approximation is a rather large effect. Additionally the higher order cross term contributions of Ref. \[25\] need to be considered.
A direct comparison between the calculated and the measured charge radii should therefore be performed after an accurate analysis of these two correcting factors.
Theory of Two Correlated Particles {#sec:1}
==================================
In order to describe the structures and the distributions of nuclei we start from the following Hamiltonian: $$\label{equ.1}
H= \sum_{\alpha\beta}
\langle\alpha|t|\beta\rangle \, a^{\dagger}_{\alpha}a_{\beta}
\:+\: \sum_{\alpha\beta\gamma\delta}
\langle\Phi_{\alpha\beta}| v_{12}|\Phi_{\gamma\delta}\rangle \,
a^{\dagger}_{\alpha}a^{\dagger}_{\beta} a_{\delta}a_{\gamma}$$ were $v_{12}$ is the singular nucleon-nucleon two body potential. Since the matrix elements $|\alpha\beta\rangle$ are uncorrelated the matrix elements of $v_{12}$ are infinite. This problem can be avoided by taking matrix elements of the Hamiltonian between correlated states. In this paper the effect of correlation is introduced via the $e^{iS}$ method. In dealing with very short range correlations only the $S_2$ part of the correlation operator need to be considered.
Following Ref. \[2,3\] we therefore calculate an “effective Hamiltonian” by using only the $S_2$ correlation operator obtaining: $$\label{eq.1}
\begin{array}{l}
H_{eff}=e^{-iS_2}He^{iS_2}=\sum_{\alpha\beta}\langle\alpha|t|\beta\rangle a^{\dagger}_{\alpha}
a_{\beta}+
\sum_{\alpha\beta\gamma\delta}\langle\Psi_{\alpha\beta}|v^l_{12}|\Psi_{\gamma\delta}\rangle
a^{\dagger}_{\alpha}a^{\dagger}_{\beta} a_{\delta}a_{\gamma}\\
=\sum_{\alpha\beta}\langle\alpha|t|\beta\rangle a^{\dagger}_{\alpha}a_{\beta}+
\sum_{\alpha\beta\gamma\delta}\langle\Psi_{\alpha\beta}|v|\Psi_{\gamma\delta}\rangle a^{\dagger}_{\alpha}a^{\dagger}_{\beta} a_{\delta}a_{\gamma}
\end{array}$$ where $v_{12}^l$ is the long component of the two body interaction (note that the $v^l_{12}$ is in the following equations simply denoted as v). The $\Psi_{\alpha\beta}$ is the two particle correlated wave function: $$\label{eq.1a}
\Psi_{\alpha\beta}=e^{iS_2}\Phi_{\alpha\beta}$$
In dealing with complex nuclei however the ($S_i,~i=3\cdots n$) correlations should also be considered.
The evaluation of these diagrams is, due to the exponentially increasing number of terms, difficult in a perturbation theory.
We note however that one way to overcome this problem is to work with $e^{i(S_1+S_2+S_3+\cdots+S_n)}$ operator on the Slater’s determinant by keeping the n-body Hamiltonian unvaried.
After having performed the diagonalization of the n-body Hamilton’s operator we can calculate the form of the effective Hamiltonian which, by now, includes correlation operators of complex order.
Using Eq. (\[equ.1\]), we can compute the commutator of the Hamiltonian with the operator $(a^{\dagger}_{j_1}a^{\dagger}_{j_2})^J$ that creates a valence particle pair. By performing this calculation we shall retain the linear and the non-linear terms which are formed by coupling the shell model states with the particle-hole excitations of the core.
At this point in order to obtain a complete system of equations, we also have to calculate the commutator of the Hamiltonian (\[equ.1\]) with the non linear terms. With this calculation we introduce in the model space states which are formed by coupling the valence states with the two particle-hole excitations of the core.
The successive model equations are then formed by calculating the commutator with the operator obtained in the previous step.
The set of commutator equations above is suitable to be solved by means of a perturbation expansion. The perturbative solution of the system of commutator equations is however not easily obtainable due to the high number of diagrams one needs to calculate.
A non perturbative solution of the system of commutator equation can be obtained within the linearization method, which consists by applying the Wick’s theorem to the ((n+2)p-2h) terms and by neglecting the normal order terms.
This approximation is motivated by the consideration that the low lying spectra of nuclei the ((n+2)p-2h) terms are lying at much higher energy than that of the ((n+1)p-1h) states.
The linearized system of the commutator equations is then solved exactly in terms of the CFT which calculates the n-body matrix elements in an expedite and exact way.
In the following we give the basic formula of the method for a (nuclear) system with two valence particles. In second quantization, the two particle states are defined by: $$\label{e1}
\Phi_{2p}\longrightarrow \Phi_{j_1j_2}^J=A^{\dagger}_1(\alpha_1J)|0\rangle
= [a^{\dagger}_{j_1}a^{\dagger}_{j_2}]^J_M|0\rangle,$$ where, for the sake of simplicity, we have omitted the isospin quantum numbers and where $$\label{e3}
\begin{array}{c}
\alpha_1\leftrightarrow j_1j_2
\end{array}$$ has been introduced to ensure a compact index notation of the angular momenta of the two particles. In this notation, the operator product $a^{\dagger}_{j_1}a^{\dagger}_{j_2}$ just creates two *coupled* particles of single particle $j_1$ and $j_2$ *coupled* to the final J quantum number.
To derive the effect of the correlation on the two valence particles, we have, at this stage, to evaluate the next commutator $$\begin{aligned}
\label{e2}
& & \hspace*{-1.0cm}
[H,A^{\dagger}_1(\alpha_1J)]|0\rangle
\nonumber \\
& = &
[(\sum_{\alpha}\epsilon_{\alpha}a^{\dagger}_{\alpha}a_{\alpha}+\frac{1}{2}
\sum_{\alpha \beta \gamma \delta}
\langle\alpha\beta|v(r)|\gamma\delta\rangle
a^{\dagger}_{\alpha} a^{\dagger}_{\beta}
a_{\delta}a_{\gamma}),(a^{\dagger}_{j_1}a^{\dagger}_{j_2})^J]|0\rangle \end{aligned}$$ which, after some operator algebra becomes $$\begin{aligned}
\label{e4}
& & \hspace*{-1.0cm}
[H,A^{\dagger}_1(\alpha_1J)]|0\rangle
\nonumber \\
& = & \sum_{\beta_{1}} \Omega(2p|2p') A^{\dagger}_1(\beta_1J)]|0\rangle
+ \sum_{\beta_2J'_1J'_2} \Omega(2p|3p1h)
A^{\dagger}_2(\beta_2J'_1J'_2J)]|0 \rangle.\end{aligned}$$
In Eq. (\[e4\]) the $A^{\dagger}_1(\beta_1J)$ operators are those of Eq. (\[e1\]) and the $A^{\dagger}_2(\beta_2J'_1J'_2J)$ are defined below: $$\label{e5}
\begin{array}{l}
\Phi_{3p1h}\longrightarrow \Phi_{j'_1j'_2j'_3j'_4}^{J'_1J'_2J}
=A^{\dagger}_2(\beta_2J'_1J'_2J)|0\rangle
= ((a_{j'_1}^{\dagger}a_{j'_2}^{\dagger})^{J'_1}
(a_{j'_3}^{\dagger}a_{j'_4})^{J'_2})^J|0\rangle.
\end{array}$$ In Eq. (\[e5\]) we have used the additional convention: $$\label{e6}
\begin{array}{c}
\beta_2\longrightarrow j'_1j'_2j'_3j'_4
\end{array}$$ and we have associated: $$\label{e7}
\begin{array}{cc}
J'_1~$to the coupling of$~j'_1j'_2\\
J'_2~$to the coupling of$~j'_3j'_4
\end{array}$$ Having extended the commutator as in Eq. (\[e4\]), we have also to calculate the commutator equation for the $A^{\dagger}_2(\alpha_2J_1J_2J)$ operators as given below: $$\label{e8}
\begin{array}{l}
\displaystyle{
[H,A^{\dagger}_2(\alpha_2J_1J_2J)]|0\rangle} \\
\displaystyle{
= \sum_{\beta_{2}J'_1J'_2} \Omega(3p1h|3p'1h')
A^{\dagger}_2(\beta_2J'_1J'_2J)|0\rangle
+ \sum_{\beta_{3}J'_1J'_2J'_3} \Omega (3p1h|4p2h)
A^{\dagger}_3(\beta_3J'_1J'_2J'_3J)|0\rangle},
\end{array}$$ where we have introduced the (4p-2h) wave functions defined below: $$\label{e9}
\begin{array}{l}
\Phi_{4p2h} \longrightarrow \Phi_{j'_1j'_2j'_3j'_4j'_5j'_6}^{J'_1J'_2J'_{12}J'_3J}
=A^{\dagger}_3(\beta_3J'_1J'_2J'_3J)|0\rangle \\
= (((a_{j'_1}^{\dagger}a_{j'_2}^{\dagger})^{J'_1}
(a_{j'_3}^{\dagger}a_{j'_4})^{J'_2})^{J'_{12}}(a_{j'_5}^{\dagger}
a_{j'_6})^{J'_3})^J|0\rangle,
\end{array}$$ and where we have consistently extended the definition given in (\[e3\],\[e7\]): $$\label{e10}
\begin{array}{c}
\beta_3\longrightarrow j'_1j'_2j'_3j'_4j'_5j'_6
\end{array}$$ with: $$\label{e11}
\begin{array}{cc}
J'_1~$associated to the coupling of$~j'_1j'_2\\
J'_2~$associated to the coupling of$~j'_3j'_4\\
J'_3~$associated to the coupling of$~j'_5j'_6
\end{array}$$ In the definition of $A^{\dagger}_3(\beta_3J'_1J'_2J'_3J)$ the coupling of $J'_1$ to $J'_2$ to $J'_{12}$ has been discarded from the notation. In Eqs. (\[e4\],\[e8\]) the $\Omega$’s are the matrix elements of the Hamilton’s operator in the model wave functions. The next step would then be the computation of the commutator of the Hamiltonian with the $A^{\dagger}_3(\beta_3J'_1J'_2J'_3J)$ operators. Here we linearize these contributions by considering that in the study of the low energy spectrum and in the calculation of ground-state correlated distributions the $A^{\dagger}_3(\beta_3J'_1J'_2J'_3J)$ terms have a small contribution. The linearization is performed by applying to the (4p2h) terms: $$\label{e12}
\sum_{\alpha \beta \gamma \delta}
\langle\alpha\beta|v(r)|\gamma\delta\rangle a^{\dagger}_{\alpha}
a^{\dagger}_{\beta}a_{\delta}a_{\gamma}A^{\dagger}_3(\beta_3J'_1J'_2J'_3J)$$ the Wick’s theorem and to discard the normal order terms. Within this linearization approximation we generate non perturbative solutions of the EoM from the commutator equations of Eq. (\[e4\],\[e8\]), i.e.: the eigenvalue equations for the mixed mode system: $$\label{e13}
\begin{array}{l}
\displaystyle{
[H,A^{\dagger}_1(\alpha_1J)]|0\rangle
= \sum_{\beta_{1}} \Omega(2p|2p') A^{\dagger}_1(\beta_1J)|0\rangle } \\
\displaystyle{
+ \sum_{\beta_2J'_1J'_2} \Omega(2p|3p'1h')
A^{\dagger}_2(\beta_2J'_1J'_2J)|0\rangle },
\end{array}$$ and $$\label{e13a}
\begin{array}{l}
\displaystyle{
[H,A^{\dagger}_2(\alpha_2J_1J_2J)]|0\rangle
= \sum_{\beta_{1}} \Omega(3p1h|2p') A^{\dagger}_1(\beta_1J)|0\rangle } \\
\displaystyle{
+ \sum_{\beta_{2}J'_1J'_2} \Omega (3p1h|3p'1h')
A^{\dagger}_2(\beta_2J'_1J'_2)|0\rangle }.
\end{array}$$ Within the application of the GLA approximation we convert Eqs. (\[e4\], \[e8\]) in an eigenvalue equation for the configuration mixing wave functions (CMWFs) of the model. In fact, the linearization provides the additional matrix elements necessary to write the following identity: $$\label{e14}
\Omega(3p1h|3p'1h') = \langle j_1j_2j_3j_4|v(r)|j'_1j'_2j'_3j'_4\rangle,$$ and to introduce the off-diagonal matrix elements which couple the (2p) to the (3p1h) subspaces. Now, by writing Eqs. (\[e13\],\[e13a\]) in the following matrix form: $$\label{e14a}
\begin{array}{l}
\left ( \begin{array}{c}
\displaystyle{
[H,A^{\dagger}_1(\alpha_1J)]|0\rangle } \\
\displaystyle{
[H,A^{\dagger}_2(\alpha_2J_1J_2J)]|0\rangle} \end{array} \right ) \\
\displaystyle{
= \left ( \begin{array}{cc}
E_{2p}+\Omega(2p|2p')&\Omega(2p|3p'1h') \\
\Omega(3p1h|2p')&E_{3p1h}+\Omega(3p1h|3p'1h')\end{array} \right )
\left ( \begin{array}{c}
A^{\dagger}_1(\beta_1J)|0\rangle\\
A^{\dagger}_2(\beta_2J_1J_2J)|0\rangle \end{array} \right ) } ,
\end{array}$$ and by multiplying to the left with: $$\label{e14b}
\begin{array}{l}
\left ( \begin{array}{c}
\langle0|A_1(\alpha_1J) \\
\langle0|A_2(\alpha_2J_1J_2J) \end{array} \right )
\end{array}$$ we generate the eigenvalue equation for the dressed particles: $$\label{e15}
\begin{array}{l}
\displaystyle{\sum_{\beta_1\beta_2J'_1J'_2}
\left ( \begin{array}{cc}
E_{2p}+\langle A_1(\alpha_1J)|v(r)|A^{\dagger}_1(\beta_1J)\rangle &
\langle A_1(\alpha_1J) |v(r)| A^{\dagger}_2(\beta_2J'_1J'_2J)
\rangle \\
\langle A_2(\alpha_2J_1J_2J)|v(r)|A^{\dagger}_1(\beta_1J)\rangle &
E_{3p1h}+ \langle A_2(\alpha_2J_1J_2J) |v(r)|
A^{\dagger}_2(\beta_2J'_1J'_2J)\rangle
\end{array} \right)} \\
\cdot
\left ( \begin{array}{c}
\chi_1(\beta_1J) \\
\chi_2(\beta_2J'_1J'_2J)
\end{array}\right)
= E \left (
\begin{array}{c}
\chi_1(\alpha_1J)\\
\chi_2(\alpha_2J_1J_2J)\end{array}\right )|0\rangle.
\end{array}$$ In Eq.(\[e15\]) $E_{2p}=\epsilon^{HF}_{j_1}+\epsilon^{HF}_{j_2}$ and $E_{3p1h}=\epsilon^{HF}_{j_1}+\epsilon^{HF}_{j_2}+\epsilon^{HF}_{j_3}
-\epsilon^{HF}_{j_4}$ are the Hartre-Fock energies while the $\chi$’s are the projections of the model states: $$\label{e15a}
|\Phi^J_{2p}\rangle
= \chi_1(\alpha_1J) A_1(\alpha_1J)|0\rangle+\chi_2(\alpha_2J_1J_2J)
A_2(\alpha_2J_1J_2J)|0\rangle$$ to the basic vectors 2p, 3p1h. To conclude, although the (4p-2h) CMWFs are not active part of the model space, they are important for structure calculations. One may therefore associate the GLA approximation to a parameter which describes the degree of complexity of the model CMWFs. Within the first order linearization we obtain the EoM for the shell model while within the second and third order linearization approximations we derive the EoM of valence particles coexisting with the complex particle-hole structure of the excited states.
In this paper we solve Eq. (\[e15\]) self-consistently. The solutions for the first iteration step are obtained by diagonalizing the eigenvalue equation (\[e15\]). The first step of the iterative method generates the dynamic amplitudes for the two dressed particles, i.e. two particles coexisting with the 3p1h structures. With the calculated eigenvectors we recompute then the matrix elements $\langle j_1j_2|v(r)|j_1j_2j_3j_4\rangle$ and $\langle j_1j_2j_3j_4|v(r)|j'_1j'_2j'_3j'_4\rangle$ and we diagonalize again the eigenvalue equation. The iterations are repeated until the stabilization of the energies has been reached. Before performing the diagonalization of relative Hamilton’s operator in the CMWFs base we have to eliminate the spurious center of mass components. In the BDCM this is performed, following the calculations of Refs. \[26, 27\], by calculating the percent weights of spurious states in the model wave functions. These can be obtained by evaluating the energy of the center of mass according to the following equation: $$\label{c1}
\begin{array}{l}
\displaystyle{
E_R = \int dR\Psi^{\dagger dressed}(j_ij_jJ)(R^2)\Psi^{dressed}(j_ij_jJ)} \\
\displaystyle{+2\sum_{ij}
\int d\vec{r_i}d\vec{r_j}
\Psi^{\dagger dressed}(j_ij_jJ)(\vec{r_i}\cdot\vec{r_j})
\Psi^{dressed}(j'_ij'_jJ)}.
\end{array}$$ In Eq. (\[c1\]) the calculation of the integrals can be performed by using the CFT expansion for the (3p1h) states and by considering that for two particle states we have: $$\begin{array}{l}
\langle j_ij_jJ|(\vec{r_i}\cdot\vec{r_j})|j_ij_jJ\rangle\\
= \frac{4\pi}{3}[\hat{j_i} \hat{j_j}]
\left ( \begin{array} {ccc}
j_i & 1 &j_j\\
-\frac{1}{2}& 0 &\frac{1}{2}
\end{array}
\right )2
\left \{ \begin{array} {ccc}
j_i & j_j &J\\
j_i & j_j &1
\end{array}
\right \}
\langle l_i|r|l_j\rangle2,
\end{array}$$ where: $$\hat{j} = (2j+1).$$ By diagonalizing the above operator in the model space we obtain the energy of the center of mass. The overlap with the model space give the degree of “spuriosity” of the different components. The model space which characterize the BDCM is formed by adding even particle to a closed-shell nucleus. The closed shell configuration can be described by a single Slater determinant and one can use the Hartree-Fock’s theory to obtain the binding energy and the single-particle energies. Alternatively one can remark that for a closed shell nucleus (Z,N) the single particle energies for the states above the Fermi surface are related to the binding energies differences: $$\epsilon^>_p = BE(Z,N)-BE^*(Z+1,N),$$ and $$\epsilon^>_n = BE(Z,N)-BE^*(Z,N+1).$$ The single particle energies for the states below the Fermi surface are given by: $$\epsilon^<_p = BE^*(Z-1,N)-BE(Z,N),$$ and $$\epsilon^<_n = BE^*(Z,N-1)-BE(Z,N).$$ The BE are ground states binding energies which are taken as positive values, and $\epsilon$ will be negative for bound states. $(BE^*=BE-E_x)$ is the ground state binding energy minus the excitation energy of the excited states associated with the single particle states. Within this method, which recently has been reintroduced by B.A. Brown [@bro01], we derive the single particle energies from the known spectra of neighbor nuclei (see Table (1)).
\[b-I\]
--------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
hole $1s_{1/2}$
energy -20.58
hole/particle $1p_{3/2}$
energy 1.43
particle $1p_{1/2}$ $1d_{5/2}$ $2s_{1/2}$ $1d_{3/2}$ $1f_{7/2}$ $2p_{3/2}$ $1f_{5/2}$ $2p_{1/2}$
energy 1.73 17.21 22.23 23.69 25.23 27.18 28.33 29.67
--------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
: Single-particle scheme and single particle energies (MeV) used to form the model CMWFs for the A=6 isotopes
\[level-scheme\]
The generalization of the previously defined formalism needed to calculate the charge radii of the A=7 (one particle DCM), A=8 (four particles BDCM) and A=11 (three particles DCM) is not given explicitly in this paper, but will be presented shortly.
Results {#sec:2}
-------
In order to perform structure calculations, we have to define a single particle base with the relative single-particle energies and to choose the nuclear two-body interactions. The single-particle energies of these levels are taken from the known experimental level spectra of the neighboring nuclei and given in Table (1). For the experimentally unknown single particle energies of the fp shells we use the corresponding energies for the mass A=9 nuclei scaled accordingly the different binding energies. In this paper we perform as in \[12,13\] calculations by assuming all levels as bound for the particle-particle interaction, we use the G-matrix obtained from Yale potential [@sha67]. These matrix elements are evaluated by applying the $e^S$ correlation operator, truncated at the second order term of the expansion, to the harmonic oscillator base with size parameter b=1.76 fm. As also explained in Ref. \[4,5\] the potential used by the BDCM is separated in low and high momentum components. Therefore, the effective model matrix elements calculated within the present separation method and those calculated by Kuo \[30-33\] are pretty similar. The separation method generates matrix elements, which are almost independent from the radial shape of the different potentials generally used in structure calculations.\
The particle-hole matrix elements could be calculated from the particle-particle matrix elements via a re-coupling transformation. We prefer to use the phenomenological potential of Ref. \[34\]. The same size parameter as for the particle-particle matrix elements has been used. In Table (2) the calculated charge radii of $^6$He are compared with the radii calculated by the other theoretical models and with the radius obtained by the IS theory. In the theoretical models quoted in this table the calculations of the charge distributions and of the charge radii are performed in terms of non- correlated operators. The correlations are included only in the derivation of the $S_2$ effective Hamiltonian.
For the stable $^6$Li the calculated charge radius is equal to 2.55 fm, a value that reproduces well the charge radius of 2.55 fm obtained in Ref. \[23\] from the electron scattering experiments. The charge radii given in Table (3) for other lithium isotopes are however larger then those calculated in the other quoted theoretical models and then those obtained within the IS theory. Here also the main difference between the results obtained in the DCM and BDCM models and those of the other theoretical calculations has to be found in the treatment of the correlation operator. In \[9,10\] the charge radii are calculated in the “no core shell model” which is based on exact solutions of the two particles Schrödinger’s equation by considering large computational spaces. The calculations do not include however the $S_3$ correlations. \[12-13\] presents charge radii evaluated within an accurate Quantum Monte Carlo Method. The Hamiltonian used includes two- and three-bodies forces in a two-body correlated mechanics. The calculation method of \[36-38\] is based on a microscopic cluster method in which few particles are interacting with the rest nucleus considered in its ground state. In our model the excitations of the core are associated to the $S_3$ correlation operator which increases the charge radius of the Lithium isotopes.
\[b-II\] \[Radius6He\]
charge radius of $^6$He Model
-------------------------- ----------------------------------------------------
1.944 [@nav98] no-core shell model
2.09 [@pip01] quantum Monte Carlo technique
2.25 this work $\mathrm{BDCM}$
2.39 this work $\mathrm{BDCM}$ without the folded diagrams
2.06 this work $\mathrm{two~correlated}~1s_{\frac{1}{2}}$-protons
1.99 [@wur97] Cluster
1.99 [@fun94] Cluster
1.99 [@var94] Cluster
2.054 $\pm$ .014[@wan04] Isotopic Shift (Exp.)
: Calculated charge radii for $^6$He in fm compared with the results obtained in other theoretical models and with the radius derived within the IS theory.
Conclusions and Outlook {#conc}
=======================
In this contribution we have investigated the effect of the microscopic correlation operators on the charge distributions of $^6$He and of the Lithium isotopes. The microscopic correlation has been separated in short- and long-range correlations according the definition of Shakin \[2,3\]. The short-range correlation has been used to define the effective Hamiltonian of the model while the long-range is used to calculate the structures and the distributions of exotic nuclei. As given in the work of Shakin, only the two-body short-range correlation need to be considered in order to derive the effective Hamiltonian especially if the correlation is of very short range. For the long range correlation operator the three body component is important and should not be neglected. Within the three body correlation operator one introduces in the theory a three body interaction which compensates for the use of the genuine three body interaction of the no-core shell model.
By using generalized linearization approximations and cluster factorization coefficients we can perform expedite and exact calculations.
Within the calculated correlated distributions we obtain charge radii slightly larger than those calculated for non correlated distributions and derived by the IS experiments.
The application of the DCM [@tom05] to the two and three electron energies and distributions of the Helium and Lithium atoms respectively could serve as future motivation for a reevaluation of the IS theory. From one side, the calculation of CM of the different isotopes, could help to obtain a non-perturbative formulation of the isotopic change of the electron transition energies. From the other side the field shift theory could include the correct isotopic variation. Since the derivation of the charge radii from the two photon experiments is influenced by the precision of the theoretical calculations, this new proposed method could contribute to evaluate with better precision the charge radii of exotic nuclei.
\[b-III\]
Lithium Exp.(GSI) [@ewa05] Exp.+Theo. [@tan88] Theo. [@nav98] Theo. [@pip01] Theo. [@suz02] DCM+BDCM
----------- -------------------- --------------------- ---------------- ---------------- ---------------- ----------
rms rms rms rms rms
$^6$Li 2.51 2.47 2.22 2.54 - 2.55
$^7$Li 2.39 2.43 2.13 2.41 2.43 2.41
$^8$Li 2.30 2.42 2.13 2.26 2.34 2.40
$^9$Li 2.22 2.34 2.16 2.21 2.27 2.42
$^{11}$Li 2.47 3.01 - - 2.57 2.67
: Calculated charge radii for the Lithium isotopes in fm compared with the results obtained in other theoretical models and with the radius derived within the IS theory (Exp.).
[77]{}
Villars F., [*Proc. Enrico Fermi Int. School of Physics XXII*]{}, Academic Press N.Y. (1961).
Shakin C. M. and Waghmare Y. R., [*Phys. Rev. Lett.*]{} [**16**]{} (1966), 403. Shakin C. M., Waghmare Y. R. and Hull M. H., [*Phys. Rev.*]{} [**161**]{} (1967), 1006.
Tomaselli M., Liu L. C., Fritzsche S., K[ü]{}hl T. and Ursescu D., [*Nucl. Phys.*]{} [**A738**]{} (2004), 216. Tomaselli M., Liu L. C., Fritzsche S. and K[ü]{}hl T. [*J. Phys. G: Nucl. Part. Phys.*]{} [**30**]{} (2004), 999.
Br[ü]{}ckner K.A., [*The Many Body Problem*]{}, John Wiley and Sons, New York (1959).
Kuo T. T. S. and Osnes E., [*Folded-Diagrams Theory of the Effective Interaction in Atomic Nuclei*]{}, Springer Lecture Notes in Physics Vol. 366, Berlin (1991).
Tomaselli M., K[ü]{}hl T., Ursescu D. and Fritzsche S., [*Prog. Theor. Phys.*]{} [**116**]{} (2006), 699.
Navr$\acute{\mathrm{a}}$til P. and Barrett B. R., [*Phys. Rev.*]{} [**C57**]{} (1998), 3119.
Navr$\acute{\mathrm{a}}$til P. and Caurier E., [*Phys. Rev.*]{} [**C69**]{} (2004), 014331.
Stancu I., Barrett B. R., Navr$\acute{\mathrm{a}}$til P. and Vary J. P., [*Phys. Rev.*]{} [**C71**]{} (2005), 044325.
Pieper S. C. and Wiringa R. B., [*Ann. Rev. Part. Sci.*]{} [**51**]{} (2001), 53.
(2002), 044310.
Wang L.-B., M[ü]{}ller P., Bailey V. [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**93**]{} (2004), 142501.
W[ü]{}rzer J. and Hofmann H. M., [*Phys. Rev.*]{} [**C55**]{} (1997), 688.
Funada S., Kaneyama H. and Sakuragi Y., [*Nucl. Phys.*]{} [**A575**]{} (1994), 93.
Varga K., Suzuki Y. and Ohbayasi Y., [*Phys. Rev.*]{} [**C50**]{} (1994), 189.
Tomaselli M., K[ü]{}hl T., N[ö]{}rtersh[ä]{}user W., Ewald G., Sanchez R., Fritzsche S., and Karshenboim G. S., [*Can. J. Phys.*]{} [**80**]{} (2002), 1347.
Ewald G., N[ö]{}rtersh[ä]{}user W., Dax A., [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**93**]{} (2004), 113002. Bushaw B. A., N[ö]{}rtersh[ä]{}user W., Ewald G. [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**91**]{} (2003), 043004. S[á]{}nchez R., N[ö]{}rtersh[ä]{}user W., Ewald G., Albers D., Behr J., Bricault P., Bushaw B. A., Dax A., Dilling J., Dombsky M., Drake G. W. F., G[ö]{}tte S., Kirchner R., Kluge H.-J., K[ü]{}hl T., Lassen J., Levy C. D. P., Pearson M. R., Prime E. J., Ryjkov V., Wojtaszek A., Yan Z.-C., Zimmerman C. [*Phys. Rev. Lett.*]{} [**96**]{} (2005), 033002.
Yan Z.-C. and Drake G. W. F. [*Phys. Rev.*]{} [**A66**]{} (2002), 042504.
Li G. C., Sick I., Whitney R. R. and Yearian M. R., [*Nucl. Phys.*]{} [**A162**]{} (1971), 583.
Tomaselli M., [*Can. J. Phys.*]{} [**83**]{} (2005), 467.
Aufmuth P., [*J. Phys. B: At. Mol. Phys.*]{} [**15**]{} (1982), 3127.
Baranger E. and Lee C. W., [*Nucl. Phys.*]{} [**22**]{} (1961), 157.
Unna I. and Talmi I., [*Phys. Rev.*]{} [**112**]{} (1958), 452.
Brown B. A., [*Prog. Part. Nucl. Phys.*]{} [**47**]{} (2001), 524.
Shakin C. M., Waghmare Y. R., Tomaselli M. and Hull M. H., [*Phys. Rev.*]{} [**161**]{} (1967), 1015.
Jiang M. F., Machleidt R., Stout D. B. and Kuo T. T. S., [*Phys. Rev.*]{} [**C46**]{} (1992), 910. Lacombe M., Loiseau B., Richard J. M., Vinh Mau R., C${\hat{\mathrm o}}$t${\acute{\mathrm e}}$ J., Pir${\grave{\mathrm e}}$s P. and de Tourreil R., [*Phys. Rev.*]{} [**C21**]{} (1980) 861.
Machleidt R., [*Adv. Nucl. Phys.*]{} [**19**]{} (1989), 189.
Bogner S., Kuo T. T. S., Coraggio L., Covello A. and Itaco N., [*Phys. Rev.*]{} [**C65**]{} (2002), 051301(R).
Millener D.J. and Kurath D., [*Nucl. Phys.*]{} [**255**]{} (1975), 315.
Tanihata I., [*Phys. Lett.*]{} [**B206**]{} (1988), 592.
Suzuki Y., [*Progr. Theo. Phys. Supp.*]{} [**146**]{} (2002), 413.
Varga K., Suzuki Y. and Taniata I., [*Phys. Rev.*]{} [**C52**]{} (1995), 3013.
Varga K., Suzuki Y. and Lovas R. G., [*Phys. Rev.*]{} [**C66**]{} (2002), 041302.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the essential dimension and linkage properties of sets of Pfister forms and symbol $p$-algebras. We prove that the essential dimension of $m$-tuples of quadratic $n$-fold Pfister forms over fields of characteristic 2 is $m(n-1)+1$, and that the essential dimension of a pair of non-isometric anisotropic $n$-fold quadratic Pfister forms $\varphi_1$ and $\varphi_2$ is $n$ if and only if $\varphi_1=\psi \otimes \langle \! \langle \alpha \rangle \! \rangle $ and $\varphi_1=\psi \otimes \langle \! \langle \beta \rangle \! \rangle$ for some $\alpha,\beta \in F$ and $\psi \otimes \langle \! \langle \alpha,\beta \rangle \! \rangle=0$. We also prove that the essential 3-dimension of a pair of symbol $p$-algebras of degree 3 is 2 if and only if they become inseparably linked under a prime-to-3 extension, that three inseparably linked symbol $p$-algebras of degree $p$ over a $p$-special field $F$ of $\operatorname{char}(F)=p$ are also separably linked, and that there exists a quadruple of quaternion algebras of essential dimension 2 which are neither separably nor inseparably linked.'
address: 'Department of Computer Science, Tel-Hai Academic College, Upper Galilee, 12208 Israel'
author:
- Adam Chapman
bibliography:
- 'bibfile.bib'
title: 'Essential Dimension and Linkage of Sets of Pfister Forms and Symbol $p$-Algebras'
---
Quadratic Forms; Pfister Forms; Essential Dimension; Linkage; Symbol Algebras; $p$-Algebras; Kato-Milne Cohomology; Brauer Groups 11E04 (primary); 11E81, 16K20, 19D45 (secondary)
Introduction
============
We examine two functors: $\operatorname{PF}_{n_1,\dots,n_m} : \text{Fields}_k \rightarrow \text{Sets}$ sending each field $F$ containing $k$ to the set of $m$-tuples $(\varphi_1,\dots,\varphi_m)$ where each $\varphi_i$ is a quadratic $n_i$-fold Pfister form over a field $F$, and its subfunctor $\operatorname{LPF}_{n_1,\dots,n_m} : \text{Fields}_k \rightarrow \text{Sets}$ where all the forms in the $m$-tuple share a common separable slot, given an algebraically closed field $k$. The main results are that the essential dimension of $\operatorname{LPF}_{n_1,\dots,n_m}$ is $n_1+\dots+n_m-m+1$, and that a pair of $n$-fold Pfister forms $(\varphi_1,\varphi_2) \in \operatorname{PF}_{n,n}$ is of essential dimension $n$ if and only if $\varphi_1=\langle \! \langle \beta_1 \rangle \! \rangle \otimes \rho$ and $\varphi_2=\langle \! \langle \beta_2 \rangle \! \rangle \otimes \rho$ for a common $(n-1)$-fold factor $\rho$ and $\langle \! \langle \beta_1,\beta_2 \rangle \! \rangle \otimes \rho=0$. We also study the case of pairs of symbol $p$-algebras of degree 3.
The first result is motivated by the study of essential dimension of sums of $m$ symbols in $\operatorname{H}_2^n$ when $\operatorname{char}(k)=2$. It was bounded from above in [@McKinnie:2017] by $m (n-1)+1$ using the fact that the essential dimension of $(\mathbb{Z}/2 \mathbb{Z})^m$ over characteristic 2 fields is 1 (see [@Ledet:2004]), and from below by $m+n-1$ using an inductive argument. For $n_1=\dots=n_m=n$, each $m$-tuple $(\varphi_1,\dots,\varphi_m)$ in $\operatorname{LPF}_{n_1,\dots,n_m}(F)$ is mapped to the sum of symbols $[\varphi_1]+[\varphi_2]+\dots+[\varphi_n]$ in $\operatorname{H}_2^n(F)$, given that $\operatorname{H}_2^n(F)\cong I_q^n F/I_q^{n+1} F$ by [@Kato:1982], so insights about the essential dimension of $\operatorname{LPF}_{n_1,\dots,n_m}$ may shed light on the essential dimension of sums of $m$ symbols in $\operatorname{H}_2^n$, or at least give a reason to believe that $m(n-1)+1$ is indeed the precise value.
In addition to the above, we study two related problems: the first is proving that $(\varphi_1,\varphi_2,\varphi_3,\varphi_4) \in \operatorname{PF}_{2,2,2,2}(F)$ having essential dimension $2$ does not imply the forms are separably or inseparably 1-linked, and the second is proving that when $F$ is $p$-special and $\operatorname{char}(F)=p$, if three symbol $p$-algebras of degree $p$ over $F$ are inseparably linked then they are also separably linked.
Preliminaries
=============
Essential Dimension
-------------------
Given a field $k$ and a functor $\mathcal{F} : \operatorname{Fields}_k \rightarrow \operatorname{Sets}$, the essential dimension of an object $x\in \mathcal{F}(F)$, denoted $\operatorname{ed}_{\mathcal{F}}(x)$, is the minimum transcendence degree of any field $E$ with $k \subseteq E \subseteq F$ to which $x$ descends, i.e., there exists $y\in \mathcal{F}(E)$ such that $y_F=x$. The essential dimension of the functor, denoted $\operatorname{ed}(\mathcal{F})$, is the supremum on the essential dimension of all $x \in \mathcal{F}(F)$ for all $F \in \operatorname{Fields}_k$.
Quadratic Pfister Forms
-----------------------
The ring of Witt classes of symmetric bilinear forms is denoted by $W F$. Its fundamental ideal $I F$ of even-dimensional forms is generated by scalar multiples of bilinear 1-fold Pfister forms $\langle \! \langle \beta \rangle \! \rangle=\langle 1,-\beta \rangle$. Its powers $I^n F$ are generated by scalar multiples of bilinear $n$-fold Pfister forms $\langle \! \langle \beta_1,\dots,\beta_n \rangle \! \rangle=\langle \! \langle \beta_1 \rangle \! \rangle \otimes \dots \otimes \langle \! \langle \beta_n \rangle \! \rangle$. By [@Voevodsky:2003] and [@Kato:1982], $I^n F/I^{n+1} F \cong K_n F/2 K_n F$, where the latter stands for the $n$th Milnor $K$-group, which is the group of formal sums of symbols $\{\beta_1,\dots,\beta_n\}$ subject to the relations $$\{\beta_1,\dots,\beta_i,\dots,\beta_n\}+\{\beta_1,\dots,\beta_i',\dots,\beta_n\}=\{\beta_1,\dots,\beta_i\beta_i',\dots,\beta_n\}$$ $$\{\dots,\gamma,\dots,1-\gamma,\dots\}=0$$ $$\{\beta_1,\dots,\beta_i,\dots,\beta_n\}+\{\beta_1,\dots,\beta_i,\dots,\beta_n\}=0.$$ The isomorphism maps each symbol $\{\beta_1,\dots,\beta_n\}$ to $\langle \! \langle \beta_1,\dots,\beta_n\rangle \! \rangle$. When $\operatorname{char}(F) \neq 2$, symmetric bilinear forms correspond to quadratic forms, and we also have the isomorphism $K_n F/2 K_n F \cong H^n(F,\mu_2^{\otimes n})$ given by $\{\beta_1,\dots,\beta_n \} \mapsto (\beta_1) \cup \dots \cup (\beta_n)$.
The group of Witt classes of (even-dimensional) nonsingular quadratic forms over $F$ is denoted by $I_q F$ and is generated by quadratic 1-fold Pfister forms $\langle \! \langle \alpha ] \! ]=[1,\alpha]=u^2+uv+\alpha v^2$. When $\operatorname{char}(F)=2$, this group can be quite different form $I F$, but it is nonetheless a $W F$-module, and has an induced filtration $$(W_q F=)I_q F \supseteq I_q^2 F \supseteq I_q^3 F \supseteq \dots$$ given by $I_q^n F=I^{n-1} F \otimes I_q F$. The subgroup $I_q^n F$ is therefore generated by scalar multiples of $n$-fold Pfister forms $\langle \! \langle \beta_1,\dots,\beta_{n-1},\alpha ] \! ]=\langle \! \langle \beta_1,\dots,\beta_{n-1} \rangle \! \rangle \otimes \langle \! \langle \alpha ] \! ]$.
Recall that $\operatorname{PF}_{n_1,\dots,n_m}$ is the functor sending any field $F$ containing $k$ to the set of $m$-tuples $(\varphi_1,\dots,\varphi_m)$ where each $\varphi_i$ is an $n_i$-fold quadratic Pfister form over $F$. Write $\operatorname{BPF}_{n_1,\dots,n_m}$ for the analogous functor for bilinear forms (which coincides with the former when $\operatorname{char}(k) \neq 2$). We say that an $m$-tuple $(\varphi_1,\dots,\varphi_m) \in \operatorname{PF}_{n_1,\dots,n_m}(F)$ is separably $t$-linked if there exist a $t$-fold Pfister form $\rho$ and $(B_1,\dots,B_n) \in \operatorname{BPF}_{n_1-t,\dots,n_m-t}(F)$ such that $\varphi_i=B_i \otimes \rho$ for each $i \in \{1,\dots,m\}$. We say that $(\varphi_1,\dots,\varphi_m)$ is inseparably $t$-linked if there exist a bilinear $t$-fold Pfister form $B$ and $(\rho_1,\dots,\rho_m) \in \operatorname{PF}_{n_1-t,\dots,n_m-t}(F)$ such that $\varphi_i=B \otimes \rho_i$ for each $i \in \{1,\dots,m\}$. We denote by $\operatorname{LPF}_{n_1,\dots,n_m}$ the functor sending any field $F$ containing $k$ to the set of all separably 1-linked $m$-tuples $(\varphi_1,\dots,\varphi_m)$ in $\operatorname{PF}_{n_1,\dots,n_m}(F)$. We define $\operatorname{LBPF}_{n_1,\dots,n_m}$ in a similar way for bilinear forms.
Symbol $p$-Algebras
-------------------
When $\operatorname{char}(F)=p$, the group ${_pBr}(F)$ is generated by symbol $p$-algebras of degree $p$ by a result of Teichmüller, which are algebras of the form $$[\alpha,\beta)_{p,F}=F \langle x,y : x^p-x=\alpha,y^p=\beta, y x y^{-1}=x+1 \rangle$$ for some $\alpha \in F$ and $\beta \in F^\times$ (see [@GS Section 9]). Both these algebras and quadratic and symmetric bilinear forms over fields $F$ of $\operatorname{char}(F)=2$ are captured by Kato-Milne cohomology: let $\Omega_F^n$ be the additive group of $n$-fold absolute differential forms over $F$, and $\wp : \Omega_F^n \rightarrow \Omega_F^n/(d \Omega_F^{n-1})$ the map defined by $$\alpha \operatorname{dlog} \beta_1 \wedge \dots \wedge \operatorname{dlog} \beta_n \mapsto (\alpha^p-\alpha) \operatorname{dlog} \beta_1 \wedge \dots \wedge \operatorname{dlog} \beta_n.$$ Then $\operatorname{H}_p^{n+1}(F)$ is defined to be the cokernel of this map, and $\nu(n)_F$ is $\Omega_F^n/\ker(\wp)$. The group $\nu(n)_F$ is isomorphic to $K_n F/p K_n F$ by $ \operatorname{dlog} \beta_1 \wedge \dots \wedge \operatorname{dlog} \beta_n \mapsto \{\beta_1,\dots,\beta_n\}$, and for $p=2$ also to $I^n F/I^{n+1} F$ by $ \operatorname{dlog} \beta_1 \wedge \dots \wedge \operatorname{dlog} \beta_n \mapsto \langle \! \langle \beta_1,\dots,\beta_n \rangle \! \rangle$. The group $\operatorname{H}_p^{n+1}(F)$ is isomorphic to $I_q^{n+1}(F)/I_q^{n+2}(F)$ when $p=2$ by $\alpha \operatorname{dlog} \beta_1 \wedge \dots \wedge \operatorname{dlog} \beta_n \mapsto \langle \! \langle \beta_1,\dots,\beta_n,\alpha ] \! ]$, and to ${_pBr}(F)$ when $n=1$ by $\alpha \operatorname{dlog} \beta \mapsto [\alpha,\beta)_{p,F}$. We say that $m$ degree $p$ symbol $p$-algebras $A_1,\dots,A_m$ are cyclically linked if there exists a cyclic degree $p$ field extension $K=F[\wp^{-1}(\alpha)]=F[\lambda : \lambda^p-\lambda=\alpha]$ that embeds into all these algebras, in which case they can be written as $A_i=[\alpha,\beta_i)_{p,F}$ for some $\beta_i \in F^\times$. We say they inseparably linked if there exists a purely inseparable degree $p$ extension $K=F[\sqrt[p]{\beta}]$ of $F$ that embeds into all these algebras, in which case they can be written as $A_i=[\alpha_i,\beta)_{p,F}$ for some $\alpha_i \in F$.
Bilinear Pfister Forms
======================
Let $k$ be an algebraically closed field of arbitrary characteristic.
The essential dimension of both $\operatorname{BPF}_n$ and $\operatorname{PF}_n$ is $n$, and $\operatorname{PF}_n=\operatorname{LPF}_n$.
The facts $\operatorname{PF}_n=\operatorname{LPF}_n$ and $\operatorname{ed}(\operatorname{BPF}_n),\operatorname{ed}(\operatorname{PF}_n){\leqslant}n$ are trivial. In order to prove that $\operatorname{ed}(\operatorname{BPF}_n),\operatorname{ed}(\operatorname{PF}_n){\geqslant}n$ it is enough to note that if a bilinear or quadratic $n$-fold Pfister form descends to a field of transcendence degree $t$ over $k$ strictly smaller than $n$ then this field is a $C_t$-field, and in particular every quadratic form of dimension greater than $2^t$ is isotropic, which means that the $n$-Pfister form is isotropic. However, anisotropic bilinear and quadratic Pfister forms clearly exist, such as the generic ones, and so their essential dimension is at least $n$.
The essential dimension of $\operatorname{BPF}_{n_1,\dots,n_m}$ is $n_1+\dots+n_k$.
Clearly $\operatorname{ed}(\operatorname{BPF}_{n_1,\dots,n_m}) {\leqslant}n_1+\dots+n_m$. Take the generic $m$-tuple $(B_1,\dots,B_m)$ in $\operatorname{BPF}_{n_1,\dots,n_m}$, i.e. $F=k(\beta_{i,j} : 1 {\leqslant}i {\leqslant}m, 1{\leqslant}j {\leqslant}n_i)$ and $B_i=\langle \! \langle \beta_{i,1},\dots,\beta_{i,n_i} \rangle \! \rangle$ over $F$. Then the tensor product $B_1 \otimes \dots \otimes B_m$ is the generic element in $\operatorname{BPF}_{n_1+\dots+n_m}$, and so of essential dimension at least $n_1+\dots+n_m$.
Invariant for $m$-tuples of linked quadratic Pfister forms
==========================================================
\[useful\] Consider an anisotropic quadratic (or bilinear) $n$-fold Pfister form $\varphi$ with quadratic (bilinear, resp.) 1-fold factors $\rho_1$ and $\rho_2$. Then there exist a bilinear $(n-1)$-fold Pfister form $\psi$ such that $\varphi_1=\psi \otimes \rho_1=\psi \otimes \rho_2$.
When the forms are quadratic and $\operatorname{char}(F) =2$, write $\rho_1=\langle \! \langle \alpha ] \! ]$, $\rho_2=\langle \! \langle \gamma ] \! ]$ and $\varphi=\langle \! \langle \beta_1,\dots,\beta_{n-1},\alpha ] \! ]$. Let $(V,\omega)$ be the $(2^n-2)$-dimensional quadratic space satisfying $\omega \perp [1,\alpha]=\varphi$. Then there exist $\lambda \in F$ and $v\in V$ such that $\gamma=\alpha+\lambda^2+\lambda+\omega(v)$. If $\omega(v)=0$ then $\psi=\langle \! \langle \beta_1,\dots,\beta_{n-1} \rangle \! \rangle$. Otherwise, write $\beta_{n-1}'=\omega(v)$, and by [@ChapmanMcKinnie:2018 Proposition 5.2] there exist $\beta_1',\dots,\beta_{n-2}' \in F$ such that $\varphi=\langle \! \langle \beta_1',\dots,\beta_{n-1}',\alpha] \! ]$, but then we also have $\varphi=\langle \! \langle \beta_{1}',\dots,\beta_{n-1}',\alpha+\beta_{n_i}'+\lambda_i^2+\lambda_i] \! ]=\langle \! \langle \beta_{1}',\dots,\beta_{n_i-1}',\gamma] \! ]$, which means we can take $\varphi=\langle \! \langle \beta_1',\dots,\beta_{n-1}' \rangle \! \rangle$. When the forms are bilinear, write $\rho_1=\langle \! \langle \alpha \rangle \! \rangle$, $\rho_2=\langle \! \langle \gamma \rangle \! \rangle$ and $\varphi=\langle \! \langle \beta_1,\dots,\beta_{n-1},\alpha \rangle \! \rangle$. Since $\gamma \in D(\varphi')$, one can write $\gamma=c_{1} \beta_{1}+\dots+c_{n-1} \beta_{n-1}+c_{n} \alpha$ where each $c_{\ell} \in D(\langle \! \langle \beta_{1},\dots,\beta_{\ell-1} \rangle \! \rangle)$. Set $\beta_{\ell}'=c_{\ell} \beta_{\ell}$ when $c_{\ell} \neq 0$, and $\beta{\ell}'=\beta_{\ell}$ when $c_{i,\ell}=0$. Then take $\psi=\langle \! \langle \beta_1',\dots,\beta_{n-1}' \rangle \! \rangle$.
\[Inv\] Let $(\varphi_1,\dots,\varphi_{m}) \in \operatorname{LPF}_{n_1,\dots,n_m}(F)$ or ($\operatorname{LBPF}_{n_1,\dots,n_m}(F)$). Let $\rho$ be a common quadratic (bilinear, resp.) 1-fold factor of these forms and write $\varphi_i=\theta_i \otimes \rho$ for each $i \in \{1,\dots,m\}$. Then $\pi(\varphi_1,\dots,\varphi_m)=\theta_1 \otimes \dots \otimes \theta_m \otimes \rho \in \operatorname{PF}_{n_1+\dots+n_m-m+1}(F)$ (or $\operatorname{BPF}_{n_1+\dots+n_m-m+1}(F)$, resp.) is an invariant of the $m$-tuple.
Since $\pi(\varphi_1,\varphi_2,\dots,\varphi_m)=\pi(\pi(\varphi_1,\varphi_2,\dots,\varphi_{n-1}),\varphi_n)$, it suffices to prove the statement for $m=2$ and then it follows by induction. Suppose that $\varphi_1$ and $\varphi_2$ have another common quadratic (or bilinear) 1-fold factor $\phi$ and write $\varphi_1=\tau_1 \otimes \phi$ and $\varphi_2=\tau_2 \otimes \phi$. Then, by Lemma \[useful\] there exist $\psi_1$ and $\psi_2$ such that $\varphi_1=\psi_1 \otimes \rho=\psi_1 \otimes \phi$ and $\varphi_2=\psi_2 \otimes \rho=\varphi_2 \otimes \phi$. Therefore, $\pi(\varphi_1,\varphi_2)=\theta_1 \otimes \theta_2 \otimes \rho=\psi_1 \otimes \psi_2 \otimes \rho=\psi_1 \otimes \psi_2 \otimes \phi=\tau_1 \otimes \tau_2 \otimes \phi$.
The essential dimension of $\operatorname{LPF}_{n_1,\dots,n_m}$ and $\operatorname{LBPF}_{n_1,\dots,n_m}$ is $n_1+\dots+n_m-m+1$.
Clearly it is bounded from above by $n_1+\dots+n_m-m+1$. Consider the generic element in $\operatorname{LPF}_{n_1,\dots,n_m}$ (or $\operatorname{LBPF}_{n_1,\dots,n_m}$), i.e., the $m$-tuple $(\varphi_1,\dots,\varphi_m) \in \operatorname{LPF}_{n_1,\dots,n_m}(F)$ where $F=k(\alpha,\beta_{1,1},\dots,\beta_{n,n_m-1})$ and $\varphi_i=\langle \! \langle \beta_{i,1},\dots,\beta_{i,n_i-1},\alpha] \! ]$ or $\langle \! \langle \beta_{i,1},\dots,\beta_{i,n_i-1},\alpha \rangle \! \rangle$ for each $i \in \{1,\dots,m\}$. Then $\pi(\varphi_1,\dots,\varphi_m)$ is the generic element in $\operatorname{PF}_{n_1+\dots+n_m-m+1}$ (or $\operatorname{BPF}_{n_1+\dots+n_m-m+1}$), whose essential dimension is exactly $n_1+\dots+n_m-m+1$. Therefore, the essential dimension of $(\varphi_1,\dots,\varphi_m)$ is at least $n_1+\dots+n_m-m+1$.
Linkage by an $(n-1)$-fold Factor
=================================
Recall that when $\varphi_1$ and $\varphi_2$ are quadratic $n$-fold Pfister forms with a common quadratic $(n-1)$-fold factor $\rho$, one can write $\varphi_1=\langle \! \langle \beta_1 \rangle \! \rangle \otimes \rho$ and $\varphi_1=\langle \! \langle \beta_2 \rangle \! \rangle \otimes \rho$, and then $\Sigma_{\varphi_1,\varphi_2}$ defined to be the Witt class of $\varphi_1 \perp \varphi_2 \perp \varphi_3$, where $\varphi_3=\langle \! \langle \beta_1 \beta_2 \rangle \! \rangle \otimes \rho \equiv \varphi_1 \perp \varphi_2 \pmod{I_q^{n+1} F}$, is represented by $\langle \! \langle \beta_1,\beta_2 \rangle \! \rangle \otimes \rho \in \operatorname{PF}_{n+1}(F)$ (see [@ChapmanGilatVishne:2017]). This invariant can be similarly defined for symmetric bilinear forms. The main result of this section is that this invariant detects exactly when a pair of such forms has essential dimension $n$.
\[PropEss\] When $k$ is algebraically closed and $\operatorname{char}(k)=2$, if $(\varphi_1,\dots,\varphi_m) \in \operatorname{PF}_{n,\dots,n}(F)$ are anisotropic and inseparably $(n-1)$-linked then their essential dimension is $n$.
There exists $\alpha_1,\dots,\alpha_m,\beta_1,\dots,\beta_{n-1} \in F$ such that $\varphi_i=\langle \! \langle \beta_1,\dots,\beta_{n-1},\alpha_i ] \! ]$. Since $(\mathbb{Z}/2 \mathbb{Z})^{\times m}$ is of essential dimension 1, all the separable extensions $F[\lambda_i : \lambda_i^2+\lambda_i=\alpha_i]/F$ descend to separable extensions of one field $E_0$ of transcendence degree $1$ over $k$, and then the $m$-tuple $(\varphi_1,\dots,\varphi_m)$ descends to $E_0(\beta_1,\dots,\beta_{n-1})$ which is of transcendence degree $n$ over $k$.
The converse of this statement for arbitrary $m$ is false in general.
\[Exchar2\] By [@Chapman:2018], there exists an $m$-tuple $(\varphi_1,\dots,\varphi_m) \in \operatorname{PF}_{n,\dots,n}(F)$ for $m=2^n-1$ of anisotropic Pfister forms over the function field field $F=k(x_1,\dots,x_n)$ in $n$ algebraically independent variables over algebraically closed $k$ of $\operatorname{char}(k)=2$ which is not inseparably $1$-linked, and therefore not inseparably $(n-1)$-linked either.
\[Summary\] Two nonisomorphic anisotropic forms $(\varphi_1,\varphi_2) \in \operatorname{PF}_{n,n}(F)$ or $\operatorname{BPF}_{n,n}(F)$ which are $(n-1)$-linked satisfy $\Sigma_{\varphi_1,\varphi_2}=0$ if and only if $\varphi_1=\psi \otimes \langle \! \langle \epsilon \rangle \! \rangle$ and $\varphi_2=\psi \otimes \langle \! \langle \epsilon+1 \rangle \! \rangle$ for some quadratic (or bilinear) $(n-1)$-fold Pfister form $\psi$ and $\epsilon \in F$.
Write $\varphi_1=\psi \otimes \langle \! \langle \beta \rangle \! \rangle$ and $\varphi_1=\psi \otimes \langle \! \langle \gamma \rangle \! \rangle$. Then $\gamma=\psi(v)+\psi(w) \beta$ for some $v,w$ in the underlying vector space $V$ of $\psi$. Since $\varphi_2$ is anisotropic, $\psi(w) \neq 0$, and since $\varphi_1$ and $\varphi_2$ are nonisomorphic, $\psi(v) \neq 0$. Because of the group structure of $D(\psi)$, there exist $t,u \in V$ such $\psi(t)=\frac{1}{\psi(v)}$ and $\psi(u)=\frac{\psi(w)}{\psi(v)}$. Hence $\psi(t) \gamma=1+\psi(u) \beta$. Take $\epsilon=\psi(u) \beta$ and we are done.
The following conditions are equivalent for two given nonisomorphic anisotropic forms $(\varphi_1,\varphi_2) \in \operatorname{PF}_{n,n}(F)$ or $\operatorname{BPF}_{n,n}(F)$:
1. They are separably $(n-1)$-linked and $\Sigma_{\varphi_1,\varphi_2}=0$.
2. The essential dimension of $(\varphi_1,\varphi_2)$ in $\operatorname{PF}_{n,n}(F)$ is $n$.
And when they are in $\operatorname{PF}_{n,n}(F)$ and $\operatorname{char}(k)=2$, also
3. They are inseparably $(n-1)$-linked.
The fact that (1) and (3) are equivalent when they are in $\operatorname{PF}_{n,n}(F)$ and $\operatorname{char}(k)=2$ was proven in [@ChapmanGilatVishne:2017]. If the essential dimension of $(\varphi_1,\varphi_2)$ in $\operatorname{PF}_{n,n}$ or $\operatorname{BPF}_{n,n}$ is $n$ then they descend to a field $E$ of transcendence degree $n$ over $k$, so $E$ is a $C_n$-field, and in particular $u(E) {\leqslant}2^n$ (or $\operatorname{rank}_2(E) {\leqslant}n$). Therefore, every two quadratic (or bilinear) $n$-fold Pfister forms over $E$ are $(n-1)$-linked, and $\Sigma_{\varphi_1,\varphi_2}$ also descends to that field, but since it is represented by an $(n+1)$-fold Pfister, it is a trivial Witt class. This means $(2)$ implies $(1)$. The implication $(1) \Rightarrow (2)$ is a special case of Theorem \[Summary\].
\[threelinkedchar2\] For any algebraically closed field $k$ of $\operatorname{char}(k)=2$ and $n {\geqslant}3$, if the triple $(\varphi_1,\varphi_2,\varphi_3)$ in $\operatorname{PF}_{n,n,n}(F)$ is of essential dimension $n$ then the forms are separably $(n-1)$-linked. For $n=2$, if the triple $(\varphi_1,\varphi_2,\varphi_3)$ in $\operatorname{PF}_{n,n,n}(F)$ is of essential dimension $n$ then the forms are either separably or inseparably 1-linked.
If the essential dimension is $n$ then the forms descend to a $C_n$-field, over which every three such forms are separably $(n-1)$-linked by [@Chapman:2018]. For $n=2$, the forms are either separably or inseparably 1-linked by the same argument as in [@Chapman:2018 Proposition 4.7].
The difference between the case of $n=2$ and of $n{\geqslant}3$ is that in the former we do not know if inseparable 1-linkage for triples of quadratic 2-fold Pfister forms implies separable 1-linkage. A partial answer can be found in Section \[TripleLinkage\].
The forms described in Example \[Exchar2\] are not separably $(n-1)$-linked when $n {\geqslant}3$ either, which shows that Proposition \[threelinkedchar2\] cannot be extended to arbitrary $m$-tuples. In Section \[Examplen2\] we provide an example of a quadruple of quadratic 2-fold Pfister forms of essential dimension 2 which are neither separably nor inseparably 1-linked, showing that also for $n=2$, Proposition \[threelinkedchar2\] cannot be extended to arbitrary $m$-tuples.
For any algebraically closed field $k$, if the triple $(\varphi_1,\varphi_2,\varphi_3)$ in $\operatorname{BPF}_{n,n,n}(F)$ is of essential dimension $n$ then the forms are $(n-1)$-linked.
If the essential dimension is $n$ then the forms descend to a $C_n$-field, over which every three such forms are $(n-1)$-linked by [@Chapman:2018] and [@Becher].
This fact does not extend to $m$-tuples for $m{\geqslant}4$ in general.
In [@ChapmanTignol:2019] and [@Chapman:2018] examples of $m$-tuples $(\varphi_1,\dots,\varphi_m) \in \operatorname{BPF}_{n,\dots,n}(F)$ with $m=2^n$, $F=k(x_1,\dots,x_n)$ being the function field in $n$ algebraically independent variables over $k$ (of $\operatorname{char}(k)=0$ or 2) which are not $(n-1)$-linked were constructed.
Pairs of separably $(n-1)$-linked Pfister forms $(\varphi_1,\varphi_2) \in \operatorname{PF}_{n,n}(F)$ or $\operatorname{BPF}_{n,\dots,n}(F)$ with nontrivial $\Sigma_{\varphi_1,\varphi_2}$ can be easily constructed, e.g., $\varphi_1=\langle \! \langle \alpha_1,\beta_1,\dots,\beta_{n-1}] \! ]$ and $\varphi_2=\langle \! \langle \alpha_2,\beta_1,\dots,\beta_{n-1} ] \! ]$ or $\varphi_1=\langle \! \langle \alpha_1,\beta_1,\dots,\beta_{n-1} \rangle \! \rangle$ and $\varphi_2=\langle \! \langle \alpha_2,\beta_1,\dots,\beta_{n-1} \rangle \! \rangle$ over the function field $F=k(\alpha_1,\alpha_2,\beta_1,\dots,\beta_{n-1})$ in $n+1$ algebraically independent variables over $k$. Therefore, the essential dimension of the pair is not $n$ but rather $n+1$.
If $(\varphi_1,\varphi_2) \in \operatorname{PF}_{n,n}(F)$ or $\operatorname{BPF}_{n,n}(F)$ is of essential dimension $n+1$, does it necessarily mean they are separably $(n-1)$-linked? The answer is no:
For forms $(\varphi_1,\varphi_2) \in \operatorname{PF}_{n,n}(F)$ when $\operatorname{char}(k)=2$ one can take $\langle \! \langle \alpha_4,\dots,\alpha_n, \alpha_1,\alpha_2^{-1} ] \! ]$ and $\langle \! \langle \alpha_4,\dots,\alpha_n, \alpha_2,\alpha_3^{-1} ] \! ]$ over $F=k(\alpha_1,\dots,\alpha_n)$ which are not separably $(n-1)$-linked by [@Chapman:2019 Theorem 4.2]. For $(\varphi_1,\varphi_2) \in \operatorname{BPF}_{n,n}(F)$ and $\operatorname{char}(k)=2$, $\langle \! \langle \alpha_1,\alpha_2,\alpha_4,\dots,\alpha_n \rangle \! \rangle$ and $\langle \! \langle \alpha_2+1,\alpha_3,\alpha_4,\dots,\alpha_n \rangle \! \rangle$ which are not $(n-1)$-linked by [@Chapman:2019 Theorem 4.3]. For similar arguments to those in [@ChapmanTignol:2019], we see that also when $k=\mathbb{C}$, the same forms $\langle \! \langle \alpha_1,\alpha_2,\alpha_4,\dots,\alpha_n \rangle \! \rangle$ and $\langle \! \langle \alpha_2+1,\alpha_3,\alpha_4,\dots,\alpha_n \rangle \! \rangle$ are not $(n-1)$-linked either, which provides examples in characteristic not 2.
Symbol $p$-Algebras of Degree 3
===============================
Let $k$ be an algebraically closed field of $\operatorname{char}(k)=3$ and $\operatorname{Sym}_{3,3} : \operatorname{Fields}/k \rightarrow \operatorname{Sets}$ be the functor mapping any field $F$ containing $k$ to the set of pairs of symbol $p$-algebras of degree 3 over $F$. Recall that $\operatorname{ed}(A_1,A_2;3)$ is the minimal essential dimension of $\operatorname{ed}(A_1 \otimes L,A_2 \otimes L)$ where $L$ ranges over all prime-to-3 extensions $L/F$. Note that division algebras of degree 3 cannot descend to a field of transcendence degree less than 2, so $\operatorname{ed}(A_1,A_2) {\geqslant}2$ for any pair of division algebras $(A_1,A_2) \in \operatorname{Sym}_{3,3}(F)$.
If $F$ is a $C_2$-field and finitely generated over $k$, then it admits nontrivial quadratic extensions, which means it admits a normic quadratic form, and therefore also a normic quartic form. It is an easy consequence of the results of [@Lang:1952] that every two cyclic algebras of degree $3$ over $F$ share a common maximal subfield. This maximal subfield is either inseparable or separable. The separable common maximal subfield is either cyclic or becomes cyclic under a quadratic field extension.
Given an algebraically closed field $k$ of $\operatorname{char}(k)=3$, the following are equivalent for a pair $(A_1,A_2) \in \operatorname{Sym}_{3,3}(F)$ of division algebras of degree 3 over $F$ containing $k$:
- They become inseparably linked after a prime to 3 extension.
- $\operatorname{ed}(A_1,A_2 ; 3)=2$.
If $\operatorname{ed}(A_1,A_2 ; 3)=2$ then $A_1 \otimes L$ and $A_2 \otimes L$ descend to a field $E$ of transcendence degree $2$ over $k$. The resulting algebras share a common maximal subfield. This subfield is either inseparable or separable. If it is inseparable, then $A_1 \otimes L$ and $A_2 \otimes L$ are inseparably linked. Suppose it is separable. Then $A_1 \otimes K$ and $A_2 \otimes K$ are separably linked for $K$ which is either $E$ or a quadratic extension of $E$. The transcendence degree of $K$ over $k$ is 2, which means that $A_1 \otimes M$ and $A_2 \otimes M$ are inseparably linked for $M$ which is either $K$ or a quadratic extension of $K$ by [@Chapman:2020]. Consequently, $A_1 \otimes L \otimes M$ and $A_2 \otimes L \otimes M$ are inseparably linked.
In the opposite direction, if the algebras become inseparably linked under restriction to a prime-to-3 extension $L$ of $F$, then by [@Ledet:2004] the algebras $A_1 \otimes L$ and $A_2 \otimes L$ descend to a field of transcendence degree 2 over $k$, and so $\operatorname{ed}(A_1,A_2 ; 3)=2$.
For algebraically closed field $k$ of $\operatorname{char}(k)=3$, we have $\operatorname{ed}(\operatorname{Sym}_{3,3}) = 3$.
It cannot be greater than 3 because of [@Ledet:2004]. If it were 2 then every pair of cyclic algebras would become inseparably linked under a prime-to-3 extension. However, for $A_1=[\alpha_1,\beta_1)$ and $A_2=[\alpha_2,\beta_2)$ over $F=k(\alpha_1,\alpha_2,\beta_1,\beta_2)$, $A_1 \otimes A_2$ remains a division algebra under prime-to-3 extensions.
The last corollary can also be obtained from [@McKinnie:2017], but the argument here is simpler.
Triples of Nonlinked Quaternion Algebras of Essential Dimension 2 {#Examplen2}
=================================================================
Let $k$ be a field of $\operatorname{char}(k)=2$, $n$ an integer with $n{\geqslant}2$, $F_0=k(\alpha_1,\dots,\alpha_n)$ and $F=k(\!(\alpha_1)\!) \dots (\!(\alpha_n)\!)$. Both fields are endowed with the right-to-left $(\alpha_1,\dots,\alpha_n)$-adic valuation which we denote by $\mathfrak{v}$.
Consider the following set of quadratic $n$-fold Pfister forms taken from [@Chapman:2018]: write $I=\{0,1\}^{\times n}$, ${\bf 0}=(0,\dots,0)$, ${\bf d}=(d_1,\dots,d_n) \in I$ and ${\bf \alpha^d}=\prod_{i=1}^n \alpha_i^{-d_i}$. For each ${\bf d} \in I \setminus \{{\bf 0}\}$, write $\varphi_{\bf d}$ be $\langle \! \langle \alpha_1,\dots,\widehat{\alpha_\ell},\dots,\alpha_n \rangle \! \rangle \otimes \langle \! \langle {\bf \alpha^d}] \! ]$ where $\ell$ is the minimal integer in $\{1,\dots,n\}$ for which $d_\ell \neq 0$. We will show that the forms $\left\{\varphi_{\bf d} : {\bf d} \in I \setminus \{{\bf 0}\}\right\} \cup \{\phi \}$ do not have a common inseparable quadratic splitting field where $\phi=\langle \! \langle \alpha_1,\dots,\alpha_{n-1},\alpha_1^{-2} \alpha_2^{-1} \dots \alpha_n^{-1} ] \! ]$. These forms are defined both over $F_0$ and over $F$.
\[psidef\] Let $\varphi=\langle \! \langle \beta_1,\dots,\beta_{t-1},\gamma ] \! ]$ be a $t$-fold Pfister form over $F$. Assume the values $\beta_1,\dots,\beta_{t-1},\gamma$ modulo 2 are $\mathbb{F}_2$-independent and that $\mathfrak{v}(\gamma)<{\bf 0}$. In particular it means $\varphi$ is anisotropic (see [@ChapmanGilatVishne:2017 Lemma 10.1]). This form decomposes as $\varphi=\psi \perp [1,\gamma]$, and its underlying vector space decomposes accordingly as $V_\varphi=V_\psi \oplus F \oplus F$ where $\varphi(v \oplus a \oplus b)=\psi(v)+a^2+ab+b^2 \gamma$. The trace map $\tr : V_\varphi \rightarrow F$ is defined by mapping $v \oplus a \oplus b$ to $b$. Let $f_\varphi$ be the function $f_\varphi : V_\psi \oplus F \rightarrow F$ defined by $f_\varphi(v\oplus a)=\varphi(v\oplus a \oplus 1)$. The elements $\lambda$ that $f_\varphi$ represents correspond to the quadratic 1-fold Pfister factors $\langle \! \langle \lambda ] \! ]$ of $\varphi$. We define the extension of the valuation $\mathfrak{v}$ from $F$ to the quadratic space $(V_\varphi,\varphi)$ by $\mathfrak{v}(v)=\frac{1}{2} \mathfrak{v}(\varphi(v))$. Following [@GaribaldiPetersson:2011] and [@Tignol:1992], we define $w(\varphi)$ to be $\min \{ \mathfrak{v}(\tr(v))-\mathfrak{v}(v) : v \in V_\varphi\}$.
1. The value $2w(\varphi)$ is the negative of the maximal value of an element represented by $f_\varphi$ in the notation above.
2. In fact, $2w(\varphi)$ is the negative of the maximal value of an element represented by $h_\varphi$ where $h_\varphi: V_\psi \rightarrow F$ is defined by $h_\varphi(v)=f_\varphi(v \oplus 0)$.
If $\tr(v)=0$ then $\mathfrak{v}(\tr(v))=\infty$, and therefore the minimum of $\mathfrak{v}(\tr(v))-\mathfrak{v}(v)$ is not obtained for elements of trace zero. Now, since for every $c \in F^\times$, $\mathfrak{v}(\tr(cv))-\mathfrak{v}(cv)=\mathfrak{v}(c\tr(v))-\mathfrak{v}(cv)=\mathfrak{v}(\tr(v))+\mathfrak{v}(c)-\mathfrak{v}(v)-\mathfrak{v}(c)=\mathfrak{v}(\tr(v))-\mathfrak{v}(v)$, the minimum is obtained for an element of trace 1. The elements represented by $f_\varphi$ are $\varphi(v)$ of the elements $v$ of trace $1$, and so the first statement follows.
Since $f_\varphi(v \oplus a)=\psi(v)+a^2+a+\alpha$ and the values $\beta_1,\dots,\beta_{t-1},\alpha$ modulo 2 are $\mathbb{F}_2$-independent and that $\mathfrak{v}(\alpha)<{\bf 0}$, the value of $f_\varphi(v \oplus a)$ is minimum of $\mathfrak{v}(\psi(v))$, $\mathfrak{v}(a^2)$ and $\mathfrak{v}(\alpha)$, and the value of $h_\varphi(v)$ is the minimum of $\mathfrak{v}(\psi(v))$ and $\mathfrak{v}(\alpha)$, so $\mathfrak{v}(h_\varphi(v)) {\geqslant}\mathfrak{v}(f_\varphi(v \oplus a))$. Therefore, the maximal value of an element represented by $f_\varphi$ is the same as the maximal value of an element represented by $h_\varphi$.
1. Note that this definition applies for any $t$, including $t=n$ and $t=1$. Clearly, if $\rho$ is a quadratic $t$-fold Pfister factor of a quadratic $n$-fold Pfister form $\varphi$, then $w(\rho){\geqslant}w(\varphi)$.
2. For $t=1$, $\varphi$ is the norm form of the quadratic separable extension $K=F[\wp^{-1}(\alpha)]$ and for $t=2$, $\varphi$ is the norm form of the quaternion algebra $D=[\alpha,\beta)_{2,F}$, and in those cases, $w(\varphi)$ coincides with $w(K)$ or $W(D)$ as defined in [@Tignol:1992].
Every common quadratic 1-fold Pfister factor $\rho$ of the forms $\left\{\varphi_{\bf d} : {\bf d} \in I \setminus \{{\bf 0}\}\right\}$ has $w(\rho) {\leqslant}(\frac{1}{2},\dots,\frac{1}{2})$.
Let $\varphi=\varphi_{\bf d}$ for $d \in I \setminus \{ {\bf 0} \}$ and $\psi$ the $(2^n-2)$-dimensional form as in Definition \[psidef\] satisfying $\varphi=\psi \perp [1,{\bf \alpha^d}]$. The elements represented by $h_\varphi$ are ${\bf \alpha^d}+\psi(v)$ where $v\in V_\psi$. For each such element, its value is either $\mathfrak{v}({\bf \alpha^d})$ or $\mathfrak{v}(\psi(v))$, because the class of these two values modulo 2 are $\mathbb{F}_2$-independent. Let $\rho$ be a quadratic 1-fold Pfister factor of $\varphi$. Then there exist $v \in V_\psi$ for which $\mathfrak{v}({\bf \alpha^d}+\psi(v))=-2w(\rho)$. Assume further that $w(\rho)>(\frac{1}{2},\dots,\frac{1}{2})$. Then $\mathfrak{v}({\bf \alpha^d}+\psi(v))$ must be smaller than $(-1,-1,\dots,-1)$, which means that $\mathfrak{v}({\bf \alpha^d}+\psi(v))=\mathfrak{v}(\psi(v))=2\mathfrak{v}(v)$. Now, since $v$ is an element in $V_\psi$, the value $2\mathfrak{v}(v)$ modulo 2 must belong to $I \setminus \{{\bf 0},{\bf d}\}$. Since $2w(\rho)=-2\mathfrak{v}(v)$, the class of $2w(\rho)$ modulo 2 belongs to $I \setminus \{{\bf 0},{\bf d}\}$ too. Hence, there is no common quadratic 1-fold factor $\rho$ of all the forms in $\left\{\varphi_{\bf d} : {\bf d} \in I \setminus \{{\bf 0}\}\right\}$ with $w(\rho)>(\frac{1}{2},\dots,\frac{1}{2})$, because the classes modulo 2 of $2w(\rho)$ for the different $\varphi_{\bf d}$ disagree.
The forms $\left\{\varphi_{\bf d} : {\bf d} \in I \setminus \{{\bf 0}\}\right\} \cup \{\phi\}$ do not have a common quadratic 1-fold Pfister factor over $F$, and therefore not over $F_0$ either. Thus $I_q^n F$ is not $2^n$-linked.
By the previous lemma, all the common quadratic 1-fold Pfister factor $\rho$ of the forms in $\left\{\varphi_{\bf d} : {\bf d} \in I \setminus \{{\bf 0}\}\right\}$ must have $w(\rho) {\leqslant}(\frac{1}{2},\dots,\frac{1}{2})$. On the other hand, $w(\phi)=(1,\frac{1}{2},\dots,\frac{1}{2})>(\frac{1}{2},\dots,\frac{1}{2})$, so $\rho$ cannot be a factor of $\phi$.
The quaternion algebras $[\alpha_1^{-1},\alpha_2)_{2,F}$, $[\alpha_2^{-1},\alpha_1)_{2,F}$, $[\alpha_1^{-1} \alpha_2^{-1},\alpha_1)_{2,F}$ and $[\alpha_1^{-2} \alpha_2^{-1},\alpha_1)_{2,F}$ over $F=k(\!(\alpha_1)\!)(\!(\alpha_2)\!)$ (or $F_0=k(\alpha_1,\alpha_2)$) are neither separably nor inseparably linked.
They are not separably linked by the previous theorem, and they are not inseparably linked by [@Chapman:2018].
Triple Linkage of Symbol $p$-Algebras {#TripleLinkage}
=====================================
In [@Chapman:2015] it was proven that if two symbol $p$-algebras of degree $p$ over a field $F$ of $\operatorname{char}(F)=p$ are inseparably linked then they are also cyclically linked. Here we prove that the same holds for three algebras, under the assumption that $F$ is $p$-special, i.e. has no finite field extensions of degree prime to $p$.
Given a prime integer $p$ and a $p$-special field $F$ of $\operatorname{char}(F)=p$, if three symbol $p$-algebras of degree $p$ over $F$ are inseparably linked then they become cyclically linked under a prime-to-$p$ extension of $F$ of degree at most $2p-1$.
The algebras are inseparably linked, so they can be written as $A_1=[\alpha_1,\beta)_{p,F}$, $A_2=[\alpha_2,\beta)_{p,F}$ and $A_3=[\alpha_3,\beta)_{p,F}$. We can suppose that $\langle [A_1],[A_2],[A_3] \rangle$ is a subgroup of ${_pBr}(F)$ of order $p^3$, because otherwise it is enough to stress that two of them are cyclically linked (which is true by [@Chapman:2015]) and the third shares the same cyclic field extension of $F$ share by the first two. This means that $\alpha_1,\alpha_2,\alpha_3 \not \in \wp(F)=\{\lambda^p-\lambda : \lambda \in F\}$ and in particular they are nonzero. Moreover, $\alpha_2 \not \in \mathbb{F}_p \alpha_1$, and so $(\frac{\alpha_1}{\alpha_2})^{p-1} \neq 1$.
Now, $A_i$ can also be written as $[\alpha_i^p,\beta)_{p,F}$, and a solution to the following system will show the algebras are cyclically linked: $$\begin{aligned}
\alpha_1^p+\beta(u_1^p-u_1 v_1^{p-1}+\alpha_1^p v_1^p) & = & \alpha_2^p+\beta(u_2^p-u_2 v_2^{p-1}+\alpha_2^p v_2^p)\\
\alpha_1^p+\beta(u_1^p-u_1 v_1^{p-1}+\alpha_1^p v_1^p) & = & \alpha_3^p+\beta(u_3^p-u_3 v_3^{p-1}+\alpha_3^p v_3^p).\end{aligned}$$ Take $u_1=u_2=u_3$, and then the equations reduce to $$\begin{aligned}
\alpha_1^p+\beta(-u_1 v_1^{p-1}+\alpha_1^p v_1^p) & = & \alpha_2^p+\beta(-u_2 v_2^{p-1}+\alpha_2^p v_2^p)\\
\alpha_1^p+\beta(-u_1 v_1^{p-1}+\alpha_1^p v_1^p) & = & \alpha_3^p+\beta(-u_3 v_3^{p-1}+\alpha_3^p v_3^p).\end{aligned}$$ and then isolate $u_1$ in both equations $$\begin{aligned}
\beta(u_1 v_2^{p-1}-u_1 v_1^{p-1}) & = & \alpha_2^p-\alpha_1^p+\beta(\alpha_2^p v_2^p-\alpha_1^p v_1^p)\\
\beta(u_1 v_3^{p-1}-u_1 v_1^{p-1}) & = & \alpha_3^p-\alpha_1^p+\beta(\alpha_3^p v_3^p-\alpha_1^p v_1^p).\end{aligned}$$ Take $v_2=\frac{\alpha_1}{\alpha_2} v_1$ and $v_3=1$, and then $$\begin{aligned}
\beta\left(\left(\frac{\alpha_1}{\alpha_2}\right)^{p-1}-1\right) u_1 v_1^{p-1} & = & \alpha_2^p-\alpha_1^p\\
\beta(u_1-u_1 v_1^{p-1}) & = & \alpha_3^p-\alpha_1^p+\beta(\alpha_3^p-\alpha_1^p v_1^p).\end{aligned}$$ Plugging in the first equation in the second gives $$\beta u_1-\frac{\alpha_2^p-\alpha_1^p}{(\frac{\alpha_1}{\alpha_2})^{p-1}-1}=\alpha_3^p-\alpha_1^p+\beta(\alpha_3^p-\alpha_1^p v_1^p), \quad \text{and so}$$ $$\beta u_1=\alpha_3^p-\alpha_1^p+\frac{\alpha_2^p-\alpha_1^p}{(\frac{\alpha_1}{\alpha_2})^{p-1}-1}+\beta(\alpha_3^p-\alpha_1^p v_1^p).$$ Plugging this into the first equation gives $$(\alpha_3^p-\alpha_1^p+\frac{\alpha_2^p-\alpha_1^p}{(\frac{\alpha_1}{\alpha_2})^{p-1}-1}+\beta(\alpha_3^p-\alpha_1^p v_1^p)) \left(\left(\frac{\alpha_1}{\alpha_2}\right)^{p-1}-1 \right) v_1^{p-1}=\alpha_2^p-\alpha_1^p,$$ a degree $2p-1$ equation in one variable $v_1$, that must have a solution under some prime-to-$p$ extension of $F$ of degree up to $2p-1$.
1. For $p$-special fields, inseparable linkage of three symbol $p$-algebras of degree $p$ implies separable linkage.
2. When $p=2$, every three inseparably linked quadratic 2-fold Pfister forms over $F$ are either separably linked or become separably linked over a degree 3 extension of $F$.
The algebras $[\alpha^{-1},\beta_1)_{p,F}$, $[\alpha^{-1},\beta_2)_{p,F}$ and $[\alpha^{-1},\beta_3)_{p,F}$ over $F=k(\!(\alpha)\!)(\!(\beta_1)\!)(\!(\beta_2)\!)(\!(\beta_3)\!)$ (or over $F_0=k(\alpha,\beta_1,\beta_2,\beta_3)$) are pair-wise not inseparably linked by the same argument as in [@Chapman:2015 Example 4.2]. When $p=2$, these algebras remain pair-wise inseparably nonlinked under any odd field extension of $F$, because for each pair the associated invariant $\Sigma$ is $\langle \! \langle \beta_i,\beta_j,\alpha] \! ]$ which remains non-hyperbolic under any odd degree extension.
If $(\varphi_1,\varphi_2,\varphi_3) \in \operatorname{PF}_{2,2,2}(F)$ is of essential dimension 2, then the forms are either separably linked or become separably linked over a degree 3 extension of $F$.
By Proposition \[threelinkedchar2\] they are either separably linked or inseparably linked, and if they are separably linked then they are either separably linked or become separably linked over a degree 3 extension of $F$.
As a result, if $F$ is an odd closed $C_2$-field of $\operatorname{char}(F)=2$, such as the odd closure of $k(\alpha,\beta)$ or $k(\!(\alpha)\!)(\!(\beta)\!)$ where $k$ is algebraically closed, then every three quaternion algebras over $F$ are separably linked.
Acknowledgements {#acknowledgements .unnumbered}
================
The author thanks Jean-Pierre Tignol for his comments on the manuscript.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose and demonstrate a new read-out technique for a superconducting qubit by dispersively coupling it to a Josephson parametric oscillator. We employ a tunable quarter-wavelength superconducting resonator and modulate its resonant frequency at twice its value with an amplitude surpassing the threshold for parametric instability. We map the qubit states onto two distinct states of classical parametric oscillation: one oscillating state, with $185\pm15$ photons in the resonator, and one with zero oscillation amplitude. This high contrast obviates a following quantum-limited amplifier. We demonstrate proof-of-principle, single-shot readout performance, and present an error budget indicating that this method can surpass the fidelity threshold required for quantum computing.'
address: |
$^{1}$Microtechnology and Nanoscience, Chalmers University of Technology, Kemivägen 9, SE-41296, Gothenburg, Sweden\
$^{2}$Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA\
$^{3}$MIT Lincoln Laboratory, 244 Wood Street, Lexington, Massachusetts 02420, USA\
$^{4}$Institute of Quantum Computing, University of Waterloo, Waterloo, Ontario N2L 3G1, Canada
author:
- 'Philip Krantz$^{1}$'
- 'Andreas Bengtsson$^{1}$'
- 'Michaël Simoen$^{1}$'
- 'Simon Gustavsson$^{2}$'
- 'Vitaly Shumeiko$^{1}$'
- 'W. D. Oliver$^{2,3}$'
- 'C. M. Wilson$^{4}$'
- 'Per Delsing$^{1}$'
- 'Jonas Bylander$^{1}$'
title: |
Single-shot Readout of a Superconducting Qubit\
using a Josephson Parametric Oscillator
---
The readout scheme for quantum bits of information (qubits) constitutes one essential component of a quantum-information processor [@Divincenzo2000]. During the course of a quantum algorithm, qubit-state errors need to be corrected; in many implementations this is done by quantum error correction, where each operation is based on the outcomes of stabilizer measurements that indicate the qubit errors. The stabilizers must therefore be determined in a “single shot” – without averaging of the output signals of repeated measurements on identically prepared qubits – with fidelity exceeding approximately 99% [@Kelly2015].\
The commonly used measurement scheme for a superconducting qubit coupled to a linear microwave resonator does not, by itself, offer single-shot measurement performance. The qubit imparts a state-dependent (“dispersive") frequency shift on the resonator, which can be determined by applying a probe signal and measuring the reflected or transmitted signal, although only for weak probing, rendering an inadequate signal-to-noise ratio (SNR) [@Gambetta2006; @Boissonneault2009].\
Researchers have addressed the problem of insufficient SNR in essentially two ways. One approach is to feed the weak output signal into a following, parametric linear amplifier that adds only the minimum amount of noise allowed by quantum mechanics [@Vijay2011; @Lin2013; @Lin2014; @Bergeal2010]. Another approach is to insert a nonlinear element into the system and apply a strong drive tone, such that the resonator enters a bistable regime, hence enhancing the detection contrast [@Lupascu2007; @Mallet2009; @Murch2012; @Reed2010; @Boissonneault2010].\
In this paper we propose and demonstrate a simplified readout technique in which a superconducting qubit is directly integrated into a Josephson parametric oscillator (JPO). We map the qubit states onto the ground and excited states of the oscillator, and demonstrate proof-of-concept, single-shot readout performance (SNR $>1$). We obtain 81.5$\,\%$ qubit-state discrimination for a read-out time $\tau = 600\,$ns; however, from the error analysis we infer a read-out fidelity of $98.7\pm1.2\%$, taking into account known and reparable errors due to qubit initialisation and decoherence ($17.2\pm1.2\,\%$). A realistically achievable qubit-relaxation time, $T_1 = 50\,\mu$s, and a Purcell bandpass filter would reduce these errors from $17.2\,\%$ to $< 0.5\,\%$, as well as shorten the required read-out time to $\tau < 100\,$ns. The remaining errors, which are due to switching events in the oscillator ($1.2\pm0.3\,\%$), can be eliminated by improving the data-aquisition protocol - see Discussion and Supplementary Information. These qubit and detection improvements would bring the read-out fidelity to $\approx 99.5\,\%$.\
Our readout scheme relies on parametric pumping of a frequency-tunable resonator by modulation of its inductance. The pumping amplitude exceeds the threshold for parametric instability, the point above which the resonator oscillates spontaneously, even in the absence of an input probe signal. This instability threshold is controlled by the state of the qubit, whose ground and excited states correspond to the nonoscillating and oscillating states of the resonator, respectively. In our measurement, the oscillating state produces a steady-state resonator field corresponding to $185\pm15$ photons, whose output we can clearly distinguish from the nonoscillating state when followed by a commercial semiconductor amplifier, eliminating the need for a quantum-limited amplifier. Conceptually, this method can yield arbitrarily large contrast due to the parametric instability, and moreover, only requires a pump but no input signal.\
This readout scheme is well aligned with scalable, multi-qubit implementations. Parametric oscillators can be readily frequency-multiplexed [@Schmitt2014] and allow for a simplified experimental setup (compared to conventional microwave reflectometry) without a separate input port to the resonator or a following parametric amplifier, and consequently also without additional bulky microwave circulators that would normally route the input and parametric-pumping tones. It is also possible to manipulate the qubit via the flux-pumping line only, which further reduces the number of cables and interconnects.
![\[fig:SetupAndReadoutPrinciple\] Experimental setup and readout mechanism. **(a)** Schematic of the cryogenic microwave reflectometry setup. The transmon qubit (red) is capacitively coupled to the coplanar waveguide parametric resonator (blue). The input and output flows of photons are denoted $\left|B\right|^2$ and $\left|C\right|^2$, respectively, whereas the number of photons in the resonator is denoted $\left|A\right|^2$. The output signal is acquired using heterodyne detection of the amplified microwave signal. The components drawn in lighter gray are those that are rendered unnecessary by the JPO readout method, thereby offering a simplified experimental setup (see text). **(b)** Parametric-oscillation regions for the qubit ground state $\left|0\rangle\right.$ (solid blue line) and excited state $\left|1\rangle\right.$ (dashed blue line), respectively. These blue lines represent the instability boundaries, $\epsilon = \epsilon_{\mbox{\tiny{th}}}$, where the number of steady-state solutions to Eq. (\[eq:Eq1\]) changes — see Eq. (\[eq:ParametricInstabilityThreshold\]) in Methods. The two panels on the right are measured \[I,Q\]-quadrature voltage histograms of the device output for the pump bias point indicated by the circles, revealing two different oscillator states: **I.** Outside of the region of parametric oscillations, the resonator is “quiet" ($|A|^2 = 0$). **II.** Within the region, the resonator has two oscillating states ($|A|^2 > 0$), with a phase difference of $\pi$ radians — see further Fig. \[fig:Histograms\].](Figure-1-Krantz.pdf){width="1\linewidth"}
\[sec:Results\]Results {#secresultsresults .unnumbered}
======================
**The Josephson parametric oscillator (JPO).** Our device consists of a quarter-wavelength ($\lambda/4$), superconducting coplanar waveguide resonator, shorted to ground in one end via two parallel Josephson tunnel junctions (JJs) — see Fig. $\ref{fig:SetupAndReadoutPrinciple}$(a). The JJs form a superconducting quantum interference device (SQUID), which acts as a variable Josephson inductance, $L_J (\Phi,I(\phi)) = \Phi_0 / \left[ 2\pi \cos (\pi\Phi/\Phi_0) \sqrt{ I_0^2 - I^2(\phi)}\right]$, where $I_0$ is the critical current and $\Phi_0$ is the flux quantum. This inductance can be controlled by the external magnetic flux through the SQUID loop, $\Phi(t) = \Phi_{\mbox{\tiny{dc}}} + \Phi_{\mbox{\tiny{ac}}}(t)$, and by the superconducting phase difference across the JJs, $\phi(t)$, via its current–phase relation, $I(t) = I_0 \sin\phi(t)$.
Time-varying modulations of $\Phi$ and $\phi$ – “parametric pumping" – affect the resonator dynamics, albeit in rather different ways; moreover, the Josephson inductance is indeed both *parametric* and *nonlinear*. We explain these differences in the Discussion section below. The resonant frequency of the JPO is parametrically modulated via the magnetic flux, $\Phi(t)$, which can lead to frequency mixing as well as parametric effects such as noiseless amplification of a signal, frequency conversion, and instabilities [@Sandberg2008; @Palacios-Laloy2008; @Wilson2010; @Wilson2011; @Lin2013; @Krantz2013].
The state of the JPO has a rich dependence on several parameters, some of which was studied recently, both theoretically [@Dykman1998; @Wustmann2013] and experimentally [@Wilson2010; @Krantz2013; @Lin2014]. The equation of motion for the intra-resonator electric field amplitude, $A$, can be written as $$\label{eq:WS_equation}
i\dot{A} + \epsilon A^{*} + \delta A + \alpha \left| A\right|^2 A + i\Gamma A = \sqrt{2 \Gamma_0}B(t).
\label{eq:Eq1}$$
Here $\epsilon$ is proportional to the externally applied pump amplitude, $\Phi_{\mbox{\tiny{ac}}}$, which modulates the resonant frequency parametrically at close to twice its value, $\omega_p \approx 2 \omega_r$ (degenerate pumping), and $\delta = \omega_p/2 - \omega_r$ is the resonator’s detuning from half of the pump frequency. The field amplitude, $A$, and its complex conjugate, $A^{*}$, are slow variables in a frame rotating at $\omega_p/2$, and $|A|^2$ is the equivalent number of photons in the resonator. The Duffing parameter, $\alpha$, associated with a cubic field nonlinearity, arises from the nonlinear Josephson inductance. The linear damping rate has two components, $\Gamma = \Gamma_0 + \Gamma_R$, where $\Gamma_0/2\pi = 1.02\,$MHz is the external damping rate, associated with the photon decay through the coupling capacitor, and $\Gamma_R/2\pi = 0.30\,$MHz is the internal loss rate. The equation’s right-hand side represents the input probe signal, such that $\left|B(t)\right|^2$ has units of photons per second. The output flow of photons per second, $|C(t)|^2$, is given by $C(t) = B(t) - i\sqrt{2\Gamma_0}A$.\
For low pumping amplitude, below the parametric instability threshold, $\epsilon < \epsilon_{\mbox{\tiny{th}}}$, this device works as a phase-sensitive parametric amplifier (JPA) for an input $B(t)$ at signal frequency $\omega_s = \omega_p/2$ [@Yamamoto2008; @Sandberg2008; @Palacios-Laloy2008; @Wilson2010; @Lin2013]. Note, however, that we keep $B(t) \! = \! 0$ in the measurements reported here. For a pumping amplitude exceeding the threshold, $\epsilon > \epsilon_{\mbox{\tiny{th}}}$, spontaneous parametric oscillations set in — see Fig. \[fig:SetupAndReadoutPrinciple\](b) and Eq. (\[eq:ParametricInstabilityThreshold\]) in Methods. The resonator field builds up exponentially in time, even in the absence of an input probe signal until it becomes limited by the Duffing and pump-induced nonlinearities and reaches a steady state [@Wilson2010; @Krantz2013].\
We connected a transmon qubit capacitively to the resonator [@Koch2007] — see Fig. \[fig:SetupAndReadoutPrinciple\](a). The state of the JPO (oscillating or nonoscillating) can then be controlled by the qubit-state-dependent, dispersive frequency shift, $\chi$, which the qubit exerts on the resonator [@Blais2004; @Wallraff2004]. When the JPO is being pumped above the threshold for parametric oscillation, with amplitude $\epsilon$ and frequency detuning $\delta$, then a change of qubit state effectively pulls the resonator to a different value of the detuning, outside of the region of parametric oscillations — see Fig. \[fig:SetupAndReadoutPrinciple\](b). We denote the qubit-state dependent detunings by $\delta^{|0\rangle} = \delta - \chi$ and $\delta^{|1\rangle} = \delta + \chi$. The resulting mapping of the qubit state onto the average number of photons in the resonator provides us with a novel qubit-state readout mechanism, which we exploit in this work.\
![\[fig:SpectroscopyAndAvoidedCrossing\] Combined frequency spectrum obtained from qubit spectroscopy of the transmon qubit (in red) and through standard reflectometry of the resonator (in blue). The solid red and grey lines are fits. The dashed grey line, at resonator flux bias $F = 0.185\pi$, indicates the bias point at which we later demonstrate the readout method. **Inset:** Vacuum-Rabi splitting around the flux-bias point where the transmon frequency crosses that of the resonator. The minimum frequency splitting yields a qubit-resonator coupling $g_{01}/2\pi = 46\,$MHz.](Figure-2-Krantz.pdf){width="1\linewidth"}
**Characterisation of qubit and JPO.** The device and cryogenic experimental setup are depicted in Fig. \[fig:SetupAndReadoutPrinciple\](a). The sample is thermally anchored to the mixing chamber of a dilution refrigerator with a base temperature of 10$\,$mK. The parametric $\lambda/4$ resonator (in blue) is capacitively coupled to the transmission line ($C_c = 11.9$ fF), yielding an external quality factor $Q_{\mbox{\tiny{ext}}} = \omega_r/2\Gamma_0 = 2\,555$. A transmon qubit (in red) is also coupled near this end of the resonator.\
The resonator output signal is amplified using a 4–8 GHz high-electron-mobility transistor amplifier, with a noise temperature $T_N = 2.2\,$K, followed by two room-temperature amplifiers. We detect the outgoing signal using heterodyne mixing. The signal is first downconverted to a frequency $\left(\omega_{\mbox{\tiny{RF}}} - \omega_{\mbox{\tiny{LO}}}\right)/2\pi = 187.5\,$MHz; then the \[I,Q\]-quadrature voltages are sampled at 250$\,$MS$\,$s$^{-1}$, before they are digitally downsampled at a rate of 20$\,$MS$\,$s$^{-1}$.\
We first characterise the transmon spectroscopically — see Fig. \[fig:SpectroscopyAndAvoidedCrossing\] — from which we extract the Josephson and charging energies, $E_{J}/2\pi$ = 9.82$\,$GHz and $E_{C}/2\pi$ = 453$\,$MHz, respectively. From the vacuum-Rabi splitting, we extract a qubit$\--$resonator coupling rate $g_{01}/2\pi$ = 46$\,$MHz — see inset in Fig. \[fig:SpectroscopyAndAvoidedCrossing\].\
Next, we fit the frequency tuning curve of the resonator (with the qubit in the $\left|0\rangle\right.$-state) to the relation $$\omega_{r}^{|0\rangle}(F) = \omega_r(F) - g_{01}^2 / \Delta(F) ,
\label{eq:ResonatorTuningCurve}$$
where $F = \pi \Phi_{\mbox{\tiny{dc}}}/\Phi_0$ denotes the static flux bias, normalised to the magnetic flux quantum. The effective dispersive shift due to the qubit is $$\chi(F) = -\frac{g_{01}^2}{\Delta(F)}\left(\frac{E_{C}}{\Delta(F) - E_{C}}\right),
\label{eq:DispersiveShift}$$
which, in turn, depends on the qubit–resonator detuning, $\Delta(F) = \omega_a(F') - \omega_r(F)$, with $F' = F/8.88 + 0.58$ representing the effective magnetic flux of the transmon. Moreover, the qubit and resonator frequency spectra are well approximated by [@Koch2007; @Wallquist2006] $$\omega_a(F') \approx \sqrt{8E_{J}\left|\cos(F')\right|E_{C}} - E_{C},
\label{eq:QubitTuningCurve}$$ $$\omega_r(F) \approx \frac{\omega_{\lambda/4}}{1 + \gamma_0/\left|\cos(F) \right|},
\label{eq:ResonatorTuningCurve2}$$
![\[fig:PulseSequences\] Qubit readout by the Josephson parametric oscillator. **(a)** Pulse sequence: The qubit $\pi$-pulse (in red), with Gaussian edges and a plateau of duration $\tau_{\pi} = 52\,$ns, is followed by a short delay, $\tau_{d} = 20\,$ns, before the pump is turned on at time $t = 0$. **(b)** The solid blue and red traces show the inferred photon number, $|A|^2$, in the resonator, with and without a prior $\pi$-pulse on the qubit, respectively. Note that the resonator latches, once it has entered into the oscillating state, and remains there even if the qubit relaxes. The traces are the result of $10^4$ averages of the raw data; the inset shows a single instance of the raw data on the same time axis as the main plot. Prior to the sampling window of width $\tau_{s} = 300\,$ns, a delay $\tau_{r} = 300\,$ns is added to avoid recording the transient oscillator response. The hatched region around the average photon number represents our uncertainty, originating from the amplifier gain calibration — see Methods and Supplementary Fig. 3.](Figure-3-Krantz.pdf){width="1\linewidth"}
where $\omega_{\lambda/4}/2\pi = 5.55\,$GHz is the bare resonant frequency (in absence of the SQUID), and $\gamma_0 = L_{J}(F\!\!=\!\!0)/L_{\mbox{\tiny{r}}} = 5.3\pm0.1\,\%$ is the inductive participation ratio between the SQUID (at zero flux) and the resonator. The solid grey and red lines in Fig. \[fig:SpectroscopyAndAvoidedCrossing\] are fits to Eqs. (\[eq:ResonatorTuningCurve\]) and (\[eq:QubitTuningCurve\]), respectively.\
{width="1\linewidth"}
**Single-shot qubit readout.** We now demonstrate our method for reading out the qubit with the JPO. We choose a static-flux bias point $F = 0.185\,\pi$ for the resonator SQUID, corresponding to a resonant frequency $\omega_{r}^{|0\rangle}/2\pi = 5.218\,$GHz and qubit transition frequency $\omega_a/2\pi = 4.885\,$GHz — see dashed grey line in Fig. \[fig:SetupAndReadoutPrinciple\](a). Consequently, the qubit–resonator detuning is $\Delta/2\pi = -334\,$MHz, and the effective dispersive shift is $2\chi/2\pi = -7.258\,$MHz. We measured a Purcell-limited qubit relaxation time, $T_1 = 4.24\pm0.21\,\mu$s, and Ramsey free-induction decay time $T_{2}^{*} = 1.66\pm0.32\,\mu$s — see Methods, Supplementary Fig. 5, and Table 3.\
To operate the parametric oscillator as a high-fidelity qubit-readout device, we must be able to map the states of the qubit onto different states of the oscillator, which we must then clearly distinguish. We encode the qubit ground state $|0\rangle$ in the “quiet" state (the empty resonator) and the excited state $|1\rangle$ in the “populated" state of the resonator. Figure \[fig:PulseSequences\](a) shows the pulse sequence for qubit manipulation and readout, and Fig. \[fig:PulseSequences\](b) shows the resulting output from the JPO, operated with the pump settings $\delta^{|0\rangle}/\Gamma = -5.34, \, \epsilon/\Gamma = 3.56$.\
The populated oscillator in Fig. \[fig:PulseSequences\](b) contains 185$\pm15$ photons. We obtained this estimate from a comparison between the probe-amplitude dependence of the resonant frequency and the expected photon-number dependence of the Duffing shift — see Methods and Supplementary Fig. 3. This number of photons should be compared to $|A|^2$ = 200$\pm3$ photons, which is the solution to Eq. (\[eq:Eq1\]) in the steady state ($\dot{A}=0$).\
In order to achieve such clear qubit-state discrimination as in Fig. \[fig:PulseSequences\](b), we needed to make a judicious choice of flux-bias point, $F$, to mitigate the effects of two nonlinear shifts of the resonant frequency [@Krantz2013]. The Duffing shift dominates when $F \rightarrow \pm\pi/2$, whereas a pump-induced frequency shift dominates when $F \rightarrow 0$. These shifts can move the resonator away from the proper pump condition, thereby effectively restricting the output power — see Methods and Supplementary Fig. 2.\
Moreover, the qubit$\--$resonator detuning should be in the dispersive regime ($\Delta \gg g_{01}$), in which the qubit state controls the resonant frequency of the resonator. Yet it must yield a sufficiently large dispersive shift, $\chi > \Gamma$ (Eq. \[eq:DispersiveShift\]), to produce clearly distinguishable output levels, corresponding to the $|0\rangle$ and $|1\rangle$ states. For our chosen flux-bias point, we identify the optimal pump settings by mapping out the parametric oscillation region as a function of pump frequency and amplitude — see Fig. \[fig:ParametricRegions\](a).\
An interesting feature is present within the left half of Fig. \[fig:ParametricRegions\](a) (where the populated resonator encodes $| 1\rangle$). Here, when the qubit is initially in the $| 1\rangle$ state, the resonator latches into its oscillating state for as long as the pump is kept on, and does not transition into its quiet state when the qubit relaxes, as one might have expected. This latching is shown by the blue trace in Fig. \[fig:PulseSequences\](b). We attribute it to the existence of a tri-stable oscillation state [@Wustmann2013; @Wilson2010], associated with red detuning of the above-threshold region for the $|0\rangle$ state. When the qubit relaxes, there occurs an instantaneous shift of the pseudopotential for the amplitude $A$, from bi-stable (with two $\pi$-shifted, finite-amplitude states; see Fig. \[fig:SetupAndReadoutPrinciple\](b), panel II.) to tri-stable (with one additional zero-amplitude state). The field’s initial condition at the time of this shift, $A \neq 0$, causes the resonator to maintain its oscillating state. A separate study of this latching feature will be reported elsewhere. We evaluate the obtainable state discrimination by collecting quadrature-voltage histograms at every point within the two regions of parametric oscillations in the \[$\delta, \, \epsilon$\]-plane — see Fig. \[fig:ParametricRegions\](b). We choose the pump operation point $\delta^{|0\rangle}/\Gamma = -5.34, \, \epsilon/\Gamma = 3.56$, indicated by the black circle, and show the characterization in detail in Fig. \[fig:Histograms\]. In this point, the state discrimination has reached a plateau around 81.5$\%$. Each histogram in Fig. \[fig:Histograms\](a–b) contains in-phase ($V_{I}$) and quadrature ($V_{Q}$) voltage measurements from $10^5$ readout cycles, with each measurement being the mean quadrature voltage within the sampling time $\tau_s$ (blue window in Fig. \[fig:PulseSequences\]). We project each of the 2D-histograms onto its real axis, and thus construct 1D-histograms of the $V_I$ component — see Fig. \[fig:Histograms\](c). We can then extract a signal-to-noise ratio, SNR = $|\mu_{|1\rangle} - \mu_{|0\rangle}|/( \sigma_{|1\rangle} + \sigma_{|0\rangle})$ = 3.39, where $\mu$ and $\sigma$ denote the mean value and standard deviation, respectively, of the Gaussians used to fit the histograms. The peak separation of the histograms gives a confidence level of 99.998$\%$ for the readout fidelity. The peak appearing in the center of the blue trace arises mainly from qubit relaxation prior to and during the readout. We analyze this and other contributions in the next section, as well as in Supplementary Note 1 and Fig. 4.\
To extract the measurement fidelity from the histograms, we plot the cumulative distribution function of each of the two traces in Fig. \[fig:Histograms\](c), by summing up the histogram counts symmetrically from the center and outward, using a voltage threshold, $V_{\mbox{\tiny{th}}}$. From these sums, we obtain the S-curves of the probability to find the qubit in its ground state as a function of the voltage threshold value — see Fig. \[fig:Histograms\](d). We define the fidelity of the measurement as the maximum separation between the two S-curves.\
![\[fig:Histograms\] Quadrature voltage histograms of the parametric oscillator output, collected after digital sampling. The pump bias point was $\delta^{|0\rangle}/\Gamma = -5.34, \, \epsilon/\Gamma = 3.56$ — see Fig. \[fig:ParametricRegions\](b). In panel **(a)**, the qubit was in its ground state; in **(b)**, a $\pi$ pulse was applied prior to the readout pulse. **(c)** 1D-histograms of the in-phase voltage component, $V_{I}$, from the quadrature histograms in (a) and (b). The black and white solid lines are Gaussian fits, from which we extracted a signal-to-noise ratio of 3.39. **(d)** Cumulative distribution functions, corresponding to the $|0\rangle$ and $|1\rangle$ states, obtained by sweeping a threshold voltage, $V_{\mbox{\tiny{th}}}$, from the center of the two histograms ($V_{I} = 0$). The maximum separation between the two S-curves yields a state discrimination of 81.5$\,\%$.](Figure-5-Krantz.pdf){width="1\linewidth"}
Discussion {#discussion .unnumbered}
==========
To evaluate the fidelity of the readout itself, as compared to the fidelity loss associated with qubit errors, we now present an error budget. From the histograms in Fig. $\ref{fig:Histograms}$(c), we can account for 81.5$\,\%$ of the population, thus missing 18.5$\,\%$. To understand the remaining contributions, we run a Monte Carlo simulation of the qubit population, consisting of the same number of $10^5$ readout cycles as in the measured histograms. The simulation results are binned in the same way as the measurements, using the Gaussian fits as boundaries, and taking into account the following statistics: (i) qubit relaxation and preparation errors, (ii) thermal population of the qubit, (iii) spurious switching events by $\pi$ radians of the oscillator phase during readout (yielding a reduced sampled voltage), and (iv) peak-separation error due to the limited signal-to-noise ratio.\
We find that the main contribution to the loss of fidelity is due to qubit relaxation prior to and during the readout. From the measured relaxation time, $T_1 = 4.24\pm0.21\,\mu$s, we obtain a fidelity loss of $11.6\pm0.5\,\%$. However, this error can be reduced substantially (to $<\!0.5\,\%$) by introducing a Purcell bandpass filter [@Reed2010a; @Jeffrey2014; @Sete2015] at the output of the JPO; since the qubit is detuned from the JPO, this decreases its relaxation into the 50-$\Omega$ transmission line. Such a filter would allow us to increase the resonator damping rate, $\Gamma_0$, substantially reducing the readout time without compromising $T_1$. This is shown in Supplementary Note 2 and Table 4. Note, however, that an increased resonator damping rate yields an increased width of the parametric oscillation region: consequently, the qubit–resonator coupling, $g_{01}$, and detuning, $\Delta$, need to be chosen accordingly to render a sufficiently large dispersive frequency shift.\
From the simulation, we further attribute $4.5\pm0.3\,\%$ to qubit preparation errors. Another 1.1$\pm0.4\,\%$ can be explained from thermal population of the qubit; the effective qubit temperature is $T_q = 45\pm3\,$mK. By adding these fidelity loss contributions due to the qubit to the measured state discrimination, we can account for 81.5$\,\%$ + 11.6$\pm0.5\,\%$ + 4.5$\pm0.3\,\%$ + 1.1$\pm0.4\,\%$ = 98.7$\pm1.2\,\%$.\
There are also errors introduced by the parametric oscillator itself: Switchings between the $\pi$-shifted oscillating states reduce the overall measured voltage. We performed a separate control measurement that yielded $2.4\pm0.5\,\%$ switching probability, which translates into a maximal fidelity loss of half of that, $1.2\pm0.25\,\%$. The switching rate of the parametric oscillator depends on many parameters, including damping rates and bias points; this error can therefore, with careful engineering, be decreased even further. We could, however, eliminate the effect of phase-switching events by using a rectifying detection scheme, *e.g.*, a diode or a field-programmable gate array (FPGA), tracking the absolute value of the output field instead of its amplitude.\
The last and smallest contribution to the fidelity loss is the peak separation error, which accounts for the intrinsic overlap between the histograms. However, this contribution is $<\!0.002\%$ for our SNR of 3.39, and can therefore be neglected. For details on the error budget analysis, see Supplementary Note 1 and Fig. 4.\
By combining the above-mentioned improvements (reduced qubit relaxation rate, optimised qubit manipulations and cooling, enhanced resonator output coupling, and rectifying data acquisition), the read-out fidelity could realistically reach $\approx 99.5\,\%$, limited only by the qubit relaxation.\
Finally, we demonstrate that the relaxation time of our qubit is not measurably afflicted by the pump – see Methods and Supplementary Fig. 5. Our measurement scheme is, in principle, quantum nondemolition (QND), see Supplementary Note 3; however, a proper experimental and theoretical assessment of the back-action is outside the scope of this work.\
Table \[tab:OperationSchemes\] puts our results in the context of previous work on parametric and nonlinear Josephson amplification and detection circuits.\
A flux-pumped, parametric phase-locked oscillator (PPLO) was used as a following amplifier, also enabling sensitive qubit readout [@Lin2014]. In our work, the qubit was directly coupled to the JPO, which simplifies the experimental setup by reducing the number of microwave components needed. Also, with a pumping amplitude below the parametric instability threshold, the flux-pumped JPA has been used to read out one qubit [@Lin2013], as well as multiple qubits coupled to the same bus resonator [@Jeffrey2014].\
There is another way of operating our device: instead of pumping the flux at $\omega_p \approx 2\omega_r$, we can apply an alternating pump current ($\epsilon \! = \! 0, \, B(t) \! \neq \! 0$), now at a frequency close to resonance, $\omega_p \approx \omega_r$, and thereby directly modulate the phase difference, $\phi$. Both methods can provide linear parametric gain upon reflection of a detuned signal ($\omega_s\neq\omega_p/2$ and $\omega_s\neq\omega_p$, respectively). The flux-pumped JPA has a very wide frequency separation between pump tone and signal, because $\omega_s \approx \omega_r \approx \omega_p/2$, which is a practical advantage since it makes the resonator’s entire instantaneous bandwidth available for amplification with no need to suppress or filter out the pump tone. Moreover, the $\lambda/4$ resonator has no mode in the vicinity of $\omega_p$ that the pump might otherwise populate.\
We emphasize that there are indeed two different physical mechanisms in play, since flux and current pumping address orthogonal variables in the sense that $\Phi = (\varphi_1 - \varphi_2)\Phi_0/2\pi$ and $\phi = (\varphi_1 + \varphi_2)/2$, where $\varphi_1$ and $\varphi_2$ denote the gauge-invariant phase differences across the two parallel JJs. This distinction is also evident in Eq. (\[eq:WS\_equation\]). The parametric flux-pumping term, $\epsilon A^*$, modulates the resonant frequency; it couples the resonator field amplitude and its complex conjugate, which can provide quadrature squeezing of an input signal and enables phase-sensitive parametric amplification; and for stronger modulation there is a parametric instability threshold into the JPO regime – see Fig. \[fig:SetupAndReadoutPrinciple\](b).\
Current pumping by an input $B(t)$, on the other hand, corresponds to an external force which directly contributes to the intra-resonator field $A$ and drives its nonlinear term $\alpha |A|^2$. For zero detuning, $\omega_s=\omega_p$, this is the driven Duffing oscillator which has no gain (it offers no phase-sensitive amplification); for stronger driving there occurs a dynamical bifurcation but no internal instability or parametric oscillations.\
Current-pumping with a moderate amplitude is used for linear amplification with the JPA [@Yurke1988; @Castellanos-Beltran2008], which enabled, *e.g.*, the observation of quantum jumps in a qubit [@Vijay2011]. Current modulation is also used in the latching detection scheme of the Josephson bifurcation amplifier (JBA) [@Siddiqi2004; @Siddiqi2005; @Lupascu2007; @Mallet2009; @Schmitt2014]. There, a higher-amplitude input strongly drives the Duffing nonlinearity near its bifurcation point; the two qubit states can then be mapped onto two different resonator output field amplitudes. The JBA was used for quantum non-demolition measurement of a qubit, and in a lumped-element resonator [@Murch2012], in which a qubit-state sensitive autoresonance was observed in response to a frequency-chirped current drive. Yet another method is to couple the qubit to a linear resonator, which inherits a cross-Kerr nonlinearity from the qubit; current pumping of the resonator can then yield a strong output signal that depends on the qubit state [@Reed2010; @Boissonneault2010].\
----------------------- --------------------------------------- ---------- ---------- ------------ ----------------
Device $\epsilon$ $B_s$ $B_p$ $\#$ modes Ref.
\[0.4ex\]
\[-2ex\] JPO$^{(*)}$ $>\epsilon_{\mbox{\tiny{th}}}$ 0 0 1 This work
\[0.4ex\] JPA $\lesssim\epsilon_{\mbox{\tiny{th}}}$ $\neq 0$ 0 1 [@Lin2013]
\[0.4ex\] JPA $\lesssim\epsilon_{\mbox{\tiny{th}}}$ $\neq 0$ 0 multimode [@Simoen2015]
\[0.4ex\]
\[-2ex\] PPLO $>\epsilon_{\mbox{\tiny{th}}}$ $\neq 0$ $\neq 0$ 1 [@Lin2014]
\[0.4ex\]
\[-2ex\] JPA 0 $\neq 0$ $\neq 0$ 1 [@Vijay2011]
\[0.4ex\] JBA$^{(*)}$ 0 0 $\neq 0$ 1 [@Mallet2009]
\[0.4ex\] JPC 0 $\neq 0$ $\neq 0$ 2 [@Bergeal2010]
\[0.4ex\]
----------------------- --------------------------------------- ---------- ---------- ------------ ----------------
: \[tab:OperationSchemes\] Overview of different modes of operation for the various Josephson amplification and detection schemes. The variables refer to Eq. (\[eq:WS\_equation\]), where $\epsilon$ denotes the flux-pumping amplitude (at $\omega_p \approx 2\omega_r$), and $B_s$ and $B_p$ denote alternating-current signal and pump amplitudes, respectively (at $\omega_p \approx \omega_r$). The two readout methods marked with an asterisk ($^{*}$) have the qubit directly integrated with the detector, whereas the other devices are used as following amplifiers.
In conclusion, we have introduced a single-shot readout technique for superconducting qubits $\--$ the Josephson parametric oscillator (JPO) readout. We demonstrated proof-of-principle operation, obtaining a bare state discrimination of 81.5%. After correcting for known and reparable errors, this translates into an inferred readout fidelity of $98.7\pm1.2\,\%$, which by implementing a rectifying detection scheme can be further increased by $1.2\pm0.3\,\%$. With foreseeable improvements and optimization, this device would be an attractive candidate for implementing multi-qubit readout in the context of scalable error correction schemes. This fidelity and the readout time are both amenable to optimization.\
Our system integrates a parametric readout mechanism into the resonator to which the qubit is coupled, substantially reducing the number of components needed to perform single-shot readout in a circuit quantum electrodynamics architecture. Advantages offered by this readout technique include the potential for multiplexing and scalability with no need for signal-probe inputs, additional microwave circulators, or separate parametric amplifiers. As opposed to other integrated readout devices, our pump frequency is far outside of the resonator band and can thus easily be spectrally separated from other transition frequencies in the system.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors would like to thank Jared Cole, Göran Johansson, and Baladitya Suri for fruitful discussions. All devices were fabricated in the Nanofabrication Laboratory at MC2, Chalmers. Support came from the Wallenberg foundation, the European Research Council (ERC), the Royal Swedish Academy of Science (KVA), the European project ScaleQIT, STINT, and Marie Curie CIG. The MIT and Lincoln Laboratory portions of this work were sponsored by the Assistant Secretary of Defense for Research & Engineering under Air Force Contract \#FA8721-05-C-0002. Opinions, interpretations, conclusions and recommendations are those of the author and are not necessarily endorsed by the United States Government.
Author contributions {#author-contributions .unnumbered}
====================
P. K., A. B., M. S., C. M. W., P. D., and J.B. designed the experimental setup. P. K. modeled and fabricated the device. P. K., A. B., S. G., W. D. O., and J. B. carried out the measurements. V. S. gave input on theoretical matters. P. K., P. D., and J. B. wrote the manuscript with input from all co-authors.
Methods {#methods .unnumbered}
=======
**Device fabrication.** We fabricated our device on sapphire, using niobium for the waveguides and the transmon paddles, and shadow-evaporated aluminum for the Josephson junctions. To reduce the surface roughness prior to processing, the 2” c-plane sapphire wafer was pre-annealed at 1100$^{\circ}$C for 10 h in an atmosphere of N$_2$:O$_2$, 4:1, ramping the temperature by 5$^{\circ}$C/min. The annealed wafer was then sputtered with 80 nm of Nb in a near-UHV magnetron sputter. The first patterning of the sample consists of a photolithography step to define alignment marks and bond pads, deposited using electron beam evaporation of 3 nm Ti and 80 nm Au. Next, the resonator, the transmon islands, and the pump line were defined in the Nb layer using a standard e-beam lithography process at 100 keV, and etched using inductively coupled plasma reactive ion etching (ICP RIE) in NF$_3$-gas.\
The Al/AlO$_{\small{\mbox{x}}}$/Al Josephson junctions forming the SQUIDs, used to terminate the resonator and for connecting the transmon islands, were then defined in a second e-beam step. After exposure, the 2”-wafer was diced into separate chips, using the exposed e-beam resist as a protective resist. Prior to the first evaporation step, the surfaces of the Nb films where cleaned using in-situ Ar-ion milling inside of the Plassys evaporator. However, due to the substantially different regimes of critical currents, $I_0$, required for the Josephson junction of the transmons and the parametric resonator, two sequential evaporations and oxidations were performed within the same vacuum cycle by rotating a planetary aperture mounted inside the evaporator load-lock, effectively shielding one half of the sample at the time. Finally, a post-deposition ashing step was done to clean the surfaces from organic residues.\
**Finding the parametric oscillation threshold.** It is hard to experimentally find the parametric oscillation threshold with good precision, when only considering the parametric oscillation region, Fig. \[fig:ParametricRegions\](a), whose observed shape gets smeared by the amplified vacuum noise. In this section we present an alternative method using a weak probe signal: we probe the parametrically amplified response as we sweep the pump amplitude across the instability threshold.\
We apply a probe signal on resonance, $\omega_s = \omega_{r}^{|0\rangle}$, while applying a detuned pump signal, such that $(\omega_{p} - 2\omega_{s})/2\pi = 100\,$kHz. The signal then undergoes degenerate, phase-preserving parametric amplification (red trace in Supplementary Fig. 1), while the parametric oscillations are cancelled out since we measure the average amplitude of the field. The parametric amplification has maximum gain just at the threshold. We plot the magnitude of the reflected signal as a function of the pump power (at the generator), yielding an oscillation threshold $P_{\mbox{\tiny{th}}} = -10.8\,$dBm, as indicated by the dashed red line. As a comparison, we measure the output power of parametric oscillation (PO), for $\omega_{p} - 2\omega_{r} = 0$ and $B(t) = 0$ — see the blue trace.\
**Limits of the parametric oscillation amplitude.** As briefly discussed in the main text, there are two nonlinear effects that move the resonator away from its pump condition, by means of their associated frequency shifts [@Krantz2013], $$\Delta \omega = - \alpha |A|^2 - \beta \Gamma (\epsilon/\Gamma)^2.
\label{eq:DeltaOmegaShifts}$$
The Duffing shift dominates near flux bias $F = \pm \pi/2$; the Duffing parameter is approximated as $$\alpha(F) \approx \frac{\pi^2 \omega_{\lambda/4}Z_0}{R_K}\left( \frac{\gamma_0}{\cos(F)}\right)^3 = \alpha_0 \left( \frac{\gamma_0}{\cos(F)}\right)^3,
\label{eq:Alpha}$$
where $Z_0 = 50\,\Omega$ is the resonator’s characteristic impedance and $R_K = h/e^2$ is the quantum resistance.\
The pump-induced frequency shift dominates near $F = 0$; it is approximated as $$\beta(F) \approx \frac{\Gamma}{\omega_{\lambda/4}\gamma_0}\frac{\cos^3(F)}{\sin^2(F)} = \beta_0\frac{\cos^3(F)}{\sin^2(F)}.
\label{eq:Beta}$$
The resonator’s frequency tuning vs. $F$, Eq. (5) in the main text, is shown in Supplementary Fig. 2(a), for the parameters of our device, and Eqs. (\[eq:Alpha\]) and (\[eq:Beta\]) are plotted in Supplementary Fig. 2(b). This figure illustrates that it is essential to bias the system far enough away from the limiting points, $F = 0$ or $\pi/2$, such that neither frequency shift pulls the resonator too far from its pump condition, thereby severely limiting the attainable output power.\
The steady state solution of Eq. (1) in the main text yields an analytic expression for the expected number of photons within the region of parametric oscillations, $$\left|A\right|^2 = \frac{\Gamma}{\alpha}\left( \sqrt{\left(\frac{\epsilon}{\Gamma}\right)^2 -1} - \frac{\delta}{\Gamma}\right),
\label{eq:PhotonNumberTheory}$$
which, for our analyzed bias point, amounts to 200$\pm$3 photons in the resonator. From this number, we obtain a Duffing shift $-\alpha |A|^2/2\pi \approx -5.4\pm0.3\,$MHz (for $\alpha/2\pi = 27\pm1.5\,$kHz per photon) and a pump-induced frequency shift $-\beta \Gamma (\epsilon/\Gamma)^2/2\pi \approx -0.64\,$MHz (for $\beta = (7.5\pm0.1)\times10^{-3}$).\
The parameter $\beta$ has the effect of skewing the parametric oscillation region, yielding an expression for the thresholds plotted in Fig. 1(b), $$\frac{\epsilon}{\Gamma} = \frac{1}{\sqrt{2}\beta}\sqrt{1 - 2\beta \frac{\delta}{\Gamma} \pm \sqrt{1 - 4\beta \left( \beta + \frac{\delta}{\Gamma}\right)}}.
\label{eq:ParametricInstabilityThreshold}$$\
**Calibration of attenuation and gain via the Duffing non-linearity.** In this section, we present how we calibrated the gain of the amplifier chain, using the photon-number-dependent frequency shift of the Duffing oscillator, $-\alpha |A|^2$, which we recall from the previous section. The frequency of the resonator as a function of input probe power takes the following form, $$\omega_{r}(P_s) = \omega_{r}(0) - \frac{2 \alpha \Gamma_0}{\Gamma^2}\frac{10^{(P_s-Att-30)/10}}{\hbar \omega_{r}(0)},
\label{eq:DuffingShift}$$
where $\omega_{r}(0)$ denotes the resonant frequency with zero photons in the resonator, $\Gamma_0$ and $\Gamma$ are the external and total loss rates, respectively, and $\alpha$ is the Duffing frequency shift per photon — recall Eq. (\[eq:Alpha\]). Using Eq. (\[eq:DuffingShift\]), we can fit the extracted resonant frequencies as a function of input probe power at different flux bias points, $F$, with the attenuation, $Att$, as the only fitting parameter (since $\alpha$ can be extracted separately by fitting $\omega_{\lambda/4}$ and $\gamma_0$ — recall Eq. (5)). This is shown in Supplementary Fig. 3, where the data for five different flux bias points are fitted to attenuations presented in Supplementary Table 1. From these values, we obtain an average attenuation, $\langle Att \rangle = 127.5\,$dB, which can be compared with the installed 120$\,$dB, indicating that we have a cable loss of 7.5$\,$dB at the measurement frequency.\
Moreover, from the same measurement we can also obtain an estimate for the gain of the amplifier chain by assuming that all the signal gets reflected when it is far off resonance with the resonator *i.e.* reflection coefficient $|S_{11}|^2 = 1$. Then, the gain is obtained from the relation $$G = \left| S_{11} \right|^2 + Att.
\label{eq:Gain}$$
For the five gain estimates presented in Supplementary Table 1, we obtain a gain of $G = 81.0\pm0.37\,$dB, at our given bias point. The error bars for this gain estimation has two origins: $\pm0.17\,$dB from the residual of the linear fit to the gain values presented in Supplementary Table 1, and another $\pm0.2\,$dB from the gain drift over time, which can be compared with our 91$\,$dB of installed amplification.\
**Calibration of the resonator photon number.** From the obtained calibration of the gain of our amplifier chain, $ G $, we can now calculate the conversion factor between our measured power on the digitizer and the number of photons in the resonator, using the following relation, $$\left|A\right|^2 = \frac{P_s - P_n}{2 (\Gamma_0/2\pi) \hbar \omega_{r}^{|0\rangle} 10^{G/10}},
\label{eq:PowerConversionFactor}$$
where $P_s$ and $P_n$ denote our signal and noise power levels, respectively. We demonstrate this for the bottom panel of Fig. 3, where the resonator is probed at a frequency $\omega_{r}^{|0\rangle}/2\pi = 5.212\,$GHz. The external damping rate is $\Gamma_0/2\pi = 1.02\,$MHz, and we calculate the background power level from the end of the trace (when the pump is off). From the obtained SNR, the number of added noise photons can be estimated accordingly, $\left|A\right|^2/\mbox{SNR}^2 = 16.1\pm1.3$.\
**Quantum coherence and readout nondestructiveness.** To study how the parametric pump strength affects the qubit’s relaxation time, we here present coherence measurements for the transmon. First, we calibrate a qubit pulse duration corresponding to a $\pi$-pulse, using a Rabi measurement, where the pulse duration time is swept, for a fixed pulse amplitude. From the fit in Supplementary Fig. 5(a), a pulse length of $\tau_{\pi} = 52\,$ns was obtained, and the Rabi decay time was $T_{\mbox{\tiny{rabi}}} = 2.53\pm0.15\,\mu$s. The histograms corresponding to the first 0.5$\,\mu$s are plotted in Supplementary Fig. 5(b), using the same projective technique as for the histograms in Fig. 5(c) in the main text. Finally, we perform a set of $T_1$ measurements for different pump amplitudes $\epsilon/\Gamma$, and compare these with traditional reflection readout, where we apply a weak resonant probe signal, but no pump ($B(t)\neq 0, \epsilon=0$). The fits to the relaxation times suggest that our readout is not any more destructive to the quantum state of the transmon than the traditional readout technique is. We note, however, that our extracted relaxation time is limited by the Purcell effect, yielding $T_1 \approx [2 \Gamma_0(g_{01}/\Delta)^2]^{-1} = 4.11\,\mu$s. Also see Supplementary Note 3.\
[35]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [, ]{}[****, ()]{} @noop [, ]{}[****, ()]{} [, ](\doibase
10.1038/nature14270)[****, ()]{} [, ](\doibase 10.1103/PhysRevA.74.042318)[****, ()]{} [, ](\doibase 10.1103/PhysRevA.79.013819)[****, ()]{} [, ](\doibase 10.1103/PhysRevLett.106.110502)[****, ()]{} @noop [, ]{}[****, ()]{} [, ](\doibase
10.1038/ncomms5480)[****, ()]{} [, ](\doibase 10.1038/nature09035)[****, ()]{} [, ](\doibase
10.1038/nphys509)[****, ()]{} [, ](\doibase
10.1038/nphys1400)[****, ()]{} [, ](\doibase
10.1103/PhysRevB.86.220503)[****, ()]{} [, ](\doibase
10.1103/PhysRevLett.105.173601)[****, ()]{} [, ](\doibase 10.1103/PhysRevLett.105.100504)[****, ()]{} [, ](\doibase 10.1103/PhysRevA.90.062333)[****, ()]{} @noop [, ]{}[****, ()]{} [, ](\doibase
10.1007/s10909-008-9774-x)[****, ()]{} [, ](\doibase
10.1103/PhysRevLett.105.233907)[****, ()]{} [, ](\doibase 10.1038/nature10561)[****, ()]{} [, ](\doibase
10.1088/1367-2630/15/10/105002)[****, ()]{} @noop [, ]{}[****, ()]{} [, ](\doibase 10.1103/PhysRevB.87.184501)[****, ()]{} [, ](\doibase
10.1063/1.2964182)[****, ()]{} [, ](\doibase 10.1103/PhysRevA.76.042319)[****, ()]{} [, ](\doibase
10.1103/PhysRevA.69.062320)[****, ()]{} [, ](\doibase
10.1038/nature02831.1.)[****, ()]{} [, ](\doibase 10.1103/PhysRevB.74.224506)[****, ()]{} @noop [, ]{}[****, ()]{} @noop [, ]{}[****, ()]{} @noop [, ]{}[****, ()]{} [, ](\doibase
10.1103/PhysRevLett.60.764)[****, ()]{} [, ](\doibase 10.1038/nphys1090)[****, ()]{} [, ](\doibase
10.1103/PhysRevLett.93.207002)[****, ()]{} [, ](\doibase 10.1103/PhysRevLett.94.027005)[****, ()]{} [, ](\doibase
10.1063/1.4933265)[****, ()]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'An estimation problem of fundamental interest is that of phase (or angular) synchronization, in which the goal is to recover a collection of phases (or angles) using noisy measurements of relative phases (or angle offsets). It is known that in the Gaussian noise setting, the maximum likelihood estimator (MLE) has an expected squared $\ell_2$-estimation error that is on the same order as the Cramér-Rao lower bound. Moreover, even though the MLE is an optimal solution to a non-convex quadratic optimization problem, it can be found with high probability using semidefinite programming (SDP), provided that the noise power is not too large. In this paper, we study the estimation and convergence performance of a recently-proposed low-complexity alternative to the SDP-based approach, namely, the generalized power method (GPM). Our contribution is twofold. First, we bound the rate at which the estimation error decreases in each iteration of the GPM and use this bound to show that all iterates—not just the MLE—achieve an estimation error that is on the same order as the Cramér-Rao bound. Our result holds under the least restrictive assumption on the noise power and gives the best provable bound on the estimation error known to date. It also implies that one can terminate the GPM at any iteration and still obtain an estimator that has a theoretical guarantee on its estimation error. Second, we show that under the same assumption on the noise power as that for the SDP-based method, the GPM will converge to the MLE at a linear rate with high probability. This answers a question raised in [@boumal2016nonconvex] and shows that the GPM is competitive in terms of both theoretical guarantees and numerical efficiency with the SDP-based method. At the heart of our convergence rate analysis is a new error bound for the non-convex quadratic optimization formulation of the phase synchronization problem, which could be of independent interest. As a by-product, we give an alternative proof of a result in [@boumal2016nonconvex], which asserts that every second-order critical point of the aforementioned non-convex quadratic optimization formulation is globally optimal in a certain noise regime.'
author:
- 'Huikang Liu[^1]'
- 'Man-Chung Yue[^2]'
- 'Anthony Man-Cho So[^3]'
bibliography:
- 'as.bib'
title: On the Estimation Performance and Convergence Rate of the Generalized Power Method for Phase Synchronization
---
Introduction {#sec:intro}
============
The problem of phase synchronization is concerned with the estimation of a collection of phases[^4] based on noisy measurements of the relative phases. Formally, let $z^\star\in \mathbb{T}^n=\{w\in{\mathbb C}^n: |w_1|=\cdots=|w_n|=1\}$ be an unknown phase vector. Given noisy measurements of the form $$\label{eq:measure}
C_{j\ell} = z_j^\star \bar{z}_\ell^\star +\Delta_{j\ell} \quad\mbox{for } 1\le j<\ell\le n,$$ where $\overline{(\cdot)}$ denotes the complex conjugate and $\Delta_{j\ell}\in{\mathbb C}$ is the noise in the measurement of the relative phase $z_j^\star\bar{z}_\ell^\star$, our goal is to find an estimate $\hat{z}\in{\mathbb T}^n$ of $z^\star\in{\mathbb T}^n$ that best fits those measurements in the least-squares sense. In other words, we are interested in solving the following optimization problem: $$\label{eq:AS-ls}
\hat{z} \in \arg\min_{z\in{\mathbb T}^n} \sum_{1\le j<\ell\le n} \left| C_{j\ell} - z_j\bar{z}_\ell \right|^2.$$ Despite its simple description, the phase synchronization problem arises in a number of applications, including clock synchronization in wireless networks [@giridhar2006distributed], signal reconstruction from phaseless measurements [@ABFM14; @viswanathan2015fast], and ranking of items based on noisy pairwise comparisons [@C16]. For further discussions on the applications of phase synchronization, we refer the reader to [@BBS16] and the references therein.
Although Problem may seem to involve an objective function that is quartic in the decision variable $z$, it can actually be reformulated as a complex quadratic optimization problem with unit-modulus constraints. Indeed, by writing the measurements more compactly as $C=(z^\star)(z^\star)^H+\Delta$, where $(\cdot)^H$ denotes the Hermitian transpose and $\Delta$ is a Hermitian matrix whose diagonal entries are zero and the above-diagonal entries are given by $\{\Delta_{j\ell}:1\le j<\ell\le n\}$, and by noting that $|z_j\bar{z}_\ell|^2=1$ for $1\le j<\ell\le n$ because $z\in{\mathbb T}^n$, we see that Problem is equivalent to $$\label{opt:QP} \tag{QP}
\hat{z} \in \arg\max_{z\in{\mathbb T}^n} \left\{ f(z)=z^HCz \right\}.$$
As it turns out, Problem is NP-hard in general [@TO98]. Over the past two decades or so, many different approaches to tackling Problem have been proposed. One popular approach is to apply the *semidefinite relaxation* (SDR) technique, which will lead to a polynomial-time algorithm for computing a feasible but typically sub-optimal solution to Problem (see [@LMS+10] for an overview of the technique). Interestingly, the *approximation accuracy* of the SDR solution, measured by the relative gap between the objective value of the SDR solution and the optimal value of Problem , can be established under various assumptions on $C$ [@SZY07; @S10a]. However, since our goal is to estimate the unknown phase vector $z^\star$, a more relevant measure of the quality of the SDR solution is its *estimation error*, which intuitively can be defined as the distance between the SDR solution and the target phase vector $z^\star$. Unfortunately, the aforementioned approximation accuracy results do not automatically translate into estimation error results. In an attempt to fill this gap, Bandeira et al. [@BBS16] considered a Gaussian noise model and studied the estimation error of the SDR solution. Specifically, suppose that the measurement noise takes the form $\Delta=\sigma W$, where $W$ is a Wigner matrix (i.e., a Hermitian random matrix whose diagonal entries are zero and the above-diagonal entries are i.i.d. standard complex normal random variables) and $\sigma^2>0$ is the noise power. It is shown in [@BBS16] that if $\sigma=O(n^{1/4})$, then with high probability the standard SDR of Problem has a unique optimal solution that is of rank one; i.e., the SDR is tight. This implies that a global maximizer $\hat{z}$ of Problem , which in this case is also a maximum likelihood estimator (MLE) of the target phase vector of $z^\star$, can be found in polynomial time. Moreover, the expected squared $\ell_2$-estimation error of $\hat{z}$ is bounded above by $O(\sigma^2)$. This matches (up to constants) the Cramér-Rao lower bound developed in [@boumal2014cramer], which applies to any unbiased estimator of $z^\star$. As an aside, although large instances of the standard SDR of Problem may be costly to solve using interior-point methods, they can be solved quite efficiently in practice by numerical methods that exploit structure; see, e.g., [@WGY10; @WMS11; @WGS12]. However, unlike interior-point methods, which are known to converge in polynomial time, most of these methods do not have convergence rate guarantees.
Besides the aforementioned SDR-based method, one can also employ the *generalized power method* (GPM) [@journee2010generalized] (see also [@luss2013conditional]) to tackle Problem . When specialized to Problem , the GPM can be viewed as a gradient method on the manifold ${\mathbb T}^n$ and is much easier to implement than the SDR-based method. In a very recent work, Boumal [@boumal2016nonconvex] analyzed the convergence behavior of the GPM under the same Gaussian noise model used in [@BBS16] and showed that if $\sigma=O(n^{1/6})$, then with high probability the GPM will converge to a global maximizer of Problem when initialized by the eigenvector method in [@Sing11]. This result is significant, since in general the GPM may not even converge to a single point, let alone to a global optimizer of the problem at hand. However, it does not give the rate at which the GPM converges to the global maximizer. Moreover, compared with the result obtained for the SDR approach in [@BBS16], we see that the above result holds only in the more restrictive noise regime of $\sigma=O(n^{1/6})$. Although numerical experiments in [@BBS16; @boumal2016nonconvex] indicate that both the SDR-based method and the GPM can find a global maximizer of Problem even when $\sigma$ is on the order of $n^{1/2}/{\rm polylog}(n)$, proving this rigorously remains an elusive task.
Motivated by the preceding discussion, our goal in this paper is to gain a deeper understanding of the GPM when it is applied to Problem under the same Gaussian noise model used in [@BBS16; @boumal2016nonconvex]. The starting point of our investigation is the following curious facts: Using Proposition \[lem:q\_norm\], which first appears in an earlier version of this paper, Boumal [@boumal2016nonconvex] showed that modulo constants, the expected squared $\ell_2$-estimation error of the initial iterate of the GPM, viz. the one produced by the eigenvector method in [@Sing11], already matches the Cramér-Rao bound. Moreover, in the noise regime $\sigma=O(n^{1/6})$, we know by the results in [@BBS16; @boumal2016nonconvex] that the same is true for the limit point of the sequence of iterates generated by the GPM, as it is a global maximizer of Problem . In view of these facts, it is natural to ask whether the intermediate iterates generated by the GPM also achieve an estimation error that is on the same order as the Cramér-Rao bound, and if so, whether the GPM actually reduces the estimation error in each iteration. Our first contribution is to resolve both of these questions in the affirmative and to bound the rate at which the estimation error decreases in each iteration. Specifically, we show that even at the noise level $\sigma=O(n^{1/2})$, the expected squared $\ell_2$-estimation errors of the iterates do not exceed $(c_1+c_2\tau^k)\sigma^2$, where $c_1,c_2>0$, $\tau\in(0,1)$ are some explicitly given constants and $k$ is the iteration counter; see the discussion after Corollary \[cor:1\]. An interesting aspect of this result is that it holds regardless of whether the iterates converge or not (recall that the convergence result in [@boumal2016nonconvex] holds only for noise level up to $O(n^{1/6})$). Thus, from a statistical estimation viewpoint, one can terminate the GPM at any iteration and still obtain an estimator whose estimation error is on the same order as the Cramér-Rao bound. Moreover, the leading constant in the estimation error becomes smaller as one runs more iterations of the GPM. This explains in part the numerical observation in [@boumal2016nonconvex] that the GPM can often return a good estimate of $z^\star$ even when the noise level is close to $O(n^{1/2})$. To the best of our knowledge, the bound we obtained on the $\ell_2$-estimation error of any accumulation point generated by the GPM holds under the least restrictive noise level requirement and is the best known to date in the Gaussian noise setting.
Next, we study the convergence behavior of the GPM when it is applied to Problem . Our second contribution is to show that in the Gaussian noise setting, if $\sigma=O(n^{1/4})$ and the GPM is initialized by the eigenvector method, then with high probability the sequence of iterates generated by the GPM will converge linearly to a global maximizer of Problem (which is an MLE of $z^\star$); see Corollary \[cor:global-conv\]. The significance of this result is twofold. First, compared with the result in [@boumal2016nonconvex], the noise level requirement for the convergence of the GPM is relaxed from $O(n^{1/6})$ to $O(n^{1/4})$, thus matching the noise level requirement for the tightness of the SDR-based method. Second, our result answers a question raised in [@boumal2016nonconvex] concerning the convergence rate of the GPM and contributes to the growing literature on the design and analysis of fast algorithms for structured non-convex optimization problems (see, e.g., [@SQW15] and the references therein for an overview). Key to our analysis is a new *error bound* for Problem , which provides a computable estimate of the distance between any given point on ${\mathbb T}^n$ and the set of second-order critical points (which includes the global maximizers) of Problem ; see Propositions \[prop:err-bd\] and \[prop:EB-2crit\]. As a by-product, we show that every second-order critical point of Problem is still a global maximizer under a slightly less restrictive noise level requirement than [@boumal2016nonconvex]; see the discussion after the proof of Proposition \[prop:EB-2crit\]. We remark that error bounds have long played an important role in the convergence rate analysis of iterative methods; see, e.g., [@HZSL13; @SZ15; @ZS15; @ZZS15; @LP16; @LWS16; @Z16] for some recent developments. However, most of the error bounds in the cited works are for convex optimization problems. By contrast, our error bound is developed for the non-convex problem , which could be of independent interest.
We end this section by introducing the notations needed. Let $\mathbf{1}$ denote the vector of all ones and $\mathbb{H}^n$ denote the set of $n\times n$ Hermitian matrices. For a complex vector $v\in\mathbb{C}^n$, let $\text{Diag}(v)$ denote the diagonal matrix whose diagonal elements are given by the entries of $v$, $|v|$ denote the vector of entry-wise moduli of $v$, and $\tfrac{v}{|v|}$ denote the vector of entry-wise normalizations of $v$; i.e., $$({\rm Diag}(v))_{jj} = v_j, \quad |v|_j=|v_j|,\quad \left(\frac{v}{|v|}\right)_j=\left\{
\begin{array}{c@{\quad}l}
\tfrac{v_j}{|v_j|} &\text{if } v_j\neq 0,\\
\noalign{\smallskip}
0 & \text{otherwise}.
\end{array}
\right.$$ For a complex matrix $M\in\mathbb{C}^{n\times n}$, let $\text{diag}(M)$ denote the vector whose entries are the diagonal elements of $M$, $\|M\|_{\text{op}}$ denote its operator norm, and $\|M\|_F$ denote its Frobenius norm.
Since the measurements $\{C_{j\ell}:1\le j<\ell\le n\}$ in are invariant under multiplication of a common phase to the target phase vector $z^\star$, we can only identify $z^\star$ up to a global phase. This motivates us to define the $\ell_q$-distance (where $q\in[1,\infty]$) between two phase vectors $w,z\in\mathbb{T}^n$ by $$d_q(w,z)=\min_{\theta\in[0,2\pi)}\|w-e^{i\theta}z\|_q.$$
Preliminaries
=============
In this section, we review the GPM for solving Problem and collect some basic facts that will be used in our subsequent analysis.
The GPM is an iterative method that was introduced in [@journee2010generalized] for maximizing a convex function over a compact set. In each iteration of the GPM, an affine minorant of the objective function at the current iterate is maximized over the feasible set to obtain the next iterate. When specialized to Problem , the maximization performed in each iteration admits a closed-form solution, and the GPM takes the following form:
input: objective matrix $C\in\mathbb{H}^n$, step size $\alpha>0$, initial point $z^0\in{\mathbb T}^n$ return $z^k$ $w^k \gets \left(I+\tfrac{\alpha}{n} C\right)z^k$ $z^{k+1} \gets \tfrac{w^k}{|w^k|}$
Algorithm \[alg:GP\] can be viewed as a projected gradient method (see lines 6-7), though it is not necessarily a Riemannian gradient method on the manifold ${\mathbb T}^n$; see [@boumal2016nonconvex Remark 1]. Due to the non-convexity of Problem , given an arbitrary initial point, Algorithm \[alg:GP\] may not converge to any useful point (if it converges at all). To tackle this issue, Boumal [@boumal2016nonconvex] proposed to use the *eigenvector estimator* $v_C\in{\mathbb T}^n$ (cf. [@Sing11]) to initialize Algorithm \[alg:GP\]. Specifically, let $u\in{\mathbb C}^n$ be a leading eigenvector of $C\in\mathbb{H}^n$ and $a\in{\mathbb C}^n$ be any vector satisfying $a^Hu\not=0$. Then, the vector $v_C$ is defined by $$\label{eq:init}
(v_C)_j = \left\{
\begin{array}{c@{\quad}l}
\tfrac{u_j}{|u_j|} & \text{if}\ u_j\neq 0,\\
\noalign{\smallskip}
\tfrac{a^Hu}{|a^Hu|} & \text{otherwise}
\end{array}
\right. \quad\mbox{for } j=1,\ldots,n.$$ As shown in [@boumal2016nonconvex], the advantage of initializing Algorithm \[alg:GP\] with $z^0=v_C$ is twofold. First, the vector $v_C$ is close to the target phase vector $z^\star$ in the following sense:
([@boumal2016nonconvex Lemma 6]) \[lem:init\] Let $v_C\in{\mathbb T}^n$ be given by . Then, we have $$d_2(v_C,z^\star) \leq \frac{8\|\Delta\|_{\rm op}}{\sqrt{n}}.$$
Second, under some mild assumptions on the measurement noise $\Delta$ and step size $\alpha$, the iterates generated by Algorithm \[alg:GP\] will converge to a global maximizer of Problem :
\[fact:global-convergence\] ([@boumal2016nonconvex Theorem 3]) Suppose that (i) the measurement noise $\Delta$ satisfies $\|\Delta\|_{\rm op}=O(n^{2/3})$ and $\|\Delta z^\star\|_\infty = O(n^{2/3}\sqrt{\log n})$, (ii) the step size $\alpha$ satisfies $\alpha \le \tfrac{n}{\|\Delta\|_{\rm op}}$, and (iii) the initial point $z^0$ is given by $z^0=v_C$. Then, the iterates generated by Algorithm \[alg:GP\] will converge to a global maximizer of Problem .
It should be noted that even allowing for the multiplication of a common phase, a global maximizer $\hat{z}$ of Problem may not equal to the target phase vector $z^\star$. Thus, an immediate question is whether global maximizers of Problem are close to $z^\star$. The following result shows that the answer is affirmative:
([@BBS16 Lemma 4.1]) \[lem:1\] Suppose that $\hat{z}\in{\mathbb C}^n$ satisfies $\|\hat{z}\|_2^2=n$ and $f(z^\star) \leq f(\hat{z})$ (e.g., if $\hat{z}$ is a global maximizer of Problem ). Then, we have $$d_2(\hat{z},z^\star) = \sqrt{2\left( n-|\hat{z}^Hz^\star| \right)}\leq \frac{4 \|\Delta\|_{\rm op}}{\sqrt{n}}.$$
Lastly, let us record a useful property of Algorithm \[alg:GP\]. Recall that $\tilde{z}\in{\mathbb T}^n$ is a *second-order critical point* of Problem if $w^HS(\tilde{z})w \ge 0$ for all $w\in T_{\tilde{z}}{\mathbb T}^n$, where $$S(z) = \Re\left\{ \mbox{Diag}\left( \mbox{diag}(Czz^H) \right) \right\} - C$$ and $$T_z{\mathbb T}^n = \left\{ w \in {\mathbb C}^n : \Re\left\{ w_i\bar{z}_i \right\} = 0 \mbox{ for } i=1,\ldots,n \right\}$$ is the tangent space to ${\mathbb T}^n$ at $z\in{\mathbb T}^n$; see [@BBS16; @boumal2016nonconvex]. By considering the second-order necessary optimality conditions of Problem , it can be shown that every global maximizer of Problem is a second-order critical point. The following result asserts that a second-order critical point of Problem is (i) a fixed point of Algorithm \[alg:GP\] and (ii) close to the target phase vector $z^\star$ if the measurement noise $\Delta$ is not too large.
([@boumal2016nonconvex Lemmas 7, 14, 15, and 16]) \[lem:fix-pt\] Let $\tilde{z}\in{\mathbb T}^n$ be any second-order critical point of Problem and $\tilde{C}= \tfrac{n}{\alpha}\left( I+\tfrac{\alpha}{n}C \right) = C+\tfrac{n}{\alpha}I$. Then, for any $\alpha>0$, $$|(C\tilde{z})_j| = (C\tilde{z})_j \overline{(\tilde{z}_j)} \quad\mbox{and}\quad | (\tilde{C}\tilde{z})_j | = (\tilde{C}\tilde{z})_j \overline{(\tilde{z}_j)} \quad\mbox{for } j=1,\ldots,n.$$ Consequently, we have $\tilde{z}^HC\tilde{z}=\|C\tilde{z}\|_1$, $\tilde{z}^H\tilde{C}\tilde{z}=\|\tilde{C}\tilde{z}\|_1$, and $$\left({\rm Diag}(|\tilde{C}\tilde{z}|)-\tilde{C}\right)\tilde{z} = \left({\rm Diag}(|C\tilde{z}|)-C\right)\tilde{z} = {\mathbf 0}.$$ Moreover, if $\|\Delta\|_{\rm op} \le \tfrac{n}{16}$, then $$|(z^\star)^H\tilde{z}| \ge n-4\|\Delta\|_{\rm op} \quad\mbox{and}\quad d_2(\tilde{z},z^\star) \le \sqrt{8\|\Delta\|_{\rm op}}.$$
To distinguish the different points of interest on ${\mathbb T}^n$, we shall reserve the notations $z^\star$, $\hat{z}$, and $\tilde{z}$ to denote the target phase vector, a global maximizer of Problem , and a second-order critical point of Problem , respectively in the sequel.
Estimation Performance of the GPM
=================================
Facts \[lem:init\] and \[lem:1\] show that both the eigenvector estimator $v_C$ and global maximizers of Problem are close to the target phase vector $z^\star$. In this section, we show that the same is true for all intermediate iterates of Algorithm \[alg:GP\]. In fact, we establish a stronger result: We show that the $\ell_2$- and $\ell_\infty$-estimation errors of the iterates decrease in each iteration of Algorithm \[alg:GP\] and provide explicit bounds on the rates of decrease.
To begin, let us introduce our first result, which concerns the $\ell_2$-estimation errors of the iterates:
\[thm:2-norm\] Suppose that (i) the measurement noise $\Delta$ satisfies $\|\Delta\|_{\rm op}\le\tfrac{n}{16}$, (ii) the step size $\alpha$ satisfies $\alpha\ge2$, and (iii) the initial point $z^0$ is given by $z^0=v_C$. Then, the sequence of iterates $\{z^k\}_{k\ge0}$ generated by Algorithm \[alg:GP\] satisfies $$d_2(z^{k+1},z^\star) \le \mu^{k+1}\cdot d_2(z^0,z^\star) + \frac{\nu}{1-\mu} \frac{8\|\Delta\|_{\rm op}}{\sqrt{n}} \le \left( \mu^{k+1} + \frac{\nu}{1-\mu} \right) \frac{8\|\Delta\|_{\rm op}}{\sqrt{n}}$$ for $k=0,1,\ldots$, where $$\label{eq:mu-nu}
\mu = \frac{16(\alpha\|\Delta\|_{\rm op}+n)}{(7\alpha+8)n} < 1, \quad \nu = \frac{2\alpha}{7\alpha+8}.$$
Theorem \[thm:2-norm\] has two noteworthy features. First, it does not assume that Algorithm \[alg:GP\] converges. Second, it provides a bound on the $\ell_2$-estimation error of each iterate generated by Algorithm \[alg:GP\]. As such, one can terminate Algorithm \[alg:GP\] at any iteration and still has a guarantee on the quality of the estimator.
To further illustrate the usefulness of Theorem \[thm:2-norm\], recall from Facts \[lem:init\] and \[lem:1\] that the $\ell_2$-estimation errors of the initial point $v_C$ and the global maximizers of Problem are bounded above by $\tfrac{8\|\Delta\|_{\rm op}}{\sqrt{n}}$ and $\tfrac{4\|\Delta\|_{\rm op}}{\sqrt{n}}$, respectively. Now, if we take $\alpha=4$ in Algorithm \[alg:GP\], then under the assumptions of Theorem \[thm:2-norm\], we have $\mu \le \tfrac{5}{9}$ and $\nu = \tfrac{2}{9}$. This implies that any accumulation point $z^\infty$ generated by Algorithm \[alg:GP\] satisfies $$d_2(z^{\infty},z^\star)\leq \frac{4\|\Delta\|_{\rm op}}{\sqrt{n}},$$ which matches the bound on the $\ell_2$-estimation error of any global maximizer of Problem . Furthermore, if we let $\alpha\rightarrow\infty$, which can be interpreted as using the update $z^{k+1} \gets \tfrac{Cz^k}{|Cz^k|}$ in line 7 of Algorithm \[alg:GP\], then $$d_2(z^{k+1},z^\star) \le \left( \left(\frac{1}{7}\right)^{k+1} + \frac{1}{3} \right) \frac{8\|\Delta\|_{\rm op}}{\sqrt{n}}$$ for $k=0,1,\ldots$. In this case, our bound is even better than that in Fact \[lem:1\] when $k$ is sufficiently large.
Next, we present our result on the $\ell_\infty$-estimation errors of the iterates:
\[thm:infty-norm\] Under the same assumptions as Theorem \[thm:2-norm\], the sequence of iterates $\{z^k\}_{k\ge0}$ generated by Algorithm \[alg:GP\] satisfies $$d_\infty(z^{k+1},z^\star) \le \gamma^{k+1}\cdot d_\infty(z^0,z^\star) + \frac{\zeta\cdot\mu^k}{1-\gamma/\mu} + \frac{\omega}{1-\gamma}$$ for $k=0,1,2,\dots$, where $$\gamma = \frac{16}{7\alpha+8} < 1, \,\, \zeta = \frac{128\alpha\|\Delta\|_{\rm op}^2}{(7\alpha+8)n^{3/2}}, \,\, \omega = \frac{16\alpha}{7\alpha+8}\left( \frac{\nu}{1-\mu}\frac{8\|\Delta\|_{\rm op}^2}{n^{3/2}} + \frac{\|\Delta z^\star\|_\infty}{n} \right),$$ and $\mu,\nu$ are given in , so that $\gamma/\mu<1$.
To prove Theorems \[thm:2-norm\] and \[thm:infty-norm\], we need the following technical results:
\[lem:q\_norm\] For any $w\in \mathbb{C}^n$, $z\in \mathbb{T}^n$, and $q\in[1,\infty]$, we have $$\left\| \frac{w}{|w|}-z \right\|_{q} \leq 2 \|w-z\|_{q}.$$
Without loss of generality, we may assume that $z=\mathbf{1}$. By definition of $\tfrac{w}{|w|}$, it suffices to show that for $j=1,\dots,n$, $$\left|\left(\frac{w}{|w|}\right)_j-1 \right|\leq 2|w_j-1|.$$ The above inequality holds trivially if $w_j=0$. Hence, we may focus on the case where $w_j\neq 0$. We claim that $$|e^{i\phi}-1|\leq 2|re^{i\phi}-1|\quad \mbox{for any } \phi\in [0,2\pi) \mbox{ and } r\geq0.$$ To prove this, observe that $|re^{i\phi}-1|^2=r^2-2r\cos\phi+1$. Thus, we have $$\arg\min_{r\geq 0}|re^{i\phi}-1|^2=
\left\{
\begin{array}{c@{\quad}l}
0 & \mbox{if } \phi\in [\tfrac{\pi}{2},\tfrac{3\pi}{2}],\\
\noalign{\smallskip}
\cos\phi & \mbox{if } \phi\in[0,\tfrac{\pi}{2})\cup(\tfrac{3\pi}{2},2\pi),
\end{array}
\right.$$ from which it follows that $$\label{eq:min-val}
\min_{r\geq 0}|re^{i\phi}-1|^2=
\left\{
\begin{array}{c@{\quad}l}
1 & \mbox{if } \phi\in [\tfrac{\pi}{2},\tfrac{3\pi}{2}],\\
\noalign{\smallskip}
\sin^2\phi & \mbox{if } \phi\in[0,\tfrac{\pi}{2})\cup(\tfrac{3\pi}{2},2\pi).
\end{array}
\right.$$ Now, for $\phi\in [\tfrac{\pi}{2},\tfrac{3\pi}{2}]$, by the triangle inequality and , we have $$|e^{i\phi}-1|\leq 2\leq 2|re^{i\phi}-1|\quad \text{for any } r\geq0.$$ On the other hand, for $\phi\in[0,\tfrac{\pi}{2})\cup(\tfrac{3\pi}{2},2\pi)$, we use the half-angle formula and to get $$|e^{i\phi}-1|=\sqrt{2(1-\cos\phi)}=2\left|\sin\frac{\phi}{2}\right|\leq 2\left| \sin\phi\right|\leq 2|re^{i\phi}-1|\quad \text{for any } r\geq0.$$ Combining the above two cases, the proof is completed.
\[prop:iter-bd\] Let $\{z^k\}_{k\ge0}$ be the sequence of iterates generated by Algorithm \[alg:GP\] with $\alpha>0$. For $q\in[1,\infty]$ and $k=0,1,\ldots$, define $$\begin{aligned}
\theta_k &=& \arg\min_{\theta\in[0,2\pi)} \|z^k-e^{i\theta}z^\star\|_q, \\
\noalign{\smallskip}
\epsilon^k &=& e^{-i\theta_k}(z^k-e^{i\theta_k}z^\star), \\
\noalign{\smallskip}
\beta_k &=& 1+\alpha+\frac{\alpha}{n} (z^\star)^H(\epsilon^k).\end{aligned}$$ Then, for any $r\in{\mathbb C}$ and $k\in\{0,1,\ldots\}$, we have $$d_q(z^{k+1},z^\star) \le 2\|rg^k-z^\star\|_q,$$ where $$g^k = \beta_kz^\star + \left( I+\frac{\alpha}{n}\Delta \right)\epsilon^k + \frac{\alpha}{n}\Delta z^\star.$$
Consider a fixed $k\in\{0,1,\ldots\}$. By definition, we have $$\begin{aligned}
w^k &=& \left( I+\frac{\alpha}{n} C \right) z^k \\
\noalign{\smallskip}
&=& e^{i\theta_k} \left( I+\frac{\alpha}{n}((z^\star)(z^\star)^H+\Delta) \right)(z^\star+\epsilon^k) \\
\noalign{\smallskip}
&=&\left[\left( 1+\alpha+\frac{\alpha}{n}(z^\star)^H(\epsilon^k) \right)z^\star + \left( I+\frac{\alpha}{n}\Delta \right)\epsilon^k + \frac{\alpha}{n}\Delta z^\star\right]e^{i\theta_k} \\
\noalign{\smallskip}
&=& g^k e^{i\theta_k}.\end{aligned}$$ Since $z^{k+1}=\tfrac{w^k}{|w^k|}$, it follows from Proposition \[lem:q\_norm\] that for any $r\in{\mathbb C}\setminus\{0\}$, $$d_q(z^{k+1},z^\star) \le \left\| \frac{g^k}{|g^k|}-z^\star \right\|_q = \left\| \frac{rg^k}{|rg^k|}-z^\star \right\|_q \le 2\|rg^k-z^\star\|_q.$$ Since the above inequality holds for all $r\in{\mathbb C}\setminus\{0\}$, by taking $r\rightarrow0$, we see that it holds for $r=0$ as well.
We are now ready to prove Theorems \[thm:2-norm\] and \[thm:infty-norm\].
We prove by induction that for $k=0,1,\ldots$, the following inequalities hold: $$\begin{aligned}
\|\epsilon^k\|_2 &\le& \frac{\sqrt{n}}{2}, \label{eq:eps-bd} \\
\noalign{\smallskip}
d_2(z^{k+1},z^\star) &\le& \mu \cdot d_2(z^k,z^\star) + \nu\cdot\frac{8\|\Delta\|_{\rm op}}{\sqrt{n}}. \label{eq:recur}\end{aligned}$$ Indeed, by the definition of $\epsilon^0$, Fact \[lem:init\], and the assumption that $\|\Delta\|_{\rm op}\le\tfrac{n}{16}$, we have $\|\epsilon^0\|_2 = d_2(z^0,z^\star) \le \tfrac{\sqrt{n}}{2}$. This implies that $$\begin{aligned}
|\beta_0| &\ge& \left| 1+\alpha+\frac{\alpha}{n}\Re\left( (z^\star)^H(\epsilon^0) \right) \right| \nonumber \\
\noalign{\smallskip}
&=& \left| 1+\alpha+\frac{\alpha}{2n} \left( \|z^\star+\epsilon^0\|_2^2 - \|z^\star\|_2^2 - \|\epsilon^0\|_2^2 \right) \right| \nonumber \\
\noalign{\smallskip}
&=& \left| 1+\alpha+\frac{\alpha}{2n} \left( \|z^0\|_2^2 - \|z^\star\|_2^2 - \|\epsilon^0\|_2^2 \right) \right| \nonumber \\
\noalign{\smallskip}
&\ge& 1+\frac{7\alpha}{8}, \label{eq:beta-bd}\end{aligned}$$ where the last inequality follows from the fact that $\|z^0\|_2^2=\|z^\star\|_2^2=n$ and $\|\epsilon^0\|_2\le\tfrac{\sqrt{n}}{2}$. Hence, by taking $r=\beta_0^{-1}$ (which is well-defined) and $q=2$ in Proposition \[prop:iter-bd\] and using , we have $$\begin{aligned}
d_2(z^1,z^\star) &\le& 2\left\| \beta_0^{-1} \left( I+\frac{\alpha}{n}\Delta \right)\epsilon^0 + \beta_0^{-1}\frac{\alpha}{n}\Delta z^\star \right\|_{2} \nonumber \\
\noalign{\smallskip}
&\le& 2|\beta_0^{-1}| \cdot \left[ \left\| \left(I+\frac{\alpha}{n}\Delta\right)\epsilon^0 \right\|_{2} + \frac{\alpha}{n} \| \Delta z^\star\|_{2} \right] \nonumber \\
\noalign{\smallskip}
&\le& \frac{16}{7\alpha+8} \left[ \left(1+\frac{\alpha}{n}\|\Delta\|_{\rm op} \right)\|\epsilon^0\|_{2}+\frac{\alpha}{\sqrt{n}}\|\Delta\|_{\rm op} \right] \nonumber \\
\noalign{\smallskip}
&=& \mu\cdot d_2(z^0,z^\star) + \nu \cdot \frac{8\|\Delta\|_{\rm op}}{\sqrt{n}}. \label{eq:recur-pf}\end{aligned}$$ Now, suppose that and hold for some $k\ge0$. By the inductive hypothesis and the assumption that $\|\Delta\|_{\rm op}\le\tfrac{n}{16}$ and $\alpha\ge2$, we have $$\begin{aligned}
\|\epsilon^{k+1}\|_2 &=& d_2(z^{k+1},z^\star) \,\,\,\le\,\,\, \mu\cdot d_2(z^k,z^\star)+\nu \cdot \frac{8\|\Delta\|_{\rm op}}{\sqrt{n}} \\
\noalign{\smallskip}
&=& \mu\|\epsilon^k\|_2 + \nu \cdot \frac{8\|\Delta\|_{\rm op}}{\sqrt{n}} \\
\noalign{\smallskip}
&\le& \frac{8(\alpha\|\Delta\|_{\rm op}+n)}{(7\alpha+8)\sqrt{n}} + \frac{16\alpha\|\Delta\|_{\rm op}}{(7\alpha+8)\sqrt{n}} \\
\noalign{\smallskip}
&\le& \frac{\sqrt{n}}{2}.\end{aligned}$$ Using the same argument as the derivation of the inequality , we have $|\beta_{k+1}|\ge1+\tfrac{7\alpha}{8}$. Hence, following the same derivation as the inequality , we obtain $d_2(z^{k+2},z^\star) \le \mu\cdot d_2(z^{k+1},z^\star) + \nu\cdot \tfrac{8\|\Delta\|_{\rm op}}{\sqrt{n}}$. This completes the inductive step.
To complete the proof of Theorem \[thm:2-norm\], it remains to unroll and use Fact \[lem:init\].
From the proof of Theorem \[thm:2-norm\], we have $\|\epsilon^k\|_2\leq \tfrac{\sqrt{n}}{2}$ and $|\beta_k^{-1}|\leq\tfrac{8}{7\alpha+8}$ for $k=0,1,\ldots$. By taking $r=\beta_k^{-1}$ and $q=\infty$ in Proposition \[prop:iter-bd\] and using Theorem \[thm:2-norm\], we compute $$\begin{aligned}
& & d_\infty(z^{k+1},z^\star) \nonumber \\
\noalign{\smallskip}
&\le& 2\|\beta_k^{-1}g^k-z^\star\|_\infty \nonumber \\
\noalign{\smallskip}
&\le& 2|\beta_k^{-1}| \cdot \left[ \|\epsilon^k\|_\infty + \frac{\alpha}{n} \left( \|\Delta\epsilon^k\|_\infty + \|\Delta z^\star\|_\infty \right) \right] \nonumber \\
\noalign{\smallskip}
&\le& \frac{16}{7\alpha+8}\left[ d_\infty(z^k,z^\star) + \frac{\alpha}{n} \left( \|\Delta\epsilon^k\|_2 + \|\Delta z^\star\|_\infty \right) \right] \nonumber \\
\noalign{\smallskip}
&\le& \frac{16}{7\alpha+8}\left[ d_\infty(z^k,z^\star) + \frac{\alpha}{n} \left( \|\Delta\|_{\rm op} \cdot \|\epsilon^k\|_2 + \|\Delta z^\star\|_\infty \right) \right] \nonumber \\
\noalign{\smallskip}
&\le& \frac{16}{7\alpha+8}\left[ d_\infty(z^k,z^\star) + \frac{\alpha}{n} \left( \left( \mu^k+\frac{\nu}{1-\mu} \right) \frac{8\|\Delta\|_{\rm op}^2}{\sqrt{n}} + \|\Delta z^\star\|_\infty \right) \right] \nonumber \\
\noalign{\smallskip}
&=& \gamma \cdot d_\infty(z^k,z^\star) + \zeta\cdot\mu^k + \omega. \label{ineq:infty_norm}\end{aligned}$$ Since $\alpha\geq2$, we have $\gamma\in(0,1)$ and $$\frac{\gamma}{\mu}=\frac{16}{(7\alpha+8)}\frac{(7\alpha+8)n}{16(\alpha\|\Delta\|_{\rm op}+n)}=\frac{n}{\alpha\|\Delta\|_{\rm op}+n}\in (0,1).$$ It follows from that $$\begin{aligned}
d_\infty(z^{k+1},z^\star) &\le& \gamma \cdot d_\infty(z^k,z^\star) + \zeta\cdot\mu^k + \omega \\
\noalign{\smallskip}
&\le& \gamma^{k+1}\cdot d_\infty(z^0,z^\star) + \zeta\sum_{j=0}^k \gamma^j\mu^{k-j} + \omega\sum_{j=0}^k\gamma^j \\
\noalign{\smallskip}
&\le& \gamma^{k+1}\cdot d_\infty(z^0,z^\star) + \frac{\zeta\cdot\mu^k}{1-\gamma/\mu} + \frac{\omega}{1-\gamma}.\end{aligned}$$ This completes the proof.
By specializing the above results to the Gaussian noise setting, we obtain the following corollary:
\[cor:1\] Suppose that the measurement noise $\Delta$ takes the form $\Delta=\sigma W$, where $\sigma^2>0$ is the noise power satisfying $\sigma\in\left(0,\tfrac{\sqrt{n}}{48}\right]$ and $W\in\mathbb{H}^n$ is a Wigner matrix. Suppose further that the step size $\alpha$ satisfies $\alpha\ge2$ and the initial point $z^0$ is given by $z^0=v_C$. Then, with probability at least $1-2n^{-5/4}-2e^{-n/2}$, the sequence of iterates $\{z^k\}_{k\ge0}$ generated by Algorithm \[alg:GP\] satisfies $$\begin{aligned}
d_2(z^{k+1},z^\star) &\le& \left(\frac{\alpha+16}{7\alpha+8}\right)^{k+1}d_2(z^0,z^\star) + \frac{24\alpha}{3\alpha-4}\sigma, \\
\noalign{\smallskip}
d_\infty(z^{k+1},z^\star) &\le& \left(\frac{16}{7\alpha+8}\right)^{k+1}d_\infty(z^0,z^\star) + \left( \frac{\alpha+16}{7\alpha+8}\right)^{k+1}\frac{\sqrt{n}}{2} \\
\noalign{\smallskip}
& & \,\,+\,\, \frac{48\alpha}{7\alpha+8}\left( \sqrt{\log n}+\frac{24\alpha}{3\alpha-4}\sigma \right)\frac{\sigma}{\sqrt{n}}\end{aligned}$$ for $k=0,1,\ldots$.
By [@BBS16 Proposition 3.3], we have $\|W\|_{\rm op}\le3\sqrt{n}$ and $\|Wz^\star\|_\infty \le 3\sqrt{n\log n}$ with probability at least $1-2n^{-5/4}-2e^{-n/2}$. The result then follows by combining these estimates with the bounds in Theorems \[thm:2-norm\] and \[thm:infty-norm\].
Note that by Fact \[lem:init\] and [@BBS16 Proposition 3.3], we have $d_2(z^0,z^\star)\le24\sigma$ with high probability. Hence, for $\alpha>4$ and $k$ sufficiently large, the bound on the $\ell_2$-estimation error will be strictly less than $12\sigma$, which is better than that obtained from Fact \[lem:1\] for any maximum likelihood estimator (which is a global maximizer of Problem ) of the target phase vector $z^\star$. Furthermore, for $k=0,1,\ldots$, since $d_2(z^k,z^\star)^2 \le 2n$, we have $$\begin{aligned}
& & \mathbb{E}\left[d_2(z^{k+1},z^\star)^2\right] \\
\noalign{\smallskip}
&\le& 1152\left(1-2n^{-5/4}-2e^{-n/2}\right)\left[ \left(\frac{\alpha+16}{7\alpha+8}\right)^{2(k+1)}+ \left(\frac{\alpha}{3\alpha-4}\right)^2 \right]\sigma^2 \\
\noalign{\smallskip}
&& \,\,+\,\, \left(2n^{-5/4}+2e^{-n/2}\right)(2n) \\
\noalign{\smallskip}
&\le& (c_1+c_2\tau^k)\sigma^2\end{aligned}$$ for some constants $c_1,c_2>0$ and $\tau\in(0,1)$. This shows that the expected squared $\ell_2$-estimation errors of the iterates generated by Algorithm \[alg:GP\] are all on the order of $\sigma$, which matches the Cramér-Rao bound developed in [@boumal2014cramer]. It is worth noting that the above conclusions hold even when the noise level is $\sigma=O(n^{1/2})$, which is the least restrictive among similar results in the literature; cf. [@BBS16; @boumal2016nonconvex]. Our result explains in part the excellent numerical estimation performance of the GPM observed in [@boumal2016nonconvex] even when the noise level is close to $O(n^{1/2})$.
Convergence Rate of the GPM
===========================
Although the results in the previous section show that Algorithm \[alg:GP\] generates increasingly accurate (in the $\ell_2$ and $\ell_\infty$ sense) estimators of the target phase vector $z^\star$, they do not shed any light on its convergence behavior. On the other hand, recall from Fact \[fact:global-convergence\] that the sequence of iterates generated by Algorithm \[alg:GP\] will converge to a global maximizer of Problem under suitable assumptions on the measurement noise $\Delta$ and step size $\alpha$. However, it does not give the rate of convergence. In this section, we prove that under weaker assumptions than those of Fact \[fact:global-convergence\], both the sequence of iterates and the associated sequence of objective values generated by Algorithm \[alg:GP\] will converge *linearly* to a global maximizer and the optimal value of Problem , respectively. Specifically, we have the following result:
\[thm:convergence-rate\] Suppose that (i) the measurement noise $\Delta$ satisfies $\|\Delta\|_{\rm op}\le\tfrac{n^{3/4}}{312}$ and $\|\Delta z^\star\|_\infty \le \tfrac{n}{24}$, (ii) the step size $\alpha$ satisfies $\alpha\in\left[ 4,\tfrac{n}{\|\Delta\|_{\rm op}} \right)$, and (iii) the initial point $z^0$ is given by $z^0=v_C$. Then, the sequence of iterates $\{z^k\}_{k\ge0}$ generated by Algorithm \[alg:GP\] satisfies $$\begin{aligned}
f(\hat{z}) - f(z^k) &\le& \left( f(\hat{z})-f(z^0) \right) \lambda^k, \\
\noalign{\smallskip}
d_2(z^k,\hat{z}) &\leq& a \left( f(\hat{z})-f(z^0) \right)^{1/2} \lambda^{k/2}
\end{aligned}$$ for $k=0,1,\ldots$, where $a>0,\lambda\in(0,1)$ are quantities that depend only on $n$ and $\alpha$, and $\hat{z}$ is any global maximizer of Problem .
Theorem \[thm:convergence-rate\] improves upon Fact \[fact:global-convergence\] in two aspects. First, Theorem \[thm:convergence-rate\] holds under a less restrictive requirement on the measurement noise $\Delta$. Specifically, it requires that $\|\Delta\|_{\rm op}=O(n^{3/4})$ and $\|\Delta z^\star\|_\infty=O(n)$, while Fact \[fact:global-convergence\] requires that $\|\Delta\|_{\rm op}=O(n^{2/3})$ and $\|\Delta z^\star\|_\infty=O(n^{2/3}\sqrt{\log n})$. Second, Theorem \[thm:convergence-rate\] is more quantitative than Fact \[fact:global-convergence\] in the sense that it also gives the rate at which Algorithm \[alg:GP\] converges. Consequently, we resolve an open question raised in [@boumal2016nonconvex].
The proof of Theorem \[thm:convergence-rate\] consists of two main parts. The first, which is the more challenging part, is to establish the following *error bound* for Problem . Such a bound provides a computable estimate of the distance between any point in a neighborhood of $z^\star$ and the set of global maximizers of Problem , which could be of independent interest.
\[prop:err-bd\] Let $\Sigma:{\mathbb T}^n\rightarrow\mathbb{H}^n$ and $\rho:{\mathbb T}^n\rightarrow{\mathbb R}_+$ be defined as $$\Sigma(z)={\rm Diag}(|\tilde{C}z|)-\tilde{C}, \quad
\rho(z)= \|\Sigma(z)z\|_2 = \left\| \left({\rm Diag}(|\tilde{C}z|)-\tilde{C}\right)z \right\|_2,$$ where $\tilde{C}=C+\frac{n}{\alpha}I$. Under the assumptions of Theorem \[thm:convergence-rate\], for any point $z\in{\mathbb T}^n$ satisfying $d_2(z,z^\star)\le\tfrac{\sqrt{n}}{2}$ and any global maximizer $\hat{z}\in{\mathbb T}^n$ of Problem , we have $$d_2(z,\hat{z}) \le \frac{8}{n}\rho(z).$$
Before we prove Proposition \[prop:err-bd\], several remarks are in order. First, recall from Fact \[lem:1\] that every global maximizer $\hat{z}$ of Problem satisfies $d_2(\hat{z},z^\star)\le\tfrac{\sqrt{n}}{2}$ whenever $\|\Delta\|_{\rm op}\le\tfrac{n}{8}$. Together with Proposition \[prop:err-bd\], this shows that up to a global phase, Problem has a unique global maximizer. Second, the proof of Theorem \[thm:2-norm\] reveals that the sequence of iterates $\{z^k\}_{k\ge0}$ generated by Algorithm \[alg:GP\] satisfies $d_2(z^k,z^\star)\le\tfrac{\sqrt{n}}{2}$ for $k=0,1,\ldots$ whenever $\|\Delta\|_{\rm op}\le\tfrac{n}{16}$. Thus, the error bound in Proposition \[prop:err-bd\] applies to the entire sequence $\{z^k\}_{k\ge0}$. Third, since every global maximizer $\hat{z}$ of Problem is a second-order critical point, we have $\rho(\hat{z})=0$ by Fact \[lem:fix-pt\]. Proposition \[prop:err-bd\] shows that the converse is also true. Hence, we can view $\rho$ as a surrogate measure of optimality and use it to keep track of Algorithm \[alg:GP\]’s progress.
Since $$\label{ineq:pf-5}
\rho(z) = \|\Sigma(z)z\|_2 \geq \|\Sigma(\hat{z})z\|_2 - \|(\Sigma(z)-\Sigma(\hat{z}))z\|_2,$$ it suffices to establish an upper bound on $\|(\Sigma(z)-\Sigma(\hat{z}))z\|_2$ and a lower bound on $\|\Sigma(\hat{z})z\|_2$. Towards that end, recall that $$\tilde{C}=C+\frac{n}{\alpha}I = (z^\star)(z^\star)^H+\Delta + \frac{n}{\alpha}I$$ and let $$\hat{\theta}=\arg\min_{\theta\in[0,2\pi)} \|z-e^{i\theta}\hat{z}\|_2, \quad \hat{\theta}^\star = \arg\min_{\theta\in[0,2\pi)} \|\hat{z}-e^{i\theta}z^\star\|_2.$$ First, we bound $$\begin{aligned}
& & \|(\Sigma(z)-\Sigma(\hat{z}))z\|_2 \nonumber \\
\noalign{\smallskip}
&=& \left\|\left(\text{Diag}(|\tilde{C}z|)-\text{Diag}(|\tilde{C}\hat{z}|)\right)z\right\|_2 \nonumber \\
\noalign{\smallskip}
&=& \left( \sum_{j=1}^n \left| \left( |(\tilde{C}z)_j| - |(\tilde{C}\hat{z})_j| \right)z_j \right|^2 \right)^{1/2} \nonumber \\
\noalign{\smallskip}
&=& \left\||\tilde{C}e^{-i\hat{\theta}}z|-|\tilde{C}\hat{z}|\right\|_2 \nonumber \\
\noalign{\smallskip}
&\le& \|\tilde{C}(e^{-i\hat{\theta}}z-\hat{z})\|_2 \nonumber\\
\noalign{\smallskip}
&\le& \|(z^\star)(z^\star)^H(e^{-i\hat{\theta}}z-\hat{z})\|_2 + \|\Delta(e^{-i\hat{\theta}}z-\hat{z})\|_2 + \frac{n}{\alpha}\|e^{-i\hat{\theta}}z-\hat{z}\|_2 \nonumber \\
\noalign{\smallskip}
&\le& \sqrt{n} \cdot | (z^\star)^H(e^{-i\hat{\theta}}z-\hat{z}) | + \left( \|\Delta\|_{\rm op} + \frac{n}{\alpha} \right) d_2(z,\hat{z}). \label{ineq:pf-1}\end{aligned}$$ By definition of $\hat{\theta}$, we have $\hat{z}^H(e^{-i\hat{\theta}}z) = |\hat{z}^Hz|$, which implies that $$\label{eq:norm-hat}
\| e^{-i\hat{\theta}}z-\hat{z} \|_2^2 = 2 (n - |\hat{z}^Hz|).$$ This, together with Fact \[lem:1\], yields $$\begin{aligned}
| (z^\star)^H(e^{-i\hat{\theta}}z-\hat{z}) | &\le& | (z^\star-e^{-i\hat{\theta}^\star}\hat{z})^H(e^{-i\hat{\theta}}z-\hat{z}) | + | (e^{-i\hat{\theta}^\star}\hat{z})^H(e^{-i\hat{\theta}}z-\hat{z}) | \nonumber \\
\noalign{\smallskip}
&\le& \| z^\star-e^{-i\hat{\theta}^\star}\hat{z} \|_2 \cdot \| e^{-i\hat{\theta}}z-\hat{z} \|_2 + \left| |\hat{z}^Hz|-n \right| \nonumber \\
\noalign{\smallskip}
&\le& \frac{4\|\Delta\|_{\rm op}}{\sqrt{n}} \cdot d_2(z,\hat{z}) + \frac{1}{2} d_2(z,\hat{z})^2. \label{ineq:pf-2}\end{aligned}$$ Upon substituting into , we obtain $$\label{eq:ub}
\|(\Sigma(z)-\Sigma(\hat{z}))z\|_2 \le \left(5\|\Delta\|_{\rm op} + \frac{n}{\alpha} \right) d_2(z,\hat{z}) + \frac{\sqrt{n}}{2} d_2(z,\hat{z})^2.$$
Next, let $\hat{u}=\left( I-\tfrac{1}{n}\hat{z}\hat{z}^H \right)(e^{-i\hat{\theta}}z-\hat{z})$ be the projection of $e^{-i\hat{\theta}}z-\hat{z}$ onto the orthogonal complement of $\mbox{span}(\hat{z})$. Hence, we have $\hat{u}^H\hat{z}=0$ and $$\label{ineq:pf-8}
\|\hat{u}\|_2 \geq \left\| e^{-i\hat{\theta}}z-\hat{z} \right\|_2 - \left\|\frac{1}{n}\hat{z}\hat{z}^H(e^{-i\hat{\theta}}z-\hat{z})\right\|_2 = d_2(z,\hat{z}) - \frac{1}{2\sqrt{n}}d_2(z,\hat{z})^2,$$ where the last equality follows from . Moreover, by definition of $\hat{\theta}^\star$, we have $(z^\star)^H(e^{-i\hat{\theta}^\star}\hat{z})=|(z^\star)^H\hat{z}|$ and $$\label{eq:star-hat}
|(z^\star)^H\hat{z}| = n-\frac{1}{2}\|\hat{z}-e^{i\hat{\theta}^\star}z^\star\|_2^2.$$ Hence, $$\begin{aligned}
\hat{u}^H\Sigma(\hat{z})\hat{u} &=& \hat{u}^H\left( \mbox{Diag}(|\tilde{C}\hat{z}|)-\tilde{C} \right)\hat{u} \nonumber \\
\noalign{\smallskip}
&=& \hat{u}^H\left( \mbox{Diag}(|C\hat{z}|)-C \right)\hat{u} \label{eq:Ctilde} \\
\noalign{\smallskip}
&\ge& \left(\sum_{j=1}^n |(C\hat{z})_j|\cdot |\hat{u}_j|^2\right) - |(z^\star)^H\hat{u}|^2 - \hat{u}^H\Delta\hat{u} \nonumber \\
\noalign{\smallskip}
&\ge& \left(|(z^\star)^H\hat{z}| - \|\Delta\hat{z}\|_\infty \right) \|\hat{u}\|_2^2 - \left| \hat{u}^H(z^\star-e^{-i\hat{\theta}^\star}\hat{z}) \right|^2 - \|\Delta\|_{\rm op}\|\hat{u}\|_2^2 \nonumber \\
\noalign{\smallskip}
&\ge& \left( n - \|\Delta\hat{z}\|_\infty - \|\Delta\|_{\rm op} - \frac{3}{2}\|z^\star-e^{-i\hat{\theta}^\star}\hat{z}\|_2^2 \right) \|\hat{u}\|_2^2 \label{eq:star-hat-2} \\
\noalign{\smallskip}
&\ge& \left( n - \|\Delta\hat{z}\|_\infty - \|\Delta\|_{\rm op} - \frac{24\|\Delta\|_{\rm op}^2}{n} \right) \|\hat{u}\|_2^2, \label{ineq:pf-9}\end{aligned}$$ where follows from Fact \[lem:fix-pt\] and the fact that $\hat{z}$ is a second-order critical point of Problem , is due to , and follows from Fact \[lem:1\]. Since $\Sigma(\hat{z})\hat{z}={\mathbf 0}$ by Fact \[lem:fix-pt\], we obtain from and that $$\begin{aligned}
& & \|\Sigma(\hat{z})z\|_2 \,\,\,=\,\,\, \|\Sigma(\hat{z})\hat{u}\|_2 \nonumber \\
\noalign{\smallskip}
&\ge& \left( n - \|\Delta\hat{z}\|_\infty - \|\Delta\|_{\rm op} - \frac{24\|\Delta\|_{\rm op}^2}{n} \right) \left( d_2(z,\hat{z}) - \frac{1}{2\sqrt{n}}d_2(z,\hat{z})^2 \right). \label{ineq:pf-3}\end{aligned}$$ Now, by Fact \[lem:1\] and the assumption that $d_2(z,z^\star)\le\tfrac{\sqrt{n}}{2}$, we have $$d_2(z,\hat{z}) \le d_2(z,z^\star) + d_2(\hat{z},z^\star) \le \frac{\sqrt{n}}{2} + \frac{4\|\Delta\|_{\rm op}}{\sqrt{n}}.$$ This implies that $$\label{ineq:pf-11}
d_2(z,\hat{z})^2 \le \left( \frac{\sqrt{n}}{2} + \frac{4\|\Delta\|_{\rm op}}{\sqrt{n}} \right) d_2(z,\hat{z}).$$ Moreover, $$\begin{aligned}
\|\Delta\hat{z}\|_\infty &\le& \|\Delta z^\star\|_\infty + \|\Delta(e^{-i\hat{\theta}^\star}\hat{z}-z^\star)\|_\infty \nonumber \\
\noalign{\smallskip}
&\le& \|\Delta z^\star\|_\infty + \|\Delta\|_{\rm op}\cdot d_2(\hat{z},z^\star) \nonumber \\
\noalign{\smallskip}
&\le& \|\Delta z^\star\|_\infty + \frac{4\|\Delta\|_{\rm op}^2}{\sqrt{n}}. \label{eq:infty-bd}\end{aligned}$$ It follows from , , , , and that $$\begin{aligned}
\rho(z) &\ge& \left[ \left( \frac{1}{2}-\frac{1}{\alpha} \right)n - \frac{3\|\Delta\hat{z}\|_\infty}{4} - \frac{39\|\Delta\|_{\rm op}}{4} - \frac{16\|\Delta\|_{\rm op}^2}{n} \right] d_2(z,\hat{z}) \\
\noalign{\smallskip}
&\ge& \left[ \frac{n}{4} - \frac{3\|\Delta z^\star\|_\infty}{4} - \frac{39\|\Delta\|_{\rm op}}{4} - \frac{3\|\Delta\|_{\rm op}^2}{\sqrt{n}} - \frac{16\|\Delta\|_{\rm op}^2}{n} \right] d_2(z,\hat{z}) \\
\noalign{\smallskip}
&\ge& \frac{n}{8}d_2(z,\hat{z})\end{aligned}$$ whenever $\|\Delta\|_{\rm op} \le \tfrac{n^{3/4}}{312}$, $\|\Delta z^\star\|_\infty \le \tfrac{n}{24}$, and $\alpha\ge4$. This completes the proof.
We note that under a slightly more restrictive noise setting, one can establish an error bound similar to that in Proposition \[prop:err-bd\] to estimate the distance between any point in a neighborhood of $z^\star$ and the set of *second-order critical points* of Problem . Specifically, we have the following result:
\[prop:EB-2crit\] Let $\Sigma:{\mathbb T}^n\rightarrow\mathbb{H}^n$ and $\rho:{\mathbb T}^n\rightarrow{\mathbb R}_+$ be as in Proposition \[prop:err-bd\]. Suppose that (i) the measurement noise $\Delta$ satisfies $\|\Delta\|_{\rm op}\le\tfrac{n^{2/3}}{32768}$ and $\|\Delta z^\star\|_\infty \le \tfrac{n}{24}$, and (ii) the parameter $\alpha$ satisfies $\alpha\ge4$. Then, for any point $z\in{\mathbb T}^n$ satisfying $d_2(z,z^\star)\le\tfrac{\sqrt{n}}{2}$ and any second-order critical point $\tilde{z}\in{\mathbb T}^n$ of Problem , we have $$d_2(z,\tilde{z}) \le \frac{8}{n}\rho(z).$$
The proof is essentially the same as that of Proposition \[prop:err-bd\]. To save space, let us just highlight the key steps. Similar to , we have $$\label{eq:2crit-0}
\rho(z) = \|\Sigma(z)z\|_2 \geq \|\Sigma(\tilde{z})z\|_2 - \|(\Sigma(z)-\Sigma(\tilde{z}))z\|_2.$$ Define $$\tilde{\theta}=\arg\min_{\theta\in[0,2\pi)} \|z-e^{i\theta}\tilde{z}\|_2, \quad \tilde{\theta}^\star = \arg\min_{\theta\in[0,2\pi)} \|\tilde{z}-e^{i\theta}z^\star\|_2.$$ Then, since $d_2(\tilde{z},z^\star) = \|z^\star-e^{-i\tilde{\theta}^\star}\tilde{z}\|_2 \le \sqrt{8\|\Delta\|_{\rm op}}$ by Fact \[lem:fix-pt\], we have $$\begin{aligned}
& & \|(\Sigma(z)-\Sigma(\tilde{z}))z\|_2 \nonumber \\
\noalign{\smallskip}
&\le& \sqrt{n} \cdot | (z^\star)^H(e^{-i\tilde{\theta}}z-\tilde{z}) | + \left( \|\Delta\|_{\rm op} + \frac{n}{\alpha} \right) d_2(z,\tilde{z}) \nonumber \\
\noalign{\smallskip}
&\le& \sqrt{n}\left( \|z^\star-e^{-i\tilde{\theta}^\star}\tilde{z}\|_2\cdot\|e^{-i\tilde{\theta}}z-\tilde{z}\|_2 + \left| |\tilde{z}^Hz| - n \right| \right) \nonumber \\
\noalign{\smallskip}
&& \,\,+\,\, \left( \|\Delta\|_{\rm op} + \frac{n}{\alpha} \right) d_2(z,\tilde{z}) \nonumber \\
\noalign{\smallskip}
&\le& \left( \sqrt{8n\|\Delta\|_{\rm op}} + \|\Delta\|_{\rm op} + \frac{n}{\alpha} \right) d_2(z,\tilde{z}) + \frac{\sqrt{n}}{2}d_2(z,\tilde{z})^2. \label{eq:2crit-1}
\end{aligned}$$
Now, let $\tilde{u}=\left( I-\tfrac{1}{n}\tilde{z}\tilde{z}^H \right)(e^{-i\tilde{\theta}}z-\tilde{z})$ be the projection of $e^{-i\tilde{\theta}}z-\tilde{z}$ onto the orthogonal complement of $\mbox{span}(\tilde{z})$. Then, similar to the derivation of , we have $$\begin{aligned}
& & \|\Sigma(\tilde{z})z\|_2 \,\,\,=\,\,\, \|\Sigma(\tilde{z})\tilde{u}\|_2 \nonumber \\
\noalign{\smallskip}
&\ge& \left(|(z^\star)^H\tilde{z}| - \|\Delta\tilde{z}\|_\infty \right) \|\tilde{u}\|_2^2 - \left| \tilde{u}^H(z^\star-e^{-i\tilde{\theta}^\star}\tilde{z}) \right|^2 - \|\Delta\|_{\rm op}\|\tilde{u}\|_2^2 \nonumber \\
\noalign{\smallskip}
&\ge& \left( n - \|\Delta\tilde{z}\|_\infty - 13\|\Delta\|_{\rm op} \right) \left( d_2(z,\tilde{z}) - \frac{1}{2\sqrt{n}}d_2(z,\tilde{z})^2 \right). \label{eq:2crit-2}
\end{aligned}$$ Moreover, following the derivations of and , we have $$\begin{aligned}
d_2(z,\tilde{z}) &\le& \frac{\sqrt{n}}{2} + \sqrt{8\|\Delta\|_{\rm op}}, \label{eq:2crit-3} \\
\noalign{\smallskip}
\|\Delta\tilde{z}\|_\infty &\le& \|\Delta z^\star\|_\infty + \sqrt{8}\|\Delta\|_{\rm op}^{3/2}. \label{eq:2crit-4}
\end{aligned}$$ Upon putting together –, we obtain $$\begin{aligned}
\rho(z) &\ge& \left[ \frac{n}{4} - \frac{3\|\Delta z^\star\|_\infty}{4} - 4\sqrt{2n\|\Delta\|_{\rm op}} - \frac{43\|\Delta\|_{\rm op}}{4} - \frac{3\sqrt{2}\|\Delta\|_{\rm op}^{3/2}}{2} \right] d_2(z,\tilde{z}) \\
\noalign{\smallskip}
&\ge& \frac{n}{8}d_2(z,\tilde{z})
\end{aligned}$$ whenever $\|\Delta\|_{\rm op}\le\tfrac{n^{2/3}}{32768}$, $\|\Delta z^\star\|_\infty\le\tfrac{n}{24}$, and $\alpha\ge4$. This completes the proof.
Recall that a global maximizer of Problem is a second-order critical point. Now, under the assumptions of Proposition \[prop:EB-2crit\], we know that every second-order critical point $\tilde{z}$ of Problem satisfies $d_2(\tilde{z},z^\star) \le \sqrt{8\|\Delta\|_{\rm op}} \le \tfrac{\sqrt{n}}{2}$; see Fact \[lem:fix-pt\]. Thus, Proposition \[prop:EB-2crit\] shows that every second-order critical point of Problem is also a global maximizer, which is unique up to a global phase. This gives an alternative proof of [@boumal2016nonconvex Theorem 4] with a less restrictive requirement on $\|\Delta z^\star\|_\infty$ ($\|\Delta z^\star\|_\infty = O(n)$ in Proposition \[prop:EB-2crit\] vs. $\|\Delta z^\star\|_\infty = O(n^{2/3}\sqrt{\log n})$ in [@boumal2016nonconvex Theorem 4]). It remains an open question to determine whether the conclusion of Proposition \[prop:EB-2crit\] still holds under the same noise requirement as Proposition \[prop:err-bd\].
Now, let us proceed to the second part of the proof of Theorem \[thm:convergence-rate\]. Our goal is to prove the following proposition, which elucidates the key properties of Algorithm \[alg:GP\]:
\[prop:alg-prop\] Under the assumptions of Theorem \[thm:convergence-rate\], the sequence of iterates $\{z^k\}_{k\ge0}$ generated by Algorithm \[alg:GP\] satisfies the following for $k=0,1,\ldots$, where $a_0,a_1,a_2>0$ are quantities that depend only on $n$ and $\alpha$, and $\hat{z}$ is any global maximizer of Problem :
1. *(Sufficient Ascent)* $f(z^{k+1})-f(z^k) \ge a_0\cdot\|z^{k+1}-z^k\|_2^2$.
2. *(Cost-to-Go Estimate)* $f(\hat{z}) - f(z^k) \le a_1 \cdot d_2(z^k,\hat{z})^2$.
3. *(Safeguard)* $\rho(z^k) \le a_2\cdot \|z^{k+1}-z^k\|_2$.
We begin by proving (a). Recalling that $\tilde{C}=C+\tfrac{n}{\alpha}I$, we have $$\begin{aligned}
& & f(z^{k+1})-f(z^k) \\
\noalign{\smallskip}
&=& (z^{k+1}-z^k)^H\tilde{C}(z^{k+1}-z^k) - 2(z^k)^H\tilde{C}(z^k) + 2\Re\{ (z^{k+1})^H\tilde{C}(z^k) \}.
\end{aligned}$$ We claim that $\Re\{ (z^{k+1})^H\tilde{C}(z^k) \} \geq (z^k)^H \tilde{C} (z^k)$. This follows from the fact that $$(z^{k+1})^H\tilde{C}(z^k) = \left( \frac{\tilde{C}z^k}{|\tilde{C}z^k|} \right)^H\tilde{C}(z^k)$$ is a real number and $$z^{k+1}\in \arg\max_{z\in \mathbb{T}^n} \Re\{z^H\tilde{C}z^k\}.$$ Hence, by the assumption on $\alpha$, we have $$f(z^{k+1})-f(z^k) \ge (z^{k+1}-z^k)^H\tilde{C}(z^{k+1}-z^k) \ge a_0 \cdot \|z^{k+1}-z^k\|_2^2$$ with $a_0=\lambda_{\rm min}\left(\Delta+\frac{n}{\alpha}I\right)>0$.
Next, we prove (b). Let $\hat{\theta}_k=\arg\min_{\theta\in[0,2\pi)} \|z^k-e^{i\theta}\hat{z}\|_2$. Then, we have $$\begin{aligned}
f(\hat{z}) - f(z^k) &=& \hat{z}^H\tilde{C}\hat{z} - (z^k)^H\tilde{C}(z^k) \nonumber \\
\noalign{\smallskip}
&=& \|\tilde{C}\hat{z}\|_1 - (z^k)^H\tilde{C}(z^k) \label{eq:ctg-1} \\
\noalign{\smallskip}
&=& (z^k)^H \left( {\rm Diag}(|\tilde{C}\hat{z}|) - \tilde{C} \right) (z^k) \nonumber \\
\noalign{\smallskip}
&=& (e^{-i\hat{\theta}_k}z^k-\hat{z})^H \left( {\rm Diag}(|\tilde{C}\hat{z}|) - \tilde{C} \right) (e^{-i\hat{\theta}_k}z^k-\hat{z}) \label{eq:ctg-2} \\
\noalign{\smallskip}
&\le& \left( \|\tilde{C}\|_{\rm op} + \|\tilde{C}\hat{z}\|_\infty \right) d_2(z^k,\hat{z})^2, \nonumber
\end{aligned}$$ where both and follow from Fact \[lem:fix-pt\]. Now, observe that $$\begin{aligned}
& & \|\tilde{C}\|_{\rm op} + \|\tilde{C}\hat{z}\|_\infty \nonumber \\
\noalign{\smallskip}
&\le& \|C\|_{\rm op}+\|C\hat{z}\|_\infty + \frac{2n}{\alpha} \nonumber \\
\noalign{\smallskip}
&\le& \|(z^\star)(z^\star)^H\|_{\rm op} + \|\Delta\|_{\rm op} + \|(z^\star)(z^\star)^H\hat{z}\|_\infty + \|\Delta\hat{z}\|_\infty + \frac{2n}{\alpha} \nonumber \\
\noalign{\smallskip}
&\le& n+\|\Delta\|_{\rm op} + |(z^\star)^H\hat{z}| + \|\Delta z^\star\|_\infty + \frac{4\|\Delta\|_{\rm op}^2}{\sqrt{n}} + \frac{2n}{\alpha} \label{eq:ctg-3} \\
\noalign{\smallskip}
&\le& 2n + \|\Delta\|_{\rm op} + \|\Delta z^\star\|_\infty + \frac{4\|\Delta\|_{\rm op}^2}{\sqrt{n}} + \frac{2n}{\alpha} \label{eq:ctg-4} \\
\noalign{\smallskip}
&<& 3n, \label{eq:ctg-5}
\end{aligned}$$ where follows from and the fact that $\|(z^\star)(z^\star)^H\|_{\rm op}=n$, follows from and Fact \[lem:1\], and is due to the assumptions on $\alpha$, $\|\Delta\|_{\rm op}$, and $\|\Delta z^\star\|_\infty$. Hence, we conclude that $$f(\hat{z})-f(z^k)\le a_1\cdot d_2(z^k,\hat{z})^2$$ for some $a_1\in(0,3n)$.
Lastly, we prove (c). By definition of $z^{k+1}$, we have $${\rm Diag}(|\tilde{C}z^k|)(z^{k+1}-z^k) = \left( \tilde{C}-{\rm Diag}(|\tilde{C}z^k|) \right)z^k.$$ It follows that $$\rho(z^k) = \left\| {\rm Diag}(|\tilde{C}z^k|)(z^{k+1}-z^k) \right\|_2 \le \|\tilde{C}z^k\|_\infty\|z^{k+1}-z^k\|_2.$$ Now, recall from the proof of Theorem \[thm:2-norm\] that $d_2(z^k,z^\star)\leq \tfrac{\sqrt{n}}{2}$ for $k=0,1,\ldots$. Upon letting $\theta_k^\star=\arg\min_{\theta\in[0,2\pi)} \|z^k-e^{i\theta}z^\star\|_2$, we obtain $$\begin{aligned}
\|\tilde{C}z^k\|_\infty &\le& \|(z^\star)(z^\star)^Hz^k\|_\infty + \|\Delta z^k\|_\infty + \frac{n}{\alpha} \\
\noalign{\smallskip}
&\le& |(z^\star)^Hz^k| + \|\Delta(e^{-i\theta_k^\star}z^k-z^\star)\|_\infty + \|\Delta z^\star\|_\infty + \frac{n}{\alpha} \\
\noalign{\smallskip}
&\le& n + \|\Delta\|_{\rm op}\cdot d_2(z^k,z^\star) + \|\Delta z^\star\|_\infty + \frac{n}{\alpha} \\
\noalign{\smallskip}
&<& \frac{3n^{5/4}}{2},
\end{aligned}$$ where the last inequality is due to the assumptions on $\alpha$, $\|\Delta\|_{\rm op}$, and $\|\Delta z^\star\|_\infty$. It follows that $$\rho(z^k) \le a_2 \cdot \|z^{k+1}-z^k\|_2$$ for some $a_2\in\left( 0,\tfrac{3n^{5/4}}{2} \right)$.
Armed with Propositions \[prop:err-bd\] and \[prop:alg-prop\], we are now ready to prove Theorem \[thm:convergence-rate\].
\[Proof of Theorem \[thm:convergence-rate\]\] By Propositions \[prop:err-bd\] and \[prop:alg-prop\], we have $$\begin{aligned}
f(\hat{z}) - f(z^{k+1}) &=& \left( f(\hat{z}) - f(z^k) \right) - \left( f(z^{k+1}) - f(z^k) \right) \nonumber \\
\noalign{\smallskip}
&\le& a_1\cdot d_2(z^k,\hat{z})^2 - \left( f(z^{k+1}) - f(z^k) \right) \nonumber \\
\noalign{\smallskip}
&\le& \frac{64a_1}{n^2} \rho(z^k)^2 - \left( f(z^{k+1}) - f(z^k) \right) \nonumber \\
\noalign{\smallskip}
&\le& \frac{64a_1a_2^2}{n^2}\|z^{k+1}-z^k\|_2^2 - \left( f(z^{k+1}) - f(z^k) \right) \nonumber \\
\noalign{\smallskip}
&\le& \left( \frac{64a_1a_2^2}{a_0n^2} - 1 \right) \left( f(z^{k+1}) - f(\hat{z}) + f(\hat{z}) - f(z^k) \right).
\end{aligned}$$ Since $f(\hat{z})\ge f(z^k)$ for $k=0,1,\ldots$, we may assume without loss of generality that $a'=\tfrac{64a_1a_2^2}{a_0n^2}>1$. It then follows that $$f(\hat{z})-f(z^{k+1}) \le \frac{a'-1}{a'}\left( f(\hat{z}) - f(z^k) \right),$$ which yields $$ f(\hat{z})-f(z^k) \le \left( f(\hat{z})-f(z^0) \right) \lambda^k$$ with $\lambda=\tfrac{a'-1}{a'}\in(0,1)$. Furthermore, we have $$\begin{aligned}
d_2(z^k,\hat{z})^2 &\le& \frac{64}{n^2}\rho(z^k)^2 \\
\noalign{\smallskip}
&\le& \frac{64a_2^2}{n^2} \|z^{k+1}-z^k\|_2^2 \\
\noalign{\smallskip}
&\le& \frac{64a_2^2}{a_0n^2}\left( f(z^{k+1})-f(z^k) \right) \\
\noalign{\smallskip}
&\le& \frac{64a_2^2}{a_0n^2}\left( f(\hat{z})-f(z^k) \right) \\
\noalign{\smallskip}
&\le& \frac{64a_2^2}{a_0n^2} \left( f(\hat{z})-f(z^0) \right) \lambda^k,
\end{aligned}$$ which implies that $$d_2(z^k,\hat{z}) \le a \left( f(\hat{z})-f(z^0) \right)^{1/2} \lambda^{k/2}$$ with $a=\sqrt{\tfrac{64a_2^2}{a_0n^2}}$. This completes the proof.
Again, we can specialize Theorem \[thm:convergence-rate\] to the Gaussian noise setting. This leads to the following corollary, which can be proven by combining Theorem \[thm:convergence-rate\] with the probabilistic estimates in [@BBS16 Proposition 3.3]; cf. Corollary \[cor:1\]:
\[cor:global-conv\] Suppose that the measurement noise $\Delta$ takes the form $\Delta=\sigma W$, where $\sigma^2>0$ is the noise power satisfying $\sigma\in\left(0,\tfrac{n^{1/4}}{936}\right]$ and $W\in\mathbb{H}^n$ is a Wigner matrix. Suppose further that the step size $\alpha$ satisfies $\alpha\in\left[ 4,312n^{1/4} \right)$ and the initial point $z^0$ is given by $z^0=v_C$. Then, with probability at least $1-2n^{-5/4}-2e^{-n/2}$, the sequence of iterates $\{z^k\}_{k\ge0}$ generated by Algorithm \[alg:GP\] satisfies $$\begin{aligned}
f(\hat{z}) - f(z^k) &\le& \left( f(\hat{z})-f(z^0) \right) \lambda^k, \\
\noalign{\smallskip}
d_2(z^k,\hat{z}) &\leq& a \left( f(\hat{z})-f(z^0) \right)^{1/2} \lambda^{k/2}
\end{aligned}$$ for $k=0,1,\ldots$, where $a>0,\lambda\in(0,1)$ are quantities that depend only on $n$ and $\alpha$, and $\hat{z}$ is any global maximizer of Problem .
Corollary \[cor:global-conv\] shows that in the Gaussian noise setting, Algorithm \[alg:GP\] will converge to a global maximizer of Problem at a linear rate with high probability for noise level up to $\sigma=O(n^{1/4})$. This matches the noise level requirement for the tightness of the SDR-based method established in [@BBS16 Theorem 2.1]. As the GPM typically has lower complexity than the SDR-based method in tackling Problem , we see that the former is competitive with the latter in terms of both theoretical guarantees and numerical efficiency.
Conclusion
==========
In this paper, we conducted a comprehensive analysis of the estimation and convergence performance of the GPM for tackling the phase synchronization problem. First, under the assumption that the measurement noise $\Delta$ satisfies $\|\Delta\|_{\rm op}=O(n)$, we established bounds on the rates of decrease in the $\ell_2$- and $\ell_\infty$-estimation errors of the iterates generated by the GPM. As a corollary, we showed that in the Gaussian noise setting (i.e., $\Delta=\sigma W$, where $\sigma>0$ is the noise level and $W$ is a Wigner matrix), the expected squared $\ell_2$-estimation errors of the iterates are decreasing and all are on the same order as that of the MLE even when the noise level is $\sigma=O(n^{1/2})$. The above result holds regardless of whether the iterates converge or not and yields the best provable bound on the estimation error of any accumulation point generated by the GPM under the least restrictive noise requirement currently known. Second, we showed that when the measurement noise $\Delta$ and target phase vector $z^\star$ satisfy $\|\Delta\|_{\rm op}=O(n^{3/4})$ and $\|\Delta z^\star\|_\infty=O(n)$, the GPM will converge linearly to a global maximizer of Problem . This not only resolves an open question in [@boumal2016nonconvex] concerning the convergence *rate* of the GPM but also improves upon the noise requirement $\|\Delta\|_{\rm op}=O(n^{2/3})$ and $\|\Delta z^*\|_\infty=O(n^{2/3}\sqrt{\log n})$ that is imposed in [@boumal2016nonconvex] to establish just the convergence of the GPM. Our result implies that in the Gaussian noise setting, the GPM will converge linearly to a global maximizer of Problem in the noise regime $\sigma=O(n^{1/4})$. This is the same regime for which the computationally heavier SDR-based method in [@BBS16] is provably tight. To establish our convergence rate result, we developed a new error bound for the non–convex problem . As a by-product, we showed that every second-order critical point of Problem is globally optimal if $\|\Delta\|_{\rm op}=O(n^{2/3})$ and $\|\Delta z^*\|_\infty=O(n)$. This slightly improves upon the corresponding result in [@boumal2016nonconvex]. An interesting future direction would be to extend the GPM and the machinery developed in this paper to design and analyze first-order methods for other (non-convex) quadratic optimization problems.
Acknowledgement {#acknowledgement .unnumbered}
===============
We thank Nicolas Boumal for his helpful comments on an earlier version of our manuscript.
[^1]: Department of Systems Engineering and Engineering Management, The Chinese University of Hong Kong, Shatin, N. T., Hong Kong. E-mail: [[email protected]]{}
[^2]: Department of Systems Engineering and Engineering Management, The Chinese University of Hong Kong, Shatin, N. T., Hong Kong. E-mail: [[email protected]]{}
[^3]: Department of Systems Engineering and Engineering Management, and, by courtesy, CUHK-BGI Innovation Institute of Trans-omics, The Chinese University of Hong Kong, Shatin, N. T., Hong Kong. E-mail: [[email protected]]{}
[^4]: Throughout the paper, the term “phase” refers to a complex number with unit modulus.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that one-sided Alexandrov embedded constant mean curvature cylinders of finite type in the 3-sphere are surfaces of revolution. This confirms a conjecture by Pinkall and Sterling that the only embedded constant mean curvature tori in the 3-sphere are rotational.'
address:
- 'Department of Mathematics, University College Cork, Cork, Ireland'
- 'Institut für Mathematik, Universität Mannheim, 68131 Mannheim, Germany'
author:
- 'M. Kilian'
- 'M. U. Schmidt'
title: 'On the moduli of constant mean curvature cylinders of finite type in the 3-sphere'
---
[^1]
Introduction
Alexandrov [@Ale0] proved that there are no compact embedded surfaces with constant mean curvature () in Euclidean 3-space ${\mathbb R}^3$ other than round spheres. However, while there are no compact minimal surfaces in ${\mathbb R}^3$, there is an abundance of such in the 3-sphere ${\mathbb S}^3$. For instance 2-spheres in the 3-sphere are minimal precisely when they are great 2-spheres, and Lawson proved that compact embedded minimal surfaces in ${\mathbb S}^3$ exist for every genus [@Law:compact; @Law:S3]. Lawson further showed [@Law:unknot] that any embedded minimal torus in ${\mathbb S}^3$ is unknotted, and conjectured that up to isometry the Clifford torus is the only embedded minimal torus in ${\mathbb S}^3$. Hsiang and Lawson [@HsiL] proved that the only embedded minimal torus of revolution is the Clifford torus. Further results suggest that an embedded minimal torus indeed has additional symmetries: Montiel and Ros [@MonR:min] showed that the only minimal torus immersed into ${\mathbb S}^3$ by the first eigenfunctions is the Clifford torus, and Ros [@Ros] proved that the normal surface of an embedded minimal torus in ${\mathbb S}^3$ is also embedded. Various methods for obtaining minimal surfaces in ${\mathbb S}^3$ have been employed to study specific classes, as in Karcher, Pinkall and Sterling [@KarPS], and more recently by Kapouleas and Yang [@KapY].
Wente’s discovery [@Wen] of tori provided the first compact examples other than spheres in Euclidean 3-space. The studies of Abresch [@Abr; @Abr:twi], Wente [@Wen:twi] and Walter [@Wal] on special classes of tori in ${\mathbb R}^3$ concluded in the classification by Pinkall and Sterling [@PinS], and their algebro-geometric description by Bobenko [@Bob:tor; @Bob:cmc]. In fact, Bobenko gave explicit formulas for tori in ${\mathbb R}^3,\,{\mathbb S}^3$ and hyperbolic 3-space $\mathbb{H}^3$ in terms of theta–functions, and provided a unified description of tori in the 3-spaceforms in terms of algebraic curves and spectral data. Independently, Hitchin [@Hit:tor] classified harmonic 2-tori in the 3-sphere, and thus also as a special case the harmonic Gauss maps of tori. These ideas culminated in the description of harmonic tori in symmetric spaces by Burstall, Ferus, Pedit and Pinkall [@BurFPP], and the generalized Weierstra[ß]{} representation by Dorfmeister, Pedit and Wu [@DorPW].
Associated to a torus in the 3-sphere is a hyperelliptic Riemann surface, the so called spectral curve. The structure equation for tori is the $\sinh$-Gordon equation. Hitchin [@Hit:tor], and Pinkall and Sterling [@PinS] independently proved that all doubly periodic solutions of the $\sinh$-Gordon equation correspond to spectral curves of finite genus. The genus of the spectral curve is called spectral genus. Ercolani, Knörrer and Trubowitz [@ErcKT] proved that for every even spectral genus $g\geq 2$ there exists a hyperelliptic curve which corresponds to an immersed torus in ${\mathbb R}^3$. The remaining cases of odd genera $g>1$ was settled by Jaggy [@Jag]. Adapting these results, Carberry [@Car:thesis] showed that minimal tori in ${\mathbb S}^3$ exist for every spectral genus. Note that while a torus in ${\mathbb R}^3$ has at least spectral genus 2, there is no such restriction for tori in ${\mathbb S}^3$. In particular, tori of revolution in ${\mathbb S}^3$ have spectral genus $g \leq
1$. Pinkall and Sterling [@PinS] conjectured that the only embedded tori in ${\mathbb S}^3$ are tori of revolution.
In parallel the global theory of embedded (especially minimal) surfaces in space forms was developed using geometric methods. Meeks [@Mee] proved that a properly embedded end of a surface is cylindrically bounded, which was used by Korevaar, Kusner and Solomon [@KorKS] to prove that the only embedded cylinders in ${\mathbb R}^3$ are surfaces of revolution - either a standard round cylinder (spectral genus $g=0$), or a Delaunay unduloid (spectral genus $g=1$). There are analogous surfaces of revolution in ${\mathbb S}^3$, some of which close up into tori, see Figure \[fig:tori\] for some simple examples.
![Stereographic projections of constant mean curvature tori of revolution in $\mathbb{S}^3$: On the left the Clifford torus, an embedded minimal torus with spectral genus 0. In the middle, an embedded non-minimal torus of spectral genus 1. On the right, a cutaway view of a non-embedded minimal torus of spectral genus 1. Images were created with cmclab [@Sch:cmclab].[]{data-label="fig:tori"}](clifford.eps "fig:") ![Stereographic projections of constant mean curvature tori of revolution in $\mathbb{S}^3$: On the left the Clifford torus, an embedded minimal torus with spectral genus 0. In the middle, an embedded non-minimal torus of spectral genus 1. On the right, a cutaway view of a non-embedded minimal torus of spectral genus 1. Images were created with cmclab [@Sch:cmclab].[]{data-label="fig:tori"}](4_lobes.eps "fig:") ![Stereographic projections of constant mean curvature tori of revolution in $\mathbb{S}^3$: On the left the Clifford torus, an embedded minimal torus with spectral genus 0. In the middle, an embedded non-minimal torus of spectral genus 1. On the right, a cutaway view of a non-embedded minimal torus of spectral genus 1. Images were created with cmclab [@Sch:cmclab].[]{data-label="fig:tori"}](3_lobes_minimal_cutaway.eps "fig:")
Kapouleas [@Kap1; @Kap2; @Kap3] proved the existence of compact surfaces in ${\mathbb R}^3$ for any genus greater than 1, as well as many new classes of non-compact surfaces in ${\mathbb R}^3$, but not much is known about the moduli of surfaces in general. Progress on understanding the moduli of immersions of punctured spheres has recently been made in the case of three punctures by Grosse-Brauckmann, Kusner and Sullivan [@GroKS:Tri] and by Schmitt et. al. [@SKKR]. Kusner, Mazzeo and Pollack [@KusMP] show that the moduli space of surfaces is an analytic variety. The local linearization of the moduli space is described by Jacobi fields which correspond to a normal variation of the surface which preserve the constant mean curvature property. Recently Korevaar, Kusner and Ratzkin [@KorKR] studied Jacobi fields on a class of surfaces with the additional property of being Alexandrov immersed. An Alexandrov immersed surface in ${\mathbb R}^3$ is a complete noncompact properly immersed surface $f:\Sigma \to {\mathbb R}^3$ that is the boundary of a 3-manifold $M$ with two additional features: The mean curvature normal of $\Sigma$ points into $M$, and $f$ extends to a proper immersion of $M$ into ${\mathbb R}^3$. When the target is the 3-sphere, we replace properness by completeness, and as Lawson [@Law:unknot] we consider in analogy a smooth immersion $f:\Sigma \to {\mathbb S}^3$ that we call a one-sided Alexandrov embedding if $\Sigma$ is the boundary of a connected 3-manifold $M$ and the following two conditions hold: The mean curvature of $\Sigma$ with respect to the inward normal is non-negative. Secondly, the manifold $M$ is complete with respect to the metric induced by $f$. We prove that the property of one-sided Alexandrov embeddedness is stable under continuous deformation, which allows us to study continuous families of one-sided Alexandrov embedded surfaces.
In this paper we consider cylinders which have constant Hopf differential, and whose metric is a periodic solution of the $\sinh$-Gordon equation of finite type. Such cylinders are said to be of finite type. We describe such finite type cylinders in Section \[sec:deformation\] and \[sec:moduli\] by spectral data, and show that the spectral data can be deformed in such a way that the corresponding family of surfaces are all topologically cylinders. It turns out that the corresponding moduli space of spectral data of genus $g$ is $g+1$-dimensional. Furthermore, we can control the spectral genus under the deformation, and by successively coalescing branch points of the spectral curve, we continuously deform the spectral curve in Lemma \[thm:connected\] into a curve of genus zero. In Section \[sec:AE space\] we show that one-sided Alexandrov embedded surfaces with constant mean curvature have collars with depths uniformly bounded from below. For this purpose we use a ’maximum principle at infinity’ which was communicated to us by Harold Rosenberg [@Ros:com]. This allows us to show in Theorem \[thm:spec\_AE\] that a large class of continuous deformations of cylinders of finite type preserve the one-sided Alexandrov embeddedness. In Lemma \[thm:continuous genus 1\] we continuously deform any one-sided finite type cylinder in ${\mathbb S}^3$ into a one-sided Alexandrov embedded flat cylinder in ${\mathbb S}^3$ with spectral genus zero. These are classified in Theorem \[thm:onesided revolution\]. Finally this classification is extended to all possible deformations of these flat cylinders in ${\mathbb S}^3$ in Theorem \[thm:main1\]. Since an embedded torus in the 3-sphere is covered by a one-sided Alexandrov embedded cylinder, our result confirms the conjecture by Pinkall and Sterling, and since the only embedded minimal torus of revolution is the Clifford torus, also affirms Lawson’s conjecture.
**Acknowledgments.** We thank Fran Burstall, Karsten Grosse-Brauckmann, Ian McIntosh, Rob Kusner, Franz Pedit and Ulrich Pinkall for useful discussions. This work was mostly carried out whilst M Kilian was a research assistant at the University of Mannheim, and he would like to thank the Institute of Mathematics there for providing excellent research conditions.
We had several beneficial conversations with Laurent Hauswirth, and we are especially grateful to Antonio Ros and Harold Rosenberg who helped us close a gap in a first draft of this paper, and Univerité Paris 7 for its hospitality during these discussions.
Conformal cmc immersions into ${\mathbb S}^3$ {#sec:CONF}
=============================================
This preliminary section recalls the relationship between immersed surfaces in ${\mathbb S}^3$ and solutions of the $\sinh$-Gordon equation, before considering the special case of cylinders in ${\mathbb S}^3$, the notion of monodromy and the period problem.
The $\sinh$-Gordon equation
---------------------------
We identify the 3-sphere ${\mathbb S}^3 \subset {\mathbb R}^4$ with ${\mathbb S}^3 \cong
{\mathrm{SU}_{\mbox{\tiny{$2$}}}}$. The Lie algebra of the matrix Lie group ${\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ is ${\mathfrak{su}_{\mbox{\tiny{2}}}}$, equipped with the commutator $[\,\cdot,\,\cdot\,]$. For $\alpha,\,\beta \in \Omega^1({\mathbb R}^2,{\mathfrak{su}_{\mbox{\tiny{2}}}})$ smooth $1$–forms on ${\mathbb R}^2$ with values in ${\mathfrak{su}_{\mbox{\tiny{2}}}}$, we define the ${\mathfrak{su}_{\mbox{\tiny{2}}}}$–valued $2$–form $$[ \alpha \wedge \beta ] (X,\,Y) =
[\alpha(X),\,\beta(Y)] - [\alpha(Y),\,\beta(X)],$$ for vector fields $X,\,Y$ on ${\mathbb R}^2$. Let $L_g:h \mapsto gh$ be left multiplication in ${\mathrm{SU}_{\mbox{\tiny{$2$}}}}$. Then by left translation, the tangent bundle is $T{\mathrm{SU}_{\mbox{\tiny{$2$}}}}\cong {\mathrm{SU}_{\mbox{\tiny{$2$}}}}\times {\mathfrak{su}_{\mbox{\tiny{2}}}}$ and $\theta : T{\mathrm{SU}_{\mbox{\tiny{$2$}}}}\to {\mathfrak{su}_{\mbox{\tiny{2}}}},\,
v_g \mapsto dL_{g^{-1}}(v_g)$ is the (left) Maurer–Cartan form. It satisfies the [**[Maurer-Cartan-equations]{}**]{} $$\label{eq:MC_equation}
2\, d\theta + [ \theta \wedge \theta ] = 0.$$ For a map $F:{\mathbb R}^2 \to {\mathrm{SU}_{\mbox{\tiny{$2$}}}}$, the pullback $\alpha = F^{\ast}\theta$ also satisfies , and conversely, every solution $\alpha \in \Omega^1({\mathbb R}^2,{\mathfrak{su}_{\mbox{\tiny{2}}}})$ of integrates to a smooth map $F:{\mathbb R}^2 \to {\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ with $\alpha =
F^{\ast}\theta$.
Complexifying the tangent bundle $T{\mathbb R}^2\cong T{\mathbb C}$ and decomposing into $(1,0)$ and $(0,1)$ tangent spaces, and writing $d=\partial +
\bar{\partial}$, we may split $\omega \in \Omega^1(M,{\mathfrak{su}_{\mbox{\tiny{2}}}})$ into the $(1,0)$ part $\omega^\prime$, the $(0,1)$ part $\omega^{\prime
\prime}$ and write $\omega = \omega^\prime + \omega^{\prime\prime}$. We set the $*$–operator on $\Omega^1(M,{\mathfrak{sl}_{\mbox{\tiny{$2$}}}})$ to $*\omega = - i
\omega^\prime + i \omega^{\prime \prime}$.
We denote by $\langle \cdot \, , \cdot \rangle$ the bilinear extension of the Ad–invariant inner product $(X,\,Y) \mapsto
-\tfrac{1}{2} {\mathrm{tr}}(XY)$ of ${\mathfrak{su}_{\mbox{\tiny{2}}}}$ to ${\mathfrak{su}_{\mbox{\tiny{2}}}}^{\mbox{\tiny{${\mathbb C}$}}} =
{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})$. The double cover of the isometry group $\mathrm{SO}(4)$ is ${\mathrm{SU}_{\mbox{\tiny{$2$}}}}\times {\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ via the action $((F,G),X) \mapsto FXG^{-1}$.
Now let $g:{\mathbb R}^2 \to {\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ be an immersion and $\omega = g^{-1}dg =
\omega' + \omega''$. Then $g$ is conformal if and only if the $(1,0)$-part of $\omega$ is isotropic $$\label{eq:conformal}
\langle \omega' ,\, \omega' \rangle = 0\,.$$ If $g$ is a conformal immersion then there exists a smooth function $v:{\mathbb R}^2 \to {\mathbb R}^\times = {\mathbb R}\setminus \{ 0 \}$, called the [**[conformal factor]{}**]{} of $g$ such that $$\label{eq:conformal_factor}
v^2 = 2\,\langle \omega',\,\omega'' \rangle\,.$$ The mean curvature function $H$ of $g$ (see e.g. [@SKKR]) is given by $$\label{eq:H_S3}
2\,d*\omega = H \,[ \omega \wedge \omega ].$$ Recall the following observation of Uhlenbeck [@Uhl], based on an earlier result by Pohlmeyer [@Poh], and suppose that in the following $g:{\mathbb R}^2 \to {\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ is a conformal immersion with non-zero constant mean curvature $H_0$ and conformal factor $v$. Then and $2d\omega + [\,\omega \wedge \omega \,]=0$ combined give $d\omega+ H_0^{-1} d*\omega =0$, or alternatively $$\label{eq:w'}
(1-iH_0^{-1})\,d\omega' + (1+iH_0^{-1})\,d\omega'' = 0.$$ Inserting $d\omega'' = -d\omega' - [\,\omega' \wedge \omega'' \,]$ respectively $d\omega' = -d\omega'' - [\,\omega' \wedge \omega''
\,]$ into gives $2d\omega' = (iH_0-1)[\,\omega' \wedge
\omega''\,]$ and $2d\omega'' = -(1+iH_0)[\,\omega' \wedge
\omega''\,]$. Then an easy computation shows that $$\alpha_\lambda = \tfrac{1}{2}(1+\lambda^{-1})(1+iH_0)\,\omega'
+ \tfrac{1}{2}(1+\lambda)(1-iH_0)\,\omega''$$ satisfies the Maurer-Cartan-equations $$2\,d\alpha_\lambda + [\,\alpha_\lambda \wedge \alpha_\lambda \,] =
0 \mbox{ for all $\lambda \in {\mathbb C}^{\times} = {\mathbb C}\setminus \{ 0\}$. }$$ The Maurer-Cartan-equations are an integrability condition, so we can integrate and obtain a corresponding [**[extended frame]{}**]{} $F_\lambda : {\mathbb R}^2 \times {\mathbb C}^\times \to {\mathrm{SL}_{\mbox{\tiny{$2$}}}}({\mathbb C})$ with $dF_\lambda =
F_\lambda\,\alpha_\lambda$ and $F_\lambda(0) = \mathbbm{1}$. Since $\omega$ takes values in ${\mathfrak{su}_{\mbox{\tiny{2}}}}$, we conclude that $F_\lambda$ takes values in ${\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ when $\lambda \in {\mathbb S}^1$. Now define for $\lambda_0,\,\lambda_1 \in {\mathbb S}^1,\, \lambda_0 \neq \lambda_1$, the following map $f:{\mathbb R}^2 \to {\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ by the [**[Sym-Bobenko-formula]{}**]{} $$\label{eq:Sym_S3}
f = F_{\lambda_1} F_{\lambda_0}^{-1}\,.$$ Then for $\Omega = f^{-1}df = \mathrm{Ad}\,F_{\lambda_0} \left(
\alpha_{\lambda_1} - \alpha_{\lambda_0} \right)$ we obtain $\Omega' = \tfrac{1}{2}(\lambda_1^{-1} - \lambda_0^{-1})(1+iH_0)\,
\mathrm{Ad}\,F_{\lambda_0} \,\omega'$ so the conformality of $f$ follows from the conformality of $g$, since $$\begin{aligned}
\langle \Omega',\,\Omega' \rangle &=
\tfrac{1}{4}(\lambda_1^{-1} - \lambda_0^{-1})^2(1+iH_0)^2\,
\mathrm{Ad}\,F_{\lambda_0} \,\langle \omega',\,\omega' \rangle =
0\\
2\,\langle \Omega',\,\Omega'' \rangle &=
\sin^2(t_1-t_0)(1+H_0^2)\,v^2\,. \label{eq:conf_Om}\end{aligned}$$ Here we have written $\lambda_{0,1} = e^{2it_{0,1}}$, and $v$ is the conformal factor of $g$. Furthermore $$\begin{aligned}
d\ast\Omega &= \tfrac{i}{4}(1+H_0^2)
\left( \lambda_0^{-1}\lambda_1 - \lambda_0\lambda_1^{-1}
\right) \mathrm{Ad}\,F_{\lambda_0} [\omega' \wedge \omega'']\,, \\
[\Omega \wedge \Omega] &=
\tfrac{1}{2}(1+H_0^2)(\lambda_1^{-1}-\lambda_0^{-1})(\lambda_1 -
\lambda_0)\,\mathrm{Ad}\,F_{\lambda_0}\,[\omega' \wedge
\omega'']\,.\end{aligned}$$ Hence by , the map $f:{\mathbb R}^2 \to {\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ given by has constant mean curvature $$\label{eq:H_mu}
H = i \,\frac{\lambda_0 + \lambda_1}{\lambda_0 -\lambda_1}\,.$$ In summary, by starting with one non-minimal conformal immersion $g$, we have just seen how to obtain a whole ${\mathbb C}^\times$-family of solutions of the Maurer-Cartan-equations, and from the corresponding extended frame we then obtained another conformal immersion $f$. Since the mean curvature and the conformal factor of $f$ in only depend on the angle between $\lambda_0,\,\lambda_1$, we in fact get a whole ${\mathbb S}^1$-family of isometric conformal immersions, called an [**[associated family]{}**]{}, which is obtained by simultaneously rotating $\lambda_0,\,\lambda_1$ while keeping the angle between them fixed.
We next recall the following version of Theorem 14.1 in Bobenko [@Bob:cmc], which provides a correspondence between solutions of the $\sinh$-Gordon equation and associated families of surfaces in the 3-sphere.
[@Bob:cmc] \[thm:sinh\] Let $u:{\mathbb R}^2 \to {\mathbb R}$ be a smooth function and define $$\label{eq:general_alpha}
\alpha_\lambda = \frac{1}{2}\,\begin{pmatrix}
u_z\,dz-u_{\bar{z}}\,d\bar{z} &
i\,\lambda^{-1}e^u\,dz + i\,e^{-u}\,d\bar{z}\\
i\,e^{-u}\,dz + i\,\lambda\,e^{u}\,d\bar{z}&
-u_z\,dz+u_{\bar{z}}\,d\bar{z}
\end{pmatrix}\,.$$ Then $2\,d\alpha_\lambda + [\,\alpha_\lambda \wedge
\alpha_\lambda\,] = 0$ if and only if $u$ is a solution of the $\sinh$-Gordon equation $$\label{eq:sinh-Gordon}
\partial \bar{\partial}\,2u + \sinh (2u) = 0.$$ For any solution $u$ of the $\sinh$-Gordon equation and corresponding extended frame $F_\lambda$, and $\lambda_0,\,\lambda_1
\in {\mathbb S}^1,\, \lambda_0 \neq \lambda_1$, the map defined by the Sym-Bobenko-formula is a conformal immersion with constant mean curvature $H$ , conformal factor $v =
e^u/\sqrt{H^2 + 1}$, and constant Hopf differential $Q\,dz^2$ with $Q = i\,(\lambda_1^{-1}-\lambda_0^{-1})/4$.
Decomposing $\alpha_\lambda = \alpha'_\lambda\,dz +
\alpha''_\lambda\,d\bar{z}$ into $(1,\,0)$ and $(0,\,1)$ parts, we compute $$\begin{split}
&\bar{\partial} \alpha'_\lambda = \frac{1}{2}\,
\begin{pmatrix} u_{z\bar{z}} & i\lambda^{-1}u_{\bar{z}}e^u \\
-iu_{\bar{z}}e^{-u} & -u_{z\bar{z}} \end{pmatrix}\,,
\quad \partial \alpha''_\lambda = \frac{1}{2}\,
\begin{pmatrix} -u_{z\bar{z}} & -iu_z e^{-u} \\
i\lambda u_z e^u & u_{z\bar{z}} \end{pmatrix}\,, \\
&\left[ \alpha'_\lambda,\,\alpha''_\lambda \right] =
\frac{1}{4} \begin{pmatrix} -e^{2u} + e^{-2u} &
2iu_{\bar{z}}\lambda^{-1}e^u + 2iu_ze^{-u} \\
-2i\lambda u_ze^u - 2iu_{\bar{z}}e^{-u} &
e^{2u} - e^{-2u} \end{pmatrix}\,. \end{split}$$ Now $2\,d\alpha_\lambda + [\,\alpha_\lambda \wedge \alpha_\lambda\,]
= 0$ is equivalent to $\bar{\partial}\alpha'_\lambda - \partial
\alpha''_\lambda = [ \alpha'_\lambda,\,\alpha''_\lambda ]$, which holds if and only if $u$ solves the sinh-Gordon equation .
If $u$ is a solution of the sinh-Gordon equation, then we may integrate $dF_\lambda = F_\lambda\,\alpha_\lambda$ to obtain a map $F_\lambda:{\mathbb R}^2 \times \mathbb{S}^1 \to {\mathrm{SU}_{\mbox{\tiny{$2$}}}}$. Let $\lambda_0,\,\lambda_1 \in \mathbb{S}^1,\, \lambda_0 \neq
\lambda_1$, and consider the map $f:{\mathbb R}^2 \to {\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ defined by the Sym-Bobenko-formula . Conformality is a consequence of the fact that the complexified tangent vector $$f^{-1}\partial f = \mathrm{Ad}\,F_{\lambda_0} (\alpha'_{\lambda_1} -
\alpha'_{\lambda_0}) =
\frac{i}{2}\,e^u \,(\lambda_1^{-1}-\lambda_0^{-1})\,
\mathrm{Ad}\,F_{\lambda_0}\,\bigl( \begin{smallmatrix}0&1\\0&0\end{smallmatrix}\bigr)$$ is isotropic with respect to the bilinear extension of the Killing form. The mean curvature can be computed using formula . The conformal factor is obtained from $$v^2 = 2\,\langle f^{-1}\partial f,\,f^{-1}\bar{\partial} f \rangle =
\tfrac{1}{4}\,e^{2u}\,(\lambda_1^{-1}-\lambda_0^{-1})
(\lambda_1-\lambda_0).$$ From we have $(H^2+1)(\lambda_1^{-1}-\lambda_0^{-1})
(\lambda_1-\lambda_0) = 4$, which proves the formula for the conformal factor.
Define the normal $N = F_{\lambda_1}
\,\varepsilon\,F_{\lambda_0}^{-1}$ with $\varepsilon = \bigl(
\begin{smallmatrix} i&0\\0&-i \end{smallmatrix} \bigr)$. Then $\partial N = F_{\lambda_1} (\alpha'_{\lambda_1}\, \varepsilon
- \varepsilon\,\alpha'_{\lambda_0})\, F_{\lambda_0}^{-1}$ and $$\alpha'_{\lambda_1}\,\varepsilon -
\varepsilon\,\alpha'_{\lambda_0} = \begin{pmatrix}
0&\tfrac{1}{2}e^u(\lambda_1^{-1}+\lambda_0^{-1})\\-e^{-u}&0
\end{pmatrix}\,.$$ Consequently, $Q := -\langle \partial \partial f,\, N \rangle =
\langle \partial f,\, \partial N \rangle =
\langle F_{\lambda_1}^{-1}\,\partial f \,F_{\lambda_0},\,
F_{\lambda_1}^{-1}\,\partial N \,F_{\lambda_0} \rangle
= \tfrac{i}{4}\,(\lambda_1^{-1} - \lambda_0^{-1})$, which proves the formula for the Hopf differential, and concludes the proof.
There is an analogous but more general theorem (see e.g Bobenko [@Bob:2x2]) than the one above, which asserts that if functions $(u,\,Q,\,H \equiv const.)$ satisfy the Gauss-Codazzi equations, then one obtains a ${\mathbb C}^\times$-family of solutions of the Maurer-Cartan-equations, thus an extended frame and consequently an associated family via the Sym-Bobenko-formula.
Monodromy and periodicity condition
-----------------------------------
The condition implies that the Hopf differential is a holomorphic quadratic differential [@Hop]. On the cylinder ${\mathbb C}^{\times}$ there is an infinite dimensional space of holomorphic quadratic differentials, large classes of which can be realized as Hopf differentials of cylinders [@KilMS]. On a torus the Hopf differential is constant (and non-zero). Since we are ultimately interested in tori, we restrict our attention to cylinders considered via Theorem \[thm:sinh\] which have constant non-zero Hopf differentials on the universal covering ${\mathbb C}$ of ${\mathbb C}^{\times}$. Note that for given solution $u$ of the sinh-Gordon equation an extended frame $\lambda \mapsto F_\lambda$ is holomorphic on ${\mathbb C}^{\times}$ and has essential singularities at $\lambda = 0,\,\infty$.
Let $F_\lambda$ be an extended frame for a immersion $f:{\mathbb R}^2 \to {\mathbb S}^3$ such that holds for two distinct unimodular numbers $\lambda_0,\,\lambda_1$. Let $\tau:{\mathbb R}^2
\to {\mathbb R}^2,\,z\mapsto z+\tau$ be a translation, and assume that $\alpha_\lambda = F_\lambda^{-1}dF_\lambda$ has period $\tau$, so that $\tau^* \alpha_\lambda = \alpha_\lambda \circ \tau =
\alpha_\lambda$. Then we define the [**[monodromy]{}**]{} of $F_\lambda$ with respect to $\tau$ as $$\label{eq:monodromy}
M_\lambda(\tau) = \tau^{\ast}(F_\lambda)\,F_\lambda ^{-1}\,.$$ Periodicity $\tau^{\ast} f = f$ in terms of the monodromy is then $\tau^{\ast} f = M_{\lambda_1}(\tau) F_{\lambda_1}
F_{\lambda_0}^{-1} M_{\lambda_0}^{-1}(\tau)$, so $\tau^{\ast} f = f$ if and only if $$\label{eq:periodicity}
M_{\lambda_0}(\tau) = M_{\lambda_1}(\tau) = \pm \mathbbm{1}\,.$$ If $\Delta(\lambda)$ is the trace of $M_{\lambda}(\tau)$ then $\tau^{\ast} f = f$ if and only if $\Delta(\lambda_0)=\Delta(\lambda_1) = \pm 2$.
Finite type solutions of the $\sinh$-Gordon equation {#sec:finite type}
====================================================
In this section we introduce the solutions of the $\sinh$-Gordon equation which are called finite type solutions. Finite type solutions of the $\sinh$-Gordon equation are in one-to-one correspondence with maps called polynomial Killing fields. These polynomial Killing fields take values in certain $2\times 2$-matrix polynomials, and solve a non-linear partial differential equation, but they are uniquely determined by one of their values. We shall call these values initial values of polynomial Killing fields or just initial values. The Symes method calculates the solutions in terms of the initial values with the help of a loop group splitting. The eigenvalues of these matrix polynomials define a real hyperelliptic algebraic curve, which is called [**[spectral curve]{}**]{}. One spectral curve corresponds to a whole family of finite type solutions of the $\sinh$-Gordon equation. We call the sets of finite type solutions (or their initial values), which belong to the same spectral curve, isospectral sets. The eigenspaces of the matrix polynomials define a holomorphic line bundle on the spectral curves called eigenbundle. These holomorphic line bundles completely determine the corresponding initial value and the corresponding solution of the $\sinh$-Gordon equation. Consequently, the isospectral sets can be identified with one connected component of the real part of the Picard group. In case the spectral curve has singularities, then the isospectral set can be identified with the real part of the compactification of a generalized Jacobian. These compactifications have stratifications, whose strata are the orbits under the action of the generalized Jacobian. In our case the spectral curves are hyperelliptic and we shall describe the corresponding stratifications of the isospectral sets.
Polynomial Killing fields {#sec:killing}
-------------------------
For some aspects of the theory untwisted loops are advantageous, and avoiding the additional covering map $\lambda \mapsto \sqrt\lambda$ simplifies for example the description of Bianchi-Bäcklund transformations by the simple factors [@TerU; @KilSS]. For the description of polynomial Killing fields on the other hand, the twisted loop algebras as in [@BurFPP; @BurP_adl; @BurP:dre; @DorPW; @McI:tor] are better suited, but we remain consistent and continue working in our ’untwisted’ setting.
Let $\varepsilon_+ = \bigl( \begin{smallmatrix} 0 & 1 \\ 0 & 0
\end{smallmatrix} \bigr)$ and $\varepsilon_- = \varepsilon_+^t$, and consider for $g \in {\mathbb N}_0$ the finite dimensional vector space $${\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}= \left\{
\xi=\sum_{d=-1}^{g}\xi_d\lambda^d
\mid\xi_{-1}\in\mathbb{C}\varepsilon_+,\, \xi_d=-\bar{\xi}^t_{g-1-d}\in{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})\mbox{ for }d=-1,\ldots,g
\right\}.$$ Clearly ${\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$ is a real $3g+2$-dimensional vector space and has up to isomorphism a unique norm $\|\cdot\|$. These Laurent polynomials define smooth mappings from $\lambda\in{\mathbb S}^1$ into ${\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})$. Note that $\sqrt{\lambda}\mapsto\lambda^{\frac{1-g}{2}}\xi$ belongs to the loop Lie algebra ${\Lambda {\mathfrak{su}_{\mbox{\tiny{2}}}}}$ of the loop Lie group ${\Lambda {\mathrm{SU}_{\mbox{\tiny{$2$}}}}}$. For the resulting solution of the $\sinh$-Gordon equation to be of finite type, we need in addition the conditions ${\mathrm{tr}}(\xi_{-1}\varepsilon_-)\neq 0
\neq {\mathrm{tr}}(\xi_0\varepsilon_+)$. These conditions ensure that $\xi_{-1}$ and the lower left entry of $\xi_0$ do not vanish, and therefore that $\xi_{-1}+{\mathrm{tr}}(\xi_0\varepsilon_+)\,\epsilon_-$ is semisimple. This is the same as semisimplicity of the leading order term in the twisted setting. We thus define $${\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}=\left\{\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}\mid
{\mathrm{tr}}(\xi_{-1}\varepsilon_-)\neq 0\neq{\mathrm{tr}}(\xi_0\varepsilon_+)
\right\}.$$ By the Symes method [@Symes_80], elucidated by Burstall and Pedit [@BurP_adl; @BurP:dre], the extended framing $F_\lambda:{\mathbb R}^2
\to \Lambda {\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ of a immersion of finite type is given by the unitary factor of the Iwasawa decomposition of $$\label{eq:FB}
\exp(z\,\xi) = F_\lambda\,B$$ for some $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ with $g\in\mathbb{N}_0$. Due to Pressly and Segal [@PreS], the Iwasawa decomposition is a diffeomorphism between the loop group ${\Lambda {\mathrm{SL}_{\mbox{\tiny{$2$}}}}}({\mathbb C})$ of ${\mathrm{SL}_{\mbox{\tiny{$2$}}}}({\mathbb C})$ into point wise products of elements of ${\Lambda {\mathrm{SU}_{\mbox{\tiny{$2$}}}}}$ with elements of the loop group $\Lambda^+{\mathrm{SL}_{\mbox{\tiny{$2$}}}}({\mathbb C})$ of holomorphic maps from $\lambda\in\mathbb{D}$ to ${\mathrm{SL}_{\mbox{\tiny{$2$}}}}({\mathbb C})$, which take at $\lambda=0$ values in the subgroup of ${\mathrm{SL}_{\mbox{\tiny{$2$}}}}({\mathbb C})$ of upper-triangular matrices with positive real diagonal entries. For every $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$ there exists a unique $\alpha(\xi)\in\Omega^1({\mathbb R}^2,\,\Lambda_{-1}^1{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C}))$, such that $\xi dz-\alpha(\xi)$ takes values in the Lie algebra of $\Lambda^+{\mathrm{SL}_{\mbox{\tiny{$2$}}}}({\mathbb C})$ of the right hand factor in the Iwasawa decomposition .
A [**[polynomial Killing field]{}**]{} is a map $\zeta:{\mathbb R}^2\to{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$ which solves $$\label{eq:pKf}
d\zeta =[\,\zeta,\,\alpha(\zeta)\,]\quad\mbox{ with }\quad
\zeta(0)=\xi.$$ For each [**initial value**]{} $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$, there exists a unique polynomial Killing field given by $$\label{eq:solution pk}
\zeta=B\xi B^{-1}=F^{-1}_\lambda\xi F_\lambda\quad
\mbox{ with }F_\lambda\mbox{ and }B\mbox{ as in \eqref{eq:FB}.}$$ For $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ with ${\mathrm{tr}}(\xi_{-1}\varepsilon_-)\in \mathbb{R}^+i$ and ${\mathrm{tr}}(\xi_0\varepsilon_+)\in\mathbb{R}^+i$ the corresponding $\Lambda_{-1}^1{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})$-valued 1-form $\alpha(\zeta)$ is the $\alpha$ as in for that particular solution $u$ of the $\sinh$-Gordon equation corresponding to the extended frame $F_\lambda$ of . For general initial values $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ the leading term $\zeta_{-1}+{\mathrm{tr}}(\zeta_0\varepsilon_+)\,\epsilon_-$ of the corresponding polynomial Killing field does not depend on the surface parameter $z$. The corresponding $\alpha(\zeta)$ differs from by multiplication of $\lambda$ and $dz$ with constant unimodular complex numbers. Given a polynomial Killing field $\zeta$, we set the initial value $\xi=\zeta|_{z=0}$ in . Thus $\zeta$, or the initial value $\xi$, gives rise to an extended frame, and thus to an associated family.
\[def:finite type\] A solution of the $\sinh$-Gordon equation is called a [**finite type**]{} solution if and only if it corresponds to a polynomial Killing field $\zeta:{\mathbb R}^2 \to {\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ with $g\in{\mathbb N}_0$.
Roots of polynomial Killing fields
----------------------------------
If an initial value $\xi$ has a root at some $\lambda=\alpha\in{\mathbb C}^{\times}$, then the corresponding polynomial Killing field has a root at the same $\lambda$ for all $z\in{\mathbb C}$. In this case we may reduce the order of $\xi$ and $\zeta$ without changing the corresponding extended frame $F$ . The following polynomials transform under $\lambda\mapsto\bar{\lambda}^{-1}$ as $$\begin{aligned}
\label{eq:transformation}
p(\lambda)&=\begin{cases}
i(\sqrt{\bar{\alpha}}\lambda-\sqrt{\alpha})
&\mbox{ for }\alpha\bar{\alpha}=1\\
(\lambda-\alpha)(1-\bar{\alpha}\lambda)&
\mbox{ for all }\alpha\in{\mathbb C}\end{cases}&
\overline{\lambda^{\deg(p)}p\left(\bar{\lambda}^{-1}\right)}&=p(\lambda).\end{aligned}$$ If the polynomial Killing field $\zeta$ with initial value $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ has a simple root at $\lambda=\alpha\in{\mathbb C}^{\times}$, then $\zeta/p$ does not vanish at $\alpha$ and is the polynomial Killing field with initial value $\xi/p\in\Lambda_{-1}^{g-\deg(p)}{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^\times$. Furthermore, obviously $\zeta$ and $\zeta/p$ commute, and we next show that both polynomial Killing fields $\zeta$ and $\zeta/p$ give rise to the same extended frame $F_\lambda$ .
\[th:pKf\_min\] If a polynomial Killing field $\zeta$ with initial value $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$ has zeroes in $\lambda\in{\mathbb C}^{\times}$, then there is a polynomial $p(\lambda)$, such that the following two conditions hold:
1. $\zeta/p$ is the polynomial Killing field with initial value $\xi/p\in\Lambda_{-1}^{g-\deg p}{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})$, which gives rise to the same associated family as $\zeta$.
2. $\zeta/p$ has no zeroes in $\lambda \in {\mathbb C}^{\times}$.
An appropriate Möbius transformation transforms any root $\alpha\in{\mathbb C}^{\times}$ into a negative root. For such negative roots the corresponding initial values $\xi$ and $\xi/p$ are related by multiplication with a polynomial with respect to $\lambda$ with positive coefficients. In the Iwasawa decomposition this factor is absorbed in $B$. Hence the corresponding extended frames coincide, which proves (i). Repeating this procedure for every root $\lambda \in {\mathbb C}^\times$ ensures (ii).
Hence amongst all polynomial Killing fields that give rise to a particular surface of finite type there is one of smallest possible degree (without adding further poles), and we say that such a polynomial Killing field has *minimal degree*. A polynomial Killing field has minimal degree if and only if it has neither roots nor poles in $\lambda \in {\mathbb C}^{\times}$. We summarize two results by Burstall and Pedit [@BurP_adl; @BurP:dre]. The first part is a variant of Theorem 4.3 in [@BurP_adl], the second part follows immediately from results in [@BurP:dre].
\[thm:PKF\] (i) A immersion $f:{\mathbb R}^2 \to \mathbb{S}^3$ is of finite type if and only if there exists a polynomial Killing field $\zeta$ with initial value $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ such that the map $F_\lambda$ obtained from is an extended frame of $f$.
\(ii) In particular there exists a unique polynomial Killing field of minimal degree that gives rise to $f$. Thus we have a smooth 1-1 correspondence between the set of immersions of finite type and the set of polynomial Killing fields without zeroes.
Point (i) is a reformulation of Theorem 4.3 in [@BurP_adl]. (ii) We briefly outline how to prove the existence and uniqueness of a minimal element.
If the initial value $\xi$ gives rise to $f$, then the corresponding polynomial Killing field $\zeta$ can be modified according to Proposition \[th:pKf\_min\] so that $\tilde{\zeta}$ is of minimal degree, and still giving rise to $f$. Hence there exists a polynomial of least degree giving rise to $f$.
For the uniqueness, assume we have two initial values $\xi,\,\tilde{\xi}$ of least degree $g$ both giving rise to $f$. Putting Proposition 3.3 and Corollary 3.8 in [@BurP:dre] together gives: Two finite type initial values give rise to the same associated family if and only if they commute and have equal residues. Since the residues coincide and both $\xi,\,\tilde{\xi}$ are of minimal degree, we conclude that $\xi \equiv \tilde{\xi}$. The unique minimal polynomial Killing field is thus $\zeta =
F^{-1}_\lambda\xi \,F_\lambda$.
Since the Iwasawa factorization is a diffeomorphism, and all other operations involved in obtaining an extended frame from $\zeta$ are smooth, the resulting surface depends smoothly on the entries of $\zeta$.
Spectral curves I
-----------------
Due to the characteristic equation $$\label{eq:characteristic0}
\det\left(\nu\,\mathbbm{1}-\zeta\right)=\nu^2+\det(\zeta)=0$$ of a polynomial Killing field $\zeta$ with initial value $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$ does not depend on $z\in{\mathbb C}$ and agrees with the characteristic equation of the initial value $\xi$. If $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$ then we may write $-\det\xi = \lambda^{-1}a$ for a polynomial $a$ of degree at most $2g$ which satisfies the reality condition $$\label{eq:a_reality}
\lambda^{2g} \overline{a(\bar{\lambda}^{-1})} =
-a(\lambda).$$ Consequently the hyperelliptic curve has three involutions $$\begin{aligned}
\label{eq:involutions}
\sigma&:(\lambda,\nu)\mapsto(\lambda,-\nu)&
\rho&:(\lambda,\nu)\mapsto(\bar{\lambda}^{-1},\bar{\lambda}^{-g}\bar{\nu})&
\eta&:(\lambda,\nu)\mapsto(\bar{\lambda}^{-1},-\bar{\lambda}^{-g}\bar{\nu})&\end{aligned}$$ If $a$ has $2g$ pairwise distinct roots, then $\nu^2 =
\lambda^{-1}a(\lambda)$ is a spectral curve of genus $g$ of a not necessarily periodic solution of the $\sinh$-Gordon equation. The genus $g$ is called the [**[spectral genus]{}**]{}.
\[thm:compact\] Let $a$ be a polynomial of degree $2g$ satisfying . Then the [**isospectral set**]{} $$\mathcal{K}_a = \left\{ \xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}\mid
\det \xi(\lambda) = -\lambda^{-1}a(\lambda) \right\}$$ is compact. Furthermore, if the $2g$ roots of $a$ are pair wise distinct, then $\mathcal{K}_a \cong \left({\mathbb S}^1\right)^g$.
For the compactness it suffices to show that all Laurent coefficients of a $\xi\in\mathcal{K}_a$ are bounded, since $\mathcal{K}_a$ is a closed subset of the $(3g+2)$-dimensional vector space ${\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$. For $d=(1-g)/2$ the product $\lambda^d\xi$ is skew hermitian on $|\lambda|=1$. The negative determinant of traceless skew hermitian $2 \times 2$ matrices is the square of a norm. Hence for all $\xi\in\mathcal{K}_a$ the Laurent polynomial of $\lambda^d\xi(\lambda)$ with respect to $\sqrt{\lambda}$ is bounded on $|\lambda |=1$. Thus the Laurent coefficients are bounded.
If $a$ has $2g$ pairwise distinct roots, then $\xi$ has no roots since at all roots of $\xi$, the determinant $\det \xi$ has a root of order two. If $\alpha$ is a root of $a$, then $\det\xi(\alpha)$ vanishes and $\xi(\alpha)$ is nilpotent. For a nonzero nilpotent $2\times 2$-matrix $\xi(\alpha)$ there exists a $2\times 2$-matrix $Q$ such that $\xi(\alpha) =[Q,\,\xi(\alpha)]$. Hence for every root $\alpha$ of $a$, there exists a $2\times 2$-matrix $Q$, such that $$\dot{\xi}(\lambda) = \frac{\xi^{-1}\det
\xi+[Q,\,\xi]}{\lambda-\alpha}=
\frac{-\xi+[Q,\,\xi]}{\lambda-\alpha}$$ has no pole at $\lambda=\alpha$. The corresponding derivative of $a=-\lambda\det(\xi)$ is equal to $\dot{a}=\frac{2a}{\lambda-\alpha}$. Furthermore, $\lambda\dot{\xi}$ is polynomial with respect to $\lambda$ of degree $g$. Two appropriate linear combinations with the analogous tangent element at the root $\bar{\alpha}^{-1}$ of $\xi$ change the roots $\alpha$ and $\bar{\alpha}^{-1}$ and fixes all other roots of $a$ and respects the reality condition of ${\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$. These two linear combinations belong to the tangent space of ${\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$. Hence the derivatives of all the coefficients of $a$ as functions on ${\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$ are non-zero at all $\xi\in\mathcal{K}_a$. By the implicit function theorem this set is therefore a $g$-dimensional submanifold. The corresponding eigenspaces of $\xi$ depend holomorphically on the solutions $(\lambda,\nu)$ of and define a holomorphic line bundle on the spectral curve. These [**[eigenbundles]{}**]{} have degree $g+1$, they are non-special in the sense that they have no holomorphic sections vanishing at one of the points at $\lambda=0$ or $\lambda=\infty$, and finally they obey some reality condition. Vice versa, all holomorphic line bundles obeying these three conditions correspond to one $\xi\in\mathcal{K}_a$ (see McIntosh [@McI:tor Section 1.4]). Hitchin has shown in [@Hit:tor], that the third condition implies the second condition. Therefore $\mathcal{K}_a$ can be identified with the real part of one connected component of the Picard group of the spectral curve, which is a $g$-dimensional torus.
If $a$ has multiple roots, then the real part of the Jacobian of the corresponding hyperelliptic curve still acts on $\mathcal{K}_a$, but not transitively. More precisely, in case of non-unimodular multiple roots of $a$ the set $\mathcal{K}_a$ has a stratification, whose strata are the orbits of the action of the real part of the generalized Jacobian of the singular hyperelliptic curve defined by $\nu^2=\lambda\,a(\lambda)$. The elements of different strata have different orders of zeroes at the multiple roots of $a$. In Proposition \[th:pKf\_min\] we have seen that all $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$ are products of $\tilde{\xi}\in\Lambda_{-1}^{\tilde{g}}{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})$ of lower degree $\tilde{g}<g$ with polynomials of the form .
\[def:bubbletons\] Every finite type solution of the $\sinh$-Gordon equation corresponds to a unique polynomial Killing field $\zeta$ without zeroes and initial value $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$. The curve defined by $\nu^2=-\det \xi$ has a unique compactification to a projective curve without singularities at $\lambda=0$ and $\lambda=\infty$. If $\det(\xi)$ has multiple roots, then we say that the solution contains [**[bubbletons]{}**]{}. The arithmetic genus of this hyperelliptic curve is equal to $g$.
Bubbletons
----------
We briefly motivate Definition \[def:bubbletons\] above, and refer the reader to [@Bur:iso; @BurP:dre; @KilSS; @McI:tor; @SteW:bub] for further details. If an initial value $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ gives rise to a cylinder with extended frame $F_\lambda$ and monodromy $M_\lambda$, and $\beta \in \{\lambda \in {\mathbb C}: 0 < |\lambda | < 1
\}$ is a point at which $M_\beta = \pm \mathbbm{1}$, we define a simple factor $$g = \begin{pmatrix}
\sqrt{\frac{\lambda-\beta}{1-\bar{\beta}\,\lambda}} & 0 \\
0 & \sqrt{\frac{1-\bar{\beta}\,\lambda}{\lambda-\beta}}
\end{pmatrix}\,.$$ Then for any $0 < r < |\beta |$ the dressed extended frame, obtained from the $r$-Iwasawa factorization [@BurP:dre] of $g F_\lambda$, is an extended frame of a cylinder with a bubbleton. On the initial value level this dressing action corresponds to $g\,\xi\,g^{-1}$ which obviously has singularities at $\beta,\,1/\bar{\beta}$. To eliminate these, consider $\tilde{\xi} =
(\lambda-\beta)(1-\bar{\beta}\,\lambda)\,g\,\xi\,g^{-1}$. If $a =
-\lambda \det \xi$, then $\tilde{a} = -\lambda \det \tilde{\xi} =
(\lambda-\beta)^2(1-\bar{\beta}\,\lambda)^2\,a$, so the polynomial $\tilde{a}$ of a bubbleton has double zeroes.
\[thm:dense stratum\] If $a$ has multiple roots, then $\mathcal{K}_a^\circ=\{\xi\in\mathcal{K}_a \,\mid
\mbox{ all roots of $\xi$ are unimodular }\}$ is open and dense in $\mathcal{K}_a$. If $a$ has no unimodular zeroes then it is a $g$-dimensional submanifold. If $a$ has unimodular zeroes, then let $\tilde{a}$ denote the quotient of $a$ by all real zeroes. Then $\mathcal{K}_a^\circ$ is the image of the multiplication with an appropriate rational function $p$ from $\mathcal{K}_{\tilde{a}}^\circ $ to $\mathcal{K}_a^\circ$.
Similar arguments as in the proof of Lemma \[thm:compact\] carry over to this situation.
\[thm:hom\_S\] Suppose $a$ is polynomial of degree $2g$ satisfying the reality condition . Assume $a$ has precisely $2\tilde{g}$ pairwise distinct non-unimodular roots and $g-\tilde{g}$ pairs of unimodular roots of order 2. Then $\mathcal{K}_a \cong ({\mathbb S}^1)^{\tilde{g}}$.
Since $-\det$ is the square of a norm on all skew-hermitian $2\times
2$ matrices, all $\xi\in\mathcal{K}_a$ have a zero at the unimodular double roots of $a$. Let $a(\lambda)=\tilde{a}(\lambda)p^2(\lambda)$ be the corresponding decomposition of $a$ into an $\tilde{a}$ with pairwise distinct roots and the corresponding factors . Due to Proposition \[th:pKf\_min\] the one-to-one correspondence $\zeta\leftrightarrow\tilde{\zeta}$ between polynomial Killing fields $\zeta$ with roots and polynomial Killing fields without roots $\tilde{\zeta}$ induces an isomorphism $\mathcal{K}_a\simeq\mathcal{K}_{\tilde{a}}$. The assertion now follows from Lemma \[thm:compact\].
Spectral curves II
------------------
We also utilize the description of finite type surfaces in ${\mathbb S}^3$ via spectral curves due to Hitchin [@Hit:tor], and relate this to our previous definition of spectral curves due to Bobenko [@Bob:cmc]. While Hitchin defines the spectral curve as the characteristic equation for the holonomy of a loop of flat connections, Bobenko defines the spectral curve as the characteristic equation of a polynomial Killing field. We shall use both of these descriptions, and briefly recall their equivalence: Due to , the monodromy ${\mathbb C}^\ast \to {\mathrm{SL}_{\mbox{\tiny{$2$}}}}({\mathbb C}),\,\lambda \mapsto M_\lambda$ is a holomorphic map with essential singularities at $\lambda = 0,\,\infty$. By construction the monodromy takes values in ${\mathrm{SU}_{\mbox{\tiny{$2$}}}}$ for $|\lambda |=1$. The monodromy depends on the choice of base point, but its conjugacy class and hence eigenvalues $\mu_\lambda,\,\mu_\lambda^{-1}$ do not. With $\Delta(\lambda)={\mathrm{tr}}(M_\lambda)$ the characteristic equation reads $$\label{eq:characteristic1}
\mu_\lambda ^2 - \Delta(\lambda) \,\mu_\lambda + 1 =0\,.$$ The set of solutions $(\lambda,\,\mu)\in{\mathbb C}^2$ of yields another definition of the [**spectral curve**]{} of periodic (not necessarily finite type) solutions of the $\sinh$-Gordon equation. Moreover, the eigenspace of $M_\lambda$ depends holomorphically on $(\lambda,\,\mu)$ and defines the [**eigenbundle**]{} on the spectral curve. Let us compare this with the previous definition of a spectral curve of periodic finite type solutions of the $\sinh$-Gordon equations. Let $\zeta$ be a polynomial Killing field with initial value $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$, with period $\tau$ so that $\zeta(p+\tau) = \zeta(p)$ for all $p \in {\mathbb R}^2$. Then also the corresponding $\alpha(\zeta)$ is $\tau$-periodic. Let $dF_\lambda= F_\lambda \alpha(\zeta),\,F_\lambda(0)=\mathbbm{1}$ and $M_\lambda = F_\lambda(\tau)$ be the monodromy with respect to $\tau$. Then for $z=0$ we have $\xi= \zeta(0) = \zeta(\tau) =
F_\lambda^{-1}(\tau) \,\xi\,F_\lambda(\tau) = M_\lambda^{-1} \xi
\,M_\lambda$ and thus $$[\,M_\lambda,\,\xi\,] = 0\,.$$ All eigenvalues of holomorphic $2\times 2$ matrix valued functions depending on $\lambda\in{\mathbb{CP}^1}$ and commuting point wise with $M_\lambda$ or $\xi$ define the sheaf of holomorphic functions of the spectral curve. Hence the eigenvalues of $\xi$ and $M_\lambda$ are different functions on the same Riemann surface. Furthermore, on this common spectral curve the eigenspaces of $M_\lambda$ and $\xi$ coincide point-wise. Consequently the holomorphic eigenbundles of $M_\lambda$ and $\xi$ coincide.
A finite type solution of the $\sinh$-Gordon equation is periodic if and only if
1. There exists a meromorphic differential $d\ln\mu$ on the spectral curve with second order poles without residues at the two points $\lambda=0$ and $\lambda=\infty$.
2. This differential is the logarithmic derivative of a function $\mu$ on the spectral curve which transforms under the involutions as $\sigma^\ast\mu=\mu^{-1}$, $\rho^\ast\mu=\bar{\mu}^{-1}$ and $\eta^\ast\mu=\bar{\mu}$.
Conversely, a periodic solutions of the $\sinh$-Gordon equation is of finite type if and only if the monodromy fails at only finitely many points $\lambda\in{\mathbb C}^\times$ to be semisimple.
Due to Krichever [@Kri_77], the translations by $z\in{\mathbb C}$ act on the eigenbundle by the tensor product with a one-dimensional subgroup of the Picard group. In Sections 1.4-1.7 McIntosh [@McI:tor] describes this Krichever construction for finite type solutions of the $\sinh$-Gordon equation. The line bundle corresponding to $\tau\in{\mathbb C}$ is trivial if and only if there exists a non-vanishing holomorphic function $\mu$ on the compactified spectral curve with essential singularities at $\lambda=0$ and $\lambda=\infty$, whose logarithm has a first order pole at $\lambda=0$ and $\lambda=\infty$ with singular part equal to $\tau/\sqrt{\lambda}$ and $\bar{\tau}\sqrt{\lambda}$. This implies the characterization of periodic finite type solutions.
At all simple roots of $\Delta^2-4$ the monodromy cannot be semisimple. Furthermore, at a double root of $\Delta^2-4$ the monodromy fails to be semisimple, if and only if it is dressed by a simple factor and contains a corresponding bubbleton. An asymptotic analysis shows that there can exists at most finitely many roots of $\Delta^2-4$ of order larger than two.
Pinkall and Sterling [@PinS], and independently Hitchin [@Hit:tor] proved that doubly periodic solutions of the sinh-Gordon are of [**[finite type]{}**]{}. Thus all metrics of tori are of finite type. We enlarge this class by relaxing one period, and make the following
The cylinders with constant Hopf differential and whose metric is a periodic solution of finite type of the $\sinh$-Gordon will be called [**[ cylinders of finite type]{}**]{}.
Examples
--------
We compute some examples of initial values, polynomial Killing fields and extended frames for spheres, and spectral genus $g=0,\,1$ surfaces. Formulas for all finite type surfaces in terms of theta-functions are given by Bobenko [@Bob:tor].
### Spheres
We start with a discussion of spheres. Since the Hopf differential vanishes identically, spheres constitute a degenerate case since their conformal factor is a solution to the Liouville equation rather than the $\sinh$-Gordon equation, a fact also reflected in the initial value which does not satisfy the semi-simplicity condition. Consider $$\label{eq:sphere_alpha}
\alpha_\lambda = \frac{1}{2}\begin{pmatrix}
u_z\,dz-u_{\bar{z}}\,d\bar{z} &
2\lambda^{-1}e^u\,dz \\
-2\lambda\,e^ud\bar{z} &
-u_z\,dz+u_{\bar{z}}\,d\bar{z}
\end{pmatrix}\,.$$ Then $2\,d\alpha_\lambda + [\,\alpha_\lambda \wedge
\alpha_\lambda\,] = 0$ if and only if $u$ solves the Liouville equation $\partial \bar{\partial}\,u + e^{2u} = 0$. The solution is $u(z,\,\bar{z}) = - \log(1+z\bar{z})$. Plugging this into the $\alpha_\lambda$ in , and solving $dF_\lambda
= F_\lambda \,\alpha_\lambda,\,F_\lambda(0) = \mathbbm{1}$ gives the extended framing $$F_\lambda = \frac{1}{\sqrt{1+z\bar{z}}}\,
\begin{pmatrix} 1 & z\lambda^{-1} \\ -\lambda\bar{z} & 1
\end{pmatrix}\,.$$ Using the initial value $\xi = \lambda^{-1}\varepsilon_+ -
\lambda\,\varepsilon_-$, the corresponding polynomial Killing field is $$\zeta = \frac{1}{1+z\bar{z}}\,\begin{pmatrix} z-\bar{z} & \lambda^{-1}(1+z^2) \\
-\lambda\,(1+\bar{z}^2) & \bar{z}-z \end{pmatrix}\,.$$
### Flat cylinders {#sec:clifford}
We next discuss flat surfaces of revolution in $\mathbb{S}^3$, and compute the closing conditions for the Clifford torus.
\[th:vaccuum\] After a possible isometry, for any flat surface $f: {\mathbb R}^2 \to
\mathbb{S}^3$ with constant mean curvature $H$, there exists a $t_0
\in {\mathbb R}$ such that $H= \cot(2t_0)$, and with $\lambda_0 = e^{i t_0}$ we have $$\label{eq:vaccuumF}
f = F_{\lambda_0^{-1}}\,F_{\lambda_0}^{-1}
\quad \mbox{ with }\quad
F_\lambda = \exp \left( \,\frac{i}{2}
\begin{pmatrix} 0 & z \lambda^{-1} + \bar{z}\\
z + \bar{z}\lambda & 0 \end{pmatrix} \,\right)\,.$$
Solving $dF_\lambda = F_\lambda \alpha_\lambda$ with $u \equiv 0$ in $\alpha_\lambda$ of proves that after a possible isometry, any flat constant mean curvature immersion is framed by such $F_\lambda$ as in . Hence there exist distinct $\lambda_0,\,\lambda_1 \in \mathbb{S}^1$, and a frame of the given form such that after a possible isometry $f =
F_{\lambda_1}\,F_{\lambda_0}^{-1}$. By , the two distinct unimodular numbers $\lambda_j= e^{2i t_j},\,j=0,\,1$ must be chosen so that for the constant mean curvature $H$ of $f$ we have $H= \cot(t_0-t_1)$. Now $\lambda_0,\,\lambda_1$ are determined only up to a phase, and since rotations in the $\lambda$-plane correspond to rotations in the $z$-plane via a unitary gauge, we may adjust the phase so that $\lambda_1 = \lambda_0^{-1}$.
For such spectral genus zero surfaces, the polynomial Killing field is constant and equal to the initial value $\xi = \tfrac{i}{2}\left(
(\lambda^{-1}+1)\,\varepsilon_+ + (1+\lambda)\,\varepsilon_-
\right)$. The factor $1/2$ is a consequence of the choices made in Theorem \[thm:sinh\]. For $\lambda_0 = i$ and $\lambda_1 = -i$ we obtain a minimal flat surface. We determine the simple periods such that the restriction to a fundamental domain gives the [**[Clifford torus]{}**]{}: The eigenvalues of $F_\lambda(z)$ are of the form $\exp(\pm
\mu)$ with $\mu(z,\,\lambda) = \tfrac{i}{2}(z\,\lambda^{-1/2} +
\bar{z}\,\lambda^{1/2})$. Simple periods are numbers $\omega_1,\,\omega_2 \in {\mathbb C}$ with smallest possible modulus satisfying $\mu(\omega_1,\,\pm i) = \pi i$ and $\mu(\omega_2,\,\pm i) = \pm \pi i$, and compute to $\omega_1 =
\pi \sqrt{2} $ and $\omega_2 =\pi i \sqrt{2} $.
### Delaunay surfaces
For $a,\,b \in {\mathbb R}$ the 1-parameter family of conformal metrics of Delaunay surfaces $v^2(x) (dx^2 + dy^2)$ is given by the Jacobian elliptic functions $v(x) = 2b\, \mathrm{dn}(2bx\,|\,1-a^2/b^2)$. For the derivation of this conformal factor from the Gauss equation in the rotational case we refer to [@BurK; @SKKR]. Note that in the limiting cases $a=\pm b$ we have $v \equiv 1$, which is the flat case, while when $a=0$ we have $v(x) = \mathrm{sech}(x)$. The initial value [@Kil:del; @SKKR] is $\xi = (a\lambda^{-1} +
b)\,i\varepsilon_+ + (b + a\lambda)\,i\varepsilon_-$, and the polynomial Killing field computes to $$\zeta = i\begin{pmatrix} -\tfrac{v'(x)}{2v(x)} &
\tfrac{2ab\lambda^{-1}}{v(x)} +\tfrac{v(x)}{2}\\
\tfrac{2ab\lambda}{v(x)} +\tfrac{v(x)}{2} & \tfrac{v'(x)}{2v(x)}
\end{pmatrix}\,.$$
Deformation of spectral data {#sec:deformation}
============================
In this section we describe the spectral curves of periodic finite type solutions of the $\sinh$-Gordon equation by two polynomials $a$ and $b$. The first polynomial defines the hyperelliptic curve and the second polynomial the meromorphic differential $d\ln\mu$ on this curve. Not all polynomials $a$ and $b$ correspond to spectral curves of periodic solutions of the $\sinh$-Gordon equation. In order to describe the subsets of all such $a$ and $b$, we derive vector fields on the space of coefficients of $a$ and $b$, which leave these subsets invariant. We will be using the usual spectral parameter $\lambda$ in which $\lambda = 0,\,\infty$ are singularities, as well as the transformed spectral parameter $$\kappa = i\,\frac{1-\lambda}{1+\lambda}\,.$$ (Hence $\lambda = (i-\kappa)/(i+\kappa)$.) The part of the spectral curve over $\{\lambda \in {\mathbb C}: |\lambda |=1 \}$ corresponds to real $\kappa$. These parameters are fixed only up to Möbius transformations $$\begin{aligned}
\label{eq:moebius}
\lambda &\mapsto e^{2i\varphi}\lambda&
\kappa \mapsto \frac{\sin
\varphi + \kappa \cos \varphi}{\cos \varphi - \kappa \sin \varphi}\end{aligned}$$ This degree of freedom allows us to assume that no branch point or otherwise significant point, which we shall introduce later (for example a zero of $d\ln\mu$), lies at $\kappa = \infty$. The spectral curve is then a hyperelliptic surface which we describe with the equation $$\label{eq:def_a}
\nu^2 = (\kappa^2 +1)\,a(\kappa)\,.$$ Here $a$ is a real polynomial of degree $2g$ which has highest coefficient equal to one, and which is non-negative for $\kappa \in
{\mathbb R}$. Thus $a$ only possesses real roots of even order.
In the following we will consider finite type cylinders in ${\mathbb S}^3$. These possess a monodromy whose eigenvalue $\mu$ is a holomorphic function on the spectral curve, and has essential singularities at $\kappa = \pm i$. Then $d\ln \mu$ is an abelian differential of the second kind of the form $$\label{eq:def_b}
d\ln\mu = 2\pi i \frac{b(\kappa)\,d\kappa}{(\kappa^2+1)\,\nu}\,,$$ where $b$ is a real polynomial of degree $g+1$.
As a consequence of the work of Bobenko [@Bob:tor; @Bob:cmc], our starting point is the definition of what we call the spectral data of a cylinder of finite type in the 3-sphere.
\[thm:spec\_bobenko\] Let $a$ be a real polynomial of degree $2g$ with highest coefficient equal to one, and let $b$ be a real polynomial of degree $g+1$, and $\kappa_0,\,\kappa_1 \in {\mathbb R}$ two [**[marked points]{}**]{}.
The [**[spectral data]{}**]{} of a cylinder of finite type in ${\mathbb S}^3$ with mean curvature $$\label{eq:H kappa}
H = \frac{1 + \kappa_0 \kappa_1}{\kappa_0 - \kappa_1}$$ consists of a quadruple $(a,\,b,\,\kappa_0,\,\kappa_1)$ with the following properties:
1. $a(\kappa) \geq 0$ for $\kappa \in {\mathbb R}$.
2. On the hyperelliptic surface $\nu^2 = (\kappa^2+1)\,a(\kappa)$ there is a single valued holomorphic function $\mu$ with essential singularities at $\kappa = \pm i$ with logarithmic differential , that transforms under the three involutions $$\begin{aligned}
\sigma:(\kappa,\,\nu)&\mapsto (\kappa,\,-\nu),&
\rho:(\kappa,\,\nu)&\mapsto (\bar{\kappa},\,\bar{\nu}),&
\eta:(\kappa,\,\nu)&\mapsto (\bar{\kappa},\,-\bar{\nu}),\end{aligned}$$ as $\sigma^{\ast}\mu = \mu^{-1}$, $\rho^{\ast}\mu=\bar{\mu}^{-1}$ and $\eta^{\ast}\mu=\bar{\mu}$.
3. $\mu(\kappa_0) = \mu(\kappa_1) = \pm 1$.
We call the fixed point set of $\rho$ the [**real part**]{}. While a Möbius transformation of the parameter $\kappa$ changes the spectral data $(a,\,b,\,\kappa_0,\,\kappa_1)$, it changes neither the corresponding periodic solutions of the $\sinh$-Gordon equation nor the corresponding cylinders in ${\mathbb S}^3$. Hence the moduli space of spectral data is the set of equivalence classes of spectral data up to the action on the spectral data induced by , prompting the following
[[(i)]{}]{} For all $g\in{\mathbb N}_0$ let $\hat{{\mathcal{M}}}_g$ be the space of equivalence classes of spectral data $(a,\,b)$ obeying conditions [[([A]{})]{}]{}-[[([B]{})]{}]{} up to the action of on $(a,b)$.
[[(ii)]{}]{} For all $g\in{\mathbb N}_0$ let ${\mathcal{M}}_g$ be the space of equivalence classes of spectral data $(a,\,b,\,\kappa_0,\,\kappa_1)$ obeying conditions [[([A]{})]{}]{}-[[([C]{})]{}]{} up to the action of on $(a,b,\kappa_0,\kappa_1)$.
Thus $\hat{{\mathcal{M}}}_g$ is the moduli space of spectral data of periodic solutions of the $\sinh$-Gordon equation of arithmetic genus $g$, and ${\mathcal{M}}_g$ is the moduli space of spectral data of finite type cylinders in ${\mathbb S}^3$.
We now derive vector fields on open sets of spectral data $\left\{
(a,\,b,\,\kappa_0,\,\kappa_1)\right\}$ and show that their integral curves are differentiable families of spectral data of periodic finite type solutions of the $\sinh$-Gordon equation. We parameterize such families by one or more real parameters, which we will denote by $t$. From condition [[([B]{})]{}]{} in Definition \[thm:spec\_bobenko\] we conclude that $\partial_t\ln \mu$ is meromorphic on the corresponding family of spectral curves. If we view these functions locally in terms of $\kappa$ and $t$, then $\partial_t\ln \mu$ can only have poles at the branch points, or equivalently at the zeroes of $a$, and at $\kappa = \pm i$. If we assume that for such a family of spectral curves the genus $g$ is constant, then $\partial_t \ln \mu$ can at most have poles of first order at simple roots of $a$. In general we have $$\label{eq:def_c}
\partial_t\ln \mu = \tfrac{2 \pi i}{\nu} \,c(\kappa)$$ with a real polynomial $c$ of degree at most $g+1$.
To compute the corresponding vector field on the space of spectral data we view $\mu$ locally as a function of the parameters $\kappa$ and $t$. Differentiating and gives $$\partial^2_{t\kappa}\ln \mu = 2\pi i
\frac{2\dot{b}a - b\dot{a}}{2\nu^3}\,,\quad
\partial^2_{\kappa t}\ln \mu = 2\pi
i\frac{2(\kappa^2+1)ac' - 2\kappa ac -
(\kappa^2+1)a'c}{2\nu^3}\,.$$ Second partial derivatives commute if and only if $$\label{eq:integrability_1}
2\dot{b}a - b\dot{a} = 2(\kappa^2+1)ac' - 2\kappa ac -
(\kappa^2+1)a'c\,.$$ The highest coefficient on the right hand side vanishes, so both sides are polynomials of at most degree $3g+1$. As the highest coefficient of $a$ does not depend on $t$ we conclude that $\dot{a}$ is a real polynomial of degree $2g-1$, and $\dot{b}$ a real polynomial of degree $g+1$. Thus we have to determine $3g+2$ real coefficients. In case $a$ and $b$ have no common roots, equation uniquely determines the values of $\dot{a}$ at the roots of $a$ and the values of $\dot{b}$ at the roots of $b$. Since the highest coefficient on the right hand side depends only on the highest coefficient of $\dot{b}$, in this case uniquely determines a tangent vector on the space of spectral data of periodic finite type solutions of the $\sinh$-Gordon equation. By defining such polynomials $c$ we obtain vector fields on the space of real polynomials $a$ of degree $2g$ and highest coefficient one and real polynomials $b$ of degree $g+1$.
For spectral data of cylinders in ${\mathbb S}^3$ we have to deform in addition to the polynomials $a$ and $b$ the two marked points, such that the closing condition [[([C]{})]{}]{} of Definition \[thm:spec\_bobenko\] is preserved. As long as $\kappa_0
\neq \kappa_1$, and thus $| H | <\infty$, we preserve the closing condition if $\partial_t \ln \mu(\kappa_j(t),\,t) = 0$, which holds precisely when $\partial_{\kappa_j} \ln
\mu(\kappa_j(t),\,t)\,\partial_t\kappa_j +
\partial_t \ln \mu(\kappa_j(t),\,t) = 0$. Using equations and , the closing conditions are therefore preserved if and only if $$\label{eq:integrability_2}
\dot{\kappa}_j = -\frac{(\kappa_j^2 +1)
\,c(\kappa_j)}{b(\kappa_j)}\,.$$ The equations and define rational vector fields on the space of spectral data $(a,b,\kappa_0,\kappa_1)$ not necessarily obeying conditions [[([A]{})]{}]{}-[[([C]{})]{}]{} of Definition \[thm:spec\_bobenko\]
\[thm:deformation\] [[(i)]{}]{} Let $U$ be an open subset of spectral data $(a,\,b)$, with $a,\,b$ having no common roots. Let $c$ be a smooth function from $U$ to the real polynomials of degree $g+1$. Then equations define a smooth vector field on $U$. The corresponding flow leaves invariant the subset of spectral data obeying conditions [[([A]{})]{}]{}-[[([B]{})]{}]{} of Definition \[thm:spec\_bobenko\].
[[(ii)]{}]{} Let $U$ be an open subset of spectral data $(a,\,b,\,\kappa_0,\,\kappa_1)$, with $a,\,b$ having no common roots, and real $\kappa_0 \neq \kappa_1$ Let $c$ be a smooth function from $U$ to the real polynomials of degree $g+1$. Then equations and define a smooth vector field on $U$. The corresponding flow leaves invariant the subset of spectral data obeying conditions [[([A]{})]{}]{}-[[([C]{})]{}]{} of Definition \[thm:spec\_bobenko\].
We only prove (i). The proof of (ii) is similar. The solutions $\dot{a}$ and $\dot{b}$ of equation are rational expressions of the coefficients of $a$, $b$ and $c$. If $a$ and $b$ have no common roots, then the Taylor coefficients of $\dot{a}$ and $\dot{b}$ at the roots of $a$ and $b$ up to the order of the roots minus one, respectively, and the highest coefficient of $\dot{b}$ are uniquely determined by equation . Hence, in this case the denominators of the rational expressions for $\dot{a}$ and $\dot{b}$ do not vanish. Hence for smooth $c$ the corresponding $\dot{a}$ and $\dot{b}$ are smooth too.
Due to $\partial_t\ln \mu$ is a meromorphic function on the hyperelliptic curve. Hence the periods of the meromorphic differential $d\ln\mu$ do not depend on $t$. The transformation rules of $\mu$ under $\sigma$, $\rho$ and $\eta$ are preserved under the flows of the vector field corresponding to $c$. Hence the integrals of $d\ln\mu$ along any smooth path from one root of $a$ to another root of $a$ is preserved too. This implies that the subset of spectral data $(a,b)$, which determine by a single valued function $\mu$ with $\sigma^{\ast}\mu=\mu^{-1}$, $\rho^{\ast}\mu=\bar{\mu}^{-1}$ and $\eta^{\ast}\mu=\bar{\mu}$ is preserved under this flow.
The space of real hyperelliptic curves of genus $g$ is up to Möbius transformations is $2g-1$ dimensional. All such curves correspond to real solutions of the $\sinh$-Gordon equation. The subset $\hat{{\mathcal{M}}}_g$ of curves corresponding to periodic solutions has codimension growing with $g$. The space of real polynomials $c$ of degree at most $g+1$ is $g+2$-dimensional. In case $c$ is proportional to $b$, then the deformation corresponds to an infinitesimal Möbius transformation . Hence $\hat{{\mathcal{M}}}_g$ is $g+1$-dimensional.
Moduli of spectral data of periodic solutions of the $\sinh$-Gordon equation {#sec:moduli}
============================================================================
In this section we switch to another description of the spectral data of Definition \[thm:spec\_bobenko\]. We shall use the values of the trace $\Delta$ of the monodromy at the roots of the derivative of $\Delta$, as in Grinevich and Schmidt [@GriS1], to determine local parameters on the moduli of spectral curves of genus $g$. In doing so, and switching to the parameter $\kappa$ of the previous section, we describe the twice-punctured Riemann sphere ${\mathbb{CP}^1}\setminus\{\pm i\}$, which is the domain of definition of the covering map $\Delta$, as the result of gluing infinitely many copies of ${\mathbb C}$ along cuts to be specified by the branch points and combinatorial data. The combinatorial data specifies which sheets are joined by which branch points and branch cuts. We call this combinatorial data the [**[gluing rules]{}**]{}.
The meromorphic function $\kappa$ and the holomorphic function $\mu$ with essential singularities at $\kappa=\pm i$ fulfill an equation of the form $$\label{eq:characteristic}
\mu^2 - \Delta(\kappa)\,\mu + 1=0\,.$$ Hence $\Delta$ is a holomorphic function $$\label{eq:covering}
\Delta\,:\,{\mathbb{CP}^1}\setminus\{\pm i\}\rightarrow {\mathbb C}\,, \qquad \kappa
\mapsto\Delta(\kappa)$$ The curve is hyperelliptic with hyperelliptic involution $\sigma:(\kappa,\mu)\mapsto(\kappa,\mu^{-1})$. The branch points are the odd-ordered roots of $\Delta^2 -4$. We next characterize those maps $\Delta$ which correspond to spectral data $(a,\,b)$ described in Definition \[thm:spec\_bobenko\].
Let us indicate to what extent the function $\Delta$ is determined by its values at the branch points, and the gluing rules. We first recall the simpler, but essentially comparable situation of finitely sheeted coverings investigated by Hurwitz [@Hur]. Finitely sheeted covering maps ${\mathbb{CP}^1}\rightarrow{\mathbb{CP}^1}$ are determined by their branch points and their gluing rules up to Möbius transformations of both copies of ${\mathbb{CP}^1}$. In most cases the parametrization of the image of the covering map is fixed by the values of this parameter at the branch points. By fixing the values of the parameter of the domain at some marked points, we can fix also the parameter of the domain. For some fixed parametrization of the domain and the image such a finitely sheeted covering map is a rational function. Consequently a rational function can be characterized by the values of the function at the zeroes of the derivative, some gluing rules, and some conditions on the parametrization of the domain.
Now we return to our infinitely sheeted covering map . In this article we shall be concerned only with spectral curves of finite geometric genus. In this case $\ln\mu$ extends to a meromorphic function on two neighbourhoods of $\kappa=\pm i$ with first order poles at these two points. In particular $(\ln\mu)^{-2}=-(\arccos(\Delta/2))^{-2}$ is a local parameter on $\kappa\in{\mathbb{CP}^1}\setminus\{\pm i\}$ at $\kappa=\pm i$. With this local parameter the domain of can be compactified to ${\mathbb{CP}^1}$. Since $\kappa$ takes at the two marked points the values $\pm i$ and transforms under $\rho$ and $\eta$ as $\kappa\mapsto\bar{\kappa}$, the parameter $\kappa$ is determined up to Möbius transformations .
\[thm:Delta\] The functions $(\kappa,\,\mu)$ correspond to spectral data $(a,\,b)$ obeying conditions [[([A]{})]{}]{} and [[([B]{})]{}]{} in Definition \[thm:spec\_bobenko\] if and only if the function $\Delta$ satisfies the following conditions:
1. All but finitely many roots of $\Delta^2 -4$ are roots of even order.
2. A branch of the function $(\ln\mu)^{-2}=-(\arccos(\Delta/2))^{-2}$ extends to a holomorphic function on two neighbourhoods of $\kappa=\pm i$ with simple zeroes at $\kappa=\pm i$.
3. The function $\Delta$ transforms as $\Delta(\bar{\kappa})=\bar{\Delta}(\kappa)$ and obeys $\Delta^2(\kappa)\leq 4$ for all $\kappa\in\mathbb{R}$.
The roots of the corresponding polynomial $a$ have to be a subset of the roots of $\Delta^2-4$, such that $(\Delta^2-4)/a$ has only roots of even order. All such real $a$ are possible choices. Conversely, the function $\Delta$ is uniquely determined by the spectral data $(a,\,b)$.
The odd order roots of $\Delta^2-4$ are the branch points of the two-sheeted covering over $\kappa\in{\mathbb{CP}^1}$. Hence only $\Delta$ obeying condition [[([D]{})]{}]{} correspond to polynomials $a$. Since $\kappa=\pm i$ is a branch point, condition [[([E]{})]{}]{} follows from condition [[([B]{})]{}]{}. Moreover, the involutions $\rho$ and $\eta$ in condition [[([B]{})]{}]{} induce the involution in Condition [[([F]{})]{}]{}.
Vice verse, Condition [[([F]{})]{}]{} implies that all real roots of $\Delta^2-4$ are roots of even order. Locally on $\kappa\in{\mathbb{CP}^1}\setminus\{\pm i\}$ the function $\mu$ is holomorphic in terms of $\Delta$ and a square root of $\Delta^2-4$. Hence for all real $a$, such that $(\Delta^2-4)/a$ has only roots of even order, the differential $d\ln\mu$ is a meromorphic differential on the hyperelliptic curve defined by $\nu^2=(\kappa^2+1)a(\kappa)$. This differential is antisymmetric with respect to the hyperelliptic involution $\sigma:(\kappa,\nu)\mapsto(\kappa,-\nu)$. Furthermore, due to condition [[([E]{})]{}]{} it has second order poles at $\kappa=\pm i$ without residues and no other poles. Hence it is of the form . Due to condition [[([F]{})]{}]{} the involution $\kappa\mapsto\bar{\kappa}$ induces on the hyperelliptic curve two antilinear involutions $\rho$ and $\eta$, one of which denoted by $\eta$ has no fixed points. The function $\mu$ transforms as $\rho^{\ast}\mu=\bar{\mu}^{-1}$ and $\eta^{\ast}\mu=\bar{\mu}$.
We shall see later that condition [[([E]{})]{}]{} implies [[([D]{})]{}]{}. The function $\Delta$ defines an infinitely sheeted covering map with essential singularities at $\kappa=\pm i$. The branch points of this covering map are the zeroes of $\Delta'(\kappa) =0$ and thus precisely the roots of $d\ln \mu$ together with the set of singularities of the spectral curve defined by equation .
We shall see that essentially we can move all the branch points independently without destroying the periodicity. The reality condition [[([F]{})]{}]{} in Theorem \[thm:Delta\] imposes the only restriction. Consequently the moduli space is a covering space over the parameter space of the values of $\Delta$ at all branch points. We shall see that for spectral curves of finite geometric genus all of them with the exception of finitely many are fixed. Consequently we can assume this parameter space to have finite dimension. The gluing rules of this covering are completely determined by the gluing rules of the covering . In Hurwitz [@Hur] the analogous deformations of finitely sheeted coverings are investigated.
\[thm:Delta 1\] For spectral data $(a,\,b)$ of genus $g$ obeying conditions [[([A]{})]{}]{}-[[([B]{})]{}]{} of Definition \[thm:spec\_bobenko\], for which none of the pairwise distinct roots of $b$ are roots of $a$, the element $[(a,\,b)]\in\hat{{\mathcal{M}}}_g$ possesses an open neighbourhood in $\hat{{\mathcal{M}}}_g$, which is uniquely parameterized by the values of $\Delta$ at the roots of $b$.
By assumption there exists for every root $\beta_i$ of $b$ a unique polynomial $c_i$ that vanishes at every root of $b$ except at $\beta_i$, where it attains a value such that $$\partial_t \Delta(\beta_i)
= 2\sinh(\ln(\mu(\beta_i)))\partial_t \ln\mu(\beta_i) = 1\,.$$ The corresponding vector fields in a neighbourhood $U$ of spectral data in the moduli space commute: the map $U \to {\mathbb C}^{g+1},\,\beta_i
\mapsto \Delta(\beta_i)$ sends these vector fields to coordinate vector fields. Since these vector fields are linearly independent on $U$, they generate the quotient of all polynomials $c$ modulo the polynomials $b$. Now the Lemma follows from Theorem \[thm:deformation\].
In the following we shall investigate how the above parametrization is affected if the roots of the polynomial $b$ are either not all distinct, or some of the roots of $b$ coincide with the roots of $a$. Recall that the local parameters are the values of $\Delta$ at the branch points of the cover $\kappa \mapsto \Delta(\kappa)$. Describing this cover by means of the branch cuts we also have a description of the moduli space as a cover of the parameter space, thus obtaining a global picture of the moduli space. Assume we are given two branch points of the covering map $\kappa \mapsto
\Delta(\kappa)$ that connect a given sheet of the cover with two different sheets. If in the process of a continuous deformation one of these branch points circumvents the other branch point, then the sheets which these branch points connect permute. Hence higher order roots of $b$ are branch points of the moduli space.
If on the other hand, a root of $b$ coincides with a root of $a$, then the derivative of $\kappa \mapsto \Delta(\kappa)$ has a higher order root there and $\Delta^2(\kappa) =4$ there. Thus a root of $b$ coincides with a root of $\Delta'$ that arises from a singularity of the spectral curve . Hence also in this case two branch points of the covering map $\kappa \mapsto
\Delta(\kappa)$ coalesce.
For any branch point of a covering map we can choose small open neighbourhoods $U$ and $V$ in the domain and in the image respectively, such that the restriction $\Delta|_U$ of $\Delta$ to $U$ is a finitely sheeted covering over $V$. Small movements of the branch points of $\Delta|_U$ over a closed subset $A\subset V$ do not change the restriction of $\Delta$ to the pre-image of $V\setminus A$ under $\Delta|_U$. Moreover, if we consider higher order branch points as coalescing first order branch points, we can move all these first order branch points independently over $A$, without changing the restrictions of $\Delta$ to the pre-image of $V\setminus A$ under $\Delta|_U$. More precisely, if $\beta$ is a branch point at $\Delta_0=\Delta(\beta)$ of order $k$, then $w=\sqrt[k+1]{\Delta-\Delta_0}$ is a local parameter of the covering space on an open neighbourhood of $\beta$. All polynomials $P(w)$ of degree $k+1$ with highest coefficient $1$ and small lower order coefficients describe small perturbations of the covering map $w\mapsto\Delta=w^{k+1}+\Delta_0$. All of them are covering maps with $k$ branch points, which are the zeroes of the derivative $P'$. The parameter $w$ of such coverings is determined up to Möbius transformations $w\mapsto aw+b$ with $a\in{\mathbb C}^{\times},b\in{\mathbb C}$. Hence the values of $\Delta$ at the $k$ branch points together with the gluing rules of these covering maps $w\mapsto P(w)$ determine such $P$ uniquely up to a reparametrization $\tilde{P}(w)=P(aw+b)$ with $a\in{\mathbb C}^{\times},b\in{\mathbb C}$. Furthermore, all small values of $P$ at the $k$ branch points are realized by such polynomials. By gluing the deformed covering $\tilde{\Delta}|_U$ along the pre-image of $V\setminus A$ under $\Delta|_U$ with the restriction of the undeformed $\Delta$ to ${\mathbb{CP}^1}\setminus(\{i,-i\}\cup A)$ we obtain a deformation $\tilde{\Delta}$ of the covering $\Delta$, as a covering map from an abstract Riemann surface without fixed parameter $\kappa$ onto ${\mathbb C}$ with fixed parameter $\Delta$.
Finite combinations of such deformations we call [**[local deformations]{}**]{}. If the movements of the branch points respect condition [[([F]{})]{}]{} in Theorem \[thm:Delta\], then we call the corresponding deformations [**[real local deformations]{}**]{}.
In Section \[sec:deformation\] we described deformations of spectral data by real polynomials $c$ of degree $g+1$. We shall calculate the corresponding deformation of . If we consider also $\Delta$ and $\mu$ as functions depending on $\kappa$ and $t$, then we have due to and $$\begin{aligned}
\Delta'=2\sinh(\ln\mu)(\ln\mu)'&= 4\pi
i\frac{\sinh(\ln\mu)b}{(\kappa^2+1)\nu}&
\dot{\Delta}=2\sinh(\ln\mu)\dot{(\ln\mu)}&= 4\pi
i\frac{\sinh(\ln\mu)c}{\nu}.\end{aligned}$$ Hence $\dot{\Delta}$ is given by $$\label{eq:dotDelta}
\dot{\Delta}=\frac{(\kappa^2+1)c}{b}\Delta'.$$
\[thm:Delta 2\] Local deformations preserve conditions [[([D]{})]{}]{}-[[([E]{})]{}]{} in Theorem \[thm:Delta\]. In particular, real local deformed $\tilde{\Delta}$ corresponds to a parameterized covering map , but the parameter $\kappa$ is determined only up to Möbius transformations . Those real local deformations $\tilde{\Delta}$ of the covering map $\Delta$ corresponding to spectral data $(a,b)$ obeying conditions [[([A]{})]{}]{}-[[([B]{})]{}]{} of Definition \[thm:spec\_bobenko\], which move only the branch points of $\Delta$ corresponding to roots of $b$, correspond to unique $[(\tilde{a},\,\tilde{b})]\in\hat{{\mathcal{M}}}_g$.
Since local deformations move only finitely many branch points, they preserve conditions [[([D]{})]{}]{}-[[([E]{})]{}]{} in Theorem \[thm:Delta\]. Due to condition [[([E]{})]{}]{} the covering space corresponding to a local deformed $\tilde{\Delta}$ can be compactified to ${\mathbb{CP}^1}$. A parameter $\kappa$, which takes at the two added points the values $\pm i$ and transforms under $\Delta\mapsto\bar{\Delta}$ as $\kappa\mapsto\bar{\kappa}$ is unique up to . Due to this Theorem all real local deformations $\tilde{\Delta}$ of $\Delta$ corresponding to spectral data $(a,b)$ correspond also to spectral data $(\tilde{a},\tilde{b})$. The branch points of $\tilde{\Delta}$ are in one-to-one correspondence with the zeroes of $d\ln\mu$ and the singularities of . Hence a real local deformation, which moves only the branch points corresponding to zeroes of $b$ corresponds to deformed spectral data $(\tilde{a},\tilde{b})$, which are unique up to the Möbius transformations .
Since the space of local deformations of the covering map are manifolds, this Lemma can be used to make $\hat{{\mathcal{M}}}_g$ into a real $(g+1)$-dimensional manifold. In the sequel we shall call those real local deformations, which can be realized as combinations of continuous movements of single branch points, [**[continuous deformations]{}**]{}. It is not difficult to introduce a topology on ${\mathcal{M}}_g$, such that these deformations corresponds to continuous paths. In fact the space of real polynomials of fixed degree have a natural topology. As a quotient space of such spaces, ${\mathcal{M}}_g$ also has a natural topology.
We shall first compute the map $\kappa \mapsto \Delta(\kappa)$ for the periodic solutions of the $\sinh$-Gordon equation of spectral genus zero: Then $a \equiv 1$ and $$\label{eq:a_vac}
\ln \mu = 2\pi i \frac{b_0\kappa -b_1}{\nu} \,.$$ For the anti-linear fix point free involution $\eta$ of Definition \[thm:spec\_bobenko\] we have $\eta^{\ast} \overline{d\ln\mu} =
d\ln\mu$. As $b_0$ and $b_1$ are real we obtain $$\label{eq:b_vac}
d\ln\mu = 2\pi i \tfrac{b_0 + b_1
\kappa}{(\kappa^2 +1)\,\nu}\,.$$ Then $$\label{eq:genus zero}
\Delta(\kappa) = 2\cosh(\ln\mu) = 2 \cos\left( 2\pi \tfrac{b_0\kappa
-b_1}{\nu}\right) = 2\cos\left(\sqrt{4\pi^2 \tfrac{b_0^2\kappa^2
- 2b_0b_1\kappa +b_1^2}{\kappa^2 +1}} \right)$$ which is the composition of the two maps $\kappa \mapsto\delta=
4\pi^2 \tfrac{b_0^2\kappa^2
- 2b_0b_1\kappa +b_1^2}{\kappa^2 +1}$ and $\delta \mapsto\Delta= 2\cos\sqrt{\delta}$. Of the second of these maps we choose the cuts along the lines $$\sqrt{\delta} \in \pi n +i\,{\mathbb R}\Longleftrightarrow
\left\{ \begin{array}{ll}
\Delta \in (-\infty,\,-2] & \mbox{ for $n$ odd}\,,\\
\Delta \in [2,\,\infty) & \mbox{ for $n$ even}\,. \end{array} \right.$$ Besides the branch points that arise for each $n \in {\mathbb Z}^{\times}$ we have two additional branch points of $\kappa \mapsto
\delta(\kappa)$, situated at $\delta =0$ and $$\kappa = -\tfrac{b_0}{b_1} \Longleftrightarrow \delta = 4\pi^2
\tfrac{b_0^4+2b_0^2b_1^2 + b_1^4}{b_0^2 + b_1^2} = 4\pi^2 (b_0^2 +
b_1^2)\,.$$ Between these latter two points we cut along the line segment joining them. The covering map $\kappa\mapsto
2\cos\sqrt{\pi^2\kappa^2/(\kappa^2+1)}$ has sheets numbered by two copies of ${\mathbb N}$. We denote these sheets by $l^+$ and $l^-$ with $l\in
{\mathbb N}$. In order to describe the movements of the branch points it is convenient to choose only parallel cuts on each sheet. Besides the branch cut along the real part we choose on the sheets $l^\pm$ $$\label{eq:branch point}\begin{aligned}
\mbox{for $l\geq 1$ a cut along }&\Delta\in(-1)^l2+ i\mathbb{R}^+_0
&\mbox{ connecting the sheets }&l^+\mbox{ and }(l+1)^+\\
\mbox{for $l\geq 1$ a cut along }&\Delta\in(-1)^l2- i\mathbb{R}^+_0
&\mbox{ connecting the sheets }&l^-\mbox{ and }(l+1)^-\\
\mbox{for $l> 1$ a cut along }&\Delta\in-(-1)^l2+ i\mathbb{R}^+_0
&\mbox{ connecting the sheets }&l^+\mbox{ and }(l-1)^+\\
\mbox{for $l> 1$ a cut along }&\Delta\in-(-1)^l2- i\mathbb{R}^+_0
&\mbox{ connecting the sheets }&l^-\mbox{ and }(l-1)^-\\
\end{aligned}$$ Hence besides the sheets $1^\pm$ each sheet has at $\Delta=2$ and $\Delta=-2$ a branch point. Those sheets, whose labels have exponents $+$, do not have branch points and cuts at small imaginary values of $\Delta$. Those sheets, whose labels have exponent $-$, do not have branch points and cuts at large imaginary values of $\Delta$. Each sheet has exactly one branch point at $\Delta=-2$. This branch point connects the sheet with another sheet, to which we pass along large circles in the $\Delta$-plane from small imaginary parts to small imaginary parts, if the label has exponent $+$ and from large imaginary parts of $\Delta$ to large imaginary parts, if the label has exponent $-$, respectively. More precisely, for all $l\in\mathbb{N}$ the branch point at $\Delta=-2$ connects the sheet with label $$\label{eq:label sheets}\begin{aligned}
(2l-1)^+\mbox{ with the sheet reached by traversing large circles }&2l-1
\mbox{ times anti-clockwise,}\\
(2l)^+\mbox{ with the sheet reached by traversing large circles }&2l-1
\mbox{ times clockwise,}\\
(2l-1)^-\mbox{ with the sheet reached by traversing large circles }&2l-1
\mbox{ times clockwise,}\\
(2l)^-\mbox{ with the sheet reached by traversing large circles }&2l-1
\mbox{ times anti-clockwise.}\\
\end{aligned}$$ The labels of the sheets are completely determined by this rule. If we choose $b_0=1/2$ and $b_1 =0$, then we have an additional real branch cut that joins the sheets $1^-$ and $1^+$ with two additional real branch points at $\Delta=\pm 2$. This is the spectral data of the standard round cylinder in $\mathbb{R}^3$. The spectral data of the standard round cylinders in ${\mathbb S}^3$ are obtained by moving the real branch point at $\Delta=-2$ in $\Delta\in[-2,2]$ along the sheets $2^\pm, 3^\pm,\ldots$ to and fro. In doing so, the real part contains branch cuts connecting the sheets $1^+$ with $1^-$, $2^+$ with $2^-,\ldots$ and $l^+$ with $l^-$. We introduce the following class of covering maps $\kappa \mapsto \Delta(\kappa)$ connecting the sheets $\left(l^-\right)_{l\in{\mathbb N}}$ and $\left(l^+\right)_{l\in{\mathbb N}}$ at branch cuts such that the following hold:
1. There exists an $L \in {\mathbb N}$ such that for all $l>L$ the sheets $l^{\pm}$ only have the branch points and branch cuts .
2. Along large circles in the $\Delta \in {\mathbb C}$-plane we get the following sequence of sheets: $$\begin{aligned}
\ldots,\,(2l+1)^+,\,(2l-1)^+,\,\ldots,
\,3^+,\,&1^+,\,2^+,\,4^+,\,\dots,\,(2l)^+,\,(2l+2)^+,\,\ldots\end{aligned}$$ traversing anti-clockwise from small imaginary parts to small imaginary parts, and $$\begin{aligned}
\ldots,\,(2l+1)^-,\,(2l-1)^-,\,\ldots,
\,3^-,\,&1^-,\,2^-,\,4^-,\,\dots,\,(2l)^-,\,(2l+2)^-,\,\ldots\end{aligned}$$ traversing clockwise from large imaginary parts to large imaginary parts. Reversing the anti-clockwise and clockwise order gives the same sequences in reverse order.
3. There exists an $L\in\mathbb{N}$ such that for $l\geq
L$ the total number of all branch points that join two sheets from the set $\{1^-,1^+,\,\ldots,\,l^-,l^+\}$ is equal to $2l$. Furthermore, $2l-2$ of the corresponding branch cuts terminate at $2l-2$ additional branch points at infinity.
4. The configuration is invariant under $l^+ \mapsto l^-,\,l^-\mapsto l^+,\,\Delta \mapsto \bar{\Delta}$. Furthermore, the branch order of real branch points at $\Delta=\pm2$ is odd. The fixed points of the corresponding anti-linear involution is called real part and is a branch cut along $\Delta\in[-2,2]$ between sheets $l^+$ and $l^-$
5. Besides the branch cuts along the real part all branch cuts run along unbounded lines parallel to the imaginary axis in direction to very large or small imaginary parts according to the superscript $\pm$ of the corresponding sheets. These branch cuts start either at branch points or at the real part.
Due to the second condition [[([H]{})]{}]{} fixes the labeling of the sheets. Together with condition [[([G]{})]{}]{} it ensures condition [[([E]{})]{}]{} in Theorem \[thm:Delta\]. With the holomorphic coordinates of this setting the corresponding covering can be compactified to a compact Riemann surface. Hence condition [[([G]{})]{}]{} is the finite type condition. Condition [[([I]{})]{}]{} ensures that the compactified covering space has genus zero. The analogous condition for finite sheeted coverings ${\mathbb{CP}^1}\rightarrow{\mathbb{CP}^1}$ is that the branching order is two times the number of sheets minus two [@Hur]. Condition [[([J]{})]{}]{} ensures the reality condition [[([F]{})]{}]{} in Theorem \[thm:Delta\]. This endows the spectral curve with the involutions $\sigma$, $\rho$ and $\eta$, the last of which is without fixed points. Finally condition [[([K]{})]{}]{} describes a choice of the corresponding branch cuts. All branch points away from the real part are the starting point of a unique branch cut running to very large or very small imaginary parts. In general, the branch cuts, which start at the real part can be moved along the real part without changing $\Delta$. In fact, if $\Delta_0\in(-2,2)$ is the starting point of a branch cut connecting the sheets $l^+$ and $m^+$, and if the sheets $l^+$ and $l^-$ and the sheets $m^+$ and $m^-$ are connected along a real branch cut running along some neighbourhood of $\Delta_0\in(-2,2)$, then this $\Delta_0$ is no branch point. Moreover the corresponding branch cut can be moved along the real part until it reaches either on the sheets $l^\pm$ or on the sheets $m^\pm$ a real branch point of $\Delta$ without changing $\Delta$.
\[thm:characterization\] All $\Delta$ obeying conditions [[([D]{})]{}]{}-[[([F]{})]{}]{} in Theorem \[thm:Delta\] fulfill conditions [[([G]{})]{}]{}-[[([J]{})]{}]{}. Moreover, the branch cuts may be chosen as described in condition [[([K]{})]{}]{}.
For all $r>0$ only finitely many of the infinite branch points of $\lambda\mapsto 2\cos(\sqrt{\lambda})$ belong to $B(0,r)\subset{\mathbb C}$. Hence condition [[([E]{})]{}]{} of Theorem \[thm:Delta\] implies condition [[([G]{})]{}]{}. For all $n\in\mathbb{N}$ the absolute value of the derivative $\sin$ of $\cos$ is not smaller than $1$ on the circles $|\ln\mu|=(n+\frac{1}{2})\pi$. Hence, due to Rouche’s Theorem, the number of zeroes of $\Delta'$ in the complement of small discs around $\kappa=\pm i$ with appropriate radius is the same as the corresponding number of the map . This argument is a slight variation of the Counting Lemma 2 in Chapter 2 of Pöschel and Trubowitz [@PoeT]. Therefore condition [[([E]{})]{}]{} of Theorem \[thm:Delta\] implies also condition [[([I]{})]{}]{}. In particular, the values of $\Delta$ at the branch points are bounded. Now condition [[([H]{})]{}]{} just fixes the labeling of the sheets. Condition [[([F]{})]{}]{} in Theorem \[thm:Delta\] implies condition [[([J]{})]{}]{}. Due to condition [[([I]{})]{}]{} at all non-real branch points there ends an unbounded branch cut, which may be chosen according to condition [[([K]{})]{}]{}. Moreover with the exception of two real branch points all real branch points correspond to an additional unbounded branch cut starting at the real part. Condition [[([K]{})]{}]{} describes a choice of these unbounded branch cuts.
In order to show the converse we shall show that we can deform all $\Delta$ obeying conditions [[([G]{})]{}]{}-[[([K]{})]{}]{} into the covering and use Lemma \[thm:Delta 2\]. It would be natural to do this with decreasing geometric genus. In Theorem \[thm:spec\_AE\] we shall concentrate on deformations decreasing the following number $$\label{eq:G}\begin{aligned}
G&=\mbox{\rm geometric genus}\quad + \sum\limits{_{\mbox{\rm\scriptsize{real
singularities}}}}\delta-\mbox{invariant
(see e.g. \cite{Ser})}\\
&=\frac{1}{2}\#\{\mbox{non real roots of }d\ln\mu\}+ \#\{\mbox{\rm
real branch points of }\Delta\}-1.
\end{aligned}$$
\[thm:connected\] Any covering map obeying conditions [[([G]{})]{}]{}-[[([K]{})]{}]{} may be continuously deformed with decreasing $G$ within this class into a covering map .
In a first step we deform without changing the genus any covering map obeying the five conditions [[([G]{})]{}]{}-[[([K]{})]{}]{} into a covering map, which has only branch points of first order. Nearby branch points of order $k$ at $\Delta=\pm 2$ any deformation of $\Delta\mp2=z^{k+1}$ into a polynomial $p(z)$ with at least $\frac{k}{2}$ distinct double roots deforms the branch point at $\Delta=\pm 2$ into simple branch points at $\Delta=\pm2$. A combination of such [**[continuous deformations]{}**]{} deforms all branch points at $\Delta=\pm 2$ into simple branch points at $\Delta=\pm 2$. The movement of all other branch points at $\Delta\not=\pm 2$ in $\Delta\in{\mathbb C}\setminus\{-2,2\}$ does not change the genus at all.
In a second step we increase the values of the real branch points at $\Delta=-2$ and decrease the values of the real branch points at $\Delta=2$ by moving them into $\Delta\in(-2,2)$. As a result all real singularities of are deformed into real zeroes of $d\ln\mu$ and the geometric genus becomes equal to $G$. In particular, $\Delta$ takes on the real part only values in $(-2,2)$. In the remaining steps we shall decrease the geometric genus.
In a third step we increase the values at those real branch points on $\Delta\in(-2,\,2)$, which are on the real part local minima, and decrease the values of $\Delta$ at those real branch points which are on the real part local maxima. If two real branch points coalesce, on $\Delta\in(-2,\,2)$, then we move them away from the real part. We may continue to shrink the real part, until only two real branch points of $\kappa\mapsto\Delta(\kappa)$ remain and converge against each other. In the limit we would obtain a spectral curve with two connected components. But we stop shortly before this happens. Consequently the sheets of the covering divide into the two groups labeled by two copies of ${\mathbb N}$, which are joined only by a small circle between the two remaining real branch points. Since all other simple branch points do not belong to the real part, they occur in complex conjugate pairs, which can be moved in complex conjugate directions.
In a fourth step we show that we can move the complex conjugate branch points over $\Delta\in{\mathbb C}\setminus\{\pm 2\}$ in such a way that the genus becomes at most equal to one. Let $L$ denote the minimum of all $l\in\mathbb{N}$, such that for all $k>l$ the sheets $l^{\pm}$ contain the branch points and possibly the two remaining real branch points, but no other branch point. In the subsequent discussion we neglect these real branch points, which in this step are not moved at all. Due to condition [[([G]{})]{}]{} the number $L$ is finite. Now we claim that we may inductively decrease this number $L$ until it is equal to $0$. In order to avoid branch points on vertical branch cuts we move all branch points at $\Delta\not=\pm 2$ to places with pairwise different real values of $\Delta$. Moreover, we can achieve that the real part of $\Delta$ takes at these simple branch points at $\Delta\not=\pm 2$ pairwise different values in $(-2,\,2)$, which simplifies the subsequent argument. Furthermore if the branch points on sheets with exponents $+$ cross the lines with real part of $\Delta$ equals to $\pm 2$ along negative imaginary values of $\Delta$ and on sheets with exponents $-$ along positive imaginary values of $\Delta$, the branch points do not cross the branch cuts described in . Consequently the number $L$ is preserved under this deformation. At $\Delta=(-1)^L2$ the sheets with labels $(L+1)^{\pm}$ join a unique branch point with sheets with labels not larger than $L$. Let $k^\pm$ denote the labels of these sheets. These sheets have besides these branch points at $\Delta=(-1)^L2$ only branch points connecting with sheets in $\{1^-,1^+,\ldots,L^-,L^+\}$. Due to condition [[([H]{})]{}]{} they have at least one other branch point. If $L$ is even and the sheets $k^{\pm}$ contain more than two branch points, then we move all branch points with the exception of those with minimal and maximal real parts of $\Delta$ starting with the smaller real parts through the vertical branch cut with the lowest real parts of $\Delta$. If $L$ is odd, then we start with the branch points with larger real parts of $\Delta$ and move these branch points through the branch cut with the largest real part of $\Delta$. Finally the sheets with label $k^{\pm}$ contain besides the branch point at $\Delta=(-1)^L2$ exactly one other branch point. If we move this branch point to $\Delta=-(-1)^L 2$ the number $L$ decreases. This proves the claim. Hence we can decrease the number $L$, until it is equal to $0$ and arrive at a covering map that corresponds to a spectral curve of genus one.
In a fifth step we finally move one of the two remaining real branch points away from the other real branch point several times along $\Delta\in[-2,\,2]$ to and fro until it eventually reaches $\Delta=2$ at the sheet with label one. In order to do so we distinguish two cases: If the real branch points are located on sheets with odd label, then we move the branch point with larger value of $\Delta$. If both real branch points are located on a sheet with even label, then we move the branch point with smaller value of $\Delta$. We thus obtain a covering map of the form corresponding to a spectral curve of geometric genus zero.
We remark, that it is also possible to deform all $\Delta$ obeying conditions [[([G]{})]{}]{}-[[([K]{})]{}]{} with decreasing geometric genus into $\Delta$ of geometric genus zero, but the proof is more complicated. Moreover, in Lemma \[thm:2 real zeroes\] and Lemma \[thm:continuous genus 1\] we shall deform those $\Delta$ corresponding to finite type cylinders, and again decrease $G$ instead of the geometric genus.
\[thm:weights\] All spectral curves of real finite type periodic solutions of the $\sinh$-Gordon equation correspond uniquely to covering maps that satisfy conditions [[([G]{})]{}]{}-[[([K]{})]{}]{} above. The geometric genus $g+1$ of the corresponding spectral curve is equal to the sum of weighted branch orders with the following weights: $$\mbox{weight} = \left\{ \begin{array}{cll} \mbox{branch order}
&\mbox{
at } \Delta \neq \pm 2, & \\
\tfrac{1}{2}(\mbox{branch order}) &\mbox{ at } \Delta = \pm 2
&\mbox{for
even branch order}, \\
\tfrac{1}{2}(\mbox{branch order}-1) &\mbox{ at } \Delta = \pm 2
&\mbox{for
odd branch order}. \end{array} \right.$$
Lemma \[thm:characterization\] shows that all $\Delta$ of periodic solutions of the $\sinh$-Gordon equation of finite type fulfill [[([G]{})]{}]{}-[[([K]{})]{}]{}. Due to Lemma \[thm:connected\] we can by suitable movements continuously deform an arbitrary $\Delta$ obeying conditions [[([G]{})]{}]{}-[[([K]{})]{}]{} into the family of covering maps , which corresponds to spectral curves of geometric genus zero. In Lemma \[thm:Delta 2\] it is shown that these deformations preserve those $\Delta$, which correspond to periodic solutions of the $\sinh$-Gordon equation of finite type. The form $d\ln\mu$ has $2g+2$ zeroes on the spectral curve. The formula for the genus is obtained by computing the order of the roots of $d\ln\mu$ on $\mu^2-\mu\Delta(\kappa) + 1 =0$.
One-sided Alexandrov embeddings in ${\mathbb S}^3$ {#sec:AE space}
==================================================
In this section we consider one-sided Alexandrov embeddings of general manifolds $N$ and $M$. Some statements apply only to one-sided Alexandrov embeddings with constant mean curvature, but we do not use the special properties of cylinders of finite type. We provide sufficient conditions which allow us to perturb surfaces which are one-sided Alexandrov embedded inside of a collar of the unperturbed surfaces into surfaces which remain one-sided Alexandrov embedded. For this purpose we have to ensure that the surfaces have collars with depths uniformly bounded from below. Making use of the fact that there are no complete stable minimal surfaces in ${\mathbb S}^3$, we present in Lemma \[th:Rosenberg\_Lemma\] a crucial technical result communicated to us by Harold Rosenberg [@Ros:com]: If both principal curvatures of an Alexandrov embedded surface are uniformly bounded, then the cut locus function is bounded from below by a positive number.
We consider one-sided Alexandrov embedded cylinders in ${\mathbb S}^3$. In the literature we only found the notion of Alexandrov embeddings for compact domains on the one hand, and the concept of properly Alexandrov embedded immersions from open manifolds into open Riemannian manifolds on the other hand. Since we are interested in immersions of open manifolds into the compact Riemannian manifold ${\mathbb S}^3$, we make the following
A [**[one-sided Alexandrov embedding]{}**]{} in ${\mathbb S}^3$ is a smooth immersion $f$ from a connected 3-manifold $N$ with connected boundary $M=\partial N$ to ${\mathbb S}^3$ with the following properties:
1. The mean curvature of $M$ in ${\mathbb S}^3$ with respect to the inward normal is non-negative everywhere.
2. The manifold $N$ is complete with respect to the metric induced by $f$.
An immersion $f:N\rightarrow{\mathbb S}^3$ obeying condition (ii) is called an [**[Alexandrov embedding]{}**]{}.
A fixed orientation of ${\mathbb S}^3$ induces on $N$ and $M=\partial N$ an orientation. Conversely, if $M$ is endowed with an orientation, then there exists a unique normal, which points inward to the side of $M$ in ${\mathbb S}^3$, which induces on the boundary $M$ the given orientation of $M$. In this sense the orientation of $M$ determines the inner normal of $N$.
For each point $p \in M$ of a hypersurface of a Riemannian manifold $N$ there exists a unique arc length parameterized geodesic $\gamma(p,\cdot)$ emanating from $p = \gamma(p,0)$ and going in the direction of the inward normal at $p$. Such geodesics are called [**[inward $M$-geodesics]{}**]{} [@Heb].
Let $\gamma(p,\cdot)$ be an inward $M$-geodesic. Points $q \in N$ in the ambient manifold that are ’close to one side’ of $M$ can thus be uniquely parameterized by $(p,\,t)$ where $p \in M$ and $q =
\gamma(p,t)$ for some inward $M$-geodesic $\gamma(p,\cdot)$ and some $t \in {\mathbb R}_0^+$. The value of $t$ is the geodesic distance of $q$ to $M$. Extending the geodesic further into $N$ it might eventually encounter a point past which $\gamma(p,t)$ no longer minimizes the distance to $M$. Such a point is called a cut point. The cut locus of $M$ in $N$ consists of the set of cut points along all inward $M$-geodesics. We define the [**[cut locus function]{}**]{} as the geodesic distance of the cut point to $M$: $$\label{eq:cutlocus}
c:M\rightarrow{\mathbb R}^+,\quad p\mapsto c(p),\quad\mbox{ such that }
\gamma(p,c(p))\mbox{ is the cut point.}$$ If we want to stress the dependence on $f$ we decorate $\gamma$ and $c$ with index $f$. A known fact from Riemannian Geometry ([@Heb Lemma 2.1]) asserts that a cut point is either the first focal point on an inward $M$-geodesic, or is the intersection point of two shortest inward $M$-geodesics of equal length.
For a one-sided Alexandrov embedding $f:N \to {\mathbb S}^3$ of a 3-manifold $N$ with boundary $\partial N = M$, the inward $M$-geodesics give us a parametrization of $N$, which we call [**[generalized cylinder coordinates]{}**]{}: $$\label{eq:cylinder}
\gamma_f : \{(p,t)\in M\times{\mathbb R}\mid 0\leq t\leq c_f(p)\}
\rightarrow N\,.$$ The restriction of $\gamma_f$ to $\{(p,t)\in M\times{\mathbb R}\mid 0\leq t<
c_f(p)\}$ is a diffeomorphism onto the complement of the cut locus. The cut locus is homeomorphic to the quotient space $M/\sim_f$ with the following equivalence relation on $M$: $$\begin{aligned}
p&\sim_f q&&\Longleftrightarrow&
c_f(p)&=c_f(q)\quad\mbox{ and }&\gamma_f(p,c_f(p))&=\gamma_f(q,c_f(q))\\
&&&\Longleftrightarrow&&&\gamma_f(p,c_f(p))&=\gamma_f(q,c_f(q)).\end{aligned}$$ For all $p\in M$ we denote the corresponding equivalence classes by $$\begin{aligned}
\label{eq:ec cutlocus}
[p]_f&=\{q\in M\mid \gamma_f(p,c_f(p))=\gamma_f(q,c_f(q))\}.\end{aligned}$$
\[thm:unique embedded 1\] Let $f:N\rightarrow{\mathbb S}^3$ and $\tilde{f}:\tilde{N}\rightarrow{\mathbb S}^3$ be two one-sided Alexandrov embeddings. The orientation of ${\mathbb S}^3$ induces orientations on $N$, $\partial N$, $\tilde{N}$ and $\partial\tilde{N}$. If the two oriented boundaries $\partial
N=M=\partial\tilde{N}$ and the two restrictions to the boundaries $f|_M=\tilde{f}|_M$ coincide, then there exists a diffeomorphism $\Psi:\tilde{N}\rightarrow N$, whose restriction to $M$ is the identity map of $M$, such that $\tilde{f}=f\circ \Psi$.
The immersions $f$ and $\tilde{f}$ induce on $N$ and $\tilde{N}$ two Riemannian metrics $g$ and $\tilde{g}$, which coincide on $M$. The Riemannian metrics induce on $N$ and $\tilde{N}$ two metrics $d$ and $\tilde{d}$. Let $c_f$ and $c_{\tilde{f}}$ denote the cut locus functions of the submanifold $M$ in the two Riemannian manifolds $(N,g)$ and $(\tilde{N},\tilde{g})$. The generalized cylinder coordinates define diffeomorphisms $\gamma_f$ and $\gamma_{\tilde{f}}$ of $$L=\{(p,t)\in M\times{\mathbb R}\mid 0\leq t<\min\{c_f(p),c_{\tilde{f}}(p)\}\}$$ onto open subsets of $N$ and $\tilde{N}$. First we claim that these diffeomorphisms $\gamma_f$ and $\gamma_{\tilde{f}}$ together with the metrics $d$ and $\tilde{d}$ induce on $L$ the same metrics. In fact, the subset of $L \times L$ on which the metrics coincide is open and closed and therefore all of $L\times L$. Consequently both cut locus functions $c_f$ and $c_{\tilde{f}}$ coincide, and for all points $p\in M$ the corresponding classes $[p]_f=[p]_{\tilde{f}}$ coincide too. Hence the diffeomeorphism $\gamma_f\circ\gamma^{-1}_{\tilde{f}}$ from the complement of the cut locus in $\tilde{N}$ onto the complement of the cut locus in $N$ extends to a homeomorphism $\Psi$ from $\Tilde{N}$ to $N$. By definition of the cylinder coordinates the immersions $f\circ\gamma_f$ and $\tilde{f}\circ\gamma_{\tilde{f}}$ from $L$ into ${\mathbb S}^3$ coincide. Hence we have $\tilde{f}=f\circ\Psi$. Since $f$ and $\tilde{f}$ are immersions, $\Psi$ is a diffeomorphism form $\tilde{N}$ onto $N$.
We call an immersion $f:W\rightarrow{\mathbb S}^3$ of a connected 3-manifold $W$ with connected boundary $V=\partial W$ a [**[local one-sided Alexandrov embedding]{}**]{}, if the following hold:
1. The mean curvature of $V$ in ${\mathbb S}^3$ with respect to the inward normal is non-negative everywhere.
2. All inward $V$-geodesics exist in $W$ until they reach the cut locus .
3. The generalized cylinder coordinates $\gamma_f$ are surjective.
An immersion obeying conditions (ii)-(iii) is called a [**[local Alexandrov embedding]{}**]{}.
If $f:N\rightarrow{\mathbb S}^3$ is an Alexandrov embedding, and $V\subset
M=\partial N$ is an open subset, which contains for all $p\in V$ the classes $[p]_f$ , then the restriction $f|_W$ of $f$ to $$W=\{\gamma_f(p,t)\in N\mid p\in V\mbox{ and }0\leq t\leq c_f(p)\}$$ is a local Alexandrov embedding. The proof of Lemma \[thm:unique embedded 1\] carries over to the following situation:
\[thm:unique embedded 2\] Let $f:W\rightarrow{\mathbb S}^3$ and $\tilde{f}:\tilde{W}\rightarrow{\mathbb S}^3$ be two local Alexandrov embeddings. If the oriented boundaries $\partial W=V=\partial\tilde{W}$ and $f|_V=\tilde{f}|_V$ coincide, then there exists a diffeomeorphism $\Psi:\tilde{W}\rightarrow W$, whose restriction to $V$ is the identity map of $V$, such that $\tilde{f}=f\circ \Psi$.
We shall prove that ’one-sided Alexandrov embeddedness’ is an open condition, which will allow us to study deformation families of one-sided Alexandrov embeddings. The main tool is a general perturbation technique of Alexandrov embeddings, which we call [**[collar perturbation]{}**]{}. We consider perturbations $\tilde{f}$ of a given smooth immersion $f:M\rightarrow{\mathbb S}^3$, which are ‘small’ with respect to the $C^1$-topology on the space of immersions from $M$ into ${\mathbb S}^3$. For this purpose we use the trivialization of the tangent bundle $T{\mathbb S}^3 \cong {\mathrm{SU}_{\mbox{\tiny{$2$}}}}\times {\mathfrak{su}_{\mbox{\tiny{2}}}}$ by left invariant vector fields.
For sake of simplicity we denote the derivative considered as a smooth function into ${\mathfrak{su}_{\mbox{\tiny{2}}}}$ by $f'$ instead of $f^{-1}df$, and endow ${\mathfrak{su}_{\mbox{\tiny{2}}}}$ with the norm $\|X\|=(-\frac{1}{2}{\mathrm{tr}}(X^2))^{\frac{1}{2}}$.
\[thm:injectivity radius\] Let $f:M\rightarrow{\mathbb S}^3$ be an immersion inducing on the 2-manifold $M$ a complete Riemannian metric. If the absolute values of both principal curvatures are bounded by $\kappa_{\max}>0$, then for some $r>0$ depending only on $\kappa_{\max}$ the exponential map $\exp_p$ is at all points $p\in M$ a diffeomorphism from $B(0,r)\subset T_pM$ onto an open neighbourhood of $p$. Furthermore, for all $\epsilon>0$ there exists a $\delta>0$ depending only on $\epsilon$ and $\kappa_{\max}$, such that $$\|(f\circ\exp_p)'(v)-(f\circ\exp_p)'(0)\|<\epsilon\quad
\mbox{ for all }p\in M\mbox{ and all }v\in B(0,\delta)\subset
T_pM.$$
Choose two radii $r,R\in(0,\pi)$ such that the following two conditions hold:
\(i) $r+R\leq\arctan(\kappa_{\max}^{-1})$.
\(ii) Let $q^\pm$ be the centers of two spheres of radius $R$ in ${\mathbb S}^3$ touching each other at the center $p$ of an open ball $B(p,r)\subset{\mathbb S}^3$. All geodesics in ${\mathbb S}^3$ emanating from $q^+$ or $q^-$ can intersect transversally the sphere of radius $R$ around $q^-$ and $q^+$ respectively, inside of $B(p,r)$.
For $R<\arctan(\kappa_{\max}^{-1})$ both conditions can be fulfilled for small $r$. For all $p\in M$ there are exactly two points $q^\pm$ on the $M$-geodesic through $p$, whose distance to $p$ are equal to $R$. Both spheres of radius $R$ and centers $q^\pm$ in ${\mathbb S}^3$ touch $M$ at $p$. The absolute values of both principal curvatures are not larger than the principal curvatures of all spheres in ${\mathbb S}^3$ with radius smaller than $\arctan(\kappa_{\max}^{-1})$. Hence on all geodesics of $M$ emanating from $p$ the distances to both centers $q^\pm$ are monotone increasing, as long as these distances are smaller than $\arctan(\kappa_{\max}^{-1})$. Due to condition (i) this is the case for all points on the geodesic, whose geodesic distance to $p$ is not larger than $r$. Hence the ball $B(p,r)\subset M$ is mapped by $f$ into the complement of both balls $B(q^\pm,R)\subset{\mathbb S}^3$. We parameterize the points of $B(p,r)\subset M$ by the intersection points of the shortest geodesics connecting these points with $q^\pm$ with the spheres of radius $R$ around $q^\pm$, respectively. Due to the second condition (ii) these parameters are smooth. Hence $\exp_p$ is on $B(0,r)\subset T_pM$ a diffeomorphism onto an open neighbourhood of $p$. Since on all $M$-geodesics through $p$ the distances to both centers $q^\pm$ is monotone increasing, the uniform estimate $\|(f\circ\exp_p)'(v)-(f\circ\exp_p)'(0)\|<\epsilon$ holds for all $p\in M$ and all sufficiently small $v\in T_pM$.
\[thm: cut locus function bound\] Let $f:N\rightarrow{\mathbb S}^3$ be a one-sided Alexandrov embedding with uniform upper bound $\kappa{_{\mbox{\rm\scriptsize{max}}}}$ on both principal curvatures. Then there exist constants $0<{c{_{\mbox{\rm\scriptsize{max}}}}}<\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})$ and $L>0$ depending only on $\kappa{_{\mbox{\rm\scriptsize{max}}}}$ such that for all $p\in M=\partial N$ with $c_f(p)<{c{_{\mbox{\rm\scriptsize{max}}}}}$ the set $[p]_f$ contains exactly two points. Furthermore, the angle between both corresponding inward $M$-geodesics at the cut locus $\gamma_f(p,c_f(p))$ is larger than $\pi-L\cdot c_f(p)$.
All 2-spheres of radius $t \in (0,\frac{\pi}{2}]$ in ${\mathbb S}^3$ have principal curvatures both equal to $\kappa =\cot(t)$ (for the computations see e.g. Montiel and Ros [@MonR:alex]). For every $p \in M$ there is a unique inward $M$-geodesic $\gamma(p,\cdot)$. Evaluating these geodesics at some $t \in {\mathbb R}^+_0$ gives a hypersurface $$\label{eq:hypersurface}
M_t = \bigcup_{p \in M} \gamma(p,t)$$ in $N$. This hypersurface is smooth as long as $\gamma(p,t)$ does not pass through a focal point. If $\kappa_1$ and $\kappa_2$ are the two principal curvatures of $M$ in $p$ with respect to the inner normal, then the corresponding principal curvatures of $M_t$ at $\gamma(p,t)$ are given by $\cot(\arctan(\kappa_1^{-1})-t)$ and $\cot(\arctan(\kappa_2^{-1})-t)$. The focal point on $\gamma(p,\cdot)$ is at the value of $t$ given by $$\label{eq:focalpoint}
t{_{\mbox{\rm\scriptsize{foc}}}} =
\arctan\left((\max\{\kappa_1,\,\kappa_2\})^{-1}\right) \geq
\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1}).$$ At all $p\in M$ with $c_f(p)<\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})$ the cut locus $\gamma_f(p,c_f(p))$ cannot be a focal point. Hence $[p]_f$ has to contain at least one other element $q\in
M\setminus\{p\}$. Choose two radii $r,R$ as in the proof of Lemma \[thm:injectivity radius\]. The two spheres in ${\mathbb S}^3$ of radius $R$, which touch $M$ at $p$ and $q$ outside of $N$, must not intersect each other at distances $<r$ to $p$ and $q$, respectively. Hence there exist ${c{_{\mbox{\rm\scriptsize{max}}}}}>0$ and $L>0$, such that for $c_f(p)<{c{_{\mbox{\rm\scriptsize{max}}}}}$ the angle between the inward $M$-geodesics $\gamma_f(p,\cdot)$ and $\gamma_f(q,\cdot)$ at the cut locus is larger than $\pi-L\cdot c_f(p)$. If ${c{_{\mbox{\rm\scriptsize{max}}}}}$ is small enough, this implies that any element of $[p]_f$ belongs to one of the balls $B(p,r)$ or $B(q,r)$ in $M$. If $c_f(p)$ is smaller than $R$, then due to Lemma \[thm:injectivity radius\] $[p]_f\cap B(p,r)=\{p\}$ and $[p]_f\cap B(q,r)=\{q\}$.
We thank Harold Rosenberg for communicating the following result to us. This arose through discussions with Antonio Ros and Harold Rosenberg, and makes use of a technique for constructing a global non-negative, non-trivial Jacobi field that was recently also employed by Meeks, Perez and Ros [@MeePR:limit].
\[th:Rosenberg\_Lemma\] Let $f:N \to {\mathbb S}^3$ be a one-sided Alexandrov embedding with constant mean curvature, and principal curvatures bounded by $\kappa{_{\mbox{\rm\scriptsize{max}}}}>0$. Then the cut locus function is bounded from below by $\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})$.
The mean curvature of the hypersurface $$\label{eq:mean curvature}
H(t) = \tfrac{1}{2}\left(
\cot\left(\arctan(\kappa_1^{-1})-t\right) +
\cot\left(\arctan(\kappa_2^{-1})-t\right) \right)$$ is positive for all $t \in (0,\,t{_{\mbox{\rm\scriptsize{foc}}}})$, and strictly monotone increasing, since $$H'(t) = \tfrac{1}{2}\left(\sin^{-2}(\arctan(\kappa_1^{-1})-t) +
\sin^{-2}(\arctan(\kappa_2^{-1})-t)\right) > 0\,.$$ Let $c_f$ denote the cut locus function . If there exists a point $p \in M$ for which $c_f(p) <
\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})\leq t{_{\mbox{\rm\scriptsize{foc}}}}$, then two inward $M$-geodesics $\gamma(p,\cdot),\,\gamma(q,\cdot)$ through $p,\,q \in
M$ respectively, have to intersect at a distance of $c_f(p)$ from $M$, and thus $c_f(p) = c_f(q)$. Hence, if there exists a point $p
\in M$ with $c_f(p) <\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})$ then $M_t$ intersects itself for a value of $t < \arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})$ over two points $p,\,q \in M$. Let $${c}_0 = \inf\{t\,\left| \right. M_t \mbox{ intersects over two
points of } M\}.$$ Since over all points $p \in M$ the mean curvature of $M_t$ is positive for all $0<t<\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})$ with respect to the inner normals, the surfaces $M_t$ cannot touch each other from different sides over two points for $0<t<\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})$.
A normal graph in ${\mathbb S}^3$ over a domain in a geodesic sphere ${\mathbb S}^2$ is the graph of the composition of the exponential map with a section of the normal bundle on this domain, see Fornari, deLira and Ripoll [@ForLR]. By Lemma \[thm:injectivity radius\] there exists a $r>0$ depending only on $\kappa{_{\mbox{\rm\scriptsize{max}}}}$ such that all $p \in M$ have open neighbourhoods in $M$, which are normal graphs over the ball $B(p,\,r)$ inside the unique geodesic 2-sphere ${\mathbb S}^2$, which touches $M$ at $p$. Due to Arzelà-Ascoli, and the a priori gradient bound from Proposition 4.1 in [@ForLR], every bounded sequence of normal graphs over $B(p,\,r) \subset {\mathbb S}^2$ has a convergent subsequence. Now let $(p_k)_{k\in{\mathbb N}}$ be a sequence in $M$ with $$\lim_{k \to\infty}c_f(p_k) = c_0 = \inf\left\{c_f(p) \mid p \in
M \right\}\,.$$ Then there exists a sequence $\Theta_k$ of isometries of ${\mathbb S}^3$ which transform each point $p_k$ into a fixed reference point $p_0
\in {\mathbb S}^3$, and the tangent plane of $M$ at $p_k$ into the tangent plane of a fixed geodesic sphere ${\mathbb S}^2_{p_0}\subset {\mathbb S}^3$ which contains $p_0$. This sequence of isometries transforms neighbourhoods $U_k$ of $p_k\in M$ into normal graphs $\Theta_k[U_k]$ over $B(p_0,\,r)$. By passing to a subsequence we may achieve that these graphs converge to a normal graph $U$ over $B(p_0,\,r) \subset {\mathbb S}^2_{p_0}$, which is tangent to ${\mathbb S}^2_{p_0}$ at $p_0$. Due to Corollary \[thm: cut locus function bound\], the sets $[p_k]_f$ contain besides $p_k$ another point $q_k$ for large $k$. Furthermore, the sequence of isometries $\Theta_k$ transforms the sequence of geodesic 2-spheres tangent to $M$ at $q_k$ into a converging sequence of spheres with limit ${\mathbb S}^2_{q_0}$. This sphere contains the limit $q_0=\lim\Theta_k(q_k)$ with distance $\mathrm{dist}(p_0,\,q_0) =
2c_0$. The shortest geodesic connecting $p_0$ and $q_0$ intersects orthogonally both geodesics spheres ${\mathbb S}^2_{p_0}$ and ${\mathbb S}^2_{q_0}$. By Corollary \[thm: cut locus function bound\], for large $k$ the points $q_k$ have neighbourhoods $V_k$, whose transforms $\Theta_k[V_k]$ are normal graphs over $B(q_0,\,r)\subset{\mathbb S}^2_{q_0}$. By passing again to a subsequence the normal graphs $\Theta_k[V_k]$ converge to a normal graph $V$ tangent to $Sp^2_{q_0}$ at $q_0$.
The transformed inward $M$-geodesics nearby $p_k$ and $q_k$ converge to normal geodesics of these two limiting surfaces $U$ and $V$ in ${\mathbb S}^3$. If we shift both limiting surfaces by $c_0$ along their normal geodesics as in we obtain two surfaces touching each other from different sides at the limit of the transformed cut points $\lim\Theta_k(\gamma_f(p_k,\,c(p_k)))=
\lim\Theta_k(\gamma_f(q_k,\,c(q_k)))$. Hence the shifted surfaces cannot have positive mean curvature with respect to the inner normal. This implies ${c}_0=0$ and $H=0$. In this case ${\mathbb S}^2_{p_0}={\mathbb S}^2_{q_0}$ and the two minimal surfaces $U$ and $V$ are tangent at $p_0=q_0$ and are graphs over the same domain $B(p_0,\,r)$. By Hopf’s maximum principle both limiting minimal normal graphs $U$ and $V$ coincide.
In this case, due to Corollary \[thm: cut locus function bound\], for large $k$ neighbourhoods $W_k\subset M$ of $q_k$ are normal graphs over $U_k\subset M$. Let ${n}_k$ be the corresponding sequence of non-negative sections of the inward normal bundle of the sequence $\Theta_k[U_k]$ of minimal graphs over $B(p_0,\,r)$. Then $\Theta_k[W_k]$ are normal graphs over the sequence $\Theta_k[U_k]$ of normal graphs over $B(p_0,r)$. Then $$\lim_{k\to\infty} \frac{{n}_k}{\sup
\left\{|{n}_k(q)| \mid q \in \Theta_k[U_k] \right\}}$$ converges to a non-trivial non-negative Jacobi field on the limiting minimal graph $U$ over $B(p_0,\,r)$. This argument is a slight variation of an observation by Meeks, Perez and Ros [@MeePR:limit].
Since the limits of both sequences of normal minimal graphs coincide, the function $|{n}_k|$ converges on $B(p_0,r)$ uniformly to zero. Hence we may repeat the same line of argument at all points of the boundary of these minimal normal graphs $\Theta_k[U_k]$ over $B(p_0,\,r)$. By passing to a diagonal subsequence like in the Arzelà-Ascoli theorem we may extend $U$ to a complete minimal surface in ${\mathbb S}^3$ with a non-negative Jacobi field which does not vanish identically. (Note that due to unique continuation of Jacobi fields [@Wol:elliptic] the limit of the re-scaled difference of both normal graphs stays bounded on all connected compact subsets of the complete minimal surface, whenever it is bounded on one compact subset.) A non-trivial non-negative Jacobi field implies that the spectrum of the Jacobi operator is non-negative (Allegretto-Piepenbrink Theorem, see Fischer-Colbrie and Schoen [@Fis-ColS] Theorem 1 and Davies [@Dav] Lemma 4.1). Therefore the limiting complete minimal surface is stable. But by a result of Fischer-Colbrie [@Fis-ColS] ( see also Corollary 3 in Schoen [@Sch:stable] ) there exist no complete stable minimal surfaces in ${\mathbb S}^3$. This contradicts $c_0 <
\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})$.
Later we apply the collar deformation to bounded open subsets $W\subset N$ of one-sided Alexandrov embeddings $f:N\rightarrow{\mathbb S}^3$. Hence we shall find bounded open subsets $V\subset M=\partial N$, which contain for all $p\in V$ the classes $[p]_f$ . For this purpose we need a [**[chord-arc]{}**]{} bound:
\[thm:ec diameter bound\] Let $f:N\rightarrow{\mathbb S}^3$ be a one-sided Alexandrov embedding with second fundamental form ${\mathfrak{h}}$ with respect to the inner normal $\mathfrak{N}$. If ${c}$ is a lower bound on the cut locus functions $c_f$ and ${D}$ a uniform bound on the covariant derivative of the second fundamental form: $$\label{eq:upper bound derivative of h}
|(\nabla_X{\mathfrak{h}})(X,X)|=|\nabla_X({\mathfrak{h}}(X,X))-2{\mathfrak{h}}(\nabla_XX,X)|\leq{D}|X|^3
\quad\mbox{ for all }X\in TM,$$ then there exists $C>0$ depending only on ${c}$ and ${D}$, such that $$\operatorname{dist}_N(p,q)\leq\operatorname{dist}_M(p,q)\leq C\operatorname{dist}_N(p,q)\quad
\mbox{ for all }p,q\in M.$$
For all $p,q\in M$ we have $\operatorname{dist}_N(p,q)\leq\operatorname{dist}_M(p,q)$. In general, these distances do not coincide. We shall construct a path from $p$ to $q$ of length at most $C\operatorname{dist}_N(p,q)$. Due to ( Rinow [@Rin], pages 172 and 141) the points $p$ and $q$ are joined by a shortest path in $N$. In case this shortest path touches at some points the boundary [@AA Theorem 1.] we decompose it into pieces. The boundary points of a shortest path might have accumulation points. But any point of a shortest path, which is not a boundary point, belongs to a unique geodesic piece in $N$, which has only two boundary points at both ends. Hence it suffices to construct such a path for two points $p$ and $q$, which are connected by a geodesic in $N$ with only two boundary points at both ends.
Due to [@Heb Lemma 2.1] the cut locus function $c_f(p)$ is for all $p\in M$ not larger than the first focal point $t{_{\mbox{\rm\scriptsize{foc}}}}$ . Hence both principal curvatures are uniformly bounded by $\kappa{_{\mbox{\rm\scriptsize{max}}}}=\cot({c})$. If $\operatorname{dist}_N(p,q)\leq{c}$, then $$\begin{aligned}
\operatorname{dist}_N(\gamma_f(p,\operatorname{dist}_N(p,q)),q)&\geq\operatorname{dist}_N(p,q)\quad\mbox{ and }&
\operatorname{dist}_N(\gamma_f(q,\operatorname{dist}_N(p,q)),p)&\geq\operatorname{dist}_N(p,q).\end{aligned}$$ Hence the angles between the geodesic connecting $p$ and $q$, and the inward $M$-geodesics at $p$ and $q$ are not smaller than the angles of the triangle in ${\mathbb S}^2$ with three sides of length $\operatorname{dist}_N(p,q)$, which is larger than $\frac{\pi}{3}$. If we reduce the length of $\operatorname{dist}_N(p,q)$ along the gradient flow of $\operatorname{dist}_N$ on $M\times M$, then we obtain a path in $M\times M$ from $(p,q)$ to the diagonal in $M\times M$ of length smaller than $2\operatorname{dist}_N(p,q)$. This shows the claim for short distances $\operatorname{dist}_N(p,q)\leq{c}$. On the other hand, all geodesics $\gamma$ in $N$ starting at $p$, which do not meet $M$ for distances $d\in(0,\pi)$ meet each other at the antipode of $p$. The exponential map of $N$ maps a unique half space of $T_pN$ into $N$. If the pre-image of $B(p,\pi)\subset N$ with respect to $\exp_p$ contains this half space, then due to Hopf’s maximum principle (see e.g. [@Esc]) $B(p,\pi)\subset M$ is a geodesic sphere in ${\mathbb S}^3$ and the statement is obvious. Otherwise there exists a geodesic starting at $p$, which touches $M$ for some $t\in(0,\pi)$. In particular, if $\operatorname{dist}_N(p,q)\geq\pi$, then there exists $\tilde{q}\in M$ with $$\operatorname{dist}_N(p,\tilde{q}) < \pi\quad\mbox{ and }\quad
\operatorname{dist}_N(p,q)\geq\operatorname{dist}_N(p,\tilde{q})+\operatorname{dist}_N(\tilde{q},q).$$ Hence we may assume ${c}<\operatorname{dist}_N(p,q)<\pi$. Let $\chi_p,\,\chi_q
\in [0,\frac{\pi}{2}]$ denote the angles in $T_pN$ and $T_qN$ between the inward geodesic $\gamma$ connecting $p$ and $q$ and the inward normal to $M$, respectively. In this proof we shall consider smooth families of geodesics $\gamma$ connecting two smooth paths $s\mapsto p(s)$ and $s\mapsto q(s)$ in $M$ parameterized by a real parameter $s$. For fixed $s$ the geodesic is parameterized by the real parameter $t$. The derivatives with respect to $s$ are denoted by prime and the derivatives with respect to $t$ by dot. Let $|p'|$ and $|q'|$ denote the lengths of the tangent vectors $p'$ and $q'$ with respect to the Riemannian metric. The function $$\label{eq:gradient}
|\nabla\operatorname{dist}_N(p,q)|=
\sup\left\{\tfrac{|\operatorname{dist}_N'(p,q)|}{|p'|+|q'|}\mid p'\in T_pM,q'\in
T_qM\right\}= \max\{|\tan(\chi_p)|,|\tan(\chi_q)|\}$$ is the length of the gradient of the function $\operatorname{dist}_N$ on $M\times
M$. The geodesic $\gamma$ extends in ${\mathbb S}^3$ to a closed geodesic. For any tangent vector $(p',q')\in T_pM\times T_qM$ there exists a Killing field $\vartheta$ on ${\mathbb S}^3$, which moves the closed geodesic $\gamma$ in such a way, that the intersection points at $p$ and $q$ moves along $p'$ and $q'$, respectively. Conversely, all Killing fields $\vartheta$ generate a one-dimensional group of isometries of ${\mathbb S}^3$. Let $s\mapsto\gamma_\vartheta(s,\cdot)$ denote the corresponding family of geodesics and $s\mapsto(p_{\vartheta}(s),q_{\vartheta}(s))$ the corresponding intersection points with $M$. In order to proceed we need
\[thm:second derivative\] For all $\kappa{_{\mbox{\rm\scriptsize{max}}}}>0$ and $0<{c}<\pi$ there exist $\epsilon,\delta>0$ and $0<s_0<\frac{3}{2}$ with the following property: For all $(p,q)\in M\times M$ with ${c}\leq\operatorname{dist}_N(p,q)<\pi$ and $\max\{\tan(\chi_p),\,\tan(\chi_q)\}\leq\epsilon$ there exists a non trivial Killing field $\vartheta$, such that $d:s\mapsto
d(s)=\operatorname{dist}_N(p_{\vartheta}(s),q_{\vartheta}(s))$ obeys $$\begin{aligned}
\label{eq:second derivatives}
d'(s)&\leq 0& d''(s)&\leq-\delta\cos\left(\tfrac{d(s)}{2}\right)&
|p'_\vartheta(s)|+|q'_\vartheta(s)|&\leq 3\cos\left(\tfrac{d(s)}{2}\right)&
\mbox{ for all }&s\in[0,s_0].\end{aligned}$$
We shall construct a Killing field $\vartheta$ with the desired properties, which rotates $\gamma$ around two antipodes of $\gamma$. The corresponding rotated geodesics $\gamma_\vartheta(s,\cdot)$ belong to a unique geodesic 2-sphere ${\mathbb S}^2\subset{\mathbb S}^3$. The corresponding paths $s\mapsto p(s)$ and $s\mapsto q(s)$ move along the intersection of this sphere ${\mathbb S}^2$ with $M$. Hence we can calculate all derivatives on this sphere.
We parameterize this sphere by the real parameter $s$ of the family $s\mapsto\gamma_{\vartheta}(s,\cdot)$ of rotated geodesics, and the real parameter $t$ of these geodesics. We choose the equator as the start points $\gamma_{\vartheta}(s,0)$ with distance $\frac{\pi}{2}$ to the rotation axis. The vector fields $\vartheta$ and the geodesic vector field $\dot{\gamma}$ along the geodesics $\gamma_\vartheta(s,\cdot)$ form an orthogonal base of the tangent space $T{\mathbb S}^2$ of this sphere away from the zeroes of $\vartheta$. The vector fields $\vartheta$ and $\dot{\gamma}$ have at $(s,t)$ the scalar products $$\begin{aligned}
g(\vartheta,\vartheta)&=\cos^2(t)&
g(\vartheta,\dot{\gamma})&=0&
g(\dot{\gamma},\dot{\gamma})&=1.\end{aligned}$$ Since $\dot{\gamma}$ is a geodesic vector field the derivative $\nabla_{\dot{\gamma}}\dot{\gamma}$ vanishes. Moreover, the mean curvature of the integral curve of $\vartheta$ starting at $(s,t)$ is equal to $\tan(t)$. Therefore at $(s,t)$ we have $$\begin{aligned}
\nabla_{\vartheta}\vartheta&
=\cos^2(t)\tan(t)\dot{\gamma}
=\cos(t)\sin(t)\dot{\gamma}\,,&
\nabla_{\vartheta}\dot{\gamma}&
=-\tan(t)\vartheta \,,\\
\nabla_{\dot{\gamma}}\vartheta&
=-\tan(t)\vartheta \,,&
\nabla_{\dot{\gamma}}\dot{\gamma}& =0\,.\end{aligned}$$ We parameterize a neighbourhood of the geodesic from $p$ to $q$ in such a way that the corresponding vector field $\dot{\gamma}$ points inward to $N$ at $p$ and outward of $N$ at $q$, respectively. Let $(s_p,t_p)$ and $(s_q,t_q)$ be the coordinates of $p$ and $q$, respectively. The Killing field $\vartheta$ induces along the paths $s\mapsto p(s)$ and $s\mapsto q(s)$ the vector fields $$\begin{aligned}
p'&=\vartheta(p)-\dot{\gamma}(p)
\frac{g(\mathfrak{N}(p),\vartheta)(p)}
{g(\mathfrak{N}(p),\dot{\gamma}(p))}\quad\mbox{ and }&
q'&=\vartheta(q)-\dot{\gamma}(q)
\frac{g(\mathfrak{N}(q),\vartheta(q))}
{g(\mathfrak{N}(q),\dot{\gamma}(q))}\end{aligned}$$ with lengths $|p'|={}\frac{|\cos(t_p)|}{\cos(\chi_p)}$ and $|q'|={}\frac{|\cos(t_q)|}{\cos(\chi_q)}$. The derivative $d'$ is equal to $$d'=
\frac{g(\mathfrak{N}(p),\,\vartheta(p))}
{g(\mathfrak{N}(p),\,\dot{\gamma}(p))}-
\frac{g(\mathfrak{N}(q),\,\vartheta(q))}
{g(\mathfrak{N}(p),\,\dot{\gamma}(q))}=
\frac{g(\mathfrak{N}(p),\,\vartheta(p))}
{\cos(\chi_p)}-
\frac{g(\mathfrak{N}(q),\,\vartheta(q))}
{\cos(\chi_q)}.$$ Along the paths $p$ and $q$ with $X=p'$ and $X=q'$, respectively, we have at $(s,t)$ $$\begin{aligned}
\nabla_X\frac{g(\mathfrak{N},\,\vartheta)}
{g(\mathfrak{N},\,\dot{\gamma})}&=
\frac{g(\nabla_X\mathfrak{N},\,\vartheta)+
g(\mathfrak{N},\,\nabla_X\vartheta)}
{g(\mathfrak{N},\,\dot{\gamma})}-
\frac{g(\mathfrak{N},\vartheta)(
g(\nabla_X\mathfrak{N},\,\dot{\gamma})+
g(\mathfrak{N},\,\nabla_X\dot{\gamma}))}
{(g(\mathfrak{N},\,\dot{\gamma}))^2}\\
&=\frac{g(\nabla_X\mathfrak{N},X)+g(\mathfrak{N},\nabla_X\vartheta)}
{g(\mathfrak{N},\dot{\gamma})}-
\frac{g(\mathfrak{N},\theta)
g(\mathfrak{N},\nabla_X\dot{\gamma})}
{g(\mathfrak{N},\dot{\gamma})^2}\\
&=-\frac{{\mathfrak{h}}(X,\,X)}{g(\mathfrak{N},\dot{\gamma})}
+\cos(t)\sin(t)
+2\tan(t)\left(\frac{g(\mathfrak{N},\,\vartheta)}
{g(\mathfrak{N},\,\dot{\gamma})}\right)^2.\end{aligned}$$ Hence the second derivative is equal to $$\begin{aligned}
d''&=-\frac{{\mathfrak{h}}(p',\,p')}{\cos(\chi_p)}-\frac{{\mathfrak{h}}(q',\,q')}{\cos(\chi_q)}
+\frac{\sin(2t_p)-\sin(2t_q)}{2}\\
&+2\tan(t_p)\left(\frac{g(\mathfrak{N}(p),\,\vartheta(p))}
{g(\mathfrak{N}(p),\,\dot{\gamma}(p))}\right)^2
-2\tan(t_q)\left(\frac{g(\mathfrak{N}(q),\,\vartheta(q))}
{g(\mathfrak{N}(q),\,\dot{\gamma}(q))}\right)^2.\end{aligned}$$ If along the rotation of the geodesic for $s\in[0,s_0]$ the following inequalities are satisfied $$\begin{aligned}
\label{eq:assumption 1}
-\tfrac{\pi}{2}\leq t_p&\leq 0,&
0\leq t_q&\leq\tfrac{\pi}{2},&
\tfrac{{c}}{2}&\leq t_q-t_p,\quad\mbox{ and }&
\max\left\{\tan(\chi_p),\tan(\chi_q)\right\}&\leq\tfrac{1}{\sqrt{3}},\end{aligned}$$ then $\frac{{c}}{2}\leq d=t_q-t_p$, $\min\{\cos(\chi_p),\cos(\chi_q)\}\geq\tfrac{\sqrt{3}}{2}$ and the last two terms of $d''$ are bounded by $$\begin{aligned}
\left|\tan(t_p)\left(\frac{g(\mathfrak{N}(p),\,\vartheta(p))}
{g(\mathfrak{N}(p),\,\dot{\gamma}(p))}\right)^2\right|&\leq
\sin(|t_p|)\cos(t_p)\tan^2(\chi_q)
\leq\frac{\sin(2|t_p|)}{2\cdot 3}\\
\left|\tan(t_q)\left(\frac{g(\mathfrak{N}(q),\,\vartheta(q))}
{g(\mathfrak{N}(q),\,\dot{\gamma}(q))}\right)^2\right|&\leq
\sin(|t_p|)\cos(t_p)\tan^2(\chi_q)
\leq\frac{\sin(2|t_q|)}{2\cdot 3}.\end{aligned}$$ Due to $\sin(2t_q)-\sin(2t_p) = 2\sin(t_q-t_p)\cos(t_p+t_q)$ and $\cos(t_p)+\cos(t_q) = 2\cos(\tfrac{t_q-t_p}{2})\cos(\tfrac{t_p+t_q}{2})$ we arrive at $$\begin{aligned}
d''(s)&\leq
-\frac{{\mathfrak{h}}(p',\,p')}{\cos(\chi_p)}-\frac{{\mathfrak{h}}(q',\,q')}{\cos(\chi_q)}
-\sin(d)\cos(t_p+t_q)\left(1-\tfrac{2}{3}\right),\\
|p'|+|q'|&={}\frac{\cos(t_p)}{\cos(\chi_p)}
+{}\frac{\cos(t_p)}{\cos(\chi_p)}
\leq{}\tfrac{4}{\sqrt{3}}\cos\left(\tfrac{d}{2}\right)
\cos\left(\tfrac{t_p+t_q}{2}\right)\leq3{}\cos(\tfrac{d}{2}).\end{aligned}$$ The assumption implies $t_p\leq-d+\frac{\pi}{2}$ and $d-\frac{\pi}{2}\leq t_q$. For $d\in[\frac{\pi}{2},\pi)$ we get $\cos(t_p)\leq\sin(d)$ and $\cos(t_q)\leq\sin(d)$. For $d\in[\frac{{c}}{2},\frac{\pi}{2})$ we use $\cos(t_p)\leq 1$ and $\cos(t_q)\leq 1$ and obtain $$\begin{aligned}
\sin\left(\tfrac{{c}}{4}\right)\leq\tfrac{1}{2}\sin\left(\tfrac{{c}}{2}\right)
\leq\tfrac{3\sqrt{3}}{8}\sin\left(\tfrac{{c}}{2}\right)&
\leq\sin(d)\min\left\{\tfrac{\cos(\chi_p)}{|p'|^2},
\tfrac{\cos(\chi_q)}{|q'|^2}\right\}.\end{aligned}$$ The third inequality of implies $2\sin(\frac{{c}}{4})\cos(\frac{d}{2})\leq\sin(d)$. To sum up, the second and the third inequalities of are implied by and $$\begin{aligned}
\label{eq:assumption 2}
\delta&\leq\frac{1}{3}\sin\left(\tfrac{{c}}{4}\right)\cos(t_p+t_q),&
-\frac{{\mathfrak{h}}(p',\,p')}{|p'|^2}&
\leq\tfrac{1}{2}\delta\quad\mbox{ and }&
-\frac{{\mathfrak{h}}(q',\,q')}{|q'|^2}&
\leq\tfrac{1}{2}\delta.\end{aligned}$$ We shall show first that there exists a vector field $\vartheta$ obeying at $s=0$ $$\begin{aligned}
\delta&\leq\frac{1}{6}\sin\left(\tfrac{{c}}{4}\right)\cos(t_p+t_q),&
-\frac{{\mathfrak{h}}(p',\,p')}{|p'|^2}&
\leq\tfrac{1}{4}\delta\quad\mbox{ and }&
-\frac{{\mathfrak{h}}(q',\,q')}{|q'|^2}&
\leq\tfrac{1}{4}\delta.\end{aligned}$$ The Killing field $\vartheta$ is uniquely determined by two choices:
Firstly the choice of a geodesic sphere ${\mathbb S}^2\subset{\mathbb S}^3$, which contains the closed geodesic from $p$ to $q$.
Secondly a choice of the zeroes of $\vartheta$, or equivalently a choice of the coordinates $t_p$ and $t_q$ with $t_q-t_p=d\mod \pi$. We start with $t_q=-t_p=\frac{d}{2}$.
We first choose a 2-sphere ${\mathbb S}^2\subset{\mathbb S}^3$. This 2-sphere is uniquely determined by a choice of a line in $T_p{\mathbb S}^2 \cap T_pM$ which includes $p'$, or equivalently a choice of a line in $T_q{\mathbb S}^2
\cap T_qM$ which includes $q'$. Since $f$ is a one-sided Alexandrov embedding and both principal curvatures are uniformly bounded by $\kappa{_{\mbox{\rm\scriptsize{max}}}}$, the cone angles of those double cones in $T_pM$ and $T_qM$, on which ${\mathfrak{h}}(X,X)\geq-\frac{1}{4}\delta|X|^2$, are not smaller than $\tfrac{\pi}{2}+{{\rm O}}(\delta)$. For sufficiently small $\epsilon\geq\max\{\tan(\chi_p)\tan(,\chi_q)\}$ these double cones correspond in the plane orthogonal to $\dot{\gamma}(p)$ in $T_pN$ and in the plane orthogonal to $\dot{\gamma}(q)$ in $T_qN$ to double cones with cone angles not smaller than $\tfrac{\pi}{2}$. Hence the intersection of both double cones is non empty.
In a second step we shall show that the inequalities and are satisfied for $s\in[0,\,s_0]$ with some $s_0>0$. Due to Lemma \[thm:injectivity radius\] and the assumption $\max\{\tan(\chi_p),\tan(\chi_q)\}\leq\epsilon$ there exists $s_0$ such that $t_p$ and $t_q$ do not reach the roots of $\vartheta$ for $s\in[0,\,s_0]$. Since the derivatives of $\cos(\chi_p)$, $\cos(\chi_q)$, $t_p$ and $t_q$ with respect to $s$ are uniformly bounded, there exists $s_0>0$ such that the inequalities and the first inequality of are satisfied for $s\in[0,\,s_0]$. Due to also the derivatives of ${\mathfrak{h}}(p',p')$ and ${\mathfrak{h}}(q',q')$ are uniformly bounded. Hence there exists $s_0>0$ only depending on ${c}$ and $\kappa{_{\mbox{\rm\scriptsize{max}}}}$, such that the second and the third inequality of are satisfied for $s\in[0,\,s_0]$.
Finally we have to satisfy the first inequality of . At the start point $s=0$ this is always the case for one choice of the sign of $\vartheta$. Now the second inequality of implies the first.
*Continuation of the proof of Lemma \[thm:ec diameter bound\].* By Lemma \[thm:second derivative\] there exists $\epsilon,\delta>0$ and $0<s_0<\frac{2}{3}$, such that for all all $p,q\in M$ with ${c}\leq\operatorname{dist}_N(p,q)<\pi$ and $\left|\nabla\operatorname{dist}_N(p,q)\right|\leq\epsilon$ there exists a Killing field $\vartheta$, along which the length $d$ is reduced for $0\leq
s\leq s_0$. The inequality $-d'(s)\leq|p'(s)|+|q'(s)|\leq
3\cos(\frac{d(s)}{2})\leq 3\cos(\frac{d(s_0)}{2})$ implies $\cos(\frac{d(0)}{2})\geq(1-\frac{3}{2}s_0)\cos(\frac{d(s_0)}{2})$. Hence we get $$\begin{aligned}
\frac{d(0)-d(s_0)}{\operatorname{dist}_M(p(s_0),p(0))+\operatorname{dist}_M(q(s_0),q(0))}&\geq
\frac{\delta\frac{s_0^2}{2}\cos\left(\frac{d(0)}{2}\right)}
{3s_0\cos\left(\frac{d(s_0)}{2}\right)}
\geq\delta\frac{s_0}{6}\left(1-\tfrac{3}{2}s_0\right).\end{aligned}$$ We apply this procedure again and again, until either $\operatorname{dist}_N(p,q)<{c}$ or $\left|\nabla\operatorname{dist}_N(p,q)\right|>\epsilon$. As long as the gradient $\left|\nabla\operatorname{dist}_N(p,q)\right|>\epsilon$, the same estimate holds for the corresponding gradient on the Riemannian manifold $M\times
M$. Furthermore, the corresponding gradient flow reduces $\operatorname{dist}_N$ with monotonic decreasing $\frac{\epsilon}{\sqrt{2}}\operatorname{dist}_M(p(s),p(0))+
\frac{\epsilon}{\sqrt{2}}\operatorname{dist}_M(q(s),q(0))+\operatorname{dist}_N(p(s),q(s))$. In summary, all points $p,q\in M$ obey $\operatorname{dist}_M(p,q)\leq\max\{2,\,\frac{\sqrt{2}}{\epsilon},\,
\frac{12}{\delta(2s_0-3s_0^2)}\}\operatorname{dist}_N(p,q)$.
(Collar perturbation)\[thm:collar deformation 1\] For given $\kappa{_{\mbox{\rm\scriptsize{max}}}}>0$ and $0\leq H{_{\mbox{\rm\scriptsize{max}}}}\leq\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})$ there exist $\epsilon>0$ and $R>0$ with the following property: If $f:N\rightarrow{\mathbb S}^3$ is a one-sided Alexandrov embedding with constant mean curvature $0\leq H\leq H{_{\mbox{\rm\scriptsize{max}}}}$ and principal curvatures bounded by $\kappa{_{\mbox{\rm\scriptsize{max}}}}>0$, $p\in M$ is some point, and $\tilde{f}:M\rightarrow{\mathbb S}^3$ is an immersion with constant mean curvature $\tilde{H}\geq 0$ obeying $$\begin{aligned}
\label{eq:immersion bound}
\operatorname{dist}(f(q),\,\tilde{f}(q)) &< \epsilon\mbox{ and}&
\|f'(q)-\tilde{f}'(q)\| &< \epsilon
&\mbox{for all }q&\in B(p,R)\mbox{ and}&
|H-\tilde{H}|&< \epsilon.\end{aligned}$$ Then $\tilde{f}$ extends to a local one-sided Alexandrov embedding $\tilde{f}:W\rightarrow{\mathbb S}^3$, whose boundary $V=\partial W\subset B(p,R)$ is an open neighbourhood of $p$.
Due to Rosenberg’s Lemma \[th:Rosenberg\_Lemma\] ${c}=\arctan(\kappa{_{\mbox{\rm\scriptsize{max}}}}^{-1})$ is a lower bound of the cut locus function . The generalized cylinder coordinates define a diffeomorphism $\gamma_f$ of $M\times[0,{c})$ onto an open subset of $N$, which is a collar. Any lower bound on the cut locus function is also a lower bound on the distance to the first focal point on the inward $M$-geodesics. Since $f$ is a one-sided Alexandrov embedding, the absolute values of the negative principal curvatures are smaller than the positive principal curvatures. Consequently, due to the formula , the distances to the first focal points on the outward $M$-geodesics are not smaller than the distances to the first focal points on the inward $M$-geodesics. Hence the normal variation defines an immersion of $(-{c},{c})\times M$ into ${\mathbb S}^3$. In particular, the induced metric makes $(-{c},{c})\times M$ into a Riemannian manifold with constant sectional curvature equal to one. For all elements of this manifold the cylinder coordinates, i.e. the distances to $M\simeq\{0\}\times M$ and the nearest point in $M$, are uniquely defined. Hence we can glue $(-{c},{c})\times M$ along $\gamma_f([0,{c})\times M)$ to $N$, and obtain a larger 3-manifold $\hat{N}\supset N$ without boundary, such that the generalized cylinder coordinates extend to a diffeomorphism $\hat{\gamma}_f: (-{c},{c})\times
M\rightarrow\hat{N}$ and the immersion $f:N\rightarrow{\mathbb S}^3$ extends to an immersion $\hat{f}:\hat{N}\rightarrow{\mathbb S}^3$.
The balls with radius ${c}$ around all elements of $M\subset\hat{N}$ are isometric to open balls in ${\mathbb S}^3$ of radius ${c}$. For $\epsilon<{c}$ all immersions $\tilde{f}$ obeying the first inequality of can be represented uniquely as the composition of a smooth section $B(p,R) \to T\hat{N}|_{B(p,R)}$ with $\hat{f}\circ\exp_{\hat{N}}$. Here $\exp_{\hat{N}}$ denotes the exponential map of $\hat{N}$. Hence for all immersions $\tilde{f}$ obeying the first inequality of there exists a smooth embedding $i:B(p,R)\hookrightarrow\hat{N}$, such that $\tilde{f}$ is equal to $\hat{f}\circ i$. Due to $\Phi$ induces on $B(p,R)$ a Riemannian metric denoted by $\operatorname{dist}_O$ nearby the original metric. Hence there exists a constant $C_1$ depending only on $\epsilon$ such that $$\operatorname{dist}_O(i(q),i(q'))\leq C_1\operatorname{dist}_M(q,q')
\quad\mbox{ for all }q,q'\in B(p,R).$$ We denote the image of this embedding $i$ by $O$ as an oriented submanifold of $\hat{N}$. We shall identify $\tilde{f}|_{B(p,R)}$ with the immersion $\hat{f}|_O$. For all $q\in O$ let $t\mapsto\gamma_{\tilde{f}}(q,t)$ denote the inward $O$-geodesics in $\hat{N}$. Since $O$ is not complete we have to be careful with the cut locus.
Due to Lemma \[thm:injectivity radius\] and Theorem 1.2 in [@ForLR] there exists $r>0$ depending only on $\kappa{_{\mbox{\rm\scriptsize{max}}}}$ and $H{_{\mbox{\rm\scriptsize{max}}}}$, such that all $q\in M$ belong to open domains in $M$, which are normal graphs over the ball $B(q,r)$ in the unique geodesic sphere ${\mathbb S}^2\subset{\mathbb S}^3$, which touches $M$ at $q$. Furthermore, for sufficiently small $\epsilon$ the same is true for all $q\in\overline{B(p,R-r)}$ and the perturbed immersion $\tilde{f}|_{B(p,R)}$. We conclude that at all $q\in M$ the second fundamental form ${\mathfrak{h}}$ of $f$ obeys with ${D}$ depending only on $\kappa{_{\mbox{\rm\scriptsize{max}}}}$ and $H{_{\mbox{\rm\scriptsize{max}}}}$. Moreover, for sufficiently small $\epsilon$ the principal curvatures of the perturbation $\tilde{f}$ are bounded at all $q\in\overline{B(p,R-r)}$ by a constant depending only on $\kappa{_{\mbox{\rm\scriptsize{max}}}}$ and $H{_{\mbox{\rm\scriptsize{max}}}}$. Finally, again at all $q\in\overline{B(p,R-r)}$ the second fundamental form $\tilde{{\mathfrak{h}}}$ of $\tilde{f}$ obeys with a constant ${D}$ depending only on $\kappa{_{\mbox{\rm\scriptsize{max}}}}$ and $H{_{\mbox{\rm\scriptsize{max}}}}$. Due to Lemma \[thm:ec diameter bound\] there exists a corresponding bound on the chord-arc ratio $C_2>0$ depending only on $\kappa{_{\mbox{\rm\scriptsize{max}}}}$ and $H{_{\mbox{\rm\scriptsize{max}}}}$.
All cut locus functions of one-sided Alexandrov embeddings are uniformly bounded from above by $\frac{\pi}{2}$, since otherwise a sphere with negative principal curvatures touches $M$ inside of $N$ contradicting Hopf’s maximum principle. Now we choose $R=C_1(C_2
2\pi+\epsilon+r)$ and denote $U=B(i(p),C_2 \pi+\epsilon)\subset O$. For all $q\in U\subset O$, we have $$\left\{ q' \in O \mid \exists t \in [0,\,\tfrac{\pi}{2}] \mbox{ with
} \operatorname{dist}_{\hat{N}} (\gamma_{\tilde{f}}(q,\,t),\,q') \leq t \right\}
\subset \left\{ q' \in O \mid \operatorname{dist}_{\hat{N}}(q,\,q') \leq \pi
\right\} \subset B(q,\,C_2\pi)\,.$$ Therefore, for all $q\in U$ the cut locus function $c_{\tilde{f}}$ is well defined. For all such $q\in U$ let $[q]_{\tilde{f}}$ denote the set $$[q]_{\tilde{f}}=\left\{q'\in O\mid
\operatorname{dist}_{\hat{N}}\left(\gamma_{\tilde{f}}(q,c_{\tilde{f}}(q)),\,q' \right)=
c_{\tilde{f}}(q)\right\}.$$ For any closed subset $A\subset O$ the set $\{q\in U\mid [q]_{\tilde{f}}\cap A\not=\emptyset\}$ is a closed subset of $U$. Hence $V=\{q\in U\mid [q]_{\tilde{f}}\subset U\}$ is an open subset of $O$. Furthermore $W=\{\gamma_{\tilde{f}}(q,t)\mid q\in V
\mbox{ and }0\leq t\leq c_{\tilde{f}}(q)\}$ is a submanifold of $\hat{N}$ with boundary $V$. By construction $\hat{f}_W$ is a local one-sided Alexandrov embedding. By choice of $R$, $V$ is contained in the image of $B(p,R)$ under $i$.
\[thm:collar deformation 2\] Let $f:N\rightarrow{\mathbb S}^3$ be a one-sided Alexandrov embedding with constant mean curvature $H\leq H{_{\mbox{\rm\scriptsize{max}}}}$ and principal curvatures bounded by $\kappa{_{\mbox{\rm\scriptsize{max}}}}$. Assume that $\tilde{f}:M\rightarrow{\mathbb S}^3$ is an immersion of the boundary $M=\partial N$ with constant mean curvature $\tilde{H}\geq 0$ obeying for all $q\in M$ instead of all $q\in B(p,R)$ with $\epsilon$ as in Lemma \[thm:collar deformation 1\]. Then $\tilde{f}$ extends to a one-sided Alexandrov embedding from $N$ to ${\mathbb S}^3$.
We apply Lemma \[thm:collar deformation 1\] to all $p\in M$ and obtain a covering of $M$ by open subsets $V$, which are boundaries of local one-sided Alexandrov embeddings. From we deduced in the proof of Lemma \[thm:collar deformation 1\] a bound on both principal curvatures of $f$ and $\tilde{f}$. Due to Lemma \[thm:ec diameter bound\] this implies a uniform bound $C_2$ on the chord-arc ratio. The choice of the radius $R$ in Lemma \[thm:collar deformation 1\] ensures that for any $p\in M$ the constructed local one-sided Alexandrov embedding nearby $p$ is not affected by the immersion $\tilde{f}$ restricted to $M\setminus
B(p,R)$. The corresponding local one-sided Alexandrov embeddings fit together to an Alexandrov embedding $\tilde{f}:\tilde{N}\rightarrow{\mathbb S}^3$ of a submanifold $\tilde{N}\subset\hat{N}$ with boundary $\partial\tilde{N}=M$. It remains to show that $\tilde{N}$ is complete with respect to the Riemannian metric induced by $\tilde{f}$. Due to Rosenberg’s Lemma \[th:Rosenberg\_Lemma\] the cut locus function is uniformly bounded by ${c}$ from below. For all $r<{c}$, the submanifolds $$\{\gamma_f(p,t)\in \Hat{N}\mid p\in M\mbox{ and }
-r\leq t\leq c_f(p)\}\subset \hat{N}$$ are complete with respect to the Riemannian metric induced by $\hat{f}$. By construction, the Riemannian manifold $\tilde{N}$ with the metric induced by $\tilde{f}$ is a closed submanifold of one of these complete submanifolds of $\hat{N}$, and therefore also complete.
Deformation of cylinders in ${\mathbb S}^3$ {#sec:deformation of cylinders}
===========================================
In this section we consider the subspace in the moduli space of periodic finite type solutions of the $\sinh$-Gordon equation that contain the spectral data of cylinders in ${\mathbb S}^3$, and begin with an investigation of the subset of spectral data of one-sided Alexandrov embedded cylinders. We shall determine all continuous deformations of spectral data, which preserve the one-sided Alexandrov embeddedness. In order to do this, we combine the description of finite type cylinders in ${\mathbb S}^3$ in terms of polynomial Killing fields with the investigation of one-sided Alexandrov embeddings. From Definition \[thm:spec\_bobenko\] and \[thm:weights\] we immediately conclude
A covering map $\kappa \mapsto \Delta(\kappa)$ satisfying conditions [[([G]{})]{}]{}-[[([K]{})]{}]{} corresponds to the spectral data of a cylinder if and only if
1. there are two branch points $\kappa_0,\,\kappa_1$ on the real part (i.e the fix point set of the involution $\rho$) at which $\Delta(\kappa_0) = \Delta(\kappa_1) =
\pm 2$.
From the point of view of the covering map the two distinguished points are two real branch points at $\Delta=\pm
2$, which are specified by the sheets they connect. A continuous deformation of a covering map preserves condition [[([L]{})]{}]{}, if the two distinguished real branch points at $\Delta=\pm 2$ are not moved and stay inside the real part. The coordinates of the two distinguished points $\kappa_0,\,\kappa_1$ are given by a parametrization of the cover $\kappa \mapsto
\Delta(\kappa)$, which is unique up to Möbius transformations . As a consequence of conditions [[([G]{})]{}]{}-[[([I]{})]{}]{} the covering map is biholomorphic to ${\mathbb{CP}^1}$, and condition [[([J]{})]{}]{} equips it with an antiholomorphic involution. In general it is difficult to read off the coordinates from the parameters of the moduli space, that is the values of $\Delta$ at the branch points. An interesting parameter is the value of the mean curvature $H=(1+\kappa_0\kappa_1)/(\kappa_0-\kappa_1)$.
\[thm:H continuous\] For continuous deformations of covering maps preserving conditions [[([G]{})]{}]{}-[[([L]{})]{}]{} the mean curvature depends continuously on the deformation parameter.
We describe the compactification of the covering space of $\Delta$ by an open cover $U_0,\,\ldots,\,U_n$. Here $U_1,\,\ldots,\,U_n$ are small disjoint discs around all branch point of $\Delta$ which are to be moved, and $U_0$ is the complement of the union of smaller closed discs $A_j\subset U_j$ such that each $A_j$ is still a neighbourhhod of the corresponding branch point. Let $w_j$ be a holomorphic coordinate on $U_j$, such that $\Delta-\Delta_j$ is a polynomial $P_j(w_j)$ with respect to $w_j$. The coefficients of $P_j$ are the parameters of the continuous deformations. For sufficiently small values of these parameters the covering space is a compact Riemann surface biholomorphic to ${\mathbb{CP}^1}$. We describe such a biholomorphic map $\Phi$ by a meromorphic function of degree one. This is unique up to Möbius transformations.
We need to show that if we eliminate the freedom of Möbius transformations by imposing three additional conditions, that $\Phi$ depends continuously on the deformation parameter. It suffices to show that a sequence of meromorphic functions $\Phi_k$ corresponding to a convergent sequence of parameters converges to the limit $\Phi$ that corresponds to the limit of the parameters. For this it suffices to show that any such sequence of meromorphic functions contains a convergent subsequence, and that the limit is the unique limit that corresponds to the limit of the parameters. Pick on each open set $U_j$ a Möbius transformation $M_j$ whose composition with $\Phi$ maps the branch point of $\Delta$ contained in $U_j$ to the origin, and with derivative at the branch point is equal to one, and such that each one of the two points $\kappa=\pm i$ is mapped to $\infty$. Since the space of schlicht functions is compact (see Proposition 7.15 in Section 14.7 of Conway [@Con2]), a subsequence of $M_j \circ \Phi_k$ converges to a biholomorphic map $U_j \to {\mathbb C}$. Since the parameters converge, the limit is equal to $M_j \circ \Phi$.
The special situation $H=\infty$ corresponding to cylinders in ${\mathbb R}^3$ has a simple description. In the case of cylinders in ${\mathbb R}^3$, the above correspondence via spectral curves also holds with one notable exception. The closing condition for cylinders in ${\mathbb R}^3$ require one value $\kappa_0 \in {\mathbb R}$ at which the monodromy is equal to $\pm \mathbbm{1}$ and its derivative vanishes there [@DorH:per]. Assume we have the spectral data for a cylinder in ${\mathbb S}^3$ in which $\kappa_0$ and $\kappa_1$ coalesce. Then the mean curvature is infinite and the cylinder is shrunk to a point. A blow-up then yields a cylinder in ${\mathbb R}^3$, see Umehara and Yamada [@UmeY:tori].
A covering map satisfying conditions [[([G]{})]{}]{}-[[([K]{})]{}]{} corresponds to a cylinder in ${\mathbb R}^3$ if and only if there is a real branch point of order at least 3.
Let us first show that all one-sided Alexandrov embedded cylinders in ${\mathbb S}^3$ of finite type obey the assumptions of Rosenberg’s Lemma \[th:Rosenberg\_Lemma\] and the chord-arc bound in Lemma \[thm:ec diameter bound\].
\[thm:killing bound\] For every compact subset $\mathcal{K}\subset{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ and $H{_{\mbox{\rm\scriptsize{max}}}}>0$ there exist constants $\kappa{_{\mbox{\rm\scriptsize{max}}}}>0$ and ${D}>0$ with the following property: If $f:\mathbb{R}^2\rightarrow{\mathbb S}^3$ is an immersion of finite type with constant mean curvature $|H|\leq H{_{\mbox{\rm\scriptsize{max}}}}$, whose polynomial Killing field $\zeta$ takes at some point $p\in\mathbb{R}^2$ a value $\zeta(p)\in \mathcal{K}$. Then the absolute values of both principal curvatures of $f$ are uniformly bounded by $\kappa{_{\mbox{\rm\scriptsize{max}}}}$, and the second fundamental form ${\mathfrak{h}}$ of $f$ satisfies .
The coefficients of $a(\lambda)=-\lambda\det(\xi(\lambda))$ depend continuously on $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$. Hence on every compact subset $\mathcal{K}\subset{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ these coefficients are uniformly bounded. In particular, the union $\hat{\mathcal{K}}$ of all isospectral sets $\mathcal{K}_a$ of Lemma \[thm:compact\] not disjoint from $\mathcal{K}$ is compact too. If $f:\mathbb{R}^2\rightarrow{\mathbb S}^3$ is a immersion of finite type, whose polynomial Killing field $\zeta$ takes at some $p\in\mathbb{R}^2$ a value in $\mathcal{K}$, then the other values belong to the corresponding isospectral set. Since all derivatives of $f$ at some $p\in\mathbb{R}^2$ depend continuously on $\zeta(p)$, the lemma follows.
\[thm:immersion bound\] For all compact subsets $\mathcal{K}\subset{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$, $\epsilon>0$, $R>0$ and $H{_{\mbox{\rm\scriptsize{max}}}}>0$ there exists $\delta>0$ with the following property: If $(\zeta,\,\lambda_0,\,\lambda_1)$ and $(\tilde{\zeta},\,\tilde{\lambda}_0,\,\tilde{\lambda}_1)$ are polynomial Killing fields and marked points of two cylinders $f,\,\tilde{f}:M\rightarrow{\mathbb S}^3$ obeying at some point $p\in M$ $$\begin{aligned}
\label{eq:killing bound 1}
\zeta(p)\in\mathcal{K}\mbox{ and }
|H|=\left|\frac{\lambda_0+\lambda_1}
{\lambda_0-\lambda_1}\right| &\leq H{_{\mbox{\rm\scriptsize{max}}}}
\quad\mbox{ or }&
\tilde{\zeta}(p)\in\mathcal{K}\mbox{ and }
|\tilde{H}|=\left|\frac{\tilde{\lambda}_0+\tilde{\lambda}_1}
{\tilde{\lambda}_0-\tilde{\lambda}_1}\right| &\leq H{_{\mbox{\rm\scriptsize{max}}}}\\
\label{eq:killing bound 2}
\|\zeta(p)-\tilde{\zeta}(p)\|&<\delta\,,&
|\lambda_0-\tilde{\lambda}_0|<\delta\mbox{ and }
|\lambda_1-\tilde{\lambda}_1|&<\delta.\end{aligned}$$ Then the corresponding immersions $f$ and $\tilde{f}$ satisfy on $B(p,\,R)\subset M$.
Since ${\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ is an open subset of ${\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})}$, there exists a $\delta>0$ and a compact subset of ${\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$, containing all balls $B(\xi,\delta)$ with $\xi\in\mathcal{K}$. Furthermore, due to Lemma \[thm:killing bound\] there exists also a compact subset $\hat{\mathcal{K}}\subset{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ containing all isospectral sets $\mathcal{K}_a$ not disjoint from $\cup_{\xi\in\mathcal{K}}B(\xi,\delta)$. Due to Lemma \[thm:killing bound\] for given $\mathcal{K}\subset{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ and $H{_{\mbox{\rm\scriptsize{max}}}}>0$ there exists $\delta>0$, $\kappa{_{\mbox{\rm\scriptsize{max}}}}>0$ and ${D}>0$, such that the conditions and imply that the corresponding immersions $f$ and $\tilde{f}$ have principal curvatures bounded by $\kappa{_{\mbox{\rm\scriptsize{max}}}}$ and second fundamental forms with covariant derivatives bounded by . In particular there exists $C>0$ such that the Riemannian metrics induced by $f$ and $\tilde{f}$ on $M$ are uniformly bounded by $C$ times the Euclidean metric on the conformal parametrization $\mathbb{R}^2$. The abelian group $\mathbb{R}^2$ of spatial translations acts continuously on all elements of $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$. Therefore all elements of the closed ball $B(0,CR)\subset\mathbb{R}^2$ act uniformly continuous on $\hat{K}$. Hence for all $\tilde{\epsilon}>0$ there exists $\delta>0$, such that the inequalities imply $\|\zeta(q)-\tilde{\zeta}(q)\|<\tilde{\epsilon}$ for all $q\in
B(p,R)$. Since all derivatives of $f(q)$ and $\tilde{f}(q)$ depend continuously on $\zeta(q)$ and $\tilde{\zeta}(q)$, respectively, the lemma follows.
\[thm:convergency\] For all compact subsets $\mathcal{K}\subset{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ and $H{_{\mbox{\rm\scriptsize{max}}}}>0$ there exists $\delta>0$ with the following property: Let $\tilde{f}:M\rightarrow{\mathbb S}^3$ be a cylinder with non-negative mean curvature $\tilde{H}$ (with respect to normal uniquely determined by the orientation of $M$) and with polynomial Killing field $\tilde{\zeta}$ and marked points $\tilde{\lambda}_0$ and $\tilde{\lambda}_1$. If for all $p\in M$ there exists a one-sided Alexandrov embedding $f:N\rightarrow{\mathbb S}^3$ with polynomial Killing field $\zeta$ and marked points $\lambda_0$ and $\lambda_1$ obeying conditions and , then $\tilde{f}$ extends to a one-sided Alexandrov embedding $\tilde{f}:\tilde{N}\rightarrow{\mathbb S}^3$.
As in the proof of Lemma \[thm:immersion bound\] there exists $\delta>0$ and a compact subset $\hat{\mathcal{K}}\subset{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ containing all isospectral sets not disjoint from $\cup_{\xi\in\mathcal{K}}B(\xi,\delta)$. Due to Lemma \[thm:killing bound\] there exists a bound $\kappa{_{\mbox{\rm\scriptsize{max}}}}$ on the absolute values of the corresponding principal curvatures. From Lemma \[thm:collar deformation 1\] we get a $R>0$ and an $\epsilon>0$ and from Lemma \[thm:immersion bound\] a $\delta>0$, such that the immersions $f$ and $\tilde{f}$ obey on $B(p,R)$ if the corresponding polynomial Killing fields $\zeta$ and $\tilde{\zeta}$ and marked points $(\lambda_0,\,\lambda_1)$ and $(\tilde{\lambda}_0,\,\tilde{\lambda}_1)$ obey and . Therefore these conditions imply that all $p\in M$ have a neighbourhood, on which $\tilde{f}$ extends to a local one-sided Alexandrov embedding. We remark that due to the uniform bound on the chord-arc ratio from Lemma \[thm:ec diameter bound\] and the choice of $R>0$ in Lemma \[thm:collar deformation 1\] these local one-sided Alexandrov embeddings are not affected by the restriction of $\tilde{f}$ to $M\setminus B(p,R)$. Due to Corollary \[thm:unique embedded 2\] these local one-sided Alexandrov embeddings can be glued together to an immersion $\tilde{f}:\tilde{N}\rightarrow{\mathbb S}^3$ with boundary $\partial\tilde{N}=M$. A subsequence of any Cauchy sequence in $\tilde{N}$ with Riemannian metric induced by $\tilde{f}$ is contained in one of the local Alexandrov embeddings from $O$ to ${\mathbb S}^3$. Hence $\tilde{N}$ with the Riemannian metric induced by $\tilde{f}$ is complete. Since $\tilde{f}$ has non-negative mean curvature, it is a one-sided Alexandrov embedding.
For $g\in\mathbb{N}_0$ let ${\mathcal{C}_g}$ denote the subset of triples $(\xi,\lambda_0,\lambda_1)\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}\times{\mathbb S}^1\times{\mathbb S}^1$ containing the initial values (not necessarily without roots) together with both marked points of finite type cylinders $f:M\rightarrow{\mathbb S}^3$. Let ${\mathcal{C}_g}^+$ denote the subset of triples $(\xi,\lambda_0,\lambda_1)\in{\mathcal{C}_g}$ corresponding to finite type cylinders $f:M\rightarrow{\mathbb S}^3$ with non-negative mean curvature (with respect to the normal determined by the orientation of $M$).
\[thm:spec\_AE\] Let $\xi\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ be the initial value and $(\lambda_0,\lambda_1)$ the marked points of a one-sided Alexandrov embedded cylinder $f:N\rightarrow{\mathbb S}^3$. Then a cylinder of finite type $\tilde{f}:M\rightarrow{\mathbb S}^3$ with non-negative mean curvature and with initial value $\tilde{\xi}\in{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ and marked points $(\tilde{\lambda}_0,\tilde{\lambda}_1)$ extends to one-sided Alexandrov embedding, if one of the following conditions are satisfied:
1. The polynomial $a(\lambda)=-\lambda\det(\xi)$ has $2g$ pairwise distinct roots and $\tilde{\xi}$ belongs to the same isospectral set $\mathcal{K}_a\subset{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ as $\xi$ with the same marked points.
2. The initial value $\xi$ has only unimodular roots and $\tilde{\xi}$ belongs to the same isospectral set $\mathcal{K}_a\subset{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ as $\xi$ with the same marked points.
3. The polynomial $a(\lambda)=-\lambda\det(\xi)$ has $2g$ pairwise distinct roots and $(\tilde{\xi},\tilde{\lambda}_0,\tilde{\lambda}_1)$ belongs to the same connected component of ${\mathcal{C}_g}^+$ as $(\xi,\lambda_0,\lambda_1)$.
4. The initial value $\xi$ has only unimodular roots and $(\tilde{\xi},\tilde{\lambda}_0,\tilde{\lambda}_1)$ belongs to the same connected component of ${\mathcal{C}_g}^+$ as $(\xi,\lambda_0,\lambda_1)$.
In particular, real continuous deformations of spectral curves with non-negative mean curvature and decreasing $G$ preserve one-sided Alexandrov embeddedness.
The initial values $\xi$ and $\tilde{\xi}$ extend to unique polynomial Killing fields $\zeta$ and $\tilde{\zeta}$, respectively.
We first show [[([M]{})]{}]{}. We obtain the map $f:M \to {\mathbb S}^3$ by composing an affine immersion of $M \cong {\mathbb S}^1 \times {\mathbb R}$ into the real part of the Jacobian with the Sym-Bobenko formula. In case that $a$ has pairwise distinct roots, the real part of the Jacobian is a real $g$ dimensional compact torus, and acts freely and transitively on the corresponding isospectral set $\mathcal{K}_a$. Due to Lemma \[thm:compact\] the isospectral sets are compact. A continuous action of a finite dimensional Lie group on a compact space is uniformly continuous. Hence for all $\tilde{f}$ in a small neighbourhood of $f$ in the isospectral set is satisfied on $q\in M$ instead of $q\in B(p,R)$. Due to Lemma \[thm:killing bound\] and Rosenberg’s Lemma \[th:Rosenberg\_Lemma\] the cut locus function $c_f$ is uniformly bounded from below. Corollary \[thm:collar deformation 2\] implies that the set of one-sided Alexandrov embedded cylinders in the isospectral set is open and closed. This proves [[([M]{})]{}]{}.
Now we prove [[([N]{})]{}]{}. If the polynomial $a$ of the initial value $\xi$ has roots of higher order, then the corresponding isospectral set $\mathcal{K}_a\subset{\Lambda_{-1}^g{\mathfrak{sl}_{\mbox{\tiny{$2$}}}}({\mathbb C})^{\times}}$ has a stratification into strata, on which the real part of the Jacobian acts transitively. Due to Lemma \[thm:dense stratum\] the initial value $\xi$ belongs to the stratum of highest dimension, which is dense in the whole isospectral set. In this case the arguments above concerning the case [[([M]{})]{}]{} carry over and show that whenever a stratum contains a one-sided Alexandrov embedded cylinder, then all elements of this stratum correspond to one-sided Alexandrov embedded cylinders. Lemma \[thm:convergency\] now implies [[([N]{})]{}]{}.
Now we prove part [[([O]{})]{}]{}. Again we show that the set of all spectral curves of arithmetic genus $g$ (i.e. the set of polynomials $a(\lambda)=-\lambda\det(\xi)$ of degree $2g$), whose isospectral sets contain one-sided Alexandrov embedded cylinders, is open and closed. Since the isospectral sets are compact, the Hausdorff distance between the isospectral sets defines a metric on the space of spectral data of given arithmetic genus. Lemma \[thm:convergency\] implies that the isospectral sets of one-sided Alexandrov embeddings are closed and open in ${\mathcal{C}_g}^+$. This proves [[([O]{})]{}]{}.
Due to [[([N]{})]{}]{} the arguments of the proof of [[([O]{})]{}]{} carry over to [[([P]{})]{}]{}, and concludes the proof.
One-sided Alexandrov embedded cylinders in ${\mathbb S}^3$ of finite type {#sec:AE}
=========================================================================
In this concluding chapter we classify one-sided Alexandrov embedded cylinders of finite type in the 3-sphere. We first show in Theorem \[thm:onesided revolution\] that there is a 1-parameter family of flat one-sided Alexandrov embedded cylinders, and a 2-parameter family of one-sided Alexandrov embedded rotational cylinders of spectral genus $g=1$. We then prove (Lemma \[thm:continuous genus 1\]) that an arbitrary one-sided Alexandrov embedded cylinder of finite type in the 3-sphere can be continuously deformed into a flat one-sided Alexandrov embedded cylinder, while preserving the condition of one-sided Alexandrov embeddedness throughout the whole deformation. We next prove (Lemma \[thm:not genus two\]) that a surface with bubbletons is not one-sided Alexandrov embedded. Putting the above results together in Theorem \[thm:main1\] gives us the following classification: A one-sided Alexandrov embedded cylinder of finite type in the 3-sphere is a surface of revolution.
Since an embedded torus in the 3-sphere is covered by a one-sided Alexandrov embedded cylinder of finite type, this result confirms the conjecture by Pinkall and Sterling [@PinS] that the only embedded tori in the 3-sphere are tori of revolution. In particular, since by a result by Hsiang and Lawson [@HsiL], the only embedded minimal torus of revolution is the Clifford torus, this affirms the Lawson Conjecture. We conclude the paper with a generalization of an ’unknottedness’ result by Lawson [@Law:unknot], and show that for a one-sided Alexandrov embedded cylinder of finite type in the 3-sphere, the 3-manifold is diffeomorphic to the cartesian product $\overline{\mathbb{D}}\times{\mathbb R}$, where $\overline{\mathbb{D}}$ is the closed unit disk.
We first turn our attention to the spectral data of one-sided Alexandrov embedded rotational cylinders of spectral genus $g \leq 1$.
\[thm:onesided revolution\] There exists a family of spectral data of one-sided Alexandrov embedded cylinders with spectral genus $g\leq 1$ parameterized by the mean curvature $H \geq 0$ and $\alpha\in[0,1)$. The corresponding spectral curves are given by $a(\kappa) = \kappa^2 + \alpha$. The boundary of the moduli $(H,\,\alpha) \in [0,\,\infty) \times [0,\,1)$ consists of
flat cylinders in ${\mathbb S}^3$
: $H \in [0,\,\infty ],\,\alpha = 0$,
minimal cylinders in ${\mathbb S}^3$
: $H=0,\,\alpha \in [0,\,1)$,
In case of flat cylinders in ${\mathbb S}^3$ the spectral genus zero curves have a real double point at $\kappa =0$ and no other real double points. This case contains all spectral data of one-sided Alexandrov embedded cylinders of geometric genus zero. In case $\alpha \neq 0$ the spectral curves have no real double points. As $\alpha \to 0$ the two branch points coalesce into a double point at $\kappa =0$.
We recall the spectral data for cylinders of spectral genus zero, and the expressions for $\ln \mu$ of and $d \ln
\mu$ of . At $\kappa = \pm \kappa_0$ we require the closing conditions that $\ln \mu \in \pi i {\mathbb Z}$. Thus for some integers $m,\,n \in {\mathbb Z}$ we have $$\begin{split} \label{eq:b_eq}
\left. \ln \mu \right|_{\kappa_0} \in \pi i {\mathbb Z}&\Longleftrightarrow
4(b_0\kappa_0 - b_1)^2 = n^2(\kappa_0^2 +1)\,, \\
\left. \ln \mu \right|_{-\kappa_0} \in \pi i {\mathbb Z}&\Longleftrightarrow
4(b_0\kappa_0 + b_1)^2 = m^2(\kappa_0^2 +1)\,.
\end{split}$$ We make the following claim: If a cylinder is one-sided Alexandrov embedded then $m = n =\pm 1$ (for this ensures that the surface is simply wrapped with respect to the rotational period). To prove this claim first note that any spectral genus zero cylinder is a covering of a flat embedded torus. The complement of this flat embedded torus with respect to ${\mathbb S}^3$ consists of two connected components $\mathcal{D}_{\pm}$, both diffeomorphic to $\mathbb{D} \times
{\mathbb S}^1$. Assume the mean curvature vector points into $\mathcal{D}_+$. For a one-sided Alexandrov embedded cylinder the extension $f:N \to {\mathbb S}^3$ is a surjective immersion onto the closure $\bar{\mathcal{D}}_+$ of $\mathcal{D}_+$. Hence this map is a covering map. The fundamental group of $\bar{\mathbb{D}}\times{\mathbb S}^1$ is isomorphic to ${\mathbb Z}$. Now the covers of a topological space are in one-to-one correspondence with subgroups of the fundamental group [@St §14], and all non-trivial subgroups of $\pi_1(\bar{\mathbb{D}}\times{\mathbb S}^1) \cong {\mathbb Z}$ correspond to compact covers. Hence the only non-compact cover of $\bar{\mathbb{D}}\times{\mathbb S}^1$ is the universal cover $\bar{\mathbb{D}}\times{\mathbb R}$. Therefore $f$ is the universal covering map. In particular the period of the cylinder is a primitive period in the kernel of $$H_1({\mathbb S}^1 \times {\mathbb S}^1,\,{\mathbb Z}) \to
H_1(\bar{\mathbb{D}}\times{\mathbb S}^1,\,{\mathbb Z})\,.$$ This implies $m=n=\pm 1$ and proves the claim.
Returning to the proof of the theorem, then with $m^2=n^2=1$ simplify to $b_0b_1\kappa_0 = 0$. Thus if $\kappa_0 \neq
0$ then $b_0b_1=0$. Now $|H|$ attains all values in $[0,\,\infty)$ when either $\kappa_0$ ranges over values $\kappa_0 \in (0,\,1]$, or when $\kappa_0$ ranges over values $\kappa_0 \in [1,\,\infty)$. Hence we can pick a $\kappa_0$ either from $(0,\,1]$ or from $[1,\,\infty)$. The Möbius transformation $\kappa\mapsto-\frac{1}{\kappa}$ interchanges these two intervals and the two cases $b_0=0$ and $b_1=0$. Hence we may assume without loss of generality that $b_1=0$. Then $4\kappa_0^2 b_0^2 = \kappa_0^2+1$.
Now we claim that with this choice only $\kappa_0\in[1,\,\infty)$ correspond to one-sided Alexandrov embedded cylinders. In fact, we have seen above that the period of the cylinder should correspond to an element in the kernel of $H_1(\partial\mathcal{D}_+,\mathbb{Z})\rightarrow
H_1(\bar{\mathcal{D}}_+,\mathbb{Z})$. The Möbius transformation $\kappa\mapsto-\frac{1}{\kappa}$ interchanges the cases $\kappa_0\in(0,\,1]$ and $\kappa_0\in[1,\,\infty)$. Hence we may consider both families of flat cylinders as two copies of one family of tori considered as cylinders with respect to different periods. In the limit $\kappa_0\rightarrow 0$ the length of the period tends to infinity, and in the limit $\kappa_0\rightarrow\infty$ the length of the period is bounded. Hence the period of the first family are the rotation period of $\bar{\mathcal{D}}_-$, i.e. a primitive element of the kernel of $H_1(\partial\mathcal{D}_-,\mathbb{Z})\rightarrow
H_1(\bar{\mathcal{D}}_-,\mathbb{Z})$ and therefore the translation period of $\bar{\mathcal{D}}_+$. The period of the second family are the rotation period of $\bar{\mathcal{D}}_+$. Therefore only the family $\kappa_0\in[1,\infty)$ corresponds to one-sided Alexandrov embedded cylinders.
For a double point $\kappa$ this means that $(\kappa_0^2 +1)\,\kappa^2 = n^2(\kappa^2
+1)\,\kappa_0^2$ for some $n \in {\mathbb Z}$. We have $$\begin{aligned}
0\leq\frac{\kappa^2}{\kappa^2+1}&\leq 1\quad
\mbox{ for real }\kappa\mbox{ and }&
\frac{1}{2}&\leq\frac{\kappa_0^2}{1+\kappa_0^2}<1
\quad\mbox{ for }\kappa_0 \in[1,\,\infty).\end{aligned}$$ Hence for $\kappa_0\in[1,\,\infty)$ this family has only the real double points $\kappa=\pm\kappa_0$ with $n=\pm 1$ and $\kappa=0$ with $n=0$. Due to the closing condition (ii) in Definition \[thm:spec\_bobenko\] the former has to be preserved. If we open the latter double points then $\ln \mu$ remains a meromorphic function on the genus 1 spectral curve. We thus altogether obtain families parameterized by $\kappa_0\in[1,\,\infty)$ and $\alpha\in[0,1)$: $$\begin{aligned}
&a(\kappa) = \kappa^2 + \alpha\,,
\qquad \qquad \qquad \nu^2=(\kappa^2+1)(\kappa^2 +\alpha)\,,&\\
&\ln\mu = 2\pi i \,b_2\,\frac{\kappa^2 +\alpha}{\nu}\,,
\qquad d\ln\mu = 2\pi i \,b_2\,\frac{\kappa^2 +1-\alpha}{(\kappa^2 +1)\nu}\,.\end{aligned}$$ At $\pm \kappa_0$ the closing conditions $\ln\mu = \pm \pi i$ must hold and thus $4(\kappa_0^2 +\alpha)^2b_2^2 = (\kappa_0^2 +\alpha)(\kappa_0^2 +1)$ giving $4(\kappa_0^2 +\alpha)\,b_2^2 = \kappa_0^2+1$. Therefore double points have to satisfy $$\frac{\kappa^2+\alpha}{\kappa^2 +1} =
n^2\frac{\kappa_0^2+\alpha}{\kappa_0^2 +1}\,\mbox{ for some }n\in{\mathbb Z}\,.$$ But for $\kappa_0 \in [1,\,\infty)$ and $\alpha\in(0,1)$ this equation has no real solutions besides $n=\pm 1$ and $\kappa=\pm\kappa_0$. Hence these families have no real double points.
In addition to the two boundary components there exists two limiting cases: When $H=\infty,\,\alpha \in [0,\,1]$ we obtain unduloidal cylinders in ${\mathbb R}^3$; When $H\in[0,\infty),\, \alpha=1$, the resulting surfaces are [*chain of spheres*]{}. Note that as a consequence of Theorem \[thm:onesided revolution\], when $\alpha
\neq 0$ there is no real branch point at $\Delta = \pm 2$. Hence it is only possible to increase the genus by opening two conjugate double points in this case.
\[thm:2 real zeroes\] Every spectral data of a one-sided Alexandrov embedded cylinder in ${\mathbb S}^3$ can be deformed into spectral data of a one-sided Alexandrov embedded cylinder in ${\mathbb S}^3$ with arbitrary large mean curvature $H$ by a continuous deformation that increases $H$ and decreases $G$ .
The real part is the fixed point set of the anti-linear involution $\rho$. It is a compact one-dimensional submanifold of the domain of $\Delta$, which is isomorphic to ${\mathbb S}^1$. Since the values of $\Delta$ at both $\kappa_0,\,\kappa_1$ are equal to $\pm 2$ both segments of the real part between $\kappa_0$ and $\kappa_1$ contain an odd number of branch points of $\Delta$ (counted with multiplicities). The simultaneous movements of $\kappa_0$ and $\kappa_1$ directed inward to one of these two segments, increases the mean curvature with respect to the inward normal (determined by the orientation of $M$). This segment we call the [**[short segment]{}**]{}, and the other segment the [**[long segment]{}**]{}. We use a Möbius transformation to ensure that $\kappa = \infty$ is contained in the long segment.
We claim, that the movement of a simple real branch point of $\Delta$ within the short segment, which increases the value of $\Delta$ at the simple branch point, if it is a local minimum, and decreases the value of $\Delta$ otherwise, increases the mean curvature. In fact, due to the signs of $\dot{\kappa}_i$ are equal to the signs of $-c(\kappa_i)/b(\kappa_i)$. We may assume that the simple branch point is a zero of $b$. Hence $c$ has the same zeroes as $b$ with the exception of the simple branch point, whose value of $\Delta$ is changed. Hence the sign of the function $\kappa\mapsto-c(\kappa)/b(\kappa)$ on the short segment changes the sign only at the simple branch point. The two points $\kappa_0$ and $\kappa_1$ sit on different sides of this simple branch point. Hence the sign of $\dot{H}$ does not depend on the position of $\kappa_0$ and $\kappa_1$, as long they stay on the corresponding side of the simple branch point. If $\kappa_0$ and $\kappa_1$ sit nearby the simple branch point, then due to and the movement, which increases the value of $\Delta$ at local minima and decreases the values of $\Delta$ at local maxima, moves $\kappa_0$ and $\kappa_1$ towards each other. This proves the claim.
Due to condition [[([F]{})]{}]{} the real branch point at $\Delta=\pm 2$ have odd orders. As a first step we shall move each such real branch point of odd order in the short segment at $\Delta=\pm 2$ by a small movement into $\Delta\in(-2,2)$. If we choose $a$ to have at all real branch points at $\Delta=\pm 2$ roots of the same order as $\Delta^2-4$, then $b$ has at all real branch points roots of the same order as $\Delta'$. Moreover, the corresponding $\frac{c}{b}$ changes the sign at the real branch point, and the arguments concerning simple real branch points in the short segment carry over. This deformation shortens the short segment, and therefore increases the mean curvature. These movements do not change $G$ and increases the geometric genus into $G$. All other deformations shall decrease the geometric genus.
In a second step we increase the values of $\Delta$ of all simple real branch points within the short segment, which are local minima, and decrease the values of $\Delta$ of all simple real branch points within the short segment, which are local maxima. This again shortens the short segment and increases the mean curvature. All higher order real branch points within the short segment we deform into pairs of complex conjugate branch points close to the real part and eventually one simple real branch point. Similar as in the third step of the proof of Lemma \[thm:connected\] we can achieve with increasing mean curvature a situation with only one simple real branch point within the short segment. If we move this real branch point arbitrarily close to the two marked points $\kappa_0$ and $\kappa_1$, then the mean curvature becomes arbitrarily large with decreasing $G$ .
The arguments show that we can deform the spectral data of a one-sided Alexandrov embedded cylinder in ${\mathbb S}^3$ into the spectral data of a cylinder in ${\mathbb R}^3$. If the two marked points $\kappa_0$ and $\kappa_1$ coalesce, then the corresponding surface in ${\mathbb S}^3$ shrinks to a point. But if we enlarge simultaneously all distances of ${\mathbb S}^3$ on an appropriate scale, then the corresponding cylinders converge to a cylinder in ${\mathbb R}^3$. In fact, the enlargement of the distances of ${\mathbb S}^3$ corresponds to spheres in ${\mathbb R}^4$ of enlarged radius and therefore also to three-dimensional space forms of diminishing constant sectional curvature. Our arguments can be used to show, that all spectral data of one-sided Alexandrov embedded cylinders in ${\mathbb S}^3$ can be continuously deformed within this class into spectral data of one-sided Alexandrov embedded cylinders in ${\mathbb R}^3$. In [@KorKS] it is shown that they have genus at most equal to one. Hence we could have used [@KorKS] instead of the subsequent Lemmas to prove the main Theorem \[thm:main1\]. Conversely, our arguments can be used to show that all one-sided Alexandrov embedded cylinders in ${\mathbb R}^3$ of finite type are Delaunay surfaces. We expect that our arguments extend to all one-sided Alexandrov embedded cylinders in ${\mathbb R}^3$ with constant Hopf differential (compare [@MaOs; @Sch; @Ta]).
\[thm:continuous genus 1\] All spectral data of one-sided Alexandrov embedded cylinders in ${\mathbb S}^3$ can be continuously deformed within the class of one sided Alexandrov embedded cylinders into spectral data of flat cylinders in ${\mathbb S}^3$ described in Theorem \[thm:onesided revolution\].
In a first step, due to Lemma \[thm:2 real zeroes\], we can deform the spectral data of one-sided Alexandrov embedded cylinders in ${\mathbb S}^3$ continuously with decreasing $G$ and increasing $H$ into spectral data with only one simple real branch point in the short segment and arbitrary large mean curvature.
In a second step we apply a similar procedure on the real branch points in the long segment. We can control the mean curvature by moving the only real branch point within the short segment closer to the two marked points $\kappa_0$ and $\kappa_1$. Hence we obtain spectral data with only two real branch points arbitrary close to the two marked points $\kappa_0$ and $\kappa_1$ with arbitrary large mean curvature.
In a third step we apply the deformation described in the fourth step of the proof of Lemma \[thm:connected\]. We obtain spectral data of genus at most equal to one with two simple real branch points very close to the two branch points at $\kappa_0$ and $\kappa_1$.
In a fourth step we move the real branch point, which connects the sheets with lower labels, away from the two real branch points at $\kappa_0$ and $\kappa_1$ along $\Delta\in[-2,2]$ to and fro until it reaches the place $\Delta=2$ on the sheets with labels $1^\pm$. We claim that we can preserve positive mean curvature by moving, if necessary, the other real branch point in the other direction along $\Delta\in[-2,2]$ to and fro. The function $\Delta$ is equal to $\mu+\mu^{-1}=2\cosh(\ln\mu)$. Hence it suffices to show that we preserve the values of $\kappa$ at both marked points by changing the values of $\ln\mu$ at the other real branch point. The polynomial $c$, which corresponds to the deformation preserving the mean curvature is proportional to $c(\kappa)=(\kappa-\kappa_0)(\kappa-\kappa_1)$. The derivatives $\dot{\ln\mu}=\dot{\mu}\mu^{-1}$ at the real branch points, that is the zeroes of the corresponding polynomial $b$, are equal to the values of $\tfrac{c(\kappa)}{\nu}$ at the real branch points due to . We remark that since $d\ln\mu$ vanishes at these real branch points any change of the values of $\kappa$ at these real branch points has no influence on the derivatives $\dot{\ln\mu}$ of the values of $\ln\mu$ at these real branch points. As long as the real branch point has not reached $\Delta=2$ on the sheets with labels $1^\pm$, the function $\nu$ has no roots on the real part. The two integrals of $d\ln\mu$ along the short and long segments are preserved under the deformation. Both real roots of $b$ are local extrema of $\tfrac{1}{2\pi i}\ln\mu$ on the real part. During the deformation the value of $\tfrac{1}{2\pi i}\ln\mu$ is increased in case of a local maximum, and decreased in case of a local minimum. Hence during the deformation one of the real roots of $b$ stays in the short segment, while the other real root stays in the long segment. Therefore we can move, without changing the mean curvature, the real branch point arbitrarily close to $\Delta=2$ on the sheets with labels $1^\pm$. Due to Lemma \[thm:H continuous\] there exist for any $\epsilon>0$ a $\delta >0$ such that the movement from $\Delta=2-\delta$ to $\Delta =2$ changes the mean curvature by at most $\epsilon$. This proves the claim.
Due to Theorem \[thm:spec\_AE\] the whole deformation preserves the one-sided Alexandrov embeddedness. Hence the final spectral data belong to the flat cylinders in ${\mathbb S}^3$ described in Theorem \[thm:onesided revolution\].
\[thm:not genus two\] The cylinders with bubbletons of finite type, whose $\Delta$ corresponds to one of the flat cylinders described in Theorem \[thm:onesided revolution\] are not one-sided Alexandrov embedded.
Due to Theorem \[thm:spec\_AE\] [[([O]{})]{}]{} it is enough to show that the cylinders, which are obtained from the continuous deformation of a $\Delta$ corresponding to a flat cylinder described in Theorem \[thm:onesided revolution\] by moving two simple complex conjugate branch points from $\Delta=\pm 2$ into $\Delta\in{\mathbb C}\setminus[-2,2]$ and increasing the genus by two, are not one-sided Alexandrov embedded. Afterwards these two places at $\Delta=\pm 2$ have no branch points. In the following first two steps we can move the unique real branch point in the short segment arbitrarily close to the two marked points and make the mean curvature arbitrarily large.
In a first step we move these branch points to the long segment of the real part between the two marked points $\kappa_0$ and $\kappa_1$, which contains the real branch point at $\Delta=2$. Afterwards this segment contains two additional real branch points.
In a second step we move these real branch points on the real part away from each other. Then between them two new sheets become connected through a new segment of the real part. Hence we can move one of these two real branch points several times along $\Delta\in[-2,2]$ to and fro until it reaches one of the two places at $\Delta=\pm 2$ without branch points. Hence the genus is reduced by one.
In a third step we move the only real branch point in the short segment away from the two marked points along $\Delta\in[-2,\,2]$ to and fro until it reaches the other place $\Delta=\pm 2$ without other branch points. In order to preserve positivity of the mean curvature, we move the other real root of $b$ in the long segment away from the two marked points along $\Delta \in [-2,\,2]$ to and fro. The arguments of the fourth step in the proof of Lemma \[thm:continuous genus 1\] ensure this is always possible. The whole deformation has twice deformed one zero of $b$, and two zeroes of $a$ into a real double point, and thus arrive at a flat cylinder of spectral genus zero. It cannot belong to the family of flat cylinders described in Theorem \[thm:onesided revolution\], since the corresponding family of spectral genus one curves described in this Theorem do not have real double points. Due to Theorem \[thm:spec\_AE\] [[([O]{})]{}]{}-[[([P]{})]{}]{} these deformations preserve one-sided Alexandrov embeddedness. Hence all spectral data obtained by adding to the flat cylinders in ${\mathbb S}^3$ described in Theorem \[thm:onesided revolution\] a bubbleton do not correspond to one-sided Alexandrov embedded cylinders in ${\mathbb S}^3$.
As an example of the technique used in the proof of Lemma \[thm:not genus two\] we demonstrate how to deform the simplest case of bubbletons at the branch point over $\Delta =2$ connecting the sheets with labels $2^\pm$ and $3^\pm$: We describe the movement of the branch point connecting the sheets which have labels with superscript $+$. The other branch point is moved along the conjugate path. First we move the branch point connecting the sheets with labels $2^+$ and $3^+$ counter clockwise around the interval $\Delta \in [-2,\,2]$ until it meets the origin $\Delta =0$ from the lower half-plane. Our branch point now meets the conjugate branch point at the real part and connects sheets with labels $1^+$ and $4^+$. Further, the branch point over $\Delta =-2$ which initially connected the sheets with labels $1^+$ and $2^+$, now connects the sheets with labels $4^+$ and $2^+$, and the branch point over $\Delta =-2$ which initially connected the sheets with labels $3^+$ and $4^+$, now connects the sheets with labels $3^+$ and $1^+$. Since the two coalescing branch points are now real, we can separate them along the real part. We move one of them connecting sheets with labels $1^+$ and $1^-$ to the left until $\Delta=-2$, and then move it further connecting the sheets with labels $3^+$ and $3^-$ to the right until $\Delta =2$. Now the genus is reduced to one. Next we simultaneously move the vertical cut between the sheets with labels $1^+$ and $4^+$ and the vertical cut between the sheets with labels $1^-$ and $4^-$ along the same way until it reaches $\Delta =-2$ and continue on until it reaches the unique real branch point in the short segment. This has the effect that over $\Delta
=-2$, the branch point which previously connected the sheets with labels $4^\pm$ and $2^\pm$, now connects the sheets with labels $1^\pm$ and $2^\pm$, and the branch point over $\Delta =-2$ which previously connected the sheets with labels $3^\pm$ and $1^\pm$, now connects the sheets with labels $3^\pm$ and $4^\pm$. The vertical cuts now connect the sheets with labels $3^\pm$ and $2^\pm$. Now we move the real branch point in the short segment to $\Delta =2$. To preserve positivity of the mean curvature, we move the other real branch point, which connects the sheets with labels $4^+$ and $4^-$, towards $\Delta =2$ and possibly further to and fro. The genus is thus reduced to zero. We arrive at a constellation described in the proof of Theorem \[thm:onesided revolution\] with $\kappa_0 \in
(0,\,1)$, which we know not to be one-sided Alexandrov embedded.
\[thm:main1\] A one-sided Alexandrov embedded cylinder of finite type in the 3-sphere is a surface of revolution.
We will show that all spectral data of one-sided Alexandrov embedded cylinders of finite type in ${\mathbb S}^3$ are described in Theorem \[thm:onesided revolution\]. Due to Lemma \[thm:continuous genus 1\] all spectral data corresponding to a one–sided Alexandrov embedded cylinder can be continuously deformed within the class of $\Delta$ corresponding to one-sided Alexandrov embedded cylinders by a combination of continuous deformations described in Theorem \[thm:spec\_AE\] [[([N]{})]{}]{}-[[([P]{})]{}]{} into one of the spectral data of the flat cylinders in ${\mathbb S}^3$ described in Theorem \[thm:onesided revolution\]. Due to Lemma \[thm:not genus two\] the $\Delta$ corresponding to these flat cylinders in ${\mathbb S}^3$ can be continuously deformed within the class of one-sided Alexandrov embedded cylinders only into the $\Delta$ described in Theorem \[thm:onesided revolution\]. All cylinders with bubbletons of finite type, whose $\Delta$ are described in Theorem \[thm:onesided revolution\] can be deformed within the class of deformations preserving the one-sided Alexandrov embeddedness into the cylinders with bubbletons described in Lemma \[thm:not genus two\]. Hence they are not one-sided Alexandrov embedded. Consequently, all spectral data corresponding to one-sided Alexandrov embedded cylinders in ${\mathbb S}^3$ are described in Theorem \[thm:onesided revolution\]. They are all surfaces of revolution!
\[th:Pinkall-Sterling\] All one-sided Alexandrov embedded tori in the 3-sphere are tori of revolution. In particular, all embedded tori in the 3-sphere are tori of revolution.
Let $f:N\rightarrow{\mathbb S}^3$ be a one-sided Alexandrov embedding of a torus $M=\partial N\simeq\mathbb{T}^2$. We will show that there exists a cover $\tilde{N}\rightarrow N$, whose composition with $f$ is a cylinder which is also one-sided Alexandrov embedded. If the class in $\pi_1(M)$ of a smooth embedding $\gamma:{\mathbb S}^1\hookrightarrow M$ belongs to the kernel of $\pi_1(M)\rightarrow\pi_1(N)$, then, due to Dehn’s Lemma, $\gamma$ is the boundary of a smooth embedding from the closed two-dimensional disc $\bar{\mathbb{D}}\hookrightarrow N$. The intersection number of another path ${\mathbb S}^1\rightarrow M$ with $\gamma$ in $M$ is equal to the intersection number with the disc in $N$. Hence the kernel of $\pi_1(M)\rightarrow\pi_1(N)$ is isotropic with respect to the intersection form and therefore cyclic. In particular, there exists a subgroup of $\Gamma\subset\pi_1(N)$, whose pre-image under $\pi_1(M)\rightarrow \pi_1(N)$ is isomorphic to $\mathbb{Z}$. Furthermore, we can choose the subgroup $\Gamma$ such that the restriction of $\pi_1(M)\rightarrow\pi_1(N)$ to the subgroup $\mathbb{Z}\subset\pi_1(M)$ is surjective onto $\Gamma$. This subgroup $\Gamma\subset\pi_1(N)$ corresponds to a covering $\tilde{N}\rightarrow N$, whose boundary $\tilde{M}=\partial\tilde{N}$ is a cylinder [@St §14]. Hence there exists a one-sided Alexandrov embedded cylinder $\tilde{f}:\tilde{N}\rightarrow{\mathbb S}^3$, which is the composition of $f$ with a covering map. Obviously the spectral data of $f$ and $\tilde{f}$ coincide. Due to Pinkall and Sterling [@PinS] this cylinder is of finite type, and by Theorem \[thm:main1\], it is a surface of revolution.
Hsiang and Lawson [@HsiL] prove that there are no embedded minimal tori of cohomogeneity one. Hence the Clifford torus is the only embedded minimal torus of revolution. By Corollary \[th:Pinkall-Sterling\] the only embedded tori are tori of revolution, and we thus affirm Lawson’s conjecture.
The Clifford torus is the only embedded minimal torus in the 3-sphere.
Since all one-sided Alexandrov embedded cylinders in ${\mathbb S}^3$ are surfaces of revolution around a closed geodesic, the ambient 3-manifold is diffeomorphic to $\overline{\mathbb{D}}\times{\mathbb R}$, where $\overline{\mathbb{D}}$ denotes the closed unit disk. We thus have the following generalization of Lawson’s ’unknottedness’ result [@Law:unknot].
For all one-sided Alexandrov embedded cylinders of finite type in the 3-sphere, the 3-manifold is diffeomorphic to the cartesian product $\overline{\mathbb{D}}\times{\mathbb R}$.
\[2\][ [\#2](http://www.ams.org/mathscinet-getitem?mr=#1) ]{} \[2\][\#2]{}
[10]{}
U. Abresch, *Constant mean curvature tori in terms of elliptic functions*, J. Reine U. Angew Math. **374** (1987), 169–192.
, *Old and new doubly periodic solutions of the sinh-[G]{}ordon equation*, Seminar on new results in nonlinear partial differential equations (Bonn, 1984), Aspects Math., E10, Vieweg, Braunschweig, 1987, pp. 37–73.
R. Alexander and S. Alexander, *Geodesics in [R]{}iemannian manifolds-with-boundary*, Indiana Univ. Math. J. **30** (1981), no. 4, 481–488.
A. D. Alexandrov, *Uniqueness theorems for surfaces in the large. [I]{}*, Vestnik Leningrad. Univ. **11** (1956), no. 19, 5–17.
A. I. Bobenko, *All constant mean curvature tori in [$\mathbb{R}^3$]{}, [$\mathbb{S}^3$]{}, [$\mathbb{H}^3$]{} in terms of theta-functions*, Math. Ann. **290** (1991), 209–245.
, *Constant mean curvature surfaces and integrable equations*, Russian Math. Surveys **46** (1991), 1–45.
, *Surfaces in terms of 2 by 2 matrices. old and new integrable cases*, Harmonic maps and integrable systems, Aspects of Mathematics, vol. E23, Vieweg, 1994.
F. E. Burstall, *Isothermic surfaces: conformal geometry, [C]{}lifford algebras and integrable systems*, Integrable systems, geometry, and topology, AMS/IP Stud. Adv. Math., vol. 36, Amer. Math. Soc., Providence, RI, 2006, pp. 1–82.
F. E. Burstall, D. Ferus, F. Pedit, and U. Pinkall, *Harmonic tori in symmetric spaces and commuting [H]{}amiltonian systems on loop algebras*, Ann. of Math. **138** (1993), 173–212.
F. E. Burstall and M. Kilian, *Equivariant harmonic cylinders*, Quart. J. Math. **57** (2006), 449–468.
F. E. Burstall and F. Pedit, *Harmonic maps via [A]{}dler-[K]{}ostant-[S]{}ymes theory*, Harmonic maps and integrable systems, Aspects of Mathematics, vol. E23, Vieweg, 1994.
, *Dressing orbits of harmonic maps*, Duke Math. J. **80** (1995), no. 2, 353–382.
E. Carberry, *On the existence of minimal tori in ${S}^3$ of arbitrary spectral genus*, Ph.D. thesis, Princeton University, 2002.
J. B. Conway, *Functions of one complex variable. [II]{}*, Graduate Texts in Mathematics, vol. 159, Springer-Verlag, New York, 1995.
E. B. Davies, *[$L^1$]{} properties of second order elliptic operators*, Bull. London Math. Soc **17** (1985), 417–436.
J. Dorfmeister and G. Haak, *On constant mean curvature surfaces with periodic metric*, Pacific J. Math. **182** (1998), 229–287.
J. Dorfmeister, F. Pedit, and H. Wu, *Weierstrass type representation of harmonic maps into symmetric spaces*, Comm. Anal. Geom. **6** (1998), no. 4, 633–668.
N. M. Ercolani, H. Kn[ö]{}rrer, and E. Trubowitz, *Hyperelliptic curves that generate constant mean curvature tori in [${\bf R}\sp 3$]{}*, Integrable systems (Luminy, 1991), Progr. Math., vol. 115, Birkhäuser Boston, Boston, MA, 1993, pp. 81–114.
J.-H. Eschenburg, *Maximum principle for hypersurfaces*, Manuscripta Math. **64** (1989), no. 1, 55–75.
D. Fischer-Colbrie and R. Schoen, *The structure of complete stable minimal surfaces in 3-manifolds of nonnegative scalar curvature*, CPAM **33** (1980), 199–211.
S. Fornari, J. H. S. deLira, and J. Ripoll, *Geodesic graphs with constant mean curvature in spheres*, Geom. Dedic. **90** (2002), 201–216.
P. G. Grinevich and M. U. Schmidt, *Period preserving nonisospectral flows and the moduli space of periodic solutions of soliton equations*, Phys. D **87** (1995), no. 1-4, 73–98.
K. Gro[ß]{}e-Brauckmann, R. Kusner, and J. M. Sullivan, *Triunduloids: embedded constant mean curvature surfaces with three ends and genus zero*, J. [R]{}eine [A]{}ngew. [M]{}ath. **564** (2003), 35–61.
J. Hebda, *Cut loci of submanifolds in space forms and in the geometries of [M]{}öbius and [L]{}ie*, Geom. Dedicata **55** (1995), no. 1, 75–93.
N. Hitchin, *Harmonic maps from a 2-torus to the 3-sphere*, J. Differential Geom. **31** (1990), no. 3, 627–710.
H. Hopf, *Differential geometry in the large*, Lecture Notes in Mathematics, vol. 1000, Springer-Verlag, 1983.
W-Y. Hsiang and H. B. Lawson, Jr., *Minimal submanifolds of low cohomogeneity*, J. Differential Geometry **5** (1971), 1–38.
A. Hurwitz, *Ueber [R]{}iemann’sche [F]{}lächen mit gegebenen [V]{}erzweigungspunkten*, Math. Ann. **39** (1891), no. 1, 1–60.
C. Jaggy, *On the classification of constant mean curvature tori in [${\bf
R}\sp 3$]{}*, Comment. Math. Helv. **69** (1994), no. 4, 640–658.
N. Kapouleas, *Complete constant mean curvature surfaces in [E]{}uclidean three space*, Ann. Math. **131** (1990), 239–330.
, *Compact constant mean curvature surfaces in [E]{}uclidean three space*, J. Diff. Geom. **33** (1991), 683–715.
, *Constant mean curvature surfaces by fusing [W]{}ente tori*, Proc. Natl. Acad. Sci. USA **89** (1992), 5695–5698.
N. Kapouleas and S.-D. Yang, *Minimal surfaces in the three-[S]{}phere by doubling the [C]{}lifford [T]{}orus*, arXiv:math/0702565v1.
K. Karcher, U. Pinkall, and I. Sterling, *New minimal surfaces in [$S\sp3$]{}*, J. Differential Geom. **28** (1988), 169–185.
M. Kilian, *On the associated family of [D]{}elaunay surfaces*, Proc. Amer. Math. Soc **132** (2004), no. 10, 3075–3082.
M. Kilian, I. McIntosh, and N. Schmitt, *New constant mean curvature surfaces*, Experiment. Math. **9** (2000), no. 4, 595–611.
M. Kilian, N. Schmitt, and I. Sterling, *Dressing [CMC]{} n-[N]{}oids*, Math. Z. **246** (2004), no. 3, 501–519.
N. Korevaar, R. Kusner, and J. Ratzkin, *On the nondegeneracy of constant mean curvature surfaces*, Geom. Funct. Anal. **16** (2006), no. 4, 891–923.
N. Korevaar, R. Kusner, and B. Solomon, *The structure of complete embedded surfaces with constant mean curvature*, J. Diff. Geom. **30** (1989), no. 2, 465–503.
I. M. Kri[č]{}ever, *Methods of algebraic geometry in the theory of nonlinear equations*, Uspehi Mat. Nauk **32** (1977), no. 6(198), 183–208, 287, English translation: Russian Math. Surveys 32 (1977), no. 6, 185–213.
R. Kusner, R. Mazzeo, and D. Pollack, *The moduli space of complete embedded constant mean curvature surfaces*, Geom. Funct. Anal. **6** (1996), 120–137.
H. B. Lawson, Jr., *Compact minimal surfaces in [$S\sp{3}$]{}*, Global Analysis (Proc. Sympos. Pure Math., Vol. XV, Berkeley, Calif., 1968), Amer. Math. Soc., Providence, R.I., 1970, pp. 275–282.
, *Complete minimal surfaces in [$S\sp{3}$]{}*, Ann. of Math. (2) **92** (1970), 335–374.
, *The unknottedness of minimal embeddings*, Invent. Math. **11** (1970), 183–187.
V. A. Mar[č]{}enko and I. V. Ostrovski[ĭ]{}, *A characterization of the spectrum of the [H]{}ill operator*, Mat. Sb. (N.S.) **97(139)** (1975), no. 4(8), 540–606, 633–634.
I. McIntosh, *Harmonic tori and their spectral data*, math.DG/0407248. To appear in ’Surveys on [G]{}eometry and [I]{}ntegrable [S]{}ystems’, Advanced Studies in Pure Mathematics, 2007.
W. H. [Meeks III]{}, *The topology and geometry of embedded surfaces of constant mean curvature*, J. Diff. Geom. **27** (1988), 539–552.
W. H. [Meeks III]{}, J. Perez, and A. Ros, *Limit leaves of a cmc lamination are stable*, arXive:0801.4345.
S. Montiel and A. Ros, *Minimal immersions of surfaces by the first eigenfunctions and conformal area.*, Invent. Math. **83** (1986), 153–166.
, *Compact hypersurfaces: the [A]{}lexandrov theorem for higher order mean curvatures*, Differential geometry, Pitman Monogr. Surveys Pure Appl. Math. **52** (1991), 279–296.
U. Pinkall and I. Sterling, *On the classification of constant mean curvature tori*, Ann. Math. **130** (1989), 407–451.
K. Pohlmeyer, *Integrable hamiltonian systems and interaction through quadratic constraints*, Comm. Math. Phys. **46** (1976), 207–221.
J. P[ö]{}schel and E. Trubowitz, *Inverse spectral theory*, Pure and Applied Mathematics, vol. 130, Academic Press Inc., Boston, MA, 1987.
A. Pressley and G. Segal, *Loop groups*, Oxford Science Monographs, Oxford Science Publications, 1988.
W. Rinow, *Die innere [G]{}eometrie der metrischen [R]{}äume*, Die Grundlehren der mathematischen Wissenschaften, Bd. 105, Springer-Verlag, Berlin, 1961.
A. Ros, *A two-piece property for compact minimal surfaces in a three-sphere.*, Indiana Univ. Math. J. **44** (1995), no. 3, 841–849.
H. Rosenberg, *Private communication*, 27/03/2008.
M. U. Schmidt, *Integrable systems and [R]{}iemann surfaces of infinite genus*, Mem. Amer. Math. Soc. **122** (1996), no. 581, viii+111.
N. Schmitt, *cmclab*, http://www.gang.umass.edu/software.
N. Schmitt, M. Kilian, S. Kobayashi, and W. Rossman, *Unitarization of monodromy representations and constant mean curvature trinoids in three dimensional space forms*, J. London Math. Soc. **75** (2007), 563–581.
R. Schoen, *Stable minimal surfaces in three manifolds*, Annals of Mathematics Studies **103** (1983), 111–126.
J. P. Serre, *Algebraic goups and class fields*, Graduate Texts in Mathematics, No. 117, Springer-Verlag, New York, 1988.
N. Steenrod, *The [T]{}opology of [F]{}ibre [B]{}undles*, Princeton Mathematical Series, vol. 14, Princeton University Press, Princeton, N. J., 1951.
I. Sterling and H. Wente, *Existence and classification of constant mean curvature multibubbletons of finite and infinite type*, Indiana Univ. Math. J. **42** (1993), no. 4, 1239–1266.
W. W. Symes, *Systems of [T]{}oda type, inverse spectral problems, and representation theory*, Invent. Math. **59** (1980), no. 1, 13–51.
C. Terng and K. Uhlenbeck, *Bäcklund transformations and loop group actions*, Comm. Pure and Appl. Math **LIII** (2000), 1–75.
V. Tkachenko, *Spectra of non-selfadjoint [H]{}ill’s operators and a class of [R]{}iemann surfaces*, Ann. of Math. (2) **143** (1996), no. 2, 181–231.
K. Uhlenbeck, *Harmonic maps into lie groups (classical solutions of the chiral model)*, J. Diff. Geom. **30** (1989), 1–50.
M. Umehara and K. Yamada, *A deformation of tori with constant mean curvature in [${\bf R}\sp 3$]{} to those in other space forms*, Trans. Amer. Math. Soc. **330** (1992), no. 2, 845–857.
R. Walter, *[E]{}xplicit examples to the $h$-problem of [H]{}einz [H]{}opf*, Geom. Dedic. **23** (1987), 187–213.
H. C. Wente, *Counterexample to a conjecture of [H]{}. [H]{}opf*, Pac. J. Math. **121** (1986), 193–243.
, *Twisted tori of constant mean curvature in [$\mathbb{R}^3$]{}*, Seminar on new results in nonlinear partial differential equations (Bonn, 1984), Aspects Math., E10, Vieweg, Braunschweig, 1987, pp. 1–36.
T. Wolff, *Recent work on sharp estimates in second order elliptic unique continuation*, Jour. Geom. Ana. **3** (1993), no. 6, 621–650.
[^1]: [*Mathematics Subject Classification.*]{}53A10, 53C17.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We define and study the degeneration property for ${\mathrm{BV}_\infty}$ algebras and show that it implies that the underlying $L_\infty$ algebras are homotopy abelian. The proof is based on a generalisation of the well-known identity $\Delta (e^{\xi})=e^{\xi}\left (\Delta(\xi)+\frac{1}{2}[\xi,\xi]\right )$ which holds in all BV algebras. As an application we show that the higher Koszul brackets on the cohomology of a manifold supplied with a generalised Poisson structure all vanish.'
address:
- |
Centre for Mathematical Science\
City University London\
Northampton Square\
London EC1V 0HB\
UK
- |
Department of Mathematics and Statistics\
Lancaster University\
Lancaster LA1 4YF\
UK
author:
- 'C. Braun'
- 'A. Lazarev'
bibliography:
- 'references.bib'
title: Homotopy BV algebras in Poisson geometry
---
Introduction
============
A Batalin–Vilkovisky (BV) algebra is a graded commutative algebra supplied with an odd differential operator of order two and square zero. It appears in various contexts of algebraic topology, differential geometry and mathematical physics. As is usual for most algebraic structures, there is a notion of a *homotopy*, or infinity, BV algebra structure encoding higher invariants of BV algebras. The general treatment of homotopy BV algebras is contained in [@GalvezTonksVallette12:BV], however for us the term ${\mathrm{BV}_\infty}$ algebra has a more restricted meaning, essentially equivalent to the definition in [@kravchenko2000:deformationsofbvalgebras].
An important special class of differential graded (dg) BV algebras is formed by imposing the *degeneration property*, introduced in [@katzarkovkontsevichpantev2008:hodgetheoretic] and [@terilla2008:smoothnessbv]. This property holds in e.g. the de Rham algebra of a symplectic manifold or the Dolbeault algebra of a Calabi–Yau manifold and in favourable cases it leads to a construction of a formal Frobenius manifold [@BarannikovKontsevich98:Frobenius; @Merkulov98:Frobenius].
A dg BV algebra supports the structure of a dg Lie algebra, whereas a ${\mathrm{BV}_\infty}$ algebra gives rise to a homotopy analogue of a Lie algebra, called an $L_\infty$ algebra. The ordinary degeneration property for a dg BV algebra implies that its underlying dg Lie algebra is *homotopy abelian*, i.e. that it is quasi-isomorphic to an abelian Lie algebra. We will prove a generalisation of this statement for ${\mathrm{BV}_\infty}$ algebras. The proof is based on a generalisation of the following well-known identity for ordinary dg BV algebras: $$\Delta(e^\xi) = e^\xi\left ( \Delta(\xi) + \frac{1}{2}[\xi,\xi] \right )$$ This generalisation holds, essentially, for all operators $\Delta$, not necessarily of second order.
Our main application of the degeneration property for ${\mathrm{BV}_\infty}$ algebras concerns the structure of the de Rham algebra of a manifold $M$ supplied with a generalised Poisson structure. The latter is just a multivector field on $M$ whose Schouten bracket with itself is zero. An ordinary Poisson structure (a bivector field) on $M$ gives rise to a Koszul bracket on $\Omega(M)$, the de Rham algebra of $M$, making the latter a dg Lie algebra (in fact, a dg BV algebra). It was shown in [@voronovkhudaverdian2008:higerpoissonbrackets] that a generalised Poisson structure on $M$ leads to an $L_\infty$ structure on $\Omega(M)$ (in fact, to a ${\mathrm{BV}_\infty}$ structure on $\Omega(M)$). We show that this ${\mathrm{BV}_\infty}$ algebra has the degeneration property. As a consequence the higher Koszul brackets on the cohomology of $M$ vanish.
Notation and conventions
------------------------
Throughout this paper ${\mathbf{k}}$ will denote a field of characteristic zero.
We choose to work in the supergraded context. More precisely, this means we work in the category of super vector spaces: $\mathbb{Z}/2\mathbb{Z}$–graded ${\mathbf{k}}$–linear vector spaces, with morphisms linear maps preserving the grading. This is a symmetric monoidal category with symmetry isomorphism $s{\colon}V\otimes W \to W\otimes V$ given by $$s(v\otimes w) = (-1)^{{\lvert v \rvert}{\lvert w \rvert}}w\otimes v.$$ Denote by $\Pi{\mathbf{k}}$ the one dimensional super vector space concentrated in odd degree. We denote by $\Pi V$ the functor $V\mapsto \Pi{\mathbf{k}}\otimes V$, called *parity reversion*. The space $\operatorname{\underline{Hom}}(V,W)$ denotes the super vector space with even part the space of morphisms $V\to W$ (the linear maps which preserve the grading) and odd part the space of morphisms $V\to \Pi W$ (the linear maps which reverse the grading). This is an internal $\operatorname{Hom}$ functor making the category of super vector spaces into a closed symmetric monoidal category.
For brevity we will normally suppress the adjective ‘super’. In particular by a (unital) associative/commutative[^1]/Lie algebra we will always mean the appropriate notions in this category. This means that commutativity, anti-commutativity and the Jacobi identity are to be understood in the graded sense, for example commutativity would mean $$ab = (-1)^{{\lvert a \rvert}{\lvert b \rvert}}ba.$$ The abbreviation ‘dg’ will stand for ‘differential (super)graded’ and we will abbreviate the expression ‘commutative dg algebra’ to ‘cdga’.
We note that most of our results also hold in the $\mathbb{Z}$–graded context, after making suitable cosmetic adaptations such as replacing parity reversion with suspension/desuspension as appropriate.
We will use the notion of a *complete* (dg) vector space; this is just an inverse limit of finite-dimensional (dg) vector spaces. An example of a complete vector space is $V^*$, the ${\mathbf{k}}$–linear dual to a discrete vector space $V$. A complete vector space comes equipped with a topology and whenever we deal with a complete vector space all linear maps from or into it will be assumed to be continuous; thus we will always have $V^{**}\cong
V$. Similarly, we will always have $(V\otimes V)^*\cong V^*\otimes V^*$ since tensor products of complete vector spaces $C = \lim_\leftarrow C_i$ and $B=\lim_\leftarrow B_i$ will always be assumed to mean *completed* tensor products, in other words $C\otimes B$ is the complete vector space $C\otimes B = \lim_\leftarrow C_i \otimes B_j$. If $V$ is a discrete vector space and $C = \lim_{\leftarrow} C_i$ is a complete vector space, the tensor product $C\otimes V$ will always be assumed to mean $\lim_\leftarrow C_i\otimes V$.
A *complete algebra* is an algebra in complete spaces which, in addition, is also local. A prototypical example of a complete algebra is the completed symmetric algebra $\widehat{S}V$ on a complete vector space $V$.
L-infinity algebras
===================
This introductory section fixes the terminology and standard facts about $L_\infty$ algebras relevant to the present work. More detailed discussion can be found in, e.g. [@ChuangLazarev:twisting].
Let $V$ be a vector space; then its dual is a complete vector space and we can form its complete symmetric algebra $\widehat{S}V^*$. Let us denote by ${\mathrm{Der}}(\widehat{S}V^*)$ the Lie algebra of (continuous) derivations of $\widehat{S}V^*$. Choosing a basis $x_i,i\in I$ in $V^*$, any derivation $\xi\in{\mathrm{Der}}(\widehat{S}V^*)$ can be written as $$\xi=\sum_{i\in I}f_i^0\partial_{x_i}+\sum_{i\in I} f_i^1\partial_{x_i}+\dots+\sum_{i\in I} f_i^k\partial_{x_i}+\dots$$ where $f_i^k$ is a linear combination (perhaps infinite if the indexing set $I$ is infinite) of monomials in $x_n$s of order $k$. If $\xi=\sum_{i\in I} f_i^n\partial_{x_i}$ for a fixed $n$ then we say that $\xi$ is a derivation of order $n$; this notion clearly does not depend on the choice of the basis in $V^*$. The space of derivations of order $\geq n$ will be denoted by ${\mathrm{Der}}_{\geq n}(\widehat{S}V^*)$.
Let $V$ be a vector space. An $L_\infty$ structure on $V$ is an odd element $m\nobreak\in\nobreak{\mathrm{Der}}_{\geq 1}(\widehat{S}\Pi V^*)$ which satisfies the equation $m^2=m\circ m\nobreak=\nobreak 0$. The pair $(V,m)$ will be referred to as an $L_\infty$ algebra and the algebra $\widehat{S}\Pi V^*$, supplied with the differential $m$, as its representing complete cdga.
There is a concomitant notion of an $L_\infty$ map.
Let $(V,m_V)$ and $(W,m_W)$ be two $L_\infty$ structures on $V$ and $W$. An $L_\infty$ map $f{\colon}(V,m_V)\to(W,m_W)$ is a map between their representing complete cdgas $f{\colon}\widehat{S}\Pi W^*\to \widehat{S}\Pi V^*$ such that $f\circ m_W=m_V\circ f$.
A more traditional approach to defining $L_\infty$ algebras and maps is through multilinear maps. Note that a derivation $m\in{\mathrm{Der}}_{\geq 1}(\widehat{S}\Pi V^*)$ has the form $m=m_1^*+m_2^*+\dots$ where $m_n^*$ is a derivation of order $n$. In other words, any derivation is determined by the collection of maps $m_n^* {\colon}\Pi V^*\to\left ((\Pi V^*)^{\otimes n}\right )_{S_n}$. We have an identification between $S_n$ coinvariants and $S_n$ invariants: $$i_n{\colon}\left((\Pi V^*)^{\otimes n}\right)_{S_n}\to
\left((\Pi V^*)^{\otimes n}\right)^{S_n}\cong
\left((\Pi V^{\otimes n})_{S_n}\right)^*$$ where $i_n(x_1\otimes \dots \otimes x_n)=\sum_{\sigma\in S_n}\sigma[x_1\otimes\dots\otimes x_n]$. Then the dual to the composite map $$i_n\circ m_n^*{\colon}\Pi V^*\to \left((\Pi V^{\otimes n})_{S_n}\right)^*$$ is a map $m_n {\colon}(\Pi V^{\otimes n})_{S_n}\to\Pi V$. Thus an $L_\infty$ structure on $V$ is equivalent to a collection of symmetric multilinear maps $m_n{\colon}(\Pi V)^{\otimes n}\to \Pi V$ of odd degree as above subject to appropriate conditions stemming from the equation $m\circ m=0$. For example, the linear component $m_1{\colon}\Pi V \to \Pi V$ is a differential on $\Pi V$.
A similar argument shows that an $L_\infty$ map $f{\colon}\widehat{S}\Pi W^*\to \widehat{S}\Pi V^*$ is equivalent to a collection of symmetric multilinear maps $f_n{\colon}(\Pi V)^{\otimes n}\to \Pi W$ of even degree satisfying suitable identities.
We can now define the notion of an $L_\infty$ (quasi-)isomorphism.
An $L_\infty$ map $f{\colon}(V,m_V)\to (W,m_W)$ is an $L_\infty$ (quasi-)isomorphism if its linear component $f_1{\colon}\Pi V \to \Pi W $ is a (quasi-)isomorphism where $\Pi V$ and $\Pi W$ are supplied with differentials the linear components of $m_V$ and $m_W$.
Minimal L-infinity algebras and homotopy abelian L-infinity algebras
--------------------------------------------------------------------
An important special class of $L_\infty$ algebras is formed by *minimal* $L_\infty$ algebras; these are derivations $m\in{\mathrm{Der}}_{\geq 2}(\widehat{S}\Pi V^*)$. Minimal $L_\infty$ algebras have the striking property that any $L_\infty$ quasi-isomorphism between them must be an isomorphism. Any $L_\infty$ algebra is $L_\infty$ quasi-isomorphic to a minimal one (cf. for example [@kontsevich2003:defquant Lemma 4.9]), called its *minimal model*. Any two minimal models of a given $L_\infty$ algebra are (non-canonically) isomorphic.
An $L_\infty$ algebra is *abelian* if all its higher brackets $m_n$ vanish for $n\geq 2$. It is *homotopy abelian* if it is $L_\infty$ quasi-isomorphic to an abelian $L_\infty$ algebra. Clearly an $L_\infty$ algebra is homotopy abelian if and only if its minimal model $(V,m_V)$ has its $L_\infty$ structure $m_V$ identically vanishing: $m_V=0$.
Maurer–Cartan elements in L-infinity algebras
---------------------------------------------
Let $m$ be an $L_\infty$ structure on a vector space $V$ and let $C$ be a complete cdga with maximal ideal $C_+$. Then an even element $\xi\in C_+\otimes \Pi V$ is *Maurer–Cartan* if it satisfies the Maurer–Cartan equation $$(d_C\otimes {\mathrm{id}})(\xi) +
\sum_{i=1}^\infty
\frac{1}{i!}m^C_i(\xi,\dots,\xi) = 0$$ where $m_i^C$ is obtained by extending $m_i$ multilinearly in $C$. The set of Maurer–Cartan elements in $C_+\otimes \Pi V$ will be denoted by ${\mathrm{MC}}(V,C)$. The correspondence $(V,C)\mapsto {\mathrm{MC}}(V,C)$ is functorial in $C$.
The significance of Maurer–Cartan elements stems from the following standard result.
\[MCfunctor\] Let $(V,m)$ be an $L_\infty$ algebra and $C$ be a complete cdga. Then the functor $C\mapsto{\mathrm{MC}}(V,C)$ is represented by the complete cdga $(\widehat{S}\Pi V^*,m)$.
\[rem:Yoneda\] An $L_\infty$ map $V\to W$ gives rise, for any complete cdga $C$, to a map of sets ${\mathrm{MC}}(V,C)\to{\mathrm{MC}}(W,C)$ which is functorial in $C$, in other words a natural transformation. In fact such a natural transformation is, by Yoneda’s lemma, equivalent to having an $L_\infty$ map $V\to W$. This observation is often useful for constructing $L_\infty$ maps.
BV-infinity algebras and differential operators
===============================================
For any associative algebra $A$, by a linear operator on $A$ we mean an element of the associative algebra $\operatorname{End}(A) = \operatorname{\underline{Hom}}(A,A)$. Any element of $a$ is regarded as a linear operator on $A$ by $a(x) = ax$.
Let $A$ be a unital associative algebra with pronilpotent ideal $I$. Then for any $a\in I$ we define: $$e^{a} = 1 + a + \frac{a^2}{2!} + \frac{a^3}{3!} + \dots \in A$$ Given any $b\in A$ such that $b - 1 \in I$ we define: $$\log{b} = (b-1) - \frac{(b-1)^2}{2} + \frac{(b-1)^3}{3} - \dots \in I$$ For any $a\in I$ we have $\log{e^a} = a$ and for any $b\in A$ with $b-1\in I$ we have $e^{\log{b}} = b$.
Recall the following well-known identity, which holds for any $a\in I$ and any $D$ a continuous linear operator on $A$: $$\operatorname{Ad}(e^a)(D) = e^a D e^{-a} = e^{\operatorname{ad}(a)}(D)$$ Rearranging, we obtain the following version of this identity, which will be important for us. $$\label{eq:exponentialidentity}
D e^{a} = e^{a}e^{-\operatorname{ad}(a)}(D)$$
Differential operators
----------------------
Recall that if $A$ is a unital commutative algebra then there is an increasing filtration $F_0\subset F_1 \subset \dots \subset \operatorname{End}(A)$ of $\operatorname{End}(A)$ where $F_n$ is the space of differential operators of order not higher than $n$.
More precisely, set $F_{-1} = \{0\}$ and define recursively $$F_n = \{D\in \operatorname{End}(A) {\colon}\forall a\in A, [D,a] \in F_{n-1}\}.$$ Unwrapping this recursive definition, $D$ is a differential operator of order not higher than $n$ if for any $a_1,\dots, a_{n+1}\in A$ it holds that $$[[[\dots[D,a_1]\dots],a_n],a_{n+1}] = 0.$$
Given $D\in \operatorname{End}(A)$, for $n\geq 0$ define the linear maps $m_n{\colon}A^{\otimes n} \to A$ by $$m_n(a_1,\dots, a_n) = [[[\dots[D,a_1]\dots],a_{n-1}],a_n](1).$$ Then $D$ is a differential operator of order not higher than $n$ if and only if $m_{n+1} = 0$ if and only if $m_{n+i} = 0$ for all $i\geq 1$.
The terminology ‘$n$–th order differential operator’ is a more elegant phrase, but technically speaking it could be ambiguous. We will use it to mean a differential operator of order not higher than $n$.
L-infinity algebras from linear operators
-----------------------------------------
Since $A$ is commutative the maps $m_n$ are symmetric. Given an operator $D$ then we obtain maps $m_n^*{\colon}A^* \to \left ((A^*)^{\otimes n}\right )_{S_n}$ and hence a derivation $m\in{\mathrm{Der}}(\widehat{S}A^*)$. We have the following observation due originally to Kravchenko.
\[prop:krav\] $D^2=0$ if and only if $m^2=0$.
We now have the following corollary, generalising in a certain way the construction of the Lie bracket associated to a BV operator. This can also be realised as an instance of Voronov’s construction of higher derived brackets [@voronov2005:higherderivedbrackets; @voronov2005:higherderivedbracketsarbitrary].
\[cor:linftyconstruction\] Let $A$ be a unital commutative algebra and let $D$ be a linear operator of odd degree such that $D^2=0$ and $D(1)=0$. Then $m\in{\mathrm{Der}}(\widehat{S}A^*)$ defines an $L_\infty$ structure on $\Pi A$.
If $D(1)\neq 0$ then $m$ will in fact define a *curved* $L_\infty$ algebra with curvature $D(1)$. Many of the results we state will also hold after adding the adjective ‘curved’ at appropriate places. However, since we will not make use of curved $L_\infty$ algebras we will not consider this, for the sake of simplicity.
This $L_\infty$ structure is always homotopy abelian, which we shall now show. We first prove the following straightforward lemma.
\[lem:expmc\] Let $A$ be a unital commutative algebra with a linear operator of odd degree with $D^2 =0$ and $D(1)=0$ giving rise to an $L_\infty$ structure on $\Pi A$ and let $C$ be a complete cdga with maximal ideal $C_+$. Then an even element $\xi \in C_+ \otimes A$ satisfies the Maurer–Cartan equation if and only if $$e^{-\operatorname{ad}(\xi)}(d_C + D)(1) = 0.$$
Note that since $\xi \in C_+ \otimes A$ the left hand side converges. The operations $m_n^C{\colon}(C \otimes A)^{\otimes n}\to C\otimes A$ satisfy: $$m_n^C(\xi,\dots,\xi) = (-\operatorname{ad}(\xi))^n(D)(1)$$ Therefore, unwrapping this formula explicitly and remembering that $d_C$ is in particular a first order differential operator we obtain: $$e^{-\operatorname{ad}(\xi)}(d_C + D)(1) = \sum_{i=0} \frac{1}{i!}(-\operatorname{ad}(\xi))^i (d_C+D)(1) = (d_C\otimes {\mathrm{id}})(\xi) + \sum_{i=1} \frac{1}{i!}m_i^C(\xi,\dots,\xi)$$ In other words, this equation is simply the Maurer–Cartan equation.
The proof of taken together with in fact generalises to the $L_\infty$ context the standard identity for a BV algebra with second order BV operator $D = \Delta$: $$\Delta (e^\xi) = e^\xi\left(\Delta(\xi) + \frac{1}{2}[\xi,\xi]\right)$$ Indeed, unwrapping in terms of the higher $L_\infty$ operations we obtain (under appropriate continuity and convergence conditions) for any operator $D$ with $D(1)=0$: $$D(e^\xi) = e^\xi \left( D(\xi) + \frac{1}{2!}m_2(\xi,\xi) + \frac{1}{3!}m_3(\xi,\xi,\xi) + \dots \right )$$
\[thm:homotopyabelian\] Let $A$ be a unital commutative algebra with a linear operator of odd degree with $D^2 =0$ and $D(1)=0$ giving rise to an $L_\infty$ structure on $\Pi A$. Then this $L_\infty$ algebra is homotopy abelian.
Let $C$ be a complete cdga with maximal ideal $C_+$. Since $d_C(1) = D(1)=0$, together with we have for $\xi \in C_+ \otimes A$: $$(d_C + D) (e^{\xi}-1) = e^{\xi}\left (e^{-\operatorname{ad}(\xi)}(d_C + D)(1)\right )$$ Therefore, for $\xi$ of even degree, by $e^\xi - 1$ is a $(d_C + D)$–cycle if and only $\xi$ is Maurer–Cartan. So the invertible map $\xi \mapsto e^\xi - 1$ gives rise to a natural isomorphism of Maurer–Cartan sets $${\mathrm{MC}}(\Pi A, C)\cong {\mathrm{MC}}(\mathfrak{h},C)$$ where $\mathfrak{h}$ is the abelian $L_\infty$ algebra with underlying space $\Pi A$, differential $m_1 = D$ and $m_n=0$ for all $n\geq 2$. It follows from that $\Pi A$ and $\mathfrak{h}$ are $L_\infty$ isomorphic.
The proof of simply constructs an automorphism of commutative algebras $\widehat{S}A^* \to \widehat{S}A^*$, commuting with two differential operators on $\widehat{S}A^*$. In particular, this construction does not depend in an essential way on these differential operators being of square zero, however if one of them squares to zero then so must the other. Since one operator is the derivation $m$ associated to $D$ and the other is just $D^*$, we could obtain in this way a simple, conceptual and non-computational proof of . Indeed, the $L_\infty$ algebra obtained in this way is merely a gauge transformation of the abelian $L_\infty$ algebra with differential $D$.
Commutative BV-infinity algebras
--------------------------------
First we recall the definition of a dg BV algebra.
A dg BV algebra is a unital commutative algebra $A$ with odd operators $d,\Delta$ with the following properties:
- $d^2 = \Delta^2 = d\Delta + \Delta d = 0$
- $d(1) = \Delta(1) = 0$
- The operator $d$ is a first order differential operator and the operator $\Delta$ is a second order differential operator.
Note that the first two conditions are equivalent to saying that $D = d + h\Delta\in \operatorname{End}(A)[[h]]$ satisfies $D^2 = 0 $ and $D(1)=0$.
We make the following definition, which is a slight modification of the original definition proposed by [@kravchenko2000:deformationsofbvalgebras].
A *(commutative) ${\mathrm{BV}_\infty}$ algebra* is a unital commutative algebra $A$ with odd operators $D_0, D_1, D_2,\dots$ with the following properties:
- $D = D_0 + hD_1 + h^2D_2 + \dots\in \operatorname{End}(A)[[h]]$ satisfies $D^2=0$ and $D(1)=0$.
- Each $D_i$ is an $(i+1)$–th order differential operator.
One of the simplest examples of a BV algebra is the Chevalley–Eilenberg *homological* complex of a Lie algebra. Let $V$ be a Lie algebra and denote by $\operatorname{CE}_\bullet(V)$ the complex for which $\operatorname{CE}_n(V)=\Lambda^n(V)$ and the differential $\Delta$ is given by the standard formula, cf. for example [@loday1992:cyclichombook Section 10.1]. Then $\Delta$ is not a derivation of the wedge product on $\operatorname{CE}_\bullet(V)$, but rather an operator of order two. Furthermore, if $V$ itself has a differential then $\operatorname{CE}_\bullet(V)$ becomes a dg BV algebra.
Now let $V$ be an $L_\infty$ algebra and $(\widehat{S}\Pi V^*,m)$ be its representing complete cdga. In this case $\operatorname{CE}_\bullet(V)$ is the symmetric (but not completed) algebra $S\Pi V$ and the differential $\Delta$ is simply the dual to $m$ under the duality isomorphism $(\widehat{S}\Pi V^*)^*\cong S\Pi V$. Note that if $m=\sum_{i=1}^\infty m_i$ then $\Delta=\sum_{i=1}^\infty\Delta_i$ where $\Delta_i$ is the operator dual to $m_i$. Note that the operator $m_i$ is a sum of operators which are compositions of a derivation of order zero followed by an operator of multiplication with a monomial of order $i$. It follows that its dual $\Delta_i$ is a differential operator of order $i$ and since $\Delta^2=0$ we conclude that $(\operatorname{CE}_\bullet(V),\Delta_1, \Delta_2, \dots)$ is a ${\mathrm{BV}_\infty}$ algebra.
The following generalisation of the degeneration property for dg BV algebras will be a crucial property for us.
A ${\mathrm{BV}_\infty}$ algebra is said to have the *degeneration property* if for every $N$ the homology of $A[h]/(h^N)$ with respect to the differential induced by $D$ is a free ${\mathbf{k}}[h]/(h^N)$–module.
The degeneration property clearly depends only on the properties of the operator $D = D_0 + hD_1 + h^2D_2 + \dots\in \operatorname{End}(A)[[h]]$, but not on the commutative algebra structure of $A$. It is easy to see that the degeneration property is equivalent to the collapsing at the $E_1$ term of the spectral sequence associated with the filtration of $A[[h]]$ by the powers of $h$. It is further equivalent to the existence of so-called Hodge to de Rham degeneration data on the multicomplex $(A,D_0,D_1,\dots)$, see [@dotsenkoshadrinvallette:derhamhomotopyfrob] concerning this terminology. Yet another way to formulate the degeneration property is to require that the differential $D$ is a *trivial* formal deformation of the differential $D_0$.
Given a ${\mathrm{BV}_\infty}$ algebra $A$, then by there is an associated $L_\infty$ structure $ m_1 = D , m_2 , m_3 , \dots$ on $\Pi A[[h]]$ arising from the operator $D$. By this $L_\infty$ algebra is homotopy abelian.
However, since $m_n|_{h=0}=0$ for all $n\geq 2$, this $L_\infty$ structure is not quite the correct generalisation of the Lie algebra associated to a dg BV algebra. We instead wish to consider the $L_\infty$ structure $D, m_2/h, m_3/h^2, \dots$. At first glance this may not appear to define an $L_\infty$ structure on $\Pi A[[h]]$, however it turns out to do so.
\[prop:linftyrescale\] Let $A$ be a ${\mathrm{BV}_\infty}$ algebra and let $m_1, m_2, m_3, \dots$ be the associated $L_\infty$ structure on $\Pi A[[h]]$. Then the sequence of maps $m_1, m_2/h, m_3/h^2, \dots$ also defines an $L_\infty$ structure of $\Pi A[[h]]$.
We first need to check that each of the maps $m_n/h^n$ do indeed give maps $A[[h]]\to A[[h]]$ (as opposed to just maps $A((h))\to A((h))$) and secondly that they do indeed give an $L_\infty$ structure.
Since each $D_i$ is an $(i+1)$–th order differential operator the operation $m_n$ is given by a formula involving only the operators $h^{i-1}D_{i-1}$ for $i \geq n$. But this means that the expression for $m_n$ has a factor of $h^{n-1}$ and so $m_n/h^{n-1}$ is a map $A[[h]]\to A[[h]]$.
To see that this does indeed define an $L_\infty$ structure observe that it is obtained by conjugating the derivation $m= m_1^* + m_2^* + \dots$ with the automorphism $\widehat{S}A((h))^* \to \widehat{S}A((h))^*$ given by setting $a \mapsto h a$ on $A$, extending $h$–linearly and then extending to an automorphism of $\widehat{S}A((h))^*$.
\[rem:genericfibre\] The two $L_\infty$ structures on $\Pi A[[h]]$ in of course give rise to $L_\infty$ structures on $\Pi A((h))$ and the proof of shows that these $L_\infty$ structures are isomorphic. However, it does not follow that this is necessarily the case for the two $L_\infty$ structures on $\Pi A[[h]]$. More precisely, we can view these $L_\infty$ structures as formal deformations of two different $L_\infty$ structures on $\Pi A$. The first deformation is always trivial, by , whereas the second is trivial, as we shall see, in the presence of the degeneration condition.
\[def:linfbv\] Let $A$ be a ${\mathrm{BV}_\infty}$ algebra. We denote by $\mathfrak{g}[[h]] = \Pi A[[h]]$ the $L_\infty$ algebra with $L_\infty$ structure given by $m_1 = D, m_2/h, m_3/h^2, \dots$. We denote by just $\mathfrak{g} = \Pi A$ the $L_\infty$ algebra obtained by setting $h=0$. In particular the differential is then just $m_1 = D_0$.
The $L_\infty$ structure on $\mathfrak{g} = \Pi A$ is given explicitly by the following formulae for the maps $m_n{\colon}A\to A$. $$m_n(a_1,\dots, a_n) = [[[\dots [D_{n-1},a_1]\dots],a_{n-1}],a_n](1)$$
The above explicit formulae imply that the $L_\infty$ operations $m_n$ are derivations in each variable (or multiderivations).
The following theorem, which is the central result of this section, will be our main tool. It says that $L_\infty$ algebras arising in this way from ${\mathrm{BV}_\infty}$ algebras with the degeneration property are homotopy abelian. This is an $L_\infty$ generalisation of the theorem proved in [@katzarkovkontsevichpantev2008:hodgetheoretic; @terilla2008:smoothnessbv], which says that dg Lie algebras arising from dg BV algebras with the degeneration property are homotopy abelian.
Let $A$ be a ${\mathrm{BV}_\infty}$ algebra with the degeneration property. Then the associated $L_\infty$ algebra $\mathfrak{g}=\Pi A$ is homotopy abelian.
First note that for any ${\mathrm{BV}_\infty}$ algebra it is the case that $\mathfrak{g}((h)) = \mathfrak{g}[[h]][h^{-1}]$ is homotopy abelian, since by $\mathfrak{g}((h))$ is isomorphic to the $L_\infty$ structure on $\Pi A((h))$ associated to the operator $D$, regarded as an operator on $A((h))$. But by this is homotopy abelian.
The degeneration property means that there exists a ${\mathbf{k}}[[h]]$–linear deformation retract of the chain complex $\mathfrak{g}[[h]]$ onto its homology. Therefore the $L_\infty$ operations of the minimal model of $\mathfrak{g}[[h]]$ are also ${\mathbf{k}}[[h]]$–linear and tensoring with ${\mathbf{k}}((h))$ over ${\mathbf{k}}[[h]]$ gives the $L_\infty$ minimal model of $\mathfrak{g}((h))$. However, since $\mathfrak{g}((h))$ is homotopy abelian, all the $L_\infty$ operations of this minimal model are zero, and so the same is true for the operations of the minimal model of $\mathfrak{g}[[h]]$ and hence, setting $h=0$, also for the operations of the minimal model of $\mathfrak{g}$. Therefore $\mathfrak{g}$ is homotopy abelian.
BV-infinity structure on the de Rham algebra
============================================
Let $M$ be a (super)manifold of dimension $d$ with space of functions $C^\infty(M)$. Denote by ${\mathfrak A}(M)$ the space of global sections of the super vector bundle $\bigwedge^\bullet TM$, in other words ${\mathfrak A}(M) = \bigwedge^\bullet_{C^\infty(M)}{\mathrm{Der}}(C^{\infty}(M))$. This is the space of multivector fields on $M$.
Recall that $\mathfrak{A}(M)$ has the structure of a Gerstenhaber algebra, with the commutative product given by the wedge product and the antibracket given by the Schouten bracket, which is the unique way of extending the Lie derivative to make $\mathfrak{A}(M)$ into a Gerstenhaber algebra. More precisely in terms of the Lie bracket on vector fields we have $$[ v_1\wedge\dots \wedge v_n, w_1\wedge\dots \wedge v_m ] = \sum_{i,j}(-1)^{i+j}[v_i,w_j]\wedge v_1\wedge\dots\wedge \hat v_i \wedge \dots\wedge v_n \wedge w_1 \wedge \dots \wedge \hat w_j \wedge \dots \wedge v_m$$ for vector fields $v_i$, $w_j$ and for a function $f$ and vector field $v$ $$[v,f] = v(f).$$ Recall that a Poisson structure on $M$ is a bivector field on $M$ whose Schouten square is zero. Considering general multivector fields, we obtain the notion of a generalised (or higher) Poisson structure, cf. [@voronovkhudaverdian2008:higerpoissonbrackets].
A generalised Poisson structure on $M$ is an even element $P\in\mathfrak A(M)$ for which $[P,P]=0$.
Any generalised Poisson structure $P$ can be written as $P=P_{-1}+P_0+\dots$ where $P_i\in\bigwedge^{i+1}_{C^\infty(M)}{\mathrm{Der}}(C^{\infty}(M))$. For simplicity (i.e. in order to exclude considering curved $L_\infty$ algebras) we make the blanket assumption that $P_{-1}=0$. Furthermore, if $M$ is a purely even manifold of dimension $d$, then $P_0$ must likewise be zero and $P=P_1+\dots+P_d$.
For a generalised Poisson structure $P=\sum_nP_n$ as above let $P(h)\in \mathfrak{A}(M)[[h]]$ be defined by the formula: $$P(h)=\sum_nP_nh^n$$ Then clearly we have $[P(h),P(h)]=0$ in $\mathfrak{A}(M)[[h]]$.
Note that the $n$–vector field $P_n$ acts by the Lie derivative on the de Rham algebra $\Omega(M)$. Recall that the operator of the Lie derivative along a multivector field $Q$ is defined as $L_Q=[i_Q,d]$ where $i_Q$ is the operation of the interior derivative and $d$ is the de Rham differential. Since for two multivector fields $Q_1$ and $Q_2$ we have $i_{Q_1\wedge Q_2}=\pm i_{Q_1}\circ i_{Q_2}$ and since $d$ is a derivation of $\Omega(M)$, we conclude that $L_{P_n}$ is a differential operator of order $n$ on $\Omega(M)$. Let $L_{P}(h)=L_{P(h)}$; it is an operator on $\Omega(M)[[h]]$. The identity $[P(h),P(h)]=0$ implies $L_P(h)\circ L_P(h)=0$.
The sequence of operators $L_{P_n}$ on $\Omega(M)$ determines the structure of a ${\mathrm{BV}_\infty}$ algebra structure on $\Omega(M)$, which will be referred to as the de Rham–Koszul ${\mathrm{BV}_\infty}$ algebra of $M$. The $L_\infty$ algebra on $\Omega(M)$ associated to this ${\mathrm{BV}_\infty}$ structure according to will be called the de Rham–Koszul $L_\infty$ algebra of $M$.
The de Rham–Koszul $L_\infty$ structure on $\Omega(M)$ defined above was introduced in [@voronovkhudaverdian2008:higerpoissonbrackets] and the corresponding $L_\infty$ operations were called ‘higher Koszul brackets’ there. Recall that $\Omega(M)[[h]]$ possesses another $L_\infty$ structure, which is a trivial deformation of the homotopy abelian structure on $\Omega(M)$, cf. . The latter structure was considered in [@bruce:higherpoission].
\[thm:generalisedPoisson\] Let $M$ be a manifold with a generalised Poisson structure; then the ${\mathrm{BV}_\infty}$ algebra $\Omega(M)$ satisfies the degeneration property.
For simplicity we assume that $M$ is a purely even manifold, although the arguments carry over with obvious modifications to the supergraded case. Let $\Omega$ be the graded sheaf of differential forms on $M$; thus for an open set $U\in M$ the group $\Omega(U)$ is the de Rham algebra on $U$. For a generalised Poisson structure $P= P_1 + P_2 + \dots$ on $M$ the operators $L_{P_i}$ are determined locally and can be viewed as endomorphisms of the sheaf $\Omega$. Adjoining the formal variable $h$ we can, therefore, consider a dg sheaf $\Omega[[h]]$ with the differential $d + h P_1+\dots$ as well as its truncated version $\Omega[h]/h^N$ for $N=1,2,\dots$; here $d$ is the de Rham differential. Since the sheaf $\Omega$ is fine, the degeneration property for the ${\mathrm{BV}_\infty}$ algebra $\Omega(M)$ is equivalent to the statement that the hypercohomology of $\Omega[h]/h^N$ is a free ${\mathbb R}[h]/h^N$–module for any $N=1,2,\dots$.
Let $U$ be a contractible open set in $M$. Then, filtering the dg space $(\Omega(U)[h]/h^N, d+hL_{P_1}+\dots+h^{N-1}L_{P_{N-1}})$ by the powers of $h$ and using the fact that the de Rham differential $d$ is acyclic, we conclude that it is quasi-isomorphic to ${\mathbb R}[h]/h^N$ concentrated in degree zero. It follows that the dg sheaf $\Omega$ is quasi-isomorphic to the constant sheaf ${\mathbb R}[h]/h^N$. Therefore its hypercohomology is isomorphic to $H(M)[h]/h^N$ and is free over ${\mathbb R}[h]/h^N$.
In the case of an ordinary Poisson structure a statement equivalent to was proved in [@dotsenkoshadrinvallette:derhamhomotopyfrob]. The following corollary generalises the corresponding results of [@sharygintalalaev2008:formalitypoisson] and [@fiorenzamanetti2012:koszulbrackets] formulated for ordinary Poisson manifolds.
The de Rham–Koszul $L_\infty$ algebra of $M$ is homotopy abelian. In particular, the higher Koszul brackets on $H(M)$ vanish.
Note that $\Omega(M)$ can be viewed as a double complex with $\Omega^{p,q}=\Omega^{p-q}$ and with two commuting differentials $d$ and $L_P$. Then we have the following corollary which was obtained in the case of ordinary Poisson manifold in [@fernandezetal1998:canonicalspectralsequence].
The spectral sequence of the double complex $\Omega(M)$ collapses at the $E_1$ term $E_1=H(\Omega(M),d)$.
[^1]: Commutative algebras are always assumed to be associative.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Riccardo Campana , Enrico Massaro , Teresa Mineo , Giancarlo Cusumano'
date: 'Received ..., accepted ...'
title: |
The multicomponent model of the Crab Pulsar\
at energies above 25 GeV
---
Introduction
============
Phase resolved energy distributions of pulsars’ signals are important to obtain information on the radiation processes and geometry of the emission regions in the magnetosphere. At $\gamma$-ray energies, at variance with other energy bands, the three brightest sources (Vela, Crab and Geminga) show remarkably similar pulse structures with two main peaks at a large phase separation ranging from 0.4 to 0.5. This pattern is confirmed by the very recent discovery of pulsed emission from PSR J2021+3651 (Halpern et al. 2008), that has a peak separation of 0.47. The double peak structure of the Crab Pulsar (PSR B0531+21) is well observed across the entire electromagnetic spectrum, but the intensity ratio between the two peaks changes with energy. In particular, in the X and soft $\gamma$-ray ranges the emission of the second peak (P2) becomes higher than the first one (P1), and a relevant emission from the region between the two peaks (interpeak or bridge, Ip) increases like P2. This behaviour is observed up to a few MeV, where the pulse shape turns almost sharply to be similar, although not equal, to the optical light curve.
On the basis of a large collection of data, covering the frequency interval from the optical to the GeV band, we proposed a model (Massaro et al. 2006a, hereafter MCCM) able to describe the spectral and phase distributions by means of a double two-component model. The energy spectra of these components are not described by a simple power law, because of the continuous spectral steepening towards high energies. We found that a very satisfactory model is a parabolic law in a double-logarithmic plot, corresponding to a log-normal spectral distribution.
A useful test to verify the goodness of this multicomponent model, and in particular the existence of the two high energy components, is the study of the pulse shape at energies higher than a few GeV, where P2 is expected to be again the dominant feature as in hard X/soft $\gamma$ rays. Some hints in this direction are given by the EGRET pulse profile (Thompson 2004), but statistics above 5 GeV are so poor that no firm conclusion can be obtained. Very recently, the MAGIC telescope (e.g. Lorenz 2004) has detected pulsed emission from the Crab (Aliu et al. 2008), at energies above 25 GeV and with 6.4 standard deviations significance, showing the very well established two peak profile, with similar amplitudes of the two peaks.
In this Note we compare the MAGIC data to the predictions of the MCCM model, and use them to constrain the high energy components. The model is synthetically described in Sect. 2, the extension to MAGIC results is presented in Sect. 3 and some hypotheses about the origin of these components on the basis of some recent works on the high energy emission from young pulsars are discussed in Sect. 4.
The multicomponent model of Crab pulsar
========================================
{width="50.00000%"} {width="50.00000%"}
The two-component model: optical to hard X-rays
-----------------------------------------------
As early presented in Massaro et al. (2000), the spectral and phase changes of Crab X-ray pulse shape are well reproduced by two components. The first component, called $C_O$, is assumed to have the same pulsed profile observed at optical frequencies, while the second component, $C_X$, is described by an analytical model to reproduce the observed pulse profiles. The latter component dominates at the interpeak (Ip) and second peak (P2) phase regions. This choice can be justified also from theorical models like that proposed by Eastlund et al. (1997), who considered the synchrotron emission from marginally clamped electrons in a shell at the boundary of the closed magnetosphere and calculated pulse profile very similar to $C_O$.
By means of spectral fits of BeppoSAX data, we obtained that the $C_O$ and $C_X$ components were described by a log-parabolic spectral law, $$F(E) = KE^{-(a+b~\mathrm{Log}E)}$$ where $E$ is the energy in keV, $K$ is the photon flux at 1 keV and the parameter $b$ describes the “curvature” of the log-parabola. Best fit estimate of the parameters allowed to derive the peak energies of the spectral energy distributions (SED), $E_p = 10^{(2-a)/2b}$, that were found at 12 keV and 178 keV for $C_O$ and $C_X$, respectively, and $b$ was found the same for both components and equal to 0.16.
The model at $\gamma$-ray energies
----------------------------------
Observations performed by the COMPTEL and EGRET experiments onboard the *Compton Gamma Ray Observatory* (CGRO; Kuiper et al. 2001; Thompson 2004) provided light curves above 10 MeV of a good statistical quality, which show that the pulse shape returns to be similar to that of $C_O$, i.e. the optical/soft X-ray one, although some minor differences are present, for instance in the shape of P2. At energies higher than $\sim$500 MeV there is some indication that the emission from Ip and P2 appears to increase, in analogy with the X-ray band. Left panel of Figure \[figure1\] shows the energy evolution of the P2/P1 flux ratio, computed in the phase intervals defined by Kuiper et al. (2001) that are reported in the caption.
To model the $\gamma$-ray emission and the change of the pulse shape, MCCM assumed that there are two more, high-energy spectral components, $C_{O\gamma}$ and $C_{X\gamma}$, with the same pulse shape of the corresponding lower-energy components and with spectral distributions also given by Eq. (1). This extended model has six new parameters, i.e. the peak energies, curvatures and normalizations of the $C_{O\gamma}$ and $C_{X\gamma}$ components, that should be determined by data fitting. Given the statistical quality of CGRO data, the resulting estimates are much more uncertain than in the X-rays. Therefore, we assumed that the curvature parameters were equal to the $C_O$ and $C_X$ ones ($b = 0.16$), and adjusted the normalizations and peak energies to reproduce the observed total (phase-averaged) spectrum. Peak energies of $C_{O\gamma}$ and $C_{X\gamma}$ were found to lie around 300 MeV and 2 GeV, respectively. To be consistent with the upper limits to the TeV pulsed emission (e.g. Lessard et al. 2000) we added also an exponential cutoff to both $C_{O\gamma}$ and $C_{X\gamma}$, at the energy $E_c=15$ GeV. This value was not compelled by observational or physical reasons, but was simply a guess to take into account the upper limits. The MCCM model was then able to reproduce both the broadband energy spectrum of the total pulse (Figure \[figure1\], right panel), and the spectra of the P1, Ip and P2 phase intervals. In the left panel of Figure \[figure1\] we plotted also the fitted and extrapolated P2/P1 ratio for two values of the $C_{O\gamma}$ cut-off energy, i.e. 11 and 15 GeV. This ratio depends on the normalizations and shapes of the $C_O$ and $C_X$ components in the proper phase ranges, and this is not immediately apparent in the phase averaged spectrum of Figure \[figure1\] (see MCCM for details).
The pulse shape and spectrum in the GeV band
============================================
The pulse shape observed above 25 GeV by the MAGIC telescope (Aliu et al. 2008) is shown in Figure \[magic\], where it is normalized to the counts of the first peak. Despite the statistical quality of data, is well apparent that P2 has an amplitude comparable and possibly higher than P1, in agreement with the trend barely apparent from the EGRET data (Thompson 2004).
In Figure \[magic\] it is also plotted the pulse shape obtained from our model, with the same parameters’ values used in MCCM, for an energy of 25 GeV. The cut-off energy (15 GeV) is the same for both the $C_{O\gamma}$ and $C_{X\gamma}$ components. The amplitude of the peaks is very well reconstructed; the pulse shape in the MAGIC band predicted by our model has small variations with energy, with the height of P2 that varies by about 15% between 25 and 60 GeV.
The spectral distribution in the MAGIC range is still poorly determined, mainly because of the uncertainties in its energy scale. Aliu et al. (2008) combined the very high energy MAGIC data with the COMPTEL and EGRET points and evaluated that a possibile exponential cut-off can be at $E_{c} = 17.7 \pm 2.8 \pm 5.0$ GeV, thus compatible to the value assumed in the MCCM model. However, the CGRO-MAGIC data are compatible with both an exponential or a super-exponential shape, with a preference for the former. We reported the flux estimate in our SED model (Figure \[figure1\], right panel) and found that it is in a substantial agreement with the extrapolation which was introduced to take properly into account the EGRET data. Finally, note that in the MCCM model the flux above $\sim$10 GeV is mainly due to the $C_{X\gamma}$ component (Figure 2). This is not in contrast with the evidence of P1 in the pulse profile, because its flux is concentrated in a quite narrow phase interval, and $C_{X\gamma}$ has however a wide pedestal extending below P1 (Figure \[magic\]).
The good matching of the predicted pulse shape with the MAGIC one implies that the cut-off energy should be similar for $C_{O\gamma}$ and $C_{X\gamma}$, because a lower cut-off for $C_{O\gamma}$ would imply an higher P2/P1 ratio (as shown in Fig. \[figure1\]) and therefore a higher amplitude for the second peak.
One of the assumptions of the MCCM model was that the phase distributions of the various components do not change with the energy. It was essentially motivated by the need to have a low number of parameters and supported by the satisfactory agreement of the computed pulse profiles with data. However, in Massaro et al. (2000) it was already shown that the introduction of a mild energy dependence of two parameters of the analytical description of $C_X$ would improve the pulse profile modelling, particularly in the Ip region. Our extrapolation at 25 GeV predicts a higher Ip than that given by the MAGIC data, especially in the leading wing of P2. We expect a number of detected events in the Ip phase interval, in excess to the off-pulse level, around 2600, whereas the measured excess is $\sim$900 counts, but still compatible with the off-pulse emission, due to the large uncertainty. Note that the S/N ratio does not allow to reach any firm conclusion: there is, for example, a dip at phase $0.3$ (another one is at phase $-0.15$), well below the mean off-pulse level, that reduces the Ip content. The assumption that the shape of $C_{X\gamma}$ is equal to that of $C_{X}$ could not be exactly verified at such different energies. They could be different, with a shallower profile for the high-energy counterpart, as already suggested in MCCM upon considering the Ip spectrum.
Discussion
==========
On the basis of a large collection of data, from the optical band to $\gamma$-ray energies, we developed a model (MCCM) that describes with a good accuracy the spectral and phase distributions. This model has essentially an heuristic value, being aimed to establish a consistent scenario for the development of a realistic physical model. The recent detection of the pulsed emission from Crab at energies above 25 GeV by the MAGIC collaboration is a relevant result that completes our knowledge on the spectral properties of this important source. We have shown here that these results are in a good agreement with the expectations from the MCCM model and confirm its validity.
It would be interesting to investigate theoretically the physical plausibility for the presence of two couples of emission components producing the observed spectra and phase distributions. A possibility to be further developed arise from the similar curvature parameter for the X-ray and $\gamma$-ray components, that suggests a synchrotron-self-Compton (SSC) mechanism for the emission (Morini 1983, Cheng & Wei 1995). A stream of relativistic electrons moving along the magnetic field lines would undergo a Compton upscattering on the synchrotron photon field. A different location and azimuthal distribution in the magnetosphere would explain the different pulse shape for the $C_{O}$ and $C_{X}$ components. An alternative hypothesis is that the $\gamma$-ray components are photons emitted by primary electrons via the curvature radiation mechanism, and the softer ones are the synchrotron emission from secondary pairs. However, in this case one would expect a more pronounced curvature of the log-parabolic spectra for the latter components. In fact, if the electron spectrum is a log-parabola with a curvature parameter $r$, in the $\delta$-approximation the synchrotron curvature $b$ would be equal to $r/4$ (Massaro et al., 2006b), while for curvature radiation this parameter would be $r/9$, because of the dependence on the electrons’ Lorentz factor as $\gamma^{3}$ instead of $\gamma^{2}$ for the synchrotron process. About the log-parabolic (or log-normal) energy distribution of electrons, we recall that it can originate by stochastic acceleration processes or when the acceleration probability is energy-dependent.
Several models have been proposed to explain the high-energy emission from pulsars, both in the outer gap (Cheng et al., 1986, 2000) and in the polar/slot gap (Muslimov & Harding 2003, 2004) frameworks. A common feature of these models is that the broadband emission comes from components originating from different physical processes.
Takata & Chang (2007) developed a 3D outer gap model based on the 2D analytical solution of the accelerating field and particle motion by Takata et al. (2004, 2006) and Hirotani (2006). In this model the X-ray emission of P1 is due to two separate components, with curved, roughly log-parabolic, spectra and originating by synchrotron emission of secondary pairs in different regions of the outer gap, i.e. below and beyond the null-charge surface. Their explanation of the $\gamma$-ray spectrum invokes other two components due to Inverse Compton scattering of secondary pairs and curvature radiation from primary electrons. Their modelling of P2 and Ip, unlike the MCCM model, has different weights for these two components. The increase of P2 and Ip with respect to P1 in this model is due to the fact that the emission comes from regions in the lower magnetosphere, where the high magnetic field produces harder spectra.
More recently, Harding et al. (2008) performed a 3D simulation of the emission from the Crab pulsar, assuming a Slot Gap accelerator, in which the emission is due to two distribution of particles, primary electrons and electron-positron pairs. The X-ray emission is due to the synchrotron radiation from the pairs, while the $\gamma$-rays are curvature radiation and the synchrotron radiation resulting from the resonant cyclotron absorption of radio beam photons by the primary particles.
Future observations, in particular with the Fermi Gamma-ray Space Telescope, will be very useful to confirm the spectral cut-off observed by MAGIC, to improve the estimate of $E_c$, and to bridge the gap between the 100–500 MeV EGRET lightcurves and the $\ge$25 GeV MAGIC one. It will be possible to follow in detail the evolution of the P2/P1 flux ratio and to verify with a much higher statistics whether the Ip emission is actually related to that of P2.
Another future interesting test will come from phase-resolved X-ray polarimetry, where MCCM expect that the hard X-rays polarization properties of P2 should become increasingly similar to those of Ip, because of the higher contribution of the $C_{X}$ component.
Aliu, E., Anderhub, H., Antonelli, L. A. et al. (The MAGIC Collaboration), 2008, Science, 322, 1221 \[arXiv:0809.2998\] Cheng, K. S., Ho, C., Ruderman M. 1986, ApJ, 300, 500 Cheng, K. S., Wei, D. M. 1995, ApJ, 448, 281 Cheng, K. S., Ruderman, M., Zhang, L. 2000, ApJ, 537, 964 Eastlund, B. J., Miller, B., Curtis Michel, F. 1997, ApJ, 483, 857 Halpern, J. P., Camilo, F., Giuliani, A. et al., 2008, ApJ, in press \[arXiv:0810.0008\] Harding, A. K., Stern, J. V., Dyks, J., Frackowiak, M. 2008, ApJ, 680, 1378 Hirotani, K. 2006, ApJ, 652, 1475 Kuiper, L., Hermsen, W., Cusumano, G. et al. 2001, A&A, 378, 918 Lessard, R. W., Bond, I. H., Bradbury, S. M. et al. 2000, ApJ, 531, 942 Lorenz, E. 2004, New Astron. Rev., 48, 339 Massaro, E., Cusumano, G., Litterio, M., Mineo, T. 2000, A&A, 361, 695 Massaro, E., Campana, R., Cusumano, G., Mineo, T. 2006a, A&A, 459, 859 Massaro, E., Tramacere, A., Perri, M., Giommi, P., Tosti, G. 2006b, A&A, 448, 861 Morini, M. 1983, MNRAS, 202, 495 Muslimov, A. G., Harding, A. K. 2003, ApJ, 588, 430 Muslimov, A. G., Harding, A. K. 2004, ApJ, 606, 1143 Takata, J., Shibata, S., Hirotani, K. 2004, MNRAS, 354, 1120 Takata, J., Shibata, S., Hirotani, K., Chang, H. K. 2006, MNRAS, 366, 1310 Takata, J. & Chang, H. K. 2007, ApJ, 670, 677 Thompson, D. J. 2004, Proc. Cosmic Gamma-ray Sources, K. S. Cheng & G. E. Romero eds., ASSL, Kluwer Acad Publ., vol 304, p. 149 \[astro-ph/0312272\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A fully self-consistent model of the neutron star inner crust based upon models of the nucleonic equation of state at zero temperature is constructed. The results nearly match those of previous calculations of the inner crust given the same input equation of state. The extent to which the uncertainties in the symmetry energy, the compressibility, and the equation of state of low-density neutron matter affect the composition of the crust are examined. The composition and pressure of the crust is sensitive to the description of low-density neutron matter and the nuclear symmetry energy, and the latter dependence is non-monotonic, giving larger nuclei for moderate symmetry energies and smaller nuclei for more extreme symmetry energies. Future nuclear experiments may help constrain the crust and future astrophysical observations may constrain the nuclear physics input.'
author:
- 'Andrew W. Steiner'
bibliography:
- 'paper.bib'
title: 'The Neutron Star Inner Crust: Nuclear Physics Input'
---
Introduction
============
The inner crust of a cold neutron star can be defined as the region between the density where neutrons drip out of nuclei (about 4 $\times
10^{11}$ g/cm$^{3}$) and the density for the transition to homogeneous nucleonic matter at about half of the nuclear saturation density. This region is sensitive to the nuclear physics input because the nature of the crust is determined by the structure of neutron-rich nuclei and the energetics of the surrounding dripped neutrons. In this work, the dependence of the description of the neutron-rich nuclei and the dripped neutrons on the equation of state (EOS) of homogenous nucleonic matter is examined.
The inner crust is of broad interest because a large variety of astrophysical observations are dependent on and sensitive to the properties of the neutron star crust. One recent motivation is the suggestion that the giant flares in Soft Gamma-Ray repeaters trigger seismic events in the neutron star crust and are sensitive to the shear modulus of the crust crust [@Thompson95; @Strohmayer06; @Samuelsson07]. The shear modulus, in turn, is sensitive to the composition of the neutron star crust and the relative magnitude of the proton and neutron numbers of the nuclei in the inner crust. The moment of inertia of rotating neutron stars is also sensitive to the inner crust and depends on the transition density between the crust and the core [@Link99]. Neutrino and photon opacities are also sensitive to the properties of the nuclei in the inner crust. For example, neutrino-nucleus scattering, which scales like $A^2$, is the most important neutrino process during the lepton-trapped phase of a Type II supernova (see Ref. [@Burrows06] for a recent review). Finally, the cooling and evolution of neutron star crusts depends on the both the size of the crust [@Lattimer94] and by its transport properties [@Potekhin97; @Brown00], which are both related to the composition. These astrophysical connections motivate the study of the magnitude of the uncertainty of the properties of the inner crust which come from present uncertainties in the nuclear physics inputs.
In this article, several models of the neutron star inner crust are constructed systematically using inputs from with the current experimental information while allowing the range of uncertainty allowed due to the uncertainty in the EOS of homogeneous matter. Of particular importance, is that the symmetry energy is varied in [*both*]{} the description of the nuclei and the description of the neutron matter at the same time. The composition depends on the symmetry energy, but is nearly independent of the compressibility. This means that astrophysical observations which are connected to properties in the crust can constrain the nuclear symmetry energy.
The inner crust is quite sensitive to the EOS of neutron matter at sub-saturation densities. At sufficiently low densities, neutron matter is somewhat well understood because three-body interactions are small, and the two-body neutron-neutron interaction is strongly constrained by the experimentally measured neutron-neutron scattering phase shifts [@Carlson03] (see also the review in Ref. [@Heiselberg00]). Many of the currently available EOSs, however, do not respect this understanding of low-density neutron matter because they are fit to the properties of nuclei which are more sensitive to matter near saturation densities. The neutron matter EOSs used in this work are designed to have a realistic behavior below the saturation density, within the precision required for the description of the crust.
The Mass Models
===============
While microscopically-based models of the nuclei are of great interest because they can disentagle important effects which are not easily treated in a classical approach, a microscopic approach can also make it more difficult to understand the physical principles which guide the nature of the inner neutron star crust. In addition, it is not clear that a classical approach is significantly less effective at estimating the magnitude of uncertainties originating in the nuclear physics input (it may even be [*more*]{} effective). In any case, since the purpose is only to estimate the uncertainties from the nuclear physics input to the EOS, a liquid-drop model quite similar to that described in Refs. [@Baym71b; @Lattimer85] is used. More microscopic models for the crust have been developed (see the pioneering work of Ref. [@Negele83] and recent efforts in Refs. [@Baldo07; @Newton07]) and it is expected that these results on the sensitivity to the EOS of homogeneous nucleonic matter will apply to some extent in these models as well.
The liquid-drop model for this work consists of a bulk energy contribution which is determined from the EOS of homogenous nucleonic matter together with surface and Coulomb contributions. This will be compared to the finite-range droplet model described in Ref. [@Myers69] and used in Ref. [@Moller95].
The binding energy per baryon of a nucleus with proton number $Z$ and atomic number $A$ is given by $$\begin{aligned}
B(Z,A)/A &=& B_{\mathrm{bulk}}(n_n,n_p)/A +
\sigma {\cal B}(n_n,n_p)
\left(\frac{36 \pi}{n^2 A}\right)^{1/3} \nonumber \\
&& + {\cal C}~\varepsilon_{\mathrm{Coulomb}}/n
\label{eq:massform}\end{aligned}$$ where $n_n$ and $n_p$ are the average neutron and proton densities inside the nucleus with the given $Z$ and $A$. The binding energy of bulk matter, $B_{\mathrm{bulk}}$ (about $-$16 MeV in isospin symmetric matter) is given by $$B_{\mathrm{bulk}}= \frac{A}{n}
\left[\varepsilon(n_n,n_p) - n_n m_n - n_p m_p\right]$$ where $m_n$ and $m_p$ are the neutron and proton masses (which is taken to be 939 MeV), $n=n_n+n_p$ is the average baryon number density in the nucleus, and $\varepsilon(n_n,n_p)$ is the energy density of homogeneous matter evaluated at the given neutron and proton density. The expression for $\varepsilon$ may be given by any EOS of homogeneous matter and several different models are employed. Note that the energy of the dripped neutrons which is added later will always be determined with the same EOS as is used to describe the bulk part of the nuclear energy.
The average baryon density will be determined from $$n = n_n+n_p = n_0 + n_1 I^2
\label{eq:n}$$ where $I = 1 - 2 Z/A$. The parameter $n_0$ is analogous to the saturation density of nuclear matter and is expected to be near 0.16 fm$^{-3}$. The parameter $n_1$ subsumes (in a very schematic way) two effects: the decrease in the saturation density with the isospin asymmetry and the increase in the saturation density due to the Coulomb interaction. These effects are both explicitly present in the finite range droplet model (see Eq. 49 of Ref. [@Moller95]). The decrease in the saturation density with isospin asymmetry is typically larger and thus $n_1$ is always negative in these models.
The individual average neutron and proton number densities are given by $$\begin{aligned}
n_n & = & n (1 + \delta) / 2 \nonumber \\
n_p & = & n (1 - \delta) / 2
\label{eq:np}\end{aligned}$$ and the density asymmetry $\delta= 1 - 2 n_p / (n_n + n_p) $ is given by $\delta = \zeta I $ where $\zeta$ is a constant parameter of the model. Neutron and proton radii (“squared-off” radii, not root-mean-square radii) are given simply by $4 \pi n_n R_n^3 = 3N$, and $4 \pi n_p R_p^3 = 3Z$. The presence of a neutron skin is determined from $\zeta$. If $\zeta$ is unity, then all nuclei have no neutron skin ($R_n=R_p$), while if $\zeta$ is less than unity, then all nuclei with $N>Z$ will have a neutron skin ($R_n>R_p$).
The surface energy contribution is proportional to the surface tension $\sigma$, $A^{2/3}$ (the surface energy scales as $A^{2/3}$ so that the surface energy per baryon scales like $A^{-1/3}$ as in Eq. \[eq:massform\]), and a unitless function ${\cal B}$. Typically this latter function is quadratic in the isospin symmetry $${\cal B}(n_n,n_p) = 1 - {\sigma}_{\delta} \delta^2$$ where $\sigma_{\delta}$ is a positive parameter representing the surface symmetry energy. This is essentially the approach taken in Ref. [@Myers69]. For a neutron star inner crust model, this can be modified to ensure that the surface energy vanishes in the limit $\delta \rightarrow 1$ as it must. One possible approach (and the one used here) is that from Ref. [@Lattimer85] $${\cal B}(n_n,n_p) = \frac{16 + b}{\left[1/x^3 + b + 1/(1-x)^3 \right]}$$ where $x=n_p/n$ and $b$ is a simple function of the parameter $\sigma_{\delta}$ and is related through $\sigma_{\delta}=96 \sigma
/(b+16)$. This is an approximate scheme for taking into account the isospin properties of the surface energy which may suffice for the present purpose, but note the more the more detailed discussion in Ref. [@Steiner05]. In particular there is still an unresolved ambiguity associated with how the surface energy is handled as discussed in this reference. The slope of the correlation between the surface symmetry energy, $\sigma_{\delta}$ and the symmetry energy at the saturation density depends on the mass formula used. This model, like all other present models for the neutron star crust, effectively chooses a particular slope for this correlation.
The Coulomb energy density of a 3-dimensional droplet of protons can be written [@Baym71b; @Ravenhall83] (modulo an overall factor of $\chi$, the volume fraction of matter present in nuclei, which is included later), $$\varepsilon_{\mathrm{Coulomb}}
= \frac{2 \pi}{5} n_p^2 e^2 R_p^2 \left( 2-3 \chi^{1/3}+\chi \right)$$ where $e^2$ is the usual Coulomb coupling $\sim \hbar c/137$. In the final term in parenthesis, the first term corresponds to the standard Coulomb contribution, the second term corresponds to the “lattice contribution” [@Baym71] in the Wigner-Seitz approximation, and the last term to a further finite-size correction relevant at higher densities when $\chi$ is comparable to unity. Note that this last term is quite important near the crust-core transition and tends to delay the transition to nuclear matter to higher densities. The Coulomb contribution is multiplied by a parameter ${\cal C}$ to take into account the fact that the proton density does not fall off sharply at a finite radius and this surface diffusiveness maybe dependent on the input symmetry energy. This parameter will always be nearly unity. As noted in Ref. [@Chamel07], the Wigner-Seitz approximation fails when describing the low-temperature transport properties, but will suffice for describing the composition and the equation of state as done here.
In summary, there are six free parameters in this model (outside of the input equation of state of bulk nuclear matter, which is a kind of parameter in itself) are the surface tension in MeV/fm$^{2}$, $\sigma$, the surface symmetry energy $\sigma_{\delta}$, the correction factor to the Coulomb energy, $ {\cal C} $, the asymmetry parameter $\zeta$, and the central density parameters, $n_0$ and $n_1$ which are expressed in units of fm$^{-3}$. These six parameters will be fit to experimental masses for each input EOS.
The Equations of State
======================
The EOS from Ref. [@Akmal98] (APR) is used, which was obtained from variational chain summation calculations of the equation of state using a realistic nucleon-nucleon interaction. Also, a “typical” relativstic field-theoretical model is utilized (review in Ref. [@Serot89]), NL4 [@Nerlo-Pomorska04] which was fit to nuclei. In order to compare with the model of Ref. [@Douchin01] the Skyrme [@Skyrme59] model SLy4 [@Chabanat95] is used, and in order to compare with the model from Ref. [@Lattimer06] the Skyrme model SkM$^{*}$ [@Bartel82] is used.
APR is expected to be particularly good for neutron matter at low densities, because it is directly computed from an interaction which reproduces the two-body nucleon-nucleon phase shifts. The model SLy4 also has a good neutron matter EOS because it was fit to both nuclei and low-density neutron matter. The NL4 and SkM$^{*}$ models were only fit to nuclei and low-density neutron matter are less constrained. SkM$^{*}$ happens to have a neutron matter EOS which is somewhat closer to APR than NL4. Like the SLy4 interaction, relativistic models are also able to reproduce, at some level, the more accurate low-density neutron matter EOS found in APR and SLy4, as was demonstrated by the RAPR model in Ref. [@Steiner05] and the FSUGold model [@Todd-Rutel05; @Piekarewicz07].
In order to examine the importance of having an accurate EOS for low-density neutron matter, the low-density neutron matter EOS of NL4 is modified and compared to the original. Three new models are constructed. The first model, NL4Q, is a modification of NL4 which treats the symmetry energy at low densities to be exactly quadratic. This approximation is quite good at lower densities, and the NL4 crust is nearly indistinguishable from the NL4Q crust. The other models, NL4QN and NL4QN2, are versions of NL4Q which reproduces the neutron matter EOS of APR at densities below a specified density, $$E_{\mathrm{neut}}^{\mathrm{NL4QN}}=E_{\mathrm{neut}}^{\mathrm{APR}}+
\frac{ E_{\mathrm{neut}}^{\mathrm{NL4Q}}-
E_{\mathrm{neut}}^{\mathrm{APR}} }
{ 1+ e^{(n_t-n)/\nu} }`$$ where $n_t$ is 0.08 (0.04) fm$^{-3}$ and $\nu$ is 0.0105 (0.016) fm$^{-3}$ for model NL4QN (NL4QN2). Both NL4QN and NL4QN2 have pressures and neutron chemical potentials which monotonically increase with density. The results below show that these two models give signficantly different results for the inner crust.
Finally, several schematic models of the EOS are constructed so that the effect of the compressibility and the symmetry energy can be examined. The schematic EOS for neutron matter is $$\begin{aligned}
E_{\mathrm{neut}}/A &=& \left[1-0.6~k_{F,n}^{0.4} + \eta_1 \left(\frac{n}{n_0}\right) +
\eta_2 \left(\frac{n}{n_0}\right)^2 \right] \nonumber \\
&& \times \frac{k_{F,n}^5}{10 \pi^2 m_n}
\label{eq:neut}\end{aligned}$$ where $k_F = (3 \pi^2 n)^{1/3}$ is the neutron Fermi momentum and the last term is just the free Fermi gas energy density. The first two terms inside the square brackets are designed to reproduce the expectation from equations of state at low densities obtained from two-body potentials which reproduce the experimental phase-shift data on neutron-neutron scattering [@Carlson03; @Gezerlis07]. It is expected that the interacting neutron matter EOS is about half the free Fermi gas energy at $k_{F,n}=0.5$ fm$^{-1}$. The term proportional to $k_{F,n}^{0.4}$, qualitatively reproduces this low-density behavior and the other parameters $\eta_1$ and $\eta_2$ can be adjusted for densities near the saturation density where the EOS is more uncertain.
Model $\eta_1$ (MeV) $\eta_2$ (MeV) $E_{\mathrm{sym}} (MeV)$ $\gamma$
-------------- ---------------- ---------------- -------------------------- ----------
Sch -0.307 0.481 31 0.9
SchS28 -0.487 0.578 28 0.9
SchS34 -0.127 0.385 34 0.9
Sch$\gamma$1 -0.0308 0.198 31 0.6
Sch$\gamma$2 -0.793 0.979 31 1.1
: The values of $\eta_1$ and $\eta_2$ (c.f. Eq. \[eq:neut\]), parameters controlling the neutron matter equation of state in the schematic models and the corresponding values of the symmetry energy at saturation density and the exponent $\gamma$. In the schematic models, $\eta_0$ is fixed at 0.5.
It is useful to connect this description with the more traditional description neutron matter in terms of a symmetry energy with the form $ E_{\mathrm{sym}} = A (n/n_0)^{2/3} + B (n/n_0)^{\gamma}$. For an effective mass of about 0.7 $M$, $A$ is about 17 MeV, and then $B$ and $\gamma$ dictate the magnitude of the symmetry energy at the saturation density and the density dependence of the symmetry energy, respectively. The base model schematic model, “Sch”, has values of $\eta_1$ and $\eta_2$ appropriate for $B=14$ MeV (a symmetry energy of 31 MeV) and $\gamma=0.9$. These values are given in Table I, as well as the values of $\eta_1$ and $\eta_2$ for the schematic models whose symmetry energy is different from the baseline model. The variation of the value of the symmetry energy at saturation density between 28 and 34 MeV is consistent with the observation that most modern equations of state fall within this range. When the symmetry energy is taken to be a pure power law (the A=0 limit), the limit of the variation of $\gamma$ ($0.6<\gamma<1.1$) is inferred from the experimental information from intermediate-energy heavy-ion collisions [@Tsang04; @Chen05]. The models Sch$\gamma$1 and Sch$\gamma$2 are constructed by fitting a symmetry energy with the given exponent with A=0 to express the expected range. The extraction of values of $\gamma$ from heavy-ion collisions is non-trivial, and there may be systematic uncertainties that are not yet understood. These uncertainties would mean that the range of variation presented here is overly conservative, and that the true range might be larger.
The schematic equation of state for nuclear matter is $$\begin{aligned}
E_{\mathrm{nuc}}/A &=& M + B + \frac{K}{18 n_0^2} \left(n - n_0\right)^2
\nonumber \\
&& + \frac{K^{\prime}}{162 n_0^3} \left(n - n_0\right)^3\end{aligned}$$ where $M$ is the nucleon mass, $B$ is the binding energy, $n_0$ is the saturation density, $K$ is the compressibility and $K^{\prime}$ is the “skewness”. Isospin asymmetric matter is computed assuming that the symmetry energy is exactly quadratic in the isospin asymmetry $\delta$ (this approximation may fail at high density, see Ref. [@Steiner06]). Note that varying the compressibility in this model is not precisely equal to varying the quantity which might be obtained from giant resonances, as the latter are only sensitive to the equation of state in the neighborhood of saturation density whereas the compressibility is applied to nuclear matter at all densities below the saturation density, thus the variation in the compressibility is a bit larger than that recently suggested in Refs. [@Agrawal03; @Colo04; @Garg06]. The baseline schmatic model “Sch”, has a binding energy of $-$16 MeV, a saturation density of 0.16 fm$^{-3}$, a symmetry energy of 31 MeV, a compressibility of 230 MeV. In all of the models, the skewness parameter is fixed by ensuring that the energy per baryon of nuclear matter vanishes at zero density, as it ought. In addition to variations of the symmetry energy as discussed above, Two models “SchK210” and “SchK250”, are constructed to be the same as the baseline model, except that they have different compressibilities.
![A survey of the equations of state used in this work. Plotted in each panel are the energy per baryon of nuclear matter (lower set of curves) and neutron matter (upper set of curves) as a function of baryon density, $n_{B}$. The upper left panel shows the models APR, SLy4, and SkM$^{*}$, the upper right panel shows the model NL4Q (based on NL4) and shows how neutron matter was modified to match APR at low densities, and the lower panels show how the symmetry energy was modified in the schematic equation of state.[]{data-label="fig:eos"}](eosplot2.eps)
A survey of some of the equations of state is given in Fig. \[fig:eos\]. The upper right panel shows how neutron matter in the NL4QN and NL4QN2 models interpolates between APR at low densities and the normal NL4Q model at higher densities, while leaving the nuclear matter EOS unmodified.
The Mass Fits
=============
The liquid drop model is fit to the experimental nuclear masses from Ref. [@Audi03] using the fitting formula $$\Delta_{\mathrm{RMS}}= \left[ \frac{1}{N} \sum_{i=1}^{N}
\left(M^{exp}_i-M^{th}_i\right)^2 \right]^{1/2}$$ where $N$ is the number of nuclei, and $M^{exp}$ and $M^{th}$ are the experimental and theoretical values of the mass excess. Ref. [@Moller95] points out that this fitting formula can be improved and that it overestimates the actual model error, but these considerations will not be important at the level of the results presented here. The fitting results are given in Table II. The fitting results for NL4QN and NL4QN2 are not given because they were found to be nearly equal to those from NL4Q.
\[tab:fit\]
Model $\zeta$ $\sigma$ $\sigma_{\delta}$ ${\cal C}$ $n_1$ $n_0$ $\Delta_{\mathrm{RMS}}$
-------------- --------- ---------- ------------------- ------------ ----------- ------- -------------------------
APR 0.886 1.19 1.72 0.885 $-$0.128 0.181 2.61
SkM$^{*}$ 0.888 1.14 1.16 0.899 $-$0.0612 0.17 2.61
SLy4 0.885 1.19 1.57 0.882 $-$0.11 0.181 2.6
NL4Q 0.89 1.15 2.67 0.915 $-$0.234 0.169 2.66
Sch 0.897 1.19 1.67 0.903 $-$0.12 0.176 2.7
SchK210 0.897 1.19 1.69 0.907 $-$0.117 0.174 2.72
SchK250 0.897 1.19 1.64 0.9 $-$0.12 0.176 2.69
SchS28 0.891 1.2 1.18 0.892 $-$0.0488 0.179 2.66
SchS34 0.9 1.19 2.44 0.909 $-$0.206 0.175 2.68
Sch$\gamma$1 0.891 1.2 1.49 0.884 $-$0.103 0.182 2.63
Sch$\gamma$2 0.911 1.11 0.978 0.954 $-$0.028 0.154 2.75
: The nuclear mass fits corresponding to the models described in the text. The values of $\sigma$ are given in MeV/fm$^{2}$ and $n_0$ and $n_1$ are given in fm$^{-3}$.
The mass fit is performed by minimizing $\Delta_{\mathrm{RMS}}$ for all the experimentally measured mass excesses from Ref. [@Audi03]. For each nucleus, this involves computing the neutron and proton densities using Eqs. \[eq:n\] and \[eq:np\], computing the bulk energy from the EOS of homoegenous matter at these densities, then inserting this bulk energy into the nuclear mass formula to compute the mass excess using Eq. \[eq:massform\].
As expected, there is a correlation between the surface symmetry energy and the symmetry energy as shown by the increase in $\sigma_{\delta}$ when going from model SchS28 to model SchS34. This is also the reason why NL4 gives a larger value of $\sigma_{\delta}$ than the other models. The value of $\sigma_{\delta}$ is nearly unchanged by modifying $\gamma$, which changes the density dependence of the symmetry. The values of $n_0$ in Table \[tab:fit\] are not quite equal to the saturation density for homogeneous nuclear matter, and this can be attributed to finite-size effects not captured in Eq. \[eq:n\]. The other parameters are nearly unchanged between models except for $n_1$ which is also sensitive to the symmetry energy, as well as the Coulomb interaction.
The Crusts
==========
In order to determine the composition and properties of the crust, the energy at a fixed density as a function of the proton number and atomic number of nuclei, and the number density of dripped neutrons, $n_{n,drip}$, is minimized. The energy of matter in the neutron star crust is given by $$\begin{aligned}
\varepsilon(Z,A,n_{n,\mathrm{drip}}) &=& (n_n+n_p) \chi B(Z,A)/A +
\nonumber \\
&& (1-\chi) {\varepsilon}_{\mathrm{drip}}(n_{n,\mathrm{drip}})+
{\varepsilon}_{\mathrm{el}} (n_e)\end{aligned}$$ This energy is minimized over the three parameters $Z, A$, and $n_{n,\mathrm{drip}}$ at each density. The volume fraction of matter inside nuclei, $\chi$, is determined from the relation $$n_B = \chi (n_n+n_p) + n_{n,\mathrm{drip}} (1-\chi)$$
![A comparison of the composition of the neutron star crusts in models NL4Q, NL4QN, and NL4QN2, designed to demonstrate the sensitivity of the composition to the equation of state of low-density neutron matter. The proton nuber, Z, the atomic number A (left axis) and the number density of the dripped neutrons $n_{\mathrm{n,drip}}$ are displayed (right axis). []{data-label="fig:ld"}](ldplot.eps)
The inner crust implied by the models NL4Q, NL4QN and NL4QN2 are compared in Fig. \[fig:ld\]. While the actual number density of dripped neutrons is not strongly modified by modifying the neutron matter EOS, the nuclear size is modified by 50 % or more. The larger energy cost of creating neutron matter in with a more realistic neutron matter EOS is reflected in moving neutrons into nuclei so as not to pay the energy cost. Because the largest difference is between the models NL4Q and NL4QN2, it is clear that most of the dependence on the low-density neutron matter EOS lies at densities below 0.04 fm$^{-3}$.
![A comparison the composition given the two different mass models. The curve labeled LDM (solid lines) is obtained from Eq. \[eq:massform\], and the curve labeled FRDM (dashed lines) is obtained from Ref. [@Moller95]. The left axis is for $A$ and $Z$ and the right axis is for $n_{\mathrm{n,drip}}$.[]{data-label="fig:mass"}](massplot.eps)
In order to show the effect due to changing the mass model, Fig. \[fig:mass\] shows the results for the APR EOS with the two different models, the liquid drop model and the FRDM. The results are qualitatively the same but quantitatively different. The form of the mass model remains a significant uncertainty in the nature of the neutron star crust, and is comparable to the other uncertainty obtained from the symmetry energy as described below.
![A comparison of the present work (solid lines) with the results of Ref. [@Douchin01] (dashed lines) using the same input equation of state, SLy4. The left axis is for $A$ and $Z$ and the right axis is for $n_{\mathrm{n,drip}}$.[]{data-label="fig:dh"}](dhplot.eps)
Comparisons of the present model to that of Refs. [@Douchin01; @Lattimer06] are given in Figs. \[fig:dh\] and \[fig:jl\]. The same input EOS for homogeneous matter as the original reference is used in both cases. The results agree qualitatively with the aforementioned works. The remaining differences lie within the nuclear mass formula used, and they are within the range of variation which is suggested by Fig. \[fig:mass\]. This model (like all other models of the neutron star crust presently available) cannot precisely predict the composition of the inner crust. Nevertheless, it is qualitatively correct and is thus useful for estimating the uncertainties due to the input EOS of homogeneous matter.
![A comparison of the present work (solid lines) with the results from Ref. [@Lattimer06] (labelled “JML” and plotted with dashed lines) using the same input equation of state, Sk$M^{*}$. The sharp dropoff in JML at high-densities is due to the transition to homogeneous nucleonic matter. The slightly jagged nature of the JML results is due to the naive interpolation employed in this work and is not necessarily present in the original table. The left axis is for $A$ and $Z$ and the right axis is for $n_{\mathrm{n,drip}}$. []{data-label="fig:jl"}](jlplot.eps)
To compare the effect of the uncertainty in the symmetry energy, Fig. \[fig:se\] shows the composition for the neutron star crust as a function of density for the schematic equations of state with different symmetry energies. The naive expectation is that a stronger symmetry energy tends to encourage nuclei to become more isospin-symmetric. This is coupled, however, with the fact that an increased symmetry energy will also raise the energy cost for the dripped neutrons. These two effects together could force larger, more symmetric nuclei, but this also affects the Coulomb and surface energy contributions. The variation of the composition with the value of the symmetry energy is not so clear, as the baseline model predicts larger nuclei than either models with smaller or larger values of the symmetry energy.
![A comparison of the composition of the crust given different symmetry energies. The bold solid line is the baseline model, the dashed and dotted lines show the variation in $\gamma$, and the dashed-dotted and thin solid line give the variation in the magnitude of the symmetry energy at the saturation density.[]{data-label="fig:se"}](seplot.eps)
In order to disentangle this result, more detailed results for schematic models with different symmetry energies are given in Fig. \[fig:seden\] at a fixed density of $n_{B}=0.01$ fm$^{-3}$. Beginning with the larger symmetry energy (with a value at saturation of 34 MeV) and proceeding downward, the expected result is obtained: lower symmetry energies allow the system to create more isospin-asymmetric nuclei. At low enough symmetry energies, however, this becomes too costly as the electron contribution to the energy increases (the proton number decreases, but the volume fraction occupied by nuclei increases, thus the electron density must increase). Instead, the system reponds by moving neutrons out of the nuclei, which lowers the electron contribution, even though it increases the contributions from nuclei and the dripped neutrons. This is allowed, in part, because the nuclei are able to maintain a relatively constant energy. They can do this because the surface and Coulomb energy cost is cancelled by the bulk energy gain which results from making nuclei with a larger (in absolute magnitude) bulk binding energy.
![A comparison of the composition of the crust given different values of the symmetry energy at the saturation density at a fixed density of $n_{B}=0.01$ fm$^{-3}$. The top panel gives the total binding energy per baryon, and the separate contributions from dripped neutrons (“n”), electrons (“e”), and nuclei (“Nuc”). The bottom panel shows the neutron and proton number of nuclei as well as the volume fraction, $\chi$.[]{data-label="fig:seden"}](seden2.eps)
Finally, Fig. \[fig:ep\] summarizes the pressure as a function of the baryon density. The upper left panel shows the variation from the different mass formulas, which is larger at lower densities. The upper right panel shows the results for SkM$^{*}$ and SLy4. The lower-left panel shows the variation allowed by the symmetry energy, and variations of up to a factor of two in the pressure are implied by the uncertainty in the symmetry energy. The pressure appears sensitive to the magnitude of the symmetry energy and its dependence on density. Finally, the lower-right panel shows the pressure for the NL4Q-related models. Note that for model NL4Q, the anomalously small EOS of low-density neutron matter underestimates the pressure in the inner crust. All of the models are connected (sometimes discontinuously) to the EOS from Ref. [@Baym71] at low densities thus giving the scatter in the pressure at the lowest densities given in this figure. Particularly interesting is that the pressure at the crust at the higher densities is nearly independent of the mass model (as shown in the upper left panel), which may indicate that astrophysical observables which are sensitive to the pressure of the crust rather than the composition are good probes of the nuclear symmetry energy, assuming that the description of low-density neutron matter is correct.
![A survey of the pressure of the crust as a function of density, scaled by the baryon density, $n_B^{4/3}$. The curves are nearly flat where the adiabatic index is expected to be nearly 4/3 since the baryon density nearly scales with the energy density. []{data-label="fig:ep"}](epplot2.eps)
The transition density to homogeneous nuclear matter is computed by noting the density at which the energy per baryon of nuclear matter becomes smaller than that of the heterogeneous phase. The results for the transition density are given in Table III. The transition densities from this work are given in the second column and third column contains the transition densities for previous works with similar input EOSs for comparison. These transition densities may be underestimates, as the ability of nuclei to deform slightly will decrease the energy of the heterogeneous phase and thus increase the transition density. This will be addressed in further work. Note that, as in the composition discussed above, the transition densities depend non-trivially on the symmetry energy.
\[tab:nt\]
Model $n_{t}$ fm$^{-3}$
-------------- ------------------- ---------------------
APR 0.0522
SkM$^{*}$ 0.0434 0.045 [@Lattimer06]
SLy4 0.0669 0.076 [@Douchin01]
NL4Q 0.0344
NL4QN 0.0409
NL4QN2 0.0333
Sch 0.0584
SchK210 0.0585
SchK250 0.0591
SchS28 0.0368
SchS34 0.0416
Sch$\gamma$1 0.0676
Sch$\gamma$2 0.0641
: Caption here.
Conclusions
===========
The composition of the neutron star crust is still partially unknown, due to uncertainties in the nuclear mass formula and the equation of state. The composition (and to a lesser extent, the overall pressure) is quite sensitive to the equation of state of low-density neutron matter, and the nuclear symmetry energy, both its magnitude and its density dependence. The dependence of the composition on the symmetry energy is not monotonic, as models with moderate symmetry energies can have larger nuclei than models with lower or higher symmetry energies. To the extent to which neutron stars depend on the composition, this means that it is important to explore the full range of variation in the crust allowed by the present knowledge of the input nuclear physics, while ensuring that the EOS is constrained by what is already known about the EOS of low-density neutron matter. Nuclear experiments will continue to provide better constraints on the symmetry energy, including from the PREX experiment [@Michaels00; @Horowitz01b] to measure the neutron skin thickness of lead at Jefferson Lab and from intermediate-energy heavy-ion collisions as has been done in Ref. [@Tsang04].
It remains to be seen if these results persist in the more microscopic models which include pairing, corrections beyond the Wigner-Seitz approximation, long-range correlations, and better treatments of the nuclear structure.
Acknowledgements
================
The author would like to thank Ed Brown, Joe Carlson, Alex Gezerlis, Jim Lattimer, Bill Lynch, Sanjay Reddy, Sergio Souza, and Anna Watts for useful discussions related to this work. This work is supported by Joint Institute for Nuclear Astrophysics at MSU under NSF-PFC grant PHY 02-16783.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Most countries are suffering severely from the ongoing covid-19 pandemic despite various levels of preventive measures. A common question is if and when a country or region will reach herd immunity $h$. The classical herd immunity level $h_C$ is defined as $h_C=1-1/R_0$, where $R_0$ is the basic reproduction number, for covid-19 estimated to lie somewhere in the range 2.2-3.5 depending on country and region. It is shown here that the disease-induced herd immunity level $h_D$, after an outbreak has taken place in a country/region with a set of preventive measures put in place, is actually substantially smaller than $h_C$. As an illustration we show that if $R_0=2.5$ in an age-structured community with mixing rates fitted to social activity studies, and also categorizing individuals into three categories: low active, average active and high active, and where preventive measures affect all mixing rates proportionally, then the disease-induced herd immunity level is $h_D=43\%$ rather than $h_C=1-1/2.5=60\%$. Consequently, a lower fraction infected is required for herd immunity to appear. The underlying reason is that when immunity is induced by disease spreading, the proportion infected in groups with high contact rates is greater than that in groups with low contact rates. Consequently, disease-induced immunity is stronger than when immunity is uniformly distributed in the community as in the classical herd immunity level.'
author:
- 'Tom Britton$^{1}$, Frank Ball$^{2}$ and Pieter Trapman$^{3}$'
bibliography:
- 'herdref.bib'
title: 'The disease-induced herd immunity level for Covid-19 is substantially lower than the classical herd immunity level'
---
plus1pt minus1pt
Introduction {#sec-Intro .unnumbered}
============
Covid-19 is spreading in most countries of the world and many different preventive measures are put in place to reduce transmission. Some countries aim for suppression by means of a total lockdown, and others for mitigation by slowing the spread using certain preventive measures in combination with protection of the vulnerable [@FMG20]. An important question for both policies is when to lift some or all of the restrictions. A highly related question is if and when herd immunity is obtained. Some regions and countries have already reached high estimates for the population immunity level, with 26% infected in metropolitan Stockholm region as per May 1 2020 [@FHM20], while by the end of March [@FMG20] estimates for Italy and Spain as a whole were already around or above 10% and more recent. It is debated if the (classical) herd immunity level $h_C=1-1/R_0$, which for Covid-19 is believed to lie between 50% and 75% since common estimates of $R_0$ for Covid-19 typically lie in the range 2-4 (e.g. [@FMG20]), is at all realistic to achieve without too many case fatalities [@FLN20; @BAB20]
Herd immunity is defined as a level of population immunity such that disease spreading will decline and stop also after all preventive measures have been relaxed. If all preventive measures are relaxed when the immunity level (from people having been infected) is clearly below the herd immunity level, then a second wave of infection will start once restrictions are relaxed.
The classical herd immunity level $h_C$ is defined as $h_C=1-1/R_0$, where $R_0$ is the basic reproduction number defined as the average number of new infections caused by a typical infected individual during the early stage of an outbreak [@DHB13]. This definition originates from vaccination considerations: if a fraction $v$ is vaccinated (with a vaccine giving 100% immunity) and vaccinees are selected uniformly in the community, then the new reproduction number is $R_v=(1-v)R_0$. From this it is clear that the critical vaccination coverage $v_c=1-1/R_0$; if at least this fraction is vaccinated, the community has reached herd immunity, as $R_v\le 1$, and no outbreak can take place.
An epidemic model for an age and activity-level structured population {#an-epidemic-model-for-an-age-and-activity-level-structured-population .unnumbered}
=====================================================================
The simplest of all epidemic models is to assume a homogeneously mixing population in which all individuals are equally susceptible, and equally infectious if they become infected. We let $\lambda $ denote the average number of infectious contacts an individual who becomes infected has before recovering and becoming immune (or dying). An infectious contact is defined as one close enough to infect the other individual if this individual is susceptible (contacts with already infected individuals have no effect). For this simple model the basic reproduction number $R_0$ equals $R_0=\lambda$ [@DHB13].
To this simple model we add two important features known to play an important role in disease spreading (the model is described in full detail in the Supplementary Information, SI). The first is to include age structure by dividing the community into different age cohorts, with inhomogeneous mixing between the different age cohorts. More precisely, we divide the community into 6 age groups and fit contact rates from an empirical study of social contacts [@WTK06]. Consequently, the person-to-person infectious contact rate between two individuals depends on the age groups of both individuals. The average number of infectious contacts an infected in age group $i$ has with individuals in age group $j$ now equals $\lambda a_{ij}\pi_j$, where $a_{ij}$ reflects both how much an $i$-individual has contact with a specific $j$-individual but also typical infectivity of $i$-individuals and susceptibility of $j$-individuals, and $\pi_j$ denotes the population fraction of individuals belonging to age cohort $j$.
The second population structure added categorizes individuals according to their social activity levels. A common way to do this is by means of network models (e.g. [@PV01]). Here we take a simpler approach and assume that individuals can be categorized into three different activity levels: 50% of each age-cohort have normal activity, 25% have low activity corresponding to half as many contacts compared to normal activity, and 25% have high activity corresponding to twice as many contacts as normal activity. By this we mean that, for example, a high-activity individual in age group $i$ on average has $2*a_{ij}\pi_j*0.5*0.25$ infectious contacts with low-activity individuals of type $j$. The factor 2 comes from the infective having high activity, the factor 0.5 from the contacted person having low activity, and the factor 0.25 from low-activity individuals making up 25% of each age cohort. The basic reproduction number $R_0$ for this model is given by the dominant eigenvalue to the (next-generation) matrix $M$ having these elements as its entries. [@DHB13].
The final fraction getting infected in the epidemic, starting with few initial infectives, is given as the unique non-negative solution to a set of equations (the final-size equations) given in the SI. This solution also agrees with the final fraction infected of different types for the corresponding stochastic epidemic model assuming a large population [@BC93]. In order to also say something about the time evolution of the epidemic we assume a classical SEIR epidemic model. More precisely, we assume that individuals who get infected are initially latent for a period with mean 3 days, followed by an infectious period having mean 4 days, thus approximately mimicking the situation for Covid-19 (e.g. [@FMG20]). During the infectious period an individual makes infectious contacts at suitable rates such that the mean number of infectious contacts agree with that of the next-generation matrix $M$.
The epidemic model with preventive measures put in place {#the-epidemic-model-with-preventive-measures-put-in-place .unnumbered}
========================================================
We assume that the basic reproduction number satisfies $R_0=2.5$ (a few other values are also evaluated) and that the epidemic is initiated with a small fraction of infectious individuals on February 15. On March 15, when the fraction infected is still small, preventive measures are implemented such that all averages in the next-generation matrix are scaled by the same factor $\alpha <1$, so the next-generation matrix becomes $\alpha M$. Consequently, the new reproduction number is $\alpha R_0$. These preventive measures are kept until the ongoing epidemic is nearly finished. More precisely, all preventive measures are relaxed thus setting $\alpha$ back to 1 on June 30. If herd immunity is not reached there will then be a second wave, whereas the epidemic continues to drop if herd immunity has been reached.
In the results section we investigate the effect of the preventive measures and for a couple of different scenarios numerically analyse whether or not a given level of preventive measures will yield disease-induced herd immunity. We do this for populations that a) are homogeneous population b) have individuals categorized into different age groups but no activity levels, c) have no age groups but different activity levels, and d) have both age and activity structures.
Results {#results .unnumbered}
=======
For each of the four population structures, we first show the overall disease-induced herd immunity level in Table \[tab\_herd\_levels\]. The level is obtained by assuming that preventive measures having factor $\alpha<1$ are implemented at the start of an epidemic, running the resulting epidemic to its conclusion and then exposing the population to a second epidemic with $\alpha=1$. We find $\alpha_*$, the greatest value of $\alpha$ such that the second epidemic is subcritical; $h_D$ is then given by the fraction of the population that is infected by the first epidemic. This approximated the situation where preventive measures are implemented early and lifted late in an outbreak. Note that $h_D$ is independent of the distributions of the latent and infectious periods.
-------------------------- ------- ------- ------- ------- ------- -------
Population structure $h_D$ $h_C$ $h_D$ $h_C$ $h_D$ $h_C$
\[0.5ex\]Homogeneous 50.0 50.0 60.0 60.0 66.7 66.7
Age structure 46.0 50.0 55.8 60.0 62.5 66.7
Activity structure 37.7 50.0 46.3 60.0 52.5 66.7
Age & Activity structure 34.6 50.0 43.0 60.0 49.1 66.7
\[1ex\]
-------------------------- ------- ------- ------- ------- ------- -------
: Disease-induced herd immunity level $h_D$ and classical herd immunity level $h_C=1-1/R_0$ for different population structures, for $R_0=2.0$, 2.5 and 3.0. Numbers correspond to percentages.
\[tab\_herd\_levels\]
As seen in the table all three structured population have lower disease-induced herd immunity $h_D$ compared to the classical herd immunity $h_C$, which assumes immunity is uniformly distributed among the different types of individual. From the table it is clear that the different activity levels have a greater effect on reducing $h_D$ than age structure.
In Table \[tab\_fin\_sizes\] the final fractions infected in the different age activity groups for $\alpha=\alpha_*$ just barely reaching disease-induced herd immunity are given. This is done for the age and activity group structure and assuming $R_0=2.5$. The overall fraction infected equals $h_D=43.0\%$, in agreement with Table \[tab\_herd\_levels\].
---------------------- -------------- ------------------ ---------------
Age-group Low activity Average activity High activity
\[0.5ex\]0 - 5 years 17.6 32.1 53.9
6 - 12 years 25.8 44.9 69.7
13 - 19 years 31.4 52.9 77.8
20 - 39 years 27.4 47.2 72.1
40 - 59 years 22.8 40.3 64.4
$\ge$ 60 years 14.6 27.0 46.7
\[1ex\]
---------------------- -------------- ------------------ ---------------
: Final outcome fractions infected in different groups assuming $R_0=2.5$ and preventive measures put in place such that $\alpha=\alpha_*$ just barely reaching herd immunity for $R_0=2.5$. Population structure includes both age and activity and fractions infected are given as percentages.
\[tab\_fin\_sizes\]
We also illustrate the time evolution of the epidemic for $R_0=2.5$, assuming both age and activity structure, and starting with a small fraction externally infected in mid-February. For this we show the epidemic over time for four different levels of preventive measures put in place early in the epidemic outbreak (mid-March) and being relaxed once transmission has dropped to low levels (June 30). In Figure \[Fig\_incidence\] the community proportion that is infectious is plotted during the course of the epidemic.
![Plot of the overall fraction infected over time for the age and activity structured community with $R_0=2.5$, for four different preventive levels inserted March 15 (day 30) and lifted June 30 (day 135). The black, red, yellow and purple curves corresponds to no, light, moderate and severe preventive measures, respectively.[]{data-label="Fig_incidence"}](Fig_Density "fig:"){width="100.00000%" height="1\textheight"} -7cm
It is seen that preventive measures reduce the size and delay the time of the peak. On March 15 preventive measures (at four different levels for $\alpha$) are put in place and it is seen that the growth rate is reduced except for the black curve which has no preventive measures ($\alpha=1$). Sanctions are lifted on June 30 putting transmission rates back to their original levels, but only in the curve with highest sanctions is there a clear second outbreak wave, since the remaining curves have reached (close to) herd immunity. The yellow curve finishes below 50% getting infected. The reason it has more than the 43% infected shown in Table \[tab\_herd\_levels\] is that preventive measures were not from the very start and were also lifted before the epidemic was over. An interesting observation is that the purple curve ends up with a *higher* overall fraction infected even though it had *more restrictions* than those of the yellow. The explanation is that this epidemic was further from completion when sanctions were lifted.
Figure \[Fig\_cumulative\] plots the corresponding cumulative fraction infected as a function of time.
![Plot of the cumulative fraction infected over time for the age and activity structured community and $R_0=2.5$, for a four different preventive levels inserted March 15 and lifted June 30. The black curve corresponds to no preventive measures, the red with light preventive measure, the yellow to moderate preventive measures and the purple corresponding to severe preventive measures.[]{data-label="Fig_cumulative"}](Fig_cdf "fig:"){width="100.00000%" height="1\textheight"} -7cm
Observe that the first three curves see no (strong) second wave of outbreak once preventive measures are lifted – it is only the curve corresponding to highest preventive measures that has a severe second wave. when restrictions are lifted. Note that also the yellow curve, having overall fraction infected well below the classical herd immunity level $h_C=60\%$ is protected by herd immunity since no second wave appears. This clearly illustrates that the disease-induced herd immunity level $h_C$ is well below 60% – it is 43% (see Table \[tab\_herd\_levels\]). In the SI we show additional plots for the situation where restrictions are lifted continuously between June 1 and August 31, and also study how the effective reproduction number evolves as a function of the time when restrictions are lifted.
Discussion {#discussion .unnumbered}
==========
The main conclusion is that the disease-induced herd immunity level may be substantially lower then the classical herd immunity level. Our illustration indicates a reduction from 60% down to 43% (assuming $R_0=2.5$) but this should be interpreted as an illustration, rather than an exact value or even a best estimate.
The current model took age cohorts and social activity levels into account. However, more complex and realistic models have many other types of heterogeneities: for instance increased spreading within households (of different sizes) or within schools and workplaces; and spatial aspects with rural areas having lower contact rates than metropolitan regions. It seems reasonable to assume that most such additional heterogeneities will have the effect of *lowering* the disease-induced immunity level $h_DS$ even further, in that high spreading environments (metropolitan regions, large households, big workplaces, ...) will have a higher fraction infected thus resulting in immunity being concentrated even more on highly active individuals. Some complex models do not categorize individuals into different activity levels, or the related feature in an underlying social network with varying number of acquaintances. As illustrated in our results section, differences in social activity plays a greater role in reducing the disease-induced herd immunity level than inhomogeneous age-group mixing. Thus models not having such features will see smaller difference between $h_D$ and $h_C$. Our choice to have 50% having average activity, 25% having half and 25% having double activity is of course very arbitrary. An important future task is hence to determine how size of differences in social activity within age groups. The more social heterogeneity there is between groups, the bigger difference between $h_D$ and $h_C$.
An assumption of our model is that preventive measures acted proportionally on all contact rates. This may not always hold. For example, most countries have a pronounced ambition to protect elderly (and other risk groups), which does not obey this assumption. Again, we expect the effect would be to *lower* the disease-induced immunity level had this type of preventive measure been considered, because the oldest age group is the one having fewest contacts. For a model having schools and workplaces, it is however not obvious what effect school closure and strong recommendations to work from home would have on the disease-induced herd immunity level.
There are of course other more efficient exit-strategies than to lift all restrictions simultaneously. In fact, most countries are currently employing a gradual lifting of preventive measures. Such slower lifting of preventive measures will avoid seeing the type of overshoot illustrated by the purple curve in Figure \[Fig\_cumulative\], which results in a greater fraction infected than the yellow curve, even though the latter has milder restrictions. The effect of such gradual lifting of restrictions will result in the final fraction infected reaching close to the disease-induced herd immunity level.
Different forms of immunity levels have been discussed previously in the literature although, as far as we know, not when considering early-introduced preventions that are lifted towards the end of an epidemic outbreak. Anderson and May [@AM91] concludes that immunity level may differ between uniformly distributed, disease-induced and optimally-located immunity, and vaccination policies selecting individuals to immunize in an optimal manners have been discussed in many papers, e.g. [@BBL04]. A very recent and independent preprint make similar observations to those in the present paper, but where heterogeneities in terms of e.g. susceptibilities vary continuously [@GAC20]. The correspondning epidemic model is solved numerically similar to our Figure \[Fig\_incidence\], but analytical results for final sizes and $h_C$ are missing.
The present study highlights that the disease-induced herd immunity level $h_D$ is substantially smaller than the classical herd immunity level $h_C$. To try to quantify more precisely the size of this effect remains to be done, and we encourage more work in this area.
**Supplementary information**
Materials and Methods {#materials-and-methods .unnumbered}
=====================
A deterministic SEIR model and the fraction of the population infected {#a-deterministic-seir-model-and-the-fraction-of-the-population-infected .unnumbered}
----------------------------------------------------------------------
In this supplementary information we describe the deterministic SEIR (Susceptible, Exposed, Infectious, Removed) epidemic model in a population partitioned by age and activity level. For reasons of notational convenience we label the types (the combination of age and activity level) from $1$ to $m$, where $m$ is the product of the number of age classes and the number of activity levels. A more detailed exposition as the one presented here can be found in \[1, Sections 5.5 and 6.2\].
We assume that for all $j \in \{1,\cdots,m\}$ the population consists of $n_j$ people of type $j$. We set $n = \sum_{j=1}^m n_j$ and $\pi_j = n_j/n$. We assume that the population is large and closed, in the sense that we do not consider births, deaths (other than possibly the deaths caused by the infectious disease) and migration. Throughout the epidemic $n_i$ is fixed. So, people who die by the infectious disease are still considered part of the population. For $j,k \in \{1,\cdots,m\}$, every given person of type $j$ makes infectious contacts with every given person of type $k$ independently at rate $\alpha a_{jk}/n$. If at the time of such a contact the type $j$ person is infectious and the type $k$ person is susceptibe then the latter becomes latently infected (Exposed). People of the same type may infect each other, so $a_{jj}$ may be strictly positive. Because the definition of an infectious contact includes that the contact leads to transmission of the disease, it is not necessarily the case that $a_{jk}$ is equal to $a_{kj}$. The parameter $\alpha$ is a scaling parameter, used to quantify the impact of control measures in the main paper, without measures $\alpha$ is set equal to 1. Exposed individuals become Infectious at constant rate $\sigma$ and infectious individuals recover or die (are Removed) at constant rate $\mu$. The rates of becoming infectious and removal are assumed to be independent of type. It is straightforward to extend the model to make those rates age or activity level dependent.
In the described multi-type SEIR model, the expected number of people of type $k$ that are infected by an infected person of type $j$ during the early stages of the infection is $n_k \times (\alpha a_{jk}/n) \times (1/\mu) = \pi_k \alpha a_{jk}/\mu$, where $1/\mu$ is the expected duration of an infectious period. The next generation matrix $M$ has for $j,k \in \{1,\cdots,m\}$ as element in the $j$-th row and $m$-th column the quantity $\pi_k \alpha a_{jk}/\mu$. We define the basic reproduction number, $R_0$ as the largest eigenvalue of $M$, which is necessarily real and positive. If $R_0>1$, then a large outbreak is possible with positive probability, while if $R_0 \leq 1$ an outbreak stays small with probability 1.
We set $S_j(t)$ to be the number of people of type $j$ that are susceptible to the disease at time $t$, $E_j(t)$ the number of people of type $j$ that are latently infected, $I_j(t)$ the number of infectious people of type $j$ and $R_j(t)$ the number of removed people of type $j$ ($j \in \{1,\cdots,m\}$). Note that $S_j(t)+E_j(t)+I_j(t)+R_j(t) = n_j = \pi_j n$ for all $t\geq 0$, because the population is closed. Again for $j \in \{1,\cdots,m\}$ we define $s_j(t)=S_j(t)/n_j$, $e_j(t)=E_j(t)/n_j$, $i_j(t)=I_j(t)/n_j$ and $r_j(t)=R_j(t)/n_j$.
Theory on Markov processes \[2, Chapter 11\] (see also \[1, Section 5.5\] for the single type counterpart) gives that for large $n$ the above model can be described well by a system of differential equations (again for $j \in \{1,\cdots,m\}$). $$\begin{array}{rllll}
\dot{s}_j(t) & = & - \frac{1}{n_j} \displaystyle\sum_{k=1}^m \alpha \frac{a_{kj}}{n} S_j(t) I_k(t) &= &
-\displaystyle\sum_{k=1}^m \alpha \pi_k a_{kj} s_j(t) i_k(t),\\
\dot{e}_j(t) & = & \frac{1}{n_j}\left( \displaystyle\sum_{k=1}^m \alpha \frac{a_{kj}}{n} S_j(t) I_k(t) - \sigma E_j(t) \right)&= &
\displaystyle\sum_{k=1}^m \lambda \pi_k a_{kj} s_j(t) i_k(t) -\sigma e_j(t),\\
\dot{i}_j(t) & = & \frac{1}{n_j}\left( \sigma E_j(t) - \mu I_j(t) \right) &= &
\sigma e_j(t)-\mu i_j(t),\\
\dot{r}_j(t) & = & \frac{1}{n_j} \mu I_j(t) &= & \mu i_j(t).
\end{array}$$ To be complete we use for the analysis in the main text that for some $j_* \in \{1,\cdots,m\}$, $s_{j_*}(0)=1-\epsilon$, $e_{j_*}(0)=\epsilon$ and $i_{j_*}(0)=r_{j_*}(0)=0$ while $s_{j}(0)=1$, $e_{j}(0)=i_{j}(0)=r_{j}(0)=0$ for $j \in \{1,\cdots,m\}\setminus j_*$. In the analysis below we do not impose assumptions on the initial conditions.
Because the population is closed, the epidemic will ultimately go extinct and therefore for all $j \in \{1,\cdots,m\}$ we have that $e_j(t) \to 0$ and $i_j(t) \to 0$ as $t \to \infty$. So $s_j(t) + r_j(t) \to 1$ as $t \to \infty$. Furthermore $s_j(t)$ is non-increasing. Therefore $s_j(\infty) = \lim_{t \to \infty} s_j(t)$ exists.
It can be shown in the spirit of \[1, Equation (6.2)\] that for $j \in \{1,\cdots,m\}$, $$\label{finalsize}
\frac{s_j(\infty)}{s_j(0)} = \exp\left[-\lambda \sum_{k=1}^m a_{kj} \pi_k \left(1-r_k(0)-s_k(\infty)\right)/\mu\right].$$ To understand this identity we observe first that $\frac{s_j(\infty)}{s_j(0)}$ is the fraction of initially susceptible people of type $j$ escapes the epidemic. while the exponent in the right hand side can be written as $$\sum_{k=1}^m n \pi_k \left(1-r_k(0)-s_k(\infty)\right) \times \lambda a_{kj}/n \times \frac{1}{\mu}
= \sum_{k=1}^m (n_k-R_k(0)-S_k(\infty)) \times \lambda a_{kj}/n \times \frac{1}{\mu}.$$ In words the summands read as the number of people of type $k$ that were infectious at some moment during the epidemic, times the rate at which a type $k$ person makes infectious contacts with someone of type $j$, times the expected time an infected person is infectious. In other words, the right hand side is the cumulative force of infection during the entire epidemic on a person of type $j$. Standard theory on epidemics gives that minus the natural logarithm of the probability that a given initially susceptible person of type $j$ avoids infection is the cumulative force of infection on the person.
So gives that the fraction of initially susceptible people that are ultimately still susceptible is equal to the probability that a given initially susceptible person avoids infection.
The population matrix {#the-population-matrix .unnumbered}
---------------------
In the main text we analyse an age structured population. Contact intensities between different age groups we took from \[3\]. The age groups are 0-5, 6-12, 13-19, 20-39, 40-59 and 60+. The contact matrix, i.e. the matrix with elements $\{a_{jk};j,k \in\{1,\cdots,6\}\}$ is taken from Table 1 of \[3\]. Note that the expected number of contacts from a person of type $j$ with people of type $k$ is $n_k a_{jk}/n = \pi_k a_{jk}$. Therefore we divide the elements of Table 1 by $\pi_k$ to obtain the contact matrix. We further multiply this matrix by a constant such that the largest eigenvalue is equal to $2.5$,t he value we have chosen for $R_0$. The contact matrix is $$\begin{pmatrix}
2.2257 & 0.4136 & 0.2342 & 0.4539 & 0.2085 & 0.1506\\
0.4139 & 3.6140 & 0.4251 & 0.4587 & 0.2712 & 0.1514\\
0.2342 & 0.4257 & 2.9514 & 0.6682 & 0.4936 & 0.1972\\
0.4539 & 0.4592 & 0.6676 & 0.9958 & 0.6510 & 0.3300\\
0.2088 & 0.2706 & 0.4942 & 0.6508 & 0.8066 & 0.4341\\
0.1507 & 0.1520 & 0.1968 & 0.3303 & 0.4344 & 0.7136
\end{pmatrix}.$$ As explained in the main text we can use this matrix to generate the 18 by 18 contact matrix for the model in which we take both age and activity level into account.
Additional figures {#additional-figures .unnumbered}
------------------
In the main text we studied effects of lifting restrictions of different levels $\alpha$ on June 30 (day 135) going back to the situation of no restrictions corresponding to setting $\alpha$ back to 1. Below are the corresponding plots but where restrictions are relaxed gradually (linearly) between June 1 (day 105) and August 31 (day 195). In Figure \[Fig\_Gradual\_inc\] the we plot the fraction of infectious individuals.
![Plot of the overall fraction infected over time for the age and activity structured community with $R_0=2.5$, for four different preventive levels inserted March 15 (day 30) and lifted gradually between June 1 (day 105) and August 31 (day 195). The black, red, yellow and purple curves corresponds to no, light, moderate and severe preventive measures, respectively.[]{data-label="Fig_Gradual_inc"}](Fig_Grad_dens.pdf "fig:"){width="100.00000%" height="1\textheight"} -7cm
The plot looks quite similar to that of Figure \[Fig\_incidence\] in the main text except that the purple curve with highest restrictions no longer has a pronounced second wave. The reason for this is that now restrictions are lifted slowly and graduella during a 3 month period rather than discretely back to normal from one day to another.
![Plot of the overall fraction infected over time for the age and activity structured community with $R_0=2.5$, for four different preventive levels inserted March 15 (day 30) and lifted gradually between June 1 (day 105) and August 31 (day 195). The black, red, yellow and purple curves corresponds to no, light, moderate and severe preventive measures, respectively.[]{data-label="Fig_Gradual_cdf"}](Fig_Grad_cdf.pdf "fig:"){width="100.00000%" height="1\textheight"} -7cm
In Figure \[Fig\_Gradual\_cdf\] the corresponding plot for cumulative fraction of infected over time is given. Compared to Figure \[Fig\_cumulative\] tha main difference is that the purple curve (severe restrictions) has fewer finally infected since there is no longer as big over-shoot above $h_C=43\%$ caused by a second wave.
Finally, in Figure \[Fig\_Eff\_R\] we consider the situation when preventive measures with level $\alpha$ are implemented 30 days after introduction of the disease and relaxed (so $\alpha$ returns to 1) at time $t>30$. The parameters are again chosen so that $R_0=2.5$ when $alpha=1$. The graphs show the effective $R_0$ (incorporating disease-induced immunity) as a function of time $t$, for four different choices of preventive level $alpha$. Thus all four curves coincide until day 30. The effective $R_0$ with no preventive measures ($\alpha=1$) reaches the critical value of one on about day 57 (mid-April), whilst that for $\alpha=0.8$ does so on about day $68$ (April 23). The stronger preventive measures ($\alpha=0.6$ and $\alpha=8/15$) are such that herd immunity is never reached even if they are retained indefinitely.
![Plot of the effective reproduction number (incorporating disease-induced immunty) if restrictions for different $\alpha$ are put in place Day 30 and relaxed on day $t>30$.[]{data-label="Fig_Eff_R"}](Fig_Eff_R.pdf "fig:"){width="100.00000%" height="1\textheight"} -7cm
Additional References {#additional-references .unnumbered}
=====================
\[1\] H. Andersson and T. Britton, *Stochastic epidemic models and their statistical analysis*. New York: Springer Verlag, 2000.
\[2\] S.N. Ethier and T.G. Kurtz. *Markov processes: characterization and convergence*, vol. 282. John Wiley & Sons, 2009.
\[3\] J. Wallinga, P. Teunis and M. Kretzschmar. Using data on social contacts to estimate age-specific transmission parameters for respiratory-spread infectious agents. *American Journal of Epidemiology*. Vol 164: 936-944, 2006.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Astrodynamical Space Test of Relativity using Optical Devices (ASTROD) is a multi-purpose relativity mission concept. ASTROD’s scientific goals are the measurement of relativistic and solar system parameters to unprecedented precision, and the detection and observation of low-frequency gravitational waves to frequencies down to $5\times10^{-6}$ Hz. To accomplish its goals, ASTROD will employ a constellation of drag-free satellites, aiming for a residual acceleration noise of (0.3-1)$\times$ 10$^{-15}$ m s$^{-2}$ Hz$^{-1/2}$ at 0.1 mHz. Noise sources and strategies for improving present acceleration noise levels are reported.'
address: |
Center for Gravitation and Cosmology, Purple Mountain Observatory,\
Chinese Academy of Sciences, Beijing West Road 2,\
Nanjing 210008, China.\
[email protected]
author:
- 'A. PULIDO PATÓN'
title: Current prospects for ASTROD Inertial Sensor
---
Introduction
============
The classical concept of a drag-free satellite[@Lange] consists of a small proof mass inserted inside a large spacecraft. The larger spacecraft shields external forces, allowing the proof mass to move in free fall. The relative position and orientation of the proof mass with respect to the main spacecraft is sensed along its trajectory. This information is feedback to thrusters on the main spacecraft, which are subsequently fired to maintain the proof mass-spacecraft relative position and orientation. In this way, the coupling of external forces to the proof mass is minimized.
Numerous space missions have already employed drag-free technology. The first example of a “drag-free” mission to test fundamental physics is Gravity Probe B (GP-B).[@GPB] GP-B was launched in April 2004 and has experimentally measured frame-dragging and geodetic effects. By spring of 2007, the results of the data analysis will become public. The GP-B inertial sensor achieved a level of free fall below $2\times10^{-12}$ m s$^{-2}$ Hz$^{-1/2}$ at $5\times10^{-3}$ Hz.
The Laser Interferometer Space Antenna, LISA, is another fundamental physics mission requiring drag-free performance. The LISA mission concept[@LISA] consists of three spacecraft in heliocentric orbits, forming a nearly equilateral triangle formation of side 5 $\times$ 10$^{6}$ km and inclined with respect to the ecliptic by 60$^{\circ}$. LISA will monitor the separation between free falling proof masses, which are shielded within the spacecraft, by using interferometric techniques, to detect and observe gravitational waves. LISA’s aims include studying the role of massive black holes in galaxy evolution, testing relativistic gravity, determining the population of ultra-compact galactic binaries, probing the physics of the early universe, observing supermassive and intermediate black holes mergers and mapping spacetime by observing gravitational captures. The LISA drag-free performance goal is 3 $\times$ 10$^{-15}$ m s$^{-2}$ Hz$^{-1/2}$ at 0.1 mHz. By 2009, the LISA Technology Package (LTP) on board the LISA Pathfinder (LPF) ESA mission, with NASA contributions, aims to demonstrate drag-free performance to a level one order of magnitude lower than the LISA requirement, approximately 3 $\times$ 10$^{-14}$ m s$^{-2}$ Hz$^{-1/2}$ in the frequency bandwidth between 1 mHz and 30 mHz.[@LTP] If LTP is successful, it will achieve the best drag-free performance up to present date.
The Astrodynamical Space Test of Relativity using Optical Devices, ASTROD, is a mission concept[@ASTROD; @Ni] that consists of a constellation of drag-free spacecraft employing laser interferometric techniques with Earth orbiting satellites, to provide high precision measurements of the relativistic parameters {$\gamma$, $\beta$}; improved determination of the orbits of major asteroids; measurement of solar angular momentum via the Lense-Thirring effect and the detection of low-frequency gravitational waves and solar oscillations. ASTROD aims to improve on the LISA drag-free goal by a factor of between 3 and 10, i.e. (0.3-1) $\times$ 10$^{-15}$ m s$^{-2}$ Hz$^{-1/2}$ at 0.1 mHz.[@NiGRG; @NiNPBProc; @NiIJMPD07; @NiLiao; @Pulido1; @Pulido2] It is worth noting that a ten-fold acceleration noise improvement with respect to LISA, would allow ASTROD to explore relativistic gravity to an uncertainty level of 1 ppb.[@NiGRG; @NiNPBProc; @PPBASTROD; @NiGravEmp] A simple version of ASTROD, ASTROD I, has been studied as the first step to ASTROD. ASTROD I concept consists of one spacecraft in a solar orbit, carrying out interferometric ranging and pulse ranging with ground stations.[@NiGRG; @NiNPBProc; @NiIJMPD07] ASTROD I also aims to measure relativistic and solar system parameters and test fundamental laws of spacetime although with less precision than ASTROD.
Other mission concepts to follow on from LISA are BBO[@BBO] (Big Bang Observer) and DECIGO[@DECIGO] (DECihertz Interferometry Gravitational wave Observatory). These missions aim to observe the cosmic gravitational wave background produced by standard inflation, having optimum sensitivity around 0.1 Hz, where white dwarf binaries confusion level is thought to be very low. The drag-free force noise requirement for BBO and DECIGO is approximately a hundredth of the LISA force noise target.
In section 2 we report general ideas that ASTROD could adopt to improve drag-free performance, we summarize acceleration and sensor back action disturbances, emphasizing the most significant low-frequency noise sources. Charging disturbances and discharging schemes are briefly discussed in section 3. Finally in section 4 we discuss the performance and problems associated with inertial sensors for other follow on LISA missions. Gravitational interactions are described in terms of multipole moments and a detailed calculation of capacitances for the case of a three dimensional capacitive sensor/actuator are given in appendix A and B, respectively.
ASTROD Inertial Sensor
======================
ASTROD will face new challenges in drag-free technology compared with LISA. First, ASTROD aims to improve the LISA drag-free performance target by a factor of between 3 and 10 at 0.1 mHz. On the other hand, ASTROD will extend the gravitational wave observational bandwidth to frequencies below 0.1 mHz, which is the lowest frequency on the LISA observational bandwidth, down to 5 $\times$ 10$^{-6}$ Hz. To achieve these goals, there are key problems that need to be evaluated and their solutions optimized.
The initial design concept proposed for LISA was that laser beams from remote spacecraft would directly illuminate the proof masses. Issues like laser beam pointing, actuation forces to correctly orientate the proof mass with respect to laser beam, and cross coupling of proof mass degrees of freedom, make this scheme less atractive. A new scheme has been recently discussed, in which the laser beam from the remote spacecraft illuminates a fiducial point in the inertial sensor or spacecraft, rather than the proof mass. The inertial sensor provides positioning and attitude reference, and is also used to actuate the proof mass along the other degrees of freedom. The proof mass could be monitored by employing heterodyne laser metrology. This scheme, so-called separate interferometry, has been adopted recently by LISA.[@Heinzel] However LISA will not fully exploit all its advantages as it may still use two proof masses per spacecraft as references for two different interferometer arms. ASTROD will employ one proof mass, minimizing disturbances and simplifying control.
Replacing capacitive sensing with optical sensing, for drag-free missions following LISA, has also been widely debated. Optical sensing is more sensitive than electrostatic sensing and it requires almost no coupling between the proof mass and surroundings. Towards that direction many efforts to implement optical sensing have been made in the last few years,[@Acernese; @KXSun; @Speake; @Xu; @KXSunAng] but further laboratory research is needed to develop a space qualified optical sensing scheme. Light pressure could also be used for active control. A more conservative design would combine an optical sensor and capacitive active control.[@Clive] Larger gaps between the proof mass and electrodes could then be used to minimize disturbances.
Ultimately local gravitational gradients are the limiting factor for drag-free performance. In that context the influence of proof mass geometries (spherical, cylindrical, cubic, polyhedral, etc) on sensitivity and its repercussion for the overall inertial sensor design merits further discussion. For monitoring and correcting length changes due to thermal effects and slow relaxations, ASTROD will use an absolute metrology system.[@NiASR]
Acceleration noise sources
--------------------------
To discuss the acceleration noise, we consider a simplified control loop model of a spacecraft and a single proof mass. The acceleration noise is given by,[@Schumaker0; @Schumaker; @Sachie]
$$\label{accnoise}
a_{n}\approx-K
X_{nr}+\frac{F_{str}}{m_{p}}+\left(\frac{F_{ns}+TN_{t}}{M}\right)\frac{K}{\omega^{2}u}$$
where $F_{str}$ are stray forces directly acting on the proof mass of mass $m_{p}$, $F_{ns}$ and $TN_{t}$ are external forces and thruster force noise acting on the outer spacecraft of mass $M$. $X_{nr}$ is the sensor readout sensitivity, $u$ is the control loop gain and $\omega\equiv2\pi f$, where $f$ is the frequency. External forces, thruster noise and readout sensitivity contribute to acceleration noise because of the proof mass-spacecraft coupling $K$. If the proof mass and spacecraft are highly decoupled, then the acceleration noise will be given by stray forces directly acting on the proof mass.
In Table 1 direct acceleration noise sources with the exception of sensor back action disturbances are listed.[@Schumaker] Environmental disturbances can be divided into different groups depending on their origin. There are disturbances caused by impacts. Cosmic rays which penetrate the spacecraft shielding and residual gas can deposit momentum onto the proof mass. There are disturbances of magnetic origin. Proof mass magnetic susceptibility, $\chi$, and residual permanent moment, $M_{r}$, can interact with the residual local and/or the interplanetary magnetic field, $B_{SC}$ and $B_{ip}$, respectively. There are also disturbances associated with charge. Residual charge accrued on the proof mass can interact with the interplanetary magnetic field via Lorentz force. Finally, there are disturbances associated with thermal fluctuations on the spacecraft. These include the radiometer and outgassing effects, thermal radiation pressure and gravitational gradients caused by thermally induced spacecraft distortions.
Several key factors needed to quantify the total acceleration noise, such as sensor readout noise, back action forces, and stiffness terms, differ for different types of sensor. As an example, optical sensing is nearly stiffness free, providing high readout sensitivity with very low back action forces. On the other hand, with the widely used capacitive sensor, high sensitivity is achieved at the expense of increased acceleration noise and stiffness. Also, because of the fact that close metallic surfaces are needed (3-4 mm gaps in the case of LISA), other noise contributions due to, for example, patch effects and dielectric losses, become significant.
In the best scenario, the proof mass will ultimately be coupled to the spacecraft by gravitational gradients. Force gradients of electrostatic origin can be made negligible by implementing large gaps between the proof mass and the surrounding metallic surfaces. An analysis of proof mass geometries and thermal-gravitational modeling of the spacecraft and payload are necessary to account for gravitational disturbances and stiffness terms.
A preliminary quantitative analysis of acceleration noise parameter requirements for ASTROD is given in Refs. and .
Low-frequency acceleration noise sources
----------------------------------------
The LISA observational bandwidth extends from 0.1 mHz to 0.1 Hz. It has been pointed out that gravitational wave observations extended to frequencies below 0.1 mHz, are desirable in the study of certain astrophysical sources like massive black holes (MBH) binaries at high redshift.[@Bender]
The ASTROD free falling proof masses will be separated by distances of 30 to 60 times longer than those of LISA. ASTROD gravitational wave sensitivity curve will therefore be shifted to lower frequencies than the target LISA bandwidth. At low frequencies, spurious forces acting directly on the proof mass are the dominant source of noise. This is the reason why, for a mission like ASTROD, it is particularly important to identify these sources of noise. An extended discussion of low-frequency acceleration noise sources and low-frequency sensitivity curve for gravitational waves for ASTROD is given in Ref. .
Thermal, magnetic and electrostatic effects are sources of low-frequency acceleration noise. Thermal noise arises due to radiometer effect; fluctuating outgassing and thermal radiation pressure assymetries; thermal distortion of the spacecraft and residual gas impacts. The magnitude of these effects for a particular mission are dependent on the mission orbit, which dictates the thermal environment. Suppressing thermal disturbances requires thermal diagnostics,[@Lobo] stable electronics, passive and, in some cases, active thermal isolation, thermally conductive electrodes (for the case of electrostatic sensing/actuation) and high vacuum. A preliminary evaluation of thermal disturbances for ASTROD shows that the outgassing effect, thermal radiation pressure and thermally induced gravity gradients could be at levels of about $f_{og}\approx$ 1.1 $\times$ 10$^{-17}$ m s$^{-2}$ Hz$^{-1/2}$, $f_{tp}\approx$ 8 $\times$ 10$^{-18}$ m s$^{-2}$ Hz$^{-1/2}$ and $f_{gg}\approx$ 5.4 $\times$ 10$^{-17}$ m s$^{-2}$ Hz$^{-1/2}$, respectively, at 0.1 mHz. A vacuum pressure of the order of 10$^{-6}$ Pa, and a thermal isolation factor, $\xi_{TS}$, of about 150, were assumed for these estimates. Below 0.1 mHz, solar irradiance fluctuations become the main cause of temperature fluctuations. Solar irradiance fluctuations become more acute when approaching solar rotational period, which is of the order of 25 days.[@Bender] This aspect will be a crucial factor for the thermal diagnostic and thermal isolation system design for ASTROD.
Magnetic noise at low frequencies is caused by interplanetary magnetic field fluctuations, local gradients and magnetic field fluctuations, eddy current damping, magnetic impurities, Lorentz forces due to proof mass residual charge, etc. Suppression of magnetic disturbances requires further reduction and shielding of permanent magnets in the payload, improving magnetic shielding, adopting a magnetic clean wiring (i.e, solar array rewiring, etc) and power system. The most significant magnetic low-frequency noise source is due to the interaction of proof mass magnetic susceptibility with the interplanetary field fluctuations, $f_{m2}$ (see Table 1). Assuming parameter values given in Ref. , $f_{m2}\approx2\times 10^{-17}$ m s$^{-2}$ Hz$^{-1/2}$ at 0.1 mHz. Disturbance $f_{m2}$ increases at low frequencies as $f^{-2/3}$.
Other relevant noise sources at low frequencies are caused by electrostatic effects, i.e., due to voltage noise, charge fluctuations, DC voltages, dielectric losses, actuation noise, etc. When employing capacitive sensing, the strategies to suppress these noise sources are: increasing the separation between the proof mass and the electrodes, active compensation of DC voltages,[@Weber] avoidance of DC voltages applied to drag-free degrees of freedom, high quality surface coatings to minimize dielectric losses, high stability power supplies and continuous discharging of the proof mass. Electrostatic noise is caused by the capacitive sensor/actuator.
An obvious way for suppression of these noise sources is to replace capacitive sensing by optical sensing. Nevertheless, other concerns will arise if optical sensing is to be used. Thermal distortion of optical components and changes in refractive index with temperature modify optical paths. Because these noise sources are due to thermal fluctuations, they will also be of significance at low frequencies. New ideas addressing these problems, such as using all reflective optics, by using gratings, has been extensively discussed in the literature.[@KXSun] Diffractive gratings have also been considered to enhance angular sensitivity, compared with standard angular sensors based on laser reflection.[@KXSunAng]
Gravitational modelling
-----------------------
Drag-free performance will ultimately be limited by local gravitational fields and field gradients. Because of structural distortion of the spacecraft due to thermal fluctuations, thermal and gravitational disturbances need to be modelled together. For the present discussion, we will be concerned only with the local gravitational interaction between the test mass and a simplified spacecraft structure.
We consider three different proof mass geometries: spherical, cylindrical and cuboid. For simplicity, the spacecraft will be assumed to be a hollow cylinder, as a first approximation.
Following appendix A, we can analyze the gravitational interaction by means of inner, $q_{lm}$, and outer multipole moments, $Q_{lm}$, that describe the proof mass, and the spacecraft and payload mass distribution, respectively. Using expression \[appB3\], the first non-zero outer multipole contribution of a hollow cylinder is $Q_{20}$. Expression \[appB5\] shows that $Q_{20}$ couples to the proof mass inner multipoles $q_{00}$, $q_{1\pm1}$, $q_{20}$, $q_{2\pm1}$ and $q_{2\pm2}$.
The first non-zero inner multipole moments for a parallelepiped proof mass of sides 2a, 2b and 2c are $q_{00}=m_{p}/\sqrt{4\pi}$, $q_{20}=2/3\sqrt{5/4\pi}m_{p}\left(2c^{2}-a^{2}-b^{2}\right)$ and $q_{2\pm2}= 1/12\sqrt{15/2\pi}m_{p}(a^{2}-b^{2})$. On the other hand for a cylindrical proof mass of radius $R$ and semi-height $h$, we have $q_{00}= m_{p}/\sqrt{4\pi}$ and $q_{20}=m_{p}\sqrt{5/4\pi}\left(h^{2}/3-R^{2}/4\right)$. Given these values, it can be seen that the quadrupole moments vanish for a cubic proof mass (as the one adopted by LISA). In the case of ASTROD I, a proof mass of dimensions $50\times50\times35$ mm is considered. In that case $q_{2\pm2}$ vanishes but not $q_{20}$. To first order in the gravitational interaction with a cylindrical tube the $q_{20}$ term appears as a constant energy term and will not contribute to the gravitational force. Disturbances proportional to proof mass cross section area can then be minimized by shortening one of the dimensions (as is the case of ASTROD I proof mass). A trade off between the acceleration disturbances which are proportional to proof mass cross sectional area, and the gravitational interaction needs to be done. If a cylindrical proof mass is utilized we can also suppress quadrupole gravitational interaction by choosing $h/R=\sqrt{3}/2$. In the same way disturbances proportional to cross sectional area can be suppressed by shortening $h$, without altering gravitational interaction with “far away” gravitational asymmetries. A preliminary analysis of gravitational force gradients for ASTROD I is given in Ref. .
Another issue that needs to be considered for ASTROD is the fact that the relative distances and angles between the spacecraft are not constant along their orbits. Therefore, telescopes utilized for laser beam pointing need to be steered during the mission. For active gravitational compensation ASTROD will employ dummy telescopes.[@NiGRG; @Pulido2]
Back action disturbances
------------------------
When deciding which type of sensor to use, there are two main options to consider. First, we could consider a low-stiffness sensor. In that case, the proof mass is highly decoupled from the spacecraft, at the expense of losing sensitivity. The other option is to employ a high-stiffness sensor to achieve better readout sensitivity, at the expense of a high level of coupling between the proof mass and the surrounding structure. Capacitive sensing exemplifies this issue. To improve sensitivity we need to place the electrodes closer to the proof mass. By doing that, the stiffness and back action disturbances increase (see Table 2).
To understand the principle behind stiffness and back action disturbances due to electrostatic sensing and actuation, we first consider the total mechanical energy for a capacitive sensor-proof mass system. Following Refs. and , the total mechanical energy is given by
$$W=-\frac{1}{2}\sum_{i}C_{i}(V_{i}-V_{s})^{2}+
\frac{1}{2}\frac{q^{2}}{C}+qV_{s}$$
\[electrostaticenergy\] where $q$ is the net charge of the proof mass; $C$ is the coefficient of capacitance of the proof mass: $C=\sum_{i}C_{i}$, where $i$ = $x_{1}, x_{2}, y_{1}, y_{2}, z_{1}, z_{2}, g$ defines the capacitances formed by an electrode facing a proof mass side, and capacitance to ground, $C_{g}$; $V_{s}$ is the voltage induced on the proof mass due to the applied voltages, $V_{i}$, and voltage to ground, $V_{g}$: $V_{s}=C^{-1}\sum_{i}C_{i}V_{i}$.
The force acting along a generic direction, assuming neither charge and voltage gradients, is given by
$$F=\frac{1}{2}\sum_{i}C'_{i}(V_{i}-V_{M})^{2}$$
where $V_{M}=V_{s}+\frac{q}{C}$, and $C'_{i}$ is the derivative of capacitance $C_{i}$ along the generic direction.
Force disturbances can then be considered of two types: a) position dependent disturbances, caused by fluctuating position and attitude of the proof mass (stiffness terms), and b) position independent disturbances, caused by voltage and charge fluctuations. Along a generic sensitive (drag-free) axis, we can write the fluctuating force terms due to charge and voltage fluctuations as
$$\delta F_{\delta V,1}=\left(\sum_{i}
C_{i}'(V_{i}-V_{s})-V_{m_{1}}'C\right)\delta V_{i}$$
$$\delta F_{\delta
V,2}=\left(\sum_{i}C_{i}'\frac{q}{C}+V_{m_{2}}'\sum_{i}
C_{i}\right)\delta V_{i}$$
$$\delta F_{\delta q,1}=V_{m_{1}}'\delta q$$
$$\delta F_{\delta q,2}=V_{m_{2}}'\delta q$$
where
$$V_{m_{1}}'=\frac{1}{C}\sum_{i}C_{i}'(V_{i}-V_{s})$$
and $$V_{m_{2}}'=-\frac{q}{C^{2}}\sum_{i}C_{i}'$$
Table 2 shows the back action disturbances given above for the special case of a one dimensional capacitive sensor and one translational degree of freedom.[@Schumaker] Disturbances due to dielectric losses and patch fields are also listed. An extended discussion of electrostatic back action disturbances for ASTROD I and ASTROD can be found in Ref. and , respectively.
Position dependent disturbances (stiffness terms) can be obtained by calculating the variations in capacitances, $\delta
C_{i}$, and capacitance gradients, $\delta C_{i}'$.
In appendix B formulae for capacitances, capacitance gradients and their fluctuations, given by $C_{i}$, $C_{i}'$, $\delta C_{i}$ and $\delta C_{i}'$, respectively, are obtained for the special case of a 6-degree of freedom, cubic, capacitive sensor. These expressions codify cross-coupling effects between translational and rotational degrees of freedom.
On the other hand, optical sensing offers advantages in terms of high sensitivity and low back action forces. Optical readout sensitivity is ultimately limited by laser shot noise. Laser shot noise is proportional to $P^{-1/2}$, where $P$ is the laser power. Lasing back action force is proportional to the laser power, and is given by $2P/c$, where $c$ is the speed of light. This force can, in principle, be compensated to a high degree of accuracy.
Charging disturbances
=====================
Galactic cosmic rays (GCR) and solar energetic particles (SEP) incident on the spacecraft will result in the accumulation of charge on the proof mass. Charge accrued on the proof mass leads to numerous sources of noise. Some of these noise sources have been described and discussed above. First, when summarizing direct acceleration noise sources, the Lorentz force due to the movement of the charged proof mass through the interplanetary magnetic field was discussed. We have also discussed, in the particular case of employing capacitive sensing/actuation, how the proof mass charge couples to sensing/actuation voltages to induce spurious forces and stiffness terms that will affect the performance of the inertial sensor.
Disturbances associated with charging can be divided into three types: a) those which are proportional to charge accrued, $q$, b) those proportional to $\delta q$, which are so-called “shot noise” terms and c) mixed terms proportional to $q\delta q$. This division is of importance in understanding disturbance suppression schemes (see discussion below).
Time dependent forces also contribute to the spectral noise density. Proof mass charging is a time dependent process and both Coulomb and Lorentz forces give rise to coherent Fourier signals (CHS). Assuming a linear increase of proof mass charge with time, the total charge can be written as $q(t)= \bar{\dot{q}}t+\delta
q$, where $\bar{\dot{q}}$ is the mean charging rate. Following Ref. the acceleration noise terms due to Coulomb and Lorentz interactions are given by
$$a_{CHS}= h_{k}(t)=\left(\phi_{k}+\Theta_{k}\right)t+\Xi_{k}t^{2}$$
where
$$\label{coherentonedim}
\phi_{k}=\frac{\bar{\dot{q}}vB_{ip}}{m_{p}\xi_{e}},
\hspace{0.2cm}\Theta_{k}=\frac{\bar{\dot{q}}C_{x}}{m_{p}Cd}V_{d},\hspace{0.1cm}\textrm{and}
\hspace{0.2cm}\Xi_{k}=\frac{2C_{x}}{m_{p}}\left(\frac{\bar{\dot{q}}}{Cd}\right)^{2}
\Delta d$$
using a parallel plate approximation to estimate capacitances and capacitance derivatives. The parameters used above are defined as follows: $v$ is the orbital velocity of the proof mass, $B_{ip}$ is the interplanetary magnetic field, $\xi_{e}$ is the electrostatic shielding factor, $C_{x}$ is the capacitance along the sensitive axis, $V_{d}$ is the voltage difference between opposite sensor sides, and $d$ and $\Delta d$ are the capacitance gap and gap asymmetry, respectively.
Inspection of (\[coherentonedim\]) shows that geometrical and electrostatic asymmetries in the inertial sensor contribute to the appearance of these signals. Geometrical asymmetry arises due to limited machining accuracy and it is represented by an asymmetry in the capacitance gap. Electrostatic asymmetry is due to a stray DC potential imbalance between opposite sides of the sensor. These residual DC stray potentials are dependent on the work function of the metallic surfaces. These potentials are measurable in average, for each electrode, and can be balanced by appropriate applied bias voltages.[@Weber] Ultimately, voltage offset compensation will depend on voltage measurement precision required and work function domains stability in periods of time comparable with the measurement integration time. In the context of LISA mission, these coherent Fourier signals can exceed the instrumental noise target, for typical parameter values.[@Shaul]
At low frequencies, charging disturbances and coherent charging signals are of particular concern. Those so-called “shot noise” charging disturbances and coherent signals scale, roughly speaking, with frequency as $1/f$.[@Shaul] Of special note is the acceleration disturbance proportional to residual voltage difference across opposite sides of the sensor and charge fluctuations, $f_{\delta q,1}$ (see Table 2). By active compensation, the potential difference across opposite sensor sides can be balanced to $\sim$1 mV.[@Weber] To improve on the LISA acceleration noise by a factor of 10 at 0.1 mHz, ASTROD would need to reduce this value to below 0.7 mV,[@Pulido2] giving a disturbance level $f_{\delta q,1}\approx$ 1.4 $\times$ $10^{-16}$ m s$^{-2}$ Hz$^{-1/2}$.
Discharging schemes
-------------------
When discussing discharging schemes we have to keep in mind that not only charge accrued by the proof mass but also the charging rate are potential causes of noise.[@Shaul2; @KXSun2] Discharging the proof mass will suppress some charging disturbances and will reduce the coupling of the proof mass with its surroundings. Coherent signals are proportional to mean charging rate. By “mean charging rate” we mean that charging and discharging rates are added linearly to give a net rate. Therefore by accurately matching charging and discharging rates, these signals can be suppressed.[@Shaul] Nevertheless the fact that proof mass discharging involves the transport of charge “packages”, will cause additional shot noise. Shot noise due to proof mass charging and discharging cannot partially cancel each other because the charging and discharging processes are statistically independent. Therefore their respective shot noise terms have to be added quadratically.
For LISA, a continuous discharging scheme will be adopted. The continuous discharging process will consist basically of two steps. Firstly, the proof mass charge needs to be accurately measured. That can be done by applying a sinusoidal dither voltage and measuring the displacement along a non drag-free axis. This proof mass displacement is proportional to proof mass charge. Secondly, UV light will shine on the proof mass and/or surrounding electrodes to discharge the proof mass via the photoelectric effect.[@Shaul; @Shaul2; @KXSun2]
The ASTROD strategy to suppress charging noise will depend on which sensing/actuation device is employed. ASTROD could benefit from the replacement of capacitive sensing by optical sensing. Even if a capacitive scheme is employed for force actuation, charging disturbances and coherent signals could be suppressed by increasing the gaps between the test mass and surrounding surfaces. If optical force actuation is employed, then only Lorentz type disturbances need to be considered. In that case, the charging requirements can be relaxed and the discharging scheme can be simplified.
Inertial sensors for other missions to test fundamental physics
===============================================================
Follow on LISA mission concepts have been proposed, not only to extend the observational bandwidth to lower frequencies, but also to fill the gap between space antennae and ground based interferometric facilities (LIGO, GEO600, VIRGO, Advanced LIGO, LCGT, etc). Follow-on mission concepts include the Big Bang Observer (BBO) and the Japanese antenna DECi-hertz Interferometer Gravitational wave Observer (DECIGO). Both BBO and DECIGO are to be designed to have an optimum sensitivity between 0.1 Hz and 10 Hz. The objectives of the decihertz antennae are: measuring the expansion rate of the universe; determining the equation of state of dark energy, by observing the coalescence of binary neutron stars and stellar mass black holes; and shedding light on the growth of supermassive black holes, by studying the merger of intermediate mass black holes.[@BBO; @DECIGO]
These missions will also measure the relative distance, or maintain the distance by feedback, between test masses in nearly free fall. To achieve their objectives they need to reduce residual spurious forces to approximately a hundredth of the LISA goal.
These missions are conceptually different. To shift the gravitational wave sensitivity curve towards the decihertz level, the effective interferometer arm length has to be, approximately, a hundredth of the LISA arm length (5 $\times$ 10$^{6}$ km). The BBO preliminary conceptual design consists of a constellation of spacecraft with a LISA-type design and an arm length of 5 $\times$ 10$^{4}$ km. On the other hand, DECIGO plans to place in space a Fabry-Perot cavity of length 1000 km and finesse of approximately 10. To accomplish this, DECIGO would place massive mirrors of 100 kg mass and 1 meter diameter in free fall. Large actuation forces, to keep the cavity in resonance, will be required. This condition and the stringent residual acceleration noise levels, of the order of 4 $\times$ 10$^{-19}$ m s$^{-2}$ Hz$^{-1/2}$, seem to be difficult to reconcile. Nevertheless one could employ a large control loop gain to minimize the acceleration disturbances due to actuation forces. Other technological difficulties common to both missions are related to the requirement for low residual gas pressure. Extremely low pressure could be achieved by venting some of the gas to outer space. However, this could cause other problems such as drag of the proof mass because of residual gas flow, and undesired particles coming from the thruster propellant, brought into the proof mass housing.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author thanks W.-T. Ni for his useful comments on this work and the manuscript and D. N. Shaul for discussing issues related to charge management. This work was funded by the National Natural Science Foundation (Grant No 10475114) and the Foundation of Minor Planets.
[0]{}
B. Lange, The Drag-free Satellite, [*AIAA Journal*]{} [**2(9)**]{}, 1950 (1964); B. Lange, The Control and Use of Drag-free Satellites, Ph.D SUDAER 194, June 1964. http://www.dragfreesatellite.com.
Gravity Probe B. http://einstein.stanford.edu.
LISA, Laser Interferometer Space Antenna: A Cornerstone Mission for the Observation of Gravitational Waves, ESA System and Technology Study Report, ESA-SCI 11, 2000.
S. Anza et al. [*Class. Quantum Grav.*]{} [**22**]{} S125 (2005).
A. Bec-Borsenberger, J. Christensen-Dalsgaard, M. Cruise, A. Di Virgilio, D. Gough, M. Keiser, A. Kosovichev, C. Lämmerzahl, J. Luo, W.-T. Ni, A. Peters, E. Samain, P. H. Scherrer, J.-T. Shy, P. Touboul, K. Tsubono, A.-M. Wu and H.-C. Yeh, Astrodynamical Space Test of Relativity using Optical Devices ASTROD — A Proposal Submitted to ESA in Response to Call for Mission Proposals for Two Flexi-Missions F2/F3, January 31, 2000; and references there in.
W.-T. Ni, [*Int J. Mod. Phys D*]{} [**11(7)**]{}, 947 (2002) ; and references therein.
W.-T. Ni, ASTROD and ASTROD I: an overview, to appear in [*Gen. Rel. Grav.*]{} [**39**]{} (2007).
W.-T. Ni, ASTROD and ASTROD I, to appear in [*Nuclear Physics B (Proceedings Supplements). Spacepart 06 Conf. Proc.*]{}-in press (2007).
W.-T. Ni, ASTROD and ASTROD I: overview and progress, [*Int. J. Mod. Phys. D*]{}, xxx, this issue (2007).
W.-T. Ni, S. Shiomi and A.-C. Liao, [*Class. Quantum Grav.*]{} [**22**]{} S269 (2005).
A. Pulido Patón and W.-T. Ni, The low-frequency sensitivity to gravitational waves for ASTROD, to appear in [*Gen. Rel. Grav.*]{} [**39**]{} (2007).
A. Pulido Patón, ASTROD Gravitational Reference Sensor (GRS): goal and requirements, to appear in [*Nuclear Physics B (Proceedings Supplements). Spacepart 06 Conf. Proc.*]{} Reference: NUPHBP 11611-in press (2007).
Wei-Tou Ni, Antonio Pulido Patón and Yan Xia. Testing Relativistic Gravity to One Part per Billion, in Lasers, Clocks, and Drag-Free: Exploration of Relativistic Gravity in Space, Eds. Hansjörg Dittus, Claus Lämmerzahl and Slava G. Turyshev. Springer (2007).
W.-T. Ni, [*Int. J. Mod. Phys. D*]{}, [**14**]{}(6), 901 (2005).
J. Crowder and N.J. Cornish. [*Phys. Rev. D*]{} [**72**]{}, 083005 (2005); and references therein.
S. Kawamura, et al. [*Class. Quantum Grav.*]{} [**23**]{} S125 (2006).
G. Heinzel et al. [*Class. Quantum Grav.*]{} [**23**]{} S119 (2006).
F. Acernese et al. [*Class. Quantum Grav.*]{} [**22**]{} S279 (2005).
Ke-Xun Sun et al., [*Class. Quantum Grav.*]{} [**22**]{} S287 (2005).
C.C. Speake and S.M. Aston, [*Class. Quantum Grav.*]{} [**22**]{}, S269 (2005).
X. Xu and W-T. Ni, [*Adv. Space Res.*]{} [**32 (7)**]{}, 1443 (2003).
Ke-Xun Sun, Saps Buchman and Robert Byer, [*Journal of Physics: Conference Series*]{} [**32**]{}, 167, (2006).
Clive Speake and Stuart Aston, in Gravitational Wave and Particle Astrophysics Detectors, Eds. James Hough, Gary H. Sanders, [*Proc. of SPIE*]{} [**5500**]{}, 120 (2004).
W.-T. Ni et al, [*Adv. Space Res.*]{} [**32**]{}, 1437 (2003).
S. Vitale et al., [*Nucl. Phys. B (Proc. Suppl.)*]{} [**110**]{}, 209 (2002).
B. L. Schumaker, [*Class. Quantum Grav.*]{} [**20**]{}, S239 (2003).
S. Shiomi and W.-T. Ni, [*Class. Quantum Grav.*]{} [**23**]{}, 4415 (2006).
P. L. Bender, [*Class. Quantum Grav.*]{} [**20**]{}, S301 (2003).
A. Lobo et al. [*Class. Quantum Grav.*]{} [**23**]{} 5177 (2006).
W. J. Weber et al, arXiv : gr-qc / 0309067 v1 (13 Sept 2003).
Sachie Shiomi, [*Journal of Physics: Conference Series*]{} [**32**]{}, 186, (2006).
D. N. A. Shaul et al, [*Int. J. Mod. Phys. D*]{} [**14**]{}, 51 (2005).
D. N. A. Shaul et al, Charge management for LISA and LISA Pathfinder, [*Int. J. Mod. Phys. D*]{}, xxx, this issue, (2007).
Ke-Xun Sun et al. [*Class. Quantum Grav.*]{} [**23**]{} S141 (2006).
J. D. Jackson, [*Classical Electrodynamics*]{}, 3rd edn. (John Wiley and Sons, Inc, 1998), p. 145.
Christian D’Urso and E. G. Adelberger, [*Phys. Rev. D*]{} [**55**]{} (12), 7970 (1997).
The gravitational potential can be written in terms of multipole moments as[@GRAVSachie; @Jackson]
$$V = -4\pi G
\sum_{l=0}^{\infty}\sum_{m=-l}^{l}\frac{1}{2l+1}q_{lm}Q_{lm}
\label{appB1}$$
where $q_{lm}$ ($Q_{lm}$) are the inner (outer) moments defined respectively by
$$q_{lm}=\int_{v_{t}}\rho_{t}(\vec{x}')r'^{l}Y^{*}_{lm}(\theta',\phi')d^{3}\vec{x}'\label{appB2}$$
and
$$Q_{lm}=\int_{v_{s}}\rho_{s}(\vec{x})r^{-(l+l)}Y_{lm}(\theta,\phi)d^{3}\vec{x}
\label{appB3}$$
If the inner multipoles of the test mass, $q_{lm}$, are known in a given reference frame, then the inner multipoles with respect to a new reference frame, can be obtained. Let us consider a position $\vec{r}"=\vec{r}'+\vec{r}$. The test mass position with respect to the reference frame, of origin O, in which the multipoles are known are denoted by $\vec{r}$. On the other hand, $\vec{r}'$ denotes the position vector of the origin O, with respect to the new reference frame in which we wish to work out the multipoles. In the special case of pure translations,[@D'Urso]
$$\begin{aligned}
r"^{L}Y^{*}_{LM}(\theta,\phi)=
\sum_{l,l'=0}^{L}\sum_{m,m'}\sqrt{\frac{4\pi(2L+1)!}{(2l'+1)!(2l+1)!}}r'^{l'}r^{l}\times\nonumber\\
\delta_{L,l+l'}C(l',m',l,m,L,M)Y^{*}_{l'm'}(\theta',\phi')
Y_{lm}(\theta,\phi) \label{appB4}\end{aligned}$$
Using eq. \[appB4\] we can rewrite the gravitational potential energy \[appB1\] in terms of the known inner multipoles and the translational parameters between the two reference frames as
$$\begin{aligned}
V(\vec{r})= -4\pi
G\sum_{L=0}^{\infty}\sum_{M=-L}^{L}\frac{1}{2L+1}Q_{LM}\times\nonumber\\
\sum_{l,l'=0}^{L}\sum_{m,m'}\sqrt{\frac{4\pi(2L+1)!}{(2l'+1)!(2l+1)!}}C(l',m',l,m,L,M)\delta_{L,l+l'}r^{l'}Y^{*}_{l'm'}(\theta,\phi)
q_{lm} \label{appB5}\end{aligned}$$
The gravitational force can be then easily calculated by taking the derivatives of \[appB5\] with respect to the translational parameters.
A cubical inertial sensor. Capacitance calculations
===================================================
We consider a six-dimensional degree of freedom model for the capacitive sensing/actuation device. A cuboid proof mass of side lengths $(2L_{x}, 2L_{y}, 2L_{z})$ is inserted into a three dimensional capacitive sensor. The gaps at the equilibrium position between the proof mass and the electrodes are denoted by $(D_{x}, D_{y}, D_{z})$. The proof mass translational degrees of freedom are denoted by $(d_{x}, d_{y}, d_{z})$. The proof mass rotational degrees of freedom are given by the Euler angles $(\phi, \theta, \psi)$, in the so-called “x-convention”.
We first consider the electrostatic energy density to work out capacitances for this model. If the electric field between two conducting surfaces is defined by $\vec{E}$, then the electrostatic energy density is given by $\omega=\frac{1}{2}\varepsilon_{0}E^{2}$. The proof mass faces and the electrodes will define the capacitances. In the case in which the displacement and rotation of the proof mass are infinitesimal, we can approximate the electric field between conducting surfaces by $\vec{E}_{i}\simeq\frac{(V_{M}-V_{i})}{\triangle
x_{i}}\vec{u}_{i}$, where $V_{M}$ ($V_{i}$) and $\triangle x_{i}$ denote the proof mass (electrode) potential and the capacitance gap in the $i$-direction, respectively. The electrostatic energy is then given by
$$W^{\pm}(d_{x}, d_{y}, d_{z}, \phi, \theta,
\psi)\simeq\frac{1}{2}\varepsilon_{0}(V_{M}-V_{i})^{2}\int\frac{dV}{\triangle
x_{i}^{\pm 2}}$$
where by the simbol $\pm$ we differentiate between the gaps at opposite sides of the sensor.
By integrating along the gap (in this case we choose the gap along the z-axis), the electrostatic energy can be written as
$$W^{\pm}(d_{x}, d_{y}, d_{z}, \phi, \theta,
\psi)\simeq-\frac{1}{2}\varepsilon_{0}(V_{M}-V_{i})^{2}\int\frac{dxdy}{\triangle
x_{i}^{\pm}(x,y,d_{x}, d_{y}, d_{z}, \phi, \theta, \psi)}$$
and we can define the capacitances by
$$C_{x_{i}}^{\pm}(d_{x}, d_{y}, d_{z}, \phi, \theta,
\psi)\simeq\varepsilon_{0}\int dx_{j}dx_{k}\frac{1}{\triangle
x_{i}^{\pm}(x_{j},x_{k},d_{x}, d_{y}, d_{z}, \phi, \theta,
\psi)}\label{cap1}$$
The information about how different degrees of freedom couple to each other is codified in \[cap1\].
To explicitly work out capacitances, we translate and rotate the proof mass by the parameters $(d_{x}, d_{y}, d_{z}, \phi, \theta,
\psi)$.
The rotation matrix in terms of the Euler angles $(\phi, \theta,
\psi)$ is written as $$\begin{aligned}
N(\phi, \theta, \psi)=\hspace{5cm}\nonumber\\
\left(%
\begin{array}{ccc}
\cos \psi \cos \phi -\cos \theta \sin \phi \sin \psi & \cos \psi \sin \phi+\cos \theta \cos \phi \sin \psi & \sin \psi \sin\theta \\
-\sin \psi \cos \phi-\cos \theta \sin \phi \cos \psi & -\sin \psi \sin \phi+\cos \theta \cos \phi \cos \psi & \cos \psi \sin \theta \\
\sin \theta \sin \phi & -\sin \theta \cos \phi & \cos \theta \\
\end{array}%
\right)\end{aligned}$$
Given this matrix we can define the normal vectors to the proof mass faces after a $(\phi, \theta, \psi)$-rotation, with respect to a fixed reference frame. This reference frame has its origin in the geometrical center of the sensor and its axis orthogonal to electrode surfaces. The normal vectors can be written as
$$n_{1}=\left(%
\begin{array}{c}
n_{1x} \\
n_{1y} \\
n_{1z} \\
\end{array}%
\right)=\left(%
\begin{array}{c}
\cos \psi \cos \phi -\cos\theta \sin \phi \sin \psi\\
\cos \psi \sin \phi+\cos \theta \cos \phi \sin \psi\\
\sin \psi \sin \theta\\
\end{array}%
\right)$$
$$n_{2}=\left(%
\begin{array}{c}
n_{2x} \\
n_{2y} \\
n_{2z} \\
\end{array}%
\right)=\left(%
\begin{array}{c}
-\sin \psi \cos \phi-\cos \theta \sin \phi \cos \psi \\
-\sin \psi \sin \phi+\cos \theta \cos \phi \cos \psi\\
\cos \psi \sin \theta\\
\end{array}%
\right)$$
$$n_{3}=\left(%
\begin{array}{c}
n_{3x} \\
n_{3y} \\
n_{3z} \\
\end{array}%
\right)=\left(%
\begin{array}{c}
\sin \theta \sin \phi \\
-\sin \theta \cos \phi\\
\cos \theta\\
\end{array}%
\right)$$
The equations of the proof mass faces, are defined by $$\left(\vec{x}-\vec{P}_{i}\right)\cdot \vec{n}_{i}=0$$ where $\vec{P}_{i}=\pm L_{i}n_{i}+\vec{d}$, being $\vec{d}\equiv(d_{x}, d_{y}, d_{z})$ a displacement, and $L_{i}$ defines the semi-length of the proof mass along the three axes x, y and z.
Using this expressions we can obtain the capacitance gaps for the three directions. These are given by
$$\triangle
x^{\pm}=F^{\pm}_{x}\pm\left(y+A_{x}^{\pm}\right)n_{1}^{x}\pm\left(z+B_{x}^{\pm}\right)n_{2}^{x}$$
$$\triangle
y^{\pm}=F^{\pm}_{y}\pm\left(x+A^{\pm}_{y}\right)n_{1}^{y}\pm\left(z+B^{\pm}_{y}\right)n_{2}^{y}$$
$$\triangle
z^{\pm}=F_{z}^{\pm}\pm\left(x+A_{z}^{\pm}\right)n_{1}^{z}\pm\left(y+B_{z}^{\pm}\right)n_{2}^{z}$$
We define the capacitances as
$$C_{x,up}^{\pm}=\varepsilon_{0}\int_{0}^{L_{y}}
dy\int_{-L_{z}}^{L_{z}}dz\frac{1}{\triangle x^{\pm}(y,z)}$$
$$C_{x,down}^{\pm}=\varepsilon_{0}\int_{-L_{y}}^{0}
dy\int_{-L_{z}}^{L_{z}}dz\frac{1}{\triangle x^{\pm}(y,z)}$$
$$C_{y,up}^{\pm}=\varepsilon_{0}\int_{-L_{x}}^{L_{x}}
dx\int_{0}^{L_{z}}dz\frac{1}{\triangle y^{\pm}(y,z)}$$
$$C_{y,down}^{\pm}=\varepsilon_{0}\int_{-L_{x}}^{L_{x}}
dx\int_{-L_{z}}^{0}dz\frac{1}{\triangle y^{\pm}(y,z)}$$
$$C_{z,left}^{\pm}=\varepsilon_{0}\int_{0}^{L_{x}}
dx\int_{-L_{y}}^{L_{y}}dy\frac{1}{\triangle z^{\pm}(x,y)}$$
$$C_{z,right}^{\pm}=\varepsilon_{0}\int_{-L_{x}}^{0}
dx\int_{-L_{y}}^{L_{y}}dy\frac{1}{\triangle z^{\pm}(x,y)}$$
where by up, down, left, right, we indicate that two electrodes face each proof mass face (see Fig. \[CS\]).
![Schematic view of capacitive sensing electrodes.[]{data-label="CS"}](Capacitivesensor.eps){height="8cm"}
The capacitances in the x-direction are given by
$$\begin{aligned}
C_{x,up}^{+}=\frac{\varepsilon_{0}}{n_{2}}L_{y}\ln\frac{F_{x}^{+}+(L_{z}+B_{x}^{+})n^{x}_{2}+(A_{x}^{+}+L_{y})n^{x}_{1}}{F_{x}^{+}+(B_{x}^{+}-L_{z})n^{x}_{2}+(A_{x}^{+}+L_{y})n^{x}_{1}}\nonumber\\
-\frac{\varepsilon_{0}}{n^{x}_{2}}\frac{F_{x}^{+}+(L_{z}+B_{x}^{+})n^{x}_{2}+A_{x}^{+}n^{x}_{1}}{n^{x}_{1}}\ln\frac{F_{x}^{+}+(L_{z}+B_{x}^{+})n^{x}_{2}+(A_{x}^{+}+L_{y})n^{x}_{1}}{F_{x}^{+}+(L_{z}+B_{x}^{+})n^{x}_{2}+A_{x}^{+}n^{x}_{1}}\nonumber\\
+\frac{\varepsilon_{0}}{n^{x}_{2}}\frac{F_{x}^{+}+(B_{x}^{+}-L_{z})n^{x}_{2}+A_{x}^{+}n^{x}_{1}}{n^{x}_{1}}\ln\frac{F_{x}^{+}+(B_{x}^{+}-L_{z})n^{x}_{2}+(A_{x}^{+}+L_{y})n^{x}_{1}}{F_{x}^{+}+(B_{x}^{+}-L_{z})n^{x}_{2}+A_{x}^{+}n^{x}_{1}}\end{aligned}$$
$$\begin{aligned}
C_{x,up}^{-}=-\frac{\varepsilon_{0}}{n^{x}_{2}}L_{y}\ln\frac{F_{x}^{-}-(L_{z}+B_{x}^{-})n^{x}_{2}-(A_{x}^{-}+L_{y})n^{x}_{1}}{F_{x}^{-}-(B_{x}^{-}-L_{z})n^{x}_{2}-(A_{x}^{-}+L_{y})n^{x}_{1}}\nonumber\\
-\frac{\varepsilon_{0}}{n^{x}_{2}}\frac{F_{x}^{-}-(L_{z}+B_{x}^{-})n^{x}_{2}-A_{x}^{-}n^{x}_{1}}{n^{x}_{1}}\ln\frac{F_{x}^{-}-(L_{z}+B_{x}^{-})n^{x}_{2}-(A_{x}^{-}+L_{y})n^{x}_{1}}{F_{x}^{-}-(L_{z}+B_{x}^{-})n^{x}_{2}-A_{x}^{-}n^{x}_{1}}\nonumber\\
+\frac{\varepsilon_{0}}{n^{x}_{2}}\frac{F_{x}^{-}-(B_{x}^{-}-L_{z})n^{x}_{2}-A_{x}^{-}n^{x}_{1}}{n^{x}_{1}}\ln\frac{F_{x}^{-}-(B_{x}^{-}-L_{z})n^{x}_{2}-(A_{x}^{-}+L_{y})n^{x}_{1}}{F_{x}^{-}-(B_{x}^{-}-L_{z})n^{x}_{2}-A_{x}^{-}n^{x}_{1}}\end{aligned}$$
$$\begin{aligned}
C_{x,down}^{+}=\frac{\varepsilon_{0}}{n^{x}_{2}}L_{y}\ln\frac{F_{x}^{+}+(L_{z}+B_{x}^{+})n^{x}_{2}+(A_{x}^{+}-L_{y})n^{x}_{1}}{F_{x}^{+}+(B_{x}^{+}-L_{z})n^{x}_{2}+(A_{x}^{+}-L_{y})n^{x}_{1}}\nonumber\\
+\frac{\varepsilon_{0}}{n^{x}_{2}}\frac{F_{x}^{+}+(L_{z}+B_{x}^{+})n^{x}_{2}+A_{x}^{+}n^{x}_{1}}{n^{x}_{1}}\ln\frac{F_{x}^{+}+(L_{z}+B_{x}^{+})n^{x}_{2}+(A_{x}^{+}-L_{y})n^{x}_{1}}{F_{x}^{+}+(L_{z}+B_{x}^{+})n^{x}_{2}+A_{x}^{+}n^{x}_{1}}\nonumber\\
-\frac{\varepsilon_{0}}{n^{x}_{2}}\frac{F_{x}^{+}+(B_{x}^{+}-L_{z})n^{x}_{2}+A_{x}^{+}n^{x}_{1}}{n^{x}_{1}}\ln\frac{F_{x}^{+}+(B_{x}^{+}-L_{z})n^{x}_{2}+(A_{x}^{+}-L_{y})n^{x}_{1}}{F_{x}^{+}+(B_{x}^{+}-L_{z})n^{x}_{2}+A_{x}^{+}n^{x}_{1}}\end{aligned}$$
$$\begin{aligned}
C_{x,down}^{-}=-\frac{\varepsilon_{0}}{n^{x}_{2}}L_{y}\ln\frac{F_{x}^{-}-(L_{z}+B_{x}^{-})n^{x}_{2}-(A_{x}^{-}-L_{y})n^{x}_{1}}{F_{x}^{-}-(B_{x}^{-}-L_{z})n^{x}_{2}-(A_{x}^{-}-L_{y})n^{x}_{1}}\nonumber\\
+\frac{\varepsilon_{0}}{n^{x}_{2}}\frac{F_{x}^{-}-(L_{z}+B_{x}^{-})n^{x}_{2}-A_{x}^{-}n^{x}_{1}}{n^{x}_{1}}\ln\frac{F_{x}^{-}-(L_{z}+B_{x}^{-})n^{x}_{2}-(A_{x}^{-}-L_{y})n^{x}_{1}}{F_{x}^{-}-(L_{z}+B_{x}^{-})n^{x}_{2}-A_{x}^{-}n^{x}_{1}}\nonumber\\
-\frac{\varepsilon_{0}}{n^{x}_{2}}\frac{F_{x}^{-}-(B_{x}^{-}-L_{z})n^{x}_{2}-A_{x}^{-}n^{x}_{1}}{n^{x}_{1}}\ln\frac{F_{x}^{-}-(B_{x}^{-}-L_{z})n^{x}_{2}-(A_{x}^{-}-L_{y})n^{x}_{1}}{F_{x}^{-}-(B_{x}^{-}-L_{z})n^{x}_{2}-A_{x}^{-}n^{x}_{1}}\end{aligned}$$
where
$$\begin{aligned}
F_{x}^{\pm}\equiv L_{x}(1-n_{1x})+D_{x}\mp d_{x}\\
A_{x}^{\pm}\equiv \mp L_{x}n_{1y}-d_{y}\\
B_{x}^{\pm}\equiv\mp L_{x}n_{1z}-d_{z}\\
n^{x}_{1}\equiv\frac{n_{1y}}{n_{1x}}\\
n^{x}_{2}\equiv\frac{n_{1z}}{n_{1x}}\end{aligned}$$
The capacitances in the y-direction are
$$\begin{aligned}
C_{y,up}^{\pm}=C_{x,up}^{\pm}((L_{y},L_{z},A_{x}^{\pm},B_{x}^{\pm},n^{x}_{1},n^{x}_{2})\rightarrow
(L_{z},L_{x},B_{y}^{\pm},A_{y}^{\pm},n^{y}_{2},n^{y}_{1}))\\
C_{y,down}^{\pm}=C_{x,down}^{\pm}((L_{y},L_{z},A_{x}^{\pm},B_{x}^{\pm},n^{x}_{1},n^{x}_{2})\rightarrow
(L_{z},L_{x},B_{y}^{\pm},A_{y}^{\pm},n^{y}_{2},n^{y}_{1}))\end{aligned}$$
where
$$\begin{aligned}
F_{y}^{\pm}\equiv L_{y}(1-n_{2y})+D_{y}\mp d_{y}\\
A_{y}^{\pm}\equiv \mp L_{y}n_{2x}-d_{x}\\
B_{y}^{\pm}\equiv\mp L_{y}n_{2z}-d_{z}\\
n^{y}_{1}\equiv\frac{n_{2x}}{n_{2y}}\\
n^{y}_{2}\equiv\frac{n_{2z}}{n_{2y}}\end{aligned}$$
The capacitances in the z-direction are given by
$$\begin{aligned}
C_{z,left}^{\pm}=C_{x,up}^{\pm}(L_{z}\rightarrow L_{y};L_{y}\rightarrow L_{x})\\
C_{z,right}^{\pm}=C_{x,down}^{\pm}(L_{z}\rightarrow
L_{y};L_{y}\rightarrow L_{x})\end{aligned}$$
where now
$$\begin{aligned}
F_{z}^{\pm}\equiv L_{z}(1-n_{3z})+D_{z}\mp d_{z}\\
A_{z}^{\pm}\equiv \mp L_{z}n_{3x}-d_{x}\\
B_{z}^{\pm}\equiv\mp L_{z}n_{3y}-d_{y}\\
n^{z}_{1}\equiv\frac{n_{3x}}{n_{3z}}\\
n^{z}_{2}\equiv\frac{n_{3y}}{n_{3z}}\end{aligned}$$
We can approximate the expressions of capacitances by taking into account that $F\gg$ $A$ and $B$. Then we have
$$C_{x,up}^{+}=2\varepsilon_{0}L_{y}L_{z}\frac{1}{F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}+L_{y})n^{x}_{1}}$$
$$C_{x,up}^{-}=2\varepsilon_{0}L_{y}L_{z}\frac{1}{F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}+L_{y})n^{x}_{1}}$$
$$C_{x,down}^{+}=2\varepsilon_{0}L_{y}L_{z}\frac{1}{F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}-L_{y})n^{x}_{1}}$$
$$C_{x,down}^{-}=2\varepsilon_{0}L_{y}L_{z}\frac{1}{F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}-L_{y})n^{x}_{1}}$$
To work out the force and the force disturbance along the sensitive axis we also need to work out capacitance gradients, which are given by
$$\frac{\partial C_{x,up}^{+}}{ \partial
d_{x}}=\frac{2\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}+L_{y})n^{x}_{1}\right]^{2}}$$
$$\frac{\partial C_{x,up}^{-}}{ \partial
d_{x}}=-\frac{2\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}+L_{y})n^{x}_{1}\right]^{2}}$$
$$\frac{\partial C_{x,down}^{+}}{ \partial
d_{x}}=\frac{2\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}-L_{y})n^{x}_{1}\right]^{2}}$$
$$\frac{\partial C_{x,down}^{-}}{ \partial
d_{x}}=-\frac{2\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}-L_{y})n^{x}_{1}\right]^{2}}$$
Along the y-axis we have
$$\frac{\partial C_{y,up}^{+}}{ \partial
d_{x}}=-\frac{2\varepsilon_{0}L_{x}L_{z}(\phi
+\psi)}{\left[F_{y}^{+}+A_{y}^{+}n^{y}_{1}+(B_{y}^{+}+L_{z})n^{y}_{2}\right]^{2}}$$
$$\frac{\partial C_{y,up}^{-}}{ \partial
d_{x}}=\frac{2\varepsilon_{0}L_{x}L_{z}(\phi
+\psi)}{\left[F_{y}^{-}-A_{y}^{-}n^{y}_{1}-(B_{y}^{-}+L_{z})n^{y}_{2}\right]^{2}}$$
$$\frac{\partial C_{y,down}^{+}}{ \partial
d_{x}}=-\frac{2\varepsilon_{0}L_{x}L_{z}(\phi
+\psi)}{\left[F_{y}^{+}+A_{y}^{+}n^{y}_{1}+(B_{y}^{+}-L_{z})n^{y}_{2}\right]^{2}}$$
$$\frac{\partial C_{y,down}^{-}}{ \partial
d_{x}}=\frac{2\varepsilon_{0}L_{x}L_{z}(\phi
+\psi)}{\left[F_{y}^{-}-A_{y}^{-}n^{y}_{1}-(B_{y}^{-}-L_{z})n^{y}_{2}\right]^{2}}$$
Along the z-axis,
$$\frac{\partial C_{z,left}^{+}}{ \partial
d_{x}}=\frac{2\varepsilon_{0}L_{x}L_{y}\phi\theta}{\left[F_{z}^{+}+B_{z}^{+}n^{z}_{2}+(A_{z}^{+}+L_{x})n^{z}_{1}\right]^{2}}$$
$$\frac{\partial C_{z,left}^{-}}{ \partial
d_{x}}=-\frac{2\varepsilon_{0}L_{x}L_{y}\phi\theta}{\left[F_{z}^{-}-B_{z}^{-}n^{z}_{2}-(A_{z}^{-}+L_{x})n^{z}_{1}\right]^{2}}$$
$$\frac{\partial C_{z,right}^{+}}{ \partial
d_{x}}=\frac{2\varepsilon_{0}L_{x}L_{y}\phi\theta}{\left[F_{z}^{+}+B_{z}^{+}n^{z}_{2}+(A_{z}^{+}-L_{x})n^{z}_{1}\right]^{2}}$$
$$\frac{\partial C_{z,right}^{-}}{ \partial
d_{x}}=-\frac{2\varepsilon_{0}L_{x}L_{y}\phi\theta}{\left[F_{z}^{-}-B_{z}^{-}n^{z}_{2}-(A_{z}^{-}-L_{x})n^{z}_{1}\right]^{2}}$$
The terms useful for stiffness calculations are variations of capacitances and capacitance gradients. These are given by
$$\delta
C_{x,up}^{+}=\frac{-2\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}+L_{y})n^{x}_{1}\right]^{2}}\delta
\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}+L_{y})n^{x}_{1}\right]$$
$$\delta
C_{x,up}^{-}=\frac{-2\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}+L_{y})n^{x}_{1}\right]^{2}}\delta\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}+L_{y})n^{x}_{1}\right]$$
$$\delta
C_{x,down}^{+}=\frac{-2\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}-L_{y})n^{x}_{1}\right]^{2}}\delta\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}-L_{y})n^{x}_{1}\right]$$
$$\delta
C_{x,down}^{-}=\frac{-2\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}-L_{y})n^{x}_{1}\right]^{2}}\delta
\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}-L_{y})n^{x}_{1}\right]$$
Finally the last useful formulae for stiffness calculations are those of the type $\delta C_{i}'$. On the x-axis we have
$$\delta
C_{x,up}^{+'}=\frac{-4\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}+L_{y})n^{x}_{1}\right]^{3}}\delta
\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}+L_{y})n^{x}_{1}\right]$$
$$\delta
C_{x,up}^{-'}=\frac{4\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}+L_{y})n^{x}_{1}\right]^{3}}
\delta\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}+L_{y})n^{x}_{1}\right]$$
$$\delta
C_{x,down}^{+'}=\frac{-4\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}-L_{y})n^{x}_{1}\right]^{3}}
\delta\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}-L_{y})n^{x}_{1}\right]$$
$$\delta
C_{x,down}^{-'}=\frac{4\varepsilon_{0}L_{y}L_{z}}{\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}-L_{y})n^{x}_{1}\right]^{3}}\delta
\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}-L_{y})n^{x}_{1}\right]$$
On the y-axis the expressions are as follows:
$$\delta C_{y,up}^{+'}= \frac{4\varepsilon_{0}L_{x}L_{z}(\phi
+\psi)}{\left[F_{y}^{+}+A_{y}^{+}n^{y}_{1}+(B_{y}^{+}+L_{z})n^{y}_{2}\right]^{3}}
\delta\left[F_{y}^{+}+A_{y}^{+}n^{y}_{1}+(B_{y}^{+}+L_{z})n^{y}_{2}\right]$$
$$\delta C_{y,up}^{-'}= -\frac{4\varepsilon_{0}L_{x}L_{z}(\phi
+\psi)}{\left[F_{y}^{-}-A_{y}^{-}n^{y}_{1}-(B_{y}^{-}+L_{z})n^{y}_{2}\right]^{3}}
\delta\left[F_{y}^{-}-A_{y}^{-}n^{y}_{1}-(B_{y}^{-}+L_{z})n^{y}_{2}\right]$$
$$\delta C_{y,down}^{+'}= \frac{4\varepsilon_{0}L_{x}L_{z}(\phi
+\psi)}{\left[F_{y}^{+}+A_{y}^{+}n^{y}_{1}+(B_{y}^{+}-L_{z})n^{y}_{2}\right]^{3}}
\delta
\left[F_{y}^{+}+A_{y}^{+}n^{y}_{1}+(B_{y}^{+}-L_{z})n^{y}_{2}\right]$$
$$\delta C_{y,right}^{-'}= -\frac{4\varepsilon_{0}L_{x}L_{z}(\phi
+\psi)}{\left[F_{y}^{-}-A_{y}^{-}n^{y}_{1}-(B_{y}^{-}-L_{z})n^{y}_{2}\right]^{3}}
\delta\left[F_{y}^{-}-A_{y}^{-}n^{y}_{1}-(B_{y}^{-}-L_{z})n^{y}_{2}\right]$$
And finally on the z-axis we have
$$\delta
C_{z,left}^{+'}=-\frac{4\varepsilon_{0}L_{x}L_{y}\phi\theta}{\left[F_{z}^{+}+B_{z}^{+}n^{z}_{2}+(A_{z}^{+}+L_{x})n^{z}_{1}\right]^{3}}
\delta\left[F_{z}^{+}+B_{z}^{+}n^{z}_{2}+(A_{z}^{+}+L_{x})n^{z}_{1}\right]$$
$$\delta
C_{z,left}^{-'}=\frac{4\varepsilon_{0}L_{x}L_{y}\phi\theta}{\left[F_{z}^{-}-B_{z}^{-}n^{z}_{2}-(A_{z}^{-}+L_{x})n^{z}_{1}\right]^{3}}
\delta\left[F_{z}^{-}-B_{z}^{-}n^{z}_{2}-(A_{z}^{-}+L_{x})n^{z}_{1}\right]$$
$$\delta
C_{z,right}^{+'}=-\frac{4\varepsilon_{0}L_{x}L_{y}\phi\theta}{\left[F_{z}^{+}+B_{z}^{+}n^{z}_{2}+(A_{z}^{+}-L_{x})n^{z}_{1}\right]^{3}}
\delta\left[F_{z}^{+}+B_{z}^{+}n^{z}_{2}+(A_{z}^{+}-L_{x})n^{z}_{1}\right]$$
$$\delta
C_{z,right}^{-'}=\frac{4\varepsilon_{0}L_{x}L_{y}\phi\theta}{\left[F_{z}^{-}-B_{z}^{-}n^{z}_{2}-(A_{z}^{-}-L_{x})n^{z}_{1}\right]^{3}}
\delta\left[F_{z}^{-}-B_{z}^{-}n^{z}_{2}-(A_{z}^{-}-L_{x})n^{z}_{1}\right]$$
where for the x, y and z axes we have
$$\delta\left[F_{x}^{+}+B_{x}^{+}n^{x}_{2}+(A_{x}^{+}\pm
L_{y})n^{x}_{1}\right]=-\delta d_{x}\pm L_{y}(\delta
\phi+\delta\psi)$$
$$\delta\left[F_{x}^{-}-B_{x}^{-}n^{x}_{2}-(A_{x}^{-}\pm
L_{y})n^{x}_{1}\right]=+\delta d_{x}\mp L_{y}(\delta
\phi+\delta\psi)$$
$$\delta\left[F_{y}^{+}+A_{y}^{+}n^{y}_{1}+(B_{y}^{+}\pm
L_{z})n^{y}_{2}\right]=-\delta d_{y}\pm L_{z}\delta \theta$$
$$\delta\left[F_{y}^{-}-A_{y}^{-}n^{y}_{1}-(B_{y}^{-}\pm
L_{z})n^{y}_{2}\right]=\delta d_{y}\mp L_{z}\delta \theta$$
$$\delta\left[F_{z}^{+}+B_{z}^{+}n^{z}_{2}+(A_{z}^{+}\pm
L_{x})n^{z}_{1}\right]=-\delta d_{z}$$
$$\delta\left[F_{z}^{-}-B_{z}^{-}n^{z}_{2}-(A_{z}^{-}\pm
L_{x})n^{z}_{1}\right]=\delta d_{z}$$
Capacitance as a position sensor.
---------------------------------
By measuring and combining capacitances along the different axis, we can obtain the position and attitude of the proof mass. The parameters $(d_{x},d_{y},d_{z},\phi, \theta, \psi)$ are obtained by the following combination of capacitances,
$$(C_{z,r}^{+}-C_{z,r}^{-})+(C_{z,l}^{+}-C_{z,l}^{-})\simeq8\varepsilon_{0}L_{x}L_{y}\frac{d_{z}}{D_{z}^{2}}$$
$$(C_{z,r}^{+}-C_{z,r}^{-})-(C_{z,l}^{+}-C_{z,l}^{-})\simeq8\varepsilon_{0}\frac{L_{x}^{2}L_{y}}{D_{z}^{2}}\theta\phi$$
$$(C_{x,up}^{+}-C_{x,up}^{-})+(C_{x,down}^{+}-C_{x,down}^{-})\simeq8\varepsilon_{0}L_{z}L_{y}\frac{d_{x}}{D_{x}^{2}}$$
$$(C_{x,up}^{+}-C_{x,up}^{-})-(C_{x,down}^{+}-C_{x,down}^{-})\simeq-8\varepsilon_{0}L_{y}^{2}L_{z}(\phi+\psi)\frac{1}{D_{x}^{2}}$$
$$(C_{y,up}^{+}-C_{y,up}^{-})+(C_{y,down}^{+}-C_{y,down}^{-})\simeq8\varepsilon_{0}L_{x}L_{z}\frac{d_{y}}{D_{y}^{2}}$$
$$(C_{y,up}^{+}-C_{y,up}^{-})-(C_{y,down}^{+}-C_{y,down}^{-})\simeq-8\varepsilon_{0}L_{x}L_{z}^{2}\theta\frac{1}{D_{y}^{2}}$$
Capacitances, capacitance derivatives and their variations
----------------------------------------------------------
For the special case in which the proof mass is in the equilibrium position, $\vec{0}$, with no translational and rotational offsets,
$$C_{x,up}^{\pm}=C_{x,down}^{\pm}=\frac{2\varepsilon_{0}L_{y}L_{z}}{D_{x}}$$
$$C_{y,up}^{\pm}=C_{y,down}^{\pm}=\frac{2\varepsilon_{0}L_{x}L_{z}}{D_{y}}$$
$$C_{z,left}^{\pm}=C_{z,right}^{\pm}=\frac{2\varepsilon_{0}L_{x}L_{y}}{D_{z}}$$
$$\frac{\partial C_{x,up}^{\pm}}{\partial d_{x}}=\frac{\partial
C_{x,down}^{\pm}}{\partial
d_{x}}=\pm\frac{2\varepsilon_{0}L_{y}L_{z}}{D_{x}^{2}}$$
$$\frac{\partial C_{y,up}^{\pm}}{\partial d_{x}}=\frac{\partial
C_{y,down}^{\pm}}{\partial d_{x}}=\frac{\partial
C_{z,left}^{\pm}}{\partial d_{x}}=\frac{\partial
C_{z,right}^{\pm}}{\partial d_{x}}\approx0$$
$$\delta
C_{x,up}^{\pm}=\mp\frac{2\varepsilon_{0}L_{y}L_{z}}{D_{x}^{2}}\left[-\delta
d_{x}+ L_{y}(\delta \phi+\delta\psi)\right]$$
$$\delta
C_{x,down}^{\pm}=\mp\frac{2\varepsilon_{0}L_{y}L_{z}}{D_{x}^{2}}\left[-\delta
d_{x}- L_{y}(\delta \phi+\delta\psi)\right]$$
$$\delta
C_{y,up}^{\pm}=\mp\frac{2\varepsilon_{0}L_{x}L_{z}}{D_{y}^{2}}\left[-\delta
d_{y}+L_{z}\delta \theta\right]$$
$$\delta
C_{y,down}^{\pm}=\mp\frac{2\varepsilon_{0}L_{x}L_{z}}{D_{y}^{2}}\left[-\delta
d_{y}-L_{z}\delta\theta\right]$$
$$\delta C_{z,left}^{\pm}=\delta
C_{z,right}^{\pm}=\pm\frac{2\varepsilon_{0}L_{x}L_{y}}{D_{z}^{2}}\delta
d_{z}$$
$$\delta
C_{x,up}^{\pm'}=\frac{-4\varepsilon_{0}L_{y}L_{z}}{D_{x}^{3}}\left[-\delta
d_{x}+ L_{y}(\delta \phi+\delta\psi)\right]$$
$$\delta
C_{x,down}^{\pm'}=\frac{-4\varepsilon_{0}L_{y}L_{z}}{D_{x}^{3}}\left[-\delta
d_{x}- L_{y}(\delta \phi+\delta\psi)\right]$$
$$\delta C_{y,up}^{\pm'}=\delta C_{y,down}^{\pm'}=\delta
C_{z,left}^{\pm'}=\delta C_{z,right}^{\pm'}\approx0$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We develop a theory of the valley Hall effect in high-quality graphene samples, in which strain fluctuation-induced random gauge potentials have been suggested as the dominant source of disorder. We find a near-quantized value of valley Hall conductivity in the band transport regime, which originates from an enhanced side jump of a Dirac electron when it scatters off the gauge potential. By assuming a small residue charge density our theory reproduces qualitatively the temperature- and gap-dependence of the observed valley Hall effect at the charge neutral point. Our study suggests that the valley Hall effect in graphene systems represents a new paradigm for the anomalous Hall physics where gauge disorder plays an important role.'
author:
- 'Wen-Yu Shan'
- Di Xiao
title: 'Strain-Fluctuation-Induced Near-Quantization of Valley Hall Conductivity in Graphene Systems'
---
Introduction
============
Charge carriers in graphene can be described by the two-dimensional (2D) Dirac equation, which exhibit a slew of interesting electronic properties [@castro2009]. One of the consequences is that strains behave as pseudo-magnetic fields for Dirac electrons, and carefully designed lattice deformation pattern can result in the formation of Landau levels [@guinea2010]. Even without the engineered strains, random strain fluctuations are inevitable in 2D materials [@meyer2007; @fasolino2007; @vozmediano2010; @amorim2016]. They appear in the form of either out-of-plane corrugations due to thermal ripples or in-plane displacement from the interaction with substrates. Strain fluctuations then act like random magnetic (gauge) field, and can significantly affect transport behaviors. Their effect has been extensively explored in the longitudinal transport phenomena [@ludwig1994; @couto2014; @engels2014; @ochoa2013; @vicent2017], such as weak localization [@couto2014; @engels2014] and spin relaxation phenomena [@ochoa2013; @vicent2017].
In this work, we investigate the role of random strain fluctuations on a particular type of transverse transport phenomena—the valley Hall effect of 2D Dirac electrons [@xiao2007]. The valley Hall effect can be regarded as two opposite copies of the anomalous Hall effect of a pair of gapped Dirac points related by time-reversal symmetry [@nagaosa2010]. That is, carriers in the two valleys will flow in the opposite transverse direction upon the application of a longitudinal electric field. Our motivation is twofold. First, in gapped monolayer and bilayer graphene systems, the valley Hall effect has been observed experimentally [@gorbachev2014; @sui2015; @shimazaki2015]. Thus, detailed experimental study of Hall-type transport in the presence of strain fluctuations is feasible. Secondly, all recent valley Hall measurements in graphene are carried out in high-quality devices in which strain fluctuations are the dominant source of disorder [@couto2014; @engels2014]. However, no existing theories have discussed its effect on the Hall transport. We will show that, the strain fluctuations are essential to understand the valley Hall effect of 2D Dirac electrons, and provide a new insight to recent debates [@lensky2015; @kirczenow2015; @li2011; @zhu2017; @song2018; @brown2018] on the observed nonlocal signals at the charge neutral point [@gorbachev2014; @sui2015; @shimazaki2015].
Our main results are summarized below. Focusing on the band transport regime, we find that the valley Hall conductivity exhibits a singular behavior in the presence of strain fluctuation-induced long-range gauge disorder: as the Fermi level sweeps across the band edge, it jumps from zero to a *nearly* quantized value, $2e^2/h$ for monolayer graphene and $4e^2/h$ for bilayer graphene. The origin of this singular behavior is traced back to an enhanced side jump of a Dirac electron when it scatters off the gauge potential [@yang2011; @yang2011a]. Furthermore, at the charge neutral point, by assuming a small residue charge density we calculate the temperature- and gap-dependence of the valley Hall conductivity, which qualitatively agrees with the experiment [@sui2015]. We also find that strain-induced long-range scalar potential can reduce the valley Hall conductivity from its quantized value. Our study suggests that the valley Hall effect in graphene systems represents a new paradigm for the anomalous Hall physics where gauge disorder plays an important role. Our theory can also be applied to other 2D valley Hall materials such as transition metal dichalcogenides as well [@wu2018; @hung2019].
The paper is organized as follows. In Sec. \[sec:intrinsic\] we present the intrinsic valley Hall effect. In Sec. \[side\_jump\] we study the random strain-induced side jump and the resulting valley Hall effect. Temperature dependence and rigorous numerical analysis are shown in Sec. \[temperature\] and Sec. \[numerics\], respectively. In Sec. \[scalar\], effect of long-range scalar potential is investigated. Finally, discussion and conclusion are made in Sec. \[conclusion\]. Technical details are relegated to the appendixes.
Intrinsic valley Hall effect {#sec:intrinsic}
============================
We begin with the following effective Hamiltonian $$\label{ham}
H_0 = \hbar v\bm k \cdot \bm\sigma + \Delta\sigma_z \;,$$ which describes the low-energy electron dynamics in one of the Dirac valleys in gapped graphene. Here $v$ is the velocity, $\bm k = (k_x, k_y)$ is the two-dimensional wave vector, $2\Delta$ is the band gap opened by inversion symmetry breaking, and $\bm\sigma$ represents the sublattice indices. The Hamiltonian for the other valley can be obtained by performing a time-reversal operation on $H_0$. The energy dispersion is given by ${\varepsilon}_{c,v} = \pm {\varepsilon}_{\bm k} = \pm (\hbar^2v^2k^2 + \Delta^2)^{1/2}$ with the corresponding eigenstates $$\label{effH}
{|u_{\bm k}^c\rangle} = \binom{\cos\frac{\theta_{\bm k}}{2}}{\sin\frac{\theta_{\bm k}}{2}e^{i\phi_{\bm k}}} \;, \quad
{|u_{\bm k}^v\rangle} = \binom{\sin\frac{\theta_{\bm k}}{2}e^{-i\phi_{\bm k}}}{-\cos\frac{\theta_{\bm k}}{2}} \;,$$ where the superscript $c$ and $v$ label the conduction and valence bands, respectively, and the angular variables $\theta_{\bm k}$ and $\phi_{\bm k}$ are defined as $\theta_{\bm k} \equiv \cos^{-1}(\Delta/{\varepsilon}_{\bm k})$ and $\phi_{\bm k} \equiv \tan^{-1}(k_y/k_x)$. The opening of the band gap gives rise to nonzero Berry curvature [@xiao2007], defined by $\bm\Omega_n(\bm k) = \Omega_n(\bm k) \hat z = i{\langle \nabla_{\bm
k}u_{\bm k}^n|\times|\nabla_{\bm k}u_{\bm k}^n \rangle}$ [@xiao2010]. For the two-band model given in Eq. we have $$\Omega_c(\bm k) = -\Omega_v(\bm k) = -\frac{\hbar^2v^2\Delta}{2(\hbar^2v^2k^2+\Delta^2)^{3/2}} \;.$$ The Berry curvature in the other valley has opposite sign, as required by time-reversal symmetry.
Assuming weak inter-valley scattering, we can decouple the valley Hall effect into two copies of the anomalous Hall effect for each valley species. However, this decoupling must be treated with care. In the anomalous Hall effect, there is an intrinsic contribution to the Hall conductivity, given by the summation of the Berry curvature over all occupied Bloch states [@nagaosa2010]. It can be divided into two parts. One comes from fully occupied bands. This part manifests as chiral edge states at the Fermi energy, and gives rise to a quantized contribution to the Hall conductivity. On the other hand, the valley Hall systems considered here are topologically trivial without protected edge states. Therefore no electronic transport is possible when the Fermi energy is inside the band gap [^1]. Consequently, we shall drop this contribution in the calculation of the valley Hall conductivity. This leaves us with the contribution from partially occupied bands. Haldane has argued that this contribution can be written as the Berry phase of quasiparticles moving on the Fermi surface, and thus it can be regarded as a Fermi surface property [@haldane2004]. For an electron-doped sample, the Fermi surface contribution is $$\label{intrins}
\sigma_H^\text{int} = 4\frac{e^2}{\hbar}\sum_{\bm k} \Omega_c(\bm k) \Theta({\varepsilon}_F - {\varepsilon}_{\bm k}) = \frac{2e^2}{h}(1 - \cos\theta_F) \;,$$ where the factor of 4 counts the spin and valley degeneracy, ${\varepsilon}_F$ is the Fermi energy, and $\theta_F \equiv \theta_{k_F}$. When ${\varepsilon}_F$ approaches the band edge, $\sigma_H^\text{int}$ vanishes. If the sample is hole doped, then one should consider the Fermi surface contribution of holes, which is opposite to that of electrons. Obviously, the intrinsic contribution alone cannot explain the observed valley Hall effect around the charge neutral point [@gorbachev2014; @sui2015; @shimazaki2015].
Random strain-induced side jump {#side_jump}
===============================
To remedy this situation, we consider the effect of random strain-induced gauge disorder. In graphene systems, strain can be induced either by out-of-plane corrugations or by in-plane displacements of the carbon atoms. Both will generate a random gauge potential for Dirac electrons [@vozmediano2010], $$\label{gauge_dis}
V_\text{imp}(\bm r) = d_x(\bm r) \sigma_x + d_y(\bm r) \sigma_y \;.$$ The point group symmetry of graphene requires that $\bm d(\bm
r)\propto(u_{xx}-u_{yy},-2u_{xy})$, where $u_{\alpha\beta}$ is the strain tensor defined in terms of the deformation field $\bm u(\bm r)$, $u_{\alpha\beta} = (\partial_\alpha u_\beta + \partial_\beta u_\alpha)/2$. The Fourier transform of $\bm d(\bm r)$ has the form [@couto2014] $$\label{dq}
d_{\pm}(\bm q)= F(\bm q) q^2 e^{\mp 2i\varphi_{\bm q}} \;,$$ where $d_{\pm}(\bm q)=d_x(\bm q)\pm i d_y(\bm q)$, and $F(\bm q)$ is a prefactor depending on the details of the strain field. The appearance of the phase angle $2\varphi_{\bm q}$, defined by $\varphi_{\bm q}\equiv\tan^{-1}(q_y/q_x)$, is due to the fact that $u_{\alpha\beta}$ is a second-order derivative of either the height field (out-of-plane corrugations) or random potentials from substrates (in-plane displacements) [@couto2014; @vozmediano2010]. Both modes are long-wavelength elastic modes, as indicated by the $\bm q$-dependence of $F(\bm q)$.
![ (a) Schematic view of coordinate shift $(\delta\bm r)_y$ for an incident wave packet accelerated by electric field $E_x$ and scattered by a gauge impurity. (b) Allowed transition process of Dirac fermions selected by different pseudospin orientation on Fermi surface and various types of disorder. Close to (Far from) the band edge, the pseudospin of electrons is aligned out of (in) the $k_x-k_y$ plane. Blue (red) dots label scalar (red) disorder. (c) Vally Hall conductivity (in units of $e^2/h$) plotted as a function of Fermi energy $\epsilon_F$ in the presence of gauge disorder. Red (blue) curve corresponds to total (side-jump) valley Hall conductivity. $2\Delta$ is the band gap and $m$ is the winding number. $m=1 (m=2)$ for monolayer (bilayer) graphene.[]{data-label="fig:sidejump"}](side_jump.pdf){width="43.00000%"}
To reveal the effect of long-range gauge disorder on the Hall conductivity, we invoke a recently developed semiclassical Boltzmann theory [@sinitsyn2006; @sinitsyn2007]. Such theory has been widely used to investigate Hall-type transport under scalar disorder in various systems, whereas our work for the first time generalizes it to the long-range gauge disorder. Without loss of generality, we consider electrons scattering in the conduction band. The key quantity here is a sudden coordinate shift [@berger1970] experienced by an electron wave packet as it scatters off an impurity (see Fig. \[fig:sidejump\] (a)), given by [@sinitsyn2006] $$\label{shift}
\delta\bm r^c_{\bm k'\bm k} = \bm A^c_{\bm k'} - \bm A^c_{\bm k} - \hat{\bm D}_{\bm k',\bm k} \arg(V_{\bm k'\bm k}^c) \;,$$ where $\bm A^c_{\bm k} = {\langle u_{\bm k}^c|i\bm\nabla_{\bm k}u_{\bm
k}^c \rangle}=-\sin^2\frac{\theta_{\bm k}}{2}(\hat{z}\times\bm k)/k^2$ is the Berry connection of the conduction band and $\hat{\bm D}_{\bm k',\bm k} =
\bm\nabla_{\bm k'} + \bm\nabla_{\bm k}$. The information of the impurity is encoded in the quantity $$V^c_{\bm k'\bm k} = \frac{1}{\cal S}\int d\bm r\,
e^{-i\bm q\cdot\bm r}{\langle u^c_{\bm k'}|V_\text{imp}(\bm r)|u^c_{\bm k} \rangle} \;,$$ where $\cal S$ is the system area and $\bm q\equiv\bm k'-\bm k$ is the momentum transfer of electrons. The overall effect of the disorder is obtained by taking the disorder average, under which the last term of Eq. becomes $${\langle \hat{\bm D}_{\bm k',\bm k}\arg(V_{\bm k'\bm k}^{c}) \rangle}_\text{dis}
=\mathrm{Im}\frac{{\langle V_{\bm k\bm k'}^{c}\hat{\bm D}_{\bm k',\bm k}V_{\bm k'\bm k}^{c} \rangle}_\text{dis}}{{\langle |V_{\bm k'\bm k}^{c}|^2 \rangle}_\text{dis}} \;,$$ where ${\langle \dots \rangle}_\text{dis}$ stands for disorder or thermal average.
To proceed further, it is convenient to write $V^c_{\bm k'\bm k}$ using the chiral basis $$V^{c}_{\bm k'\bm k}=d_-(\bm q){\langle u^c_{\bm k'}|\sigma_+|u^c_{\bm k} \rangle}
+d_+(\bm q){\langle u^c_{\bm k'}|\sigma_-|u^c_{\bm k} \rangle} \;,$$ where $\sigma_{\pm}=(\sigma_x\pm i\sigma_y)/2$. The correlation ${\langle |V_{\bm k'\bm k}^{c}|^2 \rangle}_\text{dis}$ then breaks up into terms with opposite chirality ${\langle d_{\mp}(\bm q)d_{\pm}(-\bm q) \rangle}_\text{dis}$, and terms with the same chirality ${\langle d_{\pm}(\bm q)d_{\pm}(-\bm q) \rangle}_\text{dis}$. The former contain no phase factor, whereas the latter, according to Eq. , is proportional to $\exp(\mp 4i\varphi_{\bm q})$. Note that such term should be small after angular average, and we can neglect the term. The accuracy of this approximation is demonstrated by rigorous numerical analysis in Appendix \[sec:numerical\]. By such approximation, the correlation reduces to $$\begin{split}
\langle|V^{c}_{\bm k'\bm k}|^2\rangle_\text{dis}&=\frac{1}{2}\langle d_-(\bm q)d_+(\bm q)\rangle_\text{dis}\sin^2\theta_{\bm k},
\end{split}$$ where we have used the condition of elastic scattering, i.e., $\theta_{\bm k}=\theta_{\bm k'}$.
This approximation also applies to the correlation ${\langle V_{\bm k\bm k'}^{c}\hat{\bm D}_{\bm k',\bm k}V_{\bm k'\bm k}^{c} \rangle}_\text{dis}$. Making use of the fact that $\hat{\bm D}_{\bm k',\bm k}d_{\pm}(\bm q) = (\bm\nabla_{\bm k} +
\bm\nabla_{\bm k'})d_{\pm}(\bm k'-\bm k) = 0$, we obtain $$\begin{split}\label{r_shift}
\delta\bm r^c_{\bm k'\bm k}&=\frac{\bm\Omega_c(\bm k)\times(\bm k-\bm k')}{\sin^2\theta_{\bm k}}.
\end{split}$$ Similar calculation can be applied to electrons from valence band: $$\begin{split}
\delta\bm r^v_{\bm k'\bm k}&=\frac{\bm\Omega_v(\bm k)\times(\bm k-\bm k')}{\sin^2\theta_{\bm k}}.
\end{split}$$
We can see that in the coordinate shift $\delta\bm r^c_{\bm k'\bm k}$ or $\delta\bm r^v_{\bm k'\bm k}$, the strain-related perfactor $F(\bm q)$ drops out completely, thus this expression is generally applicable for both out-of-plane and in-plane modes of strain fluctuations. Physically, $\delta\bm r_{\bm k'\bm k}$ describes a coordinate shift transverse to the momentum change $\bm k-\bm k'$, leading to a Hall-like current.
It is useful to compare with the short-range scalar disorder [@sinitsyn2007]. In that case, the coordinate shift is given by $$\begin{split}
\delta\bm r^n_{\bm k'\bm k}&=-\frac{\bm \Omega_n(\bm k)}{|\langle u^n_{\bm k'}|u^n_{\bm k}\rangle|^2}\times(\bm k-\bm k'),
\end{split}$$ where $n=c/v$ refers to conduction (valence) band. In the denominator $|\langle u^n_{\bm k'}|u^n_{\bm k}\rangle|^2\approx1$ near the band edge, whereas for gauge disorder in Eq. (\[r\_shift\]) $\sin^2\theta_{\bm k}=4|\langle u^c_{\bm k'}|\sigma_+|u^c_{\bm
k}\rangle|^2\approx0$, which makes a significant difference. Physical meaning for this difference is that close to the band edge, pseudospin of electrons is almost fixed, and the probability of a spin-flipping transition driven by gauge disorder is vanishingly small (see Fig. \[fig:sidejump\] (b)).
Once the coordinate shift $\delta\bm r_{\bm k'\bm k}$ is derived, one can calculate its contribution, known as the side jump, to the valley Hall conductivity. There are two different types of side-jump effects: the direct side-jump contribution $\sigma_{xy}^{direct}$ and the anomalous distribution-induced contribution $\sigma_{xy}^{adist}$ [@sinitsyn2006; @sinitsyn2007]. We can first write down the scattering rate $$\begin{split}
\omega_{\bm k'\bm k}&=\frac{2\pi}{\hbar}\langle|V^{c}_{\bm k'\bm k}|^2\rangle_\text{dis}\delta(\epsilon_{c,\bm k}-\epsilon_{c,\bm k'})
\end{split}$$ and the transport time $$\begin{split}\label{transport}
&\frac{1}{\tau^{tr}}
=\frac{2\pi}{\hbar}\sum_{\bm k'}\langle|V^{c}_{\bm k'\bm k}|^2\rangle_\text{dis}(1-\cos(\phi_{\bm k}-\phi_{\bm k'}))\delta(\epsilon_F-\epsilon_{c,\bm k'}).
\end{split}$$ For monolayer or bilayer graphene, point group symmetry requires that random gauge potential follows $d_{\pm}(\bm q)=F(\bm q)q^2e^{-2i\varphi_{\bm q}}$, which means $\langle d_-(\bm q)d_+(\bm q)\rangle_\text{dis}$ becomes a function of $\bm q=\bm k-\bm k'$. Therefore the transport time $\tau^{tr}$ is isotropic for all $\bm k$ on the Fermi surface.
The coordinate shift $\delta\bm r^c_{\bm k'\bm k}$ leads to an average side-jump velocity $\bm v^{sj}(\bm k)$ $$\begin{split}
&v^{sj}_x(\bm k)=\sum_{\bm k'}\omega_{\bm k'\bm k}(\delta\bm r^c_{\bm k'\bm k})_x=\frac{\cos\theta_{\bm k}}{2k_F}\frac{1}{\tau^{tr}}\sin\phi_{\bm k},
\end{split}$$ where $k_F$ is the Fermi wave vector. In the presence of an external electric field $E_y$, a nonequilibrium correction to the distribution function is given by $$\begin{aligned}
g_{\bm k}=-\frac{\partial n_e}{\partial\epsilon_{c,\bm k}}eE_yv^c_y(\bm k)\tau^{tr}, \end{aligned}$$ where $n_e=1/[\exp((\epsilon_{c,\bm k}-\epsilon_F)/k_BT)+1]$ is the Fermi distribution function, and $v_y^c(\bm k)=v\sin\theta_{\bm k}\sin\phi_{\bm k}$ is the bare velocity of electrons along the electric field. As a result, at $T=0$ K, the valley Hall conductivity $\sigma_{xy}^{direct}$ (for each valley and spin) reads $$\begin{split}
\sigma_{xy}^{direct}&=e\int\frac{d^2\bm k}{(2\pi)^2}\frac{g_{\bm k}}{E_y}v^{sj}_x(\bm k)=\frac{e^2}{4h}\cos\theta_{F}.
\end{split}$$ The physical process can be understood as follows: in the weak disorder limit, the scattering rate $\omega_{\bm k'\bm k}$ is tiny, which gives rise to a small anomalous velocity $v^{sj}_x(\bm k)$ on average during the scattering events. On the other hand, weak disorder means long lifetime, i.e., electrons can be accelerated by an electric field for more time until they are stopped by disorder scattering. This creates a large correction to the Fermi distribution $g_{\bm k}$, that is, more electrons contribute to the transverse transport. As a result, a product of the small anomalous velocity and the large number of electrons lead to a disorder-independent valley Hall conductivity $\sigma_{xy}^{direct}$ as the leading-order term of the disorder potential.
In addition, $\delta\bm r^c_{\bm k'\bm k}$ can cause an anomalous distribution $g^{adist}_{\bm k}$ that also contributes to the Hall current, i.e., $\sigma_{xy}^{adist}$ term. To find its form, let us solve the equation $$\begin{aligned}
\sum_{\bm k'}\omega_{\bm k'\bm k}(g^{adist}_{\bm k}-g^{adist}_{\bm k'}+(-\frac{\partial n_e}{\partial \epsilon_{c,\bm k}})eE_y(\delta\bm r^c_{\bm k'\bm k})_y)=0\end{aligned}$$ to derive the nonequilibrium distribution function $g^{adist}_{\bm k}$. Take the ansatz $g^{adist}_{\bm k}=\gamma_{\bm k}k_x$, we find $$\begin{split}
\gamma_{\bm k}=-(\frac{\partial n_e}{\partial \epsilon_{c,\bm k}})eE_y\frac{\cos\theta_{\bm k}}{2k^2}.
\end{split}$$ Then at $T=0$ K, $\sigma_{xy}^{adist}$ is given by $$\begin{split}
\sigma_{xy}^{adist}&=e\int\frac{d^2\bm k}{(2\pi)^2}\frac{g^{adist}_{\bm k}}{E_y}v_x^c(\bm k)=\frac{e^2}{4h}\cos\theta_{F}.
\end{split}$$
Finally, the total side-jump valley Hall conductivity $\sigma_{xy}^{sj}$ is given by $$\begin{split}\label{result}
\sigma_{xy}^{sj}&=4(\sigma_{xy}^{direct}+\sigma_{xy}^{adist})=\frac{2e^2}{h}\cos\theta_{F},
\end{split}$$ where the factor of 4 counts the valley and spin degeneracy. We notice that $\sigma_{H}^\text{sj}$ reaches its maximum value at the band edge, then decreases gradually as the Fermi energy moves away (see Fig. \[fig:sidejump\] (c)). By Eq. (\[intrins\]) and (\[result\]), the total valley Hall conductivity then becomes $\sigma_{H}^v=\sigma_{H}^\text{int}+\sigma_{H}^\text{sj}=2e^2/h$, which is quantized. We have also obtained the same valley Hall conductivity by adopting a different diagrammatic approach in Appendix \[sec:diagram\]. Intriguingly, this gives us the same result as obtained for short-range gauge disorder using the diagrammatic approach [@yang2011; @yang2011a]. This indicates that valley Hall conductivity when a Dirac electron meets with gauge disorder is actually a universal quantity, which is in striking contrast to the case of scalar disorder. Note that at zero temperature our finding proposes a new type of geometric quantization of the Fermi surface, in contrast to the well-known topological quantization of the Fermi sea. This is one of the main results of our paper.
We may also apply our theory to bilayer graphene, which can be modeled by $$H_\text{BLG} = (\hbar v)^2[(k_x^2-k_y^2)\sigma_x+2k_xk_y\sigma_y] + \Delta\sigma_z \;.$$ After some algebra, we find $\sigma_{H}^\text{sj}=(4e^2/h)\cos\theta_{F}$ and the total valley Hall conductivity $\sigma_H^v = 4e^2/h$. The doubling of $\sigma_H^v$ can be traced back to the phase winding number of $2$ of the electrons around the Dirac point in bilayer graphene. Note that we have neglected the weaker fluctuation of next-nearest-neighbor interlayer coupling in bilayer graphene [@son2011].
Temperature dependence {#temperature}
======================
So far we have studied the valley Hall effect at zero temperature and its dependence on the Fermi energy. Next we will focus on the charge neutral point and consider the dependence of the valley Hall effect on the band gap and temperature. We will consider bilayer graphene systems to compare with the experiment [@sui2015]. If the system is perfectly uniform (chemical potential $\mu=0$), $\sigma_H^v$ due to thermally-activated carriers will be exponentially small. For example, take $2\Delta\sim 100$ meV in a bilayer graphene, then $\sigma_H^v \sim 2 \times 10^{-3} e^2/h$ at $T = 70$ K, several orders of magnitude smaller than experimental values. However, in graphene systems the charge density typically fluctuates due to the formation of electron-hole puddles or gate voltage fluctuation [@dean2010]. By assuming a small residue charge density $\delta n_0$, we can move the chemical potential into the conduction or valence bands according to $$\delta n_0 =\sum_{\bm k}[n_e(\epsilon_{\bm k},\mu,T)-n_h(\epsilon_{\bm k},\mu,T)] \;,$$ where $n_{e/h}(\epsilon_{\bm k},\mu,T)=1/[\exp(\pm(\epsilon_{c/v,\bm k}-\mu)/k_BT)+1]$ is the Fermi distribution function for electrons (holes) and $\mu$ is the chemical potential. At finite temperatures, both electrons and holes contribute to the transport. Including all these effects, we find the extrinsic and intrinsic valley Hall conductivity $$\begin{split}\label{temperature2}
&\sigma_{xy}^{direct}=\sigma_{xy}^{adist}=\frac{e^2}{4h}\frac{\Delta}{k_BT}\\
&\times\int^{\infty}_{\Delta/k_BT}dx[\frac{e^{x-\frac{\mu}{k_BT}}}{(1+e^{x-\frac{\mu}{k_BT}})^2}\frac{1}{x}+\frac{e^{x+\frac{\mu}{k_BT}}}{(1+e^{x+\frac{\mu}{k_BT}})^2}\frac{1}{x}],\\
&\sigma_{xy}^{int}=\frac{e^2}{2h}\frac{\Delta}{k_BT}\\
&\times\int^{\infty}_{\Delta/k_BT}dx[\frac{1}{1+e^{x-\frac{\mu}{k_BT}}}\frac{1}{x^2}+\frac{1}{1+e^{x+\frac{\mu}{k_BT}}}\frac{1}{x^2}],
\end{split}$$ where the two terms in the square brackets originate from electrons and holes, respectively. For monolayer graphene, we obtain a much enhanced $\sigma_H^v$: $$\begin{split}
&\sigma_{H}^{v}=4[\sigma_{xy}^{direct}+\sigma_{xy}^{adist}+\sigma_{xy}^{int}]\\
&=\frac{2e^2}{h}[\frac{1}{1+e^{(\Delta-\mu)/k_BT}}+\frac{1}{1+e^{(\Delta+\mu)/k_BT}}].
\end{split}$$ For bilayer graphene, we will acquire an extra factor 2, as compared to monolayer case.
Figure \[fig:temperature\] shows the calculated $\sigma_H^v$ as a function of temperature for $\delta n_0\sim1.0\times10^{10}$ cm$^{-2}$, a value well below the density resolution of the nonlocal peak in the experiment [@gorbachev2014]. The downwards trend at medium temperature range $T\sim60-80$ K reproduces precisely the temperature dependence in bilayer graphene reported in one of the experiments (see Supplementary Figure 9b of reference [@sui2015]). Another trend found in the experiment [@sui2015], i.e., smaller $\sigma_H^v$ with larger gap, is also reproduced. At higher temperature we find that $\sigma_H^v$ will increase again, although there is no available experimental data to compare with. As far as we know, there are no counterparts of such peculiar temperature behaviors reported in anomalous Hall physics [@nagaosa2010; @xiao2018]. The origin of such non-monotonic temperature behavior is the competition between intrinsic and side-jump contributions under thermal activation. The intrinsic contribution, as a summation of Berry curvature over occupied states, favors high-temperature regime with more occupied states, whereas the side-jump contribution, suffering from a suppression of thermally activated carriers by $1/k_BT$, favors low-temperature regime, as shown in Eq. (\[temperature2\]). Nevertheless, neither a gap nor temperature dependence of valley Hall conductivity has been discussed by previous theoretical papers [@lensky2015; @kirczenow2015; @li2011; @zhu2017; @song2018; @brown2018], and our work represents a first step towards understanding these peculiar behaviors. Moreover, the newly predicted non-monotonic temperature dependence can be used as a test for our theory.
![ Temperature dependence of valley Hall conductivity $|\sigma_{H}^v|$ (in units of $e^2/h$) of bilayer graphene for $\Delta=50$, $40$, $30$meV, respectively. Parameters: $\delta n_0=1.0\times10^{10}$cm$^{-2}$.[]{data-label="fig:temperature"}](VHC_new.pdf){width="38.00000%"}
![ Magnitude of valley Hall conductivity $|\sigma_{H}^v|$ (in units of $e^2/h$) as functions of ratio $\tau_{sc}/\tau_a$ at $T=0$ K. $\tau_{sc}$, $\tau_a$ refer to relaxation time for scalar and gauge disorder scattering, respectively. Larger $\tau_{sc}/\tau_a$ implies more gauge disorder (high mobility sample); smaller implies more scalar charge disorder (low mobility sample). Red solid (cyan dashed) curve corresponds to doping density $n=0$ ($n=10^{10}$ cm$^{-2}$), whose Fermi surface is indicated in the inset. Parameters are adopted for monolayer graphene: $\Delta=20$ meV [@gorbachev2014], and $v=10^6$ m/s. []{data-label="fig:valleyhall"}](VHC4.pdf){width="45.00000%"}
Numerical analysis {#numerics}
==================
A crucial approximation we made in the semiclassical approach is assuming that $ {\langle d_{\pm}(\bm q)d_{\pm}(-\bm q) \rangle}_\text{dis} \propto \exp(\mp
4i\varphi_{\bm q})$ is negligible after angular average. The accuracy of this approximation needs to be examined since it introduces a long-range anisotropic correlation function. The effect of such correlation function is unclear yet, and an analytical treatment seems impractical. To this end, in the following part we develop a new numerical approach which treats the correlation function rigorously in Appendix \[sec:numerical\]. We consider a particular type of random strain, i.e., out-of-plane corrugations with $\bm d(\bm q)=g_1|\bm q|^{-4}\mathcal{F}(\bm
q)(q_x^2-q_y^2,-2q_xq_y)$, where $\langle\mathcal{F}(\bm q)\mathcal{F}(-\bm q)\rangle_\text{dis}$ $\propto|\bm q|^{2}$, and $g_{1}$ quantifies the electron-phonon coupling strength in graphene [@couto2014]. We find that $\sigma_{H}^{sj}$ is very close to $ \frac{2me^2}{h}\cos\theta_{F}$ ($m=1$ for monolayer; $m=2$ for bilayer), differing by $1.69\%$ for monolayer graphene and $0.91\%$ for bilayer graphene. Such small deviation from the quantized values justifies our previous treatment of dropping the fast-oscillating part $\langle d_{\pm}(\bm q)d_{\pm}(-\bm q)\rangle_\text{dis}$.
Effect of long-range scalar potential {#scalar}
=====================================
Besides the gauge potential $\bm d(\bm q)$, a long-range scalar potential may also arise due to random strain fluctuations. For out-of-plane corrugations [@couto2014], the scalar potential is given by $V(\bm q)=g_2|\bm q|^{-2}\mathcal{F}(\bm q)$, where $g_{2}$ is an electron-phonon coupling parameter. Different from the gauge potential $\bm d(\bm q)$, which represents a pseudomagnetic field, $V(\bm q)$ is electrostatic in nature and thus it needs to be screened by the static dielectric function $\epsilon(\bm q)=1+V_c(\bm q)N_F$ within the Thomas-Fermi approximation [@ando2006; @dassarma2011], where $N_F$ is the density of states, $V_c(\bm q)=2\pi e^2/\epsilon_0|\bm q|$ is the two-dimensional Coulomb potential, and $\epsilon_0$ is the background (including the substrate) dielectric constant. Here we will focus on the strong screening regime (see Appendix \[sec:screen\]): $\epsilon_0\ll e^2N_F/k_F$, which characterizes the charge neutral point of gapped graphene systems.
Taking into account both the gauge and scalar potentials, we find that the total valley Hall conductivity of a gapped graphene at $T=0$ K is (see Appendix \[sec:mix\]): $$\label{scalar_pot}
\sigma_{H}^{v} =\frac{2e^2}{h}[1-\frac{(4+\sin^2\theta_{F})\cos\theta_{F}}{\frac{4}{w_0}\frac{\tau_{sc}}{\tau_a}+4-3\sin^2\theta_{F}}].$$ Here $w_0=\sqrt{(k_FW/\pi)^2-1}/\pi$ is a cutoff-related factor with the device width $W$. For $W=1$ $\mu$m and residue charge density $\delta n_0=1.0\times10^{10}$ cm$^{-2}$, we estimate $w_0\approx2.2$. The ratio $\tau_{sc}/\tau_a=(2e^2g_1/\hbar\epsilon_0g_2)^2w_0$ defines the relative strength between the scalar and gauge disorder. In the limit $\tau_{sc}/\tau_a\rightarrow\infty$, i.e., in the absence of scalar disorder, $\sigma_H^v$ reduces to $2e^2/h$. Figure \[fig:valleyhall\] shows the calculated $\sigma_H^v$ as a function of $\tau_{sc}/\tau_a$ at two different Fermi energies. From this figure, one can immediately understand why nonlocal signal is only measured in high-quality graphene on hBN rather than on SiO$_2$ substrate [@zyb2015], since in the former (latter) case gauge (scalar) disorder is the dominant source of disorder scattering, corresponding to the limit $\tau_{sc}/\tau_a\rightarrow\infty$ ($\tau_{sc}/\tau_a\rightarrow0$).
Conclusion and discussion {#conclusion}
=========================
We have provided an alternative scenario to understand the large valley Hall conductivity observed in experiments, based on scattering from random strain fluctuations. The origin is intimately related to an enhanced coordinate shift under gauge disorder scattering in Dirac systems. Temperature and gap dependence is qualitatively reproduced. Our work paves the way for studying the effect of classical strain modes, or phonon modes [@zhang2015], on the transverse transport of broad classes of 2D materials and van der Waals heterostructures.
A few remarks are in order. Note that our theory is only valid for the band transport regime; the valley Hall effect in the phonon-assisted variable-range-hopping regime [@zou2010] still remains an open question. In addition, we would like to point out that a direct comparison of our result to the experiments requires a careful extraction of the valley Hall conductivity from the nonlocal measurement [@gorbachev2014; @sui2015; @shimazaki2015]. In particular, an accurate determination of the valley diffusion length $\ell_v$ is crucial [@sui2015] since the nonlocal signals depend on $\ell_v$ exponentially [@abanin2009]. Then the “smoking gun" validation of our theory, i.e., a non-monotonic temperature behaviors of valley Hall conductivity, can be examined. As a supplement, weak-localization magneto-resistance measurement [@couto2014] and Raman spectroscopy [@neumann2015] can also be used to uncover the role of strain fluctuations. A complete understanding of the valley Hall effect thus requires further experimental and theoretical efforts.
A universal transverse Imbert-Fedorov (IF) shift of electrons was also discovered in Weyl semimetals [@jiang2015; @yang2015]. The universal coordinate shift proposed in our work differs from the IF shift in two ways. First, the IF shift appears in the three-dimensional massless Weyl fermions, and has no counterpart in 2D, while the coordinate shift is relevant to the 2D massive Dirac fermions. Second, the IF shift occurs at normal interface, while the coordinate shift becomes universal only when scattered by gauge disorder. Appealingly, the considerations in this work can be generalized to describe other 2D systems with topological properties, such as superconductors [@yu2018], excitons [@onga2017], plasmons [@shili2018], polaritons [@gutierrez2018], or under magnetic field [@komatsu2018].
Recently, experimental observations of valley Hall transport have also been made in atomically thin MoS2 systems [@wu2018; @hung2019]. The fact that monolayer and trilayer MoS2 share qualitatively similar behaviors of valley Hall signals [@wu2018] indicates that the details of Berry curvature distribution in the conduction bands may have little influence on the final result. This is actually consistent with our theoretical prediction of the article.
Acknowledgements {#acknowledgements .unnumbered}
=================
We are grateful to Qian Niu, Shengyuan A. Yang, Mengqiao Sui, and Yuanbo Zhang for stimulating discussions. This work is supported by DOE BES Pro-QM EFRC (DE-SC0019443). W. Y. S. also acknowledges the support of a startup grant from Guangzhou University.
Diagrammatic approach to mixed gauge and scalar disorder {#sec:diagram}
========================================================
In this section we present a full quantum mechanical treatment using the diagrammatic approach [@sinitsyn2007; @shan2013] to study the valley Hall effect in the presence of out-of-plane corrugations [@couto2014]. The purpose of this section is two fold. The diagrammatic approach provides an additional check of the semiclassical result. In addition, this approach is systematic, and can be applied to mixed gauge and scalar disorder. It is also convenient for numerical calculations when we consider the anisotropic long-range correlation functions. In the following we will focus on monolayer graphene.
Correlation function
--------------------
For out-of-plane corrugation the induced vector potential $\bm d(\bm q)$ and scalar potential $V(\bm q)$ are given by [@couto2014] $$\begin{split}\label{strain}
\bm d(\bm q)&=g_1\frac{1}{|\bm q|^4}\mathcal{F}(\bm q)(q_x^2-q_y^2,-2q_xq_y) \;, \\
V(\bm q)&=g_2\frac{1}{|\bm q|^2}\mathcal{F}(\bm q) \;,
\end{split}$$ where $$\begin{aligned}
\label{strain2}
\mathcal{F}(\bm q)&=-\int d\bm q_1h(\bm q_1)h(\bm q-\bm q_1)(\bm q\times\bm q_1)^2 \;.\end{aligned}$$ $h(\bm q)$ is the Fourier transform of the height field $h(\bm r)$, and $g_{1,2}$ quantify the electron-phonon coupling strength in graphene. We assume the height correlation is $\langle h(\bm q)h(-\bm q)\rangle_\text{dis}\propto|\bm q|^{-4}$, from which one finds $\langle\mathcal{F}(\bm q)\mathcal{F}(-\bm q)\rangle_\text{dis}=C|\bm q|^2$, where $C$ is a material-dependent parameter. The Born scattering amplitude is given by $$\begin{split}
U_{cc}^A(\bm q)&=\int\frac{d\bm r}{S}e^{-i\bm q\cdot\bm r}\langle u_{\bm k'}^c|\bm d(\bm r)\cdot\bm\sigma|u_{\bm k}^c\rangle\\
&=\frac{\sin\theta_{\bm k}}{2S}[d_-(\bm q)e^{i\phi_{\bm k}}+d_+(\bm q)e^{-i\phi_{\bm k'}}],\\
U_{cc}^V(\bm q)&=\int\frac{d\bm r}{S}e^{-i\bm q\cdot\bm r}\langle u^c_{\bm k'}|V(\bm r)|u^c_{\bm k}\rangle\\
&=\frac{V(\bm q)}{S}[\cos^2\frac{\theta_{\bm k}}{2}+\sin^2\frac{\theta_{\bm k}}{2}e^{i(\phi_{\bm k}-\phi_{\bm k'})}],
\end{split}$$ where $d_{\pm}(\bm q)=g_1q^2_{\mp}\mathcal{F}(\bm q)/|\bm q|^4$ and $V(\bm q)=g_2\mathcal{F}(\bm q)/|\bm q|^2$. The identity $\bm q=\bm k'-\bm k$ leads to a useful relation $q_{\pm}=k(e^{\pm i\phi_{\bm k'}}-e^{\pm i\phi_{\bm k}})$. The correlation functions are $$\begin{aligned}
&\langle U_{cc}^A(\bm q)U_{cc}^A(-\bm q)\rangle_\text{dis}=\frac{g_1^2\sin^2\theta_{\bm k}}{4S^2}\langle\mathcal{F}(\bm q)\mathcal{F}(-\bm q)\rangle_\text{dis}\nonumber\\
&\times[\frac{q_+^4}{|\bm q|^8}e^{i(\phi_{\bm k}+\phi_{\bm k'})}+\frac{q_-^4}{|\bm q|^8}e^{-i(\phi_{\bm k}+\phi_{\bm k'})}+\frac{2}{|\bm q|^4}], \label{correlation1} \\
&\langle U_{cc}^V(\bm q)U_{cc}^V(-\bm q)\rangle_\text{dis}=\frac{g_2^2}{|\bm q|^4S^2}\langle\mathcal{F}(\bm q)\mathcal{F}(-\bm q)\rangle_\text{dis}\nonumber\\
&\times(\cos^4\frac{\theta_{\bm k}}{2}+\sin^4\frac{\theta_{\bm k}}{2}+2\cos^2\frac{\theta_{\bm k}}{2}\sin^2\frac{\theta_{\bm k}}{2}\cos(\phi_{\bm k}-\phi_{\bm k'})).\label{correlation2}\end{aligned}$$ for gauge and scalar disorder, respectively. Similar to what we did in the main text, we will ignore the first two terms in the square bracket on the right-hand side of Eq. (\[correlation1\]) and (\[correlation2\]). These terms originate from $\langle d_{\pm}(\bm q)d_{\pm}(\bm q)\rangle_\text{dis}$, whose contributions are vanishingly small and can be neglected. The validity of this approximation will be demonstrated in the next section. Under this approximation, the correlation function for gauge disorder becomes $$\begin{aligned}
\langle U_{cc}^A(\bm q)U_{cc}^A(-\bm q)\rangle_\text{dis}&=\frac{g_1^2\sin^2\theta_{\bm k}}{2S^2|\bm q|^4}\langle\mathcal{F}(\bm q)\mathcal{F}(-\bm q)\rangle_\text{dis}.\label{correlation3}\end{aligned}$$
Relaxation time and longitudinal conductivity {#sec:screen}
---------------------------------------------
We can use the correlation function in Eq. (\[correlation2\]) and (\[correlation3\]) to derive the relaxation time. According to Fermi’s golden rule, the relaxation time for gauge disorder reads $$\begin{split}
\frac{1}{\tau_A}&=\frac{2\pi}{\hbar}\sum_{\bm k'}\langle U_{cc}^A(\bm q)U_{cc}^A(-\bm q)\rangle_\text{dis}\delta(\epsilon_F-\epsilon_{c,\bm k'})\\
&=\frac{\pi N_FCg_1^2}{2S\hbar k_F^2}w_0\sin^2\theta_{\bm k},\label{tau}
\end{split}$$ where $N_F=\epsilon_F/2\pi\hbar^2v^2$ is the density of states per spin and valley. The cutoff factor $w_0$, given by $$\begin{aligned}
\label{omega0}
w_0&=\int_{\phi_0}^{2\pi-\phi_0}\frac{d\phi_{\bm k}}{2\pi}\frac{1}{1-\cos\phi_{\bm k}}=\frac{\cot\frac{\phi_0}{2}}{\pi} \;,\end{aligned}$$ is introduced to remove the divergence at small momentum [@vozmediano2010], whose physical origin is due to the finite-size effect of samples. Consider a sample with width $W$, then $|\bm q|\geq q_0=\frac{\pi}{W}$. This corresponds to a cutoff angle $\phi_0=2\arcsin(q_0/2k_F)$, and hence $w_0=\sqrt{(k_FW/\pi)^2-1}/\pi$.
On the other hand, for scalar potential, we need to take into account the screening effect. Based on the Thomas-Fermi approximation [@ando2006; @dassarma2011], we can replace $U^V(\bm q)$ by $U^V(\bm q)/\epsilon(\bm q)$, where $\epsilon(\bm q)$ is the dielectric function, $\epsilon(\bm q)=1+g_sg_vV_c(\bm q)N_F$, $V_c(\bm q)=2\pi e^2/\epsilon_0|\bm q|$ is the two-dimensional Coulomb potential, $\epsilon_0$ is the background (including the substrate) dielectric constant, and $g_s=2$ $(g_v=2)$ refers to the spin (valley) degeneracy. Now we can write down the relaxation time $$\begin{split}
&\frac{1}{\tau_V}=\frac{2\pi}{\hbar}\sum_{\bm k'}\frac{\langle U_{cc}^V(\bm q)U_{cc}^V(-\bm q)\rangle_\text{dis}}{\epsilon^2(\bm q)}\delta(\epsilon_F-\epsilon_{c,\bm k'}) \\
&=\frac{2\pi g_2^2C}{\hbar S}N_F\int\frac{d\varphi_{\bm k'}}{2\pi}\frac{\cos^4\frac{\theta_{\bm k}}{2}+\sin^4\frac{\theta_{\bm k}}{2}+2\cos^2\frac{\theta_{\bm k}}{2}\sin^2\frac{\theta_{\bm k}}{2}\cos\varphi_{\bm k'}}{(\sqrt{2}k_F\sqrt{1-\cos\varphi_{\bm k'}}+\frac{8\pi e^2N_F}{\epsilon_0})^2} \\
&\approx\frac{\epsilon_0^2g_2^2C}{32\pi\hbar Se^4N_F}(1-\frac{1}{2}\sin^2\theta_{\bm k}),\ \ \epsilon_0\ll e^2N_F/k_F.
\end{split}$$ In the last step of above derivation, we have taken the strong-screening limit: $\epsilon_0\ll e^2N_F/k_F$. The physics around the charge neutral point belongs to this limit, within which the correlation effectively becomes a short-range one.
Next we need to figure out the modified velocity $(\tilde{v}_{x}^{\bm k})_{cc}$ due to the intraband vertex correction. According to Fig \[fig:diagram\], we have $$\begin{aligned}
(\tilde{v}_{x}^{\bm k})_{cc}&=(v_x^{\bm k})_{cc}\nonumber\\
&+\sum_{\bm k'}\langle U_{cc}^{A}(\bm q)U_{cc}^{A}(-\bm q)\rangle_\text{dis} G^r_{c,\bm k'}G^a_{c,\bm k'}(\tilde{v}_{x}^{\bm k'})_{cc} \label{vertex1}\end{aligned}$$ and $$\begin{aligned}
(\tilde{v}_{x}^{\bm k})_{cc}&=(v_x^{\bm k})_{cc}\nonumber\\
&+\sum_{\bm k'}\frac{\langle U_{cc}^{V}(\bm q)U_{cc}^{V}(-\bm q)\rangle_\text{dis}}{\epsilon^2(\bm q)}G^r_{c,\bm k'}G^a_{c,\bm k'}(\tilde{v}_{x}^{\bm k'})_{cc}\label{vertex2}\end{aligned}$$ for gauge and scalar disorder, respectively. Here the bare velocity is given by $(v_x^{\bm k})_{cc}\equiv\langle u^c_{\bm k}|\hat{v}_x|u^c_{\bm k}\rangle=v\sin\theta_{\bm k}\cos\varphi_{\bm k}$, and retarded (advanced) Green’s function is $G_{c,\bm k}^{r/a}=1/(\epsilon_F-\epsilon_{c,\bm k}\pm i\hbar/2\tau_j)$, with $j=A,V$. To solve the equations, we take the ansatz $(\tilde{v}_{x}^{\bm k})_{cc}=\eta(v_{x}^{\bm k})_{cc}$ and substitute it into Eqs. (\[vertex1\]) and (\[vertex2\]). We find $\eta=w_0$ for gauge disorder and $\eta=\frac{4(1-\frac{1}{2}\sin^2\theta_{\bm k})}{1+3\cos^2\theta_{\bm k}}$ for scalar disorder, respectively. Based on this, we can obtain the transport time $$\begin{split}
\tau^{tr}_A&\equiv\eta\tau_A=\frac{4S\hbar \epsilon_F}{Cg_1^2},\\
\tau^{tr}_V&\equiv\eta\tau_V\approx\frac{128\pi\hbar Se^4N_F}{\epsilon_0^2g_2^2C}\frac{1}{1+3\cos^2\theta_{\bm k}},
\end{split}$$ which agree with previous results by using the Boltzmann approach [@couto2014].
![ (a) Ladder diagram correction to the velocity vertex. (b)-(e) Diagrams corresponding to the side jump contribution to the Hall conductivity. $\tilde{v}_x^{\bm k}$ ($v_x^{\bm k}$) refers to the modified (unmodified) velocity. Red (blue) solid line corresponds to the retarded (advanced) Green’s function and dashed line represents the disorder-averaged correlation function.[]{data-label="fig:diagram"}](diagram_strain1){width="40.00000%"}
Valley Hall conductivity {#sec:mix}
------------------------
For scalar disorder in the strong-screening limit $\epsilon_0\ll e^2N_F/k_F$, the correlation function $\langle U^V(\bm q)U^V(-\bm q)\rangle_\text{dis}/\epsilon^2(\bm q)$ exhibits a short-range behavior, whose extrinsic valley Hall conductivity has been obtained before [@sinitsyn2007; @yang2011]: $$\begin{split}
\sigma_{xy}=-\frac{2e^2}{h}\frac{\sin^2\theta_{\bm k}\cos\theta_{\bm k}}{1+3\cos^2\theta_{\bm k}}
\end{split}$$ for each spin and valley. In this expression, higher-order side-jump or skew-scattering contributions have been ignored since we are interested in the low doping regime. Next we turn to the gauge disorder, which has not been studied before. Fig. \[fig:diagram\] (b)-(e) correspond to the leading-order side-jump contributions to the valley Hall conductivity in the weak scattering limit, given by $$\begin{split}\nonumber
\sigma_{xy}^{b,i}&=\frac{e^2\hbar}{2\pi S}\sum_{\bm k,\bm k'}(\tilde{v}_{x}^{\bm k})_{cc}(v_{y}^{\bm k})_{cv}\\
&\times G^r_{c,\bm k}G^a_{c,\bm k}G^a_{v,\bm k}G^a_{c,\bm k'}\langle U_{vc}^A(-\bm q)U_{cc}^A(\bm q)\rangle_\text{dis},\\
\sigma_{xy}^{c,i}&=\frac{e^2\hbar}{2\pi S}\sum_{\bm k,\bm k'}(\tilde{v}_{x}^{\bm k})_{cc}(v_{y}^{\bm k})_{vc}\\
&\times G^r_{c,\bm k}G^r_{v,\bm k}G^r_{c,\bm k'}G^a_{c,\bm k}
\langle U_{cv}^A(\bm q)U_{cc}^A(-\bm q)\rangle_\text{dis},\\
\sigma_{xy}^{d,i}&=\frac{e^2\hbar}{2\pi S}\sum_{\bm k,\bm k'}(\tilde{v}_{x}^{\bm k})_{cc}(v_{y}^{\bm k'})_{cv}\\
&\times G^r_{c,\bm k}G^a_{c,\bm k}G^r_{c,\bm k'}G^a_{v,\bm k'}\langle U_{vc}^A(\bm q)U_{cc}^A(-\bm q)\rangle_\text{dis},
\end{split}$$
$$\begin{split}\nonumber
\sigma_{xy}^{e,i}&=\frac{e^2\hbar}{2\pi S}\sum_{\bm k,\bm k'}(\tilde{v}_{x}^{\bm k})_{cc}(v_{y}^{\bm k'})_{vc}\\
&\times G^r_{c,\bm k}G^a_{c,\bm k}G^r_{v,\bm k'}G^a_{c,\bm k'}\langle U_{cv}^A(-\bm q)U_{cc}^A(\bm q)\rangle_\text{dis}.
\end{split}$$
By making use of $[U^i_{nn'}(\bm q)]^*=U^i_{n'n}(-\bm q)$, $i=A,V$, $n,n'=c,v$, we find the following symmetry properties: $$\begin{aligned}
\sigma_{xy}^{c,i}&=[\sigma_{xy}^{b,i}]^*,\ \ \sigma_{xy}^{e,i}=[\sigma_{xy}^{d,i}]^*.\end{aligned}$$ After some algebra, we find that for gauge disorder $$\begin{split}
\sigma_{xy}^{b,A}+\sigma_{xy}^{c,A}&=\frac{e^2}{4h}w_0\cos\theta_{\bm k}, \\
\sigma_{xy}^{d,A}+\sigma_{xy}^{e,A}&=\frac{e^2}{4h}(1-w_0)\cos\theta_{\bm k},
\end{split}$$ and thus $$\begin{aligned}
\sigma_{xy}^{b,A}+\sigma_{xy}^{c,A}+\sigma_{xy}^{d,A}+\sigma_{xy}^{e,A}&=\frac{e^2}{4h}\cos\theta_{\bm k}.\end{aligned}$$ Moreover, there are equivalent contributions from diagrams by rotating Fig. \[fig:diagram\] (b)-(e) by $180^{\circ}$, then exchanging the subscript $x$, $y$ [@sinitsyn2007]. Therefore the total extrinsic valley Hall conductivity (including spin degeneracy) reads $\frac{2e^2}{h}\cos\theta_{\bm k}$, which reproduces Eq. (\[result\]) by applying the semiclassical approach.
Furthermore we can study the situation with mixed gauge and scalar disorder. In this case, the total relaxation time is given by $$\begin{split}
\frac{1}{\tau}&=\frac{1}{\tau_A}+\frac{1}{\tau_V}.
\end{split}$$ For convenience, we can introduce a Fermi-energy-independent relaxation time $\tau_a$, $\tau_{sc}$ for gauge and scalar disorder, respectively: $$\begin{split}
\tau_A&\equiv\frac{\tau_a}{\cos\theta_{\bm k}},\ \ \tau_a=\frac{4S\hbar\Delta}{Cg_1^2w_0}, \\
\tau_V&\equiv\frac{\tau_{sc}}{\cos\theta_{\bm k}(1-\frac{1}{2}\sin^2\theta_{\bm k})},\ \ \tau_{sc}=\frac{16Se^4\Delta}{\epsilon_0^2g_2^2C\hbar v^2}.
\end{split}$$ For such mixed disorder, we can follow the same procedure as above and derive $$\begin{split}
&\sigma_{xy}^{b}+\sigma_{xy}^{c}=-\frac{e^2}{4h}\eta\cos^2\theta_{\bm k}[-\frac{\tau}{\tau_a}+\frac{\tau}{2\tau_{sc}}\sin^2\theta_{\bm k}],\\
&\sigma_{xy}^{d}+\sigma_{xy}^{e}=-\frac{e^2}{4h}\eta\cos^2\theta_{\bm k}
[\frac{\tau}{\tau_a}(1-\frac{1}{w_0})+\frac{\tau}{2\tau_{sc}}\sin^2\theta_{\bm k}],\\
&\sigma_{xy}^{b}+\sigma_{xy}^{c}+\sigma_{xy}^{d}+\sigma_{xy}^{e}\\
&=-\frac{e^2}{4h}\eta\cos^2\theta_{\bm k}
[-\frac{\tau}{\tau_a}\frac{1}{w_0}+\frac{\tau}{\tau_{sc}}\sin^2\theta_{\bm k}],
\end{split}$$ where the correction factor $$\begin{aligned}
\eta&=\frac{1}{1-\frac{\tau}{\tau_a}(1-\frac{1}{w_0})\cos\theta_{\bm k}-\frac{\tau}{4\tau_{sc}}\cos\theta_{\bm k}\sin^2\theta_{\bm k}}.\end{aligned}$$ Finally, by adding the intrinsic term, we find the total contribution is $$\begin{aligned}
\sigma_{xy}^{v}&=\frac{2e^2}{h}[1-\frac{(4+\sin^2\theta_{\bm k})\cos\theta_{\bm k}}{\frac{4}{w_0}\frac{\tau_{sc}}{\tau_a}+4-3\sin^2\theta_{\bm k}}],\end{aligned}$$ which gives Eq. (\[scalar\_pot\]) in the main text.
\[sec:numerical\]Numerical treatment of anisotropic correlation function
========================================================================
In our derivation, we have made use of the approximation that terms $\langle d_{\pm}(\bm q)d_{\pm}(\bm q)\rangle_\text{dis}$ in the correlation function vanish after angular average. In this section, we test this approximation by treating the correlation function exactly, i.e., keeping all the terms in Eq. (\[correlation1\]) and (\[correlation2\]). We focus on a generic type of gauge disorder satisfying $\langle\mathcal{F}(\bm q)\mathcal{F}(-\bm q)\rangle_\text{dis}=C|\bm q|^{2\epsilon+2}$, where the value of $\epsilon$ depends on the microscopic details. For example, $\epsilon=0$ and $1$ corresponds to thermally excited and substrate-induced ripples, respectively [@vozmediano2010]. We also consider a generic chiral model $$\begin{aligned}
H=\begin{pmatrix}
\Delta & Ak_-^m \\
Ak_+^m & -\Delta \\
\end{pmatrix},\end{aligned}$$ where $m=1$ and $2$ correspond to monolayer and bilayer graphene, respectively. This leads to the eigenvalue and eigenstates $$\begin{split}
\epsilon_{c/v,\bm k}&=\pm\sqrt{\Delta^2+A^2k^{2m}},\\
|u^c_{\bm k}\rangle&=\left(\begin{array}{cc}
\cos\frac{\theta_{\bm k}}{2} \\
\sin\frac{\theta_{\bm k}}{2}e^{im\phi_{\bm k}} \\
\end{array}\right),\ \ |u^v_{\bm k}\rangle=\left(\begin{array}{cc}
\sin\frac{\theta_{\bm k}}{2} \\
-\cos\frac{\theta_{\bm k}}{2}e^{im\phi_{\bm k}} \\
\end{array}\right),
\end{split}$$ where $\cos\theta_{\bm k}=\Delta/\epsilon_{c,\bm k}$, $\sin\theta_{\bm k}=Ak^m/\epsilon_{c,\bm k}$. The density of states is given by $N_F=\epsilon_F/(2\pi mA^2k_F^{2m-2})$. We can write down a complete form of the correlation function $$\begin{split}
&\langle U_{cc}^A(\bm q)U_{cc}^A(-\bm q)\rangle_\text{dis}=\frac{Cg_1^2}{4S^2|\bm q|^{6-2\epsilon}}\sin^2\theta_{\bm k}\\
&\times[k^4e^{im(\phi_{k'}-\phi_k)}e^{i(4+2m)\phi_{\bm k}}(e^{i(\phi_{\bm k'}-\phi_k)}-1)^4\\
&+k^4e^{-im(\phi_{k'}-\phi_k)}e^{-i(4+2m)\phi_{\bm k}}(e^{-i(\phi_{\bm k'}-\phi_k)}-1)^4+2q^4].
\end{split}$$ For monolayer graphene with thermal ripples: $m=1$, $\epsilon=0$, it reduces to Eq. (\[correlation1\]). Similar to Eq. (\[tau\]), we can evaluate the relaxation time by $$\begin{split}
\frac{1}{\tau_{\bm k}}&=\frac{\pi N_FCg_1^2}{2^{1-\epsilon}S\hbar k^{2-2\epsilon}}\sin^2\theta_{\bm k}[w_{m+2}\cos(4+2m)\phi_{\bm k}+w_0],
\end{split}$$ where a cutoff factor is introduced $$\begin{aligned}
\label{omegam}
w_{m}&=\int_{\phi_0}^{2\pi-\phi_0}\frac{d\phi_{\bm k}}{2\pi}\frac{\cos m\phi_{\bm k}}{(1-\cos\phi_{\bm k})^{1-\epsilon}}.\end{aligned}$$ For such anisotropic problem, it is convenient to establish a self-consistent equation for the mean free path $(L_{x}^{\bm k})_{cc}$ [@Tokura1998] $$\begin{split}\label{self}
\frac{1}{\tau_{\bm k}}(L_{x}^{\bm k})_{cc}&=(v_x^{\bm k})_{cc}\\
&+\frac{2\pi N_FS}{\hbar}\int\frac{d\phi_{\bm k'}}{2\pi}\langle U_{cc}^{A}(\bm q)U_{cc}^{A}(-\bm q)\rangle_\text{dis}(L_{x}^{\bm k'})_{cc}.
\end{split}$$ Since the bare velocity follows $(v_x^{\bm k})_{cc}=\frac{mA}{\hbar}k^{m-1}\sin\theta_{\bm k}\cos\phi_{\bm k}$, we can take the following ansatz, $$\begin{aligned}
(L_{x}^{\bm k})_{cc}&=\frac{2^{1-\epsilon}k^{1-2\epsilon+m}S}{\pi N_FCg_1^2\sin\theta_{\bm k}}mA(\sum_{n=0}^{\infty}f_n\cos n\phi_{\bm k}).\end{aligned}$$ By substituting it into Eq. (\[self\]), we find that the coefficients $f_n$ satisfy $$\begin{aligned}
\left(\begin{array}{cccccc}
f_1 \\
f_{2m+3} \\
f_{2m+5} \\
f_{4m+7} \\
f_{4m+9} \\
\vdots
\end{array}\right)=T^{-1}
\left(\begin{array}{cccccc}
2 \\
0 \\
0 \\
0 \\
0 \\
\vdots
\end{array}\right),\end{aligned}$$ where the matrix $T$ has a non-closed form
$$\begin{aligned}
T=\begin{pmatrix}
2(w_0-w_1) & w_{m+2}-w_{m+1} & w_{m+2}-w_{m+3} & & & & \\
\\
& 2(w_0-w_{2m+3}) & 0 & w_{m+2}-w_{3m+5} & & 0 &\\
\\
& & 2(w_0-w_{2m+5}) & 0 & w_{m+2}-w_{3m+7} & & \\
\\
& * & & 2(w_0-w_{4m+7}) & 0 & w_{m+2}-w_{5m+9} & \\
\\
& & & & \ddots & \ddots & \ddots \\
\end{pmatrix}.\end{aligned}$$
Fortunately, numerics indicates that the result converges really fast. Base on this observation, we find that the total valley Hall conductivity become $$\begin{split}
8\sum_{i=b,c,d,e}\sigma_{xy}^{i,A}&=\frac{2me^2}{h}\cos\theta_{\bm k}\\
&\times[f_1(w_0-w_1)+f_{2m+3}(w_{m+2}-w_{m+1})].
\end{split}$$ To gain some insight, we consider two special cases that may be relevant to experiments.
$\epsilon=0$
------------
First we consider $\epsilon=0$, i.e., out-of-plane corrugations defined by Eq. (\[strain\]) and (\[strain2\]). By definition (\[omegam\]), we have an iterative relation $$\begin{aligned}
w_{m}+w_{m-2}&=2w_{m-1}.\end{aligned}$$ Since $w_{1}=w_0-1$, we obtain $w_{m}=w_0-m$. Then the matrix $T$ can be simplified, and the valley Hall conductivity reads $$\begin{split}
8\sum_{i=b,c,d,e}\sigma_{xy}^{i,A}&=\frac{1.9662e^2}{h}\cos\theta_{\bm k}
\end{split}$$ for $m=1$ (monolayer graphene) and $$\begin{split}
8\sum_{i=b,c,d,e}\sigma_{xy}^{i,A}&=\frac{3.9634e^2}{h}\cos\theta_{\bm k}
\end{split}$$ for $m=2$ (bilayer graphene). Note that these results are very close to $\frac{2me^2}{h}\cos\theta_{\bm k}$, differing by less than $2\%$, confirming the validity of the approximation used in Sec. \[sec:diagram\] and the main text.
$\epsilon=0.821$
----------------
In this section, we consider $\epsilon=0.821$, which is a more realistic value for thermally excited ripples [@doussal1992], i.e., out-of-plane corrugations. By definition (\[omegam\]), we have an iterative relation $$\begin{split}
w_{m}&=\epsilon\frac{1}{2(m-1)}(w_{m-2}-w_{m})\\
&+w_{m-1}-\frac{1}{2}w_{m-2}+\frac{1}{2}w_m,\ \ m\geq2
\end{split}$$ and $$\begin{aligned}
w_1=(\frac{1}{\epsilon}-1)w_0.\end{aligned}$$ This leads to a solution $$\begin{aligned}
w_{m}-w_0&=P_m(w_{1}-w_{0}),\end{aligned}$$ where $$\begin{split}
P_m&=\frac{\prod_{ii=1}^{m-1}(ii-\epsilon)}{\prod_{ii=1}^{m-1}(ii+\epsilon)}
+\frac{\prod_{ii=1}^{m-2}(ii-\epsilon)}{\prod_{ii=1}^{m-2}(ii+\epsilon)}\\
&+\cdots+\frac{\prod_{ii=1}^{1}(ii-\epsilon)}{\prod_{ii=1}^{1}(ii+\epsilon)}+1.
\end{split}$$ By numerics, we find that $$\begin{split}
8\sum_{i=b,c,d,e}\sigma_{xy}^{i,A}&=\frac{1.9994e^2}{h}\cos\theta_{\bm k}
\end{split}$$ for $m=1$ and $$\begin{split}
8\sum_{i=b,c,d,e}\sigma_{xy}^{i,A}&=\frac{3.9998e^2}{h}\cos\theta_{\bm k}
\end{split}$$ for $m=2$. Again the results are very close to $\frac{2me^2}{h}\cos\theta_{\bm k}$, differing by less than $0.1\%$.
[52]{}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/RevModPhys.81.109) [****, ()](\doibase 10.1038/nphys1420) [****, ()](\doibase
10.1038/nature05545) [****, ()](\doibase 10.1038/nmat2011) [****, ()](\doibase 10.1016/j.physrep.2010.07.003) [****, ()](\doibase
10.1016/j.physrep.2015.12.006) [****, ()](\doibase 10.1103/PhysRevB.50.7526) [****, ()](\doibase
10.1103/PhysRevX.4.041019) [****, ()](\doibase
10.1103/PhysRevLett.113.126801) [****, ()](\doibase 10.1103/PhysRevB.88.195417) [****, ()](\doibase 10.1103/PhysRevB.95.195402) [****, ()](\doibase
10.1103/PhysRevLett.99.236809) [****, ()](\doibase
10.1103/RevModPhys.82.1539) [****, ()](\doibase 10.1126/science.1254966) [****, ()](\doibase 10.1038/NPHYS3485) [****, ()](\doibase
10.1038/NPHYS3551) [****, ()](\doibase 10.1103/PhysRevLett.114.256601) [****, ()](\doibase 10.1103/PhysRevB.92.125425) [****, ()](\doibase
10.1038/nphys1822) [****, ()](\doibase 10.1038/ncomms14552) [ ](http://adsabs.harvard.edu/abs/2018arXiv180505955S) [****, ()](\doibase
10.1103/PhysRevLett.120.026802) [****, ()](\doibase
10.1103/PhysRevB.83.125122) [****, ()](\doibase
10.1209/0295-5075/95/67001) [****, ()](\doibase
10.1038/s41467-019-08629-9) [****, ()](https://advances.sciencemag.org/content/5/4/eaau6478) [****, ()](\doibase 10.1103/RevModPhys.82.1959) [****, ()](\doibase 10.1103/PhysRevLett.93.206602) [****, ()](\doibase 10.1103/PhysRevB.73.075318) [****, ()](\doibase 10.1103/PhysRevB.75.045315) [****, ()](\doibase 10.1103/PhysRevB.2.4559) [****, ()](\doibase
10.1103/PhysRevB.84.155410) [****, ()](\doibase 10.1038/nnano.2010.172) @noop [ ]{} [****, ()](\doibase http://dx.doi.org/10.1143/JPSJ.75.074716) [****, ()](\doibase 10.1103/RevModPhys.83.407) @noop [****, ()](\doibase 10.1103/PhysRevLett.115.115502) [****, ()](\doibase 10.1103/PhysRevB.82.081407) [****, ()](\doibase 10.1103/PhysRevB.79.035304) [****, ()](\doibase 10.1038/ncomms9429) [****, ()](\doibase
10.1103/PhysRevLett.115.156602) [****, ()](\doibase 10.1103/PhysRevLett.115.156603) [****, ()](\doibase
10.1103/PhysRevLett.121.176602) [****, ()](\doibase
10.1038/nmat4996) [****, ()](\doibase 10.1103/PhysRevX.8.021020) [****, ()](\doibase
10.1103/PhysRevLett.121.137402) [****, ()](http://advances.sciencemag.org/content/4/5/eaaq0194) [****, ()](\doibase
10.1103/PhysRevB.88.125301) [****, ()](\doibase 10.1103/PhysRevB.58.7151) [****, ()](\doibase 10.1103/PhysRevLett.69.1209)
[^1]: Specifically, if one integrates the Berry curvature over the valence band of the gapped Dirac Hamiltonian, one would obtain a Hall conductivity of $\pm e^2/h$ for the two valleys (counting the spin degeneracy). Naively taking their difference seems to suggest that the valley Hall conductivity should be $2e^2/h$ when the Fermi energy is in the band gap. However, since there is no topological edge states, this result is inconsistent with the general notion that electronic transport should be a Fermi surface property and fully occupied topologically trivial bands should not contribute to electronic transport. The nonzero result is an artifact due to treating the two Dirac valleys separately, which cannot capture the global topology of the bands.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a characterization of the many-body lattice wave functions obtained from the conformal blocks (CBs) of the Ising conformal field theory (CFT). The formalism is interpreted as a matrix product state using continuous ancillary degrees of freedom. We provide analytic and numerical evidence that the resulting states can be written as BCS states. We give a complete proof that the translationally invariant 1D configurations have a BCS form and we find suitable parent Hamiltonians. In particular, we prove that the ground state of the finite-size critical Ising transverse field (ITF) Hamiltonian can be obtained with this construction. Finally, we study 2D configurations using an operator product expansion (OPE) approximation. We associate these states to the weak pairing phase of the $p+ip$ superconductor via the scaling of the pairing function and the entanglement spectrum.'
author:
- 'Sebasti['' a]{}n Montes'
- 'Javier Rodríguez-Laguna'
- Germán Sierra
title: 'The BCS wave function, matrix product states, and the Ising conformal field theory'
---
Introduction
============
Even though superconductivity was discovered experimentally in 1911 by Kamerlingh Onnes [@Onnes], a sufficiently predictive microscopic theory was not available until the work of Bardeen, Cooper and Schriefer, published in 1957 [@BCSoriginal] (known today as BCS theory). One of the fundamental features of this construction is the realization of the ground state of the system as a grand canonical state of fermionic pairs. Following these steps, several other many-body systems have benefited from these insights and extended the result to other non-trivial Gaussian states.
Another seminal landmark of many-body physics is Onsager’s solution of the two-dimensional Ising model, published more than a decade earlier [@Onsager]. Despite its cumbersome original formulation, current understanding of this model is closely related to BCS theory. More concretely, the ground-state of the associated one-dimensional quantum spin chain can also be constructed from a condensate of fermionic pairs [@Henkel; @Sachdev; @Mussardo]. This is a remarkable result if we consider that both models have very different descriptions and applications.
BCS theory has remained an important starting point for the analysis of more exotic phenomena. For instance, in the past few decades two-dimensional superconductors have become testbeds for novel topological features, some of them closely related to the fractional quantum Hall (FQH) effect. Read and Green [@ReadGreen; @Miguel] established a connection between the weak pairing regime of the $p+ip$ superconductor and the topological phase defined by the Moore-Read Pfaffian state [@MooreRead91]. The robustness of these phases to local perturbations have turned them into strong candidate schemes for quantum computing [@NayakAnyons].
One of the shared tools for the study of these strongly correlated quantum systems is conformal field theory (CFT) [@BPZ; @Tsvelik; @QuantumGroupsCFT; @diFrancesco; @Gogolin; @Mussardo]. Due to its symmetry constraints, these theories have powerful algebraic structures that may allow for exact solutions. This has been exploited in the construction of trial wave functions for many-body systems, both in the lattice and the continuum. This is done by computing correlators in the CFT and using them as variational wave functions. The most famous applications have been in FQH physics [@MooreRead91], however it has also been used to study 1D spin systems using infinite matrix product states (iMPS) [@iMPS; @iMPS2; @KL; @FQH-MPS]. In this latter case, the entanglement structure has some features that cannot be easily obtained from finite matrices, such as logarithmic scaling of the entanglement entropy [@iMPS].
It has been argued that conformal blocks (CBs) of rational CFTs can be used to construct wave functions for lattice spin systems [@WFfromCB]. Even if there is no straightforward spin-like structure arising from the representation of internal symmetries (for instance, in the case of minimal CFTs [@BPZ]), the physical degrees of freedom can still be encoded in the different fusion channels of non-Abelian operators. This was illustrated using the Ising CFT, where the relevant CBs were obtained from chiral correlators of several spin operators $\sigma$, grouped in pairs to describe two-level systems.
In this paper, we provide further characterization of the many-body lattice wave functions obtained from the Ising CFT. In particular, we show both analytic and numerical evidence that states describing $N$ spins obtained from the CBs of $2N$ $\sigma$ fields (dubbed $\ket{\psi_{ee}}$) can be understood as BCS wave functions.
This article is organized as follows: Sect. II presents a general short review of BCS states. Sect. III and Sect. IV introduce the notion of vertex operators. We use this formalism to write $\ket{\psi_{ee}}$ and other related states as matrix product states with continuous ancillary degrees of freedom. In Sect. V, we develop a first-order operator product expansion (OPE) that allows us to write $\ket{\psi_{ee}}$ as an explicit (albeit approximate) BCS state. Sect. VI reviews the exact formulas for the Ising CBs. In Sect. VII, Sect. VIII and Sect. IX, we study in detail translationally invariant 1D states. We prove that in this case $\ket{\psi_{ee}}$ can be written as a BCS state in an exact manner and find suitable parent Hamiltonians. In particular, we prove that the ground-state of the finite critical Ising transverse field (ITF) spin chain corresponds to $\ket{\psi_{ee}}$ for a homogeneous configuration. In Sect. X, we study 1D excitations by means of wave functions obtained from CBs with different asymptotic boundary conditions. Sect. XI presents some general remarks about the problem of writing $\ket{\psi_{ee}}$ as a BCS state for an arbitrary coordinate configuration. Finally, in Sect. XII, we provide a brief study of the 2D states obtained from the OPE regime. We use both the entanglement spectrum [@LiHaldane] and the scaling of the entanglement entropy to relate these states to the weak pairing phase of the $p+ip$ superconductor.
BCS wave functions: a short review
==================================
Given a collection of (spinless) fermionic modes $\{c_n\}_{n=1}^N$ on a lattice, we can define the BCS many-body wave function $$\ket{\psi_\text{BCS}} = \prod_{n<m}\left(u_{nm}+v_{nm}c_n^{\dagger}c_m^{\dagger}\right)\ket{0}_c,$$ where $\ket{0}_c$ is the state annihilated by all the operators $c_n$, and $u_{nm}$, $v_{nm}$ are complex numbers that satisfy the normalization condition $|u_{nm}|^2+|v_{nm}|^2 = 1$. Furthermore, we impose $u_{nm}=u_{mn}$ and $v_{nm}=-v_{mn}$. This state can be written as $$\ket{\psi_\text{BCS}} = C_N \exp\left(\sum_{n<m}g_{nm}c_n^{\dagger}c_m^{\dagger}\right)\ket{0}_c,$$ where $g_{nm}=v_{nm}/u_{nm}$ is the pairing function (or more generally pairing matrix) and $C_N=\prod_{n<m} u_{nm}$ is a normalization constant. Note that $g_{nm}$ is a (generally complex) antisymmetric tensor $g_{nm}=-g_{mn}$.
We can interpret this wave function as a grand canonical state of pairs created by the operator $P=\sum_{n<m}g_{nm}c_n^{\dagger}c_m^{\dagger}$. From the fermionic anticommutation relations, it can be shown that the wave function amplitude for $2M$ fermions occupying sites $r(1)<\cdots<r(2M)$ is given by $$\Psi(r(1),\cdots,r(2M))= C_N \text{Pf}(\textbf{M}),$$ where $\textbf{M}$ is the $2M\times 2M$ antisymmetric matrix $$(\textbf{M})_{ij} = g_{r(i), r(j)},$$ and we make use of the Pfaffian $$\text{Pf}(\textbf{Q}) = \frac{1}{2^{M}M!}\sum_{\sigma\in S_{2M}} \text{sgn}(\sigma) \prod_{j=1}^M (\textbf{Q})_{\sigma(2j-1),\sigma(2j) }.
\label{Pfaffian}$$ BCS wave functions are Gaussian states that arise naturally from mean-field solutions of Hamiltonians describing superconductivity [@Sachdev; @ReadGreen]. In that context, both $u_{ij}$ and $v_{ij}$ can be written in terms of single-particle energies $\epsilon_k$ and the pairing interaction potential $V_{k,k'}$. Being spinless fermions, we say that these states correspond to $p$-wave superconductivity, due to the fact that the wave functions for the spatial degrees of freedom are antisymmetric.
The aim of this paper is to provide an alternative route to the BCS state using CFT. More precisely, we shall consider states obtained from the chiral conformal blocks of the critical Ising model and relate them to known BCS states.
Vertex operators in the chiral Ising CFT
========================================
The (chiral) Ising CFT is a minimal RCFT that consists of three primary fields, ${\mathds{1}}$, $\chi$ (Majorana), and $\sigma$ (spin) with conformal weights $0$, $1/2$, and $1/16$, respectively. They have the (non-trivial) fusion rules [@diFrancesco] $$\sigma\times\sigma = {\mathds{1}}+ \chi, \qquad \chi\times\chi = {\mathds{1}}, \qquad \sigma\times\chi = \sigma.$$ A conformal block (CB) in a RCFT is a chiral correlator that encodes an allowed fusion channel for a given set of primary fields. If we start with $N$ primaries $\{\phi_{j_n}\}$, a CB can be written as [@QuantumGroupsCFT; @MooreSeiberg; @MooreSeibergNotes] $${\mathcal{F}}_{\mathbf{p}}(z_1,\cdots,z_N) = \braket{\prod_{n=1}^N \phi_{j_n}(z_n)}_{\mathbf{p}}.$$ where ${\mathbf{p}}$ labels the internal channels. The number of conformal blocks of this type depends on the possible allowed fusion channels of the $\phi_{j_n}$ fields.
The exact formulas for the CBs obtained from the Ising primary fields have been calculated in Ref.[@NayakWilczek; @IsingCB]. We will be interested mainly in CBs containing only $2N$ spin field operators $${\mathcal{F}}^{(2N)}_{\mathbf{p}}(z_1,\cdots,z_{2N}) = \braket{{\sigma}(z_1)\cdots{\sigma}(z_{2N})}_{\mathbf{p}}.
\label{sigmaCBs}$$ Given the fusion rules, a pair of $\sigma$ fields can be seen as a single degree of freedom [@WFfromCB]. This allows us to write the fusion channels in terms of local binary variables. In order to see this explicitly, we group the fields in reference pairs $[\sigma(z_{2n-1}),\sigma(z_{2n})]$. When they are fused pairwise, the different channels can be labeled using the vector ${\mathbf{m}}=(m_1,\cdots,m_N)$, with $m_i=0$ ($1$) representing an identity operator ${\mathds{1}}$ (a fermion $\chi$). In this representation, there is an enforced parity coming from the preservation of fermion parity, so that $\sum_i m_i \equiv 0 (\text{mod } 2)$.
The local pair-wise fusion produces bilocal chiral vertex operators $$V_{ac}^b(z_{2n-1},z_{2n}) : {\mathcal{V}}_c \to {\mathcal{V}}_a, \qquad a,b,c = {\mathds{1}}, \chi$$ where $b=m_n$ corresponds to the fusion channel of reference pair $[\sigma(z_{2n-1}),\sigma(z_{2n})]$, ${\mathcal{V}}_a$ are the Verma modules associated to the corresponding primary fields, and we require the conservation of fermionic parity at each vertex \[Fig.\]. We can use these operators to express explicitly the inner structure of each CB.
![Graphical representation of the bilocal vertex operator $V^b_{ac}(z_1,z_2)$.[]{data-label="bilocalvertex"}](bilocalvertex){width="1.0\linewidth"}
Many-body lattice states from Ising conformal blocks
====================================================
Let us now consider the $2^N$-dimensional Hilbert space ${\mathcal{H}}$ obtained from $N$ spinless fermionic modes $\{c_n\}_{n=1}^N$ and define the $2\times 2$ operator matrix $$A^{(n)}(z_{2n-1}, z_{2n}) =
\begin{pmatrix}
V_{{\mathds{1}}{\mathds{1}}}^{\mathds{1}}& c_n^{\dagger}V_{{\mathds{1}}\chi}^\chi \\
c_n^{\dagger}V_{\chi {\mathds{1}}}^\chi & V_{\chi \chi}^{\mathds{1}}\end{pmatrix}.
\label{Amatrix}$$ This yields the map $$A^{(n)}(z_{2n-1},z_{2n}) : \left(
\begin{array}{c}
{\mathcal{V}}_{\mathds{1}}\otimes {\mathcal{H}}_e \\
{\mathcal{V}}_\chi \otimes {\mathcal{H}}_o
\end{array}
\right) \to \left(
\begin{array}{c}
{\mathcal{V}}_{\mathds{1}}\otimes {\mathcal{H}}_e \\
{\mathcal{V}}_\chi \otimes {\mathcal{H}}_o
\end{array}
\right),$$ where ${\mathcal{H}}_e$ (${\mathcal{H}}_o$) is the Fock space with even (odd) number of fermions $$\begin{aligned}
{\mathcal{H}}_e &= \left\{\ket{0}_c, c_{i_1}^{\dagger}c_{i_2}^{\dagger}\ket{0}_c,\cdots\right\}, \\
{\mathcal{H}}_o &= \left\{ c_{i_1}^{\dagger}\ket{0}_c, c_{i_1}^{\dagger}c_{i_2}^{\dagger}c_{i_3}^{\dagger}\ket{0}_c,\cdots\right\}, \nonumber\end{aligned}$$ so that ${\mathcal{H}}={\mathcal{H}}_e\oplus{\mathcal{H}}_o$. The product of $N$ matrices of type $A$ gives the $2\times 2$ operator matrix $$\begin{aligned}
\Phi^{(N)} &= A^{(1)}(z_1,z_2)\cdots A^{(N)}(z_{2N-1},z_{2N})\nonumber\\
&= \begin{pmatrix}
\Phi_{ee}^{(N)} & \Phi_{eo}^{(N)} \\
\Phi_{oe}^{(N)} & \Phi_{oo}^{(N)}
\end{pmatrix}.
\label{PsiMatrix}\end{aligned}$$ Using this notation, we have that the operator $\Phi_{ee}^{(N)}$ acting on ${\mathcal{V}}_{\mathds{1}}\otimes{\mathcal{H}}_e$ defines the (unnormalized) state $$\ket{\psi_{ee}} = \braket{ 0\left| \Phi_{ee}^{(N)} \right| 0} \ket{0}_c \in{\mathcal{H}}_e,
\label{psiee}$$ where $\bra{0}\cdots \ket{0}$ corresponds to the expectation value in the vacuum of the CFT.
As noted in Ref.[@WFfromCB], this construction is very similar to matrix product states (MPS) obtained from CFT [@iMPS; @iMPS2]. In both cases, the ancillary degrees of freedom are described by a quantum field theory and the resulting many-body wave functions describes a lattice system. As a matter of fact, note that $\ket{\psi_{ee}}$ corresponds to the many-body state defined in that paper written in fermionic variables $$\Psi_{\mathbf{m}}^{(ee)} = \braket{{\mathbf{m}}| \psi_{ee}} = {\mathcal{F}}_{\mathbf{m}}(z_1,\cdots, z_{2n}),
\label{CBamplitudes}$$ where $\ket{{\mathbf{m}}} = \ket{m_1\cdots m_N}$. The present formulation highlights both the inner (i.e., entanglement) structure of these states and its relation to the physical degrees of freedom.
We can also contruct other states by adding fermions to the asymptotic states (within the operator-state correspondence [@diFrancesco]), in particular $$\begin{aligned}
\ket{\psi_{oo}} = \braket{ \chi\left| \Phi_{oo}^{(N)} \right| \chi} \ket{0}_c \in{\mathcal{H}}_e.
\label{psioo}\end{aligned}$$ As we will see in a later section, these wave functions are natural ansätze for low-energy excited eigenstates.
First-order picture: OPE Analysis
=================================
The construction we have discussed so far is quite general. In order to get a more intuitive picture of these states, we can consider a first-order approximation using the operator product expansion (OPE). This scheme will allow us to get a glimpse of the structure of state $\ket{\psi_{ee}}$ using simplified operators.
The full expression of the OPE of two $\sigma$ fields is given by [@diFrancesco] $$\begin{aligned}
\sigma(z_1)\sigma(z_2) = \frac{1}{z_{12}^{1/8}}\bigg(&\sum_{\alpha\in{\mathcal{V}}_{\mathds{1}}}z_{12}^{h_\alpha}C_{\sigma\sigma}^\alpha \alpha\left(\frac{z_1 + z_2}{2}\right)
\label{fullSigmaOPE}\\
&+\sum_{\beta\in{\mathcal{V}}_\chi}z_{12}^{h_\beta}C_{\sigma\sigma}^\beta \beta\left(\frac{z_1 + z_2}{2}\right)\bigg),\nonumber\end{aligned}$$ where $z_{12} = z_1 - z_2$, $\alpha$ $(\beta)$ are the fields with conformal weights $h_\alpha$ $(h_\beta)$ that generate the Verma module ${\mathcal{V}}_{\mathds{1}}$ $({\mathcal{V}}_\chi)$ by acting on the vacuum, and $C_{\sigma\sigma}^\alpha, C_{\sigma\sigma}^\beta$ are constants fixed by 3-point functions. (Note that we are using a symmetrized version of the OPE, instead of pinning the resulting operators on $z_2$.) If we only keep the lowest orders in the expansion, we get the familiar expression $$\sigma(z_1)\sigma(z_2) \sim \frac{1}{z_{12}^{1/8}}\left(1+\left(\frac{z_{12}}{2}\right)^{1/2}\chi\left(\frac{z_1+z_2}{2}\right)\right),
\label{OPEapprox}$$ where we used the fact that $C_{\sigma\sigma}^\chi = 1/\sqrt{2}$.
Assume now that we have $N$ pairs of $\sigma$ fields, parametrized by $$z_{2n-1} = w_n - \frac{1}{2}\delta_n, \qquad z_{2n} = w_n +\frac{1}{2} \delta_n.$$ Using this notation and the OPE, we can write the approximate expression for $$A^{(n)} \sim \frac{1}{\delta_n^{1/8}}\left[{\mathds{1}}_2 +\left( \sqrt{\frac{\delta_n}{2}}c_n^{\dagger}\chi(w_n)\right)\sigma^x \right],
\label{AmatrixOPE}$$ where $\sigma^x = \begin{pmatrix}
0 & 1 \\
1 & 0
\end{pmatrix}$ is one of the Pauli matrices. Note that this approximation implies $$V_{{\mathds{1}}{\mathds{1}}}^{\mathds{1}}= V_{\chi \chi}^{\mathds{1}}\sim \frac{1}{\delta_n^{1/8}}{\mathds{1}}, \quad V_{{\mathds{1}}\chi}^\chi = V_{\chi {\mathds{1}}}^\chi \sim \frac{1}{\sqrt{2}}\delta_n^{3/8}\chi(w_n).
\label{OPEvertex}$$ Given that $(c_n^{\dagger})^2=0$, we have $$A^{(n)} \propto \exp\left(\sqrt\frac{\delta_n}{2} c_n^{\dagger}\chi(w_n)\sigma^x\right),$$ so that becomes $$\Phi^{(N)} \propto \exp\left[\sum_{n=1}^N\left(\sqrt\frac{\delta_n}{2} c_n^{\dagger}\chi(w_n)\right)\sigma^x \right].$$ Now, using the fact that the vacuum of the Ising CFT is a free Gaussian state for the Majorana fermions [@diFrancesco], we employ the familiar identity $$\braket{\exp(A)}_\text{Gaussian} = \exp\left(\frac{1}{2}\braket{A^2}_\text{Gaussian}\right)$$ (assuming $\braket{A}_\text{Gaussian}=0$) to obtain $$\begin{aligned}
\braket{0\left|\Phi_{ee}^{(N)}\right|0} &\propto \exp\left[\sum_{n<m}\frac{\sqrt{\delta_n\delta_m}}{2}\braket{\chi(w_n)\chi(w_m)}c_n^{\dagger}c_m^{\dagger}\right]\nonumber\\
&=\exp\left[\sum_{n<m}\frac{\sqrt{\delta_n\delta_m}}{2\left(w_n - w_m\right)}c_n^{\dagger}c_m^{\dagger}\right].\end{aligned}$$ We conclude then that $\ket{\psi_{ee}}$ is a BCS state defined by the (real-space) pairing function $$g_{nm}^{(\text{OPE})} = \frac{\sqrt{\delta_n\delta_m}}{2(w_{n}-w_{m})}.
\label{OPEpairingPlane}$$ Note that this result holds for arbitrary complex coordinates and only depends on the validity of the OPE approximation. One may wonder if this expansion is really needed to guarantee the BCS structure of the lattice wave function. As we will show in a later section, numerical calculations suggests that this result extends beyond the OPE expansion, albeit with a different pairing function. We will also discuss some aspects regarding a full analytical proof of this fact.
If $|z_n|=1$, it is also convenient to use the conformal transformation that maps the plane to the cylinder $$z \mapsto \exp(i\theta).$$ In this setting, we parametrize the coordinates as $$\theta_{2n-1} = \phi_n - \frac{1}{2}\epsilon_n, \qquad \theta_{2n} = \phi_n + \frac{1}{2}\epsilon_n,
\label{cylcoord}$$ so that the OPE can be written as $$\begin{aligned}
\sigma(\theta_{2n-1})& \sigma(\theta_{2n}) \\
&\sim \left(\frac{1}{2\sin\left(\frac{\epsilon_n}{2}\right)}\right)^{1/8}\left(1+\sin^{1/2}\left(\frac{\epsilon_n}{2}\right)\chi(\phi_n)\right).\nonumber\end{aligned}$$ Given that on the cylinder we have $$\braket{\chi(\phi_1)\chi(\phi_2)}_{\text{cyl}} = \frac{1}{2\sin\left(\frac{\phi_1-\phi_2}{2}\right)},$$ a similar analysis yields a BCS state with the pairing function $$g_{nm}^{(\text{OPE, cyl})} = \frac{\sqrt{\sin\left(\frac{\epsilon_n}{2}\right)\sin\left(\frac{\epsilon_m}{2}\right)}}{2\sin\left(\frac{\phi_n - \phi_m}{2}\right)}.
\label{OPEpairingCylinder}$$ This representation is particularly useful for lattice configurations which are periodic, such as cylinders. Note again that this analysis holds for arbitrary configurations, allowing for complex $\theta_n$.
Exact expressions for the Ising conformal blocks
================================================
Finding the exact form of the CBs for an arbitrary CFT is in general an ardous task. While it is known that they must satisfy a set of well-known differential equations [@diFrancesco], it is far from obvious that they can be solved analytically for any given number of primary fields. In the case of the Ising CFT, one can find exact closed expressions by means of bosonization [@NayakWilczek; @IsingCB].
We will make use of the multiperipheral basis to write the exact formulas for the CBs. This is a canonical representation that is valid for all types of CBs [@QuantumGroupsCFT; @WFfromCB]. We will omit the $\sigma$’s in this notation and write ${\mathbf{p}}=(p_1,\cdots,p_{N-1})$, where $p_i=0$ ($1$) corresponds to an identity operator ${\mathds{1}}$ (a fermion $\chi$). (Note that ${\mathbf{p}}$ can take $2^{N-1}$ different values, as expected.)
We can easily relate the multiperipheral basis to the one obtained from the pair-wise fusion of operators (see Fig. ). The latter is the basis that we used previously in . Note that, in order to preserve the number of fermions at each vertex, there is the restriction $m_k = p_{k-1} + p_k (\text{mod}\, 2)$. (We define fixed auxiliary values $p_0=p_N=0$.)
![A conformal block using only ${\sigma}$ field operators grouped in reference pairs $({\sigma}(z_{2k-1}), {\sigma}(z_{2k}))$. The equivalance between the two representations is obtained from the relation $m_k = p_{k-1} + p_k (\text{mod}\, 2)$.[]{data-label="IsingCB"}](IsingCB){width="0.95\linewidth"}
Before stating the formulas for the CBs, we introduce some extra notation. First, we will need certain bipartitions of the ${\sigma}$ field coordinates that associate the points of each reference pairs to different groups. We call these macrogroups $\ell_{\mathbf{q}}, \ell'_{\mathbf{q}}$ and they are generated from an integer ${\mathbf{q}}=0,\cdots, 2^{N-1}-1$ according to [@WFfromCB; @IsingCB] $$\ell_{\mathbf{q}}(k) = 2k -\frac{1}{2}(1+s_k), \qquad \ell'_{\mathbf{q}}(k) = 2k -\frac{1}{2}(1-s_k),
\label{Spin macrogroup}$$ where $q_k$ are the binary digits of ${\mathbf{q}}= (q_1, q_2, \dots, q_{N-1})$, $$s_k = \prod_{i=1}^{k-1}\left(1-2q_i\right),
\label{auxspin}$$ and $s_1=1$ by definition.
Using this notation, we can define $$z_{\ell_{\mathbf{q}}}=\prod_{k<m}z_{\ell_{\mathbf{q}}(k),\ell_{\mathbf{q}}(m)}.$$ where $z_{ab} = z_a - z_b$. We will also need the sign given by $$\begin{aligned}
\epsilon_{{\mathbf{p}}{\mathbf{q}}} &\equiv (-1)^{\sum_k p_k q_k}= \prod_{k=1}^{N-1}\left(1-2p_k q_k\right)\\
&= \prod_{k=1}^{N-1}\left(1+p_k (s_{k}s_{k+1}-1)\right)\equiv \tilde\epsilon_{{\mathbf{p}}{\mathbf{s}}}.\nonumber\end{aligned}$$ using the binary expansion of both ${\mathbf{p}}$ and ${\mathbf{q}}$.
The expression for the CB can be written as [@IsingCB] $${\mathcal{F}}^{(2N)}_{\mathbf{p}}= \frac{1}{2^{\frac{N-1}{2}}}\prod_{a<b}^{2N}z_{ab}^{-1/8}\left(\sum_{{\mathbf{q}}= 0}^{2^{N-1}-1}{\epsilon}_{{\mathbf{p}}{\mathbf{q}}}\sqrt{z_{\ell_{\mathbf{q}}}z_{\ell'_{\mathbf{q}}}}\right)^{1/2}.
\label{CBsigmas}$$ Note that the sum inside the square root is the only part that depends on ${\mathbf{p}}$.
It is important to remark that we are assuming radial ordering $$|z_1|\geq |z_2| \geq \cdots \geq |z_{2N}|.$$ Moreover, if $|z_n| = |z_m|$ and $n<m$, we will assume that the angular parts in the polar decomposition are ordered with respect to the principal value of the logarithm. In other words, if $z_n = \exp(a_n + ib_n)$, whenever $a_n = a_m$, we will assume $$-\pi < b_n < b_m \leq \pi$$ if $n<m$.
The ordering of the coordinates will be important because it ensures that we consistently choose the same branches of the (complex) square root. In order to see this, let us define $$B_{\mathbf{q}}= \prod_{n<m}^N\left[\left(1 - \frac{z_{\ell_{\mathbf{q}}(m)}}{z_{\ell_{\mathbf{q}}(n)}}\right)\left(1 - \frac{z_{\ell'_{\mathbf{q}}(m)}}{z_{\ell'_{\mathbf{q}}(n)}}\right)\right]^{\frac{1}{2}}.$$ Using this notation, we note that we can write the ${\mathbf{p}}$-dependent part of using only the main branch of the square root $${\mathcal{F}}_{\mathbf{p}}^{(2N)} \propto \left(\sum_{{\mathbf{q}}= 0}^{2^{N-1}-1}{\epsilon}_{{\mathbf{p}}{\mathbf{q}}} \frac{B_{\mathbf{q}}}{B_0}\right)^{1/2}.$$ This will be particularly important for 2D spin configurations. Note that we can obtain CBs for coordinates which are not radially ordered by analytic continuation of these expressions. This can be done by means of the Ising braid matrices [@QuantumGroupsCFT; @WFfromCB; @MooreSeibergNotes].
1D wave functions
=================
We focus now on a one-dimensional configuration. For this purpose, we choose the $2N$ coordinates to be given by $z_k=\exp(i\theta_k)$, where (see Fig. ) $$\theta_k =\frac{2\pi}{2N}\left(k+(-1)^k\delta-N\right),
\label{coord1d}$$ and $\delta\in (-\frac{1}{2},\frac{1}{2})$ is a fixed parameter. Using this parametrization, we can rewrite the wavefunction amplitudes as [@WFfromCB] $$\Psi_{\mathbf{p}}^{(ee)}(\delta) = \frac{1}{\tilde N_0}\left(\sum_{{\mathbf{q}}= 0}^{2^{N-1}-1}{\epsilon}_{{\mathbf{p}}{\mathbf{q}}}\, A_{\mathbf{q}}(\delta)\right)^{1/2},
\label{varWF}$$ where $$A_{\mathbf{q}}= \prod_{n>m}^N\left[\sin\frac{\theta_{\ell_{\mathbf{q}}(n)}-\theta_{\ell_{\mathbf{q}}(m)}}{2}\sin\frac{\theta_{\ell'_{\mathbf{q}}(n)}-\theta_{\ell'_{\mathbf{q}}(m)}}{2}\right]^{\frac{1}{2}},
\label{AAA}$$ and $\tilde N_0$ is a normalization constant $$\tilde N_0^2 = \frac{N^{N/2}}{2^{(N-1)(N-2)/2}}.$$ We can also write in terms of the auxiliary spins (see the Appendix in Ref.[@WFfromCB]) $$A(\{s_k\}) = \prod_{j>i}^N \sin\left[\frac{\pi}{N}\left(j-i+\frac{1+2\delta}{4}(s_j-s_i)\right)\right].
\label{AAAspins}$$ Note that for all values of $\delta$, the resulting wave function describes a translationally invariant spin chain with periodic boundary conditions. This is a consequence of the fact that we are describing physical degrees of freedom on the lattice by means of pairs of $\sigma$ fields. The centers-of-mass of the pairs are uniformly distributed on the circle, while their size is constant for fixed $\delta$ $$\begin{aligned}
\theta_{2k}-\theta_{2k-1} &= \frac{2\pi}{N}(\frac{1}{2}+\delta)\equiv\frac{2\pi}{N}\epsilon, \\
\frac{\theta_{2k}+\theta_{2k-1}}{2} &= \frac{2\pi}{N}\left(k-\frac{2N+1}{2}\right).\nonumber\end{aligned}$$
![Coordinate configuration on the complex plane for a 1D system with 12 spins. Note that the reference pairs are uniformly distributed, so the wave function is translationally invariant for all values of $\epsilon$.[]{data-label="conf_epsilon"}](conf_epsilon){width="0.65\linewidth"}
In this representation, there is an exponentially large number of numerical operations that need to be performed. Luckily, we can obtain a determinant form for these particular configurations that simplifies the calculations. Once again, we make use of the pair-wise fusion basis ${\mathbf{m}}=(m_1,\cdots,m_N)$. It can be shown that the normalized wave function amplitudes can be written as (we will leave the details for Appendix A) $$\Psi^{(ee)}_{\mathbf{m}}(\delta) = \left(\frac{\det(F_{\mathbf{m}}*V)}{\det(V)}\right)^{1/2},
\label{psieedet}$$ where $F_{\mathbf{m}}*V$ is the element-wise matrix product (also known as the Hadamard product of matrices) $$(F_{\mathbf{m}}*V)_{rt} = (F_{\mathbf{m}})_{rt}(V)_{rt},$$ obtained from matrices $$(V)_{rt} = \exp\left(i\frac{2\pi}{N}r(t-1)\right)$$ and $$(F_{\mathbf{m}})_{rt} = \left\{
\begin{array}{l l}
\cos\left[\frac{\pi}{4N}(1+2\delta)\left(2t-N-1\right)\right], & \,\, m_r=0,\\
i\sin\left[\frac{\pi}{4N}(1+2\delta)\left(2t-N-1\right)\right], & \,\, m_r=1.
\end{array} \right.$$ Determinant expression can also allow us to write $\ket{\psi_{ee}(\delta)}$ as a BCS state. If we define the lattice momenta as $$k = \frac{\pi}{N}(2m-N-1), \qquad m=1,\cdots,N
\label{kdef}$$ we can write the normalized state as $$\ket{\psi_{ee}(\delta)} = C_N(\delta) \exp\left(\sum_{n<m}\tilde g_{nm}(\delta)c_n^{\dagger}c_m^{\dagger}\right)\ket{0}_c,$$ where $$C_N(\delta) = \prod_k \sqrt{\cos\left[\frac{(1+2\delta)}{4}k\right]},
\label{cN1D}$$ and $\tilde g_{nm}=g_{n-m}$ with $$\begin{aligned}
g_r = (-1)^r\frac{2}{N}\sum_{k>0} & \tan\left[\frac{(1+2\delta)}{4}k\right]\sin\left(kr\right).
\label{g!D}\end{aligned}$$ We will leave the details for Appendix B. Note that this result is exact and does not depend on any approximation. We also highlight that this is further evidence that $\ket{\psi_{ee}}$ as defined in has a BCS structure beyond the OPE regime.
The ground state of the critical Ising spin chain from conformal blocks
=======================================================================
In Ref.[@WFfromCB], it was argued from numerical evidence that $\ket{\psi_{ee}(\delta=0)}$ correspond exactly to the ground-state of the even sector (defined by $\braket{Q}=\braket{\prod_n\sigma^z_n}=1$) of the Ising tranverse field (ITF) critical Hamiltonian with periodic boundary conditions $$H = -\sum_{n=1}^N \sigma^x_n\sigma^x_{n+1} - \sum_{n=1}^N \sigma^z_n.
\label{ITFh1}$$ We will now present a analytical proof of this result.
The exact solution of is well-known [@Henkel; @Sachdev; @Mussardo]. The ground state can be obtained by mapping the spin variables to spinless fermions using a Jordan-Wigner (JW) transformation $$\begin{aligned}
\sigma_n^z &= 1-2c_n^{\dagger}c_n, \\
\sigma_n^x &= \prod_{m=1}^{n-1}(1-2c_m^{\dagger}c_m)(c_m^{\dagger}+c_m),\nonumber\end{aligned}$$ to obtain $$\begin{aligned}
H = -\sum_{n}(1-2c_n^{\dagger}c_n)- \sum_{n}(c_n^{\dagger}- c_n)(c_{n+r}^{\dagger}+ c_{n+r}).\end{aligned}$$ This is a translationally invariant quadratic Hamiltonian that can be solved via a Fourier transform $$c_n^{\dagger}= \frac{1}{\sqrt{N}}\sum_k e^{ikn}c_k^{\dagger},$$ where we take $k$ as in , followed by a Bogoliubov transformation. The normalized ground state has a BCS structure $$\begin{aligned}
\ket{gs} &= \prod_{k>0} \left[\cos\left(\frac{\theta_k}{2}\right)+i\sin\left(\frac{\theta_k}{2}\right)c_k^{\dagger}c_{-k}^{\dagger}\right]\ket{0}_c \\
&=\prod_{k>0} \cos\left(\frac{\theta_k}{2}\right)\exp\left[\sum_{k>0}i\tan\left(\frac{\theta_k}{2}\right)c_k^{\dagger}c_{-k}^{\dagger}\right]\ket{0}_c,\nonumber\end{aligned}$$ where we define $$\begin{aligned}
\cos\left(\frac{\theta_k}{2}\right) &=\sqrt{\frac{1+\sin\left|\frac{k}{2}\right|}{2}},\\
\sin\left(\frac{\theta_k}{2}\right) &= -\text{sgn}(k)\sqrt{\frac{1-\sin\left|\frac{k}{2}\right|}{2}}. \nonumber\end{aligned}$$ This implies that the normalization constant is $$\begin{aligned}
\prod_{k>0} \cos\left(\frac{\theta_k}{2}\right) &= \prod_{k>0}\sqrt{\frac{1+\sin\left(\frac{k}{2}\right)}{2}}\\
&=\prod_{m=1}^N \sqrt{\cos\left[\frac{\pi}{4N}\left(2m-N-1\right)\right]}. \nonumber\end{aligned}$$ We can also compute the real-space pairing function by doing a Fourier transform of $g_k=i\tan\left(\frac{\theta_k}{2}\right)$ $$\begin{aligned}
g_{r}&=\frac{2}{N}\sum_{k>0}\frac{1-\sin\left(\frac{k}{2}\right)}{\cos\left(\frac{k}{2}\right)}\sin\left(kr\right)\\
&=(-1)^r\frac{2}{N}\sum_{k>0}\tan\left(\frac{k}{4}\right)\sin\left(kr\right),\quad r\in\mathbb{Z},\nonumber\end{aligned}$$ where the second expression is obtained from the first one by replacing $k\mapsto \pi-k$.
Now, coming back to the wave functions obtained from the Ising CBs, note that these expressions correspond to the normalization constant and the pairing function obtained in the previous section when $\delta=0$, so that $$\ket{gs}=\ket{\psi_{ee}(\delta=0)}.$$ This is a remarkable result given that the expression for the CBs was obtained from the infrarred fixed point of the critical theory. It is non-trivial that it would agree with the ground state of a finite-size lattice system.
Parent Hamiltonians for 1D
==========================
We have checked numerically that for $\delta\neq 0$, we can find parent Hamiltonians that can also be mapped to a quadratic fermionic form. We consider the following family of Hamiltonian terms $$\begin{aligned}
Z &= -\sum_{n}\sigma_n^z, \nonumber\\
X_r &= -\sum_n \sigma_{n}^x\sigma_{n+1}^z\cdots\sigma_{n+r-1}^z\sigma_{n+r}^x,\label{HamTerms1D} \\
Y_r &= -\sum_n \sigma_{n}^y\sigma_{n+1}^z\cdots\sigma_{n+r-1}^z\sigma_{n+r}^y,\nonumber\end{aligned}$$ with $r=1,\cdots,N/2$. (Note that $X_1$ is the usual Ising term.) This particular choice for the family of Hamiltonian terms corresponds to those that will yield quadratic forms in fermionic variables (see Appendix C for an explicit fermionic formulation). Given that $\ket{\psi_{ee}(\delta)}$ is translationally invariant and describes a system with periodic boundary conditions, we impose the same constraints on the Hamiltonian terms.
We know that the variational wavefunctions obtained from the CB of the Ising model behave nicely under a Kramers-Wannier (KW) duality transformation [@WFfromCB; @topDefIsing]. In particular, we have that $$\ket{\psi_{ee}(\delta)} \mapsto \ket{\psi_{ee}(-\delta)}.$$ Something similar can be said about the Hamiltonian terms we are considering.
![Variational coefficients of the parent Hamiltonians of the form for $\ket{\psi_{ee}(\delta)}$ for $N=20$ spins. They are normalized so that $a_1=1$. []{data-label="coefham"}](coefham){width="1.0\linewidth"}
The action of the KW transformation can be summarized in the map $$\sigma_n^z \mapsto \sigma_n^x \sigma_{n+1}^x, \qquad \sigma_n^x \mapsto \sigma_1^z\cdots \sigma_n^z.$$ (See Appendix C for a formulation of the KW transformation in terms of Majorana fermions.)From these relations, it is easy to compute the action of the KW transformation (for the even-parity sector of the Hilbert space, defined by $\braket{Q}=\braket{\prod_n\sigma^z_n}=1$) $$\begin{aligned}
Z \mapsto X_1, \quad & X_1 \mapsto Z, \quad X_r \mapsto -Y_{r-1},\, (r=2,\cdots N/2), \nonumber\\ & Y_r \mapsto -X_{r+1},\, (r=1,\cdots, N/2).\end{aligned}$$ Note first that the KW dual of a Hamiltonian that can be written as a quadratic form in fermionic variables is once again of the same type.
We can try to take advantage of the KW duality. For our variational fits, we used the Hamiltonian family $$\begin{aligned}
&\tilde H_0 = {\mathds{1}}, \quad \tilde H_1 = X_1 + Z, \quad \tilde H_2 = X_1 - Z,\label{HamFamily1D} \\
&\tilde H_r = X_{r-1} + Y_{r-2},\, (r=3,\cdots N/2+1).\nonumber\end{aligned}$$ We need $\tilde H_0$ to be equal to the identity for the variational algorithm (see Appendix C). Notice also that $\tilde H_1$ corresponds to the critical ITF Hamiltonian . Using these definitions, the whole family is closed under a KW transformation $$\begin{aligned}
&\tilde H_0 \mapsto \tilde H_0, \quad \tilde H_1 \mapsto \tilde H_1, \\
&\tilde H_r \mapsto -\tilde H_r, \, (r=2,\cdots, N/2+1).\nonumber\end{aligned}$$ We checked numerically that the wavefunction $\ket{\psi_{ee}(\delta)}$ obtained from the Ising CBs with $|\delta|<1/2$ is the ground state of a Hamiltonian of the form (see Fig. ) $$H = -\sum_{r=1}^{N/2+1}a_r \tilde H_r.
\label{varHams}$$ The duality implies that $a_1(\delta) = a_1(-\delta)$, so we will set $a_1=1$. (Remember this is the coefficient associated to the critical ITF Hamiltonian term.) Note also that $$a_r(-\delta) = -a_r(\delta), \,(r=2,\cdots,N/2+1).$$ In Fig. , we plot the variational coefficients obtained for different values of $\delta$ and $N=20$ spins. We see that the Hamiltonian is dominated by $\tilde H_1$, namely, the ITF critical Hamiltonian . The other significant contribution comes from $\tilde H_2$, in particular for $|\delta|\approx 1/2$, so that the ground state approximates the trivial Ising fixed points in these limits (see Ref.[@WFfromCB]).
For $|\delta|\approx 0$, all the Hamiltonian terms that change sign under a KW transformation are very small compared to $\tilde H_1$. Given that the whole Hamiltonian family $\{\tilde H_r\}$ respect the basic Ising symmetries, this implies that $\ket{\psi_{ee}(\delta)}$ approximates small massive perturbations away from criticality. If we drop the Hamiltonian terms for $r>2$, we see that in this vicinity the corresponding transverse field will be given by $$h = \frac{1-a_2}{1+a_2}\approx 1-2a_2 + \mathcal{O}(a_2^2).$$ This explains the relative good agreement between $\ket{\psi_{ee}(\delta)}$ for $|\delta|\ll 1$ and the ground state of the ITF Hamiltonian close to the critical point [@WFfromCB].
Excited states
==============
We can extend the previous discussion to other states obtained from operator matrix . Let us first consider state $\ket{\psi_{oo}}$, defined in . In this case, the asymptotic states of the CFT are fermions $$\Psi^{(oo)}_{\mathbf{p}}\propto \braket{\chi | \sigma(z_1)\cdots \sigma(z_{2N}) | \chi}_{\mathbf{p}}.$$ Assuming radial ordering, we can obtain the amplitudes for this state by adding two fermions at $z=0,\infty$. Starting from the exact expression [@IsingCB] and taking the appropriate limit, the corresponding amplitudes for the associated wave function are given by $$\begin{aligned}
\Psi^{(oo)}_{\mathbf{p}}&= \frac{1}{\tilde N_2}\left(\sum_{{\mathbf{q}}= 0}^{2^{N-1}-1}{\epsilon}_{{\mathbf{p}}{\mathbf{q}}}A_{\mathbf{q}}\right)^{-1/2}\\
&\left[\sum_{{\mathbf{q}}= 0}^{2^{N-1}-1}{\epsilon}_{{\mathbf{p}}{\mathbf{q}}}A_{\mathbf{q}}\left( \sqrt{\prod_{k=1}^N\frac{z_{\ell_{\mathbf{q}}(k)}}{z_{\ell'_{\mathbf{q}}(k)}}} + \sqrt{\prod_{k=1}^N\frac{z_{\ell'_{\mathbf{q}}(k)}}{z_{\ell_{\mathbf{q}}(k)}}} \right)\right].\nonumber\end{aligned}$$ If we use the homogeneous 1D configuration , we can rewrite these amplitudes as $$\begin{aligned}
\Psi^{(oo)}_{\mathbf{p}}&= \frac{1}{N_2}\left(\sum_{\{s_k\}}\tilde\epsilon_{{\mathbf{p}}{\mathbf{s}}}\,A(\{s_k\})\right)^{-1/2}\\
&\left(\sum_{\{s_k\}}\tilde\epsilon_{{\mathbf{p}}{\mathbf{s}}}\,A(\{s_k\})\cos\left[\frac{\pi}{2N}(1+2\delta)\sum_{k=1}^N s_k\right]\right.,\nonumber\end{aligned}$$ They can also be written in terms of determinants. Define the matrix $$(J_{\mathbf{m}}(q))_{r,t} = \exp\left(i\frac{2\pi}{N}r(t-1)\right)(F_{\mathbf{m}}(q))_{r,t}\, ,$$ where $$\begin{aligned}
(F_{\mathbf{m}}(q)&)_{r,t} \\
=&\left\{
\begin{array}{l l}
\cos\left[\frac{\pi}{4N}(1+2\delta)\left(2t-N-1+q\right)\right], & \, m_r=0,\\
i\sin\left[\frac{\pi}{4N}(1+2\delta)\left(2t-N-1+q\right)\right], & \,m_r=1.
\end{array} \right. \nonumber\end{aligned}$$ We can easily show that $$\Psi^{(oo)}_{\mathbf{m}}\propto \frac{1}{\left(\det(J_{\mathbf{m}}(0)\right)^{1/2}}\left[\det(J_{\mathbf{m}}(2)+\det(J_{\mathbf{m}}(-2)\right].$$ Given that this wave function can be written in terms of real amplitudes (up to a possible overall phase that does not depend on ${\mathbf{p}}$), it is easy to compute the overlap with $\ket{\psi_{ee}}$ $$\begin{aligned}
\braket{\psi_{ee}(\delta)|\psi_{oo}(\delta)}&=\sum_{\mathbf{p}}\Psi^{(ee)}_{\mathbf{p}}\Psi^{(oo)}_{\mathbf{p}}\\
&\propto \cos\left(\frac{\pi}{2}(1+2\delta)\right) = -\sin\left(\pi\delta\right).\nonumber\end{aligned}$$ This implies that the two states will be orthogonal if $\delta=0$. (Recall that we are assuming $|\delta|<1/2$.) This is exactly the case for which $\ket{\psi_{ee}}$ describes the ground state of the critical ITF Hamiltonian .
We have checked numerically the action of this Hamiltonian on $\ket{\psi_{oo}(\delta=0)}$ for sizes up to $N=20$ spins using a Lanczos algorithm. We found that it corresponds within machine precision to the first excited state of the even-parity sector of the critical ITF Hamiltonian . This is again a remarkable result given that the amplitudes are computed from CBs obtained at the infrared limit of the thermodynamic model.
These results reflect the relation between the finite-size study of the Ising spin chain and the operator content of the Ising CFT [@Henkel; @ConformallyInvariantLimit; @CardyContent]. It is known that the spectrum of the even-parity sector of with periodic boundary conditions corresponds to the Virasoro towers of both operators ${\mathds{1}}$ and $\epsilon$. These are the primary operators of the full Ising CFT that are even under the internal $\mathbb{Z}_2$ symmetry. Starting from these states, we can in principle construct the full spectrum of the Hamiltonian by acting with the corresponding representation of the Virasoro algebra. These operators can be obtained on the lattice from the local Hamiltonian density [@SKoriginal; @SKnonintegrable].
It is tempting to extend this construction to the odd-parity sector of the ITF spin chain. Finite-size scaling using periodic boundary conditions relate this sector to the Virasoro tower of $\sigma$ [@Henkel]. We tried the natural candidates obtained from (a) using a single fermion on the asymptotic states, both at $z=0,\infty$, so that the CFT degrees of freedom are traced out by $\bra{0}\cdots\ket{\chi}$ or $\bra{\chi}\cdots\ket{0}$; (b) using a pair of $\sigma$ fields on the asymptotic states $\bra{\sigma}\cdots\ket{\sigma}$. In both scenarios, the amplitudes obtained using configuration for $\delta=0$ contained complex amplitudes that cannot be factored to an overall phase. This implies that these states cannot be used naively to describe ground states of real Hamiltonians. Moreover, using $\sigma$ for both asymptotic states can yield wave functions that are not translationally invariant even if the degrees of freedom are arranged uniformly on the circle. This suggests that there is a richer structure underlying the general framework that needs to be understood in further work.
BCS structure beyond OPE: general observations
==============================================
We have seen that the OPE expansion of the CBs yields many-body wave functions with a BCS structure, and that this remains true in the exact case for translationally invariant 1D configurations. One may wonder if this result still holds true for the exact CBs using an arbitrary configuration (assuming, of course, radial ordering). In order to check this, let us consider $N=4$ spins described by $\ket{\psi_{ee}}$. This provides the smallest system size in which a BCS wave function is non-trivial and it allows us to understand the problem in more detail.
First, consider a general BCS wave function for $N=4$. If we write it in full detail, we have $$\ket{\psi_\text{BCS}} \propto \left(1 + \sum_{n<m}g_{nm}c_n^{\dagger}c_m^{\dagger}+ g_{1234}c_1^{\dagger}c_2^{\dagger}c_3^{\dagger}c_4^{\dagger}\right)\ket{0}_c,$$ where we define for convenience $$g_{1234} = g_{12} g_{34} - g_{13} g_{24} + g_{14}g_{23} .
\label{BCS4}$$ Note that this definition relates explicitly to Wick theorem for fermions. If $\ket{\psi_{ee}}$ does indeed describe a BCS state, we expect its amplitudes to fulfill this constraint.
In order to check this, let us write the operator matrix as (we omit the coordinates for simplicity) $$A^{(n)} =
\begin{pmatrix}
V_{00} & c_n^{\dagger}V_{01} \\
c_n^{\dagger}V_{10} & V_{11}
\end{pmatrix}.$$ Using this notation, it is easy to see that condition will be fulfilled for $\ket{\psi_{ee}}$ if and only if (see Fig. ) $$\begin{aligned}
\braket{V_{00}V_{00}V_{00}V_{00}}&\braket{V_{01}V_{10}V_{01}V_{10}}=\nonumber\\
&\braket{V_{01}V_{10}V_{00}V_{00}}\braket{V_{00}V_{00}V_{01}V_{10}}
\label{N4VertexCondition} \\
&- \braket{V_{01}V_{11}V_{10}V_{00}}\braket{V_{00}V_{01}V_{11}V_{10}}\nonumber \\
&+\braket{V_{01}V_{11}V_{11}V_{10}}\braket{V_{00}V_{01}V_{10}V_{00}}.\nonumber\end{aligned}$$ Note first that this equation is trivially satisfied if all $V_{ij}$ are numbers. Also, if we use the OPE approximation , the equation reduces to the usual Wick theorem for free fermions. If we write this using the exact amplitudes in the pair-wise fusion basis, we get $${\mathcal{F}}_{0000}{\mathcal{F}}_{1111} = {\mathcal{F}}_{1100}{\mathcal{F}}_{0011} - {\mathcal{F}}_{1010}{\mathcal{F}}_{0101} + {\mathcal{F}}_{1001}{\mathcal{F}}_{0110}.$$
![Graphical representation of equation . []{data-label="N4Vertex"}](WickN4){width="1.0\linewidth"}
We have checked numerically the condition for (radially ordered) random configurations using the exact CBs and they do indeed describe BCS wave functions. Unfortunately, we cannot provide a general proof even for such a small system size. One possible route is to expand the vertex operators using the full OPE expansion . In that case, condition can be recast into a perturbative expression. Some subtleties regarding this approach are discussed in the Appendix D.
2D wave functions
=================
So far, we have used coordinate configurations for the $\sigma$ fields that are constrained to the unit circle on the complex plane. We now study 2D configurations, where the formalism for the construction of the wave function will be very similar. Unfortunately, we cannot use the same procedure we described in Appendix B to write the amplitudes using a determinant form such as . This limits the system sizes we can consider numerically. However, we can get around this impasse by considering the OPE approximation we already discussed.
We will relate $\ket{\psi_{ee}}$ for a 2D configuration to the weak pairing phase of the effective mean-field Hamiltonian that describes $p+ip$ superconductivity [@ReadGreen] $$H = \sum_{{\mathbf{k}}} \left[\xi_{\mathbf{k}}c_{\mathbf{k}}^{\dagger}c_{\mathbf{k}}+ \frac{1}{2}\left(\Delta_{\mathbf{k}}^* c_{-{\mathbf{k}}}c_{\mathbf{k}}+ h.c.\right) \right],
\label{HpSC}$$ where $$\xi_{\mathbf{k}}=\frac{1}{2m}{\mathbf{k}}^2 - \mu, \qquad \Delta_{\mathbf{k}}= \hat\Delta(k_x - ik_y),$$ $\mu$ is the chemical potential, and $\hat\Delta$ is a constant defining the gap function. The (normalized) ground state of this theory is obtained by usual BCS methods and can be written as $$\ket{gs} = \left.\prod_{\mathbf{k}}\right.' \left(u_{\mathbf{k}}+ v_{\mathbf{k}}c_{\mathbf{k}}^{\dagger}c_{-{\mathbf{k}}}^{\dagger}\right)\ket{0},$$ where the prime on the product indicates that each pair $({\mathbf{k}},-{\mathbf{k}})$ appears only once, and $u_{\mathbf{k}},v_{\mathbf{k}}$ are the Bogoliubov functions obtained from the Bogoliubov-de Gennes (BdG) equations $$\begin{aligned}
E_{\mathbf{k}}u_{\mathbf{k}}= \xi_{\mathbf{k}}u_k - \Delta_{\mathbf{k}}^* v_{\mathbf{k}}, \quad E_{\mathbf{k}}v_{\mathbf{k}}= -\xi_{\mathbf{k}}v_{\mathbf{k}}- \Delta_{\mathbf{k}}u_{\mathbf{k}}.\end{aligned}$$ This reduces to $$\begin{aligned}
E_{\mathbf{k}}& = \sqrt{\xi_{\mathbf{k}}^2 + |\Delta_{\mathbf{k}}|^2},\\
|u_{\mathbf{k}}|^2 &= \frac{1}{2}\left(1+\frac{\xi_{\mathbf{k}}}{E_{\mathbf{k}}}\right), \quad
|v_{\mathbf{k}}|^2 = \frac{1}{2}\left(1-\frac{\xi_{\mathbf{k}}}{E_{\mathbf{k}}}\right).\nonumber\end{aligned}$$ The ground state can then be rewritten as $$\ket{gs} =\left( \prod_{\mathbf{k}}|u_{\mathbf{k}}|^2\right)\exp\left(\frac{1}{2}\sum_{\mathbf{k}}g_{\mathbf{k}}c_{\mathbf{k}}^{\dagger}c_{-{\mathbf{k}}}^{\dagger}\right)\ket{0},$$ where $$g_{\mathbf{k}}= \frac{v_{\mathbf{k}}}{u_{\mathbf{k}}} = -\frac{E_{\mathbf{k}}-\xi_{\mathbf{k}}}{\Delta_{\mathbf{k}}^*}.$$ (Note there is no restriction on ${\mathbf{k}}$, except maybe for ${\mathbf{k}}=0$.) Using the fermionic statistics, the amplitudes of the ground state can be written as Pfaffians using the real-space pairing function $$g({\mathbf{r}}) = \frac{1}{N^2}\sum_{\mathbf{k}}e^{i{\mathbf{k}}\cdot{\mathbf{r}}}g_{\mathbf{k}}.$$ If $\mu>0$, the system will be in the so-called weak pairing phase [@ReadGreen; @Miguel]. For small momenta we have $\xi_{\mathbf{k}}<0$ and $$g_{\mathbf{k}}\sim -\frac{2\mu}{\hat\Delta(k_x+ik_y)}.$$ The leading behavior of the real-space pairing function is given by (see Appendix E for details) $$g({\mathbf{r}})\sim -\frac{2a^2 \mu}{2\pi i \hat\Delta}\frac{1}{x+iy},
\label{gSC}$$ where $a$ is the lattice spacing. Note that this analysis is done on a regular square lattice, assuming a very large system size. However, the leading singular term gives the qualitative infrared behavior that determines the phase of the system.
![2D configuration corresponding to 48 $\sigma$ fields arranged on a cylinder with $N_x=6$ and $N_y=4$. We represent it on the plane according to the exponential map $z\mapsto \exp(i\theta)$. For clarity, we label the 24 physical spins, each one obtained from a pair of $\sigma$ fields. Using coordinates , spins with equal $y_n$ are located at the same radius.[]{data-label="conf2D"}](fig_2d){width="0.75\linewidth"}
Pairing function is similar to the one obtained from $\ket{\psi_{ee}}$ using an OPE approximation . This suggests that $\ket{\psi_{ee}}$ can be related to the weak pairing phase of as long as the OPE regime yields a good approximation of the CBs. The set of distances between the $\sigma$ fields can then be related to the chemical potential of the $p+ip$ superconductor. We expect then that $\ket{\psi_{ee}}$ can describe the topological weak pairing phase of , which has been associated to the Moore-Read Pfaffian state in the fractional quantum Hall effect [@ReadGreen; @Miguel].
Based on the previous analysis, we focus now 2D spin systems on finite cylinders. The lattice will contain $N_y$ spins along the longitudinal direction and $N_x$ spins along the periodic one. We follow the analysis presented in Sect. V for the OPE approximation on the cylinder. We set $z_n = \exp(i\theta_n)$ using the cylinder coordinates , where $\phi_n$ corresponds to the location of the $n$-th physical spin and $\epsilon_n$ to the size of its reference pair. We parametrize them as (see Fig. ) $$\phi_n = \frac{2\pi}{N_x} (x_n - i R y_n)
\label{phi2D}$$ where $n=1,\cdots, N_x N_y$ labels the spin sites, $x_n \in \{1,\cdots N_x\}$ and $y_n \in \{1,\cdots, N_y\}$ are positive integers that define the lattice on the cylinder, and $R$ is the anisotropy factor. (We will use a regular square lattice, so we set $R=1$.) We also define the same separation for all reference pairs $$\epsilon_n = \frac{2\pi}{N_x}\epsilon.$$ We can use complex values for $\epsilon$, but the radial ordering leads to subtleties when we extrapolate to the exact regime. We will focus then on real values, noting that the OPE regime corresponds to $0<\epsilon\ll 1$.
Using this notation, the OPE pairing function becomes $$\begin{aligned}
g_{nm} = \frac{\sin\left(\frac{\pi}{N_x}\epsilon\right)} {2\sin\left(\frac{\pi}{N_x}(x_n-x_m - i R (y_n - y_m))\right)}.\end{aligned}$$ Note that, for large values of $N_x$, we can approximate this expression by a power law, so the leading singular term is similar to .
![Expected occupation per site for different values of $\epsilon$ for a cylinder with $N_x=N_y=20$. The layer $y_n$ correspond to the longitudinal $y$-direction in the cylinder. Note that being periodic along the $x$-direction, the expectation value does not depend on $x_n$.[]{data-label="dens2D"}](fig_density){width="1.0\linewidth"}
In order to characterize these wave functions, we study the entanglement spectrum [@LiHaldane] and the entanglement entropy obtained from the reduced density matrix $\rho_\text{cyl}$ of half a cylinder (corresponding to all sites with $y_n=1,\cdots,\frac{N_y}{2}$). Being a BCS state, we know that $\rho_\text{cyl}$ can be written as [@EntanglementSpinChains; @Peschel] $$\rho_\text{cyl} = \frac{1}{Z}\exp\left(\sum_m \lambda_m b^{\dagger}_m b_m \right),$$ where $\{b_m\}$ are fermionic modes and $Z$ the normalization constant. In Appendix F, we describe a general algorithm to obtain both the spectrum $\{\lambda_m\}$ and the fermionic modes from the pairing function $g_{nm}$.
We compute the expected occupation per site for different values of $\epsilon$ (see Fig. ). We see that the boundaries do not affect the physics deep inside the bulk for large enough $N_y$. Given that in the limit $\epsilon\to 0$ the state corresponds to a trivial vacuum, the occupation is small in the OPE regime.
The periodicity in the $x$-direction is preserved in $\rho_\text{cyl}$, so that we can associate a momentum $k$ to each mode. We can then write the single-body entanglement spectrum as a dispersion relation. In Fig. , we see the single-body spectrum for different values of $\epsilon$. It corresponds to a chiral free fermion. For values close to $\epsilon\to 0$, there is a gap in the dispersion that closes at around $\epsilon\sim 0.1$. This behavior is in agreement with the entanglement spectrum of $p+ip$ superconductors in the weak pairing phase [@chiralfermion].
From the entanglement spectrum, we computed the scaling of the entanglement entropy for different values of $\epsilon$ by changing the cincumference of the cylinder (see Fig. ). In all cases, the scaling follows an area law $S(N_x)\sim c N_x$, with non-universal slopes. According to the scaling, there is no topological correction in the entanglement entropy. Once again, this is in agreement with the behavior of $p+ip$ superconductors [@chiralfermion; @triplettopSC].
Conclusions
===========
We have presented a characterization of many-body states for lattice systems constructed from the CBs of the chiral Ising CFT. The basic feature of the construction is the use of pairs of $\sigma$ fields to describe single localized spins. Writing these CBs using local vertex operators enables us to relate this formalism to usual matrix product states. This rewriting makes explicit the relation between the ancillary CFT degrees of freedom and the lattice fermionic modes.
We have provided evidence that states constructed from CBs using only $\sigma$ fields can be written as BCS states. A partial proof of this fact can be obtained whenever an OPE approximation is valid. In this case, an explicit BCS form can be obtained using the local vertex operator formalism. In the case of translationally invariant 1D configurations, we can go beyond the approximation and write a full non-perturbative proof. This also allows us to obtain a whole family of quasi-local parent Hamiltonians that can be written as quadratic fermionic forms. They are closely related to the critical ITF Hamiltonian . In particular, we presented a proof that the ground state of the critical ITF Hamiltonian can be obtained exactly from this construction. The first excited state of the even-parity sector of this Hamiltonian can also be obtained using CBs with fermions in the asymptotic CFT states.
The OPE approximation can be used to study large 2D spin configurations. By placing the degrees of freedom on finite cylinders, we have related the states obtained from the CBs in the OPE regime to the weak pairing phase of the $p+ip$ superconductor. This has been done via the entanglement spectrum obtained from the reduced density matrix of half of the cylinder.
Further work is needed to deepen the connection between CBs and the ground states of finite systems. In the case of the Ising CFT, this would mean a general proof that $\ket{\psi_{ee}}$ describes a BCS wave function regardless of the coordinate configuration. A deeper understanding of the formalism may produce other physically relevant states, such as the ground state of the 1D odd parity sector of the ITF Hamiltonian, or vortices in 2D superconductors. In additional, generalizations to other rational CFTs, such as the Potts model or the $\mathbb{Z}_n$ model [@Tsvelik; @QuantumGroupsCFT; @diFrancesco], are worth studying. Due to the algebraic constraints, we expect those constructions to be related to anyon chains [@goldenChain; @AnyonTN] or parafermions [@para1; @para2; @para3].
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank H.-H. Tu for being part of this project at its early stages and for suggesting the algorithm for variational parent Hamiltonians. We would also like to thank A.E.B. Nielsen, J. Slingerland, and R. Sanchez for useful discussions. This work is supported by the Spanish Research Agency (Agencia Estatal de Investigación) through the grant IFT Centro de Excelencia Severo Ochoa SEV-2016-0597, and funded by Grant No. FIS2015-69167-C2-1-P from the Spanish government, and QUITEMAD+ S2013/ICE-2801 from the Madrid regional government. SM is supported by the FPI-Severo Ochoa Ph.D. fellowship No. SVP-2013-067869.
Appendix A: Determinant form for the 1D wave function {#appendix-a-determinant-form-for-the-1d-wave-function .unnumbered}
=====================================================
We can rewrite the wave function amplitudes as $$\Psi_{\mathbf{p}}= \frac{1}{\tilde N_0}\left(\sum_{\{s_1=1,\cdots\}}\tilde\epsilon_{{\mathbf{p}}{\mathbf{s}}}\,A(\{s_k\})\right)^{1/2}$$ where $$\begin{aligned}
A(\{s_k\}) =\bigg(& \prod_{j>i}^N \sin\left[\frac{\pi}{N}\left(j-i+\frac{1+2\delta}{4}(s_j-s_i)\right)\right]\\
&\sin\left[\frac{\pi}{N}\left(j-i-\frac{1+2\delta}{4}(s_j-s_i)\right)\right]\bigg)^{1/2}.\nonumber\end{aligned}$$ Using the fact that (see Appendix in [@WFfromCB]) $$\begin{aligned}
\prod_{j>i}^N &\sin\left[\frac{\pi}{N}\left(j-i+\frac{\alpha}{4}(s_j-s_i)\right)\right] \\
&=\prod_{j>i}^N \sin\left[\frac{\pi}{N}\left(j-i-\frac{\alpha}{4}(s_j-s_i)\right)\right],\end{aligned}$$ for arbitrary real $\alpha$, we can eliminate the square root in $A(\{s_k\})$ and lift the restriction on $s_1$ noting that $$\begin{aligned}
\sum_{\{s_1=1,\cdots\}}\tilde\epsilon_{{\mathbf{p}}{\mathbf{s}}}\, A(\{s_k\})
=\frac{1}{2}\sum_{\{s_1=\pm1,\cdots\}}\tilde\epsilon_{{\mathbf{p}}{\mathbf{s}}}\, A(\{s_k\}).\end{aligned}$$ Putting all the pieces together, we have $$\Psi_{\mathbf{p}}= \frac{1}{N_0}\left(\sum_{\{s_k\}}\tilde\epsilon_{{\mathbf{p}}{\mathbf{s}}}\,A(\{s_k\})\right)^{1/2},$$ where now $$A(\{s_k\}) = \prod_{j>i}^N \sin\left[\frac{\pi}{N}\left(j-i+\frac{1+2\delta}{4}(s_j-s_i)\right)\right]$$ and $$N_0^2 = 2 \tilde N_0^2 = 2^{N}\left(\frac{N}{2^{N-1}}\right)^{N/2}.$$ Note first that, using the identities $$\sum_{\sigma\in S_N}\text{sgn}(\sigma)\prod_{n=1}^N \alpha_n^{\sigma(n)-1} = \prod_{n>m}^N (\alpha_n-\alpha_m)$$ and $$\sin\left(\frac{\theta_n - \theta_m}{2}\right) = \exp\left(-i\frac{\theta_n + \theta_m}{2}\right)\left(\frac{z_n - z_m}{2i}\right),$$ we have $$\begin{aligned}
A(\{s_k\})&=\prod_{j>i}^N\left[ \exp\left(-i\frac{\theta_j + \theta_i}{2}\right)\left(\frac{z_j - z_i}{2i}\right)\right]\nonumber\\
&=C_N\sum_{\sigma\in S_N}\text{sgn}(\sigma)\left(\prod_{j=1}^N a_{j,\sigma(j)}\right)\left(\prod_{j=1}^N b_{j,\sigma(j)}\right)\end{aligned}$$ where $$(V)_{r,t}=a_{r,t} = \exp\left(i\frac{2\pi}{N}r(t-1)\right)
\label{Vandermonde}$$ defines a Vandermonde matrix, $$b_{r,t} = \exp\left(i\frac{\pi}{4N}s_r(1+2\delta)(2t-N-1)\right)$$ contains all the dependence on $\{s_k\}$, and $$C_N = (2i)^{-N(N-1)/2}e^{-i\frac{\pi}{2}(N^2-1)}$$
Coming back to $\Psi_{\mathbf{p}}$, we can now sum over the auxiliary spins $\{s_k\}$ $$\begin{aligned}
\sum_{\{s_k\}}&\tilde\epsilon_{{\mathbf{p}}{\mathbf{s}}}\,A(\{s_k\}) \\
&= C_N\sum_{\sigma\in S_N}\text{sgn}(\sigma)\prod_{j=1}^N a_{j,\sigma(j)} \left(\sum_{\{s_k\}}\tilde\epsilon_{{\mathbf{p}}{\mathbf{s}}}\prod_{j=1}^N b_{j,\sigma(j)}\right).\nonumber\end{aligned}$$ We can perform the sum $$\begin{aligned}
\sum_{\{s_k\}}\tilde\epsilon_{{\mathbf{p}}{\mathbf{s}}}\prod_{j=1}^N b_{j,\sigma(j)}=2^N\prod_{j=1}^N f_{j,\sigma(j)}.\end{aligned}$$ where $$f_{r,t} = \left\{
\begin{array}{l l}
\cos\left[\frac{\pi}{4N}(1+2\delta)\left(2t-N-1\right)\right], & \,\,m_r=0,\\
i\sin\left[\frac{\pi}{4N}(1+2\delta)\left(2t-N-1\right)\right], & \,\, m_r=1,
\end{array} \right.$$ and we make use again of the pair-wise basis ${\mathbf{m}}=(m1_,\cdots,m_N)$. In other words, we have a cosine whenever the $r$-th reference pair fuses to an identity and sine when it fuses to a fermion. We can now clean everything up. Note first that $$\begin{aligned}
N_0^2 &= 2^{N}\prod_{j>i}^N \sin\left(\frac{\pi}{2N}\left(j-i\right)\right)\nonumber\\
&= 2^N C_N\sum_{\sigma\in S_N}\text{sgn}(\sigma)\left(\prod_{j=1}^N a_{j,\sigma(j)}\right) \\
&= 2^N C_N\det(V), \nonumber\end{aligned}$$ where $V$ is the Vandermonde matrix defined by $a_{r,t}$. If we define the matrix $F_{\mathbf{m}}$ by the elements $f_{r,t}$, we have $$\Psi_{\mathbf{m}}= \left(\frac{\det(F_{\mathbf{m}}*V)}{\det(V)}\right)^{1/2},$$ where $F_{\mathbf{m}}*V$ is the Hadamard product $$(F_{\mathbf{m}}*V)_{r,t} = a_{r,t}f_{r,t}\,.$$
Appendix B: BCS state from determinant form for 1D wave functions {#appendix-b-bcs-state-from-determinant-form-for-1d-wave-functions .unnumbered}
=================================================================
In order to show that the wave function amplitudes correspond to a BCS state, we need to write them as Pfaffians obtained from a given pairing function. We will accomplish this by using the multilineality of the determinant.
First, consider the matrix $$(U)_{r,t} = \frac{1}{\sqrt N} \exp\left(i\frac{2\pi}{N}r\left(t-\frac{1}{2}\right)\right).$$ It is easy to show that $U$ is a unitary matrix. Also, by multilineality of the determinant, $$\Psi_{\mathbf{m}}^2 = \frac{\det(F_{\mathbf{m}}*V)}{\det(V)} = \frac{\det(F_{\mathbf{m}}*U)}{\det(U)}.$$ Now, note that we can write $$\det(F_{\mathbf{m}}*U) = C_N(\delta)^2 \det(H_{\mathbf{m}}*U),$$ where $$\begin{aligned}
C_N(\delta)^2 &= \prod_{m=1}^N \cos\left[\frac{\pi}{4N}(1+2\delta)\left(2m-N-1\right)\right],
$$ and $H_{\mathbf{m}}$ is defined by matrix elements $$(H_{\mathbf{m}})_{r,t} = \left\{
\begin{array}{l l}
1, & \,\,m_r=0,\\
i\tan\left[\frac{\pi}{4N}(1+2\delta)\left(2t-N-1\right)\right], & \,\, m_r=1.
\end{array} \right.$$ We can further simplify this expression. If we define $M_{\mathbf{m}}= \left((H_{\mathbf{m}}*U)U^{\dagger}\right)$ and $$\begin{aligned}
g_r &= (-1)^r\frac{i}{N}\sum_{m=1}^N \tan\left[\frac{\pi(1+2\delta)}{4N}\left(2m-N-1\right)\right]e^{i\frac{2\pi}{N}mr}\nonumber\\
& =(-1)^r\frac{2}{N}\sum_{k>0}\tan\left(\frac{k}{4}\right)\sin\left(kr\right),
\label{BCSpairingfun}\end{aligned}$$ (using momenta $k$ as defined in ), it is easy to see that $$(M_{\mathbf{m}})_{r,t} = \left\{
\begin{array}{l l}
\delta_{r,t}, & \,\,m_r=0,\\
g_{r-t}, & \,\, m_r=1.
\end{array} \right.$$ Note that $g_r$ is an anti-symmetric function. Taking into account that $\sum m_n=2R$ is an even number, assume that the $1$’s are located at positions $r(1)<\cdots<r(2R)$. In order to compute the determinant of $M_{\mathbf{m}}$, note that $$\det(M_{\mathbf{m}}) = \det(\textbf{G}_{\mathbf{m}})$$ where $$(\textbf{G}_{\mathbf{m}})_{ij} = g_{r(i)-r(j)},$$ is the $2R\times 2R$ anti-symmetric matrix obtained from $M_{\mathbf{m}}$ by keeping only the rows and columns corresponding to $r(1),\cdots,r(2R)$. Being anti-symmetric, note also that $$\det(\textbf{G}_{\mathbf{m}}) = \text{Pf}^2 (\textbf{G}_{\mathbf{m}}).$$ Summing up, we have $$\begin{aligned}
\Psi_{\mathbf{m}}^2 & = C_N(\delta)^2\frac{\det(H_{\mathbf{m}}*U)}{\det(U)}\nonumber\\
& = C_N(\delta)^2\det\left(M_{\mathbf{m}}\right)\\
& = \left(C_N(\delta)\,\text{Pf} (\textbf{G}_{\mathbf{m}})\right)^2.\nonumber\end{aligned}$$ Given that this result holds for all ${\mathbf{m}}$, we conclude that $\ket{\psi}=\sum_{\mathbf{m}}\Psi_{\mathbf{m}}\ket{{\mathbf{m}}}$ corresponds to a BCS state defined by pairing function .
Appendix C: Finding the 1D parent Hamiltonians {#appendix-c-finding-the-1d-parent-hamiltonians .unnumbered}
==============================================
Consider a family of Hamiltonian terms $$H_\alpha = \sum_{i_1,\cdots,i_k}h^{(\alpha)}_{i_1,\cdots,i_k}$$ which can be either local or non-local. For convenience, we set $H_0 = {\mathds{1}}$. Given a wavefunction $\ket{\Psi}$, we would like to find a linear superposition of these operators that will have $\ket{\Psi}$ as an eigenstate. In other words, we want to find coefficients $J_\alpha$ such that $$\left(\sum_\alpha J_\alpha H_\alpha\right) \ket{\Psi} = 0.
\label{superposition}$$ In order to solve this, consider the matrix $$(M)_{\alpha\beta} = \braket{\Psi\left|H_\alpha H_\beta\right|\Psi}.$$ It is easy to see that condition will be satisfied for a certain set of coefficients $\{J_\alpha\}$ if and only if $M$ has a non-trivial kernel. (Note that $M$ is positive-definite.)
Going back to the Hamiltonian terms , we can use the JW transformation, $$\begin{aligned}
\sigma_n^z &= 1-2c_n^{\dagger}c_n, \nonumber\\
\sigma_n^x &= \prod_{m=1}^{n-1}(1-2c_m^{\dagger}c_m)(c_m^{\dagger}+c_m), \\
\sigma_n^y &= i \prod_{m=1}^{n-1}(1-2c_m^{\dagger}c_m)(c_m^{\dagger}-c_m),\nonumber\end{aligned}$$ to obtain $$\begin{aligned}
Z &\mapsto -\sum_{n}(1-2c_n^{\dagger}c_n), \nonumber\\
X_{r} &\mapsto
- \sum_{n}(c_n^{\dagger}- c_n)(c_{n+r}^{\dagger}+ c_{n+r}),\\
Y_{r} &\mapsto \sum_{n} (c_n^{\dagger}+ c_n)(c_{n+r}^{\dagger}- c_{n+r}).\nonumber\end{aligned}$$ We pick the convention for the JW transformation so that $\ket{0}_c = \ket{\uparrow}^{\otimes N}$. Note that, for the even parity sector $\braket{Q}=\braket{\prod_n \sigma_n^z}=1$, we have antiperiodic boundary conditions for the fermions $c_{N+m} = -c_{m}$.
It is illuminating to write these operators in terms of Majorana fermions $$a_{2n-1} = c_n + c_n^{\dagger}, \quad a_{2n} = \frac{c_n-c_n^{\dagger}}{i}.$$ Antiperiodic boundary conditions on the fermions imply $a_{2N+r} = -a_r$. In these variables, we have $$\begin{aligned}
Z &= i\sum_n a_{2n-1}a_{2n},\nonumber\\
X_r &= i\sum_n a_{2n}a_{2(n+r)-1},\\
Y_r &= - i\sum_n a_{2n-1}a_{2(n+r)}.\nonumber\end{aligned}$$ Now, let us consider the variational Hamiltonian family defined in . Using the Majorana variables, we obtain $$\begin{aligned}
\tilde H_1 &= i\sum_n a_n a_{n+1},\\
\tilde H_r &= i \sum_n (-1)^n a_n a_{n+2(r-1)-1},\nonumber\end{aligned}$$ with $r = 2,\cdots, N/2+1$. Tha action of the KW transformation on the Majorana fermions is simply $$a_r \mapsto a_{r+1},$$ which acts as expected on the Hamiltonian family. Note that this action mimics the interpretation of the KW transformation as a consequence of the braiding of sigma fields as discussed in Ref.[@WFfromCB].
The Hamiltonian family we have used is very similar to the conserved quantities of the Ising model, seen as an integrable model [@SKoriginal]. Those can be obtained from $$E_p = (-1)^p\frac{i}{2p}\sum_n a_n a_{n+p}.$$ Note that $\tilde H_1 = -2E_1$ and that $[E_p,E_q]=0$. It has been shown that formal manipulations of $\{E_p\}$ can yield lattice representations of the Virasoro algebra [@SKoriginal; @SKnonintegrable].
Appendix D: Towards a generalized Wick theorem {#appendix-d-towards-a-generalized-wick-theorem .unnumbered}
==============================================
The OPE of two $\sigma$ fields is given by . One can easily identify the fields $\alpha(z)$ as the ones appearing in the fusion channel of the CVOs $V_{00}$ and $V_{\chi \chi}$, while the fields $\beta(z)$ are the ones appearing in the CVOs $V_{0 \chi}$ and $V_{\chi0}$. This implies that holds provided the fields $\alpha(z_i)$ and $\beta(z_i)$ satisfy the relation $$\begin{aligned}
\braket{ \alpha_1 \alpha_2 \alpha_3 \alpha_4 } \braket{ \beta_1 \beta_2 \beta_3 \beta_4 } = & \braket{ \beta_1 \beta_2 \alpha_3 \alpha_4 } \braket{ \alpha_1 \alpha_2 \beta_3 \beta_4 } \label{generalizedWick} \\
&- \braket{ \beta_1 \alpha_2 \beta_3 \alpha_4 } \braket{ \alpha_1 \beta_2 \alpha_3 \beta_4 }\nonumber \\
&+ \braket{ \beta_1 \alpha_2 \alpha_ 3 \beta_4 } \braket{ \alpha_1 \beta_2 \beta_3 \alpha_4 },\nonumber\end{aligned}$$ where $\alpha_i = \alpha(z_i)$ and $\beta_i = \beta(z_i)$. This equation coincides with the standard Wick theorem if $\alpha(z)={\mathds{1}}$ and $\beta(z) = \chi(z)$. Let us provide other examples.
Suppose $\alpha_1 = T(z_1)$ with $T(z)$ the stress tensor [@diFrancesco], $\alpha_2 = \alpha_3 = \alpha_4 = {\mathds{1}}$ and $\beta_i = \chi(z_i)$. Using this, becomes $$\begin{aligned}
\braket{ T_1 } \braket{ \chi_1 \chi_2 \chi_3 \chi_4 } = &
\braket{ \chi_1 \chi_2 } \braket{ T_1 \chi_3 \chi_4 } - \braket{ \chi_1 \chi_3 } \braket{ T_1 \chi_2 \chi_4 } \nonumber\\
&+ \braket{ \chi_1 \chi_4 } \braket{ T_1 \chi_2 \chi_3 }.\end{aligned}$$ The left hand side of the equation vanishes because on the plane $\braket{ T(z) } = 0$. To find $\braket{ T \chi \chi }$, we use Ward identities [@BPZ] to conclude $$\braket{ T(z_1) \chi({z_2}) \chi(z_3) } = \frac{ z_{23}}{ 2 z_{12}^2 z_{13}^2}.$$ Plugging these equations into yields $$\begin{aligned}
\frac{1}{ z_{12}} \frac{ z_{34}}{ 2 z_{13}^2 z_{14}^2} &- \frac{1}{ z_{13}} \frac{ z_{24}}{ 2 z_{12}^2 z_{14}^2}
+ \frac{1}{ z_{14}} \frac{ z_{23}}{ 2 z_{12}^2 z_{13}^2} = \\
&\frac{ z_{12} z_{34} - z_{13} z_{24} + z_{14} z_{23}}{ 2 (z_{12} z_{13} z_{14})^2} =0, \nonumber\end{aligned}$$ so that the condition is satisfied.
As a more elaborate example, choose $\alpha_i = {\mathds{1}}$, $\beta_1 = L_{-n} \chi(z_1)$ with $L_{-n}$ the mode operator of the stress tensor that belongs to the representation of the Virasoro algebra [@diFrancesco], and $\beta_i(z) = \chi(z_i) \; (i=2,3,4)$. Equation becomes $$\begin{aligned}
\braket{ \left(L_{-n}\chi_1\right) \chi_2 \chi_3 \chi_4 } = &
\braket{ \left(L_{-n}\chi_1\right) \chi_2 } \braket{ \chi_3 \chi_4 } \\
&- \braket{\left(L_{-n}\chi_1\right) \chi_3 } \braket{ \chi_2 \chi_4 } \nonumber\\
&+ \braket{ \left(L_{-n}\chi_1\right) \chi_4 } \braket{ \chi_2 \chi_3 },\nonumber\end{aligned}$$ where $$L_{-n} \chi_1(z_1) = \oint_{z_1} d \zeta \; (\zeta - z_1)^{ - n +1} \; T(\zeta) \, \chi(z_1) , \quad n \geq 1$$ (We have suppressed the denominator $2 \pi i$ in the integral.) Equation can be written as $$\Omega_n \equiv \oint_{z_1} d \zeta \; (\zeta - z_1)^{ - n +1} \; f(\zeta, \{ z_i \} ) = 0,
\label{OmegaCondition}$$ with $$\begin{aligned}
f(\zeta, \{ z_i \} ) =& \braket{T(\xi)\chi_1\chi_2\chi_3\chi_4}- \braket{T(\xi)\chi_1\chi_2}\braket{\chi_3\chi_4}\\
&+ \braket{T(\xi)\chi_1\chi_3}\braket{\chi_2\chi_4} - \braket{T(\xi)\chi_1\chi_4}\braket{\chi_2\chi_3}.\nonumber\end{aligned}$$ We now use the familiar identity for general fields $\phi_i$ with conformal weights $h_i$ [@diFrancesco] $$\begin{aligned}
&\braket{T(\zeta) \prod_{i} \phi_i(z_i) } =\\
&\qquad \left[ \sum_{i} \left( \frac{ h_i}{(\zeta- z_i)^2 } + \frac{1}{ \zeta - z_i} \frac{ \partial}{ \partial z_i} \right) \right]
\braket{ \prod_{i} \phi_i({z_i})}, \nonumber\end{aligned}$$ to find $$f(\zeta, \{ z_i \} ) = \left( \frac{ \zeta - z_1}{ (\zeta - z_2) (\zeta- z_3) (\zeta - z_4)} \right)^2 \frac{ z_{23} z_{24} z_{34}}{ 2 z_{12} z_{13} z_{14}}.$$ Hence, equation becomes $$\begin{aligned}
\Omega_n &= \oint_{z_1} d \zeta \; (\zeta - z_1)^{ - n +1} \; \left( \frac{ \zeta - z_1}{ (\zeta - z_2) (\zeta- z_3) (\zeta - z_4)} \right)^2 \nonumber\\
&= \oint_{z_1} d \zeta \; \frac{ (\zeta - z_1)^{ -n + 3}}{ [ (\zeta - z_2) (\zeta- z_3) (\zeta - z_4)]^2} = 0. \quad (n\geq 1)\end{aligned}$$ This equation holds for $n=1,2,3$ but for $n=4$ one has $$\Omega_4 = \oint_{z_1} d \zeta \; \frac{ (\zeta - z_1)^{ -1}}{ [ (\zeta - z_2) (\zeta- z_3) (\zeta - z_4)]^2} = \frac{ 1}{ (z_{12} z_{13} z_{14})^2}$$ It seems that $\Omega_n \neq 0$ for $n \geq 4$. Hence in these cases does not hold. What is the explanation of this fact?
The characters of the Verma modules ${\mathcal{V}}_{\mathds{1}}$ and ${\mathcal{V}}_\chi$ are given by $$\begin{aligned}
\chi_0(q) &= {\rm Tr}_{{\mathcal{V}}_{\mathds{1}}} q^{ L_0 - c/24} \\
&=q^{ - \frac{1}{48}} \left( 1 + q^2 + q^3 + 2 q^4 + 2 q^5 + 3 q^6 + \dots \right), \nonumber\\
\chi_1(q) & = {\rm Tr}_{{\mathcal{V}}_\chi} q^{ L_0 - c/24}\\
&= q^{ - \frac{1}{48}- \frac{1}{2} } \left(1 + q + q^2 + q^3 + 2 q^4 + 2 q^5 + \dots \right). \nonumber \end{aligned}$$ Notice that at level $n=4$ there are two states in the Majorana sector. As a matter of fact, the descendants we had considered above correspond to the derivatives of the field $\chi(z)$, $$(L_{-n} \chi)(0) = \frac{n+1}{2} \chi_{- n - \frac{1}{2}}$$ The conclusion is that equation reduces to equation only if the fields $\alpha$ and $\beta$ that appear in the OPE are unique at a given level. Otherwise one has to consider all the fields appearing at the same level.
Appendix E: Fourier transform of 2D pairing function {#appendix-e-fourier-transform-of-2d-pairing-function .unnumbered}
====================================================
If $\mu>0$, we have $\xi_{\mathbf{k}}>0$ for small momenta and $$g_{\mathbf{k}}\sim -\frac{2\mu}{\hat\Delta(k_x+ik_y)}.$$ Let us try to fix the constants in the Fourier transform, at least in an asymptotic way. Taking $L=aN$ to be the length of the systems (so that the total number of sites is $N\times N$), we can define $$k_x = \frac{2\pi(n-\frac{1}{2})}{aN}, \quad k_y = \frac{2\pi(m-\frac{1}{2})}{aN},$$ where $n,m=-N/2,-N/2+1\cdots,N/2$. Using this, we have $$\begin{aligned}
\frac{1}{N^2}&\sum_{k_x,k_y} \frac{\exp\left[i\left(k_x x+k_y y\right)\right]}{k_x+ik_y} \\
&\to \frac{a^2}{(2\pi)^2}\int_{-\pi/a}^{\pi/a}dk_x\int_{-\pi/a}^{\pi/a}dk_y\frac{\exp\left[i\left(k_x x+k_y y\right)\right]}{k_x+ik_y}.\nonumber\end{aligned}$$ Here we need to be careful. We will both take the limit $a\to 0$ and keep it explicitly in the prefactor. (This can be fixed by changing the normalization of the Fourier transform.) Note that for $y>0$ $$\int_{-\infty}^\infty \frac{dk_y}{2\pi i} \frac{\exp\left[i\left(k_x x+k_y y\right)\right]}{k_y-ik_x} = \Theta(k_x) \exp\left[ik_x\left(x+i y\right)\right].$$ Using this, we have $$g({\mathbf{r}})\to -\frac{2a^2 \mu}{2\pi i \hat\Delta}\frac{1}{x+iy}.$$ For a fixed number of fermions, this corresponds to the Moore-Read state for the FQHE. In this phase, the ground-state of the $p+ip$ conductor is then a grand-canonical state of fermions with this pairing.
Appendix F: Bogoliubov transformation from a BCS pairing matrix {#appendix-f-bogoliubov-transformation-from-a-bcs-pairing-matrix .unnumbered}
===============================================================
Let us consider a fermionic system with on-site creation operators $c^\dagger_i$, $i\in \{1,\cdots, N\}$ and annihilation operators $c_i$. We will adopt the following notation: $$\{C_l\}_{l=1}^{2N}=\{c_1,\cdots,c_N,c^\dagger_1,\cdots,c^\dagger_N\}.$$ Thus, creation and annihilation operators are bundled together. Let us consider a different set of creation and annihilation operators, $$\{B_l\}_{l=1}^{2N}=\{b_1,\cdots,b_N,b^\dagger_1,\cdots,b^\dagger_N\}$$ with $B_l=\sum_p M_{lp} C_p$. The linear transformation will be a Bogoliubov transformation if the $b^\dagger$ and $b$ are bona-fide creation and annihilation operators, with the expected anticommutation and adjoint relations. The first condition is that $M$ is unitary. If that is the case, the Bogolibov matrix $M$ can be naturally split: $$\begin{pmatrix}b \\ b^\dagger\end{pmatrix} =
\begin{pmatrix}D & E \\ E^* & D^*\end{pmatrix}
\begin{pmatrix}c \\ c^\dagger\end{pmatrix},$$ where $D$ and $E$ are $N\times N$ complex matrices, $D^*$ and $E^*$are their complex conjugates (not Hermitian adjoints!) and they must fulfill $$\begin{aligned}
DD^\dagger + EE^\dagger = {\mathds{1}}, \quad DE^T + ED^T = 0\end{aligned}$$ so that matrix $M$ will be unitary. Notice that $A^\dagger$ is the Hermitian adjoint, and $A^T$ is merely the transpose.
In our case, the BCS state is defined via the pairing function $g_{ij}$, which is anti-symmetric, $g_{ij}=-g_{ji}$, $$\ket{\Psi}=\exp\left( \sum_{ij} g_{ij} c^{\dagger}_i c^{\dagger}_j \right) \ket{0}_c \equiv \exp(P) \ket{0}_c$$ where the last relation defines the pairing operator $P$. This state is the vacuum of a certain Bogoliubov set of operators, $\{B_l\}_{l=1}^{2N}=\{b^\dagger_1,\cdots,b^\dagger_N,b_1\cdots,b_N\}$, which means that $$b_k\ket{\Psi}=0, \qquad k\in \{1,\cdots,N\}.$$ Let us impose that condition in order to find the Bogoliubov transformation $M$. By definition, $$b_k = \sum_i D_{ki} c_i + E_{ki}c^\dagger_i,$$ so our condition becomes $$0 = b_k \exp(P) \ket{0} = \left( \exp(P) b_k +[b_k,\exp(P)] \right)\ket{0}.$$ Remember that $b_k$ can be expanded as a linear combination of $c_i$ and $c^\dagger_i$. Using $$[c_i,f(\{c_j,c^\dagger_j\})]=\frac{\partial f}{\partial c^\dagger_i},$$ we find that $$[c_i,\exp(P)]=\exp(P) \left(\sum_j g_{ij}c^\dagger_j \right).$$ Of course, $c^\dagger_i$ commutes with $\exp(P)$. Thus, the annihilation condition becomes $$\exp(P) \left\{ \sum_i D_{ki}c_i + \sum_{ij}D_{ki}g_{ij}c^\dagger_j+ \sum_i E_{ki}c^\dagger_i \right\} \ket{0} =0.$$ which implies the following relation between $D$, $E$ and $g$: $$\sum_i D_{ki} g_{ij} + E_{kj}=0.$$ Thus, in order to find the Bogoliubov transformation given the pairing matrix $g$, we have to solve the following matrix equations: $$\begin{aligned}
Dg+E &=0, \nonumber\\
D D^{\dagger}+ E E^{\dagger}&={\mathds{1}}, \\
DE^T + ED^T&= 0,\nonumber\end{aligned}$$ From the first equation we get $E=-Dg$, which when inserted into the third equation yields $D(g^T+g)D=0$. But this relation is trivial due to the antisymmetry of $g$. Then, the only non-trivial equation becomes $$D \left( {\mathds{1}}+ g g^\dagger \right) D^\dagger = {\mathds{1}}.$$ This equation can be easily solved in the eigenbasis of ${\mathds{1}}+ g g^{\dagger}$, which is self-adjoint and positive-definite.
[10]{} \[1\][`#1`]{} \[2\]\[\][[\#2](#2)]{}
H. K. Onnes, Commun. Phys. Lab. Univ. Leiden **12**, 120 (1911).
J. Bardeen, L. N. Cooper, J. R. Schrieffer, Phys. Rev. **108**, 1175 (1957).
L. Onsager, Phys. Rev. **65**, 117 (1944).
M. Henkel, *Conformal invariance and critical phenomena*, Springer, 1999.
S. Sachdev, *Quantum Phase Transitions*, Cambridge University Press, Cambridge, 1999.
G. Mussardo, *Statistical Field Theory*, Oxford University Press, 2009.
N. Read, D. Green, Phys. Rev. B **61**, 10267 (2000).
M. Ibañez, J. Links, G. Sierra, S.-Y. Zhao, Phys. Rev. B **79**, 180501(R) (2009). C. Dunning, M. Ibañez, J. Links, G. Sierra, S.-Y. Zhao, J. Stat. Mech. P08025 (2010).
G. Moore, N. Read, Nucl. Phys. B **360**, 362 (1991). C. Nayak et al., Rev. Mod. Phys. **80**, 1083 (2008).
A. Belavin, A. Polyakov, A. Zamolodchikov, Nucl.Phys. **B241**, 33 (1984).
A.M. Tsvelik, *Quantum Field Theory in Condensed Matter Physics*, Cambridge University Press, 1995.
C. Gómez, M. Ruiz-Altaba, G. Sierra, *Quantum groups in two-dimensional physics*, Cambridge University Press, 1996. P. di Francesco, P. Mathieu, D. Sénéchal, *Conformal field theory*. Springer, 1997. A. O. Gogolin, A. A. Nersesyan, A. M. Tsvelik, *Bosonization Approach to Strongly Correlated Systems*, Cambridge University Press, 1999.
J. I. Cirac, G. Sierra, Phys. Rev. B **81**, 104431 (2010).
A. E. B. Nielsen, J. I. Cirac, G. Sierra, J. Stat. Mech. P11014 (2011). H.-H. Tu, A. E. B. Nielsen, J. I. Cirac, G. Sierra, New J. Phys. **16**, 033025 (2014). I. Glasser, J. I. Cirac, G. Sierra, A. E. B. Nielsen, Nucl. Phys. **B886**, 63 (2014). H.-H. Tu, A. E. B. Nielsen, G. Sierra, Nucl. Phys. **B886**, 328 (2014). R. Bondesan, T. Quella, Nucl. Phys. **B886**, 483 (2014). B. Herwerth, G. Sierra, Hong-Hao Tu, A. E. B. Nielsen, Phys. Rev. B **91**, 235121 (2015). H.-H. Tu, G. Sierra, Phys. Rev. B **92**, 041119(R) (2015). I. Glasser, J. I. Cirac, G. Sierra, A. E. B. Nielsen, New J. Phys. 17, 082001 (2015).
A. E. B. Nielsen, J. I. Cirac, G. Sierra, Phys. Rev. Lett. **108**, 257206 (2012). A. E. B. Nielsen, G. Sierra, J. I. Cirac, Nature Communications **4**, 2864 (2013). B. Herwerth, G. Sierra, H.-H. Tu, J. I. Cirac, A. E. B. Nielsen, Phys. Rev. B **92**, 245111 (2015). I. Glasser, J.I. Cirac, G. Sierra, A. E. B. Nielsen, arXiv:1609.02435 (2016).
B. Estienne, Z. Papić, N. Regnault, B. A. Bernevig, Phys. Rev. B **87**, 161112 (2013).
S. Montes, J. Rodr' iguez-Laguna, H.-H. Tu, G. Sierra, Phys. Rev. B **95**, 085146 (2017).
H. Li, F.D.M. Haldane, Phys. Rev. Lett. **101**, 010504 (2008).
G. Moore, N. Seiberg, Comm. Math. Phys. **123**, 177 (1989).
G. Moore, N. Seiberg, “Lectures on RCFT.” in *Physics, Geometry and Topology*, Springer, 1990.
C. Nayak, F. Wilczek, Nucl. Phys. B **479**, 529 (1996).
E. Ardonne, G. Sierra, J. Phys. A **43**, 505402, 2010. D. Aasen, R.S. Mong, P. Fendley, J. Phys. A: Math. Theor. **49**, 354001 (2016). J.L. Cardy, Nucl. Phys. **B270**, 186 (1986).
Y.A. Bashilov, S.V. Pokrovsky, Commun. Math. Phys. **113**, 115 (1987).
W. M. Koo, H. Saleur, Nucl. Phys. **B426**, 459 (1994).
A. Milsted, G. Vidal, arXiv:1706.01436.
J. I. Latorre, E. Rico, G. Vidal, Quant. Inf. Comp. **4**, 48 (2004).
I. Peschel, J. Stat. Mech. P06004 (2004).
N. Bray-Ali, L. Ding, S. Haas, Phys. Rev. B **80**,180504(R) (2009).
T. P. Oliveira, P. Ribeiro, P. D. Sacramento, Journal of Physics: Condensed Matter **26**, 425702 (2014).
A. Feiguin et al., Phys. Rev. Let. **98**, 160409 (2007).
R.N.C. Pfeifer, P. Corboz, O. Buerschaper, M. Aguado, M. Troyer, G. Vidal, Phys. Rev. B **82**, 115126 (2010).
E. Fradkin and L. P. Kadanoff, Nucl. Phys. B **170**, 1 (1980).
A.B. Zamolodchikov and V.A. Fateev, Sov. Phys. JETP **62**, 215 (1985).
P. Fendley, J. Stat. Mech. p. P11020 (2012).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Demand response is designed to motivate electricity customers to modify their loads at critical time periods. The accurate estimation of impact of demand response signals to customers’ consumption is central to any successful program. In practice, learning these response is nontrivial because operators can only send a limited number of signals. In addition, customer behavior also depends on a large number of exogenous covariates. These two features lead to a high dimensional inference problem with limited number of observations. In this paper, we formulate this problem by using a multivariate linear model and adopt an experimental design approach to estimate the impact of demand response signals. We show that randomized assignment, which is widely used to estimate the average treatment effect, is not efficient in reducing the variance of the estimator when a large number of covariates is present. In contrast, we present a tractable algorithm that strategically assigns demand response signals to customers. This algorithm achieves the optimal reduction in estimation variance, independent of the number of covariates. The results are validated from simulations on synthetic data as well as simulated building data.'
author:
- 'Pan Li, and Baosen Zhang, [^1] [^2] [^3]'
bibliography:
- 'ref.bib'
title: An Optimal Treatment Assignment Strategy to Evaluate Demand Response Effect
---
[Li: TBD]{}
INTRODUCTION
============
As more uncertain and intermittent renewable resources are integrated into the power system, operators are increasingly exploring flexibility in customers’ consumptions to balance supply and demand. This operation is commonly known as *demand response (DR)*. In a typical implementation of DR programs, customers receive a DR signal to elicit a change in their consumptions. This signal can be a modification of electricity prices or simply a message requesting a change in consumption [@Siano2014]. An effective DR program improves the efficiency and sustainability of power systems and is a central pillar of the envisioned smartgrid [@DR0; @DR2; @Dai; @LiEtAl2011].
A natural question about demand response is quantifying the impact of a DR signal. That is, if a DR signal is sent to a subset of the users, what is the change in these users’ consumptions because of that signal? An accurate estimate of this change is central to the operation of demand response programs: if not enough change in demand is elicited, other measures need to be taken; if too much change is elicited, the program is inefficient.
Most of existing work in this area of demand response have approached the problem from a market optimization point of view. For example, authors in [@LiEtAl2011; @QianEtAl2013] considered how to optimize the social welfare; and authors in [@SaadEtAl2012] have considered how to create an efficient market for demand response. In all these settings, customers’ responses are captured by well-defined utility functions. These functions are assumed to be known to the operators, or at least to the users themselves.
In practice, these market based approaches can be difficult to implement because customers often do not have a clear model of their own utilities. For example, consider a household with a smart energy management system (e.g., a NEST thermostat). This household will respond to a DR signal, but the response can be a complicated function of the current conditions in the household–e.g., temperature, appliances that are on, number of people at home and so on– and the user may not be consciously aware of the households’ utility function. Therefore the operator needs to *learn customers’ responses* from past history. Furthermore, because of the advancement of household sensors, this response need to be learned under a possibly *high dimensional setting*.
By performing enough experiments with enough customers, that is, sending enough DR signals, the operator will eventually learn the users’ response with accuracy. Repeated experimentation with large group of customers, however, is impractical for two reasons. The first is that operators only sends out DR signals if the demand need to be modified appreciably, and this event does not occur all that often in the power system. The second is that because of use fatigue [@HolyheadEtAl2015], most utilities have agreements with their customers that a single household will only receive a limited number of DR requests [@BalijepalliEtAl2011]. Under these *limited data* regimes, accurately learning the response of users is nontrivial. In this paper, we adopt an *experimental design approach* to the problem estimating the response of users to DR signals. We design user selection algorithms where by carefully choosing the users that receive DR signals, the maximum information about the system response can be learned.
In this paper we consider the linear setting, where a user’s consumption is a linear function of a set of variables and the DR signal. We refer to the former abstractly as a user’s *covariates*, where they could represent measurements such as temperature, appliance status, building type, behavior patterns and etc. Because of the explosive growth in sensing devices, we are particularly interested in the high dimensional case, that is, where users have a large number of covariates. We assume the impact of DR signals is additive to the original consumption behavior. Using the language of experimental design, we regard DR signal as a *treatment* and a user receiving a DR signal as being *assigned this treatment*. Then learning the DR response of the users is equivalent to learning the *average treatment effect*, which is the average response of the customers to the treatment [@Holland1986]. The metric with respect to the estimation of the treatment effect is the *variance* of the estimator as more experiments are performed. Randomized trial is usually thought as the “gold standard” in these types of models mainly due to the fact that randomly assigning treatments to users removes the effect of confounding factors and provides a consistent estimate of the treatment effect. In the presence of many covariates, however, random assignment can be extremely inefficient. In fact, as we show in this paper, in the high dimensional setting, random assignment *does not reduce the variance* of the estimate of the average treatment effect, even as the number of treatment grows without bound. Instead, following the outline in [@ABtest], we design a selection scheme users are picked based on their covariates to be treated.
Suppose there are $n$ total users in the system. Under the linear models considered in this paper, the best possible lower bound on the rate of variance reduction is $\Omega(1/n)$,[^4] given by considering the Fisher information [@Fisher1935]. As discussed above, under high dimensional settings, a randomized algorithm can only achieve $\Omega(1)$, even when are large number of fraction of users are assigned DR signals. The main contributions of this paper are:
1. We show if the number of users selected is a constant fraction of the total number of users, there exists user assignments that achieve a variance reduction rate of $\Theta(1/n)$. This rate is independent of the dimension of the covariates, as long as it is less than $n$.
2. We develop a tractable user assignment algorithm. This algorithm is obtained by converting the variance reduction problem in a densest-cut problem on a graph [@KortsarzPeleg8; @FeigLangber6; @SDP; @YeZhang19].
Our approach differs from previous effort in learning demand response in one important regard. In previous studies, the focus was on training the best predictive model and subtracting out the predicted consumption from the measured consumption [@Zhou16; @Brodersen15]. In our approach, we do not ever learn a predictive model, in the sense that we do not learn the relationship between the covariates and the consumption. Rather, focus on learning a single parameter: the response to the DR signal.
The results in [@ABtest] act as an impetus to this paper. The main difference is that in [@ABtest], the users are assigned a treatment of $\pm 1$, therefore some information is always conveyed by this assignment. In our model, the users are assigned either $1$ (receives DR signal) or $0$ (no signal). Therefore, for the users assigned $0$, we do not obtain any information about the impact of DR signals. This makes the problem much more technically challenging, and consequently we only consider the offline assignment problem whereas [@ABtest] also considers the online assignment problem. There are extensive literature on average treatment effect estimation, and the interested reader can refer to [@causalreview1; @causalreview2] and references within.
The rest of the paper is organized as follows. Section \[problem\] introduces the preliminaries and the problem throughout this paper. Section \[random\] presents the variance of the estimator obtained by random assignment. Section \[optimal\] further presents the variance of the estimator by optimal assignment, followed by a tractable algorithm presented in Section \[sec:SDP\]. Section \[simulation\] details the simulation results obtained by either random assignment and optimal assignment. Section \[conclusion\] concludes the paper.
PRELIMINARIES AND PROBLEM FORMULATION {#problem}
=====================================
In this paper we assume that a user’s consumption is given by a linear model. Let $x_i \in \{0,1\}$ denotes a binary DR signal, where $1$ represents that a signal is sent to user $i$ and $0$ presents that no signal is sent. A covariate vector $\bm z_i$ is also associated with a user, representing available side information. For example, side information may include local temperature, user’s household size, and number of electrical vehicles and so on. We denote the dimension of the covariate vector by $p$, and assume the last component is $1$, which is the intercept. Let $y_i$ denote the consumption of user $i$, which is given as $$\label{prob:problem}
y_i = \beta x_i + \bm{\gamma}^T \bm{z}_i +\epsilon_i,$$ where $\epsilon_i$ is white noise with variance $\sigma^2 = 1$ (for convenience). The coefficient $\beta$ is the impact of the DR signal and estimating it efficiently is the goal of the paper. The coefficient $\bm \gamma$ represents the effect of the covariate vectors. The main technical challenge is to accurately estimate the coefficient of interest $\beta$, even when $\gamma$ is high dimensional. For analytical simplicity, we assume that the entries of $\bm z_i$ are drawn as i.i.d. Gaussian random variables (possibly after centering and rescaling). In simulations (Section \[simulation\]), we show that the results holds for other types of distributions as well.
We assume there are $n$ total users. In this model, a single user that receives two demand response signals at two different times is equivalent to two users each receiving a demand response signal. Therefore, we suppress the time dimension and label all users by $i$. Note that in , all users share a common response $\beta$ to DR signals.
We denote the estimate of $\beta$ by $\hat{\beta}$. The value of $\hat{\beta}$ is a function of the DR assignments, that is, the value of the $x_i$’s. Under the linear setting in , the ordinary least square (ols) estimator $\hat{\beta}$ of $\beta$ is unbiased for all possible allocations of DR assignments, $\hat{\beta}$ is centered at the true value $\beta$. The natural measure of performance is then the variance: $\operatorname{\text{Var}}\hat{\beta}$. With some simple linear algebra, the variance of $\hat{\beta}$ is given by [@LinReg2012]: $$\label{eq:varbeta}
\operatorname{\text{Var}}\hat{\beta} = \frac{\sigma^2}{\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}},$$ where $P_{\bm{Z}^{\perp}} = I-\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}$. The $i$’th row of the data matrix $\bm{Z}$ is given by $\bm{z}_i^T$. We adopt the notation that $\bm{Z}_{n,p}$ denotes a matrix $\bm{Z}$ that has $n$ rows and $p$ columns, while $\bm{Z}_{i:j}$ denotes the $i^{th}$ to $j^{th}$ column of a matrix $\bm{Z}$, where $i \leq j$.
We are primarily interested in the setting where an operator can assign a limited number of $x_i$’s to be $1$. This setting reflects the limit in budget of an operator in sending DR signals. Specifically, let $k$ be the total number of DR signal that can be sent. The goal of the operator is to strategically assign $k$ $x_i$’s to be $1$ such that the variance of $\hat{\beta}$ is minimized. In particular, we are interested in the rate of reduction of $\hat{\beta}$ as $n$ increases and in settings where $k/n$ is a constant.
From , minimizing the variance of $\hat{\beta}$ is equivalent to maximizing the quantity $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$, and we focus on the latter quantity in the rest of the paper due to notational convenience. Two types of algorithms are of interest: i) the standard random assignment where each $x_i$ is chosen to be $1$ or $0$ with probability $k/n$, and ii) an optimal assignment procedure where $x_i$’s are chosen to maximize $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$. Both algorithms face the constrain that only $k$ out of $n$ $x_i$’s can be assigned to be $1$. We characterize *growth rate* of quantity $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$ in terms of $k,n$ and $p$, or equivalently, the decay rate of $\operatorname{\text{Var}}\hat{\beta}$.
We show that when $p$ is relatively small compared to $n$, the two strategies yield similar rates of $\Theta(n)$. In a high dimensional setting where $p$ is comparable to $n$, e.g., $p = n-1$, however, the random assignment is essentially useless in estimating $\beta$, in the sense that $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$ remains a constant in expectation as $n$ grows. Our proposed strategy, on the other hand, improves the rate to $\Theta(n)$ in this case, as long as $k/n$ is a constant. In Section \[random\], we discuss the randomized strategy. The optimal assignment algorithm is then considered in Section \[optimal\].
RANDOM ASSIGNMENT {#random}
=================
Random assignment has been extensively studied in literature, mainly because it balances the covariates in two groups and eliminate the influence of confounders [@causalreview1]. For our model in , random assignment means that a subset of $k$ $x_i$’s are chosen at random and assigned a value $1$. Theorem \[theorem1\] quantifies the rate of the increase of $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$.
\[theorem1\] Random assignment achieves a rate of $\Theta$($(n-p)\frac{k(n-k)}{n^2}$). If $\frac{k}{n} = \rho$ is a constant, then this rate is $\Theta$($n-p$). [$\blacksquare$]{}
Before proving Theorem \[theorem1\], we discuss the scaling rate under the setting when $k/n=\rho$ is a constant. In practice, this is the regime of interest since it is reasonable to suppose that a fraction (e.g. 10%) of users receives DR signals. In this case, the rate achieved by random assignment is $\Theta(n-p)$. This rate is $\Theta$($n$) when $p$ is relatively small compared to $n$. However, when $p$ is large, e.g., $p = n-1$, then this rate becomes $\Theta$($1$). This rate is not desirable as it indicates that the variance of the estimator is not decaying with $n$ even when $n$ is large. Thus we would like to design an assignment strategy which yields an estimator that still possesses a relatively good performance even when $p$ is very close to $n$. In the next section, we show that with optimal assignment, we achieve the optimal rate $\Theta$($n$) when $\frac{k}{n} = \rho$. The proof of Theorem \[theorem1\] follows.
We consider a random assignment where $\text{Pr} \{x_i=1\}=\frac{k}{n}$. Then the rate becomes: $$\label{eq:randomvar}
\begin{aligned}
& \operatorname{\mathbb{E}}\mathrm{tr}\{\bm{x}^{\text{T}}(I-\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}})\bm{x}\} \\
= & k - \operatorname{\mathbb{E}}\mathrm{tr} \{\bm{x}^{\text{T}}\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}\bm{x}\} \\
\overset{(a)} = & k - \mathrm{tr} \{\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}} \operatorname{\mathbb{E}}\bm{x} \bm{x}^{\text{T}} \}\\
\overset{(b)} = & k - \mathrm{tr} \{ \bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}} \operatorname{\mathbb{E}}(\tilde{\bm{x}} + \frac{k}{n})(\tilde{\bm{x}} + \frac{k}{n})^{\text{T}} \}\\
\overset{(c)} = & k - \mathrm{tr} \{ \bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}} ( \frac{k(n-k)}{n^2}I + 0 + \frac{k^2}{n^2} \bm{1} \bm{1}^{\text{T}}) \} \\
\overset{(d)} = & k - p\frac{k(n-k)}{n^2} - \frac{k^2}{n^2} n\\
= & (n-p) k (n-k)/n^2, \\
\end{aligned}$$ where ($a$) follows from linearity and cyclic permutation of the trace operator; ($b$) follows from defining $\tilde{\bm{x}} = \bm{x} - \frac{k}{n}$; ($c$) follows from multiplying out each terms inside $(\tilde{\bm{x}} + \frac{k}{n})(\tilde{\bm{x}} + \frac{k}{n})^{\text{T}} $ and using the fact that each element in $\tilde{\bm{x}}$ has a zero mean and a variance as $\frac{k(n-k)}{n^2}$; ($d$) follows from $\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}$ being a projection matrix onto $\bm{Z}$. Using the fact that the eigenvalues of a projection matrix are either 0 or 1 and $\bm{Z}$ has rank $p$ with probability one, then the trace of $\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}$ is $p$ with probability one. In addition, from Lemma \[lemma0\], it is shown that $\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}\bm{1}= \bm{1}$ if $\bm{Z}$ contains one column as intercept, so that the trace of $\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}\bm{1}\bm{1}^{T} = n$, which completes the equality in $(d)$.
\[lemma0\] If $\bm{Z}$ is a $n$ by $p$ matrix (where $p<n$) with one column which contains all ones, then $\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}\bm{1}= \bm{1}$. [$\blacksquare$]{}
Note that $I - \bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}$ is the projection matrix which is orthogonal to $\bm{Z}^{\text{T}}$, we then have the following: $$\label{nulleq}
\bm{Z}^{\text{T}}(I - \bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}) = \bm{0},$$ where $\bm{0}$ is a zero vector that has length $n$.
Note that $\bm{Z}$ has one column as the intercept, which suggests that $\bm{Z}^{\text{T}}$ has one row where each element takes value one. Since the equality in holds for every row, we then have:
$$\bm{1}^{\text{T}}(I - \bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}) = 0,$$
which indicates that $\bm{1}^{\text{T}} \bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}} = \bm{1}^{\text{T}}$.
OPTIMAL ASSIGNMENT {#optimal}
==================
Instead of being randomly assigned into DR programs, users can be optimally allocated to either the treatment group or the control group depending on their covariate information, in order to obtain the best estimator of $\beta$. Mathematically speaking, we optimally assign each $x_i$ to be 0 or 1, in order to minimize the variance of the estimator $\hat{\beta}$. This optimization problem is: $$\label{prob:original0}
\begin{aligned}
& \underset{\hat{\bm{x}}}{\text{maximize}}
&& \bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}\\
& \text{subject to}
&& \sum_{i=1}^{n}{x}_i = k\\
&&& {x}_i \in \{1,0\}.
\end{aligned}$$
We first discuss the upper bound on the quantity $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$ (which signifies the lower bound for $\operatorname{\text{Var}}\hat{\beta}$). We show that it is O($n$). Then we establish that under the regime of $k/n=\rho$, there exist algorithms that achieve a rate that meets the upper bound of $O(n)$.
Optimal Rate
------------
Before proceeding on analyzing the rate obtained by the proposed strategy, we first discuss the upper bound on the rate of $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$.
\[proposition0\] No assignment can achieve a better rate that O$(n)$ [@ABtest].[$\blacksquare$]{}
The basic idea is to derive the Fisher information with the linear regression model in . The inverse of the Fisher information provides a lower bound for the variance of the estimator obtained by least squares and thus an upper bound for the quantity $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$. For more details, please refer to Proposition 1 in [@ABtest].
In the next subsection we will show that when $\frac{k}{n} = \rho$ which is a constant, we achieve this upper bound.
Achievability of Optimal Rate
-----------------------------
We first present the main result of this section. We assume that each element of $\bm{Z}_{1:p-1}$ (excluding the intercept column) is drawn independently from a standard Gaussian distribution. This assumption will facilitate the calculation of the main result shown in Theorem \[theorem2\]. The algorithm associated with Theorem \[theorem2\] is presented in Algorithm \[algo0\].
\[theorem2\] Recall that the rate is the growing rate of the inverse of the variance introduced in . This rate from optimal assignment is of $\Theta$($\frac{k^2log(\frac{n}{k})}{n}$), which is independent of the dimension of covariates. More specifically, when $\frac{k}{n} = \rho$ is a constant, then this rate is linear rate, i.e., $\Theta$($n$). [$\blacksquare$]{}
Reduce the optimization problem in to using Lemma \[increasing\].
Compute the null space of $\bm{Z}_{n,n-1}^{\text{T}}$, denote it by $\bm{y}$. Each element of $\bm{y}$ should independently follow a standard Gaussian distribution, according to Lemma \[Gaussiannull\].
Find the lower bound for the $k^{th}$ largest element in $\bm{y}$ (suppose that this element is non negative). This lower bound is shown in Lemma \[lemma2\].
The optimal value of the objective function in is at least $\frac{k^2}{n}$ times this lower bound. The rate of this optimal value is stated in Theorem \[theorem2\]. The optimal assignment is to assign those $x_i$’s corresponding to the $k$ largest $y_i$’s in $\bm{y}$ to be 1’s and the rest to be 0’s.
Before proving Theorem \[theorem2\], we first show in Lemma \[increasing\] that the worst case scenario for the rate is when $p = n - 1$. This scenario provides a minimum on the quantity $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$ for every $p$ where $p<n$, which provides a maximum for $\operatorname{\text{Var}}\hat{\beta}$ for every $p < n$. Thus if we can show in Theorem \[theorem2\] that in the worst case scenario where $p = n - 1$, the growing rate of quantity $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$ is $\Theta(n)$ when $\frac{k}{n} = \rho$ is a constant, then this rate holds for all $p$ where $p<n-1$.
\[increasing\] $\operatorname{\text{Var}}\hat{\beta}$ is increasing in $p$. Consequently, if $p$ = $n-1$, the estimator yields the worse case performance [@ABtest]: $$\inf_{1 \leq p < n} \frac{\bm{x}^{\text{T}}P_{\bm{Z}_{n,p}^{\perp}}\bm{x}}{n} = \frac{\bm{x}^{\text{T}}P_{\bm{Z}_{n,n-1}^{\perp}}\bm{x}}{n}$$ [$\blacksquare$]{}
This is a general result about linear estimation and the interested reader can refer to Lemma 5 in [@ABtest].
When $p = n-1$, the rank of $\bm{Z}_{n,n-1}$ is one with probability one, thus we write $P_{\bm{Z}_{n,p}^{\perp}} = \frac{\bm{y}\bm{y}^{\text{T}}}{||\bm{y}||^2}$, where $\bm{y}$ is in the null space of $\bm{Z}_{n,n-1}^{\text{T}}$ [@ABtest], i.e., $\bm{Z}^{\text{T}}_{n,n-1}\bm{y} = \bm{0}$. Based on this observation, $\bm{x}^{\text{T}}P_{\bm{Z}^{\perp}}\bm{x}$ is written into a simpler form as $\frac{\bm{x}^{\text{T}}\bm{y}\bm{y}^{\text{T}}\bm{x}}{||\bm{y}||_2^2} = \frac{(\bm{y}^{\text{T}}\bm{x})^2}{||\bm{y}||^2}$. The problem is then to maximize $\frac{(\bm{y}^{\text{T}}\bm{x})^2}{||\bm{y}||^2}$ under the constraint that we only get to assign $k$ $x_i$’s to be 1’s and the rest to be 0’s. The optimization problem is: $$\label{prob:p=n-1}
\begin{aligned}
& \underset{\hat{\bm{x}}}{\text{maximize}}
&& \frac{(\bm{y}^{\text{T}}\bm{x})^2}{||\bm{y}||^2}\\
& \text{subject to}
&& \sum_{i=1}^{n}{x}_i = k\\
&&& {x}_i \in \{1,0\}.
\end{aligned}$$
To solve the optimization problem in , we need to find $k$ $y_i$’s in $\bm{y}$ such that their sum is maximized, where $y_i$ is the $i^{th}$ element of the vector $\bm{y}$. We observe that it actually suffices to provide a lower bound on this maximum sum to prove the rate.
To provide this lower bound we need to know the structure of $\bm{y}$. We then show in Lemma \[Gaussiannull\] that if $\bm{y}$ is in the null space of $\bm{Z}_{n,n-1}$, then each $y_i$ can be constructed to be drawn from an i.i.d. standard Gaussian distribution. Based on this observation, the problem is further reduced to find the lower bound on the $k^{th}$ largest $y_i$, assuming that $2k$ is smaller than $n$ to ensure that with overwhelming probability the $k^{th}$ largest $y_i$ is non negative. Let us refer to this statistic as the $(n-k+1)^{th}$ order statistic of $\bm{y}$ and denote it by $y_{(n-k+1)}$ such that $y_{(1)} \leq y_{(2)} \leq \dots \leq y_{(n)}$. We present a lower bound for $y_{(n-k+1)}$ in Lemma \[lemma2\] when $k$ is smaller than $\frac{n}{2}$. This lower bound facilitates the final proof for Theorem \[theorem2\]. The proof of Lemma \[Gaussiannull\] and Lemma \[lemma2\] is left in the appendix.
\[Gaussiannull\] The basis of the null space of $\bm{Z}_{n,n-1}^{\text{T}}$ can be constructed as an i.i.d. standard Gaussian vector with length $n$, when each element of $\bm{Z}_{1:n-2}$ is independently drawn from a standard Gaussian distribution and the last column of $\bm{Z}$ is an all one column. [$\blacksquare$]{}
\[lemma2\] Let $\bm{y}$ satisfies $\bm{Z}_{n,n-1}^{\text{T}}\bm{y} = 0$. If each $y_i$ is independent and follows standard Gaussian distribution, then $\operatorname{\mathbb{E}}y_{(n-k+1)} \geq C\sqrt{\log{\frac{n}{k}}}$, where C is a positive constant and $\frac{k}{n} < \frac{1}{2}$. [$\blacksquare$]{}
Now we can use the introduced lemmas to prove Theorem \[theorem2\]. A summary is presented in Algorithm \[algo0\], illustrating the procedures to obtaining the rate stated in Theorem \[theorem2\] using the proposed lemmas. This algorithm also provides the optimal assignment strategy when $p = n - 1$.
We will focus on the case when $p = n-1$ since it provides the worst case rate for every $p < n$, as stated in Lemma \[increasing\].
From lemma \[Gaussiannull\], we know that $\bm{y} \sim N(0, \text{I}_{n})$, we then obtain the following results: $$\label{eq:optimalvar}
\begin{aligned}
& \underset{\bm{x}, x_i \in \{1,0\}, \sum x_i = k}{\text{max}}\operatorname{\mathbb{E}}\frac{(\bm{y}^{\text{T}}\bm{x})^2}{||\bm{y}||_2^2} \\
& \geq \operatorname{\mathbb{E}}\frac{\{(y_{(n)}+...+y_{(n-k+1)})^2\}}{||\bm{y}||_2^2}\\
& \overset{(a)} = \frac{\operatorname{\mathbb{E}}\{(y_{(n)}+...+y_{(n-k+1)})^2\}}{\operatorname{\mathbb{E}}||\bm{y}||_2^2} + O(\frac{1}{n})\\
& \geq \frac{\operatorname{\mathbb{E}}\{(ky_{(n-k+1)})^2\}}{n} + O(\frac{1}{n})\\
& = k^2 \frac{\operatorname{\mathbb{E}}\{y^2_{(n-k+1)}\}}{n} + O(\frac{1}{n})\\
& \overset{(b)} \geq k^2 \frac{(\operatorname{\mathbb{E}}y_{(n-k+1)})^2}{n} + O(\frac{1}{n})\\
& \overset{(c)}\geq k^2 \frac{C^2log\frac{n}{k}}{n} + O(\frac{1}{n}),\\
\end{aligned}$$ where ($a$) is based on the multivariate delta method [@deltamethod], ($b$) comes from Jensen’s inequality and ($c$) is based on Lemma \[Gaussiannull\] and Lemma \[lemma2\].
Specifically, if $\frac{k}{n} = \rho$, then can be written as: $$k^2 \frac{C^2log\frac{n}{k}}{n} = C^2\rho^2\log{(\rho^{-1}})n = \Theta(n).$$
Another interesting case is when $k = \log{n}$, can be written as: $$k^2 \frac{C^2\log{\frac{n}{k}}}{n} = C^2(\log{n} - \log{\log{n}}) = \Theta(\log{n}).$$
As can be seen from Theorem \[theorem2\], we will obtain the optimal rate $\Theta(n)$ by replacing $\frac{k}{n}$ as a constant. This indicates that with optimal assignment, the estimation variance will indeed decay with $n$ instead of being a constant as shown in Theorem \[theorem1\], when the dimension of the covariates $p$ is comparable to $n$. This is very interesting as it indicates that even in a high dimensional setting, the variance of the estimator will decay optimally by solving the variance minimization problem.
A TRACTABLE ALTERNATIVE {#sec:SDP}
=======================
Algorithm \[algo0\] provides a simple way to find the optimal assignment when $p = n - 1$. When $p$ is less than $n-1$, this algorithm cannot be applied. In fact, the optimization problem in is a nonconvex quadratic optimization problem which can be NP-hard. In this section, we present a tractable approximate algorithm by relaxing the original combinatorial optimization problem into and semidefinite program. We then demonstrate that this SDP problem approximates the original problem with a performance ratio that is better than $\frac{k}{n}$ when $\frac{k}{n}$ is in the range of $(0.2, 0.9995)$. This procedure follows the results established in [@SDP].
We first revisit the original variance minimization problem in and transform $\bm{x}$ into $\bm{x} = (\hat{\bm{x}} + 1)/2$. Denote each element in $\hat{\bm{x}}$ as $\hat{x}_i$, then each $\hat{x}_i$ takes value in $\{-1,1\}$. Therefore the variance minimization problem is written as: $$\label{prob:original}
\begin{aligned}
& \underset{\hat{\bm{x}}}{\text{maximize}}
&& \frac{1}{4}(\hat{\bm{x}}+1)^{\text{T}}P_{\bm{Z}^{\perp}}(\hat{\bm{x}}+1)\\
& \text{subject to}
&& \sum_{i=1}^{n}\hat{x}_i = 2k-n\\
&&& \hat{x}_i \in \{1,-1\}.
\end{aligned}$$
This is a Dense-$k$-Subgraph (DSP) problem with existence of self edges. To illustrate this, let element at row $i$ and column $j$ in matrix $P_{\bm{Z}^{\perp}}$ be denoted as edge weight $w_{ij}$ (except that $w_{ii}$ is half of the value on the diagonals) associated with vertex $i$ and vertex $j$, then is trying to find a set of $k$ vertices such that the sum of edge weights induced by these vertices are maximized. Since the problem presented in contains binary variables, we relax this problem into a SDP formulation: $$\label{eq:SDP}
\begin{aligned}
& \underset{\bm{X}, \hat{\bm{x}}}{\text{maximize}}
& & \frac{1}{4}\sum_i\sum_j w_{ij}(1+\hat{x}_i+\hat{x}_j+X_{ij})\\
& \text{subject to}
&& \sum_i \hat{x}_i = 2k-n\\
&&& X_{ii} = 1\\
&&& \sum_i\sum_j X_{i,j} = (2k-n)^2 \\
&&& \begin{bmatrix}
1 & \hat{\bm{x}}^{\text{T}}\\
\hat{\bm{x}} & \bm{X}
\end{bmatrix} \succeq 0.
\end{aligned}$$
The original problem in is hard, we will only obtain a surrogate solution in polynomial time. We thus adopt Algorithm \[algo\] to obtain an approximate solution from SDP formulation. Let us denote this solution by $\hat{\bm{x}}^*$ based on Algorithm \[algo\]. The performance of the approximation from SDP is evaluated by the performance ratio $r$ which satisfies: $$\label{ratioex}
\operatorname{\mathbb{E}}\frac{1}{4}(\hat{\bm{x}}^*+1)^{\text{T}}P_{\bm{Z}^{\perp}}(\hat{\bm{x}}^*+1)
\geq rw^*.$$
Here the randomness in $\hat{\bm{x}}^*$ is introduced by the random rounding procedure shown in Algorithm \[algo\] and $w^*$ is the optimal value of the objective function shown in . Performance ratio $r$ can be used to quantify how close the solution from Algorithm \[algo\] is to the optimal solution by solving the original hard problem.
There exists a fruitful line of work on the approximation algorithms using either greedy algorithm or LP/SDP relaxation for DSP problems [@KortsarzPeleg12; @KortsarzPeleg8; @Asahiro3; @FeigLangber6; @FeigSeltser7; @SrivastavWolf16; @YeZhang19; @SDP; @DkP14]. Most recent research has improved the performance ratio to $O(n^{-\frac{1}{4}+\epsilon})$ with LP relaxation in [@DkP14]. However, if $\frac{k}{n}$ is not decaying with $n$, i.e., a constant, then this ratio is not desirable since it is decreasing in $n$. The authors in [@SDP] propose an improved performance ratio that is better than $\frac{k}{n}$ for a wide range of $\frac{k}{n}$. We will adopt the approximation procedure in [@SDP] and argue that the performance ratio is valid in our case here as well. In the following, we will first present the general algorithm and then show that the performance ratio in [@SDP] is still applicable in our case.
The approximation procedure in [@SDP] is presented in Algorithm \[algo\], including three main procedures:
- Solve SDP problem in (step 1). After this procedure we can obtain the optimal continuous solution for and the optimal value of the objective function is denote by $w^{SDP}$.
- Construct initial $S$, where $S$ represents the initial subgraph and is a set of indices (step 2 through step 4). The $\hat{x}_i$’s take value 1 such that $i \in S$ and the rest -1. Let us denote them by $\hat{\bm{x}}^{0}$. The value of the objective function in is written as $w(S) = \frac{1}{4}\sum_i\sum_j w_{ij}(1+\hat{x}_i^{0}+\hat{x}_j^{0}+\hat{x}_i^{0}\hat{x}_j^{0}) = \frac{1}{4}(\hat{\bm{x}}^{0}+1)^{\text{T}}P_{\bm{Z}^{\perp}}(\hat{\bm{x}}^{0}+1)$. Here $w({S})$ is the total weights of edges in the subgraph induced by ${S}$. At this point the cardinality of $S$ is not necessarily $k$.
- Resize $S$ to $\tilde{S}$ such that $\tilde{S}$ contains exactly $k$ vertices (step 5 through step 16). The final assignment of $\hat{x}_i$’s is that $\{\hat{x}_i = 1, \hat{x}_j = -1| i \in \tilde{S}, j \notin \tilde{S}\}$. Let us denote them by $\hat{\bm{x}}^{*}$. The value of the objective function is $w(\tilde{S}) = \frac{1}{4}(\hat{\bm{x}}^{*}+1)^{\text{T}}P_{\bm{Z}^{\perp}}(\hat{\bm{x}}^{*}+1)$ and is the total weights of the edges induced by $\tilde{S}$.
Solve SDP in , obtain $\bm{X}, \hat{\bm{x}}$.\
Construct $\bar{\bm{X}} = \begin{bmatrix}
1 & \hat{\bm{x}}^{\text{T}}\\
\hat{\bm{x}} & \bm{X}
\end{bmatrix}$.\
Construct covariance matrix $Y = \theta \bar{\bm{X}} + (1-\theta)\bm{P}$, where: $0\leq\theta \leq 1$, $\bm{P} = \begin{bmatrix}
1 & \chi & \chi & \dots & \chi \\
\chi & 1 & \chi^2 & \dots & \chi^2 \\
\vdots & \vdots & \vdots & \ddots & \vdots \\
\chi & \chi^2 & \dots & \chi^2 & 1
\end{bmatrix}$, $\chi = 2\frac{k}{n}-1$.\
Generate $\bm{u} \sim N(0,\bm{Y})$, $\tilde{\bm{x}} = sign(\bm{u})$, $S = \{i \geq 2:\tilde{x}_i = \tilde{x}_1\}$.\
Let $\tilde{S} = S$.\
We now demonstrate that Algorithm \[algo\] achieves the performance ratio shown in Proposition \[propositionratio\].
\[propositionratio\] The performance ratio $r$ from Algorithm \[algo\] defined as: $$\operatorname{\mathbb{E}}w(\tilde{S}) \geq r w^*$$ satisfies the conditions presented in Proposition 2 in [@SDP] and is plotted in Fig. \[figratio\], where $w^*$ is the optimal value of the objective function in problem . When $k$ is large, this ratio is better than either O($n^{-(\frac{1}{3}-\epsilon)}$) or O($n^{-(\frac{1}{4}-\epsilon)}$) obtained by LP relaxation. [$\blacksquare$]{}
As can be seen from Proposition \[propositionratio\], the performance ratio $r$ quantifies the gap between the approximated solution obtained by Algorithm \[algo\] and the optimal solution from the original problem shown in . This ratio $r$ is the direct result from the random rounding procedure and the resizing procedure shown in Algorithm \[algo\]. The ratios associated with these two procedures are presented in and :
$$\label{bounda}
\operatorname{\mathbb{E}}w(S) \geq \alpha w^*,$$
and $$\label{boundb}
w(\tilde{S}) \geq \xi w(S).$$
In [@SDP], the authors well define the parameter $\alpha$ (which depends on $k$ and $n$) and $\xi$ (depends on $k$ and $S$ obtained from the random rounding procedure) so that the performance ratio $r$ satisfies Proposition \[propositionratio\] where there are no self edges, i.e., $w_{ii} = 0$. However, in our case, the $w_{ii}'s $ are the diagonal elements of $\bm{P}_{\bm{Z}^{\perp}}$ and they are not necessarily zero, so the graph in our case contains non trivial self edges. If we show that the presence of self edges does not change the values of $\alpha$ and $\xi$, then Proposition \[propositionratio\] naturally holds in our case as well.
Let us first discuss . From [@SDP], parameter $\alpha$ does not depend on whether there are self edges in the graph, so directly applies.
Next we need to check if the same $\xi$ applies at the presence of self edges. When there are no self edges, i.e., $w_{ii} = 0$, the authors in [@SDP] show that $\xi = \frac{k(k-1)}{|S|(|S|-1)}$ when $|S|>k$ and $\xi = 1$ otherwise. We show that the same condition for $\xi$ holds even with the presence of non negative self edges, as stated in Lemma \[selfedge\].
\[selfedge\] Let $S$ and $\tilde{S}$ be obtained from random rounding procedure and resizing procedure in Algorithm \[algo\] from a graph with non negative self edges, i.e., $w_{ii} \geq 0$. Then we have: $$w(\tilde{S})=
\begin{cases}
\frac{k(k-1)}{|S|(|S|-1)}w(S),& \text{if } |S| > k\\
1, & \text{otherwise}
\end{cases}$$ [$\blacksquare$]{}
Proof of Lemma \[selfedge\] is given in the appendix. Lemma \[selfedge\] validates the $\alpha$ and $\xi$ with the presence of non negative self edges, thus the performance ratio $r$ stated in Proposition \[propositionratio\] is valid. A list of performance ratio $r$’s for different values of $\frac{k}{n}$ is shown in [@SDP] and is plotted in Fig. \[figratio\]. This rate is satisfactory since it has a rate of $O(\frac{k}{n})$ but strictly larger than $\frac{k}{n}$ [@SDP]. It is better than O($n^{-\frac{1}{4}+\epsilon}$) when $\frac{k}{n}$ is not decaying faster than $O(n^{-\frac{1}{4}+\epsilon})$. In fact, as long as $\frac{k}{n}$ lies within some constant range, then $\frac{k}{n}$ can be seen as a varying constant thus is not decaying as a function of $n$.
SIMULATION
==========
In this section we show the comparison results between random assignment and optimal assignment. We simulate the covariates from two different distributions, i.e., Gaussian distribution and uniform distribution. We also validate the claim by simulated building data from [@sim].
Gaussian Ensemble
-----------------
We first generate the covariates as they are drawn from i.i.d. Gaussian ensemble, i.e., $N(\bm{0}, {I})$. We compare two cases where $n = 3k$ and $n = 5k$ in Fig. \[Gaussian\]. Note that Fig. \[Gaussian\] is shown in semilogarithmic plot where the $y$-axis has a logarithmic scale and the $x$-axis has a linear scale. In addition, we adopt the value of $\theta$ in [@SDP], i.e., 0.9 for $n = 3k$ and 0.94 for $n = 5k$. We let $p = n -1$ to obtain the worst case performance.
Besides SDP relaxation, we also simulate a greedy based assignment to maximize the weighted edges induced by $k$ vertices. The greedy assignment sequentially eliminates vertices and works as follows: we start with the original graph and a set containing all vertex. At each elimination step, the vertex with the least weighted edges are eliminated from the set until this set contains exactly $k$ vertices. This greedy algorithm is introduced in [@Asahiro3].
In addition, we use the result from branch and bound (upper bound) to serve as the reference in order to compare the random assignment and the proposed optimal assignment. The duality gap for branch and bound is set to be 0.05 for all $n$ when $n = 3k$. Due to computational complexity and time constraints, we set this gap to be around 0.25 when $n$ is big in the case of $n = 5k$.
In Fig. \[Gaussian\], we see that the semilog plot on $\operatorname{\text{Var}}^{-1} \hat{\beta}$ ($\bm{x}^{\text{T}}P_{\bm{Z}_{n,n-1}^{\perp}}\bm{x}$) is growing with $n$ and similar to $\log n$. This suggests that $\operatorname{\text{Var}}^{-1} \hat{\beta}$ is linear in $n$, as we stated in Section \[optimal\]. In addition, It is very close to the solution obtained by branch and bound, meaning that the result from SDP relaxation is close to the optimal solution in . The empirical performance ratio from SDP relaxation is shown in Table I.
From Table I, we see that the performance ratio for $n$ between $10$ and $200$ is actually greater than $\frac{k}{n}$, which is even better than the theoretical bound in Proposition \[propositionratio\]. In addition, if we decrease $k$ with respect to $n$, i.e., change $k$ from $\frac{n}{3}$ to $\frac{n}{5}$, then the performance ratio is reduced. This is due to the fact we need to do more eliminations during the resizing procedure in Algorithm \[algo\] and the deterioration increases. However, if $\frac{k}{n}$ is well defined within a range, then this deterioration is controlled and does not change the statement in Proposition \[propositionratio\].
On the other hand, the semilog plot on $\operatorname{\text{Var}}^{-1} \hat{\beta}$ from random assignment is a constant on average across different values of $n$, whether $k$ is small or large. This validates Theorem \[theorem1\] as it states that the rate is $\Theta$($1$) when $p = n-1$. In this case, we cannot obtain an efficient estimator since the variance is not decaying even when $n$ is big.
What is more, the greedy assignment does not provide a relatively good performance as well. It yields a constant $\operatorname{\text{Var}}\hat{\beta}$ as random assignment. This suggests although both the greedy algorithm and SDP relaxation are aimed to solve an optimization problem, the solution from SDP relaxation is much better and reliable.
\[table1\]
------------------------------------------------------------------------------- -- -- -- --
**$\rho = \frac{k}{n}$ & **$n$ = 10 & **$n$ = 50 & **$n$ = 100 & **$n$ = 200\
$\frac{1}{3}$ & 0.8131 & 0.6486 & 0.7416 & 0.6588\
$\frac{1}{5}$ & 0.7026 & 0.5362 & 0. 6577 & 0.5273\
**********
------------------------------------------------------------------------------- -- -- -- --
: Empirical performance ratio for Gaussian ensemble with different values for $\frac{k}{n}$ and varying $n$.
Uniform Ensemble
----------------
Although we discuss the rate of quantity $\bm{x}^{\text{T}}P_{\bm{Z}_{n,n-1}^{\perp}}\bm{x}$ with respect to Gaussian ensemble, we also simulate the covariates where the elements are drawn from a uniform distribution in an interval \[-1,1\]. The results are shown in Fig. \[Uniform\]. Fig. \[Uniform\] is again a semilogarithmic graph. We again take $\theta$ to be 0.9 when $n = 3k$ and 0.94 when $n = 5k$. The duality gap is 0.05 when $n = 3k$. When $n = 5k$, the duality gap is around 0.1 to 0.2 for $n$ greater than 100 and is 0.25 for $n = 200$. The comparison result is shown in Fig. \[Uniform\] and the performance ratio by SDP relaxation is shown in Table II.
\[table2\]
------------------------------------------------------------------------------- -- -- -- --
**$\rho = \frac{k}{n}$ & **$n$ = 10 & **$n$ = 50 & **$n$ = 100 & **$n$ = 200\
$\frac{1}{3}$ & 0.6961 & 0.6755 & 0.8068 & 0.8094\
$\frac{1}{5}$ & 0.6101 & 0.6113 & 0.4799 & 0.5145\
**********
------------------------------------------------------------------------------- -- -- -- --
: Empirical performance ratio for uniform ensemble with different values for $\frac{k}{n}$ and varying $n$.
The observation from Fig. \[Uniform\] and Table II is similar to the analysis in the case of Gaussian ensemble, that the solution obtained from SDP relaxation is still within a constant of the branch and bound solution for $10 \leq n \leq 200$. Again, greedy algorithm fails to find a solution close to that by branch and bound and performs as poorly as the random assignment. The performance ratio is again decreased when we decrease $k$, which indicates the similar deterioration occurred during the resizing procedure.
Building data
-------------
We also validate the claim in the paper by simulated building data obtained from [@sim]. Using this software, we can generate building covariates such as environment temperature, number of occupants, appliances scheduling, etc. The software outputs the energy consumption based on the covariates that we generate. The buildings vary from a small postal office, to a large commercial hotel, with different number of covariates involved in the modeling.
For the purpose of this paper, we include significant covariates into the linear regression model and the number of those covariates are comparable to the number of users in the simulation. The values of the covariates are whitened by the covariance matrix so that the covariates have zero mean and unit variance [@CarlosEtAl2016]. The details of this procedure is given in the appendix. In addition, we fix $k = \frac{1}{3}n$ in the simulation.
The simulation results are shown in Fig. \[building\] and Fig. \[varying\_d\].
In Fig. \[building\], we observe that $\operatorname{\text{Var}}\hat{\beta}$ is decaying very fast with the optimal assignment strategy, whereas the variance stays unchanged if adopting a random assignment strategy.
The variance $\operatorname{\text{Var}}\hat{\beta}$ is a varying quantity in terms of the dimension of the covariates, i.e., $p-1$. Assume that the number of users are fixed, i.e., $n$ is fixed, and that $\rho$ is a constant. According to Theorem \[theorem1\] and in Theorem \[theorem2\], with random assignment the variance decays with $n-p$ whereas with optimal assignment the worse case scenario with $p = n - 1$ yields a decay rate of $n$.
In Fig. \[varying\_d\], we show how $\operatorname{\text{Var}}\hat{\beta}$ varies with an increasing $p$ and a fixed $n = 50$. As can be seen from Fig. \[varying\_d\], the increasing $p$ deteriorates the variance much severely from random assignment than that from the proposed optimal assignment. This again validates the efficiency of the proposed strategy in improving estimation accuracy, especially in a high dimensional setting where $p$ is comparable to $n$.
CONCLUSION
==========
In this paper, we estimate the average treatment effect of demand response (DR) signals. We adopt an additive linear regression model and discuss two different strategies to assign DR signals to users under limited assignment budgets. The first strategy randomly picks $k$ users and sends DR signals to them. The second strategy optimally assigns DR signals to $k$ users by minimizing the variance to estimate treatment effect. We show that in a high dimensional setting, the second strategy achieves order optimal rates in variance reduction, whereas random assignment does not reduce variance even as the number of users grows. We formulate the general assignment as a combinatorial optimization problem and present a tractable SDP relaxation. We show that this relaxation obtains a solution that is within bounded gap of the original optimal solution. The simulation results validate this proposition with the synthetic data on both i.i.d. Gaussian covariates and uniform covariates. This work provides a framework for further research in applying causal inference in analyzing consumption data and DR interventions.
Appendix {#appendix .unnumbered}
========
Proof of Lemma \[Gaussiannull\]
-------------------------------
This proof follows the intuition in [@nullspace]. In [@nullspace], to prove the null space of a Gaussian random matrix, the authors use the fact that a standard multivariate Gaussian distribution is invariant to any orthogonal transform. The difference here is that the matrix $\bm{Z}$ contains an extra column of 1’s as intercept.
Denote the null space of $\bm{Z}_{n,n-1}^{\text{T}}$ by $\bm{y}$. It is one dimensional and satisfies: $$\label{eq:null}
\begin{bmatrix}
\bm{Z}_{1:n-2}^{\text{T}}\\
\bm{1}^{\text{T}}
\end{bmatrix} \bm{y} = \bm{0}.$$
Now we multiply $\bm{y}$ by an orthogonal matrix $\bm{U}_{n,n}^{\text{T}}$ on the left hand side and multiply $\bm{Z}_{n,n-1}^{\text{T}}$ by $\bm{U}_{n,n}$ on the right hand side. For simplicity let us write $\bm{U}_{n,n}$ as $\bm{U}$. Since $\bm{U}\bm{U}^{\text{T}} = {I}$ then the following holds: $$\label{eq:transform}
\begin{bmatrix}
\bm{Z}_{1:n-2}^{\text{T}}\\
\bm{1}^{\text{T}}
\end{bmatrix} \bm{U}\bm{U}^{\text{T}}\bm{y} =
\begin{bmatrix}
\bm{Z}_{1:n-2}^{\text{T}}\bm{U}\\
\bm{1}^{\text{T}}\bm{U}
\end{bmatrix} \bm{U}^{\text{T}}\bm{y}
= \bm{0}.$$
Because $\bm{Z}_{1:n-2}^{\text{T}}$ has i.i.d. standard Gaussian entries and $\bm{U}$ is an orthogonal matrix, then each row of $\bm{Z}_{1:n-2}^{\text{T}}\bm{U}$ must follow a Gaussian distribution $N(\bm{0}\bm{U}, \bm{U}{I}\bm{U}^{\text{T}})$, which is still a standard Gaussian distribution. Thus $\bm{Z}_{1:n-2}^{\text{T}}\bm{U}$ has i.i.d. standard Gaussian entries. Denote $\tilde{\bm{Z}} = \bm{Z}_{1:n-2}^{\text{T}}\bm{U}$, then $\tilde{\bm{Z}} $ and $\bm{Z}_{1:n-2}$ are drawn from the same distribution and each of their entries follows i.i.d. standard Gaussian distribution.
From , we also require that the orthogonal matrix $\bm{U}$ satisfies $\bm{1}^{\text{T}}\bm{U} = \bm{1}$. If such orthogonal matrix exists (which is easy to find), then we can rewrite as: $$\label{eq:newnull}
\begin{bmatrix}
\tilde{\bm{Z}}^{\text{T}}\\
\bm{1}^{\text{T}}
\end{bmatrix} \bm{U}^{\text{T}}\bm{y}
= \bm{0}.$$
Comparing and , we see that $\bm{U}^{\text{T}}\bm{y}$ and $\bm{y}$ must be identically distributed. One distribution that satisfies this property is the standard normal distribution $N(\bm{0}, {I})$. It is easy to show that identity matrix is the only covariance matrix that satisfies such condition whereas zero mean is based on the fact that $\bm{1}^{\text{T}}\bm{y} = 0$ from . This observation concludes the final proof.
Proof of Lemma \[lemma2\]
-------------------------
The proof is based on the fact that if in expectation, at least $k$ $y_i$’s is greater than a constant, then the $k$th largest $y_i$ must be greater than this constant [@MaxGaussian]. Mathematically speaking, we want $\operatorname{\mathbb{E}}\{ |i:y_i \geq C \sqrt{\log{\frac{n}{k}}}, \forall i|\} \geq k$, where $C$ is a constant. We write $\operatorname{\mathbb{E}}\{ |\{i:y_i \geq C \sqrt{\log{\frac{n}{k}}}\}|\}$ in the following form: $$\label{lowerbound1}
\begin{aligned}
\operatorname{\mathbb{E}}\{ |\{i:y_i \geq C \sqrt{\log{\frac{n}{k}}}\}|\} & = n \text{Pr} \{y_i \geq C \sqrt{\log{\frac{n}{k}}}\} \\
& = nQ(C \sqrt{\log{\frac{n}{k}}}).\\
\end{aligned}$$
$Q$ function of Gaussian distribution does not have a close form expression, so we are interested in a tight lower bound for it in order to obtain a lower bound for . Many lower bounds are obtained in literature [@05470020; @1202.6483; @erf-approx; @j64]. In particular, [@erf-approx] provides a lower bound that is valid when the argument is small, and [@05470020] provides a lower bound that is the tightest when the argument becomes relatively large. The lower bounds (bound 1 through bound 4, from [@05470020; @1202.6483; @erf-approx; @j64] respectively) are presented in through .
$$\begin{aligned}
\label{bound1}
Q(x) & \geq \frac{1}{12}e^{-x^2}+\frac{1}{\sqrt{2\pi}(x+1)}e^{-\frac{x^2}{2}},\\
\label{bound2}
Q(x)&\geq \frac{e^{\frac{1}{\pi(\kappa-1)+2}}}{2\kappa}\sqrt{\frac{1}{\pi}(\kappa-1)(\pi(\kappa-1)+2)}e^{-\frac{\kappa x^2}{2}}, \kappa \geq 1, \\
\label{bound3}
Q(x) & \geq \frac{1-\sqrt{1-e^{-\frac{2x^2}{\pi}}}}{2},\\
\label{bound4}
Q(x) & \geq \sqrt{\frac{e(\beta-1)}{2\pi\beta^2}}e^{-\frac{\beta x^2}{2}}, \beta \geq 1.
\end{aligned}$$
Note that bound 3 (in ) is only valid when $x$ is small. A comparison of these four bounds are shown in Fig.\[lowerbound\]. It is shown in semilogarithmic plot where $y$-axis is in a logarithmic scale and $x$-axis is in a linear scale.
Assume that $C \sqrt{\log{\frac{n}{k}}}$ is small, i.e., when $\frac{k}{n} = \frac{1}{2} - \epsilon$ which is slightly smaller than $\frac{1}{2}$ but greater than $\frac{1}{4}$. In this range the lower bound provided in is the tightest. Use this lower bound we obtain the constant $C$ that is universally applicable for that holds when $\frac{n}{k}$ is small. After some simple calculation, we obtain the constant $C = \sqrt{\frac{\pi \log{(1-4\epsilon^2)}}{2\log{(\frac{1}{2}-\epsilon)}}}$.
Although provides the tightest bound when $C \sqrt{\log{\frac{n}{k}}}$ gets bigger, this lower bound includes two exponential terms which complicate the calculation. Actually when $\frac{k}{n}\leq \frac{1}{4}$, we can obtain a fairly good but conservative constant $C$ for the lower bound provided in with only one exponential term. The constant $C$ in this case when $\frac{n}{k}$ is relatively large is calculated as $\sqrt{1-\frac{\log{8\pi}-1}{2\log{4}}} \approx 0.445$.
Proof of Lemma \[selfedge\]
---------------------------
Before proving Lemma \[selfedge\], let us first introduce Lemma \[wii\] that bounds the values of the diagonals of $P_{z^{\perp}}$.
\[wii\] The diagonals of $P_{z^{\perp}}$ is non negative.
Recall that $\bm{P}_{\bm{z}^{\perp}} = \bm{I} - \bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}$, and $\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}$ is a projection matrix, then we have the following: $$\begin{aligned}
(\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}})^2 & = \bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}.\\
\end{aligned}$$
Denote the diagonals of the $\bm{Z}(\bm{Z}^{\text{T}}\bm{Z})^{-1}\bm{Z}^{\text{T}}$ by $v_{ii}$, then we have $v_{ii} = \sum_j v_{ij}^2 = v_{ii}^2 + \sum_{j \neq i} v_{ij}^2$, which implies that $0 \leq v_{ii} \leq 1$.
This suggests that the diagonals of $P_{z^{\perp}}$, denoted by $w_{ii}$, is non-negative.
Now we proceed to prove Lemma \[selfedge\].
\[Proof of Lemma \[selfedge\]\]
To prove that Lemma \[selfedge\] holds, we add up a constant to each $w_{ij}$ to make sure that there are no negative weights in the graph. This does not change the optimization problem since it only adds a constant to the objective function and the solution remains the same.
If $|S| \leq k$, then we arbitrarily add vertices to $\tilde{S}$ until it contains exactly $k$ vertices. Since the weight on each edge is non negative and we keep adding more edges into the subgraph induced by $\tilde{S}$, $\xi$ is at least 1 in this case.
Now suppose that $|S| > k$, in this case we need to eliminate vertices from $\tilde{S}$ until it only contains $k$ vertices. Assume that we want to eliminate vertex $i$ from $\tilde{S}$, then the total weights induced by $\tilde{S} \setminus\{ i\} $ are: $$w(\tilde{S}) - (\sum_{j \in \tilde{S}, i \neq j} w_{ij} + w_{ii}).$$
If each vertex is removed once, then: $$\begin{aligned}
\sum_{i \in \tilde{S}} w(\tilde{S}\setminus \{i\}) & = \sum_{i \in \tilde{S}} w(\tilde{S}) - (\sum_{j \in \tilde{S}, i \neq j} w_{ij} + w_{ii}) \\
& = (|\tilde{S}|-2)w(\tilde{S})+\sum_{i \in \tilde{S}}w_{ii}.
\end{aligned}$$
The last equality is because each non-self edge is counted twice during the removal (once at the count of each vertex), but self-edge is only counted once.
Similarly, suppose that $v$ is the node that is removed during the swapping procedure, then according to the swapping procedure in Algorithm \[algo\]: $$\sum_{j \in \tilde{S}} w_{vj} \leq \sum_{j \in \tilde{S}}w_{ij}, \forall i \neq v \in \tilde{S}.$$
Then: $$\begin{aligned}
w(\tilde{S}\setminus \{v\}) & \geq \frac{1}{|\tilde{S}|} \sum_{i \in \tilde{S}}w(\tilde{S}\setminus \{i\}) \\
& = \frac{1}{|\tilde{S}|}((|\tilde{S}|-2)w(\tilde{S})+\sum_{i \in \tilde{S}}w_{ii}) \\
& = \frac{|\tilde{S}|-2}{|\tilde{S}|}w(\tilde{S}) + \frac{\sum_{i \in \tilde{S}}w_{ii} }{|\tilde{S}|} \\
& \geq \frac{|\tilde{S}|-2}{|\tilde{S}|}w(\tilde{S}).
\end{aligned}$$
The last inequality follows because $w_{ii}$’s are non-negative, which is proved in Lemma \[wii\].
Finally by induction, we obtain the eventual $\tilde{S}$ containing $k$ vertices satisfying: $$w(\tilde{S}) \geq \frac{k(k-1)}{|S|(|S|-1)}w(S),$$ which concludes the final proof.
Whitening of covariate data
---------------------------
Suppose that the covariates generated from the simulation tool in [@sim] has a covariance matrix $\Sigma$. We can decompose $\Sigma = \bm{V}\bm{D}\bm{V}^{\top}$， where $\bm{D}$ is a diagonal matrix with diagonal elements as the eigenvalues of $\Sigma$, and $\bm{V}$ is an orthogonal matrix containing the corresponding eigenvectors. Suppose that each observation of the covariates is denoted by $\bm{z}$, then the whitened observation $\bar{\bm{z}} = \bm{D}^{-\frac{1}{2}}\bm{V}^{\top}\bm{z} \in \mathbb{R}^{p-1}$, where $p-1$ is the dimension of the covariates. For the original covariate matrix $\bm{Z}$, it is whitened as $\bar{\bm{Z}} = \bm{Z}\bm{V}\bm{D}^{-\frac{1}{2}}$. The whitened covariates are then fed into the optimization problem to find the optimal assignment strategy.
[^1]: This work was supported by NSF grant CNS-1544160 and the University of Washington Clean Energy Institute.
[^2]: The authors are with the Department of Electrical Engineering, University of Washington, Seattle, WA 98195, USA (e-mail: {pli69,zhangbao}@uw.edu).
[^3]: Partial results appeared in an earlier version of this paper presented in the Allerton Conference, 2016.
[^4]: no estimator can reduce variance faster than $1/n$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that for all $n \geq 2$, there exists a doubling linearly locally contractible metric space $X$ that is topologically a $n$-sphere such that every weak tangent is isometric to ${\mathbb{R}}^n$ but $X$ is not quasisymmetrically equivalent to the standard $n$-sphere. The same example shows that $2$-Ahlfors regularity in Theorem 1.1 of [@BK02] on quasisymmetric uniformization of metric $2$-spheres is optimal.'
address: |
Angela Wu\
Department of Mathematics\
University of California\
Los Angeles\
CA90095\
USA
author:
- Angela WU
bibliography:
- '/home/angela/Desktop/Dropbox/Research/Recurrence\_Self\_Similar\_Graphs/Biblography.bib'
date: 'June 7, 2018'
title: A metric sphere not a quasisphere but for which every weak tangent is Euclidean
---
Introduction
============
Let $(X, d_X)$ and $(Y, d_Y)$ be two metric spaces. A homeomorphism $f: X \to Y$ is a quasisymmetry if there exists a homeomorphism $\eta:[0,\infty) \to [0,\infty)$ such that for all $x,y,z \in X$, with $x \neq z$, we have $$\begin{aligned}
\frac{d_Y(f(x), f(y))}{d_Y(f(x), f(z))} \leq \eta{\left( \frac{d_X(x,y)}{d_X(x,z)} \right)}.\end{aligned}$$ In this case we say that $X$ and $Y$ are quasisymmetrically equivalent.
In [@Kinneberg17] Kinneberg characterized metric circles that are quasisymmetric to the standard circle in terms of weak tangents (defined in Section 4 below):
A doubling metric circle $\mathcal{C}$ is quasisymmetrically equivalent to the standard circle $\mathbb{S}^1$ if and only if every weak tangent of $\mathcal{C}$ quasisymmetrically equivalent to the real line ${\mathbb{R}}$ based at $0$.
Here a metric space is said to be doubling if for every $R > 0$, every ball $B$ of radius $2R$ can be covered by $N$ balls of radius $R$, where $N$ is a positive integer that does not depend on the choice of the ball $B$. In this paper we prove that Kinneberg’s result cannot be extended to higher dimensions:
\[Theorem\] For every $n \geq 2$, there exists a doubling, linearly locally contractible metric space $X$ that is topologically an $n$-sphere such that every weak tangent is isometric to ${\mathbb{R}}^n$ but $X$ is not quasisymmetrically equivalent to the standard $n$-sphere.
When $n = 2$, one can compare our result with the following Theorem:
\[Bonk\_Kleiner\] Let $Z$ be an 2-Ahlfors regular metric space homeomorphic to $\mathbb{S}^2$. Then $Z$ is quasisymmetric to $\mathbb{S}^2$ if and only if $Z$ is linearly locally contractible.
Our Theorem \[Theorem\] shows that the conclusion of Theorem \[Bonk\_Kleiner\] is false if we replace $2$-Ahlfors regularity with $Q$-Ahlfors regularity for $Q >2$.
Our study is also related to the following theorem, proven in [@Lindquist17]:
\[Intro\_Thm1\] Let $(Z,d)$ be a doubling metric space homeomorphic to $\mathbb{S}^2$. The following are equivalent:
(i) $(Z, d)$ is quasisymmetrically equivalent to the standard $2$-sphere.
(ii) For every $x \in Z$, there exists an open neighborhood $U$ of $x$ in $Z$ such that $U$ is quasisymmetrically equivalent to ${\mathbb{D}}$.
An alternative proof of Theorem \[Intro\_Thm1\] can be given using ideas in [@GW18]. Roughly speaking, Theorem \[Intro\_Thm1\] says local geometry properties promote to global property. Since weak tangents are local, one could ask the following question:
Suppose $(Z,d)$ is doubling and linearly locally connected. Are the following two statements equivalent?
(i) $Z$ is quasisymmetrically equivalent to the standard $d$-sphere $\mathbb{S}^d$
(ii) Every weak tangent of $Z$ is quasisymmetrically equivalent to ${\mathbb{R}}^d$.
When $Z$ is a doubling and linearly locally connected metric sphere, statement (i) implies statement (ii). However, our construction shows that statement (ii) does not imply statement (i).
The author would like to thank Mario Bonk and John Garnett for their support and many conversations. The author thanks Jeff Lindquist for helpful discussions and Wenbo Li for his comments and corrections.
The Modification of the Metric on ${\mathbb{R}}$
================================================
We first construct a metric $\delta$ on ${\mathbb{R}}$ so that every weak tangent of $({\mathbb{R}}, \delta)$ is isometric to $({\mathbb{R}}, 0, d)$, where $d$ is the Euclidean metric on ${\mathbb{R}}$, and so that there are segments $[a_j, b_j]$ of $({\mathbb{R}}, \delta)$ so that $\ell([a_j, b_j]) / \delta(a_j, b_j) \to +\infty$. Here, and in the remaining of the paper, $\ell(\gamma)$ denotes the length of the curve $\gamma$ with respect to the metric $\delta$. Once the metric $\delta$ is constructed, we show that for all $n \geq 2$, the weak tangents of the product space $({\mathbb{R}}, \delta) \times {\mathbb{R}}^{n-1}$ are isometric to the Euclidean space ${\mathbb{R}}^n$, but because $\ell([a_j, b_j]) / \delta(a_j, b_j) \to +\infty$, the product space $({\mathbb{R}}, \delta) \times {\mathbb{R}}^{n-1}$ cannot be quasisymmetric to the Euclidean space ${\mathbb{R}}^n$.
To construct $({\mathbb{R}}, \delta)$, we glue together segments $[a_j, b_j]$ that look like snowflake curves of Hausdorff dimension $(\alpha_j)^{-1}$ i.e. intervals $I$ equipped with the metric $d^{\alpha_j}$. Such snowflake curves will have infinite length, so we modify the metric on the segment $[a_j, b_j]$ to get a metric $d_n$ so that $d_n$ looks like the Euclidean metric when two points are close together, but it looks like $d^{\alpha_j}$ when the two points are far apart.
Note that our construction of $({\mathbb{R}}, \delta)$ only modifies the geometry of a bounded subset of ${\mathbb{R}}$, and therefore we can embed our $n$-dimensional construction into a topological $n$-sphere.
To begin the construction let $\alpha \in (0,1)$ and $c \in (0,1)$ and define $$\varphi_{\alpha, c}(x) = \begin{cases}
L(\alpha, c) x,& x \in [0,c] \\
{\left( \frac{x - c(1-\alpha)}{1 - c(1-\alpha)} \right)}^\alpha,& x \in [c,1],
\end{cases}$$ where $$L(\alpha, c) = \frac{1}{c} {\left( \frac{c\alpha}{1 - c(1-\alpha)} \right)}^\alpha.$$ The function $\varphi_{\alpha, c}$ is the only function that has the following properties:
1. $\varphi_{\alpha, c}(0) = 0$,
2. $\varphi_{\alpha, c}(1) = 1$,
3. $\varphi_{\alpha, c}$ is linear on $[0,c]$,
4. There exists $a,b \in {\mathbb{R}}$, with $a \neq 0$, such that $\varphi_{\alpha ,c}(x) = (ax - b)^\alpha$ when $x \in [c,1]$,
5. $\varphi_{\alpha, c}$ is continous and differentiable at $c$.
In addition to the above properties that uniquely define $\varphi_{\alpha, c}$, we observe that $\varphi_{\alpha, c}$ has the following extra properties:
1. $\varphi_{\alpha, c} :[0,1] \to [0,1]$ is a homeomorphism,
2. $\varphi_{\alpha, c}$ is concave on $[0,1]$.
\[lemma\_sup\_multiplicative\] For any $\alpha \in (0,1)$, $c \in (0,1]$, and $a, b \in [0,1]$, we have $$\varphi_{\alpha, c}(ab) \varphi_{\alpha, c}(a) \varphi_{\alpha, c}(b).$$
The function $L(\alpha, c)$ is decreasing in $c$ to $1$, therefore if $0 < c_1 \leq c_2 \leq 1$, then $$\varphi_{\alpha, c_1} \geq \varphi_{\alpha, c_2}.$$ Let $$\varphi_\ast(x) = \frac{\varphi_{\alpha, c}(ax)}{\varphi_{\alpha, c}(a)}.$$ The function $\varphi_\ast$ meets condition (1)-(5) listed above for some $c_\ast \in [c, 1]$, therefore $\varphi_{\ast} = \varphi_{\alpha, c_\ast}$. This implies that $\varphi_\ast \leq \varphi_{\alpha, c}(x)$ for all $x \in [0,1]$. Taking $x = b$, we have $$\varphi_{\alpha, c}(ab) \leq \varphi_{\alpha, c}(a)\varphi_{\alpha, c}(b).$$
\[lemma\_rough\_isometries\_1\] For any fixed $\alpha \in (0,1)$, for any $c \in (0,1)$, and whenever $0 \leq t \leq x \leq 1$, we have $$0 \leq \varphi_{\alpha, c}(t) + \varphi_{\alpha, c}(x-t) -\varphi_{\alpha, c}(x) \leq {\left( 2 - 2^{\alpha} \right)}\varphi_{\alpha, c}(x).$$
When $x\leq c$, for all $t \in [0,x]$, we have $\varphi_{\alpha, c}(t) + \varphi_{\alpha, c}(x-t) -\varphi_{\alpha, c}(x) = 0$. When $x \geq 2c$ is fixed, $\varphi_{\alpha, c}(t) + \varphi_{\alpha, c}(x-t) -\varphi_{\alpha, c}(x)$ is maximized when $$\varphi_{\alpha, c}'(t) - \varphi_{\alpha, c}'(x-t) = 0,$$ which is possible only if $x- t = t$ i.e. $2t = x$. Then we have $$\begin{aligned}
\varphi_{\alpha, c}(t) + \varphi_{\alpha, c}(x-t) -\varphi_{\alpha, c}(x)
&\leq 2\varphi_{\alpha, c}(x/2) - \varphi_{\alpha, c}(x) \\
&\leq 2\varphi_{\alpha, c}(x/2) - 2^{\alpha}\varphi_{\alpha, c}(x/2) \\
&= (2 - 2^{\alpha})\varphi_{\alpha, c}(x/2).\end{aligned}$$ When $c < x < 2c$, we have $$\varphi_{\alpha, c}(t) + \varphi_{\alpha, c}(x-t) \leq L(\alpha, c)x = \frac{x}{c} \varphi_{\alpha, c}(c).$$ By the concavity of $\varphi_{\alpha, c}$, we have $$\frac{\varphi_{\alpha, c}(x)}{x} \geq \frac{\varphi_{\alpha, c}(2c)}{2c}.$$ Therefore $$\begin{aligned}
\varphi_{\alpha, c}(t) + \varphi_{\alpha, c}(x-t) - \varphi_{\alpha, c}(x)
&\leq \frac{x}{c} \varphi_{\alpha, c}(c)- \frac{x \varphi_{\alpha, c}(2c)}{2c}
= \frac{x}{2c} {\left( 2 \varphi_{\alpha, c}(c) -\varphi_{\alpha, c}(2c) \right)}.\end{aligned}$$ But $$2 \varphi_{\alpha, c}(c) -\varphi_{\alpha, c}(2c) \leq (2 - 2^{\alpha})\varphi_{\alpha, c}(c) = (2 - 2^{\alpha})cL(\alpha, c).$$ Therefore $$\begin{aligned}
\varphi_{\alpha, c}(t) + \varphi_{\alpha, c}(x-t) - \varphi_{\alpha, c}(x)
\leq \frac{x}{2c} (2 - 2^{\alpha})cL(\alpha, c)
= (2 - 2^{\alpha}) \varphi_{\alpha, c}(x/2).\end{aligned}$$ In any case, we have $$0 \leq \varphi_{\alpha, c}(t) + \varphi_{\alpha, c}(x-t) - \varphi_{\alpha, c}(x) \leq (2 - 2^{\alpha}) \varphi_{\alpha, c}(x/2) \leq (2 - 2^{\alpha}) \varphi_{\alpha, c}(x).$$
The One Dimensional Construction
================================
For any $\alpha \in (0,1)$, we have $$\label{equation_1}
\lim_{c \to 0^+} L(\alpha, c) = +\infty.$$ Let $\alpha_n$ be an increasing sequence in $(0,1)$ such that $\lim_{n \to \infty} \alpha_n = 1$. By (\[equation\_1\]), we can choose $c_n \in (0,1)$ such that $c_n \to 0$, and $$\lim_{n \to \infty} L(\alpha_n, c_n) = +\infty.$$ Let $\varphi_n = \varphi_{\alpha_n, c_n}$. Choose a sequence $s_n$ so that for all $n \in {\mathbb{N}}$, $s_n < 2{\left( \frac{1}{n} - \frac{1}{n+1} \right)}$, such that $s_n L(\alpha_n, c_n)$ is decreasing and $\sum_{n \in{\mathbb{N}}} s_n L(\alpha_n, c_n) < \infty$.
For all $n \in {\mathbb{N}}$, let $a_n = \frac{1}{n} -s_n, b_n = \frac{1}{n}$. Let $I_n = [a_n, b_n]$, and equip $I_n$ with the metric $\delta_n = s_n\varphi_n \circ (s_n^{-1} d)$, where $d$ is the usual Euclidean metric on $I_n$. Note that
1. The distance between the two endpoints of $I_n$ is $\delta_n(a_n, b_n) = s_n$.
2. $I_n$ is rectifiable and the length of $I_n$ is $\ell(I_n) = s_nL(\alpha_n, c_n)$.
We construct a metric $\delta$ on ${\mathbb{R}}$ so that if $x \leq y$,
1. $\delta = \delta_n$ when restricted to $I_n \times I_n$,
2. $\delta(x,y) = d(x,y)$ when $x,y \in {\mathbb{R}}{\ifthenelse{\equal{}{}}{\backslash}{\backslash\{\}}}\bigcup_{i \geq } I_i$,
3. $\delta(x,y) = d(x, a_n) + \delta_n(a_n, y)$ when $x \in {\mathbb{R}}{\ifthenelse{\equal{}{}}{\backslash}{\backslash\{\}}}\bigcup_{i \in {\mathbb{N}}} I_i$ and $y \in I_n$,
4. $\delta(x,y) = \delta_n(x, b_n) + d(b_n, y)$ when $x \in I_n$ and $y \in {\mathbb{R}}{\ifthenelse{\equal{}{}}{\backslash}{\backslash\{\}}}\bigcup_{i \in {\mathbb{N}}} I_i$, and
5. $\delta(x,y) = \delta_n(x, b_n) + d(b_n, a_m) + d_m(a_m, y)$ when $x \in I_n$, $y \in I_m$.
The Weak Tangents of $({\mathbb{R}}, \delta)$
=============================================
A *pointed metric space* is a a triplet $(X, x_0, d_X)$, where $(X, d_X)$ is a metric space and $x_0$ is a point in $X$. Let ${\varepsilon}> 0$. A map $\psi: (X, x_0, d_X) \to (Y, y_0, d_Y)$ between two pointed metric spaces $(X, x_0, d_X)$ and $(Y, y_0, d_Y)$ is a *${\varepsilon}$-rough isometry* if
1. $\psi(x_0) = y_0$,
2. $d_Y(\psi(X), Y) \leq {\varepsilon}$, and
3. for all $x_1, x_2 \in X$, $${\left| d_Y(\psi(x_1), \psi(x_2)) - d_X(x_1, x_2) \right|} \leq {\varepsilon}.$$
Note that a ${\varepsilon}$-rough isometry may not be continuous. The *pointed Gromov-Hausdorff distance* between 2 pointed metric spaces, denoted $d_{GH}{\left( (X, x_0, d_X), (Y, y_0, d_Y) \right)}$, is defined as the infimum of ${\varepsilon}$ for which we can find a ${\varepsilon}$-rough isometry $\psi_\delta: (X, x_0, d_X) \to (Y, y_0, d_Y)$.
A pointed metric space $(T, p, d_T)$ is called a *weak tangent* of another metric space $(X, d_X)$ if there exist points $x_n \in X$ and positive integers $\lambda_n \to +\infty$ such that $(X, x_n, \lambda_n d_X)$ converges to $(T, p, d_T)$ in pointed Gromov-Hausdorff sense, i.e. for all $R > 0$, and for all ${\varepsilon}> 0$ there exists $N > 0$ such that for all $n \geq N$, $$\begin{aligned}
d_{GH}{\left( {\overline{B}}_{\lambda_n d_X}(x_n, R + {\varepsilon}), {\overline{B}}_{d_T}(p, R) \right)} \leq {\varepsilon}. \end{aligned}$$ In particular, we get $(X, x_n, \lambda_n d_X) \to (T, p, d_T)$ if for all $R > 0$, and for all ${\varepsilon}> 0$ there exists $N > 0$ such that for all $n \geq N$, $$\begin{aligned}
d_{GH}{\left( {\overline{B}}_{\lambda_n d_X}(x_n, R), {\overline{B}}_{d_T}(p, R) \right)} \leq {\varepsilon}. \end{aligned}$$ Our notion of pointed Gromov-Hausdorff convergence is adopted from [@The_Purple_Book Definition 11.3.1]. Also see [@Burago] for detailed discussion on Gromov-Hausdorff distance and Gromov-Hausdorff convergence.
Let M be a set of separable, uniformly doubling, and uniformly linearly locally contractible pointed metric spaces. The pointed Gromov Hausdorff convergence on the set $M$ induces a topology on $M$ that is metrizable[@LeDonne]. In particular, if a sequence of pointed metric space converges to a Gromov-Hausdorff limit, then the limit is unique.
The goal of this section is to prove the following proposition that describe all the weak tangents of $({\mathbb{R}}, \delta)$.
\[Prop\_1\] For all $a_n \in {\mathbb{R}}$, and for all positive integers $\lambda_n \to +\infty$, $({\mathbb{R}}, a_n, \lambda_n \delta)$ converges in pointed Gromov-Hausdorff sense to $({\mathbb{R}}, 0, d)$.
Note that Proposition \[Prop\_1\] guarentees the existence of weak tangents of $({\mathbb{R}}, \delta)$. To prove the above proposition we will use the following three lemmas:
\[lemma\_rough\_isometries\_2\] Suppose $x, y, z \in {\mathbb{R}}$ are three points so that $x \leq y \leq z$. Suppose $N = \inf\{n \in {\mathbb{N}}: \{x,y,z\} \cap I_N \neq \emptyset\} < \infty$. Then $$0 \leq \delta(x,y) + \delta(y,z) - \delta(z,x) \leq {\left( 2 - 2^{\alpha_N} \right)} \min\{ s_N, \delta(x,y) \}$$
If $N = +\infty$, then $\delta(x,y) + \delta(y,z) - \delta(z,x) = 0$. Otherwise, let $$\begin{aligned}
a &= \sup\{ a_n \leq y: n \in {\mathbb{N}}\} \vee \sup\{b_n \leq y: n \in {\mathbb{N}}\} \vee x \\
b &= \inf\{ a_n \geq y: n \in {\mathbb{N}}\} \wedge \inf\{b_n \geq y: n \in {\mathbb{N}}\} \wedge z\end{aligned}$$ We have $x \leq a \leq y \leq b \leq z$. By definition of $\delta$, we have $$\begin{aligned}
\delta(x,y) + \delta(y,z) - \delta(x,z)
&= {\left( \delta(x,a) + \delta(a,y) \right)} + {\left( \delta(y,b) + \delta(b,z) \right)} \\
&\hspace{30pt} - {\left( \delta(x,a) + \delta(a,b) + \delta(b,z) \right)} \\
&= \delta(a,y) + \delta(y,b) -\delta(a,b).\end{aligned}$$ Bu our choice of $a$ and $b$, either $a,y,b \in I_n$ for some $n \geq N$, or $(a, b) \cap \bigcup_{n \in {\mathbb{N}}} I_n = \emptyset$. In the latter case, we have $$\delta(x,y) + \delta(y,z) - \delta(z,x) = \delta(a,y) + \delta(y,b) -\delta(a,b) = 0.$$ In the former case, Lemma \[lemma\_rough\_isometries\_1\] gives $$\begin{aligned}
0 &\leq \delta(x,y) + \delta(y,z) - \delta(z,x)
= \delta(a,y) + \delta(y,b) -\delta(a,b) \\
&\leq {\left( 2 - 2^{\alpha_n} \right)} \delta(a,b)
= {\left( 2 - 2^{\alpha_N} \right)} \min\{ s_N, \delta(x,y) \}. \end{aligned}$$ Since $\alpha_N \leq \alpha_n \leq 1$, we have our desired conclusion.
\[lemma\_rough\_isometries\_3’\] Let $p \in {\mathbb{R}}$ and $r > 0$ be arbitrary. Suppose $N = \inf\{n \in {\mathbb{N}}: {\overline{B}}_{\delta}(p,r)\cap I_n \neq \emptyset\} < \infty$. Let $a = \inf\{x \in {\mathbb{R}}: \delta(x,p) \leq r\}$ and $b = \sup\{x \in {\mathbb{R}}: \delta(x,p) \leq r\}$. The map $$\begin{aligned}
\psi: ([a,b], p, \delta) \to ([-r, r],0,d)\\
\psi(x) = \begin{cases}
-\delta(p,x) ,& x \leq p \\
\delta(p,x),& x > p.
\end{cases}\end{aligned}$$ is a ${\varepsilon}$-rough isometry, where ${\varepsilon}= {\left( 2 - 2^{\alpha_N} \right)} \min\{ s_N, 2r\}$.
Since $\psi$ is surjective and fixes $p$, it remains to check that for all $x,y \in [a,b]$, we have $${\left( d(\psi(x), \psi(y)) - \delta(x,y) \right)} \leq {\varepsilon}.$$ Suppose $x \leq y$. If $x \leq y \leq p$, then $$\begin{aligned}
{\left| d(\psi(x), \psi(y)) - \delta(x,y) \right|}
&= {\left| {\left| \delta(p,x) - \delta(p,y) \right|} - \delta(x,y) \right|} \\
&= {\left| \delta(p,y) - \delta(p, x) - \delta(x,y) \right|} \\
&\leq {\left( 2 - 2^{\alpha_N} \right)} \min\{ s_N, \delta(p,x)\} \\
&= {\left( 2 - 2^{\alpha_N} \right)} \min\{ s_N, 2r\} .\end{aligned}$$ The second to last inequality is a consequence of Lemma \[lemma\_rough\_isometries\_2\]. If $x \leq p \leq y$, then $$\begin{aligned}
{\left| d(\psi(x), \psi(y)) - \delta(x,y) \right|}
&= {\left| {\left| \delta(p,x) - \delta(p,y) \right|} - \delta(x,y) \right|} \\
&= {\left| \delta(p,y) + \delta(p, x) - \delta(x,y) \right|} \\
&\leq {\left( 2 - 2^{\alpha_N} \right)} \min\{ s_N, \delta(x,y)\} \\
&= {\left( 2 - 2^{\alpha_N} \right)} \min\{ s_N, 2r\}.\end{aligned}$$ If $p \leq x \leq y$, then following a similar arguement as when $x \leq y \leq p$, we get $$\begin{aligned}
{\left| d(\psi(x), \psi(y)) - \delta(x,y) \right|}
\leq {\left( 2 - 2^{\alpha_N} \right)} \min\{ s_N, 2r\}\end{aligned}$$ This verifies that $\psi$ is a ${\varepsilon}$-rough isometry.
\[lemma\_rough\_isometries\_3\] For $r \in (0,1)$, we have $$\sup_{p \in {\mathbb{R}}} d_{GH}(({\overline{B}}_{\delta}(p,r), p, \delta),({\overline{B}}_{d}(0,r), 0, d)) = o(r)$$ as $r \to 0$.
Let $p \in {\mathbb{R}}$ and $r > 0$ be arbitrary. Let $N = \inf\{n \in {\mathbb{N}}: {\overline{B}}_{\delta}(p,r)\cap I_n \neq \emptyset\}$. If $N = +\infty$, then $\delta = d$ on ${\overline{B}}_{\delta}(p,r)$. We have $$d_{GH}(({\overline{B}}_{\delta}(p,r), p, \delta),({\overline{B}}_{d}(0,r), 0, d)) = 0.$$ If $N < \infty$, we consider 2 cases:
Case 1: $r \geq \frac{s_N}{2}\varphi_N(c_N)$. As $r \to 0$, $N \to +\infty$, therefore ${\left( 2 - 2^{\alpha_N} \right)} \to 0$. In this case Lemma \[lemma\_rough\_isometries\_3’\] implies $$\begin{aligned}
d_{GH}(({\overline{B}}_{\delta}(p,r), p, \delta),({\overline{B}}_{d}(0,r), 0, d)) \leq 2{\left( 2 - 2^{\alpha_N} \right)} r = o(r).\end{aligned}$$
Case 2: $r < \frac{s_N}{2}\varphi_N(c_N)$. In this case, $\delta$ is a length metric on $[a,b]$, and $\psi$ is an isometry between two length spaces. We have $$d_{GH}(({\overline{B}}_{\delta}(p,r), p, \delta),({\overline{B}}_{d}(0,r), 0, d)) = 0.$$
Let $\{a_n\}_{n \in {\mathbb{N}}}$ be a sequence in ${\mathbb{R}}$ and $\{\lambda_n\}_{n \in {\mathbb{N}}}$ be a sequence of positive numbers that diverges to $+\infty$. By Lemma \[lemma\_rough\_isometries\_3\], $$\begin{aligned}
&d_{GH}{\left( ({\overline{B}}_{\lambda_n \delta}(a_n, R ), a_n, \lambda_n \delta),({\overline{B}}_{\lambda_n d}(0, R), 0, \lambda_n d) \right)} \\
&\hspace{30pt}= \lambda_n d_{GH}{\left( ({\overline{B}}_{\delta}(a_n, \lambda_n^{-1}R ), a_n, \delta),( {\overline{B}}_{d}(0, \lambda_n^{-1}R), 0, d) \right)} = \lambda_n o(\lambda_n^{-1}R).\end{aligned}$$ But $({\overline{B}}_{\lambda_n d}(0, R), 0, \lambda_n d) = {\overline{B}}_{d}(0, R), 0, d))$ by symmetry of ${\mathbb{R}}$. As $n \to \infty$, $\lambda_n^{-1}R \to 0$. We have $$({\overline{B}}_{\lambda_n \delta}(a_n, R), a_n, \delta ) \to {\overline{B}}_{d}(0, R), 0, d).$$ This is true for all $R > 0$. We conclude that $({\mathbb{R}}, a_n, \lambda_n \delta) \to ({\mathbb{R}}, 0, d)$.
Linear Local Contractibility and Assouad Dimension of $({\mathbb{R}}, \delta)$
==============================================================================
In this section we establish two properties of the space $({\mathbb{R}}, \delta)$. These properties often appear in the study of quasisymmetry classes of metric spheres. Both properties are discussed in detail in [@Heinonen].
Let $C > 1$ be a constant. A metric space is *$C$-linearly locally contractible* if every small ball is contractible inside a ball whose radius is $C$ times larger. A metric space is *linearly locally contractible* if it is $C$-linearly locally contractible for some $C > 0$.
Let $N > 0$. A metric space is *$N$-doubling* if for all $R > 0$, every open ball of $2R$ can be covered by $N$ balls of radius $R$. A metric space is *doubling* if it is $N$-doubling for some $N > 0$.
Both doubling and linear local contractibility are preserved under quasisymmetry. The Euclidean spaces like ${\mathbb{R}}^n$ or $\mathbb{S}^n$ are doubling and linearly locally contractible. The doubling property also ensures the existence of weak tangents.
\[Prop\_3\] The space $({\mathbb{R}}, \delta)$ is $1$-linearly locally contractible.
Any open ball $B$ in $({\mathbb{R}}, \delta)$ is an open interval $(a,b)$. Denote $p$ the center of $B$ (in $({\mathbb{R}}, \delta)$.) Note that the map $x \mapsto \delta(p, x)$ is increasing on $\{x \in {\mathbb{R}}: x \geq p$, and decreasing on $\{x \in {\mathbb{R}}: x \leq p\}$. Therefore the map $H(x,t) = tx + (1-t)p$ is a homotopy of $(a,b)$ to $\{p\}$ in $B$. This proves that $({\mathbb{R}}, \delta)$ is $1$-linearly locally contractible.
\[Prop\_2\] The space $({\mathbb{R}}, \delta)$ is doubling.
If a metric space $X$ is doubling, then there exists $\beta > 0$ and $C > 0$ such that for all ${\varepsilon}\in (0, 1/2)$ and $r > 0$, any set of diameter $r$ in $X$ can be covered by at most $C {\varepsilon}^{-\beta}$ subsets of diameter at most ${\varepsilon}r$. The function ${\varepsilon}\mapsto C {\varepsilon}^{-\beta}$ is called the *covering function* of $X$. The Assouad dimension of $X$ is defined to be the infimum of all $\beta$ so that a covering function of the form ${\varepsilon}\mapsto C {\varepsilon}^{-\beta}$ of $X$ exists. Conversely, any metric space of finite Assouad dimension is doubling. Proposition \[Prop\_2\] will follow from the stronger proposition below.
For each $n \in {\mathbb{N}}$, the function $f_n({\varepsilon}) = 2{\varepsilon}^{-\alpha_n^{-1}}$ is a covering function of $(I_n, \delta)$.
Every subinterval of $I_n$ of $\delta$-diameter $r \in [0, s_n]$ has $d$-diameter $s_n\varphi_n^{-1}(s_n^{-1}r)$. Thus our goal is to show that for every $r \in [0, s_n]$, and every ${\varepsilon}\in (0, 1/2)$, every subset of $I_n$ of $d$-diameter $s_n\varphi_n^{-1}(s_n^{-1}r)$ can be covered by no more than $({\varepsilon}^{-\alpha_n}+1)$-many subintervals of $I_n$ of $d$-diameter at most $s_n\varphi_n^{-1}(s_n^{-1}{\varepsilon}r)$. The number of subintervals we need can be bounded from above by $$\begin{aligned}
\frac{s_n\varphi_n^{-1}(s_n^{-1}r)}{s_n\varphi_n^{-1}(s_n^{-1}{\varepsilon}r)} + 1
\leq \sup_{y \in (0,1]} \frac{\varphi_n^{-1}(y)}{\varphi_n^{-1}({\varepsilon}y)} + 1
= \sup_{y \in [\varphi_n(c_n),1]} \frac{\varphi_n^{-1}(y)}{\varphi_n^{-1}({\varepsilon}y)} + 1.\end{aligned}$$ We claim that the last supremum is attained when $y = 1$. This is equivalent to $$\label{equation_2}
\frac{\varphi_n(\varphi_n^{-1}({\varepsilon}y))}{\varphi_n(\varphi_n^{-1}(y))} \leq \varphi_n{\left( \frac{\varphi_n^{-1}({\varepsilon}y)}{\varphi_n^{-1}(y)} \right)}$$ Inequality (\[equation\_2\]) is true by Lemma \[lemma\_sup\_multiplicative\]. Suppose $\varphi_n(x_0) = {\varepsilon}$. When ${\varepsilon}> \varphi_n(c_n)$, we have $x_0 > c_n$, and $$\begin{aligned}
{\varepsilon}^{\alpha_n^{-1}} = \frac{x_0 - c(1-\alpha_n)}{1-c(1-\alpha_n)} \leq x_0.\end{aligned}$$ When $0 < {\varepsilon}< \varphi_n(c_n)$, we have $x_0 < c_n$ and ${\varepsilon}= \varphi_n(x_0) = \frac{x_0}{c} \varphi_n(c_n)$. Since $ \varphi_n(c_n)^{\alpha_n^{-1}} \leq c$, we have $$\begin{aligned}
\frac{1}{x_0}
= \frac{\varphi(c)}{c{\varepsilon}}
= {\varepsilon}^{-\alpha_n^{-1}} {\left( \frac{{\varepsilon}}{\varphi(c)} \right)}^{\alpha_n^{-1}-1} \frac{\varphi(c)^{\alpha_n^{-1}}}{c} \leq {\varepsilon}^{-\alpha_n^{-1}}. \end{aligned}$$ In any case, we can take the covering function of $I_n$ to be $$\begin{aligned}
{\varepsilon}^{-\alpha_n^{-1}}+1 \leq 2{\varepsilon}^{-\alpha_n^{-1}}. \end{aligned}$$
\[Prop\_4\] The Assouad dimension of $({\mathbb{R}}, \delta)$ is $1$.
Let $\beta > 1$ be arbitrary. There exists $N \in {\mathbb{N}}$ such that when $n \geq N$, $\alpha^{-1} <\beta$. Let $C = \max_{n < N}\{2^{\alpha_n^{-1} - \beta}\} \geq 1$. Then the function ${\varepsilon}\mapsto 2C {\varepsilon}^{-\beta}$, where ${\varepsilon}\in (0, 1/2]$, is a covering function of $(I_n, \delta)$ for all $n \in {\mathbb{N}}$. Thus ${\varepsilon}\mapsto 4C {\varepsilon}^{-\beta}$ is a covering function of $({\mathbb{R}}, \delta)$.
Since $(-\infty, 0)$ has Assouad dimension $1$, the proposition follows.
Higher Dimension Construction
=============================
Let $d \geq 2$. We will denote by $d_{Euclid}$ the Euclidean metric on ${\mathbb{R}}^d$. Let $$X_d = {\left( {\mathbb{R}}\times {\mathbb{R}}^{d-1}, \sqrt{\delta^2 + d_{Euclid}^2} \right)}$$ be the product of $({\mathbb{R}}, \delta)$ and $({\mathbb{R}}^{d-1}, d_{Euclid})$. Write $\rho_n = \sqrt{\delta^2 + d_{Euclid}^2}$. Here are some facts about $X_d$.
\[HD\_Prop\_1\]
(a) Every weak tangent of $X_d$ is isometric to $({\mathbb{R}}^d, 0, d_{Euclid})$.
(b) $X_d$ is doubling and linearly locally contractible.
<!-- -->
(a) Every weak tangent of $X_d$ is of the form $(T \times {\mathbb{R}}^{n-1}, (x, 0), d_T \times d_{Euclid})$, where $(T, x, d_T)$ is a weak tangent of $({\mathbb{R}}, \delta)$.
(b) Recall that $({\mathbb{R}}, \delta)$ is doubling (Proposition \[Prop\_2\]). $X_d$ is product of doubling metric space, hence doubling. By Proposition \[Prop\_3\], $({\mathbb{R}}, \delta)$ is $C$-linearly localy contractible for some $C > 1$. Let $x = (x_1, x_2)$ be any point in ${\mathbb{R}}\times {\mathbb{R}}^{d-1}$, and $r >0$ be arbitrary. The ball $B(x,r)$ in $X_d$ can first be contracted to $\{x_1\} \times B(x_2, r)$ within a $B(x, Cr)$, which can then be contracted to the point $\{x_1, x_2\}$.
Every finite segment in $({\mathbb{R}}, \delta)$ is rectifiable. Let $\mu_1$ be a the measure on $({\mathbb{R}}, \delta)$ given by length. For $d \geq 2$, let $\mu_d$ be the product measure $\mu_1 \times \lambda_{d-1}$ on $X_d$, where $\lambda_{d-1}$ is the $(d-1)$-dimensional Lebesgue measure on ${\mathbb{R}}^{d-1}$.
In the remaining of this section we show that $X_d$ is not quasisymmetrically equivalent to ${\mathbb{R}}^d$. To do that we consider a geometric quantity that is roughly preserved under quasisymmetry called modulus. Given a family $\Gamma$ of curves in a measured metric space $(X, d_X, \mu)$, we say that a function $\rho:X \to [0,\infty)$ is admissible if for all $\gamma \in \Gamma$, $$\begin{aligned}
\int_\gamma \rho(x) \,ds \geq 1.\end{aligned}$$ Let $Q > 0$. We define the *$Q$-modulus* of $\Gamma$ as $$\begin{aligned}
{\operatorname{mod}}_Q(\Gamma) = \inf\{ \int_X \rho^Q \,d\mu : \rho \text{ admissible }\}.\end{aligned}$$ Let $E, F \subset X$ be two disjoint nondegerate continua in $X$. Let $\Gamma_{E,F}$ to be the collection of all rectifiable curves joining $E$ and $F$. We write ${\operatorname{mod}}_Q(E,F) = {\operatorname{mod}}_Q(\Gamma_{E,F})$.
Moduli behave nicely under quasisymmetry, as illustrated by the following theorem.
\[Tyson\_theorem\] Let $X, Y$ be locally compact, connected, $Q$-Ahlfors regular metric spaces, where $Q > 1$, and $f:X \to Y$ be a quasisymmetric homeomorphism. Then there exists $C > 1$ such that for all curve family $\Gamma \subset X$, we have $$\frac{1}{C} {\operatorname{mod}}_Q(\Gamma) \leq {\operatorname{mod}}_Q(f(\Gamma)) \leq C{\operatorname{mod}}_Q(\Gamma).$$
See the next section for the definition of $Q$-Ahlfors regularity.
We can now prove that $X_d$ is not quasisymmetrically equivalent to ${\mathbb{R}}^d$. The idea is that if the two spaces are quasisymmetrically equivalent, then for any curve family $\Gamma$, $\Gamma$ and $f(\Gamma)$ should have comparable moduli. We know that ${\mathbb{R}}^d$ has the property that any disjoint nondegerate continua $E, F$ in ${\mathbb{R}}^d$ satisfy $$\label{modulus}
{\operatorname{mod}}_d(E,F) \leq \phi{\left( \frac{d(E,F)}{{\operatorname{diam}}(E) \wedge {\operatorname{diam}}(F)} \right)}.$$ for some non-increasing function $\phi:[0,\infty) \to (0,\infty)$. However, Proposition \[HD\_Prop\_3\] shows that some sequence disjoint nondegerate continua $E_n, F_n$ in $X_d$ do not behaviour as in (\[modulus\]). The only problem is that $X_d$ is not $d$-Ahlfors regular, so we cannot apply Theorem \[Tyson\_theorem\] directly. In Proposition \[HD\_Prop\_2\], however, we will show that the inequality $${\operatorname{mod}}_d(E_n,F_n) \leq C {\operatorname{mod}}_d(f(E_n), f(F_n))$$ holds for some $C$ independent of $C$.
From now on, we will denote $\Delta(E,F) = \frac{d(E,F)}{{\operatorname{diam}}(E) \wedge {\operatorname{diam}}(F)}$.
\[HD\_Prop\_3\] There exists $E_n, F_n \subset X_d$ such that $\Delta(E_n, F_n) = \frac{d(E_n, F_n)}{{\operatorname{diam}}E_n \wedge {\operatorname{diam}}F_n} = 1$ and ${\operatorname{mod}}_d(E_n, F_n) \to +\infty$.
Take $E_n = I_n \times \{0\} \times [0,s_n]^{d-2},
F_n = I_n \times \{s_n\} \times [0,s_n]^{d-2}$. Then $d(E_n, F_n) = s_n$, and ${\operatorname{diam}}E_n = {\operatorname{diam}}F_n = s_n$, therefore $$\Delta(E_n, F_n) = \frac{d(E_n, F_n)}{{\operatorname{diam}}E_n \wedge {\operatorname{diam}}F_n} = 1.$$ For each $n \in {\mathbb{N}}$, and for $x \in I_n$ and $(v_2, \ldots, v_{d-1}) \in [0,s_n]^{d-2}$, let $\gamma_{x, v_2, \ldots, v_{d-1}}$ be the path $$t \mapsto {\left( x, t, v_2, \ldots, v_{d-1} \right)}, t \in [0,s_n].$$ Let $$\Gamma_n = \{ \gamma_{x, v_2, \ldots, v_{d-1}} : x \in I_n, (v_2, \ldots, v_{d-1}) \in [0,s_n]^{d-2}\}$$ be the family of straight lines joining $E_n$ and $F_n$ that meet $E_n$ orthogonally. Then $${\operatorname{mod}}_d(E_n, F_n) \geq {\operatorname{mod}}_d(\Gamma_n).$$
Let $\rho: X_d \to {\mathbb{R}}_{\geq 0}$ be an admissible function for $\Gamma_n$. This means for all $\gamma_{x, v_2, v_3, \ldots, v_{d-1}} \in \Gamma_n$, we have $$\int_{\gamma_{x, v_2, v_3, \ldots, v_{d-1}}} \rho(t) \,dt
= \int_0^{s_n} \rho(x,t,v_2, \ldots, v_{d-1}) \,dt \geq 1.$$ Integrating over $I_n \times [0,s_n]^{d-2}$ and applying Fubini’s theorem, we get $$\begin{aligned}
\int_{I_n \times [0,s_n]^{d-2}} \,d\mu_1 \times \lambda_{d-2}
&\leq \int_{I_n \times [0,s_n]^{d-2}} \int_{\gamma_{x, v_2, v_3, \ldots, v_{d-1}}} \rho(t) \,dt \,d(\mu_1 \times \lambda_{d-2}) \\
&= \int_{I_n \times [0,s_n]^{d-1}} \rho \,d\mu_d.\end{aligned}$$ Applying Hölder’s inequality, we get $$\begin{aligned}
\int_{I_n \times [0,s_n]^{d-1}} \rho \,d\mu_d
&\leq {\left( \int_{I_n \times [0,s_n]^{d-1}} \,d\mu_d \right)}^{\frac{1}{\delta}} {\left( \int_{I_n \times [0,s_n]^{d-1}} \rho^d \,d\mu_d \right)}^{\frac{1}{d}}\end{aligned}$$ where $\delta$ is the conjugate exponent of $d$ (so $\frac{1}{\delta} + \frac{1}{d} = 1$). We have $$\begin{aligned}
\int_{I_n \times [0,s_n]^{d-1}} \rho^d \,d\mu_d
&\geq {\left( \int_{I_n \times [0,s_n]^{d-1}} \,d\mu_d \right)}^{-\frac{d}{\delta}}{\left( \int_{I_n \times [0,s_n]^{d-2}} \,d\mu_1 \times \lambda_{d-2} \right)}^d \\
&={\left( (s_n)^{(d-1)} \ell(I_n) \right)}^{-\frac{d}{\delta}} {\left( (s_n)^{d-2} \ell(I_n) \right)}^d \\
&= 2^{n} \ell(I_n).\end{aligned}$$ This is true for all admissible function $\rho$, therefore $${\operatorname{mod}}_d(\Gamma_n) \geq s_n^{-1} \ell(I_n).$$ As $n \to \infty$, $s_n^{-1} \ell(I_n) = L(\alpha_n, c_n) \to +\infty$. We have $${\operatorname{mod}}_d(E_n, F_n) \geq {\operatorname{mod}}_d(\Gamma_n) \to +\infty.$$
\[HD\_Prop\_2\] $X_d$ is not quasisymmetrically equivalent to any subset of ${\mathbb{R}}^d$.
For each $n \in {\mathbb{N}}$, and for $x \in I_n, (v_2, \ldots, v_{d-1}) \in [0,s_n]^{d-2}\}$, let $\gamma_{x, v_2, \ldots, v_{d-1}}$ be the path $$t \mapsto {\left( x, t, v_2, \ldots, v_{d-1} \right)}, t \in [0,s_n].$$ Let $$\Gamma_n = \{ \gamma_{x, v_2, \ldots, v_{d-1}} : x \in I_n, (v_2, \ldots, v_{d-1}) \in [0,s_n]^{d-2}\}$$ be the family of straight lines joining $E_n$ and $F_n$ that meet $E_n$ orthogonally. Suppose $f: X_d \to {\mathbb{R}}^d$ is a $\eta$-quasisymmetric embedding. We denote by $\gamma^*$ the image of the curve $\gamma$. For a path family $\Gamma$, define $\Gamma^* = \bigcup_{\gamma \in \Gamma} \gamma*$, and define ${\operatorname{diam}}(\Gamma) = \inf\{{\operatorname{diam}}\gamma: \gamma \in \Gamma\}$.
Choose $E_n$ and $F_n$ as in Proposition \[HD\_Prop\_3\]. We know from ${\operatorname{diam}}(E_n) \wedge {\operatorname{diam}}(F_n) = d(E_n, F_n)$ that $$\frac{d(f(E_n), f(F_n))}{{\operatorname{diam}}(f(E_n)) \wedge {\operatorname{diam}}(f(F_n))} \sim 1,$$ where the implicit constant for $\sim$ depends only on $\eta$. For the same reason, there exists $\alpha > 1$, depending only on $\alpha,\beta > 1$, depending only on $\eta$, so that $f(\Gamma_n)^* \subset B_n$ for a ball $B_n$ with diameter $r_n \leq \alpha \,d(E_n, F_n) \leq \beta {\operatorname{diam}}(f(\Gamma))$.
For each $n \in {\mathbb{N}}$, we can cover $\Gamma^*$ by squares $\{R_i\}_{i \in I_n}$ of diameter $s_n\varphi_n(c_n)$ so that their sides are either parallel to the paths in $\Gamma$ or orthogonal to the paths in $\Gamma$. We can choose $\{R_i\}$ so that these $R_i$’s don’t overlap and their union is precisely $\Gamma^*$. For each rectifiable path $\gamma$, denote by $\ell(\gamma)$ its length. Let $$\rho_n = ({\operatorname{diam}}f(\Gamma_n))^{-1} \sum_{i \in I_n} \frac{{\operatorname{diam}}(f(R_i))}{{\operatorname{diam}}(R_i)} \mathbbm{1}_{f^{-1}(B_n) \cap R_i}.$$ be a function on $X_d$. For all $\gamma \in \Gamma_n$,
We have $$\int_{\gamma} \rho_n(s) \,ds
= ({\operatorname{diam}}f(\Gamma_n))^{-1} \sum_{i \in I_n} \frac{{\operatorname{diam}}(f(R_i))}{{\operatorname{diam}}(R_i)} \ell( f^{-1}(B_n) \cap R_i \cap \gamma).$$ For each $i$, $f^{-1}(B_n) \cap R_i \cap \gamma = R_i \cap \gamma$. When $R_i \cap \gamma \neq \emptyset$, $\ell(R_i \cap \gamma) = {\operatorname{diam}}(R_i)$. As $\{R_i\}_{i \in I_n}$ covers $\Gamma^*$, we have $$\begin{aligned}
\int_{\gamma} \rho_n(s) \,ds
\geq ({\operatorname{diam}}f(\Gamma_n))^{-1} \sum_{i \in I_n, R_i \cap \gamma \neq \emptyset } {\operatorname{diam}}(f(R_i))
\geq ({\operatorname{diam}}f(\Gamma_n))^{-1} {\operatorname{diam}}(f(\gamma))
\geq 1.\end{aligned}$$ This means $\rho_n$ is admissible for $\Gamma_n$. We have $$\begin{aligned}
{\operatorname{mod}}_d \Gamma_n
&\leq \int \rho_n^d \,d\mu_d
= ({\operatorname{diam}}f(\Gamma_n))^{-d} \sum_{i \in I_n} {\left( \frac{{\operatorname{diam}}(f(R_i))}{{\operatorname{diam}}(R_i)} \right)}^d \mu_d(f^{-1}(B_n) \cap R_i) \\
&= ({\operatorname{diam}}f(\Gamma_n))^{-d} \sum_{i \in I_n} {\left( \frac{{\operatorname{diam}}(f(R_i))}{{\operatorname{diam}}(R_i)} \right)}^d \mu_d(R_i) \\
&= ({\operatorname{diam}}f(\Gamma_n))^{-d} \sum_{i \in I_n} {\left( \frac{{\operatorname{diam}}(f(R_i))}{{\operatorname{diam}}(R_i)} \right)}^d {\operatorname{diam}}(R_i)^d.\end{aligned}$$ For the last equality, we make use of the fact that $R_i$ are chosen so small that $\mu_d(R_i) = {\operatorname{diam}}(R_i)^d$. We get $$\begin{aligned}
{\operatorname{mod}}_d \Gamma_n
\lesssim ({\operatorname{diam}}f(\Gamma_n))^{-d} \sum_{i \in I_n} {\left( {\operatorname{diam}}(f(R_i)) \right)}^d
\lesssim ({\operatorname{diam}}f(\Gamma_n))^{-d} \sum_{i \in I_n} \lambda_d(f(R_i)).\end{aligned}$$ Here we use the fact that the Lebesgue $\lambda_d$ on ${\mathbb{R}}^d$ is $d$-Ahlfors regular and that $f(R_i)$ are uniform quaisdisks. Since $\{f(R_i)\}_{n\in{\mathbb{N}}}$ are disjoint subsets of $B_n$, we have $$\begin{aligned}
{\operatorname{mod}}_d \Gamma_n \lesssim ({\operatorname{diam}}f(\Gamma_n))^{-d} \lambda_d(B_n) \lesssim 1.\end{aligned}$$ where all the implicit constants for $\lesssim$ depends only on $\eta$. But this is a contradiction to Proposition \[HD\_Prop\_3\].
From Prop \[HD\_Prop\_1\] and Prop \[HD\_Prop\_2\], $X_d$ is homeomorphic to ${\mathbb{R}}^d$, doubling and linearly localy connected, it is not quasisymmetric to ${\mathbb{R}}^d$. Our proof shows that the any ball in $X_d$ centered at $0$ cannot be quasisymmetrically embedded into ${\mathbb{R}}^d$.
We conclud this section by a proof of Theorem \[Theorem\].
The $d$-dimensional unit ball $B(0,1)$ in ${\mathbb{R}}^d$, equipped with the metric $${\widetilde{\rho}}(x, y) = \frac{\rho{\left( \frac{x}{1-{\left| x \right|}}, \frac{y}{1-{\left| y \right|}} \right)}}{1+ \rho{\left( \frac{x}{1-{\left| x \right|}}, \frac{y}{1-{\left| y \right|}} \right)}}.$$ The completion of the space $(B(0,1), {\widetilde{\rho}})$ is ${\overline{B}}(0,1)$, and the metric on the boundary is same as the Euclidean metric. Glue the space ${\overline{B}}(0,1)$ with another hemisphere to form a topological $d$-sphere. This $d$-sphere is doubling, locally linearly contractible, and every weak tangent is isometric to ${\mathbb{R}}^d$, but it cannot be a quasisphere.
Ahlfors Regularity
==================
Let $Q > 0$. A measured metric space $(X, d, \mu)$ is said to be $Q$-Ahlfors regular if for all $x \in X$ and $r \leq {\operatorname{diam}}X$, we have $$\mu(B(x,r)) \sim r^Q.$$ Ahlfors regularity and doubling property of metric spaces are related notions. We record a result from [@Heinonen]:
\[Heinonen\_AR\_2\][@Heinonen Theorem 14.6] Let $X$ be a complete, connected metric space of finite Assouad dimension $\beta$. Then for each $Q > \beta$, there exists a quasisymmetric homeomorphism of $X$ onto a closed $Q$-Ahlfors regular subset of some ${\mathbb{R}}^N$.
With these facts our example gives:
\[Q-AR-counter-example\] For every $Q > 2$, there exists a $Q$-Ahlfors regular and linearly locally contractible metric space $X$ that is topologically an $2$-sphere such that every weak tangent is uniformly quasisymmetric to ${\mathbb{R}}^2$ but $X$ is not quasisymmetric to the standard $2$-sphere.
Let $Q > 2$. By Proposition \[Prop\_4\], the Assouad dimension of $X_2$ is $2$. Proposition \[Heinonen\_AR\_2\] says that there exist a distortion function $\eta:[0,\infty) \to [0,\infty)$ and a $\eta$-quasisymmetry $\varphi:X_2 \to X'$, where $X'$ is a closed $Q$-Ahlfors regular subset of ${\mathbb{R}}^N$. By Proposition \[HD\_Prop\_2\] and Proposition \[HD\_Prop\_1\](b), $X'$ is not quasisymmetric to the standard $2$-sphere, but every weak tangent of $X'$ is $\eta$-quasisymmetric to $({\mathbb{R}}^2, 0)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Boolean functions with high algebraic immunity are important cryptographic primitives in some stream ciphers. In this paper, two methodologies for constructing binary minimal codes from sets, Boolean functions and vectorial Boolean functions with high algebraic immunity are proposed. More precisely, a general construction of new minimal codes using minimal codes contained in Reed-Muller codes and sets without nonzero low degree annihilators is presented. The other construction allows us to yield minimal codes from certain subcodes of Reed-Muller codes and vectorial Boolean functions with high algebraic immunity. Via these general constructions, infinite families of minimal binary linear codes of dimension $m$ and length less than or equal to $m(m+1)/2$ are obtained. In addition, a lower bound on the minimum distance of the proposed minimal linear codes is established. Conjectures and open problems are also presented. The results of this paper show that Boolean functions with high algebraic immunity have nice applications in several fields such as symmetric cryptography, coding theory and secret sharing schemes.'
address:
- 'School of Mathematics and Information, China West Normal University, Nanchong, Sichuan, 637002, China'
- 'Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong, China'
- 'LAGA, Department of Mathematics, Universities of Paris VIII and Paris XIII, CNRS, UMR 7539 and Telecom ParisTech, France'
author:
- Hang Chen
- Cunsheng Ding
- Sihem Mesnager
- Chunming Tang
title: A Novel Application of Boolean Functions with High Algebraic Immunity in Minimal Codes
---
Boolean function ,vectorial Boolean function ,Reed-Muller code ,secret sharing ,minimal code. 05B05 ,51E10 ,94B15
Introduction
============
Secret sharing, independently introduced in 1979 by Shamir [@Shamir79] and Blakley [@Blakley79], is one of the most widely studied topics in cryptography. Relations between linear codes and secret sharing schemes were first investigated by McEliece and Sarwate in [@McESar81]. In theory every linear code can be employed to construct secret sharing schemes. Unfortunately, it is extremely hard to determine the access structures of secret sharing schemes based on general linear codes. However, the access structures of secret sharing schemes based on minimal linear codes are known and interesting [@DY03; @Mas93; @Mas95].
Minimal codes have already received a lot of attention. It was pointed out in [@ABN19] and [@TQLZ19] that minimal codes are close to blocking sets in finite geometry. Many minimal linear codes were obtained from codes with few weights [@DD15; @DY03; @M17; @MOS17; @MOS18; @TLQZT]. Recently, Ding, Heng and Zhou [@DHZI] constructed three infinite families of minimal binary linear codes using certain Boolean functions. They also constructed an infinite family of minimal ternary linear codes from ternary functions in [@DHZF]. Bartoli and Bonini [@BBI] generalized the construction of minimal linear codes in [@DHZF] from the ternary case to the odd $p$ characteristic case via $p$-ary functions. Li and Yue [@LY] obtained some minimal binary linear codes with nonlinear Boolean functions. Xu and Qu [@XQ19] constructed minimal $q$-ary linear codes from some special functions. In the recent paper [@MQRT], the authors considered minimal codes from the supports of $p$-ary functions. Lu, Wu and Cao [@LWC] obtained minimal codes with special subsets of vector spaces over finite fields. Bonini and Borello [@BB19] presented a family of minimal codes arising from some blocking sets.
The main objective of this paper is to find connections among special sets, Boolean functions with high algebraic immunity and binary minimal codes. Two general constructions of minimal binary codes with minimal codes contained in the Reed-Muller codes, subsets of finite fields and vectorial Boolean functions with high algebraic immunity are proposed. Two families of minimal codes contained in the second-order Reed-Muller code with large dimension are presented. Sets and vectorial Boolean functions with high algebraic immunity are also demonstrated. By plugging these subcodes, special sets and vectorial Boolean functions into our general construction, some infinite classes of minimal binary linear codes of dimension $m$ and length less than or equal to $m(m+1)/2$ are produced. Finally, a lower bound on the minimum distance of the proposed minimal codes is derived. Conjectures and open problems are also presented.
The rest of this paper is organized as follows. In Section \[sec:intr\], we recall some standard facts about cyclic codes, Reed-Muller codes and vectorial Boolean functions. In Section \[sec:codes-sets\], we establish some relations between binary minimal codes and subsets of finite fields without nonzero low degree annihilators. It enables us to yield minimal codes via certain subcodes of Reed-Muller codes and sets with high algebraic immunity. In Section \[sec:codes-vec-func\], we present a general construction of minimal codes from subcodes of Reed-Muller codes and vectorial Boolean functions having high algebraic immunity. In Section \[sec:conc\], we conclude this paper.
Background {#sec:intr}
==========
Boolean functions and vectorial Boolean functions
-------------------------------------------------
A *Boolean function* $f$ on ${{\mathrm{GF}}}(2^m)$ is a ${{\mathrm{GF}}}(2)$-valued function on the Galois field ${{\mathrm{GF}}}(2^m)$ of order $2^m$. The set of all Boolean functions over ${{\mathrm{GF}}}(2^m)$ forms a ring and is denoted by $\mathbb{B}_m$. The *support* of $f$, denoted by ${{\mathrm{Supp}}}(f)$, is the set of elements of ${{\mathrm{GF}}}(2^m)$ whose image under $f$ is $1$, that is, ${{\mathrm{Supp}}}(f)=\left \{ x\in {{\mathrm{GF}}}(2^m): f(x)=1 \right \}$. The *Hamming weight ${{\mathrm{wt}}}(f)$* of a Boolean function is the size of its support ${{\mathrm{Supp}}}(f)$. The *characteristic function* $f_D$ of a subset $D$ of ${{\mathrm{GF}}}(2^m)$ is the Boolean function such that $f(x)=1$ for all $x \in D$ and $f(x)=0$ for all $x \in {{\mathrm{GF}}}(2^m) \setminus D$. Thus ${{\mathrm{Supp}}}\left (f_D \right )=D$. Every nonzero Boolean function $f$ on ${{\mathrm{GF}}}(2^m)$ has a unique univariate polynomial expansion of the form $$\begin{aligned}
f(x)=\sum_{j=0}^{2^m-1} a_j x^{j},\end{aligned}$$ where $a_j\in {{\mathrm{GF}}}(2^m)$. The *algebraic degree* $\mathrm{deg}(f)$ of $f$ is then equal to the maximum $2$-weight (or Hamming weight) of an exponent $j$ for which $a_j \neq 0$, with the usual convention that the degree of the zero function is the negative infinity.
For a nonempty proper subset $D$ of ${{\mathrm{GF}}}(2^m)$, a function $g \in \mathbb{B}_m$ is called an *annihilator* of $D$ if $g f_D=0$. All annihilators of $D$ form an ideal of $\mathbb{B}_m$, denoted by $\mathrm{Ann}(D)$. The *algebraic immunity of $D$* is defined as $$\begin{aligned}
{{\mathrm{AI}}}(D)=\min \left \{ \mathrm{deg} (g) : g\in \mathrm{Ann}(D) \setminus \{0\} \right \}.\end{aligned}$$ For convenience, we define ${{\mathrm{AI}}}(\emptyset)=- \infty$ and ${{\mathrm{AI}}}({{\mathrm{GF}}}(2^m))=+ \infty$. It is easy to see that ${{\mathrm{AI}}}(\cdot)$ is monotone, which means that ${{\mathrm{AI}}}(D_1)\le {{\mathrm{AI}}}(D_2)$ for any subsets $D_1 \subseteq D_2$ of ${{\mathrm{GF}}}(2^m)$.
A vectorial Boolean $(m, r)$-function $F=(f_1, \cdots, f_r)$ is a function from ${{\mathrm{GF}}}(2^m)$ to ${{\mathrm{GF}}}(2)^r$. For any vector $v=(v_1, \cdots, v_r) \in {{\mathrm{GF}}}(2)^r$, the component function $v\cdot F$ is the Boolean function given by $v_1f_1 + \cdots + v_r f_r$. The *algebraic immunity* of $F$ is defined as $$\begin{aligned}
{{\mathrm{AI}}}(F) = \min \left \{ {{\mathrm{AI}}}\left (F^{-1}(y) \right ): y \in {{\mathrm{GF}}}(2)^r \right \},\end{aligned}$$ where $F^{-1}(y)$ is the *preimage* of $y$ under $F$. It was shown in [@DGM06] that the Hamming weight ${{\mathrm{wt}}}(f)$ of a Boolean function $f$ with prescribed algebraic immunity satisfies : $$\begin{aligned}
\sum_{i=0}^{{{\mathrm{AI}}}(f)-1} \binom{m}{i} \le {{\mathrm{wt}}}(f) \le \sum_{i=0}^{m-{{\mathrm{AI}}}(f)} \binom{m}{i}.\end{aligned}$$ It follows that ${{\mathrm{AI}}}(f) \le \lceil \frac{m}{2} \rceil$. Thus, Boolean functions attaining this upper bound are often said to have the optimal algebraic immunity. For more information on vectorial Boolean functions, the reader is referred to [@Carlet].
The *$\tau$-th order nonlinearity $\mathrm{NL}_{\tau} (f)$* of a Boolean function $f\in \mathbb B_m$ is the minimum *Hamming distance* $\mathrm{dist}(f,g)=\left | \left \{ x\in {{\mathrm{GF}}}(2^m) : f(x)\neq g(x) \right \} \right |$ between $f$ and all functions $g$ of algebraic degree at most $\tau$. The *$\tau$-th order nonlinearity $\mathrm{NL}_{\tau} (F)$* of a vectorial function $F$ is the minimum $\tau$-th order nonlinearity of its component functions. It was shown in [@Carlet08] that the $\tau$-th order nonlinearity of a vectorial $(m,r)$ function $F$ with given algebraic immunity ${{\mathrm{AI}}}(F)=t$ satisfies $$\begin{aligned}
\label{eq:NL}
\mathrm{NL}_{\tau} (F) \ge \Upsilon_{m,r,t,\tau} ,\end{aligned}$$ where $\Upsilon_{m,r,t,\tau}=2^{r-1} \sum_{i=0}^{t-\tau-1} \binom{m}{i} +2^{r-1} \sum_{i=t-2\tau}^{t-\tau-1} \binom{m-\tau}{i}$. In the particular case that $r=\tau=1$, (\[eq:NL\]) says that $$\begin{aligned}
\label{eq:NL-1-1}
\mathrm{NL}_{1} (f) \ge 2 \sum_{i=0}^{{{\mathrm{AI}}}(f)-2} \binom{m-1}{i},\end{aligned}$$ where $f\in \mathbb B_m$.
Minimal codes and cyclic codes
------------------------------
We assume that the reader is familiar with the basics of linear codes (see for instance [@MS77] for detail). A linear code of length $n$ and dimension $k$ will be referred to as an $[n, k]$ code. Further, if the code has minimum distance $d$, it will be referred to as an $[n, k, d]$ code.
The Hamming weight (for short, weight) of a vector $\mathbf v$ is the number of its nonzero entries and is denoted $\mathrm{wt}(\mathbf v)$. The minimum (respectively, maximum) weight of the code ${{\mathcal{C}}}$ is the minimum (respectively, maximum) nonzero weight of all codewords of ${{\mathcal{C}}}$, $w_{\min} = \min (\mathrm{wt}(\mathbf c))$ (respectively, $w_{\max} = \max (\mathrm{wt}(\mathbf c))$).
Let ${{\mathbf{c}}}=(c_0, \cdots, c_{n-1})$ be a codeword in ${{\mathcal{C}}}$. The *support* ${{\mathrm{Supp}}}({{\mathbf{c}}})$ of the codeword ${{\mathbf{c}}}$ is the set of indices of its nonzero coordinates: $${{\mathrm{Supp}}}({{\mathbf{c}}})=\{i: c_i \neq 0\}.$$ A codeword ${{\mathbf{c}}}$ of the linear code ${{\mathcal{C}}}$ is called minimal if its support does not contain the support of any other linearly independent codeword. ${{\mathcal{C}}}$ is called a minimal linear code if all codewords of ${{\mathcal{C}}}$ are minimal. Minimal codes are a special class of linear codes. A sufficient condition for a linear code to be minimal is given in the following lemma [@AB98].
\[lem:AB\] A linear code ${{\mathcal{C}}}$ over ${{\mathrm{GF}}}(q)$ is minimal if $\frac{w_{\min}}{w_{\max}} >\frac{q-1}{q}$.
Let $\mathcal C$ be an $[n,k,d]$ linear code over $\mathrm{GF}(q)$ and $T$ a set of $t$ coordinate locations of $\mathcal C$. Then the code ${{\mathcal{C}}}^T$ obtained from ${{\mathcal{C}}}$ by puncturing at the locations in $T$ is the code of length $n-t$ consisting of codewords of ${{\mathcal{C}}}$ which have their coordinate at the location $P$ deleted if $P\in T$ and left alone if $P \not \in T$, which is called the *punctured code* of $\mathcal C$ on $T$. The *shortened code* $\mathcal C_{T}$ is the set of codewords from ${{\mathcal{C}}}$ that are zero at locations in $T$ , with coordinates in $T$ deleted.
An $[n,k]$ linear code ${{\mathcal{C}}}$ over ${{\mathrm{GF}}}(q)$ is called *cyclic* if $(c_0,c_1,\cdots, c_{n-1})\in {{\mathcal{C}}}$ implies that the circular shift $(c_{n-1},c_0,\cdots, c_{n-2})\in {{\mathcal{C}}}$. Clearly the vector space ${{\mathrm{GF}}}(q)^n$ is isomorphic to the residue class ring ${{\mathrm{GF}}}(q)[X]/(X^n-1)$ (considered as an additive group). An isomorphism is given by $$\begin{aligned}
(c_0, c_1, \cdots, c_{n-1}) \longleftrightarrow c_0+c_1X+\cdots +c_{n-1} X^{n-1}.\end{aligned}$$ From now on we do not distinguish between codewords of ${{\mathcal{C}}}$ and polynomials of degree less than $n$ over ${{\mathrm{GF}}}(q)$. Note that the multiplication by $X$ in ${{\mathrm{GF}}}(q)[X]/(X^n-1)$ amounts to the circular right shift $(c_0, c_1, \cdots, c_{n-1}) \longrightarrow (c_{n-1},c_0,\cdots, c_{n-2})$. From this it follows that a cyclic code ${{\mathcal{C}}}$ corresponds to an ideal in ${{\mathrm{GF}}}(q)[X]/(X^n-1)$, which we also denote by ${{\mathcal{C}}}$. Every $[n,k]$ cyclic code ${{\mathcal{C}}}$ over ${{\mathrm{GF}}}(q)$ is a principal ideal generated by some polynomial $g(X)$ of degree $n-k$ that divides $X^n-1$. We shall call $g(X)$ and $h(X)=(X^n-1)/g(X)$ the *generator polynomial* and the *check polynomial* of ${{\mathcal{C}}}$, respectively. Note that the codewords $g(X), Xg(X)$, $\cdots$, $X^{k-1}g(X)$ form a basis of ${{\mathcal{C}}}$.
Let us recall the *BCH bound* on the minimum distance of cyclic codes [@HT72].
Let $h$ be an integer and $\delta$ be a positive integer with $1\le \delta<n$. Let $\alpha$ be a primitive $n$-th root of unity in the algebraic closure of ${{\mathrm{GF}}}(q)$. Let ${{\mathcal{C}}}$ be a cyclic code of length $n$ over ${{\mathrm{GF}}}(q)$ with generator polynomial $g(X)$. If $g(X)$ has $\delta$ consecutive zeros $\alpha^{h}, \cdots, \alpha^{h+\delta-1}$, then the minimum distance of ${{\mathcal{C}}}$ is greater than $\delta$.
Reed-Muller codes
-----------------
Reed-Muller (RM) codes are classical codes that have enjoyed unabated interest since their introduction in 1954 due to their simple recursive structure.
Let $\alpha$ be a primitive element of ${{\mathrm{GF}}}(2^m)$. Let $P_0=0$ and $P_j=\alpha^{j-1}$, where $1\le j \le 2^m-1$. Then $P_0,...,P_{2^m-1}$ is an enumeration of the points of the vector space ${{\mathrm{GF}}}(2^m)$. Under this enumeration, the Reed-Muller code $\mathrm{RM}(\ell, m)$ of order $\ell $ in $m$ variables is defined as $$\begin{aligned}
\mathrm{RM}(\ell, m)=\left \{(f(P_0), \cdots, f(P_{2^m-1})): f\in \mathbb{B}_m, \mathrm{deg}(f)\le \ell\right \}.\end{aligned}$$ In this paper, we index the coordinates of the code $\mathrm{RM}(\ell, m)$ with the sequence $(P_0, P_1, \ldots, P_{2^m-1})$. The general affine group over ${{\mathrm{GF}}}(2^m)$, denoted by $\mathrm{GA}(1,2^m)$, is defined by $$\mathrm{GA}(1,2^m) =\left \{\pi_{a,b} : a\in {{\mathrm{GF}}}(2^m)^*, b\in {{\mathrm{GF}}}(2^m) \right \},$$ where $\pi_{a,b}$ is the permutation on ${{\mathrm{GF}}}(2^m)$ defined by $x \mapsto ax+b$. Since $\mathrm{deg}(f(x))=\mathrm{deg}(f(ax+b))$ for any $(a, b) \in {{\mathrm{GF}}}(2^m)^* \times {{\mathrm{GF}}}(2^m)$, the Reed-Muller code $\mathrm{RM}(\ell, m)$ is invariant under the action by $\mathrm{GA}(1,2^m)$. We denote the codes obtained after the *puncturing* and *shortening* operation on $\mathrm{RM}(\ell, m)$ at the coordinate location $P_0$ as $\mathrm{PRM}(\ell, m)$ and ${{\mathrm{SRM}}}(\ell,m)$, respectively. It is easy to see that the punctured code $\mathrm{PRM}(\ell, m)$ and the shortened code ${{\mathrm{SRM}}}(\ell,m)$ of the Reed-Muller code $\mathrm{RM}(\ell, m)$ are cyclic codes of length $2^m-1$. Let $g_{\ell, \alpha}(X)$ and $g_{\ell, \alpha}^*(X)$ denote the generator polynomials of the cyclic codes ${{\mathrm{PRM}}}(\ell, m)$ and ${{\mathrm{SRM}}}(\ell, m)$, respectively.
The following proposition describes the generator polynomials of the punctured Reed-Muller codes, which are not hard to prove [@AssKey98].
\[prop:prm-generator-alpha\] The punctured Reed-Muller code $\mathrm{PRM} (\ell,m)$ is a cyclic code of dimension $\sum_{j=0}^{\ell} \binom{m}{j}$ with generator polynomial $$\begin{aligned}
g_{\ell, \alpha}(X)= \prod_{ \scriptsize{\begin{array}{c}0<i_{m-1}+\cdots+ i_0 \le m-1 -\ell
\\ i_{m-1}, \cdots, i_0 \in \{0,1\} \end{array}}} (X-\alpha^{i_{m-1}2^{m-1}+ \cdots+i_0 2^0}).
\end{aligned}$$
The following proposition is taken from Corollary 4 of [@Assmus92].
\[prop:mini-geometry\] The minimum weight of the Reed-Muller code $\mathrm{RM}(\ell,m)$ is $2^{m-\ell}$ and the minimum-weight codewords are the incidence vectors of the $(m-\ell)$-flats of the affine space $\mathrm{AG}(m,2)$ of dimension $m$ over ${{\mathrm{GF}}}(2)$. The minimum weight of the punctured code $\mathrm{PRM} (\ell,m)$ is $2^{m-\ell}-1$ and the minimum-weight codewords are the incidence vectors of the $(m-\ell-1)$-dimensional subspaces of the projective space $\mathrm{PG}(m-1,2)$ of dimension $m-1$ over ${{\mathrm{GF}}}(2)$.
\[lem:minimum-set-geometry\] The minimum weight of the shortened Reed-Muller code ${{\mathrm{SRM}}}(\ell,m)$ is $2^{m-\ell}$ and the minimum-weight codewords are the incidence vectors of the $(m-\ell)$-flats not passing through the origin in $\mathrm{AG}(m,2)$.
Note that the shortened code ${{\mathrm{SRM}}}(\ell,m)$ consists of codewords of $\mathrm{RM}(\ell,m)$ that are zero at the origin of $\mathrm{AG}(m,2)$. The desired result then follows from Proposition \[prop:mini-geometry\].
Minimal codes from sets without nonzero low-degree annihilators {#sec:codes-sets}
================================================================
In this section we present a general construction of minimal codes using subcodes of Reed-Muller codes and subsets of finite fields without nonzero low-degree annihilators.
Here and hereafter, for any subset $D$ of ${{\mathrm{GF}}}(2^m)$, let $D^*$ denote the set $D\setminus \{0\}$ and $\overline{D}$ stand for the complement of $D$ in ${{\mathrm{GF}}}(2^m)$. In particular, if $D\subseteq {{\mathrm{GF}}}(2^m)^*$, then $\overline{D}^*$ is the complement of $D$ in ${{\mathrm{GF}}}(2^m)^*$. Let $\mathbb B_m^0$ denote the set $\{f \in \mathbb{B}_m : f(0)=0\}$.
The following theorem presents a general approach to constructing binary minimal codes, and produces many classes of binary minimal codes by selecting some subcodes of Reed-Muller codes and subsets of $ {{\mathrm{GF}}}(2^m)$ with special annihilators.
\[thm:min-rm-ann\] Let ${{\mathcal{C}}}$ be a $k$-dimensional subcode of the Reed-Muller code $\rm{RM} (\ell, m)$. Let $D$ be a subset of ${{\mathrm{GF}}}(2^m)$. Then ${{\mathcal{C}}}^{\overline{D}}$ is a minimal code of dimension $k$ if and only if the following two conditions hold:
(1) the code ${{\mathcal{C}}}$ is minimal, and
(2) for any two nonzero codewords $(f_1(P_0), \cdots, f_1(P_{2^m-1}))$ and $(f_2(P_0), \cdots, f_2(P_{2^m-1}))$ of ${{\mathcal{C}}}$ (including the case $f_1=f_2$), where $f_1, f_2 \in \mathbb{B}_m$, the product $f_1f_2$ of $f_1$ and $f_2$ is not an annihilator of $D$.
Let ${{\mathcal{C}}}$ be a linear code satisfying Conditions (1) and (2). Let $(f(P_0), \cdots, f(P_{2^{m}-1}))$ be any nonzero codeword of ${{\mathcal{C}}}$, where $f \in \mathbb B_{m}$. Let $f_1=f_2=f$. Then $f_1f_2=f^2=f$, and by Condition (2), $f$ is not an annihilator of $D$, i.e., $\left ( f(P) \right )_{P\in D} \neq 0$. Consequently the punctured code ${{\mathcal{C}}}^{\overline{D}}$ has the same dimension as the original code ${{\mathcal{C}}}$. Suppose that ${{\mathcal{C}}}^{\overline{D}}$ is not minimal. Then there exist two distinct nonzero codewords $\left ( f_1(P) \right )_{P\in D}, \left ( f_2(P) \right )_{P\in D} \in {{\mathcal{C}}}^{\overline{D}}$, where $f_1, f_2 \in
\mathbb B_m$, such that ${{\mathrm{Supp}}}\left ( \left ( f_1(P) \right )_{P\in D} \right ) \subsetneq {{\mathrm{Supp}}}\left ( \left ( f_2(P) \right )_{P\in D} \right )$. This clearly forces $${{\mathrm{Supp}}}\left ( \left ( f_2(P) \right )_{P\in D} \right ) = {{\mathrm{Supp}}}\left ( \left ( (f_1 + f_2)(P) \right )_{P\in D} \right ) \dot\cup
{{\mathrm{Supp}}}\left ( \left ( f_1(P) \right )_{P\in D} \right ).$$ It follows that $f_1 (f_1 + f_2) \in \mathrm{Ann}(D)$, which is contrary to Condition (2). Therefore ${{\mathcal{C}}}^{\overline{D}}$ is minimal.
Conversely, assume ${{\mathcal{C}}}^{\overline{D}}$ is a minimal code with dimension $k$. It is clear that ${{\mathcal{C}}}$ is minimal. It remains to show that Condition (2) holds. On the contrary, suppose that there exist two nonzero codewords $(f_1(P_0), \cdots, f_1(P_{2^m-1}))$ and $(f_2(P_0), \cdots, f_2(P_{2^m-1}))$ of ${{\mathcal{C}}}$, where $f_1, f_2 \in \mathbb{B}_m$, such that $f_1f_2 \in \mathrm{Ann}(D)$. Then ${{\mathrm{Supp}}}\left ( \left ( f_1(P) \right )_{P\in D} \right )$ and ${{\mathrm{Supp}}}\left ( \left ( f_2(P) \right )_{P\in D} \right )$ are disjoint. This yields $${{\mathrm{Supp}}}\left ( \left ( f_1(P) \right )_{P\in D} \right ) \subsetneq {{\mathrm{Supp}}}\left ( \left ( (f_1 + f_2)(P) \right )_{P\in D} \right),$$ which contradicts the minimality of nonzero codewords of ${{\mathcal{C}}}^{\overline{D}}$. This completes the proof.
\[cor:ad>2l\] Let ${{\mathcal{C}}}$ be a minimal code contained in the Reed-Muller code $ \mathrm{RM} (\ell, m)$. Let $D$ be a subset of ${{\mathrm{GF}}}(2^m)^*$ with ${{\mathrm{AI}}}(D) \ge 2 \ell +1$. Then ${{\mathcal{C}}}^{\overline{D}}$ is a minimal code of dimension $k$, where $k$ equals the dimension of ${{\mathcal{C}}}$.
Let $(f_1(P_0), \cdots, f_1(P_{2^m-1}))$ and $(f_2(P_0), \cdots, f_2(P_{2^m-1}))$ be any two nonzero codewords of ${{\mathcal{C}}}$, where $f_1, f_2 \in \mathbb{B}_m$. Since ${{\mathcal{C}}}$ is minimal, $f_1f_2$ cannot be the zero function. From $\mathrm{deg}(f_1f_2)\le 2\ell$ and ${{\mathrm{AI}}}(D) \ge 2 \ell +1$, we conclude that $f_1f_2 \not \in \mathrm{Ann}(D)$. The desired result then follows from Theorem \[thm:min-rm-ann\].
\[cor:C-shortened-ad>2l\] Let ${{\mathcal{C}}}$ be a minimal code contained in the shortened Reed-Muller code $ \mathrm{SRM} (\ell, m)$. Let $D$ be a subset of ${{\mathrm{GF}}}(2^m)^*$ with ${{\mathrm{AI}}}(D \cup \{ 0 \}) \ge 2 \ell +1$. Then ${{\mathcal{C}}}^{\overline{D}^*}$ is a minimal code of dimension $k$, where $k$ equals the dimension of ${{\mathcal{C}}}$.
Denote by ${{\mathcal{C}}}'$ the codes $\{ (0, \mathbf c): \mathbf c \in {{\mathcal{C}}}\}$. Since ${{\mathrm{AI}}}(D \cup \{ 0 \}) \ge 2 \ell +1$, it follows from Corollary \[cor:ad>2l\] that ${{\mathcal{C}}}'^{\overline{D \cup \{0\}}}$ is a minimal code of dimension $k$. The desired conclusion then follows from the definitions of ${{\mathcal{C}}}'$ and shortened codes.
To deduce a lower bound on the minimum distance of the codes from sets with high algebraic immunity, we need some additional lemmas. Denote by $\mathrm{Ann}_{t-1} (g)$ the vector space of those annihilators of degrees at most $t-1$ of ${{\mathrm{Supp}}}(g)$.
\[lem:wt-dim\] Let $D\subseteq {{\mathrm{GF}}}(2^m)$ with ${{\mathrm{AI}}}(D)=t$ and $g\in \mathbb B_m$. Then $$\begin{aligned}
{{\mathrm{wt}}}(g f_D) \ge \mathrm{dim} \left (\mathrm{Ann}_{t-1} (1+g) \right ).\end{aligned}$$
Let $w= {{\mathrm{wt}}}(g f_D)$ and let $\mathcal Q$ be the set consisting of $w$ distinct points $Q_1$, $\cdots$, $Q_w$ in $D$ satisfying $g(Q_i)=1$. Consider the evaluation map $\mathrm{Ev}_{\mathcal Q}$ $$\mathrm{Ann}_{t-1} (1+g) \longrightarrow {{\mathrm{GF}}}(2)^w,$$ defined by $\mathrm{Ev}_{\mathcal Q} (h)= \left ( h(Q_1), \cdots, h(Q_w) \right )$. Then $\mathrm{Ev}_{\mathcal Q}$ is a linear transformation. Suppose the assertion of the lemma is false. Then $\mathrm{Ev}_{\mathcal Q}$ is not injective. Thus there exists a nonzero function $h$ in $\mathrm{Ann}_{t-1} (1+g) $ such that $h f_{\mathcal Q}=0$. It follows easily that $h \in \mathrm{Ann}(D)$, which contradicts the condition that ${{\mathrm{AI}}}(D)=t$. This completes the proof.
Little is known about the behavior of the annihilators of a polynomial of a given degree. Mesnager [@Sihem08] proved the following lower bound on the dimension of $\mathrm{Ann}_{t-1} (g)$ $$\mathrm{dim} \left (\mathrm{Ann}_{t-1} (g) \right ) \ge \sum_{i=0}^{t-\tau-1} \binom{m-\tau}{i},$$ where $g \in \mathbb B_m$ with $\mathrm{deg}(g)=\tau$. Lemma \[lem:wt-dim\] indicates that the following lemma holds.
\[lem:wt-bound\] Let $D\subseteq {{\mathrm{GF}}}(2^m)$ with ${{\mathrm{AI}}}(D)=t$ and $g\in \mathbb B_m \setminus \{0\}$ with $\mathrm{deg} (g)=\tau$. Then $$\begin{aligned}
{{\mathrm{wt}}}(g f_D) \ge \sum_{i=0}^{t-\tau-1} \binom{m-\tau}{i}.\end{aligned}$$
A method of explicitly constructing minimal codes by puncturing the Simplex codes is given in the following theorem.
\[thm:C-D\] Let $D$ be a subset of ${{\mathrm{GF}}}(2^m)^*$ with ${{\mathrm{AI}}}\left (D\cup \{0\} \right )= t \ge 3$. Let ${{\mathcal{C}}}(D)$ be the linear code given by $$\begin{aligned}
\label{eqn:C-D}
{{\mathcal{C}}}(D) = \left \{ \left ( {{\mathrm{Tr}}}^m_1 (a x) \right )_{x \in D} : a \in {{\mathrm{GF}}}(2^m)\right \}.
\end{aligned}$$ Then ${{\mathcal{C}}}(D)$ is a minimal code with parameters $\left [ |D|, m, \ge \sum_{i=0}^{t-2} \binom{m-1}{i} \right ]$.
Let ${{\mathcal{C}}}\subseteq {{\mathrm{SRM}}}(1, m)$ be the Simplex code defined by $$\left \{ \left ( {{\mathrm{Tr}}}^m_1 (a P_1) , \cdots, {{\mathrm{Tr}}}^m_1 (a P_{2^m-1}) \right ) : a \in {{\mathrm{GF}}}(2^m) \right \}.$$ It is well-known that ${{\mathcal{C}}}$ is a minimal code of dimension $m$. The desired conclusions are immediate from Corollary \[cor:C-shortened-ad>2l\] and Lemma \[lem:wt-bound\].
\[cor:C-Supp-f\] Let $m\ge 5$ be an integer and let $f\in \mathbb B_m$ be a Boolean function with ${{\mathrm{AI}}}\left (f \right )=t \ge 3$ and ${{\mathrm{wt}}}(f) \ge 2^{m-1}$. Let $D= {{\mathrm{Supp}}}(f) \setminus \{0\}$. Then the code ${{\mathcal{C}}}\left (D \right )$ defined by (\[eqn:C-D\]) is an $m$-dimensional minimal code with minimum distance $d$ satisfying $$\begin{aligned}
d \ge \sum_{i=0}^{t-2} \binom{m-1}{i} + \frac{1}{2}\left ( {{\mathrm{wt}}}(f)- 2^{m-1} \right ).\end{aligned}$$
It follows from Theorem \[thm:C-D\] that ${{\mathcal{C}}}\left (D \right )$ is an $m$-dimensional minimal code. It remains to prove the lower bound on the minimum distance of ${{\mathcal{C}}}\left (D \right )$. Denote by $g$ the Boolean function ${{\mathrm{Tr}}}^m_1 (a x)$, where $a \in {{\mathrm{GF}}}(2^m)^*$. Let $w= {{\mathrm{wt}}}\left ( \left ( {{\mathrm{Tr}}}^m_1 (a x) \right )_{x \in D} \right )$. Thus $$w= \left| \{x\in {{\mathrm{GF}}}(2^m): f(x)=1, g(x)=1\} \right |.$$ By the definition of the Hamming distance between $f$ and $g$, we have $$\begin{aligned}
\begin{array}{rl}
{{\mathrm{dist}}}(f,g)=&\left| \{x\in {{\mathrm{GF}}}(2^m): f(x)=0, g(x)=1 \}\right |\\
& + \left| \{x\in {{\mathrm{GF}}}(2^m): f(x)=1, g(x)=0 \}\right |\\
=& {{\mathrm{wt}}}(g) -w +{{\mathrm{wt}}}(f) -w\\
=& {{\mathrm{wt}}}(g) +{{\mathrm{wt}}}(f) -2w.
\end{array}
\end{aligned}$$ This yields $$\begin{aligned}
\label{eq:local-global}
\begin{array}{rl}
w=& \left ( {{\mathrm{wt}}}(g) +{{\mathrm{wt}}}(f) - {{\mathrm{dist}}}(f,g) \right )/2\\
=& \left ( {{\mathrm{dist}}}(f,1+g)+ {{\mathrm{wt}}}(g) + {{\mathrm{wt}}}(f) -2^m \right )/2.\\
\end{array}\end{aligned}$$ The desired conclusion then follows from (\[eq:NL-1-1\]), (\[eq:local-global\]) and the fact that ${{\mathrm{wt}}}(g) =2^{m-1}$.
Now, consider balanced Boolean functions in Corollary \[cor:C-Supp-f\].Then, we obtain the following result.
\[cor:C-Supp-f-balanced\] Let $m\ge 5$ be an integer. Let $f\in \mathbb B_m$ be a balanced Boolean function with ${{\mathrm{AI}}}\left (f \right )=t \ge 3$. Let $D= {{\mathrm{Supp}}}(f) \setminus \{0\}$. Then the code ${{\mathcal{C}}}\left (D \right )$ defined by (\[eqn:C-D\]) is a $\left [2^{m-1}, m, \ge \sum_{i=0}^{t-2} \binom{m-1}{i} \right ]$ minimal code. Moreover, if $f$ has optimum algebraic immunity, then ${{\mathcal{C}}}\left (D \right )$ is a minimal code with parameters $\left [ 2^{m-1}, m, \ge \sum_{i=0}^{ \lceil \frac{m-4}{2} \rceil } \binom{m-1}{i}\right].$
In order to apply Theorem \[thm:min-rm-ann\] to construct minimal codes, finding sets without low-degree nonzero annihilators is very important. Let $\alpha$ be a primitive element of ${{\mathrm{GF}}}(2^m)^*$, $h$ and $\delta$ be two integers with $\delta >0$. Denote then $[h; \delta]_{\alpha}=\left \{\alpha^{h}, \alpha^{h+1}, \cdots, \alpha^{h+\delta-1} \right \}$.
\[lem:ai-det-0\] Let $\delta $ be an integer with $\sum_{i=0}^{t} \binom{m}{i} \le \delta < \sum_{i=0}^{t+1} \binom{m}{i}$. Then ${{\mathrm{AI}}}\left ( [h; \delta]_{\alpha} \right )= t+1$.
Let $f$ be a function of degree at most $t$ in $ \mathrm{Ann} ([h; \delta]_{\alpha} )$ and $\tilde{f}(X) \in {{\mathrm{PRM}}}(t, m)$ be the codeword associated with $f$. By assumption, we see that $$\begin{aligned}
\label{eq:f-codew}
\tilde{f}(X)=X^h \sum_{i= \delta }^{2^m-2} c_i X^i,\end{aligned}$$ where $c_i \in {{\mathrm{GF}}}(2)$. By the definition of the generator polynomial $g_{t, m}(X)$ of ${{\mathrm{PRM}}}(t, m)$, the codeword $X^{-h} \tilde{f}(X)$ can also be written as $$\begin{aligned}
\label{eq:f-codew-g}
X^{-h} \tilde{f}(X) =\left (a_0+a_1 X+ \cdots a_{\delta_t -1 } X^{\delta_t-1} \right ) g_{t, m}(X),\end{aligned}$$ where $\delta_t= \sum_{i=0}^{t} \binom{m}{i} $ and $a_i \in {{\mathrm{GF}}}(2)$. Combining (\[eq:f-codew\]) with (\[eq:f-codew-g\]) yields $a_0= \cdots = a_{\delta_t -1 } =0$. We thus get $f=0$. Hence ${{\mathrm{AI}}}([h; \delta]_{\alpha} ) \ge t+1$.
Let $f$ be the function corresponding to the codeword $\tilde{f}(X) $ of ${{\mathrm{PRM}}}(t+1, m)$, where $\tilde{f}(X) $ is given by $$\begin{aligned}
\tilde{f}(X)=X^h
\left (X^{\delta}+ X^{\delta+1} \cdots + X^{\delta_{t+1}-1} \right ) g_{t+1, m}(X),\end{aligned}$$ where $\delta_{t+1}= \sum_{i=0}^{t+1} \binom{m}{i} $. It is easy to check that $f \in \mathrm{Ann} ( [h; \delta]_{\alpha} ) \setminus \{0\}$. It follows that ${{\mathrm{AI}}}([h; \delta]_{\alpha} ) \le t+1$.
Summarising the discussions above yields ${{\mathrm{AI}}}([h; \delta]_{\alpha} ) = t+1$, which is the desired conclusion.
The proof of the following lemma is similar to the proof of Lemma \[lem:ai-det-0\], with punctured Reed-Muller codes replaced by shortened Reed-Muller codes, and therefore is omitted.
\[lem:\[h;delta\]+0\] Let $\delta $ be an integer with $\sum_{i=1}^{t} \binom{m}{i} \le \delta < \sum_{i=1}^{t+1} \binom{m}{i}$. Then ${{\mathrm{AI}}}\left ( [h; \delta]_{\alpha} \cup \{0\}\right )= t+1$.
Now we recall some facts on Gauss sums which will be needed to derive an improved lower bound on the minimum distance of the codes from the sets $[h; \delta]_{\alpha}$. Let $\xi_{q-1}$ denote the complex primitive $(q-1)$th root of unity $e^{2\pi \sqrt{-1}/(q-1)}$. Let $\alpha $ be a primitive element of ${{\mathrm{GF}}}(q)$, and let $\chi$ be the character of ${{\mathrm{GF}}}(q)^*$ given by $$\begin{aligned}
\chi(\alpha^j)=\xi_{q-1}^j,\end{aligned}$$ where $0 \le j \le q-2$. The *Gauss sum* associated to $\chi^j$ over ${{\mathrm{GF}}}(q)$ with $q=2^m$ is defined by $$\begin{aligned}
\label{eq:Gass-number}
G(\chi^j)= \sum_{i=0}^{q-2} (-1)^{{{\mathrm{Tr}}}^{m}_1\left( \alpha^i \right ) } \chi\left (\alpha^{ij}\right ), \text{ for } j=0, \cdots, q-2.\end{aligned}$$ Then $G(\chi^0)=-1$ and $G(\chi^j)$ ($1\le j \le q-2$) satisfies the fundamental property [@IreRosen72 p. 132] $$\begin{aligned}
\label{eq:sqrt-q}
G(\chi^j) \overline{G(\chi^j)} =q,\end{aligned}$$ where the bar denotes complex conjugate. It is sometimes convenient to view the Gauss sum $G(\chi^j)$ as a function of $\chi^j$. This amounts to viewing $\chi^j \mapsto G(\chi^j)$ as the multiplicative Fourier transformation of the function $(-1)^{{{\mathrm{Tr}}}^m_1(x)}$ on ${{\mathrm{GF}}}(q)^*$. The following Fourier inversion formula allows us to recover $(-1)^{{{\mathrm{Tr}}}^m_1(x)}$ from $G(\chi^j)$ by $$\begin{aligned}
\label{eq:add-mulp-Gauss}
(-1)^{{{\mathrm{Tr}}}^m_1(\alpha^i)}=\frac{1}{q-1} \sum_{j=0}^{q-2} \overline{\chi}^j (\alpha^i) G(\chi^j) .\end{aligned}$$
We will need the following lemma, whose proof can be found in [@CFeng09].
\[lem:exp-sin-ln\] Let $q=2^m$. It holds $$\begin{aligned}
\sum_{j=1}^{2^{m-1}-1} \frac{1}{\sin \left (\pi j/(q-1) \right )} \le \frac{q-1}{2 \pi } \ln \left ( \frac{4(q-1)}{\pi} \right ).\end{aligned}$$
Combining Simplex codes with the sets $[h; \delta]_{\alpha}$, an infinite class of binary minimal codes is given in the following theorem.
\[thm:C-\[h;delta\]\] Let $q=2^m$. Let $\delta $ be an integer with $\sum_{i=1}^{t-1} \binom{m}{i} \le \delta < \sum_{i=1}^{t} \binom{m}{i}$ and $3\le t \le m$. Then the code ${{\mathcal{C}}}\left ([h; \delta]_{\alpha} \right )$ defined by (\[eqn:C-D\]) is a minimal code with parameters $[\delta, m ,d]$, where $$\begin{aligned}
d \ge \max \left \{ \sum_{i=0}^{t-2} \binom{m-1}{i}, \frac{\delta-1}{2}- \frac{\sqrt{q}}{2 \pi} \ln \left ( \frac{4(q-1)}{\pi} \right ) \right \}.
\end{aligned}$$
Let ${{\mathbf{c}}}=\left ( {{\mathrm{Tr}}}^m_1 (\lambda \alpha^{h}), \cdots, {{\mathrm{Tr}}}^m_1 (\lambda \alpha^{h+\delta-1} ) \right ) $ be a nonzero codeword of ${{\mathcal{C}}}\left ([h; \delta]_{\alpha} \right )$. Then $$\begin{aligned}
\label{eq:wt-av-F}
\begin{array}{rl}
{{\mathrm{wt}}}({{\mathbf{c}}}) = & \frac{1}{2} \sum_{ i=h }^{h+\delta-1} \left (1-(-1)^{ {{\mathrm{Tr}}}^m_1 (\lambda \alpha^{i})} \right )\\
=& \frac{1}{2}\delta-\frac{1}{2} \sum_{ i=0 }^{\delta -1} (-1)^{ {{\mathrm{Tr}}}^m_1 (\lambda \alpha^h \alpha^{i})}\\
=& \frac{1}{2} \delta- \frac{1}{2} F,
\end{array}\end{aligned}$$ where $F=\sum_{ i=0 }^{\delta -1} (-1)^{ {{\mathrm{Tr}}}^m_1 (\lambda \alpha^h \alpha^{i})}$. Set $\lambda'= \lambda \alpha^h$. Substituting (\[eq:add-mulp-Gauss\]) into $F$ yields $$\begin{aligned}
\begin{array}{rl}
|F|=& \frac{1}{q-1} |\sum_{i=0}^{\delta -1} \sum_{j=0}^{q-2} \overline{\chi}
(\lambda'^j \alpha^{ij}) G(\chi^{j})| \\
=& \frac{1}{q-1} |\sum_{j=0}^{q-2} \overline{\chi} (\lambda'^j ) G(\chi^{j}) \sum_{i=0}^{\delta -1} \overline{\chi}
( \alpha^{ij}) | \\
=&|-\frac{1}{q-1} \delta + \frac{1}{q-1} \sum_{j=1}^{q-2} \overline{\chi} (\lambda'^j ) G(\chi^{j}) \sum_{i=0}^{\delta -1} \overline{\chi}
( \alpha^{ij}) |\\
\le & \frac{1}{q-1} \delta +\frac{\sqrt{q}}{q-1} \sum_{j=1}^{q-2} \left |\sum_{i=0}^{\delta -1} \xi_{q-1}^{ij} \right |,
\end{array}\end{aligned}$$ where the last inequality follows from (\[eq:sqrt-q\]). A simple calculation yields $$\begin{aligned}
\label{eq:F-indpendentOFdelta}
\begin{array}{rl}
|F|\le & \frac{1}{q-1} \delta +\frac{\sqrt{q}}{q-1} \sum_{j=1}^{q-2} \left | \frac{\sin \left (\pi \delta j/(q-1) \right )}{\sin \left (\pi j/(q-1) \right )} \right | \\
\le & \frac{1}{q-1} \delta +\frac{2\sqrt{q}}{q-1} \sum_{j=1}^{2^{m-1}-1} \frac{1}{\sin \left (\pi j/(q-1) \right )}\\
\le & \frac{1}{q-1} \delta + \frac{\sqrt{q}}{\pi} \ln \left ( \frac{4(q-1)}{\pi} \right ),
\end{array}\end{aligned}$$ where the last inequality follows from Lemma \[lem:exp-sin-ln\]. Combining (\[eq:wt-av-F\]) with (\[eq:F-indpendentOFdelta\]), we deduce that $$\begin{aligned}
{{\mathrm{wt}}}({{\mathbf{c}}}) \ge \frac{\delta-1}{2}- \frac{\sqrt{q}}{2 \pi} \ln \left ( \frac{4(q-1)}{\pi} \right ).\end{aligned}$$ The desired conclusion then follows from Theorem \[thm:C-D\] and Lemma \[lem:\[h;delta\]+0\].
coordinates [ (36,8) (37,8) (38,8) (39,8) (40,8) (41,8) (42,8) (43,8) (44,8) (45,8) (46,8) (47,9) (48,9) (49,10) (50,10) (51,11) (52,11) (53,12) (54,12) (55,13) (56,13) (57,14) (58,14) (59,15) (60,15) (61,16) (62,16) (63,17) (64,17) (65,18) (66,18) (67,19) (68,19) (69,20) (70,20) (71,21) (72,21) (73,22) (74,22) (75,23) (76,23) (77,24) (78,24) (79,25) (80,25) (81,26) (82,26) (83,27) (84,27) (85,28) (86,28) (87,29) (88,29) (89,30) (90,30) (91,31) (92,31) (93,32) (94,32) (95,33) (96,33) (97,34) (98,34) (99,35) (100,35) (101,36) (102,36) (103,37) (104,37) (105,38) (106,38) (107,39) (108,39) (109,40) (110,40) (111,41) (112,41) (113,42) (114,42) (115,43) (116,43) (117,44) (118,44) (119,45) (120,45) (121,46) (122,46) (123,47) (124,47) (125,48) (126,48) (127,49) (128,49) (129,50) (130,50) (131,51) (132,51) (133,52) (134,52) (135,53) (136,53) (137,54) (138,54) (139,55) (140,55) (141,56) (142,56) (143,57) (144,57) (145,58) (146,58) (147,59) (148,59) (149,60) (150,60) (151,61) (152,61) (153,62) (154,62) (155,63) (156,63) (157,64) (158,64) (159,65) (160,65) (161,66) (162,66) (163,67) (164,67) (165,68) (166,68) (167,69) (168,69) (169,70) (170,70) (171,71) (172,71) (173,72) (174,72) (175,73) (176,73) (177,74) (178,74) (179,75) (180,75) (181,76) (182,76) (183,77) (184,77) (185,78) (186,78) (187,79) (188,79) (189,80) (190,80) (191,81) (192,81) (193,82) (194,82) (195,83) (196,83) (197,84) (198,84) (199,85) (200,85) (201,86) (202,86) (203,87) (204,87) (205,88) (206,88) (207,89) (208,89) (209,90) (210,90) (211,91) (212,91) (213,92) (214,92) (215,93) (216,93) (217,94) (218,99) (219,99) (220,99) (221,99) (222,99) (223,99) (224,99) (225,99) (226,99) (227,99) (228,99) (229,100) (230,100) (231,101) (232,101) (233,102) (234,102) (235,103) (236,103) (237,104) (238,104) (239,105) (240,105) (241,106) (242,106) (243,107) (244,107) (245,108) (246,120) (247,120) (248,120) (249,120) (250,120) (251,120) (252,120) (253,120) (254,127) ]{};
coordinates [ (36,8) (37,8) (38,8) (39,8) (40,8) (41,8) (42,8) (43,8) (44,8) (45,8) (46,8) (47,8) (48,8) (49,8) (50,8) (51,8) (52,8) (53,8) (54,8) (55,8) (56,8) (57,8) (58,8) (59,8) (60,8) (61,8) (62,8) (63,8) (64,8) (65,8) (66,8) (67,8) (68,8) (69,8) (70,8) (71,8) (72,8) (73,8) (74,8) (75,8) (76,8) (77,8) (78,8) (79,8) (80,8) (81,8) (82,8) (83,8) (84,8) (85,8) (86,8) (87,8) (88,8) (89,8) (90,8) (91,8) (92,29) (93,29) (94,29) (95,29) (96,29) (97,29) (98,29) (99,29) (100,29) (101,29) (102,29) (103,29) (104,29) (105,29) (106,29) (107,29) (108,29) (109,29) (110,29) (111,29) (112,29) (113,29) (114,29) (115,29) (116,29) (117,29) (118,29) (119,29) (120,29) (121,29) (122,29) (123,29) (124,29) (125,29) (126,29) (127,29) (128,29) (129,29) (130,29) (131,29) (132,29) (133,29) (134,29) (135,29) (136,29) (137,29) (138,29) (139,29) (140,29) (141,29) (142,29) (143,29) (144,29) (145,29) (146,29) (147,29) (148,29) (149,29) (150,29) (151,29) (152,29) (153,29) (154,29) (155,29) (156,29) (157,29) (158,29) (159,29) (160,29) (161,29) (162,64) (163,64) (164,64) (165,64) (166,64) (167,64) (168,64) (169,64) (170,64) (171,64) (172,64) (173,64) (174,64) (175,64) (176,64) (177,64) (178,64) (179,64) (180,64) (181,64) (182,64) (183,64) (184,64) (185,64) (186,64) (187,64) (188,64) (189,64) (190,64) (191,64) (192,64) (193,64) (194,64) (195,64) (196,64) (197,64) (198,64) (199,64) (200,64) (201,64) (202,64) (203,64) (204,64) (205,64) (206,64) (207,64) (208,64) (209,64) (210,64) (211,64) (212,64) (213,64) (214,64) (215,64) (216,64) (217,64) (218,99) (219,99) (220,99) (221,99) (222,99) (223,99) (224,99) (225,99) (226,99) (227,99) (228,99) (229,99) (230,99) (231,99) (232,99) (233,99) (234,99) (235,99) (236,99) (237,99) (238,99) (239,99) (240,99) (241,99) (242,99) (243,99) (244,99) (245,99) (246,120) (247,120) (248,120) (249,120) (250,120) (251,120) (252,120) (253,120) (254,127) ]{};
\[cor:\[m(m+1)/2,m\]\] Let $m\ge 5$ be an integer and $\alpha$ a primitive element of ${{\mathrm{GF}}}(2^m)$. Let ${{\mathcal{C}}}$ be the set given by $$\begin{aligned}
{{\mathcal{C}}}=\left \{ \left ( {{\mathrm{Tr}}}^m_1 \left (a \alpha^{0} \right ), \cdots, {{\mathrm{Tr}}}^m_1 \left (a \alpha^{m(m+1)/2-1} \right ) \right ): a \in {{\mathrm{GF}}}(2^m) \right \}.\end{aligned}$$ Then ${{\mathcal{C}}}$ is a binary minimal code of dimension $m$ and length $m(m+1)/2$.
According to our best knowledge there is only one known example of binary minimal codes with dimension $m$ and length $m(m+1)/2$ , which was introduced in [@ZYW19]. These minimal codes can be described as follows: $$\begin{aligned}
\label{eq:P and P+P}
\left \{
\begin{array}{c}
\left ( {{\mathrm{Tr}}}^m_1(a \alpha_1), \cdots, {{\mathrm{Tr}}}^m_1(a \alpha_m), \right . \\
\left . {{\mathrm{Tr}}}^m_1(a (\alpha_1+\alpha_2)), \cdots, {{\mathrm{Tr}}}^m_1(a (\alpha_{m-1}+\alpha_m)) \right )
\end{array}
: a \in {{\mathrm{GF}}}(2^m) \right \},\end{aligned}$$ where $\alpha_1, \cdots, \alpha_m$ form a basis of ${{\mathrm{GF}}}(2^m)$ over ${{\mathrm{GF}}}(2)$. Obviously, the codes obtained in (\[eq:P and P+P\]) are unique, up to equivalence. Their minimum distance $d(m)$ is equal to $m$. Many infinite families of minimal codes of dimension $m$ and length $m(m+1)/2$ can be produced form Corollary \[cor:\[m(m+1)/2,m\]\]. Denote by $d_{max}(m)$ and $d_{min}(m)$ the largest and smallest values of the minimum distances of the codes of Corollary \[cor:\[m(m+1)/2,m\]\], respectively. Figure \[fig:minimum distance\] shows that the minimum distance of the minimal code form Corollary \[cor:\[m(m+1)/2,m\]\] would be better than that of the code given in (\[eq:P and P+P\]), and suggests the following conjecture.
coordinates [ (5,5) (6,7) (7,11) (8,13) (9,16) (10,21) (11,25) (12,29) (13,34) (14,39) (15,44) (16,50) ]{};
coordinates[ (5,5) (6,6) (7,9) (8,11) (9,13) (10,14) (11,15) (12,23) (13,24) (14,27) (15,27) (16,35) ]{};
coordinates[ (5,5) (6,6) (7,7) (8,8) (9,9) (10,10) (11,11) (12,12) (13,13) (14,14) (15,15) (16,16) ]{};
Let $m\ge 7$ be an integer. Then the minimal code in Corollary \[cor:\[m(m+1)/2,m\]\] has minimum distance greater than $m$. Moreover, if two primitive elements $\alpha$ and $\alpha'$ of ${{\mathrm{GF}}}(2^m)$ satisfy $\alpha' \neq \alpha^{2^i}$ for any $0 \le i \le m-1$, then the two codes corresponding to $\alpha$ and $\alpha'$ are inequivalent.
By the definition of shortened codes, Proposition \[prop:prm-generator-alpha\] shows that the generator polynomial of the shortened second-order Reed-Muller code ${{\mathrm{SRM}}}(2,m)$ can be expressed as $$\begin{aligned}
g_{2, \alpha}^*(X)= \prod_{ \scriptsize{\begin{array}{c}i_{m-1}+\cdots+ i_0 \le m-3
\\ i_{m-1}, \cdots, i_0 \in \{0,1\} \end{array}}} (X-\alpha^{i_{m-1}2^{m-1}+ \cdots+i_0 2^0}).
\end{aligned}$$ For a positive integer $\epsilon$, let $\mathfrak C_{\epsilon}$ be the code contained in the shortened second-order Reed-Muller code ${{\mathrm{SRM}}}(2,m)$ given by $$\begin{aligned}
\label{eq:modpoly-linear}
\mathfrak C_{\epsilon} =\left \{ \left ( \sum_{i=0}^{\epsilon-1} c_i X^i \right ) g_{2, \alpha}^*(X) : c_i \in {{\mathrm{GF}}}(2) \right \}.\end{aligned}$$
The following theorem provides a way of constructing linear codes of dimension $m$ and length less than $m(m+1)/2$ via sets of algebraic immunity $2$.
\[thm:C-\[0;m(m+1)/2\]\] Let $\epsilon$ be an integer with $1\le \epsilon < m(m-1)/2$. Let ${{\mathcal{C}}}$ be the set given by $$\begin{aligned}
{{\mathcal{C}}}=\left \{ \left ( {{\mathrm{Tr}}}^m_1 \left (a \alpha^{0} \right ), \cdots, {{\mathrm{Tr}}}^m_1 \left (a \alpha^{m(m+1)/2-\epsilon-1} \right ) \right ): a \in {{\mathrm{GF}}}(2^m) \right \}.\end{aligned}$$ Then ${{\mathcal{C}}}$ is a minimal code if and only if the minimum distance of the code $\mathfrak C_{\epsilon}$ in (\[eq:modpoly-linear\]) is greater than $2^{m-2}$.
Since $1\le \epsilon < m(m-1)/2$, it follows immediately that $\mathrm{dim}({{\mathcal{C}}})=m$.
Let us first prove the sufficient condition for ${{\mathcal{C}}}$ to be a minimal code. Suppose the assertion of the theorem is false. There would be a codeword $\left ( \sum_{i=0}^{\epsilon -1} c_iX^i \right ) g^*_{2,\alpha}(X) $ of weight $2^{m-2}$ of $ \mathfrak C_{\epsilon}$ , where $c_i \in {{\mathrm{GF}}}(2)$. Set $g(X)=X^{m(m+1)/2- \epsilon}\left ( \sum_{i=0}^{\epsilon -1} c_iX^i \right ) g^*_{2,\alpha}(X)$ and let $f\in \mathbb B_{m}^0$ be the Boolean function corresponding to $g(X)$. Then $g(X)$ is a codeword of weight $2^{m-2}$ of ${{\mathrm{SRM}}}(2,m)$ and $f$ satisfies $$\begin{aligned}
\label{eq:f-annihilator}
f(\alpha^i)=0,\end{aligned}$$ where $0 \le i \le m(m+1)/2- \epsilon-1$. Lemma \[lem:minimum-set-geometry\] now leads to $f(x)={{\mathrm{Tr}}}^m_1(a_1x) {{\mathrm{Tr}}}^m_1(a_2x)$, where $a_1\neq a_2 \in {{\mathrm{GF}}}(2^m)^*$. Let $\mathbf{c}_i$ be the nonzero codeword of ${{\mathcal{C}}}$ given by $\left ( {{\mathrm{Tr}}}^m_1(a_i \alpha^{i}) \right )_{i=0}^{m(m+1)/2- \epsilon-1}$, where $i=1 \text{ or } 2$. From (\[eq:f-annihilator\]), it is a simple matter to check that ${{\mathrm{Supp}}}(\mathbf{c}_1+ \mathbf{c}_2)= {{\mathrm{Supp}}}(\mathbf{c}_1) \dot \cup {{\mathrm{Supp}}}( \mathbf{c}_2) $, which contradicts the assumption that ${{\mathcal{C}}}$ is a minimal code. Consequently, the minimum distance of $\mathfrak C_{\epsilon}$ is greater than $2^{m-2}$.
Conversely, let $\mathfrak C_{\epsilon}$ be a linear code of minimum distance greater than $2^{m-2}$. Suppose that ${{\mathcal{C}}}$ is not a minimal code. Then we could find two distinct nonzero codewords $\mathbf{c_0}$ and $\mathbf c_1$ of ${{\mathcal{C}}}$ such that ${{\mathrm{Supp}}}( \mathbf{c}_1) \subsetneq {{\mathrm{Supp}}}(\mathbf{c}_0)$, where $\mathbf c_i= \left ( {{\mathrm{Tr}}}^m_1(a_i \alpha^i) \right )_{i=0}^{m(m+1)/2-\epsilon-1}$. Let $f$ be the quadratic Boolean function ${{\mathrm{Tr}}}^m_1( a_1x) {{\mathrm{Tr}}}^m_1(a_2 x)$, where $a_2=a_0+a_1$. A trivial verification shows that $$\begin{aligned}
\label{eq:f-0-sequen=0}
f(\alpha^0)=\cdots =f(\alpha^{m(m+1)/2-\epsilon-1})=0,\end{aligned}$$ and $$\begin{aligned}
\label{eq:wt(f)=q/4}
{{\mathrm{wt}}}(f) =2^{m-2}.\end{aligned}$$ Since $g^*_{2,\alpha}(X)$ is the generator polynomial of ${{\mathrm{SRM}}}(2,m)$, the codeword $\left ( f(P_i) \right )_{i=1}^{2^m-1}$ can be uniquely expressed as $\left (\sum_{i=0}^{m(m+1)/2-1} c_iX^i \right ) g^*_{2,\alpha}(X)$, where $c_i\in {{\mathrm{GF}}}(2)$. From (\[eq:f-0-sequen=0\]) it may be concluded that $c_i=0$ for any $0 \le i \le m(m+1)/2-\epsilon-1$. Combining this with (\[eq:wt(f)=q/4\]) we deduce that $${{\mathrm{wt}}}\left (\left (\sum_{i=0}^{\epsilon-1} c_{i+m(m+1)/2-\epsilon}X^i \right ) g^*_{2,\alpha}(X) \right )=2^{m-2}.$$ This contradicts our assumption about the minimum distance of $\mathfrak C_{\epsilon}$. It completes the proof.
Let $\alpha$ be any primitive element of ${{\mathrm{GF}}}(2^m)$. Let us denote by $\epsilon_m(\alpha)$ the maximum $\epsilon$ such that the code ${{\mathcal{C}}}$ in Theorem \[thm:C-\[0;m(m+1)/2\]\] is an $m$-dimensional minimal code. The values of $\epsilon_m(\alpha)$ and their corresponding frequencies are listed in Table \[tab:defect\] for $5 \le m \le 9$. It shows that a large number of minimal codes with dimension $m$ and length less than $m(m+1)/2$ can be produced from Theorem \[thm:C-\[0;m(m+1)/2\]\].
[>lcc|lcc]{} $m$ & $\epsilon_{m}(\alpha)$ & Freq. & $m$ & $\epsilon_{m}(\alpha)$ & Freq.\
5 & 0 & 10 & 9 & 1 & 36\
5 & 1 & 20& 9 & 5 & 18\
6 & 0 & 12 & 9 & 6 & 18\
6 & 1 & 12 & 9 & 7 & 18\
6 & 3 & 12& 9 & 8 & 36\
7 & 0 & 42 & 9 & 9 & 36\
7 & 3 & 28 & 9 & 10 & 72\
7 & 4 & 42 & 9& 11 & 36\
7 & 5 & 14 & 9 & 12 &54\
8 & 6 & 48 & 9 & 13 & 18\
8 & 7 & 16 & 9 & 14 & 54\
8 & 10 & 64 & 9 & 15 &18\
9 & 0 & 18\
As a corollary of Theorem \[thm:C-\[0;m(m+1)/2\]\], we have the following.
\[cor:C-\[0;m(m+1)/2-2\]\] Let ${{\mathcal{C}}}$ be the set given by $$\begin{aligned}
{{\mathcal{C}}}=\left \{ \left ( {{\mathrm{Tr}}}^m_1 \left (a \alpha^{0} \right ), \cdots, {{\mathrm{Tr}}}^m_1 \left (a \alpha^{m(m+1)/2-2} \right ) \right ): a \in {{\mathrm{GF}}}(2^m) \right \}.\end{aligned}$$ Then ${{\mathcal{C}}}$ is minimal if and only if the Hamming weight of the generator polynomial $g_{2, \alpha}^*(X)$ of the shortened second-order Reed-Muller code ${{\mathrm{SRM}}}(2, m)$ is not equal to $2^{m-2}$.
\[cor:C-\[0;m(m+1)/2-3\]\] Let ${{\mathcal{C}}}$ be the set given by $$\begin{aligned}
{{\mathcal{C}}}=\left \{ \left ( {{\mathrm{Tr}}}^m_1 \left (a \alpha^{0} \right ), \cdots, {{\mathrm{Tr}}}^m_1 \left (a \alpha^{m(m+1)/2-3} \right ) \right ): a \in {{\mathrm{GF}}}(2^m) \right \}.\end{aligned}$$ Then ${{\mathcal{C}}}$ is a minimal code if and only if both ${{\mathrm{wt}}}\left( g_{2, \alpha}^*(X) \right )$ and ${{\mathrm{wt}}}\left( (1+X) g_{2, \alpha}^*(X) \right )$ are greater than $2^{m-2}$.
Let $q=2^5$ and $\alpha$ be a primitive element with minimal polynomial $ \alpha^5 + \alpha^3 + 1=0$. Then $g^*_{2,\alpha}(X)=X^{16} + X^{12} + X^{11} + X^{10} + X^9 + X^4 + X + 1$. Clearly, ${{\mathrm{wt}}}\left ( g^*_{2,\alpha}(X) \right )=8$. The binary linear code ${{\mathcal{C}}}$ in Corollary \[cor:\[m(m+1)/2,m\]\] is a minimal code, but that in Corollary \[cor:C-\[0;m(m+1)/2-2\]\] is not.
Let $q=2^6$ and $\alpha$ be a primitive element with minimal polynomial $\alpha^6 + \alpha^5 + \alpha^3 + \alpha^2 + 1=0$. Then $g^*_{2,\alpha}(X)=X^{42} + X^{41} + X^{39} + X^{38} + X^{37} + X^{32} + X^{31} + X^{30} + X^{29} + X^{24} + X^{19} +
X^{17} + X^{16} + X^{13} + X^{12} + X^{11} + X^{10} + X^9 + X^8 + X^7 + X^3 + X^2 + X +
1$ and $(1+X)g^*_{2,\alpha}(X)=X^{43} + X^{41} + X^{40} + X^{37} + X^{33} + X^{29} + X^{25} + X^{24} + X^{20} + X^{19} + X^{18} +
X^{16} + X^{14} + X^7 + X^4 + 1$. Clearly, ${{\mathrm{wt}}}\left ( g^*_{2,\alpha}(X) \right )=24$ and ${{\mathrm{wt}}}\left ( (1+X) g^*_{2,\alpha}(X) \right )=16$. The binary linear code ${{\mathcal{C}}}$ in Corollary \[cor:C-\[0;m(m+1)/2-2\]\] is a minimal code, but that in Corollary \[cor:C-\[0;m(m+1)/2-3\]\] is not.
Let $q=2^6$ and $\alpha$ be a primitive element with minimal polynomial $\alpha^6 + \alpha^5 + \alpha^4 + \alpha + 1=0$. Then $g^*_{2,\alpha}(X)=X^{42} + X^{41} + X^{39} + X^{37} + X^{36} + X^{35} + X^{33} + X^{30} + X^{27} + X^{26} + X^{24} +
X^{21} + X^{19} + X^{17} + X^{16} + X^{15} + X^{13} + X^{9} + X^{8} + X^7 + X^5 + X^4 + X^3
+ 1$ and $(1+X)g^*_{2,\alpha}(X)=X^{43} + X^{41} + X^{40} + X^{39} + X^{38} + X^{35} + X^{34} + X^{33} + X^{31} + X^{30} + X^{28} +
X^{26} + X^{25} + X^{24} + X^{22} + X^{21} + X^{20} + X^{19} + X^{18 }+ X^{15} + X^{14} + X^{13} +
X^{10} + X^7 + X^6 + X^3 + X + 1$. Clearly, ${{\mathrm{wt}}}\left ( g^*_{2,\alpha}(X) \right )=24$ and ${{\mathrm{wt}}}\left ( (1+X) g^*_{2,\alpha}(X) \right )=28$. Both the binary linear codes in Corollary \[cor:C-\[0;m(m+1)/2-2\]\] and \[cor:C-\[0;m(m+1)/2-3\]\] are minimal codes.
It would be interesting to know how the Hamming weight of the polynomial $ g_{2, \alpha}^*(X)$ would be affected by selecting $\alpha$. Based on our numerical experiments, we pose the following conjecture and open problem.
For any integer $m\ge 5$ , there exists a primitive element $\alpha$ of ${{\mathrm{GF}}}(2^m)$ such that the Hamming weight of the generator polynomial $ g_{2, \alpha}^*(X)$ of the shortened second-order Reed-Muller ${{\mathrm{SRM}}}(2,m)$ is greater than $2^{m-2}$.
Are there infinitely many positive integers $m$ such that ${{\mathrm{wt}}}\left ( g_{2, \alpha}^*(X) \right ) > 2^{m-2}$ for any primitive element $\alpha$ of ${{\mathrm{GF}}}(2^m)$?
Let $\left ( f(P_i) \right )_{i=0}^{q-1}$ be any codeword of the second-order Reed-Muller code ${{\mathrm{RM}}}(2,m)$. Then the corresponding Boolean function $f$ can be uniquely expressed as $$\begin{aligned}
\label{eq:rm2-trance}
f(x)=\left \{
\begin{array}{cl}
\begin{array}{l}{{\mathrm{Tr}}}^{m/2}_1\left (a_{m/2} x^{2^{m/2}+1} \right )+ \\
\sum \limits_{i=1}^{ \frac{m-2}{2} } {{\mathrm{Tr}}}^m_1\left (a_i x^{2^i+1} \right )+{{\mathrm{Tr}}}^m_1(a_0 x) +c\end{array}, & \text{ if $m$ is even}, \\
& \\
\sum \limits_{i=1}^{ \frac{m-1}{2}} {{\mathrm{Tr}}}^m_1\left (a_i x^{2^i+1} \right )+{{\mathrm{Tr}}}^m_1(a_0 x) +c, & \text{ if $m$ is odd},
\end{array}
\right .\end{aligned}$$ where $a_{m/2} \in {{\mathrm{GF}}}(2^{m/2})$, $c \in {{\mathrm{GF}}}(2)$ and $a_i\in {{\mathrm{GF}}}(2^m)$ for $0\le i \le \lfloor (m-1)/ 2 \rfloor$. Berlekamp and Sloane [@Berlekamp69] have shown that all possible weights of codewords of ${{\mathrm{RM}}}(2,m)$ are of the forms $2^{m-1}$ and $2^{m-1}\pm 2^{m-1-j}$, where $0\le j \le \lfloor m /2 \rfloor$. A compact formula of the weight distribution of ${{\mathrm{RM}}}(2,m)$ can be found in [@Li19].
Next we shall present several classes of minimal codes contained in the shortened second-order Reed-Muller code ${{\mathrm{SRM}}}(2,m)$.
\[thm:quadratic functions space-odd\] Let $m\ge 3$ be an odd integer. Let ${{\mathcal{C}}}$ be the cyclic subcode of the shortened second order Reed-Muller code ${{\mathrm{SRM}}}(2,m)$ given by $$\begin{aligned}
\left \{ \left ( \sum_{i=1}^{( m-1)/2 } {{\mathrm{Tr}}}^{m}_1 \left (a_i \alpha^{(2^i+1)j} \right ) \right )_{j=0 }^{ 2^m-2}: a_i \in {{\mathrm{GF}}}(2^m) \right \}.\end{aligned}$$ Then ${{\mathcal{C}}}$ is a minimal code with parameters $\left [ 2^m-1, m(m-1)/2, \ge 3\cdot 2^{m-3} \right ]$.
Let us first prove that there is no codeword of weight $2^{m-2}$ or $3\cdot 2^{m-2}$. If there existed a codeword $\mathbf{c} \in {{\mathcal{C}}}$ such that ${{\mathrm{wt}}}(\mathbf{c})=2^{m-2}$, by Lemma \[lem:minimum-set-geometry\] there would be two distinct elements $a, b \in {{\mathrm{GF}}}(2^m)^*$ such that $\mathbf{c}= \left ( {{\mathrm{Tr}}}^m_1(a \alpha^j) {{\mathrm{Tr}}}^m_1(b \alpha^j) \right )_{j=0}^{2^m-2}$. There is no loss of generality in assuming $a=1$ and $b\in {{\mathrm{GF}}}(2^m)\setminus {{\mathrm{GF}}}(2)$ as ${{\mathcal{C}}}$ is a cyclic code. A direct calculation shows $$\begin{aligned}
\lefteqn{ {{\mathrm{Tr}}}^m_1( \alpha^j) {{\mathrm{Tr}}}^m_1 (b \alpha^j) } \\
&=& {{\mathrm{Tr}}}^m_1 \left(b \alpha^j {{\mathrm{Tr}}}^m_1( \alpha^j) \right )\\
&=& \sum^{m-1}_{i=0} {{\mathrm{Tr}}}^m_1 \left (b \alpha^{(2^i+1)j} \right )\\
&=& \sum^{m-1}_{i=1} {{\mathrm{Tr}}}^m_1 \left (b \alpha^{(2^i+1)j} \right )+ {{\mathrm{Tr}}}^m_1 \left (\sqrt{b} \alpha^{j} \right ), \end{aligned}$$ which is impossible. Thus there is no codeword of weight $2^{m-2}$. Suppose there was a codeword $\mathbf{c} \in {{\mathcal{C}}}$ of weight $3\cdot 2^{m-2}$, then the codeword $\mathbf{1}+ \mathbf{c}$ of ${{\mathrm{RM}}}(2,m)$ has weight $2^{m-2}$. Proposition \[prop:mini-geometry\] now implies $\mathbf{1}+ \mathbf{c}= \left ((1+ {{\mathrm{Tr}}}^m_1(a \alpha^j))(1+ {{\mathrm{Tr}}}^m_1(b \alpha^j)) \right )_{j=0}^{2^m-2}$, where $a\neq b \in {{\mathrm{GF}}}(2^m)^*$. We can assume that $a=1$ and $b \not \in {{\mathrm{GF}}}(2)$ as in the previous discussion. Consequently, $\mathbf{c}$ is just the codeword given by the Boolean function ${{\mathrm{Tr}}}^m_1( x){{\mathrm{Tr}}}^m_1(b x)+{{\mathrm{Tr}}}^m_1((b+1) x)$, which can be rewritten as: $$\begin{aligned}
\begin{array}{rl}
& \sum^{m-1}_{i=1} {{\mathrm{Tr}}}^m_1 \left (bx^{2^i+1} \right )+ {{\mathrm{Tr}}}^m_1 \left ( (b+\sqrt{b}+1) x \right ).
\end{array}\end{aligned}$$ Since $m$ is an odd integer, we have $b+\sqrt{b}+1\neq 0$. This clearly forces $\mathbf{c} \not \in {{\mathcal{C}}}$, a contradiction. Hence the weight ${{\mathrm{wt}}}(\mathbf c) $ of any codeword $\mathbf{c}$ of ${{\mathcal{C}}}$ is not equal to $3\cdot 2^{m-2}$. Consequently, we conclude that for any nonzero codeword $\mathbf{c}$ of ${{\mathcal{C}}}$ its weight satisfies the following $$\begin{aligned}
3\cdot 2^{m-3} \le {{\mathrm{wt}}}(\mathbf c ) \le 5\cdot 2^{m-3}.\end{aligned}$$ Therefore ${{\mathcal{C}}}$ is a minimal code by Lemma \[lem:AB\] and has minimum distance at least $3\cdot 2^{m-3}$. It is obvious that $\mathrm{dim}({{\mathcal{C}}}) =m(m-1)/2$ from the definition of ${{\mathcal{C}}}$. This completes the proof.
The proof above gives more, namely the code ${{\mathcal{C}}}$ in Theorem \[thm:quadratic functions space-odd\] is not a minimal code if $m\ge 4$ is an even integer.
\[thm:quad+linear\] Let $m\ge 3$ be an integer. Let ${{\mathcal{C}}}$ be the cyclic subcode of the shortened second-order Reed-Muller code ${{\mathrm{SRM}}}(2,m)$ given by $$\begin{aligned}
\left \{ \left ( \sum_{i=2}^{\lfloor m /2 \rfloor } {{\mathrm{Tr}}}^{m}_1 \left (a_i \alpha^{(2^i+1)j}\right ) + {{\mathrm{Tr}}}^m_1(b \alpha^j) \right )_{j=0}^{2^{m}-2}: b, a_i \in {{\mathrm{GF}}}(2^m) \right \}.\end{aligned}$$ Then ${{\mathcal{C}}}$ is a minimal code with parameters $\left [ 2^m-1, m(m-1)/2, \ge 3\cdot 2^{m-3} \right ]$.
The statements will be proved once we prove that there are no codewords of weight $2^{m-2}$ or $3\cdot 2^{m-2}$. Suppose, contrary to our claim, that there exists a codeword with weight $2^{m-2}$ or $3\cdot 2^{m-2}$. By a similar argument in the proof of Theorem \[thm:quadratic functions space-odd\], we could find $b\in {{\mathrm{GF}}}(2^m) \setminus {{\mathrm{GF}}}(2)$ and $c \in {{\mathrm{GF}}}(2)$ such that the codeword of ${{\mathrm{SRM}}}(2,m)$ corresponding to the Boolean function $f(x)={{\mathrm{Tr}}}^m_1( x){{\mathrm{Tr}}}^m_1(b x)+c{{\mathrm{Tr}}}^m_1((b+1) x)$ lies in ${{\mathcal{C}}}$. A simple calculation yields $$\begin{aligned}
\lefteqn{ {{\mathrm{Tr}}}^m_1( x){{\mathrm{Tr}}}^m_1(b x)+c{{\mathrm{Tr}}}^m_1((b+1) x) } \\
&=& \sum_{i=0}^{m-1} {{\mathrm{Tr}}}^m_1(b x^{2^i+1}) +c{{\mathrm{Tr}}}^m_1((b+1) x)\\
&=& \sum_{i=2}^{m-2} {{\mathrm{Tr}}}^m_1(b x^{2^i+1}) +{{\mathrm{Tr}}}^m_1((\sqrt{b}+bc+c) x)\\
& &+ {{\mathrm{Tr}}}^{m}_1((b+b^2) x^{2+1}),\end{aligned}$$ which contradicts the definition of ${{\mathcal{C}}}$. This completes the proof.
In the spirit of Theorems \[thm:quadratic functions space-odd\] and \[thm:quad+linear\], we pose the following open problem.
Does there exist a minimal code ${{\mathcal{C}}}$ contained in the second-order Reed-Muller code ${{\mathrm{RM}}}(2,m)$ such that its dimension $\mathrm{dim}({{\mathcal{C}}})$ is greater than $m(m-1)/2$?
The following two theorems describe two infinite classes of minimal codes obtained by puncturing of the minimal codes in Theorems \[thm:quadratic functions space-odd\] and \[thm:quad+linear\].
Let $m\ge 5$ be an odd integer. Let $\delta $ be an integer with $\sum_{i=1}^{t-1} \binom{m}{i} \le \delta < \sum_{i=1}^{t} \binom{m}{i}$ and $5\le t \le m$. Let ${{\mathcal{C}}}$ be the binary code given by $$\begin{aligned}
\left \{ \left ( \sum_{i=1}^{ (m-1)/2 } {{\mathrm{Tr}}}^{m}_1 \left (a_i \alpha^{(2^i+1)j} \right ) \right )_{j=0 }^{ \delta-1}: a_i \in {{\mathrm{GF}}}(2^m) \right \}.\end{aligned}$$ Then ${{\mathcal{C}}}$ is a minimal code with parameters $\left [ \delta , m(m-1)/2, \ge \sum_{i=0}^{t-3} \binom{m-2}{i} \right ]$.
Combining Theorem \[thm:quadratic functions space-odd\] and Lemma \[lem:\[h;delta\]+0\] with Corollary \[cor:ad>2l\] proves the desired conclusion.
Let $m\ge 5$ be an integer. Let $\delta $ be an integer with $\sum_{i=1}^{t-1} \binom{m}{i} \le \delta < \sum_{i=1}^{t} \binom{m}{i}$ and $5\le t \le m$. Let ${{\mathcal{C}}}$ be the binary code given by $$\begin{aligned}
\left \{ \left ( \sum_{i=2}^{ \lfloor m/2 \rfloor } {{\mathrm{Tr}}}^{m}_1 \left (a_i \alpha^{(2^i+1)j} \right )+ {{\mathrm{Tr}}}^m_1 \left (a_0 \alpha^j \right ) \right )_{j=0 }^{\delta-1}: a_i \in {{\mathrm{GF}}}(2^m) \right \}.\end{aligned}$$ Then ${{\mathcal{C}}}$ is a minimal code with parameters $\left [ \delta , m(m-1)/2, \ge \sum_{i=0}^{t-3} \binom{m-2}{i} \right ]$.
Combining Theorem \[thm:quad+linear\] and Lemma \[lem:\[h;delta\]+0\] with Corollary \[cor:ad>2l\] yields the desired conclusion.
Minimal codes from vectorial Boolean functions with high algebraic immunity {#sec:codes-vec-func}
===========================================================================
In this section, we shall demonstrate that binary minimal codes can be obtained from the vector subspace spanned by certain subcodes of Reed-Muller and the component functions of vectorial Boolean functions with high algebraic immunity.
For a vectorial Boolean $(m,r)$-function $F=(f_1, \cdots, f_r)$ with ${{\mathrm{AI}}}(F)=t\ge 1$, let $\mathrm{Span}(F)$ be the linear code defined by $$\begin{aligned}
\mathrm{Span}(F)=\left \{ \left (\sum_{j=1}^{r} a_j f_j(P_i) \right)_{i=0}^{2^m-1}: a_j \in {{\mathrm{GF}}}(2) \right \}.\end{aligned}$$ Let ${{\mathcal{C}}}$ be a linear code. The sum of two linear subcodes ${{\mathcal{C}}}_1$ and ${{\mathcal{C}}}_2$ of ${{\mathcal{C}}}$ is the set, denoted ${{\mathcal{C}}}_1 + {{\mathcal{C}}}_2$, consisting of all the elements ${{\mathbf{c}}}_1+ {{\mathbf{c}}}_2$, where ${{\mathbf{c}}}_1 \in {{\mathcal{C}}}_1$ and ${{\mathbf{c}}}_2 \in {{\mathcal{C}}}_2$. If ${{\mathcal{C}}}_1 \cap {{\mathcal{C}}}_2 =\{\mathbf{0}\}$, then the sum is also called the direct sum of ${{\mathcal{C}}}_1$ and ${{\mathcal{C}}}_2$, and is written by ${{\mathcal{C}}}_1 \bigoplus {{\mathcal{C}}}_2 $. Note that the direct sum of linear subcodes of a linear code is not the same thing as the direct sum of some linear codes.
\[lem:vec-f-set\] Let $r\ge 2$ and $F$ be a vectorial Boolean $(m,r)$-function with ${{\mathrm{AI}}}(F)=t\ge 1$. Let $D$ be the subset of ${{\mathrm{GF}}}(2^m)$ given by $$D=\left \{ x\in {{\mathrm{GF}}}(2^m): v_1\cdot F(x)= \epsilon_1, v_2\cdot F(x)= \epsilon_2 \right \},$$ where $v_1$ and $ v_2 $ are two distinct nonzero elements in ${{\mathrm{GF}}}(2)^r$ and $\epsilon_1, \epsilon_2 \in {{\mathrm{GF}}}(2)$. Then ${{\mathrm{AI}}}(D) \ge t$.
By assumption, $v_1$ and $v_2$ are linearly independent over ${{\mathrm{GF}}}(2)$. It follows that there exists $y\in {{\mathrm{GF}}}(2)^r$ such that $v_1\cdot y =\epsilon_1$ and $v_2\cdot y =\epsilon_2$. We thus get $F^{-1}(y) \subseteq D$. By the definition of algebraic immunity, ${{\mathrm{AI}}}(D) \ge t$. This completes the proof.
Let $F$ be a vectorial Boolean $(m,r)$-function with ${{\mathrm{AI}}}(F)\ge t$. Then $\mathrm{Span}(F) \cap {{\mathrm{RM}}}(t-1,m)=\{\mathbf 0\}$ from Lemma \[lem:vec-f-set\]. The following theorem presents a new method to construct minimal codes from some subcodes of Reed-Muller codes and vectorial Boolean functions with high algebraic immunity.
\[thm:min-rm-Vecfun\] Let ${{\mathcal{C}}}$ be a $k$-dimensional subcode of the Reed-Muller code $\rm{RM} (\ell, m)$ such that $k>1$. Let $F$ be a vectorial Boolean $(m,r)$-function with algebraic immunity ${{\mathrm{AI}}}(F) \ge 2\ell+1$. Then ${{\mathcal{C}}}\bigoplus \mathrm{Span}(F)$ is a minimal code of dimension $k+r$.
It is clear that $\mathrm{dim}({{\mathcal{C}}}\bigoplus \mathrm{Span}(F))=k+r$. It remains to prove that ${{\mathcal{C}}}\bigoplus \mathrm{Span}(F)$ is a minimal code.
Let $\left ( v_1 \cdot F(P_i) +f_1(P_i) \right )_{i=0}^{2^m-1}$ and $\left ( v_2 \cdot F(P_i) +f_2(P_i) \right )_{i=0}^{2^m-1}$ be any two nonzero codewords of ${{\mathcal{C}}}\bigoplus \mathrm{Span}(F)$, where $f_1, f_2 \in \mathbb B_m$ are Boolean functions of algebraic degree at most $\ell$, and $v_1, v_2 \in {{\mathrm{GF}}}(2)^r$. We will complete the proof of the theorem if we prove the following: $$\begin{aligned}
\label{eq:c*c-f*f}
\left ( v_1 \cdot F(x) +f_1(x) \right ) \cdot \left ( v_2 \cdot F(x) +f_2(x) \right ) \not \equiv 0.
\end{aligned}$$ To this end, consider the following four cases.
Case 1: $\left ( v_1 \cdot F +f_1\right ) = \left ( v_2 \cdot F +f_2 \right ).$ If $f_1 \equiv 0$, then $v_1 \neq \mathbf{0}$. Applying Lemma \[lem:vec-f-set\], we see that ${{\mathrm{Supp}}}( v_1 \cdot F )$ is not the empty set, which gives (\[eq:c\*c-f\*f\]). Let $f_1 $ be a nonzero function. By the assumption of the theorem, $f_1$ does not vanish on $\left \{ x \in {{\mathrm{GF}}}(2^m): v_1 \cdot F(x)=0 \right \}$, and (\[eq:c\*c-f\*f\]) is proved.
Case 2: $f_1=f_2 \equiv 0$. In this case, none of $v_1$ and $ v_2$ is the zero vector. Lemma \[lem:vec-f-set\] now leads to ${{\mathrm{Supp}}}(v_1 \cdot F) \cap {{\mathrm{Supp}}}( v_2 \cdot F) \neq \emptyset$. Then it follows that the product of $v_1 \cdot F(x)$ and $v_2 \cdot F(x)$ is not the zero function, which establishes (\[eq:c\*c-f\*f\]).
Case 3: $f_1 \equiv 0$ and $f_2 \not \equiv 0$, or , $f_1 \not \equiv 0$ and $f_2 \equiv 0$. By symmetry, we can assume $f_1 \equiv 0$ and $f_2 \not \equiv 0$. Let $D$ be the subset of ${{\mathrm{GF}}}(2^m)$ given by $$\begin{aligned}
D=\left \{
\begin{array}{rl}
\{x\in {{\mathrm{GF}}}(2^m): v_1 \cdot F=1, v_2 \cdot F=1 \}, &\text{ if } v_1 =v_2,\\
& \\
\{x\in {{\mathrm{GF}}}(2^m): v_1 \cdot F=1, v_2 \cdot F=0 \}, &\text{ if } v_1 \neq v_2.
\end{array}
\right .
\end{aligned}$$ Note that $f_1\not \equiv 1$ since $\mathrm{dim}({{\mathcal{C}}}) >1$. Therefore none of $f_1$ and $(1+f_1)$ vanishes on $D$ from Lemma \[lem:vec-f-set\]. Thus (\[eq:c\*c-f\*f\]) holds.
Case 4: $f_1 \not \equiv 0$ and $f_2 \not \equiv 0$. Denote $D=\{x\in {{\mathrm{GF}}}(2^m): v_1 \cdot F=v_2 \cdot F=0\}$. It is obvious that $f_1f_2 \not \equiv 0$ because ${{\mathcal{C}}}$ is a minimal code. Combining Lemma \[lem:vec-f-set\] with $\mathrm{deg}(f_1f_2)\le 2\ell$ yields $f_1f_2 \not \in \mathrm{Ann} (D)$, which gives (\[eq:c\*c-f\*f\]).
Summarising the discussions in the four cases completes the proof of (\[eq:c\*c-f\*f\]) and the theorem.
In order to apply Theorem \[thm:min-rm-Vecfun\] to obtain minimal codes, we need to construct vectorial Boolean functions with hight algebraic immunity. The following result is in this direction.
\[thm:vecFUN-AI\] Let $n_0$, $n_1$, $\cdots$, $n_{2^r}$ be integers satisfying $0=n_0 < n_1 < \cdots <n_{2^r}=2^m-1$. Let $y_0, y_1, \cdots, y_{2^r-1}$ be an enumeration of the points of ${{\mathrm{GF}}}(2)^r$. Let $F$ be the function defined by $$\begin{aligned}
F(x)=\left \{
\begin{array}{rl}
y_i, & \text{ if }x\in [n_i;n_{i+1}-n_{i}]_{\alpha},\\
\\
y_0, & \text{ if } x=0.\\
\end{array}
\right .\end{aligned}$$ Then $F$ is a vectorial Boolean $(m,r)$-function with algebraic immunity $t$, where $t$ is the biggest integer $t$ such that $\sum_{j=1}^{t-1} \binom{m}{j} \le n_{1}-n_0$ and $\sum_{j=0}^{t-1} \binom{m}{j} \le n_{i+1}-n_i$ for $1\le i \le 2^r-1$.
The desired conclusion follows directly from Lemmas \[lem:ai-det-0\] and \[lem:\[h;delta\]+0\].
Note that the theorem is still true if the vector space ${{\mathrm{GF}}}(2)^r$ is replaced by the finite field ${{\mathrm{GF}}}(2^r)$. Using Theorem \[thm:vecFUN-AI\], we obtain the following explicit construction of vectorial functions with high algebraic immunity.
\[cor:vecFUN-AI\] Let $n_0$, $n_1$, $\cdots$, $n_{2^r}$ be integers satisfying $0=n_0 < n_1 < \cdots <n_{2^r}=2^m-1$. Let $F$ be the vectorial Boolean function defined by $F(0)=0$ and $F(\alpha^j)=(y_0, \cdots, y_{r-1})$, where $$\begin{aligned}
n_{ \sum_{i=0}^{r-1} y_{i} 2^{i}} \le j < n_{1+ \sum_{i=0}^{r-1} y_{i} 2^{i}} .\end{aligned}$$ Then $F$ is a vectorial Boolean $(m,r)$-function with algebraic immunity $t$, where $t$ is the biggest integer $t$ such that $\sum_{j=1}^{t-1} \binom{m}{j} \le n_{1}-n_0$ and $\sum_{j=0}^{t-1} \binom{m}{j} \le n_{i+1}-n_i$ for $1\le i \le 2^r-1$.
\[cor:simplex+vecfun\] Let $F$ be the vectorial Boolean $(m,r)$-function of Theorem \[thm:vecFUN-AI\] with $m^2+m+2\le 2^{m-r+1}$ and $n_1-n_0+1=n_2-n_1=\cdots =n_{2^r}-n_{2^r-1}=2^{m-r}$. Let ${{\mathcal{C}}}(F)$ be the binary code given by $$\begin{aligned}
{{\mathcal{C}}}(F)=
\left \{ \left ( v\cdot F(P_i) + {{\mathrm{Tr}}}^m_1 (b P_i) \right )_{i=1}^{2^m-1}:
v \in {{\mathrm{GF}}}(2)^r, b \in {{\mathrm{GF}}}(2^m) \right \}.\end{aligned}$$ Then ${{\mathcal{C}}}(F)$ is a minimal code of dimension $m+r$.
Combining Theorem \[thm:vecFUN-AI\] with Theorem \[thm:min-rm-Vecfun\] proves the desired conclusion.
Let $\delta $ be an integer with $\sum_{i=0}^{t-1} \binom{m}{i} \le \delta < \sum_{i=0}^{t} \binom{m}{i}$ and $3\le t \le m-3$. Let $f$ be the Boolean function with ${{\mathrm{Supp}}}(f)= [0;\delta]_{\alpha}$. Let ${{\mathcal{C}}}(f)$ be the binary code defined by $$\begin{aligned}
{{\mathcal{C}}}(f)=
\left \{ \left ( f(P_i) + {{\mathrm{Tr}}}^m_1 (b P_i) \right )_{i=1}^{2^m-1}:
b \in {{\mathrm{GF}}}(2^m) \right \}.\end{aligned}$$ Then ${{\mathcal{C}}}(f)$ is a minimal code of parameters $[2^m-1,m+1,\ge d]$ with $d$ being the smaller of $\delta$ and $$\begin{aligned}
\max \left \{\sum_{i=0}^{t-2} \binom{m-1}{i} +\sum_{i=0}^{t'-2} \binom{m-1}{i}, 2^{m-1}- 1 - \frac{\ln 2}{\pi} (m+1) \sqrt{2^m} \right \},\end{aligned}$$ where $t'=m-t$ when $\delta \neq \sum_{i=0}^{t-1} \binom{m}{i}$ and $t'=m-t+1$ when $\delta = \sum_{i=0}^{t-1} \binom{m}{i}$.
An easy computation shows that $$\sum_{i=0}^{m-t-1} \binom{m}{i} < 2^m-\delta \le \sum_{i=0}^{m-t} \binom{m}{i}.$$ Note that ${{\mathrm{Supp}}}(f+1) =2^m-\delta$. From Lemmas \[lem:ai-det-0\] and \[lem:\[h;delta\]+0\] we conclude that ${{\mathrm{AI}}}({{\mathrm{Supp}}}(f))=t$ and ${{\mathrm{AI}}}({{\mathrm{Supp}}}(f))=t'$. Theorem \[thm:min-rm-Vecfun\] now implies that ${{\mathcal{C}}}(f)$ is a minimal code of dimension $m+1$. We are left with the task of determining the lower bound of $d$. It is easy to see that for any $b\in {{\mathrm{GF}}}(2^m)^*$ the Hamming distance $\mathrm{dist}(f,{{\mathrm{Tr}}}^m_1(bx))$ of $f$ and ${{\mathrm{Tr}}}^m_1(bx)$ can be written as $$\begin{aligned}
\begin{array}{c}
\mathrm{dist}(f,{{\mathrm{Tr}}}^m_1(bx))={{\mathrm{wt}}}\left (f (1+{{\mathrm{Tr}}}^m_1(bx)) \right )+ {{\mathrm{wt}}}\left ((1+f) {{\mathrm{Tr}}}^m_1(bx) \right ).
\end{array}\end{aligned}$$ By Lemma \[lem:wt-bound\], we deduce that $$\begin{aligned}
\label{eq:t-t'}
\mathrm{dist}(f,{{\mathrm{Tr}}}^m_1(bx)) \ge \sum_{i=0}^{t-2} \binom{m-1}{i} +\sum_{i=0}^{t'-2} \binom{m-1}{i}.\end{aligned}$$
An easy computation yields that $$\begin{aligned}
\label{eq:dist-walsh-zhang}
\begin{array}{rl}
\mathrm{dist}(f,{{\mathrm{Tr}}}^m_1(bx)) = & 2^{m-1}+ \sum_{x\in \mathrm{Supp} (f)} (-1)^{{{\mathrm{Tr}}}^m_1 (bx)}\\
=& 2^{m-1} + \sum_{i=0}^{\delta -1} (-1)^{{{\mathrm{Tr}}}^m_1 (b \alpha^i)}\\
\ge & 2^{m-1}- 1 - \frac{\sqrt{2^m}}{\pi} \ln \left ( \frac{4(2^m-1)}{\pi} \right )\\
\ge & 2^{m-1}- 1 - \frac{\ln 2}{\pi} (m+1) \sqrt{2^m},
\end{array}\end{aligned}$$ where the first inequality follows from (\[eq:F-indpendentOFdelta\]). Combining (\[eq:t-t’\]) and (\[eq:dist-walsh-zhang\]) yields the desired conclusion.
Let $m$ be an odd integer. Let $F$ be the vectorial Boolean $(m,r)$-function of Theorem \[thm:vecFUN-AI\] with $\sum_{i=0}^{4} \binom{m}{j}\le 2^{m-r}$ and $n_1-n_0+1=n_2-n_1=\cdots =n_{2^r}-n_{2^r-1}=2^{m-r}$. Let ${{\mathcal{C}}}(F)$ be the binary code given by $$\begin{aligned}
{{\mathcal{C}}}(F)=\left \{
\begin{array}{r}
\left ( v\cdot F(P_i) + \sum_{j=1}^{(m-1)/2}{{\mathrm{Tr}}}^m_1 \left (b_j P_i^{2^j+1} \right ) \right )_{i=1}^{2^m-1}: \\
\\
v \in {{\mathrm{GF}}}(2)^r, b_j \in {{\mathrm{GF}}}(2^m)
\end{array}
\right \}.\end{aligned}$$ Then ${{\mathcal{C}}}(F)$ is a minimal code of dimension $\frac{m(m+1)}{2}+r$.
Combining Theorems \[thm:quadratic functions space-odd\] and \[thm:vecFUN-AI\] with Theorem \[thm:min-rm-Vecfun\] proves the desired conclusion.
Let $m$ be a positive integer. Let $F$ be the vectorial Boolean $(m,r)$-function of Theorem \[thm:vecFUN-AI\] with $\sum_{i=0}^{4} \binom{m}{j}\le 2^{m-r}$ and $n_1-n_0+1=n_2-n_1=\cdots =n_{2^r}-n_{2^r-1}=2^{m-r}$. Let ${{\mathcal{C}}}(F)$ be the binary code given by $$\begin{aligned}
{{\mathcal{C}}}(F)=\left \{
\begin{array}{r}
\left ( v\cdot F(P_i) + \sum_{j=2}^{\lfloor m/2 \rfloor}{{\mathrm{Tr}}}^m_1 \left (b_j P_i^{2^j+1} \right ) +{{\mathrm{Tr}}}^m_1(c P_i) \right )_{i=1}^{2^m-1}: \\
\\
v \in {{\mathrm{GF}}}(2)^r, b_j, c \in {{\mathrm{GF}}}(2^m)
\end{array}
\right \}.\end{aligned}$$ Then ${{\mathcal{C}}}(F)$ is a minimal code of dimension $\frac{m(m+1)}{2}+r$.
Combining Theorems \[thm:quad+linear\] and \[thm:vecFUN-AI\] with Theorem \[thm:min-rm-Vecfun\] yields the desired conclusion.
Let $q=2^7$ and $\alpha$ be a primitive element with minimal polynomial $ \alpha^7 + \alpha + 1=0$. Let $f_1$ be the Boolean function with ${{\mathrm{Supp}}}(f_1)= [63;64]_{\alpha}$ and let $f_2$ be the function with ${{\mathrm{Supp}}}(f_2)=[31;32]_{\alpha} \cup [95;32]_{\alpha} $. Let $F$ denote the vectorial Boolean function $(f_1, f_2)$. Then the algebraic immunity of $F$ equals $3$ and the binary linear code ${{\mathcal{C}}}(F)$ in Corollary \[cor:simplex+vecfun\] is a minimal code with parameters $[127,9,52]$.
Summary and concluding remarks {#sec:conc}
==============================
In this paper, a link between minimal linear codes and subsets of finite fields without nonzero low degree annihilators was established. This link allowed us to construct binary minimal codes with special sets, Boolean functions or vectorial Boolean with high algebraic immunity. A general construction of minimal binary linear codes from sets without nonzero low degree annihilators was proposed. Employing this general construction, minimal codes of dimension $m$ and length less than or equal to $m(m-1)/2$ were obtained, and a lower bounder on the minimum distance of the proposed minimal codes was established. A explicit construction of minimal codes using certain subcodes of Reed-Muller codes and vectorial Boolean functions with algebraic immunity was also developed. These results show that there are natural connections among binary minimal codes, sets without nonzero low degree annihilators and Boolean functions with high algebraic immunity.
The results of this paper were presented in terms of univariate representations of functions and codes. The corresponding multivariate analogies can be easily worked out. It would be interesting to generalize the results of this paper to the nonbinary cases. It would be good if the open problems and conjectures proposed in this paper could be settled. The reader is cordially invited to join this adventure.
Acknowledgements {#acknowledgements .unnumbered}
================
C. Ding’s research was supported by the Hong Kong Research Grants Council, Proj. No. 16300919. S. Mesnager was supported by the ANR CHIST-ERA project SECODE. C. Tang was supported by National Natural Science Foundation of China (Grant No. 11871058) and China West Normal University (14E013, CXTD2014-4 and the Meritocracy Research Funds).
[99]{} E. F. Assmus Jr, “On the Reed-Muller codes," Discrete Math., vol. 106, no. 107, pp. 25–33, 1992.
E. F. Assmus Jr, J. D. Key, “Polynomial codes and finite geometries," in *Handbook of Coding Theory*, pp. 1269-1343, Elsevier, Amsterdam, 1998.
A. Ashikhmin, A. Barg, “Minimal vectors in linear codes," *IEEE Trans. Inf. Theory*, vol. 44, no. 5, pp. 2010–2017, Sept. 1998. G. N. Alfarano, M. Borello, A. Neri, “A geometric characterization of minimal codes and their asymptotic performance," arXiv:1911.11738, 2019.
E. R. Berlekamp, N. J. A. Sloane, “Restrictions on weight distribution of Reed-Muller codes," *Information and Control*, vol. 14, no. 5, pp. 442–456, 1969.
G. R. Blakley, “Safeguarding cryptographic keys," In *Proceddings of the 1979 International Workshop on Managing Requirements Knowledge (MARK)*, pp. 313–318, IEEE, 1979.
D. Bartoli, M. Bonini, “Minimal linear codes in odd characteristic," *IEEE Trans. Inf. Theory*, vol. 65, no. 7, pp. 4152–4155, 2019.
M. Bonini, M. Borello, “Minimal linear codes arising from blocking sets," *J. Algebr. Comb.*, DOI: 10.1007/s10801-019-00930-6, 2020.
C. Carlet, “On the higher order nonlinearities of Boolean functions and S-boxes, and their generalizations," In *Proceedings of the International Conference on Sequences and Their Applications*, pp. 345–367, Springer, Berlin, Heidelberg, 2008.
C. Carlet, K. Feng, “An infinite class of balanced vectorial Boolean functions with optimum algebraic immunity and good nonlinearity," In *Proceedings of the International Conference on Coding and Cryptology*, pp. 1–11, Springer, Berlin, Heidelberg, 2009.
C. Carlet, “Vectorial Boolean functions for cryptography," In *Boolean Models and Methods in Mathematics, Computer Science, and Engineering*, vol. 134, pp. 398–469, 2010.
D. K. Dalai, K. C. Gupta, S. Maitra, “Notion of algebraic immunity and its evaluation related to fast algebraic attacks," In *Proceedings of the International Workshop on Boolean Functions: Cryptography and Applications*, pp. 13–15, Rouen, France, March, 2006.
K. Ding, C. Ding, “A class of two-weight and three-weight codes and their applications in secret sharing," *IEEE Trans. Inf. Theory*, vol. 61, no. 11, pp. 5835–5842, 2015.
C. Ding, J. Yuan, Covering and secret sharing with linear codes," in *Discrete Mathematics and Theoretical Computer Science*, Lecture Notes in Computer Science 2731, 2003, Springer Verlag, pp. 11–25.
C. Ding, Z. Heng, Z. Zhou, “Minimal binary linear codes," *IEEE Trans. Inf. Theory*, vol. 64, no. 10, pp. 6536–6545, 2018.
C. R. Hartmann, C. R. Tzeng, “Generalizations of the BCH bound," *Information and Control*, vol. 20, no. 5, pp. 489–498, 1972.
Z. Heng, C. Ding, Z. Zhou, “Minimal linear codes over finite fields," *Finite Fields Appl.*, vol. 54, pp. 176–196, 2018.
K. F. Ireland, M. I. Rosen, *Elements of Number Theory: Including an Introduction to Equations over Finite Fields*, Bogden & Quigley, 1972.
X. Li and Q. Yue, “Four classes of minimal binary linear codes with $W_{\min}/W_{\max}<1/2$ derived from Boolean functions," *Des. Codes Cryptogr.*, vol. 88, no. 2, pp. 257–271, 2020.
W. Lu, X. Wu, X. Cao, “The parameters of minimal linear codes," arXiv:1911.07648, 2019.
R. J. McEliece and D. V. Sarwate, “On sharing secrets and Reed-Solomon codes," *Communications of the ACM*, vol. 24, pp. 583–584, 1981.
J. L. Massey, “Minimal codewords and secret sharing," In *Proc. 6th Joint Swedish-Russian Int. Workshop on Info. Theory*, Sweden, pp. 276–279, 1993.
J. L. Massey, “Some applications of coding theory in cryptography," In *Codes and Cyphers: Cryptography and Coding IV*, Esses, England, pp. 33–47, 1995.
F. J. MacWilliams, N. J. A. Sloane, *The Theory of Error-Correcting Codes*, North-Holland, Amsterdam, 1977.
S. Mesnager, “Improving the lower bound on the higher order nonlinearity of Boolean functions with prescribed algebraic immunity," *IEEE Trans. Inf. Theory*, vol. 54, no. 8, pp. 3656–3662, 2008.
S. Mesnager, “Linear codes with few weights from weakly regular bent functions based on a generic construction," *Cryptogr. Commun.*, vol. 9, no. 1, pp. 71-84, 2017.
S. Mesnager, F. Özbudak, A. S[i]{}nak, “A new class of three-weight linear codes from weakly regular plateaued functions," In *Proceedings of the Tenth International Workshop on Coding and Cryptography (WCC)*, 2017.
S. Mesnager, F. Özbudak, A. S[i]{}nak, “Linear codes from weakly regular plateaued functions and their secret sharing schemes," *Des., Codes Cryptograph.*, vol. 87, no. 2, pp. 463–480, 2019.
S. Mesnager, Y. Qi, H. Ru, C. Tang, “Minimal linear codes from characteristic functions," arXiv:1908.01650, 2019.
S. Li, “On the weight distribution of second order Reed-Muller codes and their relatives," *Des., Codes Cryptograph.*, vol. 87, no. 10, pp. 2447–2460, 2019.
A. Shamir, “How to share a secret," *Communications of the ACM*, vol. 22, no. 11, pp. 612–613, 1979.
C. Tang, N. Li, Y. Qi, Z. Zhou, T. Helleseth, “Linear codes with two or three weights from weakly regular bent functions," *IEEE Trans. Inf. Theory*, vol. 62, no. 3, pp. 1166–1176, 2016.
C. Tang, Y. Qiu, Q. Liao, Z. Zhou, “Full characterization of minimal linear codes as cutting blocking sets," arXiv:1911.09867, 2019.
G. Xu, L. Qu, “Three classes of minimal linear codes over the finite fields of odd characteristic," *IEEE Trans. Inf. Theory*, vol. 65, no. 11, pp. 7067–7078, 2019.
W. Zhang, H. Yan, H. Wei, “Four families of minimal binary linear codes with $ w_ {\min}/w_ {\max}\le 1/2$," *Applicable Algebra in Engineering, Communication and Computing*, vol. 30, no. 2, pp. 175–184, 2019.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Weyl type-II fermions are massless quasiparticles that obey the Weyl equation and which are predicted to occur at the boundary between electron- and hole-pockets in certain semi-metals, i.e. the (W,Mo)(Te,P)$_2$ compounds. Here, we present a study of the Fermi-surface of WP$_2$ *via* the Shubnikov-de Haas (SdH) effect. Compared to other semi-metals WP$_2$ exhibits a very low residual resistivity, i.e. $\rho_0 \simeq 10$ n$\Omega$cm, which leads to perhaps the largest non-saturating magneto-resistivity $(\rho(H))$ reported for any compound. For the samples displaying the smallest $\rho_0$, $\rho(H)$ is observed to increase by a factor of $2.5 \times 10^{7}$ $\%$ under $\mu_{0}H = 35$ T at $T = 0.35$ K. The angular dependence of the SdH frequencies is found to be in very good agreement with the first-principle calculations when the electron- and hole-bands are slightly shifted with respect to the Fermi level, thus supporting the existence of underlying Weyl type-II points in WP$_2$.'
author:
- 'R. Sch[ö]{}nemann'
- 'N. Aryal'
- 'Q. Zhou'
- 'Y. -C. Chiu'
- 'K. -W. Chen'
- 'T. J. Martin'
- 'G. T. McCandless'
- 'J. Y. Chan'
- 'E. Manousakis'
- 'L. Balicas'
title: 'Fermi surface of the Weyl type-II metallic candidate WP$_2$'
---
Weyl fermions are predicted to emerge as low energy excitations in semimetals characterized by strong spin-orbit coupling and lack of inversion or time-reversal symmetries [@soluyanov_type-ii_2015; @felser; @bernevig2; @Weyl1; @Weyl2; @Hasan; @xu_discovery_2015]. Two types of Weyl semi-metallic systems (WSM) have been proposed: Type-I displays linearly dispersing bands which cross at pairs of point-like Fermi surfaces (or Weyl points), while in type-II the Weyl points appear at linearly touching points between electron and hole pockets [@soluyanov_type-ii_2015] resulting from the intersection of these bands with the Fermi level. Weyl points act as a topological charges, or as either sources or sinks of Berry-phase curvature pseudospin. Experimental fingerprints of WSM systems are topological Fermi arcs on the Fermi surface (FS) of the surface states [@xu_discovery_2015; @deng_experimental_2016; @tamai_fermi_2016] and unconventional magnetotransport properties due to the Adler-Abel-Jackiw anomaly [@adler; @bell; @nielsen], which corresponds to the pumping of charge carriers between Weyl points of opposite charge, or chirality, under the simultaneous presence of electric and magnetic fields [@adler; @bell; @nielsen; @zyuzin_topological_2012; @burkov_chiral_2015; @huang_observation_2015; @zhang_signatures_2016].
Te based transition-metal dichalcogenides such as WTe$_2$ and orthorhombic MoTe$_2$ (or $\gamma-$MoTe$_2$) were predicted to display a Weyl type-II semi-metallic state [@felser; @bernevig2]. A series of angle-resolved photoemission spectroscopy (ARPES) experiments [@tamai_fermi_2016; @arpes] on MoTe$_2$ claim to observe topological Fermi arcs and a FS whose geometry is in broad agreement with first-principle calculations [@bernevig2]. However, this contrasts with a quantum oscillatory study which reveals a far simpler Fermi surface displaying three-dimensional character [@daniel] despite the layered nature of this compound. The reason for such a discrepancy remains unclear, but it does question the validity of both the ARPES measurements and the predictions by density-functional theory calculations. Recently, a type-II WSM ground-state was predicted for the transition-metal diphosphides MoP$_2$ and WP$_2$ [@autes_robust_2016]. WP$_2$ crystallizes in two distinct structures [@ruhl_uber_1983; @martin_synthesis_1990; @mathis_reduction_1991]: the $\alpha$-phase which displays the OsGe$_2$ structure type and the $\beta$-phase which is isostructural to MoP$_2$ and belongs to the non-centrosymmetric orthorhombic space group $Cmc2_{1}$ (36) [@ruhl_uber_1983] (see Fig. \[fig:panel1\](a)) and is predicted to host Weyl points. Unlike WTe$_2$ and $\gamma-$MoTe$_2$, $\beta-$WP$_2$ is not a layered material which leads to a simpler band structure and a more robust arrangement of the Weyl points with respect to small structural changes [@autes_robust_2016]. Based on DFT calculations it was shown that the electron and hole pockets touch at two inequivalent points located at $0.471$ and $0.340$ eV below the Fermi energy $E_{\mathrm{F}}$ resulting in a total of eight Weyl points across the $k_{x}$-$k_{y}$-plane [@autes_robust_2016]. In addition, a very recent report highlights its unusual electrical transport properties including extremely large magnetoresistivity (MR) possibly related to its underlying Weyl physics [@kumar_extremely_nodate].
In this letter we present a detailed study on the Fermi surface of $\beta$-WP$_2$ *via* measurements of the Shubnikov-de Haas (SdH) effect which allows us to compare the geometry of the FS determined experimentally with the one predicted by the DFT calculations. The SdH-effect can only detect the electronic structure at $E_{\mathrm{F}}$ and is not able to directly probe the existence of the Weyl points below $E_{\mathrm{F}}$. But a good agreement between the calculated and the experimentally determined Fermi surfaces validates the band-structure calculations and therefore the predicted existence of underlying Weyl type-II points in this compound. Given that previous quantum oscillatory measurements [@daniel] unveiled strong discrepancies with the DFT predictions, our findings make $\beta$-WP$_2$ one of the most promising candidates for the realization of Weyl quasiparticles.
WP$_2$ single crystals were grown through a chemical vapor transport (CVT) method as described in detail within the Supplemental Information (SI) file [@Supplemental]. Conventional four-terminal resistivity measurements were performed in a physical property measurement system under magnetic fields up to $\mu_0H = 9$ T and temperatures as low as $T = 2$ K. The angular dependence of the Shubnikov-de Haas effect under magnetic fields up to $35\,\mathrm{T}$ was measured in a resistive Bitter magnet at the National High Magnetic Field Laboratory in Tallahassee. For further details see SI. Band structure calculations were performed by using the Quantum Espresso [@giannozzi_quantum_2009] implementation of the Density Functional Theory (DFT) in framework of the Generalized Gradient Approximation (GGA) including spin-orbit coupling (SOC). The Perdew-Burke-Ernzerhof (PBE) exchange correlation functional [@hamann_optimized_2013] was used with fully relativistic, norm conserving pseudo-potentials generated by using the optimized norm-conserving Vanderbilt pseudo-potential as described in Ref. [@hamann_optimized_2013]. For additional information see SI. The lattice parameters used in the calculations were extracted from single-crystal X-ray diffraction measurements and are in good agreement with previous reports [@noauthor_materialsproject_nodate], see SI.
Figures \[fig:panel1\](a) and \[fig:panel1\](b) display respectively, the crystallographic structure of WP$_2$ and its calculated FS within the first BZ. The electrical resistivity $\rho$ as a function of the temperature $T$ under magnetic fields $\mu_0H=0, 1, 3, 6,$ and 9 T are shown in Fig. \[fig:panel1\](c). For all of the measured samples the current $j$ was injected along the $a$-axis while the field remained perpendicular to $j$ by rotating it between the $c-$ and the $b-$axes. The crystal whose data is shown in the Fig. \[fig:panel1\](c) displays a residual resistivity ratio (RRR) of 4750 with a residual resistivity $\rho_{0}$ of $\approx 10$ n$\Omega$cm. This small value is unique among transition-metal chalcogenides/pnictides but comparable to those observed in pure metals [@noauthor_materialsproject_nodate]. For $\mu_{0}H = 9\,\mathrm{T}$ applied along the $b$-axis the magneto-resistivity (MR) exceeds $6 \times 10^{6}$ % at $T = 2\,\mathrm{K}$, which is comparable to WTe$_2$ under $\mu_0H = 60$ T [@ali_large_2014]. For all of the studied crystals, the RRR was found to vary between 1000 and $\approx 20000$. For subsequent measurements we selected crystals with $4000 \leq $ RRR $ \leq 10000$, finding no variations in the SdH signal superimposed onto $\rho (\mu_0H)$ among the different samples.
The FS shown in Fig. \[fig:panel1\](b) consists of 2 pairs of electron- and hole-pockets which, in absence of inversion symmetry, are split by the spin-orbit coupling. The electron-pockets are closed while the hole ones are corrugated cylinders responsible for the marked anisotropy in $\rho(\theta, \mu_0H)$ shown in Fig. \[fig:panel1\](d), which results from cyclotronic and open orbits on the FS. $\theta$ is the angle between $\mu_0H$ and the $b-$axis. The anisotropy in $\rho(\theta, \mu_0H)$ observed for fields along the $b$-axis and fields along the other two axes reaches $\sim 6000$ under $T = 2\,\mathrm{K}$ and $\mu_{0}H = 9\,\mathrm{T}$ and $\sim 35000$ under $T = 0.35\,\mathrm{K}$ and $\mu_{0}H = 35\,\mathrm{T}$. $\rho(\mu_0 H)$ follows an anomalous power-law when $\mu_0H \gtrsim 0.5\,\mathrm{T}$, i.e. $\rho(\mu_0 H)\propto (\mu_0 H)^{\lambda}$, with $\lambda = 1.8$, for $\mu_0 H\parallel b$ and $\lambda \approx 1.8 - 1.9$ for the other field orientations. No saturation was observed in $\rho(\theta, \mu_0H)$ for fields up to $35\,\mathrm{T}$, in agreement with an earlier report indicating no saturation up $60\,\mathrm{T}$ [@kumar_extremely_nodate]. Hall effect measurements, analyzed via a two-band model and discussed in the SI [@Supplemental], confirm that WP$_2$ is a well-compensated conductor, which explains its pronounced magnetoresistivity. Subsequently, we focus on the quantum oscillatory phenomena or on the pronounced SdH-effect superimposed onto the raw data plotted in Fig. \[fig:panel2\]. As the field is rotated from the $b-$ towards the $a$-axis the magnitude of the SdH oscillations decreases continuously and become unobservable for fields along the $a$-axis. However, we were able to observe a decrease in the longitudinal magnetoresistivity within a narrow angular window around the $a-$axis (see, SI), when the current is aligned along the magnetic field. Whether this results from the chiral anomaly [@adler; @bell; @nielsen; @zyuzin_topological_2012; @burkov_chiral_2015; @huang_observation_2015; @zhang_signatures_2016] or from current jetting (see Refs. [@pippard_magnetoresistance_2009; @ueda_anisotropy_1980]) remains unclear and will require additional studies. In WP$_2$, the effects of the chiral anomaly associated with its Weyl type-II points are predicted to be observable when fields and currents are parallel to the $b$-axis which, as already pointed out by Ref. [@kumar_extremely_nodate], is an experimentally challenging task.
Figure \[fig:panel3\] (a) shows $\rho$ as a function of $\mu_0H \parallel b$-axis for several temperatures $T$ ranging from 0.6 to 5.0 K. Figure \[fig:panel3\](b) displays the superimposed oscillatory signal after subtraction of the magnetoresistive background. The Onsager relation associates every frequency $F$ to a FS extremal cross-sectional area $A$. The fast Fourier transform (FFT), or the frequency spectra of the SdH signal, are shown in Fig. \[fig:panel3\](c). The observed FFT peaks can be assigned to individual FS pockets by comparing their positions with the frequencies predicted by our DFT calculations: the $\alpha$ and $\beta$ peaks belong to the hole-like sheets, the $\gamma$ and the $\delta$ ones to the electron pockets. Usually, there is more than just one extremal cross-sectional orbit (i.e. maxima and minima), hence multiple peaks can be assigned to a single FS sheet. Therefore, we proceed by labeling these orbits as the $\alpha_{1, 2}$ (or the $\beta_{1,2}$, etc.) frequencies. The effective cyclotron masses $\mu$ of the associated charge carriers can be extracted from the amplitude of the peaks seen in the FFT spectra as a function of the temperature. Their $T-$dependence is described by the Lifshitz-Kosevich temperature damping factor [@shoenberg_magnetic_1984], i.e. $R_{T} = BT/\sinh(BT)$, where $B = 14.69/\overline{\mu_0H}\mu$ with $\overline{\mu_0H}$ being the average inverse field. For the hole pockets we obtain $\mu$s between $\approx 0.8$ and 1.1 $m_0$, where $m_0$ is the bare electron mass, and $0.7-0.8$ $m_0$ for the electron pockets. For $\mu_0H \parallel a$-axis we obtain considerable heavier masses, between $3$ and $10$ $m_0$ for the electron pockets. See, SI for a comparison between calculated and experimentally extracted frequencies and effective masses.
Figures \[fig:panel4\](a, b) display the angular-dependence of the SdH frequencies obtained by rotating $\mu_0H$ in the $a-b$ and in the $b-c$ planes. The position of the peaks in the FFT spectra are shown as markers, whereas the theoretically predicted frequencies are indicated by solid lines. Higher harmonics of the fundamental frequencies were omitted for clarity. In Fig. \[fig:panel4\](b), the calculated electron- and hole-pockets were displaced, respectively, by $+30\,\mathrm{meV}$ and $-30\,\mathrm{meV}$ relative to $E_{F}$. This improves significantly the agreement between the calculated and the measured frequencies, particularly in what concerns the $\alpha_{2}$ and the $\beta_{2}$ branches. The two-dimensional character of the hole-pockets ($\alpha$, $\beta$ branches) is indicated by the divergence of the SdH frequencies upon approaching the $a$- or the $c$-axis. In contrast, the frequencies related to the electron pockets remain finite within the whole angular range. Large portions of the FFT spectra are dominated by the hole-frequencies, making it difficult to track the position of the much less pronounced FFT peaks associated with the electron pockets (i.e. $\gamma$, $\delta$), since they seem to partially overlap higher harmonics of the $\alpha$ and $\beta$ frequencies. SdH frequencies can be assigned to the $\gamma$ and $\delta$ branches only for angles close to the $a-$, $b-$ and the $c-$ axes. This is the reason for the absence of markers between $\theta = -60^{\circ}$ and $-20^{\circ}$ for rotations in the $b-c$ plane. Furthermore, for $F \lesssim 2$ T we were not able to detect any of the electron-like orbits predicted by DFT for rotations in the $b-c$ plane. Nevertheless, the FFT peaks assigned to the electron pockets display a good agreement with the theoretically predicted ones (depicted by blue and magenta lines).
In summary, WP$_2$ displays a very low residual resistivity and a gigantic magneto-resistivity for magnetic fields applied along its $b-$axis. These are clear indications for a very clean compound which, according to our Hall-effect measurements, is carrier compensated. The magnetoresistivity decreases by several orders of magnitude when the field is rotated towards the $b-$axis. This suppression results from two factors: open orbits on the hole-Fermi surfaces, and remarkably heavy effective masses (or lower mobilities) for the carriers performing cyclotronic motion on the electron-pockets. More importantly, our detailed study on the Shubnikov-de Haas effect reveals a Fermi surface geometry in quite good agreement with first-principle calculations. Density Functional Theory finds that WP$_2$ is a Weyl type-II “semi-metal", hence our results support the calculations. Although the Weyl type-II points are located well-below the Fermi-level, the observation of negative longitudinal magneto-resistivity might indicate that these can influence carriers at the Fermi level *via* the axial anomaly. The large number of SdH frequencies observed here precludes a reliable extraction of the Berry phase by fitting the oscillatory signal to the Lifshitz-Kosevich formalism. Hence, to confirm the topological character of this compound, angle resolved photoemission experiments would have to be performed in order to detect Fermi arcs [@soluyanov_type-ii_2015; @felser; @bernevig2; @Weyl1; @Weyl2; @Hasan; @xu_discovery_2015]. Finally, we want to point out that the density-of-carriers in WP$_2$ is 2 orders of magnitude larger those of conventional semi-metals, in agreement with its rather large Fermi surfaces. Hence, this compound should be classified as being metallic instead of “semi-metallic".
This work was supported by DOE-BES through award DE-SC0002613. The NHMFL is supported by NSF through NSF-DMR-1157490 and the State of Florida. Correspondence and requests for materials should be addressed to R.S. or to L.B.
[10]{} A. A. Soluyanov, D. Gresch, Z. Wang, Q. Wu, M. Troyer, X. Dai, and B. A. Bernevig, Nature [**527**]{}, 495 (2015). Y. Sun, S.C. Wu, M. N. Ali, C. Felser, and B. Yan, Phys. Rev. B **92**, 161107(2015). Z. Wang, D. Gresch, A. A. Soluyanov, W. Xie, S. Kushwaha, X. Dai, M. Troyer, R. J. Cava, and B. A. Bernevig, Phys. Rev. Lett. **117**, 056805 (2016). H. M. Weng, C. Fang, Z. Fang, B. A. Bernevig, and X. Dai, Phys. Rev. X **5**, 011029 (2015). S. Y. Xu, I. Belopolski, N. Alidoust, M. Neupane, G. Bian, C. Zhang, R. Sankar, G. Chang, Z. Yuan, C. -C. Lee, S. -M. Huang, H. Zheng, J. Ma, D. S. Sanchez, B. Wang, A. Bansil, F. Chou, P. P. Shibayev, H. Lin, S. Jia, M. Z. Hasan, Science **349**, 613-617 (2015). T. R. Chang, S. -Y. Xu, G. Chang, C. -C. Lee, S. -M. Huang, B. Wang, G. Bian, H. Zheng, D. S. Sanchez, I. Belopolski, N. Alidoust, M. Neupane, A. Bansil, H. -T. Jeng, H. Lin, and M. Z. Hasan, Nat. Commun. **7**, 10639 (2016). S.-Y. Xu, I. Belopolski, N. Alidoust, M. Neupane, G. Bian, C. Zhang, R. Sankar, G. Chang, Z. Yuan, C.-C. Lee, S.-M. Huang, H. Zheng, J. Ma, D. S. Sanchez, B. Wang, A. Bansil, F. Chou, P. P. Shibayev, H. Lin, S. Jia, and M. Z. Hasan, Science [**349**]{}, 613 (2015). K. Deng, G. Wan, P. Deng, K. Zhang, S. Ding, E. Wang, M. Yan, H. Huang, H. Zhang, Z. Xu, J. Denlinger, A. Fedorov, H. Yang, W. Duan, H. Yao, Y. Wu, S. Fan, H. Zhang, X. Chen, and S. Zhou, Nat. Phys. [**12**]{}, 1105 (2016). A. Tamai, Q. S. Wu, I. Cucchi, F. Y. Bruno, S. Riccò, T.K. Kim, M. Hoesch, C. Barreteau, E. Giannini, C. Besnard, A. A. Soluyanov, and F. Baumberger, Phys. Rev. X [**6**]{}, 031021 (2016). S. Adler, Phys. Rev. **177**, 2426 (1969). J. S. Bell, and R. A. Jackiw, Nuovo Cimento A **60**, 47 (1969). H. B. Nielsen, and M. Ninomiya, Phys. Lett. **130B**, 389 (1983). A. A. Zyuzin and A. A. Burkov, Phys. Rev. B [**86**]{}, 115133 (2012). A. A. Burkov, J. Phys.: Cond. Matt. [**27**]{}, 113201 (2015). X. Huang, L. Zhao, Y. Long, P. Wang, D. Chen, Z. Yang, H. Liang, M. Xue, H. Weng, Z. Fang, X. Dai, and G. Chen, Phys. Rev. X [**5**]{}, 031023 (2015). C.-L. Zhang, S.-Y. Xu, I. Belopolski, Z. Yuan, Z. Lin, B. Tong, G. Bian, N. Alidoust, C.-C. Lee, S.-M. Huang, T.-R. Chang, G. Chang, C.-H. Hsu, H.-T. Jeng, M. Neupane, D. S. Sanchez, H. Zheng, J. Wang, H. Lin, C. Zhang, H.-Z. Lu, S.-Q. Shen, T. Neupert, M. Zahid Hasan, and S. Jia, Nat. Commun. [**7**]{}, 10735 (2016). L. Huang, T. M. McCormick, M. Ochi, Z. Zhao, M.-T. Suzuki, R. Arita, Y. Wu, D. Mou, H. Cao, J. Yan, N. Trivedi, and A. Kaminski, Nat. Mater. **15**, 1155 (2016); K. Deng, G. Wan, P. Deng, K. Zhang, S. Ding, E. Wang, M. Yan, H. Huang, H. Zhang, Z. Xu, J. Denlinger, A. Fedorov, H. Yang, W. Duan, H. Yao, Y. Wu, S. Fan, H. Zhang, X. Chen, and S. Zhou, Nat. Phys. **12**, 1105 (2016); J. Jiang, Z. K. Liu, Y. Sun, H. F. Yang, C. R. Rajamathi, Y. P. Qi, L. X. Yang, C. Chen, H. Peng, C.- C. Hwang, S. Z. Sun, S.- K. Mo, I. Vobornik, J. Fujii, S. S. P. Parkin, C. Felser, B. H. Yan, and Y. L. Chen, Nat. Commun. **8**, 13973 (2017). D. Rhodes, Q. Zhou, R. Schönemann, Q. R. Zhang, E. Kampert, Y.-c. Chiu, Y. Lai, Y. Shimura, G. T. McCandless, J. Y. Chan, D. W. Paley, J. Lee, J. P. C. Ruff, S. Das, E. Manousakis, M. D. Johannes, and L. Balicas, arXiv:1605.09065 (2016). G. Autès, D. Gresch, M. Troyer, A. A. Soluyanov, and O. V. Yazyev, Phys. Rev. Lett. [**117**]{}, 066402 (2016). R. Ruhl and W. Jeitschko, Monatsh. Chem. [**114**]{}, 817 (1983). J. Martin and R. Gruehn, Solid State Ionics [**43**]{}, 19 (1990). H. Mathis, R. Glaum, and R. Gruehn, Acta Chem. Scand [**45**]{}, 781 (1991). N. Kumar, Y. Sun, K. Manna, V. Suess, I. Leermakers, O. Young, T. Foerster, M. Schmidt, B. Yan, U. Zeitler, C. Felser, and C. Shekhar, arXiv:1703.04527 (2017). See Supplemental Material at http://link.aps.org/supplemental, for the description of sample synthesis and characterization, angle dependence of the Fourier spectra, band-structure, and Hall-effect data and related analysis. P. Giannozzi [*et al.*]{}, J. Phys. Condens. Matter [**21**]{}, 395502 (2009). D. R. Hamann, Phys. Rev. B [**88**]{}, 085117 (2013). https://materialsproject.org/. M. N. Ali, J. Xiong, S. Flynn, J. Tao, Q. D. Gibson, L. M. Schoop, T. Liang, N. Haldolaarachchige, M. Hirschberger, N. P. Ong, and R. J. Cava, Nature (2014). A. B. Pippard, *Magnetoresistance in metals*, Cambridge studies in low temperature physics No. 2 (2009). Y. Ueda and T. Kino, J. Phys. Soc. Jpn. [**48**]{}, 1601 (1980). D. Shoenberg, *Magnetic oscillations in metals* (Cambridge University Press), Cambridge, (1984). M. Lenz and R. Gruehn, Chem. Rev. [**97**]{}, 2967 (1997). X. Luo, C. Fang, C. Wan, J. Cai, Y. Liu, X. Han, Z. Lu, W. Shi, R. Xiong, and Z. Zeng, Nanotechnology [**28**]{}, 145704 (2017). Q. Zhou, D. Rhodes, Q. R. Zhang, S. Tang, R. Schönemann, and L. Balicas, Phys. Rev. B [**94**]{}, 121101 (2016).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The multichannel Na-Cs interactions are characterized by a series of measurements using two atoms in an optical tweezer, along with a multichannel quantum defect theory (MQDT) with minimal input parameters. The triplet and singlet scattering lengths are measured by performing Raman spectroscopy of the Na-Cs motional states and least-bound molecular state in the tweezer. The two-scale MQDT improves accuracy over the single-scale model by incorporating the $-C_8/r_8$ potential in addition to the $-C_6/r_6$ potential. Magnetic Feshbach resonances are observed for only two atoms at fields which agree to within $1\%$ of the MQDT predictions. Our tweezer-based approach combined with an effective theory of interaction provides a new methodology for futures studies of more complex interactions, such as atom-molecule and molecule-molecule, and where the traditional high-phase-space density bulk-gas techniques are technically challenging.'
author:
- 'J.D. Hood'
- 'Y. Yu'
- 'Y.-W. Lin'
- 'J.T. Zhang'
- 'K. Wang'
- 'L.R. Liu'
- 'B. Gao'
- 'K.-K. Ni'
title: Multichannel interactions of two atoms in an optical tweezer
---
Introduction
============
Tuning interactions in ultracold gases of atoms and molecules via Feshbach resonances or optical lattices has enabled studies of many rich quantum phenomena such as the BEC-BCS crossover [@Regal2004], superfluid-to-Mott insulator transitions [@Greiner2002a], and supersolidity [@Bottcher2019; @Tanzi2019; @Chomaz2019]. Feshbach resonances have also been utilized to associate loosely-bound molecules, which has been an an important step for creating ultracold ro-vibrational ground state molecules [@Ni2008; @Molony2014; @Takekoshi2014; @Park2015; @Guo2016]. A key prerequisite for these experiments is an understanding of the underlying two-body and few-body interactions. Although the origins of these interactions are complex molecular potentials, in the low-temperature regime effective theories can describe the interactions with no reliance on the short-range potentials. For example, the single-scale multichannel quantum defect theory (MQDT) can provide an efficient description of atom-atom interactions in all spin channels and partial waves using only three parameters: the triplet and singlet scattering lengths, and the van der Waals $C_6$ coefficient [@Gao2005a].
Probing two- to few-body interactions in cold atoms has traditionally been performed by scattering experiments in bulk gases [@Weiner1999] or by spectroscopy in optical lattices [@Ospelkaus2006c; @Danzl2010; @Covey2016; @Goban2018; @Amato-Grill2019], both of which rely on initially preparing high phase-space-density (PSD) gases. More recently, single atoms trapped in optical tweezers have become widely pursued as a versatile experimental platform for studying few and many-body physics through bottom-up scaling [@Schlosser2001; @Darquie2005; @Miroshnychenko2006; @Yavuz2006; @Kaufman2015; @Thompson2013; @Norcia2018; @Saskin2019; @Covey2019]. An optical tweezer with only two atoms is a pristine environment for studying ultracold collisions [@Xu2015; @Liu2018; @Guan2019] or for producing two-particle entanglement [@Kaufman2015; @Sompet2019]. Optical tweezers are now even being used with single ultracold molecule assembly [@Liu2018; @Liu2019] and trapping [@Anderegg2019].
![\[fig:calc\] (a) A single Na and Cs atom are trapped and cooled to the motional ground state of an optical tweezer. (b) The Na-Cs interaction shifts their motional trapping frequencies in the tweezer (see inset). These trapping energies are calculated as a function of the scattering length , normalized by the relative oscillator length $\beta_{R,\text{Axial}} = 158$ nm, for various motional states, including the ground state $\Delta E_g$ (blue). The ground state energy $\Delta E_g$ (blue) is shifted to higher (lower) frequencies for positive (negative) scattering lengths, corresponding to a repulsive (attractive) potential. The motional states with relative and center-of-mass axial excitations $|m_{R},m_{\text{CM}}\rangle$ are also plotted, while radial excitations are off the scale. ](fig1.pdf){width="0.91\columnwidth"}
The natural questions arise: would measurements performed on two particles in an optical tweezer be sufficient to fully characterize two-body interaction, including the identification of Feshbach resonances? And, in the future, can such a platform offer a way to investigate interactions of more complex composite particles?
In this letter we probe the collisions of a single Na and a single Cs atom in an optical tweezer and in fully-controlled internal and external quantum states without any contribution from multi-body effects or intra-species processes. Despite previous characterization of the Na-Cs potential using Fourier-transform spectroscopy with hot atoms [@Docenko2006], the near-threshold NaCs ground-state potential and Feshbach resonances have not been directly probed. Our work combines Raman spectroscopy of trap motional states and the least-bound molecular state with a two-scale MQDT to extract the Na-Cs singlet and triplet scattering lengths. The single-scale MQDT [@Gao2005a; @Hanna2009; @Gao2011; @Makrides2014; @Cui2018b] describes low-energy alkali interactions with the fewest parameters by separation the long-range potential $-C_6/r^6$ (with length scale $\beta_6 = (2 \mu C_6/\hbar^2)^{1/4}$) from the short-range potential, which is captured by the singlet and triplet scattering lengths. But its accuracy for magnetic Feshbach resonances decreases for systems with larger hyperfine splittings due to the larger energy scale [@Makrides2014]. We introduce a two-scale MQDT that captures the larger energy variation with a the shorter length scale $-C_8/r^8$ potential. We observe the first Na-Cs Feshbach resonances, and the corresponding magnetic fields agree to within $1\%$ of our effective theory.
Our tweezer-based scheme can be extended to more complex interactions, such as atom-molecule or molecule-molecule interactions [@Mayle2012; @Mayle2013; @Gregory2019]. The diffraction-limited optical tweezer creates a high effective density for collisions of around $\rho \approx 10^{14}\, \text{cm}^{-3}$, which is useful for species where achieving high densities is otherwise experimentally challenging. Whereas traditional bulk-gas and optical lattice experiments require suitable collisional properties (including miscibility) between all species in order to obtain a high PSD, the tweezer method instead achieves a high PSD by optically cooling individual particles [@Thompson2013; @Kaufman2015; @Norcia2018; @Covey2019] before merging them together.
Calculation
===========
We perform our experiment with a single $^{23}$Na and a single $^{133}$Cs atom in the motional ground state of the same optical tweezer, as schematically shown in Fig. \[fig:calc\](a) and reported in Ref. [@Liu2019]. Initially, a single Na and Cs atom are loaded stochastically into separate tweezers from a magneto-optical trap with a combined probability of $\approx 35\%$ [@Liu2018]. The final results, however, can be post-selected to guarantee both atoms are initially present with high confidence. The single atoms are then cooled simultaneously to their 3D motional ground state using Raman sideband cooling [@Monroe1995; @Kaufman2012; @Yu2018; @Liu2019]. Subsequently, one of the atoms is transported and merged into the tweezer of the other atom, all while maintaining both atoms in the motional ground state [@Liu2019; @Wang2019]. The resulting mean separation distance is 112 nm, which gives an effective density of $\rho = 2 \times 10^{14}\, \text{cm}^{-3}$.
The strategy for extracting the triplet and singlet scattering lengths is outlined here. The triplet least-bound binding energy is measured with two-photon spectroscopy and is directly related to the triplet scattering length by a single channel quantum defect theory (QDT) [@Gao1998b; @Gao2001; @Gao2008a], but extended to two-scale by including the $-C_8/r^8$ potential. We also measure the shifts of the Na-Cs motional states in the tweezer due to the interactions in various spin combinations, and relate the shifts to the scattering lengths through a numerical calculation of two atoms interacting in a harmonic potential. Finally, we use the two-scale MQDT (which accounts for the hyperfine interaction) to extract the singlet scattering length consistent with the measured scattering lengths for the different hyperfine spins.
When both atoms are in the same trap, they interact through their molecular potential $V(\mathbf{r}_1 - \mathbf{r}_2)$, which consists of short-range molecular forces and a long-range van der Waals potential with a lowest-order term $-C_6/r^6$, where $r$ is the relative coordinate. In the low-energy limit where the de Broglie wavelengths are much larger than the molecular potential, their interaction can be well-modeled by a Fermi pseudo-potential consisting of the scattering length $a$ and a regularized $\delta$-function, $V(\mathbf{r}_1 - \mathbf{r}_2) = \frac{2\pi \hbar^2}{\mu} a \,\delta^{(3)}(\mathbf{r}_1 -\mathbf{r}_2) \frac{\delta}{\delta r} r$ [@Busch1998], where $\mu$ is the reduced mass. The validity of the pseudo-potential is characterized by the ratio of the van der Waals length $\beta_6 = (2\mu C_6 / \hbar^2 )^{1/4}$ to the relative harmonic oscillator lengths $\beta_R = \sqrt{\hbar/ \mu \omega_R }$ [@Bolda2002; @Blume2002a], where $\omega_R$ is the relative trapping frequency. In our experiment, these are $\beta_6 \approx 6$ nm, $\beta_{R,\text{Radial}} \approx 66$ nm for the radial axes, and $\beta_{R,\text{Axial}} \approx 158$ nm for the axial axis.
{width="99.00000%"}
Scattering length This Work Ref. [@Docenko2006]
------------------------------------------------------------------- ---------------------- ---------------------
$a_T$ 30.4 $\pm$ 0.6 $a_0$ 33 $\pm$ 5 $a_0$
$a_S$ 428 $\pm$ 9 $a_0$ 513 $\pm$ 250 $a_0$
Hyperfine Channel Scattering length Interaction shift
Na(2,2)Cs(4,4), ($\uparrow_{\text{Na}}\uparrow_{\text{Cs}}$) 30.4 $a_0$ 1.40 kHz
Na(2,2)Cs(3,3), ($\uparrow_{\text{Na}}\downarrow_{\text{Cs}}$) -693.8 $a_0$ -30.7 kHz
Na(1,1)Cs(3,3), ($\downarrow_{\text{Na}} \downarrow_{\text{Cs}}$) 13.7 $a_0$ 0.62 kHz
Hyperfine Channel Feshbach resonance MQDT Prediction
Na(1-1)Cs(3,-3) $s$-wave 652.1 $\pm$ 0.4 G 663 G
Na(1,-1)Cs(3,-3) $p$-wave 791.10 $\pm$ .05 G 799 G
: Summary of measured scattering parameters. The triplet $a_T$ and singlet $a_S$ scattering lengths are compared to the results in Ref. [@Docenko2006]. The measured scattering length for various hyperfine channels as well as the corresponding interaction shift. The Feshbach resonances are compared to the two-scale MQDT predictions. []{data-label="table:1"}
The two-body interaction shifts the motional trapping energies of the atoms in the tweezer, which is calculated as a function of the scattering length in Fig. \[fig:calc\](b) and probed experimentally in Fig. \[fig:int\]. When the atoms have the same trapping frequencies, as in the case of identical atoms, the Hamiltonian is separable into center-of-mass (CM) and relative coordinates, for which analytical results exist for a regularized $\delta$-function in a spherically [@Busch1998] and cylindrically [@Idziaszek2006] symmetric harmonic trap. However, Cs trapping frequencies are $19 \%$ larger than Na trapping frequencies, which are $(\omega_{\text{Na},x},\, \omega_{\text{Na},y} ,\, \omega_{\text{Na},z} ) = 2 \pi \times (109, 118, 20) $ kHz. We therefore calculate the shifted frequencies using the analytic solutions for the separable, cylindrically symmetric Hamiltonian, and then diagonalize a matrix containing the remaining non-separable [@Bertelsen2007; @Deuretzbacher2008] and anisotropic terms of the full Hamiltonian. The details are described in appendix \[app:int\].
In Fig. \[fig:calc\](b), the calculated Na-Cs motional trapping energies are plotted as a function of the scattering length $a$. The motional ground state $\Delta E_g$ (blue) shifts to higher (lower) frequencies for positive (negative) scattering lengths, corresponding to a repulsive (attractive) potential. The dashed line is the first-order perturbation theory described in Appendix \[app:pert\]. For the excited motional states (black), states with an odd relative axial quantum number $m_{R}$ have no shift because the relative wavefunction is zero at the $\delta$-function, and therefore are non-interacting, while states with even $m_{R}$ are interacting.
Interaction shift experiment
============================
Experimentally, we measure the interaction shifts by resonantly flipping the ground hyperfine spin of one but not the other atom and then comparing it to the bare hyperfine splitting measured in the absence of the other atom, as shown in Fig. \[fig:int\](a). Atoms are initially prepared in stretched state spin combinations that are stable against hyperfine changing spin collisions and denoted by ${|\!\uparrow_{\text{Cs}}\rangle}={|F\!=\!4,m_F\!=\!4\rangle_{\text{Cs}}}$, ${|\!\downarrow_{\text{Cs}}\rangle}={|F\!=\!3,m_F\!=\!3\rangle_{\text{Cs}}}$, ${|\!\uparrow_{\text{Na}}\rangle}={|F\!=\!2,m_F\!=\!2\rangle_{\text{Na}}}$, and ${|\!\downarrow_{\text{Na}}\rangle}={|F\!=\!1,m_F\!=\!1\rangle_{\text{Na}}}$. The spin flip of one atom is driven by an optical Raman pulse with co-propagating beams at a bias magnetic field of 8.8 gauss.
For the case of flipping a Cs spin in Fig. \[fig:int\](b), several interaction-shifted peaks are observed by comparing ${|\uparrow_{\text{Na}}\uparrow_{\text{Cs}}\rangle}\rightarrow{|\uparrow_{\text{Na}}\downarrow_{\text{Cs}}\rangle}$ (blue) to the one-atom case ${|\uparrow_{\text{Cs}}\rangle}\rightarrow{|\downarrow_{\text{Cs}}\rangle}$ (orange). The largest peak is the shifted ground motional state ${|m_{R},m_{\text{CM}}\rangle}$=${|0,0\rangle}$. The shift gives the difference of the interaction shifts of the initial and final spin configurations, $\Delta\nu_1 = ({\Delta E_g(\uparrow_{\text{Na}}\uparrow_{\text{Cs}})}-{\Delta E_g(\uparrow_{\text{Na}}\downarrow_{\text{Cs}})} )/\hbar = -32$.1(2)kHz. The smaller peaks at positive frequency shifts correspond to the motional excited states ${|2,0\rangle}$ and ${|4,0\rangle}$, which can be populated because they have some overlap with the initial state due to the wavefunction modification by the strong interactions. The peak near zero frequency corresponds to the initial Na and Cs population that is not prepared in the motional ground state or an interacting state. The fitted height 0.46 of the $|0,0 \rangle$ peak serves as a lower bound for the relative motional ground state population.
Similarly, for the case of flipping a Na spin ${|\uparrow_{\text{Na}}\downarrow_{\text{Cs}}\rangle}\rightarrow{|\downarrow_{\text{Na}}\downarrow_{\text{Cs}}\rangle}$ in Fig. \[fig:int\](c), we observe several interaction shifted peaks corresponding to the motional states of ${|0,0 \rangle}$, ${|2,0 \rangle}$, and ${|0,2\rangle}$, and a non-shifted peak as the initial non-interacting population.
Because interaction shifts only give the difference of the shifts between two states, we determine an absolute interaction shift of the triplet ${|\uparrow_{\text{Na}}\uparrow_{\text{Cs}}\rangle}$ by measuring the triplet least-bound ($v=-1$) binding energy, which can be related to the triplet scattering length directly through the two-scale single channel QDT. We measure the least-bound ($v=-1$) triplet binding energy with two-photon Raman spectroscopy, as schematically shown in Fig. \[fig:binding\](a). When the two-photon detuning is resonant with the binding energy, the atoms are transferred to the molecular state which is observed as simultaneous loss of both the Na and Cs atom [@Liu2019], as shown in the spectrum in Fig. \[fig:binding\](b). The resonance positions are plotted in Fig. \[fig:binding\](c) as a function of different tweezer powers in order to extrapolate the binding energies without light shift. For these experiments, the optical tweezer light is also used as the Raman beams, and is detuned $+18.2$ GHz from the $c^3\Sigma_{\Omega=1}(v\!=\!0,J\!=\!2)$ line at 288,698.2 GHz. A linear extrapolation to zero power gives a $N=0$ binding energy of 297.6(1) MHz.
The two-scale single-channel QDT relates the binding energies to the scattering lengths. For the van der Waals coefficients, we use the $C_6=3227$ a.u. and $C_8=3.681 \times 10^5$ a.u. from Refs. [@Docenko2006; @Derevianko2001; @Porsev2003]. The $N=0$ binding energy then gives a triplet scattering length $a_T = a(\uparrow_{\text{Na}}\uparrow_{\text{Cs}}) = 30.4(6) \, a_0$, where $a_0$ is the Bohr radius. The scattering lengths are summarized in Table. \[table:1\].
![\[fig:binding\] (a) Raman spectroscopy to measure the binding energy of the least-bound state $v=-1$ of the ground molecular potential $a^3\Sigma^+$. The tweezer light is also used for both branches of Raman light. (b) Example Raman spectrum, where the two-photon relative detuning $\Delta$ is scanned until the molecular resonance is observed and both atoms are lost. (c) Measured binding energies for triplet least-bound molecular state as a function of the tweezer power in units of Na+Cs trap depth. ](fig3.pdf){width="0.95\columnwidth"}
From the triplet scattering length from the binding energy measurement, the pseudo-potential model then gives an absolute interaction shift of $\Delta E_g(\uparrow_{\text{Na}}\uparrow_{\text{Cs}})/h$ = 1.40 kHz, which can be used to obtain an absolute shift for each state in the interaction shift measurements. The measurement $\Delta \nu_1$ in Fig. \[fig:int\](b) gives an absolute interaction shift $\Delta E_g(\uparrow_{\text{Na}}\downarrow_{\text{Cs}}) /h = -30.7$ kHz, corresponding to a scattering length of $a(\uparrow_{\text{Na}}\downarrow_{\text{Cs}}) = -693.8\, a_0$. The two-scale MQDT is then used to extract a singlet scattering length of $a_{S} = 428(9)$ $a_0$.
With our measured triplet $a_{T}$ and singlet $a_{S}$ scattering lengths, MQDT can now describe Na-Cs interactions in all spin configurations and all magnetic fields. As a verification, we compare the MQDT calculation to the Na spin-flip interaction shift $\Delta\nu_2=31.8(2)$ kHz in Fig. \[fig:int\](c). The shifts predominately come from ${E_g(\uparrow_{\text{Na}}\downarrow_{\text{Cs}}) /h} = -30.7$ kHz. The MQDT predicts for the final state, ($\downarrow_{\text{Na}}\downarrow_{\text{Cs}}$), a scattering length of $a(\downarrow_{\text{Na}} \uparrow_{\text{Cs}})$ = 13.7 $a_0$ and a corresponding interaction shift of 0.64 kHz, which would result in a calculated $\Delta\nu_2=31.34$ kHz, consistent with the measurement.
Feshbach resonances
===================
With Na-Cs interactions completely characterized at a low magnetic field, we use our two-scale MQDT to guide a search of Feshbach resonances at much higher magnetic fields. Feshbach resonances in ultracold atoms occur when pairs of atoms are magnetically tuned into resonance with a closed molecular bound state, as shown in Fig. \[fig:feshbach\](b). Searching for these resonances is typically done in bulk gases by measuring atom loss near the resonances due to enhanced two-body or three-body inelastic collisions. Here we demonstrate such a search starting with exactly two atoms in an optical tweezer.
Our search uses a separate optical tweezer apparatus, shown in Fig. \[fig:feshbach\](a), with Helmholtz coils that are capable of producing a magnetic field $B$ up to 1000 G. The atoms are prepared in the lower hyperfine manifolds, Na($F\!=\!1$,$m_F\!=\!-1$) and Cs($F\!=\!3$, $m_F\!=\!-3$), at about 100 $\mu$K. Despite the high temperature as compared to typical bulk-gas-based searches, the tight confinement in the optical tweezer increases the sensitivity to collisional losses. The B-field is ramped on, the atoms are merged into the same tweezer and held for 100 ms, and then the atoms are separated and imaged individually to check for loss. Because there are no lower energy states with the same total $M_F$, the inelastic loss must occur through anisotropic interactions such as the electron spin-spin interactions.
Using our measured scattering parameters, the two-scale MQDT predicts two Feshbach resonances for this hyperfine state, where the closed channel is either the rotational ground $N=0$ ($s$-wave) or excited $N=1$ ($p$-wave) of a molecular state which has the approximate quantum number $\nu=-1$ and Na($F\!=\!2$)Cs($F\!=\! 3$) in the low-field limit. Na-Cs loss spectroscopy is shown in Fig. \[fig:feshbach\](c) and (d) for two different magnetic field ranges. The $s$-wave resonance is observed at 652.1(4) G, and the narrower $p$-wave Feshbach resonance is observed at 791.10(5) G, which agree with the MQDT predictions with an accuracy of $1.4\%$ and $0.9\%$ respectively. These are summarized in Table \[table:1\]. With a colder sample for future study, inelastic confinement-induced resonances could shift the Feshbach resonance location due to coupling between center-of-mass and relative motion, as described in Ref. [@Sala2012].
![\[fig:feshbach\] (a) Apparatus for observing Feshbach resonances in an optical tweezer. The tweezer trap atoms in an ultra-high vacuum glass cell chamber between Helmholtz coils (orange), which produces a magnetic field $B$ of up to 1000 gauss at the atoms. (b) Feshbach resonances occurs when the tweezer trap states are tuned into resonance with a molecular bound state by applying a magnetic field. (c-d) An $s$-wave and $p$-wave Feshbach resonance is observed by measuring the simultaneous loss of both Na and Cs as a function of magnetic field. The dashed line is a Gaussian fit. ](fig4.pdf){width="0.99\columnwidth"}
Conclusion
==========
The agreement of the Feshbach resonance locations with the two-scale MQDT model combined with a series of tweezer-based measurements represents an important validation of the use of effective theory for interactions. The interaction of composite particles, and in particular molecules, is generally too difficult for the standard coupled-channel approach because the number of coupled-channels is simply too large [@Mayle2012; @Mayle2013; @Gregory2019]. This difficulty requires that we look for effective theories to describe the interactions that require a minimum number of parameters, and efficient ways to measure these parameters. This tweezer scheme can also be used to probe atom-molecule and molecule-molecule interactions, providing both high effective densities and exact preparation of the collisional partners. The measurement of the interaction shift of an atom in the presence of a molecule could also be used for non-destructive state-sensitive detection of the molecule, which is a challenge for molecules without closed optical transitions. For Na-Cs in particular, the observation of a Na-Cs Feshbach resonance for exactly two atoms is an important step towards creating a single Feshbach molecule, and the eventual coherent creation of a ro-vibrational ground state molecule.
Two-scale multichannel quantum defect theory {#app:mqdt}
============================================
The simplest description of the interaction between two alkali-metal atoms is a three-parameter model based on a single-scale multichannel quantum defect theory (MQDT) [@Gao2005a; @Hanna2009; @Gao2011; @Makrides2014; @Cui2018b]. Here “single-scale” corresponds to the fact the theory is built on the solutions for the long-range potential $-C_6/r^6$ which has a single length scale $\beta_6=(2\mu C_6/\hbar^2)^{1/4}$. And the three parameters can be taken as the singlet $s$-wave scattering length $a_S$, the triplet $s$-wave scattering length $a_T$, and the $C_6$ coefficient. The theory provides the simplest description of low-energy alkali interactions with or without magnetic field, including magnetic Feshbach resonances in all partial waves.
Depending on the specific system, the description of the magnetic Feshbach spectrum by this simple model has a typical accuracy of a few percent to about 8% for heavier systems with large hyperfine splittings [@Cui2018b]. To understand these deviations and their improvements, one can keep in mind that an accurate description of a magnetic Feshbach resonance requires a *simultaneous* accurate descriptions of both the open and the closed channels, in particular the bound spectrum in the closed channels over an energy range corresponding to the channel energy spacing. For alkali-metal atoms, the channels are separated by the hyperfine splittings, and the requirement for accurate Feshbach spectrum translates into the requirement of accurate bound spectrum over a binding energy range of $|\Delta E^{\mathrm{hf}}|$. Thus even when we are at ultracold temperatures where a single scattering length would suffice to describe the interaction in the open channel, the understanding of the scattering length itself, its relation to scattering lengths in other channels, and its tuning through the Feshbach spectrum, would require an underlying understanding over a much broader range of energies of at least $|\Delta E^{\mathrm{hf}}|$. The validity and the accuracy of single-scale MQDT depends on the validity that the energy variation over this energy range is due solely to the $-C_6/r^6$ potential, which is less valid for systems with large hyperfine splittings such as our system with a Cs atom that has a hyperfine splitting of around 9 GHz or 0.4 K.
There are different options for improving upon this single-scale baseline result. Focusing on effective theories that do not rely on the details of the short-range potential, the easiest and the simplest improvement to implement is to introduce parameters that characterize the energy dependence and the partial-wave dependence of the single-scale short-range parameters [@Li2014; @Li2015]. A more fundamental approach that requires the fewest number of extra parameters, an important criteria for a good effective theory of interaction, is to go to a shorter length scale through the inclusion of higher order terms in the asymptotic expansion for the potential, such as $$-\frac{C_6}{r^6}-\frac{C_8}{r^8} \;,$$ which is used as the reference potential for our 2-scale MQDT and is a more accurate representation of the real atomic interaction at shorter distances. The additional $-C_8/r^8$ terms has a corresponding length scale $\beta_8=(2\mu C_n/\hbar^2)^{1/(n-2)}=(2\mu C_8/\hbar^2)^{1/6}$ that is smaller than $\beta_6$.
A 2-scale MQDT description of alkali-metal interactions thus uses only one more parameter, the $C_8$, than the single scale theory for a total of 4 parameters that can be taken as $a_S$, $a_T$, $C_6$, and the $C_8$. It is more accurate in its description of scattering lengths and Feshbach spectrum, and can cover a greater range of energies when needed, similar to what has been demonstrated in QDT for $-C_1/r-C_4/r^4$ potential [@Fu2016a]. The theory is formally the same as the single-scale theory except for the details of the QDT functions. These details will be presented elsewhere. Figure \[fig:NaCsgslBMFm4\] shows the reduced generalized scattering lengths [@Gao2011; @Makrides2014] for $s$ and $p$ partial waves in the Na$(-1)_1$Cs$(-3)_1$ channel over a $B$ field range of 0-1000 G, giving a more complete picture of the Feshbach resonances that we have observed. Table \[tab:Na23Cs133FeshMFm4\] gives the calculated properties of the Feshbach resonances.
![The reduced generalized scattering lengths [@Gao2011; @Makrides2014] for $s$ and $p$ partial waves in the Na$(-1)_1$Cs$(-3)_1$ channel. The $s$-wave has a Feshbach resonance at 663 G, and the $p$-wave has a Feshbach resonance at 799 G. \[fig:NaCsgslBMFm4\]](appfig1.pdf){width="0.99\columnwidth"}
We point out that similar theoretical analysis can also be carried out using numerical formulations of MQDT [@Mies1984a; @Mies1984b; @Burke1998; @Ruzic2013] and other numerical methods of similar spirit which are already multiscale (see, e.g. [@vanKempen2002; @Pires2014]). The main difference of our approach is that it tries to make the best use of the existing single-scale analytic solutions [@Gao1998a], and is formulated to show explicitly how physics at different length scales are related, and the progression from the single-scale theory that is sufficient at small energies and requires fewer parameters, to multiscale theories at shorter length scales that requires more parameters. In such an approach, an extra parameter or the complexity of the theory is added only when necessary. This characteristic is important in the context of an effective theory of interactions, especially when the parameters of the theory need to be determined from a limited number of experimental measurements.
For our particular system of $^{23}$Na-$^{133}$Cs, we have taken the long-range potential parameters $C_6 = 3227$ a.u. and $C_8 = 3.681\times 10^5$ a.u. from Docenko *et al*. [@Docenko2006]. They are much more reliably determined by theory [@Derevianko2001; @Porsev2003] than the parameters $a_S$ and $a_T$, which are sensitive to short-range complex molecular interactions. With still a limited number of experimental measurements, only the latter parameters are determined experimentally as discussed in the main text. With more future measurements, especially those related to more deeply bound molecular states, we anticipate more precise determinations of all parameters including the $C_6$ and $C_8$ coefficients in a future study.
Channel $M_F$ $\ell$ $B_{0\ell}$ (G) $\Delta_{B\ell}$ (G) $\widetilde{a}_{\text{bg}\ell}/\bar{a}_{\ell}$ $\delta \mu_\ell / \mu_B$ $K^{c0}_{\text{bg}\ell}$ $g_{\text{res}}$ $d_{B\ell}$ (G) $\zeta_{\text{res}}$
---------------- ------- -------- ----------------- ---------------------- ------------------------------------------------ --------------------------- -------------------------- ------------------ ----------------- ----------------------
$(-1)_1(-3)_1$ $-4$ 0 663.0 42.54 -0.6487 -0.6124 -0.6065 2.366 -16.74 1.300
$(-1)_1(-3)_1$ $-4$ 1 799.0 6.473 -1.242 -0.5026 -4.128 3.850 -33.19 -0.1865
Interaction shift calculation {#app:int}
=============================
In the limit of small collision energy, the molecular potential $V(r)$ can be replaced with a Fermi pseudopotential consisting of the scattering length $a$ and a regularized $\delta$-potential, $$V(|\mathbf{r}|) = \frac{2 \pi \hbar^2 a}{\mu } \delta^{(3)}(\mathbf{r}) \frac{\partial}{ \partial r} r.$$ For low energies, the Fermi pseudopotential reproduces the correct wavefunction outside the range of interactions. The Fermi pseudopotential approximation accuracy is characterized by the ratio of the characteristic van der Waal’s length $\beta_6 = (2\mu C_6 / \hbar^2 )^{1/4}$ to the relative oscillator length $\beta_R = \sqrt{\hbar/ m \omega_R}$ [@Bolda2002; @Blume2002a]. The Na mass is 23 amu and Cs mass is 133 amu. The measured trapping frequencies are $(\omega_{\text{Na},x},\, \omega_{\text{Na},y} ,\, \omega_{\text{Na},z} ) = 2 \pi \times (109, 118, 20) $ kHz, and $(\omega_{\text{Cs},x},\, \omega_{\text{Cs},y} ,\, \omega_{\text{Cs},z} ) = 2 \pi \times (130, 140, 24)$ kHz. In our experiment, $\beta_6 = 6$ nm is much smaller than the relative oscillator lengths $\beta_{R,\text{Axial}} = 158$ nm and $\beta_{R,\text{Radial}} = (65,67)$ nm.
The Hamiltonian for two different atoms interacting in an anisotropic 3D harmonic trap is given by $$\begin{split}
H = \sum_{i=x,y,z} \left( \frac{1}{2} m_1 \dot{x}^2_{1, i} + \frac{1}{2} m_1 \omega^2_{1,i} x_{1, i}^2 \right) \\
+ \sum_{i=x,y,z} \left( \frac{1}{2} m_2 \dot{x}^2_{2, i} + \frac{1}{2} m_2 \omega^2_{2,i} x_{2, i}^2 \right)
+ V(|\mathbf{r}_1 - \mathbf{r}_2|),
\end{split}$$ where $V(|\mathbf{r}_1 - \mathbf{r}_2|)$ is the spherically symmetric molecular potential for the two atoms. The trapping frequencies are in general different for the two atoms due to different polarizabilities and masses. The trapping frequencies can also be different for all three axes in an anisotropic trap. In terms of the relative and center-of-mass coordinates, $\mathbf{r}_R = \mathbf{r}_1 - \mathbf{r}_2$ and $ \mathbf{r}_{C} = \frac{m_1 \mathbf{r}_1 + m_2 \mathbf{r}_2}{m_1 + m_2}$, the Hamiltonian can be re-expressed in terms of the reduced mass $\mu = \frac{m_1 m_2}{m_1 + m_2}$ , total mass $M = m_1 + m_2 $, relative frequencies $ \omega_{R,i} =\sqrt{\frac{ m_2 \omega_{1,i}^2 - m_1 \omega_{2,i}^2}{m_1 + m_2}}$, and center-of-mass frequencies $\omega_{C,i} = \sqrt{\frac{m_1 \omega_{1,i}^2 + m_2 \omega_{2,i}^2}{m_1 + m_2}}$ as $$\label{eq:H1}
\begin{split}
H = \sum_{i=x,y,z} \left( \frac{1}{2} M \dot{x}^2_{C, i} + \frac{1}{2} M \omega^2_{C,i} x_{C, i}^2 \right)\\
+ \left( \sum_{i=x,y,z} \left(\frac{1}{2} \mu \, \dot{x}^2_{R, i} + \frac{1}{2} \mu \,\omega^2_{r,i} x_{R, i}^2\right) +V_{\text{int}}(|\mathbf{r}_R|)\right) \\
+ \sum_{i=x,y,z} \mu (\omega_{1,i}^2 - \omega_{2,i}^2) x_{R,i} x_{C, i} .
\end{split}$$ The center-of-mass part of the Hamiltonian is just a single particle 3D harmonic oscillator. The second term, the relative Hamiltonian, is a 3D harmonic oscillator with a regularized $\delta$-function at the origin, for which analytic solutions exist for cases that the trap has spherical [@Busch1998] and cylindrical [@Idziaszek2006] symmetries.
The final term in the Hamiltonian in Eq. \[eq:H1\] couples the center-of-mass and relative coordinates. If the trapping frequencies of the two atoms are the same for each axis (for example in the case when the two atoms are the same species), then the last term vanishes, and then the relative and center-of-mass coordinates are separable. But when the two atoms are different species, as in our case, then this term can be important, which is discussed for the spherical case in Refs. [@Bertelsen2007; @Deuretzbacher2008]. The trapping frequencies depend on the polarizability and mass as $\sqrt{\alpha(\lambda )/ m}$. At the tweezer wavelength of $\lambda$ = 976 nm, the measured Cs trapping frequencies are $\approx 19\%$ larger than those of Na.
To find the eigenenergies of the full Hamiltonian of Eq. \[eq:H1\], we first ignore the last term so that the Hamiltonian is separable into center-of-mass and relative coordinates. In our experiment, the axial trapping frequency (along the axis of the tweezer and labeled as $z$) is approximately 5.6 times smaller than the two radial trapping frequencies ($x$ and $y$). There is also a $7 \%$ difference between the two radial axes, but we initially assume they are the same and add the difference later as a correction. Therefore, we use the cylindrical harmonic oscillator wavefunctions as the solutions of the center-of-mass Hamiltonian, and they are labeled as $|n,l,m_z\rangle$, where $n$ and $l$ are the principal and angular momentum quantum numbers for the radial part, and $m_z$ is quantum number for 1D harmonic oscillator for the axial part, and have eigenenergies $$\label{eq:Ecylin}
E_{n,l,m_z}/ (\hbar \omega_z)= ( 2 n + |l| + 1) \eta + (m_z + 1/2),$$ where we define $\eta$ as the ratio of the radial to axial trapping frequency.
For the relative Hamiltonian, we use the analytic cylindrical solutions from Ref. [@Idziaszek2006]. These solutions require that the axial trapping frequency is an integer multiple $\eta$ of the radial trapping frequency. We define $\eta = 6$, which is close to the actual values of $5.6$, and we will include the remaining terms later as a correction. The analytic solutions are given for the interacting states, but there also many relative states which have zero wavefunction at the $\delta$-function, and therefore are unaffected. For example any state with $l\ne0$ or odd $m_z$ has a zero at $\delta$-function. The complete basis includes both the interacting states from Ref. [@Idziaszek2006] as well as all of the non-interacting states. The non-interacting states are solutions to the cylindrical harmonic oscillator, and so are just cylindrical harmonic oscillator wavefunctions.
One complication is that when $\eta$, the ratio of the radial to the axial trapping frequency, is an integer, there is a subspace of cylindrical harmonic oscillator states with $l=0$ and even $m_z$ that are degenerate and have the same energy from Eq. \[eq:Ecylin\]. In each degenerate subspace with $N_{\text{deg}}$ states, the non-interacting states are a linear superposition of the degenerate eigenstates $\psi_i$. We find these amplitudes $c_i$ using a Gram-Schmidt procedure, which requires that $\sum_{i=1}^{N_{\text{deg}}} c_i \psi_i(0)$ = 0.In each subspace, there is only one interacting state, for which the analytic solution is used, and $N_{\text{deg}}-1$ non-interacting states.
For the interacting states, the energies are given by the transcendental equations [@Idziaszek2006] $$\mathcal{F}(-(E-E_0)/2 , \eta) = -\sqrt{2\pi}/a,$$ where $\mathcal{F}(x,\eta)$ is given by $$\begin{split}
\mathcal{F}(x, \eta) = \frac{\sqrt{\pi} \Gamma(x)}{\Gamma(x+\frac{1}{2})} \sum_{m=1}^{n-1} F(1,x;x+\frac{1}{2} ; e^{i(2\pi m/ \eta )} ) \\ - \frac{2 \sqrt{\pi} \Gamma(x)}{\Gamma(x-\frac{1}{2})} .
\end{split}$$ Here $F(a,b;c,x)$ denotes the hypergeometric function and $\Gamma(x)$ is the Euler gamma function. The energy $E$ and $E_0$ are in units of the axial trap energy $\hbar \omega_z$, and so the ground state energy $E_0 = \eta + 1/2$.
Now that we have solution in the separable and cylindrical case, the next step is to include the non-separable and asymmetric correction terms by diagonalizing the total matrix in the combined center-of-mass and relative cylindrical bases. For the matrix, we include all states with energies up to 20 $\omega_{R,z}$. The matrix elements are calculated numerically using the cylindrical wavefunctions, which for completeness are given here:
$$\Psi_{n,l,m_z} (\rho,\theta,z) = \Psi^{\text{radial}}_{n,l}(\rho, \theta) \Psi^{\text{axial}}_{m_z}(z) ,$$
with the normalized radial harmonic oscillator wavefunction $$\begin{split}
\Psi^{\text{radial}}_{n,l}(\rho, \theta) = \sqrt{\frac{2 n!}{a_\perp^2 (n+|l|)!}} e^{-r^2/(2 a_\perp^2)} ( r/ a_\perp)^{|l|} \\ \times L_n^{|l|}( r^2/ a_\perp^2) \frac{ e^{i l \theta}}{\sqrt{2 \pi}},
\end{split}$$ and the normalized 1D harmonic wavefunction $$\Psi^{\text{axial}}_{m_z}(z) = \frac{1}{\sqrt{2^{m_z} m_z!}} \frac{1}{\sqrt{a_z}(\pi)^{1/4}} e^{- z^2/(2 a_z)} H_{m_z}( z/a_z).$$ Here the radial and relative oscillator lengths are defined as $a_\perp = \sqrt{\hbar/( \mu \omega_\perp)}$ and $a_z = \sqrt{\hbar/ (\mu \omega_z)}$. $H_{m_z}$ are the Hermite-Gaussian functions, and $L^{|l|}_n$ are the generalized Laguerre polynomials. The eigenenergies of the matrix are calculated as a function of the scattering length, which is shown in Fig. 1 of the main text.
Perturbation theory for ground state shift {#app:pert}
==========================================
We use first-order perturbation theory to estimate the shift of the Na-Cs 3D motional ground state due to the interaction, which is approximated by the Fermi pseudopotential interaction $V(| \mathbf{r}_1 - \mathbf{r}_2|) = \frac{2 \pi \hbar^2 a}{\mu } \delta^{(3)}(\mathbf{r}_1 - \mathbf{r}_2) \frac{\partial}{ \partial r} r$. Using the Cartesian harmonic oscillator bases $|n^{\text{Na}}_x, n^{\text{Na}}_y, n^{\text{Na}}_z ; n^{\text{Cs}}_x, n^{\text{Cs}}_y, n^{\text{Cs}}_z \rangle $, first-order perturbation theory gives a ground state shift of $\Delta E_g \approx \frac{2\pi\hbar^2 a}{\mu }\langle 0,0,0;0,0,0|\delta^{(3)}(\mathbf{r}_1-\mathbf{r}_2) \frac{\partial}{ \partial r} r |0,0,0;0,0,0\rangle $, which simplifies to $$\label{eq:pert2}
\Delta E_g \approx a \left( \frac{2 \hbar^2}{\mu \sqrt{\pi}} \frac{1}{ \beta_{\text{eff}}^3 } \right) ,$$ where the effective 3D oscillator length is defined in terms of the 1D oscillator lengths, $$\beta_{\text{eff}} = \left( (\beta_{\text{Na},x}^2 +\beta_{\text{Cs},x}^2)(\beta_{\text{Na},y}^2 +\beta_{\text{Cs},,y}^2)(\beta_{\text{Na},z}^2 +\beta_{\text{Cs},,z}^2) \right)^{1/6}.$$ The oscillator lengths are $\beta = \sqrt{\hbar/ m \omega} $, where $m$ is the mass of the atom and $\omega$ is the trapping frequency.
We thank Lewis Picard, Eliot Fenton, and Frederic Condin for experimental assistance. This work is supported by the Arnold and Mabel Beckman Foundation, as well as the NSF (PHYS-1806595 and through Harvard-MIT CUA), AFOSR (FA9550-19-1-0089), and the Camille and Henry Dreyfus Foundation. J. T. Z acknowledges support from an NDSEG fellowship. The work at Toledo was supported by NSF (PHY-1607256).
[66]{}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.92.083201) [****, ()](https://doi.org/10.1038/415039a) [****, ()](\doibase
10.1103/PhysRevX.9.011051) [****, ()](\doibase
10.1103/PhysRevLett.122.130405) [****, ()](\doibase 10.1103/PhysRevX.9.021012) [****, ()](\doibase
10.1126/science.1163861) [****, ()](\doibase 10.1103/PhysRevLett.113.255301) [****, ()](\doibase
10.1103/PhysRevLett.113.205301) [****, ()](\doibase 10.1103/PhysRevLett.114.205302) [****, ()](\doibase 10.1103/PhysRevLett.116.205303) [****, ()](\doibase 10.1103/PhysRevA.72.042719) [****, ()](\doibase
10.1103/RevModPhys.71.1) [****, ()](\doibase
10.1103/PhysRevLett.97.120402) [****, ()](http://dx.doi.org/10.1038/nphys1533) @noop [****, ()]{} [****, ()](\doibase 10.1038/s41586-018-0661-6) [****, ()](\doibase
10.1103/PhysRevA.99.033612) [****, ()](\doibase 10.1038/35082512) [****, ()](\doibase 10.1126/science.1113394) [****, ()](https://doi.org/10.1038/442151a) [****, ()](\doibase
10.1103/PhysRevLett.96.063001) [****, ()](\doibase
10.1038/nature16073) [****, ()](\doibase 10.1103/PhysRevLett.110.133001) [****, ()](\doibase 10.1103/PhysRevX.8.041054) [****, ()](\doibase 10.1103/PhysRevLett.122.143002) [****, ()](\doibase
10.1103/PhysRevLett.122.173201) [****, ()](https://doi.org/10.1038/ncomms8803) [****, ()](\doibase
10.1126/science.aar7797) [****, ()](\doibase
10.1103/PhysRevLett.122.083401) [****, ()](\doibase
10.1038/s41467-019-09420-6) [****, ()](\doibase 10.1103/PhysRevX.9.021039) [****, ()](\doibase
10.1126/science.aax1265) [****, ()](http://stacks.iop.org/0953-4075/39/i=19/a=S08) [****, ()](\doibase 10.1103/PhysRevA.79.040701) [****, ()](\doibase 10.1103/PhysRevA.84.022706) [****, ()](\doibase 10.1103/PhysRevA.89.062718) [****, ()](\doibase
10.1103/PhysRevA.98.042708) [****, ()](\doibase 10.1103/PhysRevA.85.062712) [****, ()](\doibase 10.1103/PhysRevA.87.012709) [****, ()](\doibase
doi.org/10.1038/s41467-019-11033-y) [****, ()](\doibase
10.1103/PhysRevLett.75.4011) [****, ()](\doibase 10.1103/PhysRevX.2.041014) [****, ()](\doibase
10.1103/PhysRevA.97.063423) [****, ()](\doibase 10.1103/PhysRevA.100.063429) [****, ()](\doibase 10.1103/PhysRevA.58.4222) [****, ()](\doibase 10.1103/PhysRevA.64.010701) [****, ()](\doibase 10.1103/PhysRevA.78.012702) [****, ()](\doibase https://doi.org/10.1023/A:1018705520999) [****, ()](\doibase 10.1103/PhysRevA.66.013403) [****, ()](\doibase 10.1103/PhysRevA.65.043613) [****, ()](\doibase 10.1103/PhysRevA.74.022712) [****, ()](\doibase 10.1103/PhysRevA.76.043615) [****, ()](\doibase
10.1103/PhysRevA.77.032726) [****, ()](\doibase 10.1103/PhysRevA.63.052704) [****, ()](\doibase 10.1063/1.1578052) [****, ()](\doibase 10.1103/PhysRevLett.109.073201) [****, ()](\doibase
10.1103/PhysRevA.89.052704) [****, ()](\doibase 10.1103/PhysRevA.91.032702) [****, ()](\doibase
10.1088/1367-2630/18/10/103016) [****, ()](\doibase 10.1063/1.447000) [****, ()](\doibase 10.1063/1.447046) [****, ()](\doibase 10.1103/PhysRevLett.81.3355) [****, ()](\doibase 10.1103/PhysRevA.87.032706) [****, ()](\doibase 10.1103/PhysRevLett.88.093201) [****, ()](\doibase
10.1103/PhysRevA.90.012710) [****, ()](\doibase 10.1103/PhysRevA.58.1728)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The velocity of neutrons from a pulsed neutron source is well-defined as a function of their arrival time. Electromagnetic neutron accelerator/decelerator synchronized with the neutron time-of-flight is capable of selectively changing the neutron velocity and concentrating the velocity distribution. Possible enhancement of the neutron intensity at a specific neutron velocity by orders of magnitude is discussed together with an experimental design.'
address:
author:
- ', , and'
title: Concentration of the velocity distribution of pulsed neutron beams
---
Introduction
============
Increase of the production efficiency of ultra-cold neutrons (UCNs) in the energy range of below about $250$ neV is intensively attempted for the improvement of the experimental accuracies of neutron properties such as the neutron electric dipole moment (nEDM), $\beta$-decay lifetime and searches for exotic interactions with neutrons [@refEDM; @refLife]. The UCN storage density is the crucial parameter especially for the suppression of the systematic errors in the measurement of the nEDM. The combination of an accelerator-based spallation neutron source and a superthermal down-conversion from the very-cold region into the ultracold region introduced the flexibility in the design of UCN sources and has been employed for new generation UCN sources [@refUCN]. Solid deuterium has been utilized at currently operational UCN sources at Los Alamos National Laboratory [@refANL] and Paul Scherrer Institut [@refPSI].
Application of the inelastic neutron scattering due to phonon excitations in superfluid helium is expected to further improve the UCN density since the UCN loss due to the absorption and up-scattering can be sufficiently suppressed below the temperature of about $1$ K although its down-conversion rate is relatively smaller than that of the solid deuterium. Superfluid helium is adopted as UCN sources under construction at TRIUMF [@refTRIUMF], in preparation at the SNS [@refSNS], under design study at the PNPI [@refPNPI]. The down-conversion into the UCN region is dominated by the single phonon exitation induced by very-cold neutrons with the velocity $v_n$ of around $447$ m s$^{-1}$ (the wavelength of around $0.89$ nm). Therefore, UCN yield is approximately proportional to the neutron flux at $v_n \sim 447 ({\rm m}\,{\rm s}^{-1})$ and the TRIUMF and PNPI sources are designed to maximize the coupling efficiency among the neutron production target, cold moderator and the superfluid helium converter.
The SNS source is designed to accept a pulsed cold neutron beam into a superfluid helium. Only neutrons with the velocity of around $447$m/s are contributing to the UCN yield while other neutrons simply transmit the superfluid helium.
In this paper, we discuss an application of neutron accelerator/decelerator, that was demonstrated to achieve the UCN time focusing [@refRebuncher], to concentrate neutrons into the $v_n \sim 447 ({\rm m}\,{\rm s}^{-1})$ region by selectively decelerate faster neutrons and accelerate slower neutrons on the beam transportation to the superfluid helium with a series of accelerator/decelerator units synchronized with the neutron time-of-flight.
Concept of neutron velocity concentrator
========================================
We denote the longitudinal flight path length as $z$. We consider the case where neutrons are transported through a magnetic field using a neutron guide and the transverse velocity is sufficiently smaller than the longitudinal velocity, where the neutron motion can be described as a one-dimensional problem along the $z$-axis as schematically illustrated in Fig. \[fig\_pot\]. We assume that the magnetic field is sufficiently strong so that the polarity of the neutron spin about the local magnetic field is adiabatically transported. In general, the neutron polarity under a magnetic field $\bm{B}$ can be flipped by applying an rf-field with the frequency of $\nu=2| \mu \bm{B}|/h$, where $\mu$ is the neutron magnetic moment. If the incident neutron spin polarity is positive, the relation between the kinetic energy of the incident neutron ($E_{\rm in}$) and that of the exit neutron ($E_{\rm ex}$) is $
E_{\rm ex} = E_{\rm in} - h\nu,
$ and the neutron is decelerated by $h\nu$. On the other hand, negative polarity neutron is accelerated by $h\nu$.
Here we consider the case where a main rf-cavity is installed in a strong-field region on the neutron beam path and an auxiliary rf-cavity in a weak-field region as shown in Fig. \[fig\_pot\]. The frequency of the main rf-cavity ($\nu_0$) is significantly larger than that of the auxiliary rf-cavity ($\nu_1$). We refer to this device as a “spin flipper unit” below. The spin flipper unit decelerates positive polarity neutrons by $h\nu_0$ and it accelerates negative polarity neutrons by $h\nu_0$. The spin polarity is flipped to the original polarity in the auxiliary rf-cavity.
![Schematic view of a spin flipper unit. (a) A neutron with positive polarity is decelerated. (b) A neutron with negative polarity is accelerated. []{data-label="fig_pot"}](mag_pot.eps){width="10cm"}
The rf-power is applied for a certain duration time earlier than $z_i/v_{\rm f}$ when neutrons faster than $v_{\rm f}$ arrive $i$-th spin flipper unit. Positive polarity neutrons with the velocity $v > v_{\rm f}$ are decelerated and negative polarity ones accelerated.
Figure \[fig\_position\] illustrates a series of identical spin flipper units placed with an equal spacing to which we refer as the “neutron velocity concentrator”. We denote the distance between the neutron source and the main rf-cavity, at which spin polarity is flipped, of $i$-th spin flipper unit as $z_i$. We consider to concentrate neutron velocity distribution around a specific velocity which is referred to as the target velocity $v_{\rm target}$ below. The rf-power of $i$-th spin flipper unit is turned off at the time of $t=z_i/v_{\rm target}$ when neutrons with the velocity $v_{\rm target}$ arrive.
![Schematic drawing of the neutron velocity concentrator and the timing chart of rf-power application.[]{data-label="fig_position"}](position.eps){width="13cm"}
If the first spin flipper unit (SFU 1) is on at the time $z_1/v$ and the $\nu_0$ satisfies Eq. \[eq:single\], the velocity of positive polarity component is changed to $v_{\rm target}$ in the SFU 1. $$\frac{mv_{\rm target}^2}{2} = \frac{mv^2}{2}-h\nu_0
\label{eq:single}$$ Much faster neutrons can be decelerated in the vicinity of $v_{\rm target}$ by the successive deceleration through SFU 1, SFU 2, …. In the same way, slower neutrons can be accelerated into the vicinity of $v_{\rm target}$. Selecting the timing of the rf-power application, neutron velocity distribution can be concentrated into the vicinity of $v_{\rm target}$ as shown in Fig. \[fig\_phasespace\].
We note that the final polarity of concentrated neutrons can be arranged to be positive by turning off the auxiliary rf power of the last active spin flipper unit.
![Time evolution of the phase space distribution of neutrons. (1) Neutrons are generated as a pulse. (2) Neutrons spread spatially during transport. (3) Neutrons are accelerated or decelerated properly. (4) Velocity distribution in concentrated into the vicinity of the target velocity. Right box shows that of a neutron bunch in the accelerator region. This figure illustrates the behavior of polarized neutrons. []{data-label="fig_phasespace"}](phasespace2.eps){width="15cm"}
A design of neutron velocity concentrator for superthermal ultracold source with superfluid helium {#sec_simulation}
==================================================================================================
We discuss a practical design for concentrating the neutron velocity distribution into the vicinity of $v_{\rm n} \sim 447 \,({\rm m}\,{\rm s}^{-1})$ where the down-conversion into the ultracold region via phonon excitations in superfluid helium takes place. We employ a superconducting compact solenoid as the electromagnet of each spin flipper unit. The numerical calculation tells us that the field strength can be $7.5$ T and the length of the spin flipper unit can be $30$ cm as shown in Fig.\[fig\_scm\]. Corresponding rf frequency for the spin flip is about $200$ MHz and resulting neutron energy change is about $0.9$ $\mu$eV.
![Left: Conceptual design of spin flipper unit. Right-Top: A model of superconducting solenoid magnet with the symmetry about $z=0$ plane and the symmetry around $z$ axis. Gray areas represent superconducting solenoid coil. Generated flux lines are also shown. Curves represent lines of magnetic force. Right-Bottom: Calculated magnetic field strength. []{data-label="fig_scm"}](mag_unit.eps "fig:"){width="8cm"} ![Left: Conceptual design of spin flipper unit. Right-Top: A model of superconducting solenoid magnet with the symmetry about $z=0$ plane and the symmetry around $z$ axis. Gray areas represent superconducting solenoid coil. Generated flux lines are also shown. Curves represent lines of magnetic force. Right-Bottom: Calculated magnetic field strength. []{data-label="fig_scm"}](mag_cal.eps "fig:"){width="6cm"}
We employ a series of sixty identical spin flipper units as shown in Fig.\[fig\_beamline\]. We simulated the transport of pulsed neutrons by using the Monte-Carlo method. We put the target velocity as $v_{\rm target} = 447$ m s$^{-1}$. Since the maximum energy change with the sixty successive spin flipper units amounts $54$ $\mu$eV, the initial velocity of neutrons, which can be concentrated to the target velocity, distributes in the velocity region of $435$ to $447$ m s$^{-1}$. Neutrons in this velocity range with unpreferred spin can be diluted into the wider velocity range of $410$ to $490$ m s$^{-1}$. Thus we simulated the change of the velocity distribution in the velocity range of $410$ to $490$ m s$^{-1}$. For simplicity, we assumed that the velocity distribution is flat and the time width of the neutron pulse is negligibly small.
![Beamline setup for simulation. Sixty flipper units were set on the beamline at the position of 10 m from the source.[]{data-label="fig_beamline"}](beamline.eps){width="13cm"}
Figure \[fig\_nvc1\] shows the result of the simulation. Unpolarized neutrons in the velocity range shown in Fig.\[fig\_nvc1\] (a) was put in the initial condition. As shown in Fig.\[fig\_nvc1\] (b), a sharp peak appears at the target velocity of $v_{\rm target}=447$ m s$^{-1}$ with the full width at the half maximum of $0.19$ m s$^{-1}$, which corresponds to the discrete energy change in individual spin flipper unit of $0.85$ $\mu$eV. The neutron intensity is 50 times enhanced compared with the initial intensity. Since the velocity concentrator is configured with the assembly of discrete length units, some of neutrons with unpreferred spin are also concentrated as shown in Fig. \[fig\_nvc1\] (c). The resulting neutron polarization is almost 100% as shown in Fig.\[fig\_nvc1\] (d). In case polarized neutrons would be used, the net intensity enhancement is 100 in the vicinity of the target velocity.
![Result of simulation of the velocity concentration. (a) velocity distribution of flux of incident neutrons. (b) exit flux of preferred spin component. (c) exit flux of unpreferred spin component. (d) polarization of exit neutrons. []{data-label="fig_nvc1"}](nvc1.eps){width="14.5cm"}
The spin flip probability was assumed to be 100% in the above result. However, the practical flipper unit might have an incomplete spin flip probability. The adverse effect of the incomplete spin flip probability was also simulated as shown in Fig.\[fig\_nvc2\]. In these simulations, we ignored the quantum-mechanical phase relation between positive- and negative-polarity components and we assumed that positive- and negative-polarity neutrons contributes incoherently. The velocity concentration can be expected even with the spin flip efficiency as low as 90% .
![The flux of exit neutrons from the incomplete flipping probability. $P$ is spin-flipping probability.[]{data-label="fig_nvc2"}](nvc2.eps){width="14.5cm"}
Discussion
==========
The concept of the neutron velocity concentration was introduced with a possible design for the application to the enhancement of UCN production via the superthermal down-conversion in superfluid helium. A possible enhancement with the velocity width of $0.19$ m s$^{-1}$ was estimated to be as large as 100. We note that larger enhancement in narrower velocity width is achievable by employing additional spin flipper rf-cavities and/or frequency modulation synchronized with the neutron time-of-flight. The optimum velocity width should be considered
This technique can be applied to the nEDM measurement in the superfluid helium UCN converter located at the end of the neutron guide. We note that the individual spin flipper units change the velocity distribution step by step and even smaller number of spin flipper units enables corresponding gain of UCN production. We also note that the control of neutron energy with the high frequency spin flipper was successfully demonstrated [@refRebuncher] and the fast switching of rf flippers synchronized with the neutron time-of-flight is already practically applied in J-PARC [@refSFC]. These achievements indicate that the neutron velocity concentration is feasible within the existing technologies.
[9]{}
C.A. Baker, et al., Phys. Rev. Lett. [**97**]{} 131801 (2006).
A. P. Serebrov, Phys. Rev. C [**82**]{}, 035501 (2010).
R. Golub, D.J. Richardson, S.K. Lamoreaux, Ultracold Neutrons, Adam Hilger, Bristol, 1991.
A. Saunders, et al., Phys. Lett. B [**593**]{} 55 (2004).
A. Anghel, et al., Nucl. Instr. and Meth. A [**611**]{} 272 (2009).
Y. Masuda, et al., Phys. Rev. Lett. [**89**]{} 284801 (2002).
N. Fomin, et. al., Nucl. Instr. and Meth. A [**773**]{}, 45-51 (2015).
A. P. Serebrov, A.P. et al., Nucl.Instr. and Meth. A [**611**]{} 276 (2009).
Y. Arimoto, et al., Phys. Rev., A [**86**]{}, 023843 (2012).
K. Taketani, et. al., Nucl. Instr. and Meth. A [**634**]{} 134 (2011).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Nonlinear localized excitations in one-dimensional diatomic lattices with cubic and quartic nonlinearity are considered analytically by a quasi-discreteness approach. The criteria for the occurence of asymmetric gap solitons(with vibrating frequency lying in the gap of phonon bands) and small-amplitude, asymmetric intrinsic localized modes(with the vibrating frequency being above all the phonon bands) are obtained explicitly based on the modulational instabilities of corresponding linear lattice plane waves. The expressions of particle displacement for all these nonlinear localized excitations are also given. The result is applied to standard two-body potentials of the Toda, Born-Mayer-Coulomb, Lennard-Jones, and Morse type. The comparison with previous numerical study of the anharmonic gap modes in diatomic lattices for the standard two-body potentials is made and good agreement is found.'
address: |
$^{1}$Max-Planck-Institut für Physik komplexer Systeme, Nöthnitzer Strasse 38,\
D-01187 Dresden, Germany\
$^{2}$Center for Nonlinear Studies and Department of Physics, Hong Kong Baptist University,\
Hong Kong, China\
$^{3}$Department of Physics, East China Normal University, Shanghai 200062, China\
$^{4}$Department of Physics, University of Houston, Houston TX 77204, USA
author:
- 'Guoxiang Huang$^{1,2,3}$ and Bambi Hu$^{2,4}$'
title: |
Asymmetric gap soliton modes in diatomic lattices\
with cubic and quartic nonlinearity
---
=6.5truein =-0.1truein =9.0truein =-0.2truein =0.1pt =0.1pt = =0.8cm
INTRODUCTION
============
The study of the dynamics of nonlinear lattices and related solitonic excitations has greatly influenced by the pioneering works of Fermi, Pasta, and Ulam,[@fer] and of Zabusky and Kruskal. [@zab1] Most of the work in this area has focused on models of one-dimensional(1D) monatomic chains with simple interatomic potentials of polynomial [@zab2; @col; @pne1; @fly1; @fly2], which can approximate any realistic potential near the equilibrium separation distance of two atoms. This description, usually done in a continuum limit, is only valid for a zone-boundary phonon mode, [*i.e*]{}. for $q$, the wavenumber of lattice waves, being near zero or $\pi/d_0$, where $d_0$ is lattice spacing. In 1972, Tsurui[@tsu] proposed an analytical method for studying the nonlinear excitations of lattices valid in the whole Brillouin zone(BZ). Later on this approach was extended by Remoissenet[@rem] and Huang.[@hua1; @hua2] Exact analytical solutions for the nonlinear localized excitations in 1D monatomic lattices can be obtained only for the Toda[@tod] and Ablowitz and Ladik[@abl] lattices, which are discrete completely integrable systems.
In recent years, the interest in localized excitations in nonlinear lattices has been renewed due to the identification of a new type of anharmonic localized modes.[@hua1; @dol; @sie1; @pag; @bur; @kiv1; @san; @fla1; @kis1; @sie2; @fla2] These modes, called the intrinsic localized modes(ILM’s),[@sie1] or the discrete breathers,[@fla2] are the discrete analog of the envelope(or breather) solitons with their spatial extension being only of a few lattice spacing and the vibrating frequency lying above the upper cutoff of phonon bands.[@yos] Experimentally, the ILM’s have been observed in coupled pendulum lattices[@che1] and electrical lattices.[@mar] The quantum mechanical aspects of the ILM’s have also been considered.[@shi; @hua3; @ros] However, examination of the 1D lattices with standard Toda, Born-Mayer-Coulumb, Lennard-Jones, and Morse two-body interatomic potentials demonstrates that the ILM’s do not appear above the top of the plane wave spectrum. The physical reason for this is that the cubic nonlinearity in the Taylor expansion of these realistic potentials is too strong. One of the effects of the cubic nonlinearity is that increasing the magnitude of the cubic term makes the potentials softer and hence decreases the localized mode frequency. The localized mode is destroyed as it approaches the bounding plane wave spectrum.[@hua1; @bic]
Recently, much attention has been paid to the gap solitons in nonlinear diatomic lattices.[@hua2; @shi; @kiv2; @chu; @kis2; @kis3; @fra; @aok; @bon; @tei; @kon; @hua4; @kis4] The concept of the gap solitons was first introduced by Chen and Mills[@che2] when investigating the nonlinear optical response of superlattices. For a diatomic lattice, the phonon spectrum of the system consists of two branches(acoustic and optical ones), induced by mass or force-constant difference of two kinds of particles. Due to nonlinearity gap soliton modes may appear as localized excitations with vibrating frequency being in the gap of the linear spectrum. Since the gap solitons occur in perfect lattices with discrete translational symmetry, a name “anharmonic gap mode” or “intrinsic gap mode(IGM)” was given by Sievers and his collaborators.[@kis2; @kis3; @kis4] It is possible that the ILM’s and the IGM’s may be created experimentally in diatomic lattices. Refs.43 and 44 reported some experimental studies of the gap solitons, resonant kinks and the ILM’s in a damped and parametrically excited 1D diatomic pendulum lattices.
Since the standard two-body potentials of the Toda, Born-Mayer-Coulomb, Lennard-Jones, and Morse type have a strong cubic nonlinearity in their Taylor expansion near equilibrium position, it is therefore necessary to consider the nonlinear excitations in the diatomic lattices with cubic and quartic anharmoniticity. In their recent contributions, Kiselev [*et al*]{}.[@kis2; @kis3] investigated the anharmonic localized modes in 1D diatomic lattices with above mentioned two-body potentials. By using a rotating wave approximation combined with computer simulation, they showed that an ILM does [*not*]{} exist and a nonlinear optical lower cutoff gap mode([*i.e*]{}. IGM) is general feature of these diatomic lattices. Lately, Franchini [*et al*]{}.[@fra] numerically found that for a potential with the cubic and quartic nonlinearity there exists a “critical” $K_3$(cubic force constant in the potential) value. For small $K_3$, nonlinear optical and acoustic upper cutoff localized modes occur, while for large $K_3$ these modes disappear and a nonlinear optical lower cutoff mode rises. In a recent work, Bonart [*et al*]{}.[@bon1] investigated the boundary condition effects in the diatomic lattices with cubic and quartic anharmonicity. Based on a rotating wave approximation they gave existence criteria for the ILM’s and IGM’s, which are related to the stability properties of linear optical upper and lower cutoff phonon modes. These studies posed an interesting problem of how to provide an analytical approach which can give not only the explicit criteria for the existence of the ILM’s and the IGM’s as well as other possible nonlinear eacitations for both optical and acoustic branches but also the approximate analytical expressions for these nonlinear excitations in a unified way. It is this problem that will be addressed here.
There are several theoretical methods to study the nonlinear localized excitations in diatomic lattices(see Ref.11 and references therein). In this paper we use the quasi-discreteness approach(QDA) for diatomic lattices[@hua2] to investigate the ILM’s and IGM’s as well as kink-like excitations with small-amplitude in 1D diatomic lattices with cubic and quartic nonlinear interactions between their nearest- neighbor particles. The paper is organized as follows. In Sec.II, the model is introduced and an asymptotic expansion based on the QDA is made for the equations of motion. By using the results obtained in Sec.II, in Sec.III we discuss the solutions of the ILM’s and IGM’s in a simple and unified way. Some explicit criteria and expressions of particle displacement of the ILM’s and the IGM’s are also given in this section. In Sec.IV we apply our results to the standard two-body potentials from Toda to Morse type and make a comparison with existed numerical experiments. Finally, Sec.V contains a discussion and summary of our results.
MODEL AND ASYMPTOTIC EXPANSION
==============================
The model
---------
We consider a 1D diatomic lattices with a nearest-neighbor interaction between particles. The restriction to the nearest-neighbor interaction is for simplicity and the approach can be easily extended to second and higher neighbors. The Hamiltonian of the system is given by H=\_i, where $u_i=u_i(t)$ is the displacement from its equilibrium position of the $i$th particle with mass $m_i
=m\delta_{i,2k}+M\delta_{i,2k+1}\,(M>m, k$ is an integer). The potential $V(r)$ is quite general, typically it can be the standard two-body potentials of the Toda, Born-Mayer-Coulomb, Lennard-Jones and Morse type(for their detailed expressions, see Sec.IV below). We focus on displacements with smaller amplitude which can be detected experimentally without introducing reconstruction or phase transitions in the system. This allows us to Tayler expand the potential $V(r)$ at the equilibrium position $r=0$ in a power series of the displacements to four order.[@fra] Thus we obtain an approximate $K_2$-$K_3$-$K_4$ potential V(r)=K\_2r\^2+K\_3 r\^3+K\_4 r\^4, where $K_2(>0)$, $K_3$ and $K_4(>0)$ are harmonic, cubic and quartic force constants, respectively. We assume that the basic features of the weakly nonlinear localized excitations for the standard two-body potentials may be obtained by corresponding the $K_2$-$K_3$-$K_4$ potentials. Then the Hamiltonian (1) takes the following form H=\_[i]{}. Since each of the standard two-body potentials mentioned above has only one minimum, we assume that for the $K_2$-$K_3$-$K_4$ potential (2) there is the constraint <4, unless there are two minima([*i.e*]{}. double well potential) hence the system may admit some types of nonlinear excitations which will not be discussed here.
If we write $u_{2k}=v_n$(even particles) and $u_{2k+1}=w_n$(odd particles), $n$ is the index of the $n$th unit cell with a lattice spacing $d=2d_0$, $d_0$ is the equilibrium distance between two adjacent particles, the system can be split into two sublattices. The equations of motion for $v_n$ and $w_n$ are & & mv\_n= K\_2(w\_n+w\_[n-1]{}-2v\_n) +K\_3\[(w\_n-v\_n)\^2-(w\_[n-1]{}-v\_n)\^2\]\
& & +K\_4\[(w\_n-v\_n)\^3+(w\_[n-1]{}-v\_n)\^3\],\
& & Mw\_n= K\_2(v\_n+v\_[n+1]{}-2w\_n) -K\_3\[(v\_n-w\_n)\^2-(v\_[n+1]{}-w\_n)\^2\]\
& & +K\_4\[(v\_n-w\_n)\^3+(v\_[n+1]{}-w\_n)\^3\]. The linear dispersion relation of Eqs.(5) and (6) is \_(q)={I\_2+J\_2\^}\^, where $I_2=K_2/m$ and $J_2=K_2/M$. The minus(plus) sign corresponds to acoustic(optical) mode. At wave number $q=0$ the eigen frequency spectrum has a lower cutoff $\om_{-}(0)=0$ for the acoustic mode and an upper cutoff $\om_{+}(0)\equiv \om_{3}=[2(I_2+J_2)]^{1/2}$ for the optical mode. At $q=\pi/d$ there exists a frequency gap between the upper cutoff of the acoustic branch, $\om_{-}(\pi/d)\equiv \om_1=\sqrt{2J_2}$, and the lower cutoff of the optical branch, $\om_{+}(\pi/a)\equiv \sqrt{2I_2}$. The width of the frequency gap is $\sqrt{2I_2}-\sqrt{2J_2}=\sqrt{2K_2}(1/\sqrt{m}-1/\sqrt{M})$. In linear theory, the amplitudes of lattice waves are constants and linear waves can not propagate and will be damped when $\om$(the frequency of the waves) lies in the regions $\om_1<\om<\om_2$ and $\om>\om_3$. Accordingly, these regions are the “forbidden bands” of the linear waves. This property of the eigen frequency spectrum results from the discreteness of the system([*i.e*]{}. discrete translational symmery). However, when the nonlinearity in Eqs.(5) and (6) is considered, the above conclusions are no longer valid. As localized excitations, some nonlinear modes may appear, whose oscillatory frequencies can lie in these forbidden bands of the phonon spectrum.
Asymptotic expansion
--------------------
We use the QDA for diatomic lattices developed in Ref.11 to investigate the effects of nonlinearity and discreteness of the system. In this treatment one sets u\_n(t)=u\^[(1)]{}\_[n,n]{}+\^2 u\^[(2)]{}\_[n,n]{} +\^3 u\^[(3)]{}\_[n,n]{}+\
where $\ep$ is a smallness and ordering parameter denoting the relative amplitude of the excitation and $u_{n,n}^{(\nu)}=u^{(\nu)}(\xi_n,\tau;\phi_n)$. $\xi_n=º\ep (na-\la t)$ and $\tau=\ep^2t$ are two multiple-scales variables(slow variables). $\la$ is a parameter to be determined by a solvability condition. The “fast” variable, $\phi_n=qnd-\om (q)t$, representing the phase of carrier wave, is taken to be completely discrete. Substituting (8) into (5) and (6) and comparing the power of $\ep$, we obtain a hierarchy of equations about $v_{n,n}^{(j)}$ and $w_{n,n}^{(j)}$($j=1,\,2,\,3,\cdots$): & & v\_[n,n]{}\^[(j)]{}-I\_2(w\_[n,n]{}\^[(j)]{}+w\_[n,n-1]{}\^[(j)]{}-2v\_[n,n]{}\^[(j)]{}) =M\_[n,n]{}\^[(j)]{} with & & M\_[n,n]{}\^[(1)]{}=0,\
& & M\_[n,n]{}\^[(2)]{}=2v\_[n,n]{}\^[(1)]{}-I\_2dw\_[n,n-1]{}\^[(1)]{}+I\_3(w\_[n,n]{}\^[(1)]{}-v\_[n,n]{}\^[(1)]{})\^2 -I\_3(w\_[n,n-1]{}\^[(1)]{}-v\_[n,n]{}\^[(1)]{})\^2,\
& & M\_[n,n]{}\^[(3)]{}=2v\_[n,n]{}\^[(2)]{} -(2+\^2 ) v\_[n,n]{}\^[(1)]{}+I\_2 ( -d w\_[n,n-1]{}\^[(2)]{} +w\_[n,n-1]{}\^[(1)]{})\
& & +2I\_3(w\_[n,n]{}\^[(1)]{}-v\_[n,n]{}\^[(1)]{}) (w\_[n,n]{}\^[(2)]{}-v\_[n,n]{}\^[(2)]{}) -2I\_3 (w\_[n,n-1]{}\^[(1)]{}-v\_[n,n]{}\^[(1)]{}) (w\^[(2)]{}\_[n,n-1]{}-v\^[(2)]{}\_[n,n]{} -dw\^[(1)]{}\_[n,n-1]{})\
& & +I\_4 \[ (w\^[(1)]{}\_[n,n]{}-v\^[(1)]{}\_[n,n]{})\^3 + (w\^[(1)]{}\_[n,n-1]{}-v\^[(1)]{}\_[n,n]{})\^3\],\
& & and & & w\_[n,n]{}\^[(j)]{}-J\_2(v\_[n,n]{}\^[(j)]{}+v\_[n,n+1]{}\^[(j)]{} -2w\_[n,n]{}\^[(j)]{}) =N\_[n,n]{}\^[(j)]{} with & & N\_[n,n]{}\^[(1)]{}=0,\
& & N\_[n,n]{}\^[(2)]{}=2w\_[n,n]{}\^[(1)]{}+J\_2dv\_[n,n+1]{}\^[(1)]{}-J\_3(v\_[n,n]{}\^[(1)]{}-w\_[n,n]{}\^[(1)]{})\^2 +J\_3(v\_[n,n+1]{}\^[(1)]{}-w\_[n,n]{}\^[(1)]{})\^2,\
& & N\_[n,n]{}\^[(3)]{}=2w\_[n,n]{}\^[(2)]{} -(2+\^2 ) w\_[n,n]{}\^[(1)]{}+J\_2 ( d v\_[n,n+1]{}\^[(2)]{} +v\_[n,n+1]{}\^[(1)]{})\
& & -2J\_3(v\_[n,n]{}\^[(1)]{}-w\_[n,n]{}\^[(1)]{}) (v\_[n,n]{}\^[(2)]{}-w\_[n,n]{}\^[(2)]{}) +2J\_3 (v\_[n,n+1]{}\^[(1)]{}-w\_[n,n]{}\^[(1)]{}) (v\^[(2)]{}\_[n,n+1]{}-w\^[(2)]{}\_[n,n]{} +dv\^[(1)]{}\_[n,n+1]{})\
& & +J\_4 \[ (v\^[(1)]{}\_[n,n]{}-w\^[(1)]{}\_[n,n]{})\^3 + (v\^[(1)]{}\_[n,n+1]{}-w\^[(1)]{}\_[n,n]{})\^3\],\
& & which can be solved order by order. In Eqs.(9)-(16), we have defined $I_{i}=K_{i}/m$ and $J_{i}=K_{i}/M$($i=2,\,3,\,4$). The expressions of $M_{n,n}^{(j)}$ and $N_{n,n}^{(j)}$($j=4,\,5,\cdots$) need not be written down explicitly here.
Amplitude equations for acoustic and optical modes
--------------------------------------------------
In order to avoid possible divergence for zone-boundary phonon modes, we solve the acoustic and optical modes separetely. First we consider the low-frequency acoustic mode of the system. For this we rewrite Eqs.(9) and (13) in the form & & w\_[n,n]{}\^[(j)]{}=J\_2( M\_[n,n]{}\^[(j)]{}+M\_[n,n+1]{}\^[(j)]{} ) +(+2I\_2)N\_[n,n]{}\^[(j)]{},\
& & (+2I\_2)v\_[n,n]{}\^[(j)]{}=I\_2 ( w\_[n,n]{}\^[(j)]{}+w\_[n,n-1]{}\^[(j)]{})+M\_[n,n]{}\^[(j)]{}, where the operator $\hat{L}$ is defined by u\_[n,n]{}\^[(j)]{}=(+2I\_2) (+2J\_2) u\_[n,n]{}\^[(j)]{}-I\_2J\_2( u\_[n,n-1]{}\^[(j)]{}+ u\_[n, n+1]{}\^[(j)]{}+2u\_[n,n]{}\^[(j)]{}) with $u_{n,n}^{(j)}(j=1,\,2,\,3,\cdots)$ a set of arbitrary functions. For $j=1$ it is easy to get & & w\_[n,n]{}\^[(1)]{}=F\_[10]{}(,\_n)+\[F\_[11]{}(,\_n)[e]{}\^[i\_[n]{}\^[-]{}]{} +[c.c.]{}\],\
& & v\_[n,n]{}\^[(1)]{}=F\_[10]{}(,\_n) -with $\phi_n^{-}=qnd-\om_{-}(q)t$. $\om_{-}(q)$ has been given in Eq.(7) with a minus sign. The amplitude(or envelope) functions $F_{10}$ and $F_{11}$ are yet to be determined. $F_{10}$ is a real function representing the “direct current(dc)” part relative to the fast variable $\phi_n^{-}$ and $F_{11}$ is a complex amplitude of the “alternating current(ac)” part. If $K_3=0$, the dc part($F_{10}$) vanishes in this order. For $j=2$(the second order) a solvability condition determines $\la =V_{g}^{-}={\rm d}\om_{-}/{\rm d}q$([*i.e*]{}. the group velocity of the carrier waves) thus $\xi_n=\xi_n^{-}\equiv \ep (nd-V_{g}^{-}t)$. In the third order($j=3$), solvability conditions yields the evolution equations for $F_{10}$ and $F_{11}$: & & iF\_[11]{}+\_[-]{} F\_[11]{}+\_[-]{}F\_[11]{}F\_[10]{} +\_[-]{}|F\_[11]{}|\^[2]{}F\_[11]{}=0,\
& & \_[-]{}F\_[10]{} +\_[-]{}|F\_[11]{}|\^2=0. The detailed expressions of the coefficients in Eqs.(22) and (23) are given in Appendix A.
Second, we study the high-frequency optical mode excitations. In this case we recast Eqs.(9) and (13) into the form & & v\_[n,n]{}\^[(j)]{}=J\_2( N\_[n,n]{}\^[(j)]{}+N\_[n,n-1]{}\^[(j)]{} ) +(+2J\_2)M\_[n,n]{}\^[(j)]{},\
& & (+2I\_2)w\_[n,n]{}\^[(j)]{}=J\_2 ( v\_[n,n]{}\^[(j)]{}+v\_[n,n+1]{}\^[(j)]{})+N\_[n,n]{}\^[(j)]{}. By the same procedure of solving the acoustic mode given above, we obtain & & v\_[n,n]{}\^[(1)]{}=G\_[10]{}(,\_n)+\[G\_[11]{}(,\_n)[e]{}\^[i\_[n]{}\^[+]{}]{} +[c.c.]{}\],\
& & w\_[n,n]{}\^[(1)]{}=G\_[10]{}(,\_n) -with $\phi_n^{+}=qnd-\om_{+}(q)t$. The evolution equations for the amplitudes $G_{10}$(dc part of the optical mode) and $G_{11}$(complex amplitude for ac part of the optical mode) are given by & & iG\_[11]{}+\_[+]{} G\_[11]{}+\_[+]{}G\_[11]{}G\_[10]{} +\_[+]{}|G\_[11]{}|\^[2]{}G\_[11]{}=0,\
& & \_[+]{}G\_[10]{} +\_[+]{}|G\_[11]{}|\^2=0, where $\xi_{n}^{+}=\ep (nd-V_{g}^{+}t)$ with $\la =V_{g}^{+}={\rm d}\om_{+}/{\rm d}q$. The coefficients in Eqs.(28) and (29) are also given in the Appendix A.
Under the transformation & & F\_[10]{}=(1/)g\_[-]{},F\_[11]{}=(1/)f\_[-]{},\
& & G\_[10]{}=(1/)g\_[+]{},G\_[11]{}=(1/)f\_[+]{}, the nonlinear amplitude equations (22), (23), (28) and (29) can be written in the unified form & & if\_+\_ f\_+\_f\_g\_ +\_|f\_|\^[2]{}f\_=0,\
& & \_g\_ +\_|f\_|\^2=0, when returning to the original variables. In Eqs.(32) and (33), $x_{n}^{\pm}=nd-V_{g}^{\pm}t$ and plus(minus) sign corresponds to the optical(acoustic) mode, respectively.
Finally, we consider the acoustic mode at $q=0$. Noting that Eq.(32) for minus sign is invalid at $q=0$ for the description of nonlinear excitations since $\beta_{-}|_{q=0}=\gamma_{-}|_{q=0}=0$ and $\al_{-}|_{q=0}=\infty$. This breakdown is due to the fact that at $q=0$ an acoustic mode excitation is a long wavelength one. In this case a discrete long-wave approximation[@hua2] should be applied. By using the same technique used in Ref.11, for the acoutic mode at $q=0$ we obtain a long wavelength amplitude equation +P uu+Q u\^[2]{}u +H u=0, where $u=\pa A_{0}/\pa x_{n}$, $x_{n}=nd-ct$ with $c^{2}=K_{2}d^{2}/[2(M+m)]$. $A_{0}$ is the leading order approximation of $v_{n}$ and $w_{n}$. Since $\om_{-}(0)=0$, for the long wavelength acoustic mode there is no carrier wave. Thus the displacement of the lattice is purely a “direct current”. Eq.(34) without the second term $P\,\pa u/\pa x_n$ is standard modified Korteweg-de Vries(MKdV) equation. Thus (34) is a [*modified*]{} MKdV(MMKdV) equation. Its coefficients are given by & & P=()\^,\
& & Q= ()\^,\
& & H=()\^ .
ASYMMETRIC GAP SOLITONS, KINKS AND INTRINSIC LOCALIZED MODES
============================================================
When deriving the nonlinear amplitude equations (32)-(34) we have not used any kind of decoupling ansatz for the motion of two kinds of particles with different mass. This is one of the advantages of the QDA. On the other hand, the nonlinear amplitude equations, which are the reduced forms of the original equations of motion (5) and (6) for small-amplitude excitations, are valid in the whole BZ($-\pi/d<q\leq\pi/d$) except a zero-dispersion point for the optical phonon branch.[@for] Thus one can obtain the gap solitons and ILM’s as well as some possible new nonlinear excitations by solving the cutoff modes of the system in a simple and unified way.
(1). [**Optical upper cutoff mode**]{}. For the optical mode at $q=0$, we have $\om_{+}=\om_3=[2K_2(1/m+1/M)]^{1/2}$, $V_{g}^{+}=0$, $x_{n}^{+}=nd\equiv x_n$, $\al_{+}=-K_{2}d^{2}/[2(M+m)\om_{3}]$, $\beta_{+}=-K_{3}\om_{3}d/(2K_{2})$, $\ga_{+}=-3K_{4}\om_{3}(1+m/M)^2/(2K_2)$, $\de_{+}=K_{2}^{2}d^{2}/(Mm)$, and $\si_{+}=4K_{2}K_{3}d(1+m/M)^2/(Mm)$. Eqs.(32) and (33) with plus sign take the form & & i\_[+]{}+\_[+]{} \_[+]{} +\_[+]{}|\_[+]{}|\^[2]{}\_[+]{}=0,\
& & g\_[+]{}= -|\_[+]{}|\^2+C\_1, where $\ti{f}_{+}=f_{+}\exp (-i\beta_{+}C_{1}t)$ with $C_1$ an integration constant and \_[+]{}= & &\_[+]{}-\_[+]{}\
= & & (1+) ( -). Eq.(38) is standard nonlinear Schrödinger(NLS) equation. It has an uniform vibrating solution \_[+]{}=f\_[0]{}(i\_[+]{}|f\_[0]{}|\^[2]{}t),where $f_{0}$ is any complex constant, which corresponds to the linear optical upper cutoff phonon mode with a simply frequency shift $\ti{\ga}_{+}|f_{0}|^{2}$ and is a fixed point of the system. Note that it is possible to eliminate the time dependence by a simply transformation, justfying our use of the term “fixed point” for the uniform vibrating solution. In fact, the fixed point may also be written as \_[+]{}=f\_[0]{},where $f_0$ in this case is any real constant and $0\leq \bar{\phi}<2\pi$. In this sense there is a ring of fixed point characterized by the different values of the phase, $\bar{\phi}$. It is easy to show that, since $\al_{+}<0$, when $\ti{\ga}_{+}<0$ the fixed point is unstable by a long wavelength small perturbation. This kind of instability is due to a sideband modulation of the linear optical upper cutoff mode. The modulational instability for waves is called the Benjamin-Feir(BF) instability[@ben]. It is similar to the Eckhaus instability for patterns in extended dissipative systems out of equilibrium.[@eck] The BF instability in discrete lattices and related formation of solitonlike localized states were already discussed by Kivshar and Peyrard.[@kiv3] By this mechanism(usually called the Benjamin-Feir resonance mechanism[@stu]) a linear optical upper cutoff excitation will bifurcate, grow exponentially at first and then saturate due to the nonlinearity of the system. At later stage, a nonlinear localized mode— optical upper cutoff soliton is formed. In fact, for $\ti{\ga}_{+}<0$, [*e.g*]{}., <, Eq.(38) admit the envelope(breather) soliton solution \_[+]{}=()\^\_[0]{} [sech]{}\[\_0(x\_n-x\_[n\_0]{})\] \[-i|\_[+]{}|\_[0]{}\^[2]{}t-i\_0\], where $\eta_0$, $\phi_0$ and $x_{n_0}=n_0d$ are constants, $n_0$ is an arbitrary integer. Inequality (41) is just the condition of the modulational instability for the linear optical upper cutoff mode. From Eq.(39) we obtain g\_+=-\_0 , where the integration constant $C_1$ has chosen as zero as it corresponds to a constant displacement for all particles. In leading approximation the lattice configuration takes the form & & v\_[n]{}(t)=-\_0\
& & +2()\^\_[0]{} [sech]{}\[\_0(n-n\_0)d\] (\_[3s]{}t+\_0),\
& & w\_[n]{}(t)=-\_0\
& & -2()\^\_[0]{} [sech]{}\[\_0(n-n\_0)d\] (\_[3s]{}t+\_0), with \_[3s]{}=\_3+|\_[+]{}|\_[0]{}\^[2]{}, . the vibrating frequency of the localized mode is above the spectrum of the linear optical mode thus above the all phonon bands. Hence (44) and (45) represent an ILM accompanied by an asymmetric dc displacement due to the cubic anharmonicity of the system. We call it the small-amplitude [*asymmetric intrinsic localized mode*]{}.
From (44) and (45) we can see that the free parameter $\eta_0$ can be taken as an expansion parameter, [*i.e*]{}. $\eta_0=O(\ep)$. By (46) we have $\eta_0=[2(\Omega_{3s}-\omega_3)/|\al_+|]^{1/2}$. Thus in our approach, the expansion parameter $\ep$, used in (8), is propotional to the squareroot of frequency difference between the nonlinear localized mode and the linear cutoff phonon mode.
When $\ti{\ga}_{+}>0$, [*e.g*]{}., the inequality (41) takes opposite sign, the uniform vibrating solution of the NLS equation (38) is neutral stable. In this case Eq.(38) admits the dark soliton solution \_[+]{}=()\^\_[0]{} \[i|\_[+]{}|\_[0]{}\^[2]{}t-i\_0\]. From Eq.(39) we can obtain $g_{+}$ by integration. In this case we chose $C_1$ in such a way[@fly1] that $(\pa g_{+}/\pa x_n)|_{|x_n|=\infty}=0$. Then we have $C_1=\si_{+}|\al_{+}|\eta_{0}^{2}/(\de_{+}\ti{\ga}_{+})$. Hence we have g\_+=\_0 . The lattice displacement in this case takes the form & & v\_[n]{}(t)=\_0\
& & +2()\^\_[0]{} (\_[3k]{}t+\_0),\
& & w\_[n]{}(t)=\_0\
& & -2()\^\_[0]{} (\_[3k]{}t+\_0) with \_[3k]{}=\_3-\_[0]{}\^[2]{}. Since $\ga_{+}<0$, the vibrating frequency of the kink mode denoted by the expressions (49) and (50) is [*greater*]{} than $\om_3$. This is an example of a kink with the vibrating frequency above the all phonon bands due to the cubic nonlinearity of the system.
(2). [**Optical lower cutoff mode**]{}. For the optical mode at $q=\pi/d$(zone-boundary optical phonon mode), one has $\om_{+}=\om_2=\sqrt{2K_2/m}$, $V_{g}^{+}=0$, $x_{n}^{+}=x_n$, $\al_{+}=K_{2}d^{2}/[2\om_{2}(M-m)]$, $\beta_{+}=-K_{3}\om_{2}d/(2K_{2})$, $\ga_{+}=-3K_{4}\om_{2}/(2K_2)$, $\de_{+}=K_{2}^{2}d^{2}/(Mm)$ and $\si_{+}=4K_{2}K_{3}d/(Mm)$. In this case we have \_[+]{}= \_2(-). If $\tilde{\ga}_{+}>0$, [*i.e*]{}. >, \[cri2\] Eqs.(32) and (33) for plus sign have the solution & & f\_[+]{}=()\^\_[0]{} [sech]{}\[\_0(x\_n-x\_[n\_0]{})\] \[i\_[+]{}\_[0]{}\^[2]{}t-i\_0\],\
& & g\_+=-\_0 . The lattice configuration takes the form & & v\_[n]{}(t)= -\_0\
& & +(-1)\^[n]{}2()\^\_[0]{} [sech]{}\[\_0(n-n\_0)d\] (\_[2s]{}t+\_0),\
& & w\_[n]{}(t)= -\_0 with \_[2s]{}=\_2-\_[+]{}\_[0]{}\^[2]{}, lying in the frequency gap of the phonon spectra between the acoustic and the optical modes. It is a typical asymmetric nonlinear gap mode, existing in the diatomic lattices when the condition (53) is satisfied. We note that for this mode the displacement of the heavy particles only has a kink-like asymmetric dc part. But the displacement of the light particles, besides the same type of dc part, has an additional “staggered” vibrational part([*i.e*]{}. “staggered” envelope soliton). We call it the [*asymmetric optical lower cutoff gap soliton*]{}. The vibrating frequency $\Omega_{2s}$ has the parabola relation with respect to the wave amplitude, denoted by the parameter $\eta_0$. The formation of the asymmetric nonlinear gap mode is also the conclusion of the BF instability for the corresponding linear optical lower cutoff phonon mode. A further discussion for such nonlinear modes in the realistic potentials is given in the next section.
(3). [**Acoustic upper cutoff mode**]{}. For the acoustic mode at $q=\pi/d$(zone-boundary acoustic phonon mode), one has $\om_{-}=\om_1=\sqrt{2K_2/M}$, $V_{g}^{-}=0$, $x_{n}^{-}=x_n$, $\al_{-}=-K_{2}d^{2}/[2\om_{1}(M-m)]<0$, $\beta_{-}=-K_{3}\om_{1}d/(2K_{2})$, $\ga_{-}=-3K_{4}\om_{1}/(2K_2)$, $\de_{-}=K_{2}^{2}d^{2}/(Mm)$ and $\si_{-}=4K_{2}K_{3}d/(Mm)$. Similarly one can obtain the equations like (38) and (39) with $g_{+}, \tilde{f_{+}}, \si_{+}, \de_{+}, \al_{+}$ and $\tilde{\ga}_{+}$ changed by $g_{-}, \tilde{f_{-}}, \si_{-}, \de_{-}, \al_{-}$ and $\tilde{\ga}_{-}$. Here \_[-]{}=\_[-]{}-\_[-]{}= \_1(-). If $\tilde{\ga}_{-}<0$, when $\tilde{\ga}_{-}<0$, [*i.e*]{}. <, \[cri3\] due to a BF instability of the corresponding linear upper cutoff acoustic mode an [*asymmetric acoustic upper cutoff gap soliton*]{} appears, with the vibrating frequency being in the gap of the phonon spectra between the acoustic and optical modes. Otherwise an acoustic upper cutoff kink vibrational mode occurs. We can readily written down the explicit expression of the lattice configuration in this case, but for saving space it is omitted here.
(4). [**Acoustic lower cutoff mode**]{}. This is a long wavelength mode without any carrier wave, because when $q=0$ we have $\om_{-}=0$ thus $\phi_{n}^{-}=0$. The lattice displacement only has dc part and its evolution is controled by the MMKdV equation, given by (34). A single-soliton solution of Eq.(34) is given by u=, where $\kappa$ is an arbitrary constant. Making the transformation $u=-P/(2Q)+U(\zeta_{n},t)$ with $\zeta_{n}=x_n+P^2 t/(4Q)$, Eq.(34) becomes +QU\^2+H=0. It is the standard MKdV equation and can be solved by the inverse scattering transform. For the explict expressions of the kink and breather solutions of the MKdV equation we refer to Ref.11.
Needless to say, in addition to the cutoff modes considered above, our approach developed in Sec.II can also be used to discuss the nonlinear localized excitations for $q\neq 0$ and $q\neq \pi/d$([*i.e*]{}. the intra-band modes), which will be done elsewhere.
APPLICATION TO REALISTIC TWO-BODY NEAREST-NEIGHBOR POTENTIALS
=============================================================
In this section, we apply the general results obtained above to the diatomic lattices with the standard two-body nearest-neighbor potentials to see whether the anharmonic gap modes and the ILM’s can appear or not. This can be easily done by using the existence criteria given by (41)(for ILM’s), (\[cri2\])(for optical lower cutoff gap solitons), and (\[cri3\])(for acoustic upper cutoff gap solitons). Four standard interatomic potentials are:[@kis3]
1. [*Toda*]{}: V(r)=\^[-br]{}+ar-, where $a$ and $b$ are coefficients, such that $ab>0$. $r$ is the deviation of the relative interparticle separation from its zero-temperature equilibrium position.
2. [*Born-Mayer-Coulomb*]{}: V(r)=, where $\al_{M}$ is the Madelung constant, $q$ is the effective charge, $d_0$ is the zero-temperature equilibrium distance between adjacent particles, and $\rho_0$ is the constant describing the repulsion between atoms.
3. [*Lennard-Jones*]{}: V(r)=, where $\al$ is the constant determing the potential strength.
4. [*Morse*]{}: V(r)=P([e]{}\^[-ar]{}-1)\^[2]{}, where $P$ and $a$ are constants determing the strength and the curvature of the potential, respectively.
For weakly nonlinear excitations we can Taylor expand these potentials at their equilibrium position($r=0$) to obtain the force constants. They are defined by K\_[j]{}=()\_[r=0]{} with $j=2, 3, 4,\cdots$. Thus we have the values of $K_2$, $K_3$, $K_4$ and $K_3^2/(K_2K_4)$ which are given in Table 1. Notice that $K_3<0$ for these potentials except that the Toda one with $a$ and $b$ being both negative(note that we require that $K_2$ and $K_4$ are positive). The parameter $I$ for the Born-Mayer-Coulomb potential in Table 1 is defined by I=. Generally, we have $I>1$. For example[@san; @kis3; @bor], for KI, $d_0$=3.14Å, $\rho_0=0.26$Å, one has $I=1.1171$; for KBr, $d_0=3.29$Å, $\rho_0=0.334$Å, we have $I=1.1328$; for LiI, $d_0=3.0$Å, $\rho_0=0.374$Å, one has $I=1.1487$. By these results as well as the criteria given by (41), (53) and (60), we make the following conclusions:
(1). The condition (53) is satisfied by all these standard two-body potentials. Thus the asymmetric optical lower cutoff gap soliton mode given in the expressions (56) and (57) [*does*]{} exist in the diatomic lattices with the Toda, Born-Mayer-Coulomb, Lennard-Jones and Morse type interatomic interactions. This result agrees with the conclusion by numerical study presented by Kiselev [*et al*]{}.[@kis2; @kis3]
Shown in Fig.1 is the dc(denoted by the solid circles) and ac(denoted by the open circles) amplitude patterns of light particles for the optical lower cutoff gap soliton mode. The amplitude pattern for the heavy particles, which is not shown here, only has dc part similar to that of the dc part for the light particles. The Born-Mayer-Coulomb potential for KBr-like parameters is used with $m/M=39/80$. Comparing the panel (a) of Fig.1 in Ref.34 with our result shown in Fig.1 here, we see that the lattice configuration obtained analytically by our QDA is basically the same as the corresponding result given by the rotating wave approximation plus computer simulation, used by Kiselev [*et al*]{}.[@kis3] Furthermore, the parabola relation between the vibrating frequency and the amplitude of the nonlinear optical lower cutoff gap mode, given by (58), is in accordance well with the numerical results(see the Fig.3 of Ref.34).
Shown in Fig.2 is the amplitude(maximum absolute value) ratio $(|A_{dc}|/|A_{ac}|)$ of the dc part($|A_{dc}|$) to ac part($|A_{ac}|$) of the light particle displacement for the optical lower cutoff gap soliton mode. We see that the amplitude ratio is the function of the mass ratio $m/M$. The larger is $m/M$, the larger is the amplitude ratio. The amplitude ratio grows very fast as the mass ratio approaches 1.
(2). Notice that the existence criterion of an ILM(the optical upper cutoff soliton) is the inequality (41), [*i.e*]{}. the necessary condition for the occurence of the ILM is <. From Table 1 we see that in general $K_3^2/(K_2K_4)>1$. Thus for all the standard two-body potentials from Toda to Morse type, [*an ILM is impossible*]{}. This conclusion concides also with the numerical results of Refs.33 and 34.
(3). An acoustic upper cutoff gap soliton is not possible for all these standard two-body potentials, because its existence condition (60)(the same as (41)) can not be satisfied. However, acoutic and optical upper cutoff vibrating kinks are possible excitations for these diatomic lattice systems.
(4). Nonlinear acoustic lower cutoff excitations are long wavelength modes and are governed by the MMKdV equation (34). Since $Q$ and $H$ are positive and $P$ is negative(due to $K_3<0$) for all the two-body potentials(except the Toda one with $a$ and $b$ being both negative), the soliton amplitude(see (61)) in the presence of cubic nonlinearity ($K_3\neq0$) is smaller than the soliton amplitude in the absence of the cubic nonlinearity($K_3=0$).
(5). For general $K_2$-$K_3$-$K_4$ potentials, by the criteria given by (41), (53) and (60), we conclude that in weakly nonlinear approximation the “critical” value of $K_3^2/(K_2K_4)$ for the transition from an optical upper cutoff kink to an optical upper cutoff soliton(ILM) and for the occurence of an optical(acoustic) lower(upper) cutoff gap mode is 3/4, independent of $m/M$. Because one of the conditions (53) and (60) must be satisfied, we have the conclusion that [*for any nonlinear diatomic lattice, gap solitons always occur*]{}. Our analytical results support the numerical findings of Franchini [*et al*]{}.[@fra]: for small cubic nonlinearity, nonlinear optical and acoustic upper cutoff localized modes appear, while for large cubic nonlinearity a nonlinear optical lower cutoff mode rises. When $K_3=0$, the corresponding theoretical and numerical results for harmonic plus quartic potentials[@hua2; @fra] are recovered.
It is interesting to note that the criterion (41) and (60) for the occurence of the asymmetric optical and acoustic upper cutoff solitons in the diatomic lattices are the same as that for the appearence of the upper cutoff solitons in monatomic lattices, given by Tsurui[@tsu](Eq.(4.8), $\om_c^2=4$), Flytzanis [*et al*]{}.[@fly1](Eq.(5.5), $kD=\pi$) and Flach[@fla3](Eq.(3.24), $v_4=0$). This criterion was also given implicitly in Ref.10 since the envelope soliton solution (23) in Ref.10 requires ${\rm sgn}(PQ)>0$.
In addition, our results show analytically that for the potentials with cubic and quartic nonlinearity an nonlinear localized excitation always consists of dc and ac parts. The appearrence of the dc part is a direct conclusion of the asymmetry in the potentials. This fact is already known by using different approaches(see [*e.g*]{}. Refs.33 and 34).
Recently, Bonart, Rössler and Page[@bon1] considered the boundary condition effects in the diatomic lattice with cubic and quartic anharmonicity and gave existence criteria for the ILM’s and the IGM’s not restricted to small amplitudes by using a different approach. By comparison we can see that our instability threshold criterion for the linear optical upper cutoff phonon modes is in accordance with their corresponding one\[Eq.(10) in Ref.45)\] when exact to $A_{ac}^2$\[remenber that in our notation $A_{ac}$ is the amplitude of the ac part of the nonlinear excitation, [*e.g*]{}. in (44) we have $A_{ac}=2(\al_{+}/\tilde{\ga}_{+})^{1/2}\eta_0$\]. It is easy to show that our frequency formulas for the nonlinear optical (including upper and lower) cutoff modes, [*i.e*]{}. Eqs.(46) and (58), also coincide with the corresponding ones in Ref.45\[Eqs.(9) and (12)\] if the terms propotional to $A_{ac}^4$ are neglected. Our results support also the criterion (14) in Ref.45 for the appearance of nonlinear localized modes. In addition, our analytical approach based on QDA may provide much more insights such as for the occurance of nonlinear localized excitations in acoustic branch although it has the limitation of small amplitude.
DISCUSSION AND SUMMARY
======================
Based on the QDA for diatomic lattices we have [*analytically*]{} studied the nonlinear localized excitations with small amplitude in the diatomic lattices with cubic and quartic nonlinearity. The results are quite general and allows us directly to obtain many different types of nonlinear excitations in a unified way. Starting from the nonlinear amplitude equations given in (32) and (33), the existence criteria for the optical upper cutoff solitons and asymmetric gap soliton modes have been explicitly provided in (41), (53) and (60). The analytical expressions of particle displacements for all these nonlinear localized modes are also given. The theoretical results have been applied to the standard two-body nearest-neighbor potentials from Toda to Morse type and agreements well with the previous numerical findings have also been found.
Most of the existed analytical studies for the nonlinear excitations in diatomic lattices involved so-called “decoupling ansatz”, in which some relations were assumed between the displacement of light particles and that of heavy ones before solving the equations of motion(see Ref.11 and references therein). Pnevmatikos [*et al*]{}.[@pne2] investigated the soliton dynamics of nonlinear diatomic lattices by using the decoupling ansatz. For long wavelength([*i.e*]{}. $q=0$) excitations this can be done without big difficulty. But for envelope-type excitations and in the case of cubic nonlinearity, the concrete form of the decoupling ansatz is not easy to determine and the analytical calculation for the coefficients in amplitude equations, like (32) and (33), is also heavy. Except several numerical studies, such an analytical calculation has never been accomplished. Recently, it was shown that the decoupling ansatz is completely unnecessary and can be derived by the QDA[@hua2]. In addition, the QDA has many other advantages. For example, the results obtained by the QDA, though restricted to small amplitudes, are valid in the whole BZ except at the zero-dispersion point of the optical phonon branch( see Ref.46). Thus one can obtain all solutions for nonlinear cutoff and non-cutoff modes in a simple and unified way; the method is quite general and can be applied to the other lattice systems. An extension based on the QDA for magnetic gap soliton excitations in alternating Heisenberg ferromagnets has been given recently.[@hua5] In the present work, we use the QDA to consider the nonlinear localized excitations in the diatomic lattices with cubic and quartic nonlinearity. The detailed expressions of the coefficients in the amplitude equations (32) and (33) as well as (34) thus various types of nonlinear excitations are [*explicitly*]{} given.
The study of stability by using nonlinear amplitude equations is widely employed in pattern formation in systems out of equilibrium.[@stu; @cro] In present work, by a similar idea we have studied the stability of the linear optical and acoustic (upper and lower) cutoff phonon modes. For the $K_2$-$K_3$-$K_4$ potential, we have obtained the existence criteria for the occurance of linear cutoff phonon mode-related nonlinear localized excitations: If $K_{3}^2/(K_2K_4)<3/4$, we have the acoustic and optical upper cutoff solitons. Otherwise one has only the optical lower cutoff (gap) solitons. The later case happens for the standard two-body potentials from Toda to Morse type. The formation of the asymmetric gap solitons in the standard two-body potentials is the result of BF modulational instability of the optical lower cutoff phonon modes. Our results show that for any nonlinear diatomic lattice a gap soliton always occurs. The reason for this is that the curvatures of the acoustic and optical branches of the phonon spectra in the vicinity of the BZ edge have different signs. This means that if the optical mode at the BZ edge supports a lower cutoff gap soliton, the acoustic mode at the BZ edge supports an upper cutoff kink subject to the same sign of the nonlinearity, and vice vesa.
For large-amplitude excitations we should extend the present QDA to a higher-order approximation. It is expected that higher-order correction terms, likely $|f_{\pm}|^{4}f_{\pm}$ and $(\pa/\pa x_n^{\pm})|f_{\pm}|^4$(as well as some higher-order dispersion terms), will be added respectively into Eqs.(32) and (33). Then the existence criteria for the nonlinear localized excitations, [*i.e*]{}. (41), (53) and (60), will be modified to be amplitude-dependent.
The generation of an asymmetric IGM in nonlinear 3D diatomic lattices has been considered recently[@bon; @kis4]. These studies reveal the possibility of observing experimentally the IGM’s in real (3D) crystals. An macro-analogy of the diatomic lattices with cubic and quartic nonlinearity is to use a chain of magnetic pendulums. Russell [*et al*]{}.[@rus] reported moving breathers in such a system quite recently, but they did not pay attention to nonlinear cutoff modes. If the mass of the magnetic pendulums is arranged in a alternating way, one can obtain a diatomic lattice with an asymmetric intersite(dipole-dipole interaction) potential. For small amplitude excitations the intersite potential reduces to a $K_2$-$K_3$-$K_4$ one. Thus it is possible to observe the asymmetric IGM’s by using this macro-diatomic lattice system.
ACKNOWDGEMENTS {#acknowdgements .unnumbered}
==============
G.H. wishes to express his appreciation to Prof. Director P. Fulde for the warm hospitality received at the Max-Planck-Institut für Physik komplexer Systeme, where part of this work was carried out. It is a pleasure to thank Prof. A. J. Sievers for kindly sending a copy of Ref.21, Dr. S. Flach for the critical reading of the manuscript, Drs. M. Bär and A. Cohen, Profs. F. G. Mertens and M. G. Velarde for fruitful discussions. We also thank an anonymous referee for bringing our attention to Ref.45 and useful comments. This work was partly supported by the Hong Kong Research Grant Council and the Hong Kong Baptist University Faculty Research Grant.
APPENDIX A {#appendix-a .unnumbered}
==========
The detailed expressions of the coefficients of Eqs.(22) and (23) are given by $$\displaylines{
\de_{-}=I_2J_2d^2-2(I_2+J_2)\left[\fr{I_2J_2d\sin (qd)}
{2\om_{-}(I_2+J_2-\om_{-}^2)}\right ]^2,
\hfill \rlap{(A.1)}\cr
\si_{-}=4I_2I_3d\fr{\om_{-}^2-2J_2}{\om_{-}^2-2I_2}+
\fr{4J_2I_3d}{\om_{-}^2-2I_2}(\om_{-}^2+2I_2\cos qd)
+\fr{32I_2J_2I_3V_{g}^{-}\om_{-}\sin qd}
{(\om_{-}^2-2I_2)^2},
\hfill \rlap{(A.2)}\cr
\al_{-}=\fr{4(I_2+J_2-3\om_{-}^2)(V_{g}^{-})^2
-d^2[(\om_{-}^2-2I_2)(\om_{-}^2-2J_2)-2I_2J_2]}
{4\om_{-}(\om_{-}^2-I_2-J_2)},
\hfill \rlap{(A.3)}\cr
\beta_{-}=-\fr{I_3\om_{-}d}{2I_2},
\hfill \rlap{(A.4)}\cr
\ga_{-}=-\fr{3J_4\om_{-}}{2J_2}
\left [1+\fr{\om_{-}^2-2J_2}{J_2}\left(1+\fr{I_2}{\om_{-}^2-2I_2}\right )
\right ]
\hfill\cr
\qquad+\fr{\om_{-}^3º\sin^{2}(qd)}{(\om_{-}^2-2I_2)(I_2+J_2-\om_{-}^2)}
º\left[ -I_3L_{-}+\fr{J_3(I_3+I_2L_{-}\cos qd)}
{2\om_{-}^{2}-I_2}\right ],
\hfill \rlap{(A.5)}\cr
L_{-}=\fr{I_2J_3[\cos qd-(2\om_{-}^2-I_2)/J_2]}
{(2\om_{-}^{2}-J_2)(2\om_{-}^2-I_2)-I_2J_2\cos^2qd}.
\hfill \rlap{(A.6)}\cr
}$$ The coefficients of Eqs.(28) and (29) are given by $$\displaylines{
\de_{+}=I_2J_2d^2-2(I_2+J_2)\left[\fr{I_2J_2d\sin (qd)}
{2\om_{+}(I_2+J_2-\om_{+}^2)}\right ]^2,
\hfill \rlap{(A.7)}\cr
\si_{+}=4J_2J_3d\fr{\om_{+}^2-2I_2}{\om_{+}^2-2J_2}+
\fr{4I_2J_3d}{\om_{+}^2-2J_2}(\om_{+}^2+2J_2\cos qd)
+\fr{32I_2J_2J_3V_{g}^{+}\om_{+}\sin qd}
{(\om_{+}^2-2J_2)^2},
\hfill \rlap{(A.8)}\cr
\al_{+}=\fr{4(I_2+J_2-3\om_{+}^2)(V_{g}^{+})^2
-d^2[(\om_{+}^2-2J_2)(\om_{+}^2-2I_2)-2I_2J_2]}
{4\om_{+}(\om_{+}^2-I_2-J_2)},
\hfill \rlap{(A.9)}\cr
\beta_{+}=-\fr{J_3\om_{+}d}{2J_2},
\hfill \rlap{(A.10)}\cr
\ga_{+}=-\fr{3I_4\om_{+}}{2I_2}
\left [1+\fr{\om_{+}^2-2I_2}{I_2}\left(1+\fr{J_2}{\om_{+}^2-2J_2}\right )
\right ]
\hfill\cr
\qquad+\fr{\om_{+}^3º\sin^{2}(qd)}{(\om_{+}^2-2J_2)(I_2+J_2-\om_{+}^2)}
º\left[ -J_3L_{+}+\fr{I_3(J_3+J_2L_{+}\cos qd)}
{2\om_{+}^{2}-J_2}\right ],
\hfill \rlap{(A.11)}\cr
L_{+}=\fr{J_2I_3[\cos qd-(2\om_{+}^2-J_2)/I_2]}
{(2\om_{+}^{2}-I_2)(2\om_{+}^2-J_2)-I_2J_2\cos^2qd},
\hfill \rlap{(A.12)}\cr
}$$ where $\om_{\pm}(q)$ has been given in Eq.(7).
E. Fermi, J. Pasta, and S. Ulam, [*Los Alamos Nat. Lab. Report LA1940*]{}, 1955. Also in [*Collected Papers of Enrico Fermi*]{}(Univ. Chicago Press, Chicago, 1962), vol.2, p.978.
N. J. Zabusky and M. D. Kruskal, Phys. Rev. Lett. [**15**]{}, 240(1965).
N. J. Zabusky, Comput. Phys. Commun. [**50**]{}, 1(1973).
M. A. Collins, Chem. Phys. Lett. [**77**]{}, 342(1981).
St. Pnevmatikos, C. R. Acad. Sci. Paris II [**296**]{}, 1031(1983).
N. Flytzanis, St. Pnevmatikos, and M. Remoissenet, J. Phys. C: Solid State Phys. [**18**]{}, 4603(1985).
N. Flytzanis, St. Pnevmatikos, and M. Peyrard, J. Phys. A: Math. Gen. [**22**]{}, 783(1989) and references therein.
A. Tsurui, Prog. Theor. Phys. [**48**]{}, 1196(1972).
M. Remoissenet, Phys. Rev. B [**33**]{}, 2386(1986).
Guoxiang Huang, Zhu-Pei Shi, and Zaixin Xu, Phys. Rev. B [**47**]{}, 14561(1993).
Guoxiang Huang, Phys. Rev. B [**51**]{}, 12347(1995).
M. Toda, J. Phys. Soc. Jpn. [**22**]{}, 431(1967).
M. J. Ablowitz and J. F. Ladik, Stud. Appl. Phys. [**55**]{}, 213(1976).
A. S. Dolgov, Sov. Phys. Solid State [**28**]{}, 907(1986).
A. J. Sievers and S. Takeno, Phys. Rev. Lett. [**61**]{}, 970(1988).
J. B. Page, Phys. Rev. B [**41**]{}, 7835(1990).
V. M. Burlakov, S. A. Kiselev, and V. N. Pyrkov, Phys. Rev. B [**42**]{}, 4921(1990).
Yu. S. Kivshar, Phy. Rev. Lett. [**70**]{}, 3055(1993). K. W. Sandusky and J. B. Page, Phys. Rev. B [**50**]{}, 866(1994).
S. Flach, Phys. Rev. E [**51**]{}, 3579(1995); S. Flach, K. Kladko, and R. S. MacKay, Phys. Rev. Lett. [**78**]{}, 1207(1997).
S. A. Kiselev, S. R. Bickham, and A. J. Sievers, Comments Condens. Matter Phys. [**17**]{}, 135(1995).
A. J. Sievers and J. B. Page, “Unusual anharmonic local mode systems”, in [*Dynamical Properties of Solids*]{}, ed. by G. K. Horton and A. A. Maradudin (Elsevier Science B.V., Amsterdam, 1995), p.137 and references therein.
S. Flach and C. R. Wills, Phys. Rep. (1997), in press.
K. Yoshimura and S. Watanabe, J. Phys. Soc. Jpn. [**60**]{}, 82(1991).
Wei-zhong Chen, Phys. Rev. B [**49**]{}, 15063(1994).
P. Marquié, J. M. Bilbault, and M. Remoissenet, Phys. Rev. E [**51**]{}, 6127(1995).
Zhu-Pei Shi, Guoxiang Huang, and Ruibao Tao, Int. J. Mod. Phys. B [**5**]{}, 2237(1991).
Guoxiang Huang, Hongfang Li, and Xianxi Dai, Chin. Phys. Lett. [**9**]{}, 151(1992).
T. Rössler and J. B. Page, Phys. Rev. B [**51**]{}, 11382(1995).
S. R. Bickham, S. A. Kiselev, and A. J. Sievers, Phys. Rev. B [**47**]{}, 14206(1993).
Yu. S. Kivshar and N. Flytzanis, Phys. Rev. A [**46**]{}, 7972(1992).
O. A. Chubykalo, A. S. Kovalev, and O. V. Usatenko, Phys. Rev. B [**47**]{}, 3153(1993).
S. A. Kiselev, S. R. Bickham, and A. J. Sievers, Phys. Rev. B [**48**]{}, 13508(1993).
S. A. Kiselev, S. R. Bickham, and A. J. Sievers, Phys. Rev. B[**50**]{}, 9153(1994).
A. Franchini, V. Bortolani, and R. F. Wallis, Phys. Rev. B [**53**]{}, 5420(1996).
M. Aoki and S. Takeno, J. Phys. Soc. Jpn. [**64**]{}, 809(1995).
D. Bonart, A. P. Mayer, and U. Schröder, Phys. Rev. Lett. [**75**]{}, 870(1995); Phys. Rev. B [**51**]{}, 13739(1995).
J. N. Teixeira and A. A. Maradudin, Phys. Lett. A [**205**]{}, 349(1995).
V. V. Konotop, Phys. Rev. E [**53**]{}, 2843(1996).
Guoxiang Huang, Sen-yue Lou, and M. G. Velarde, Int. J. Birfurcation and Chaos [**6**]{}, 1775(1996).
S. A. Kiselev and A. J. Sievers, Phys. Rev. B [**55**]{}, 5755(1997).
Wei Chen and D. L. Mills, Phys. Rev. Lett. [**58**]{}, 160(1987).
Sen-yue Lou and Guoxiang Huang, Mod. Phys. Lett. [**9**]{}, 1231(1995).
Sen-yue Lou, Jun Yu, Ji Lin, and Guoxiang Huang, Chin. Phys. Lett. [**12**]{}, 400(1995).
D. Bonart, T. Rössler, and J. B. Page, Phys. Rev. B [**55**]{}, 8829(1997).
For the optical phonon branch a wavenumber $q=q_c$ exists at which the curvature of $\omega^{+}(q)$, $2\al_{+}={\rm d}^2\omega_{+}/{\rm d}q^2$, vanishes([*i. e*]{}. $q=q_c$ is a zero-dispersion point). Thus at $q=q_c$ the Eq.(28) is not valid for the description of nonlinear excitations. In this case a higher-order dispersion term $-\fr{i}{3!}{\rm d}^3\omega^{+}/{\rm d}q^3 G_{11}$ should be added in the Eq.(28).
T. B. Benjamin and J. E. Feir, J. Fluid Mech. [**27**]{}, 417(1967).
W. Eckhaus, [*Studies in Nonlinear Stability Theory*]{} (Springer, Berlin, 1965); M. C. Cross and P. C. Hohenberg, Rev. Mod. Phys. [**65**]{}, 851(1993).
Yu. S. Kivshar and M. Peyrard, Phys. Rev. E [**46**]{}, 3198(1992).
J. T. Stuart and R. C. DiPrima, Proc. Roy. Soc. London, Ser. A [**362**]{}, 27(1978).
M. Born and K. Huang, [*Dynamical Theory of Crystal Lattices*]{}(Oxford University Press, Oxford, 1954), P.26.
S. Flach, Physica D [**91**]{}, 223(1996).
St. Pnevmaticos, N. Flytzanis, and M. Remoissenet, Phys. Rev. B [**33**]{}, 2308(1986).
Guoxiang Huang, M. G. Velarde, and Shanhua Zhu, Phys. Rev. B [**55**]{}, 336(1997). M. C. Cross and P. C. Hohenberg, Rev. Mod. Phys. [**65**]{}, 851(1993).
F. M. Russell, Y. Zolotaryuk, and J. C. Eilbeck, Phys. Rev. B [**55**]{}, 6304(1997).
----------- ------------------------------------------- --------------------------------------------------- --------------------------------------------------- ------------------
potential $K_2$ $K_3$ $K_4$ $K_3^2/(K_2K_4)$
Toda $ab$ -$\fr{ab^2}{2}$ $\fr{ab^3}{6}$ $\fr{3}{2}$
B-M-C $\fr{\al_Mq^2(d_0-2\rho_0)}{\rho_0d_0^3}$ -$\fr{\al_Mq^2(d_0^2-6\rho_0^2)}{2\rho_0^2d_0^4}$ $\fr{\al_Mq^2(d_0^3-24\rho_0^3)}{6\rho_0^3d_0^5}$ $\fr{3}{2}I$
L-J $\fr{72\al}{d_0^2}$ $-\fr{756\al}{d_0^3}$ $\fr{6678\al}{d_0^4}$ $\fr{63}{53}$
Morse $2Pa^2$ $-Pa^3$ $\fr{Pa^4}{3}$ $\fr{3}{2}$
----------- ------------------------------------------- --------------------------------------------------- --------------------------------------------------- ------------------
: The force constants $K_2$, $K_3$, $K_4$ and the value of $K_3^2/(K_2K_4)$ for the standard two-body potentials from the Toda, Born-Mayer-Coulomb(B-M-C), Lennard-Jones(L-J), and Morse type.
| {
"pile_set_name": "ArXiv"
} |
-0.5in 0.2in 0.1in 6.0in 8.5in
\
\
a)\
b)\
c)\
[**Abstract**]{}
It is shown that the operator methods of supersymmetric quantum mechanics and the concept of shape invariance can profitably be used to derive properties of spherical harmonics in a simple way. The same operator techniques can also be applied to several problems with non-central vector and scalar potentials. As examples, we analyze the bound state spectra of an electron in a Coulomb plus an Aharonov-Bohm field and/or in the magnetic field of a Dirac monopole.
Spherical harmonics are introduced in physics courses while treating the Laplacian in spherical polar coordinates. In the context of quantum mechanics, the Schrödinger equation is separable into radial and angular parts if the potential is spherically symmetric. The angular piece of the Laplacian operator generates spherical harmonics which obey interesting and useful recursive formulae. With the advent of supersymmetric quantum mechanics (SUSYQM)\[1-3\] and the idea of shape invariance[@Gendenshtein83], study of potential problems in nonrelativistic quantum theory has received renewed interest. SUSYQM allows one to determine eigenstates of known analytically solvable potentials using algebraic operator formalism[@Dutt] without ever having to solve the Schrödinger differential equation by standard series method. However, the operator method has so far been applied only to one dimensional and spherically symmetric three dimensional problems.
The study of exact solutions of the Schrödinger equation with a vector potential and a non-central scalar potential is of considerable interest. In recent years, numerous studies\[6-10\] have been made in analyzing the bound states of an electron in a Coulomb field with simultaneous presence of Aharonov-Bohm (AB)[@Aharonov] field and/or a magnetic Dirac monopole[@Dirac]. In most of these studies, the eigenvalues and eigenfunctions are obtained via separation of variables in spherical or other orthogonal curvilinear coordinate systems. The purpose of this paper is to illustrate that the idea of supersymmetry and shape invariance can be used to obtain exact solutions of such non-central but separable potentials in an algebraic fashion. In this method, it emerges that the angular part (as well as the radial part) of the Laplacian of the Schrödinger equation can indeed be dealt with using the idea of shape invariance. This is a novel method of generating interesting recurrence properties of spherical harmonics. In standard text books on quantum mechanics[@qm-texts], properties of spherical harmonics are discussed from a different point of view. In this regard our approach is new and instructive in the sense that the radial and the angular pieces of the Schrödinger equation can both be treated within the same framework. The present work also gives insight into the solvability of certain three-dimensional problems. Basically, this is a consequence of both separability of variables and the shape invariance of the resulting one-dimensional problems.
The Schrödinger equation for a particle of charge $e$ in the presence of a scalar potential $V(r,\theta,\phi)$ and a vector potential $\vec{A}(r,\theta,\phi)$ is (in units of $\hbar=2m=1$) $$[\{-i \vec{\nabla}-e\vec{A}(r,\theta,\phi)\}^2+V(r,\theta,\phi)] \psi
= E \psi~~.$$ We will consider vector potentials of the form $\vec{A} = \frac{F(\theta)}{r\sin\theta} \hat{e}_\phi$, and scalar potentials of the form $V=V_1(r)+\frac{V_2(\theta)}{r^2}$. The Schrödinger equation reads $$\begin{aligned}
-\frac{1}{r} \frac{\partial^2}{\partial r^2}(r \psi)
-\frac{1}{r^2\sin \theta} \frac{\partial}{\partial \theta}
\left( \sin\theta \frac{\partial \psi}{\partial \theta} \right)
-\frac{1}{r^2\sin^2\theta} \frac{\partial^2\psi}{\partial \phi^2}&&\nonumber\\
+\left( \frac{e^2 F^2(\theta)}{r^2\sin^2\theta} + V_1(r) +
\frac{V_2(\theta)}{r^2} \right) \psi
+ \frac{2ie F(\theta)}{r^2\sin^2\theta} \frac{\partial\psi}{\partial \phi}
&=&E \psi.
\label{rP}\end{aligned}$$ Eq. (\[rP\]) permits a solution via separation of variables, if one writes the wave function in the form $$\psi(r,\theta,\phi)=R(r) P(\theta) e^{im \phi}~~.$$ The equations satisfied by the quantities $R(r)$ and $P(\theta)$ are:
$$\label{R}
\frac{d^2R}{dr^2} + \frac{2}{r}\frac{dR}{dr} +
\left( E-V_1(r)-\frac{l(l+1)}{r^2} \right) R = 0~~,$$
$$\label{P}
\frac{d^2P}{d\theta^2} + \cot \theta \frac{dP}{d\theta} +
\left[l(l+1)- \frac{\left\{ m-F(\theta) \right\}^2}{\sin^2\theta}-V_2(\theta)
\right]P=0.$$
Solutions of these equations for various choices of scalar and vector potentials will now be discussed.
Let us begin with the simplest case of a free particle $[V_1(r)=V_2(\theta)=F(\theta)=0.]$ This will allow us to obtain the standard properties of the spherical harmonics. In this case, differential equation (\[P\]) reduces to $$\label{PP}
\frac{d^2P}{d\theta^2} + \cot \theta \frac{dP}{d\theta}
+\left[l(l+1)-\frac{m^2}{\sin^2 \theta}\right]P = 0~~,$$ which is the equation satisfied by associated Legendre polynomials. To solve it by the SUSYQM method, we need to re-cast it into a Schrödinger-like equation. Changing the variable $\theta \rightarrow z$ through a mapping function $\theta = f(z)$, one obtains $$\frac{d^2P}{dz^2} + \left[ -\frac{f''}{f'} + f' \cot f \right]\frac{dP}{dz}
+f'^2 \left[ l(l+1)-\frac{m^2}{\sin^2 f} \right]P = 0~.$$ Since we want the first derivative to vanish, we choose $$\frac{f''}{f'} = f' \cot f~~,$$ which gives $$\theta \equiv f=2 \tan^{-1}(e^z)~~. \label{theta}$$ This transformation amounts to the replacement $\sin\theta={\rm sech} z$ and $\cos\theta=-\tanh z$. The range of the variable $z$ is $-\infty <z< \infty$. Eq. (\[PP\]) now reads $$\label{Poschl-T}
-\frac{d^2P}{dz^2}-l(l+1) {\rm sech}^2 z~P =- m^2~ P ~~.$$ This is a well-known, shape invariant, exactly solvable potential. It can be readily solved by standard SUSYQM techniques [@Cooper95; @Dutt]. The energy eigenvalues for the potential $V(z)=-a_0(a_0+1) {\rm sech}^2 z~~(a_0>0)$ are: $$E_n =-(a_0-n)^2~~; ~~~(n=0,1,2,\cdots N)~~,$$ where $N$ is the number of bound states this potential holds, and is equal to the largest integer contained in $a_0$. The eigenfunctions $\psi_n(z,a_0)$ are obtained by using supersymmetry operators[@Dutt]: $$\label{dagger}
\psi_n(z;a_0) \propto A^{\dagger}(z;a_0) A^{\dagger}(z;a_1)
\cdots A^{\dagger}(z;a_{n-1}) \psi_0(z;a_n) ~~,$$ where $A^{\dagger}(z;a) \equiv \left( -\frac{d}{dz}+a \tanh z\right)$ and $a_n=a_0-n$. The ground state wave function is $\psi_0(z;a_n)= {\rm sech}^{a_n} z$.
For our problem, $a_0=l$, $E_n=-m^2$ and consequently one has $$n=l-m~;~~~~P_{l,m}(\tanh z) \sim \psi_{l-m}(z;l),$$ where $P_{l,m}(\tanh z)$, the solutions of eq. (\[PP\]), are the associated Legendre polynomials of degree $l$. Now that these $P_{l,m}(z)$ functions can be viewed as solutions of a Schrödinger equation, we can apply all the machinery one uses for a quantum mechanical problem. For example, we know that the parity of the $n$-th eigenfunction of a symmetric potential is given by $(-1)^n$, we readily deduce the parity of $P_{l,m}$ to be $(-1)^{l-m}=(-1)^{l+m}\cite{Arfken}.$ Also, the application of supersymmetry algebra results in identities that are either not very well known or not easily available. With repeated application of the $A^\dagger$ operators, we can determine $P_{l,m}$ for a fixed value ($l-m$). As an illustration, we explicitly work out all the polynomials for $l-m=2$. The lowest polynomial corresponds to $l=2, m=0$. For a general $l$, using eq. (\[dagger\]), one gets $$\begin{aligned}
P_{l,l-2}(\tanh z) &\sim & \psi_2(z;l) \nonumber \\
&\sim & A^{\dagger}(z;l) A^{\dagger}(z;l-1) \psi_0(z;l-2) \nonumber \\
&\sim & \left( -\frac{d}{dz} +l\tanh z \right)
\left( -\frac{d}{dz} +(l-1) \tanh z \right) {\rm sech}^{l-2} z \nonumber \\
&\sim & \left[ -1 + (2l-1) \tanh^2 z \right] {\rm sech}^{l-2} z~~.\end{aligned}$$ A similar procedure is readily applicable for other values of $n$. When one converts back to the original variable $\theta~~({\rm sech} z = \sin \theta~;
\tanh z = -\cos \theta)$, the results are: $$\begin{aligned}
P_{l,l}(\cos \theta) &\sim &\sin^l \theta \\
P_{l,l-1}(\cos \theta) &\sim & \sin^{l-1} \theta \cos \theta\nonumber \\
P_{l,l-2}(\cos \theta) &\sim &
\left[ -1 +(2l-1)\cos^2\theta\right]\sin^{l-2}\theta
\nonumber \\
P_{l,l-3}(\cos \theta) &\sim &
\left[ -3 +(2l-1)\cos^2\theta\right]\sin^{l-3}\theta
\cos \theta \nonumber
\\
P_{l,l-4}(\cos \theta) &\sim &\left[ 3 -6(2l-3)\cos^2 \theta +
(2l-3)(2l-1)\cos^4 \theta\right]\sin^{l-4}\theta \nonumber \\
P_{l,l-5}(\cos \theta) &\sim &\left[ 15 -10 (2l-3) \cos^2\theta +
(2l-3)(2l-1)\cos^4 \theta
\right]\sin^{l-5}\theta \cos \theta.\nonumber\end{aligned}$$ These results are not readily available in standard books, although of course they can be obtained from the generating function for associated Legendre polynomials. Likewise, there are several recurrence relations which are easily obtainable via SUSYQM methods. In particular, by applying $A$ or $A\dagger$ once, we generate recurrence relations of varying degrees (differ in $l$): $$\begin{aligned}
\psi_{l-m}(z,l)=A^\dagger(z,l) \psi_{l-m-1}(z,l-1) &\Longrightarrow&
P_{l,m}(x)=\left((1-x^2)\frac{d}{dx}+l\,x \right) P_{l-1,m} \\
\psi_{l-m}(z,l)=A(z,l-1) \psi_{l-m+1}(z,l+1) &\Longrightarrow&
P_{l,m}(x)=\left((1-x^2)\frac{d}{dx}+(l-1)\,x \right) P_{l+1,m}.\nonumber \end{aligned}$$
Here we examine the general case with non-zero $F(\theta)$ in eq. (\[P\]). In particular, we consider $$F(\theta)=\frac{{\cal{F}}}{2\pi} + g(1-\cos \theta)~~,
~~ V_1(r)=-\frac{Ze^2}{r}~~,
~~V_2(\theta)=0~~.$$ This choice corresponds to the physically interesting problem of the motion of an electron in a Coulomb field in the presence of an Aharonov-Bohm potential $\vec{A}_{AB}=\frac{{\cal{F}}}{{2\pi}r \sin \theta} \hat{e}_\phi$, and a Dirac monopole potential $\vec{A}_D=\frac{g(1-\cos\theta)}{r \sin \theta} \hat{e}_\phi$[@Villalba]. Both potentials can be chosen to be of arbitrary strength depending on the values of the coupling constants ${\cal F}$ and $g$. Again, we transform eq. (\[P\]) via the change of variables from $\theta$ to $z$ given in eq. (\[theta\]). The result is $$\label{pnew}
\frac{d^2P}{dz^2} +[(\lambda^2-{\tilde{m}}^2)-V(z)]P=0~~,$$ with $$V(z)=(\lambda^2+q^2)\tanh^2z-2q{\tilde{m}} \tanh z~~,$$ and $$q=-ge~~,~~{\tilde{m}}=\frac{{\cal{F}}e}{2\pi}+ge-m~~,~~\lambda^2=l(l+1)~~.$$ The potential $V(z)$ has the form of the Rosen-Morse II potential, which is well-known to be shape invariant. More specifically, the potential $$V_{RM}(z)=a_0(a_0+1) \tanh ^2 z +2b_0\tanh z~~~(b_0<a_0^2)$$ has energy eigenvalues and eigenfunctions[@Dutt] $$\begin{aligned}
\label{en}
E_n&=&a_0(a_0+1) - (a_0-n)^2 - \frac{b_0^2}{(a_0-n)^2}, \\
\psi_n&=&(1-\tanh z)^{\frac{1}{2}\left[a_0-n+\frac{b_0}{a_0-n}\right]}
(1+\tanh z)^{\frac{1}{2}\left[a_0-n-\frac{b_0}{a_0-n}\right]}
P_n^{\left( a_0-n+\frac{b_0}{a_0-n},
a_0-n-\frac{b_0}{a_0-n}\right)}.\nonumber\end{aligned}$$ For our case, $E_n=\lambda^2-{\tilde{m}}^2$ and the constants $a_0$, and $b_0$ are given by $$a_0=-\frac{1}{2}+\sqrt{\frac{1}{4}+(\lambda^2+q^2)}~, ~~~~
b_0=-q{\tilde{m}}~~.$$ Using these values and stipulating that eigenfunctions given in eq. (\[en\]) be normalizable, we get $$\begin{aligned}
n&=\sqrt{\frac{1}{4}+(\lambda^2+q^2)} ~~~~ - |q|-
\frac{1}{2}&~~~~~~~~(|q|>|{\tilde{m}}|)~~,\nonumber\\
&=\sqrt{\frac{1}{4}+(\lambda^2+q^2)} ~~~~ - |{\tilde{m}}|-
\frac{1}{2}&~~~~~~~~(|q|<|{\tilde{m}}|)~~.\end{aligned}$$ Corresponding to these two cases, one gets $$\begin{aligned}
l & =-\frac{1}{2} + \sqrt{(n+|q|+\frac{1}{2})^2-q^2}
& ~~~~~~~~(|q|>|{\tilde{m}}|)~~,\nonumber\\
&=-\frac{1}{2} + \sqrt{(n+|{\tilde{m}}|+\frac{1}{2})^2-q^2}
&~~~~~~~~(|q|<|{\tilde{m}}|)~~.\end{aligned}$$ The energy eigenvalues obtained from eq. (\[R\]) for the Coulomb potential are $$\label{Coul}
E_N=\frac{-Z^2e^4}{4[N+l+1]^2}~~.$$ Therefore, our final eigenvalues for a bound electron in a Coulomb potential as well a combination of Aharonov-Bohm and Dirac monopole vector potentials are $$\begin{aligned}
E_N&=\frac{-Z^2e^4}{4 \left[ N + \frac{1}{2} +
\sqrt{(n+|q|+\frac{1}{2})^2-q^2} \right]^2} &
~~~~~~~~ (|q|>|{\tilde{m}}|)~~,\nonumber\\
&=\frac{-Z^2e^4}{4 \left[ N + \frac{1}{2} +
\sqrt{(n+|{\tilde{m}}|+\frac{1}{2})^2-q^2} \right]^2}
&~~~~~~~~ (|q|<|{\tilde{m}}|)~~,\end{aligned}$$ which agree with eqs. (32) and (33) respectively of ref. [@Villalba].
The above calculation using operator techniques in SUSYQM can also be carried out when the Coulomb potential is replaced by a harmonic oscillator. Potentials of this type have been studied by many authors using other approaches[@Kibler93].
As a final example, we consider the case of zero vector potential and a scalar piece consisting of the Coulomb potential and a non-central part $V_2(\theta)$. The angular piece satisfies a modified version of eq. (\[Poschl-T\]): $$\label{Poschl-T1}
\frac{d^2P}{dz^2}+\left[ {\rm sech}^2 z \left\{ l(l+1)-V_2(z) \right\}
-m^2\right]~P=0 ~~.$$ This equation is the Schrödinger equation for one of the known shape invariant potentials provided the function $V_2(z)$ is chosen appropriately. The following three simple choices can be made: $V_2(z) = b\sinh z , b\sinh 2z , b\cosh^2z.$ These choices give rise to the Scarf II and Rosen-Morse II potentials[@Cooper95] in eq. (\[Poschl-T1\]). In terms of $\theta$, these choices correspond to non-central potentials with angular dependences $\cot \theta, \cot \theta {\rm cosec} \theta {\rm ~and~}
{\rm cosec}^2 \theta$, respectively.
Here, we will treat the case $V_2(z)=b \sinh z$ in detail. The full potential in eq. (\[Poschl-T1\]) is $-l(l+1) {\rm sech}^2 z + b {\rm sech} z \tanh z$ and the role of energy is played by $-m^2$. Recall[@Cooper95] that the potential $V(x)=(B^2-A^2-A) {\rm sech}^2 x + B(2A+1) {\rm sech} x \tanh x$ with $A>0$ has eigenvalues $E_n=2An-n^2~~(n<A)~$. In terms of our parameters, this implies $$B(2A+1)=b~~,~~b^2-A^2-A=-l(l+1)~~,~~2An-n^2=-m^2~~.$$ Elimination of $A,B$ leads to $$l=-\frac{1}{2}+\sqrt{\frac{1}{4}+X}$$ with $$X=\frac{(n^2-m^2)}{2n}+\frac{(n^2-m^2)^2}{4n^2}-\frac{n^2b^2}
{(n^2-m^2+n)^2}~~.$$ Substitution for $l$ into eq. (\[Coul\]) gives the eigenvalues for this problem. Although we have explicit energy eigenvalues and eigenfunctions, it is not clear whether the non-central potential we have considered has physical significance.
In conclusion, we mention that in this paper we have attempted to explore the effectiveness of the SUSYQM operator methods to obtain analytic solutions of Schrödinger systems in more than one dimensions. It is clear that the factorization technique which has so far been applied to only one dimensional or spherically symmetric three dimensional problems, can be equally useful for non-central separable problems for which one dimensional equations can be recast into Schrödinger equations with shape invariant potentials. Many interesting properties of spherical harmonics emerge naturally as a simple realization of this operator technique. One should also note that although, in this paper, we have focused on spherical polar coordinates, any orthogonal curvilinear coordinate system which is separable into Schrödinger-type equations with shape invariant potentials will allow similar algebraic analysis, and will have analytically solvable eigenvalues.
This work was supported in part by the U.S. Department of Energy. Two of us, R.D. and A.G.are grateful to the Physics Department of University of Illinois at Chicago for kind hospitality. The authors are grateful to Dr. Avinash Khare for useful conversations.
[99]{} E. Witten, Nucl. Phys. [**B188**]{} (1981) 513, Nucl. Phys. [**B202**]{} (1982) 253. F. Cooper and B. Freedman, Ann. Phys. [**146**]{} (1983) 262. For a recent review of supersymmetric quantum mechanics and additional references see F. Cooper, A. Khare and U. Sukhatme, Phys. Rep. [**251**]{} (1995) 267. L. Gendenshtein, JETP Letters [**38**]{} (1983) 356. R. Dutt, A. Khare and U.P. Sukhatme, Am. Jour. Phys., [**56**]{} (1988) 163. M. Kibler and T. Negadi, Phys. Lett. [**A 124**]{} (1987) 42. A. Guha and S. Mukherjee, Jour. Math. Phys. [**28**]{} (1989) 840. Gh.E. Draganescu, C. Campiogotto, and M. Kibler, Phys. Lett. [**A 170**]{} (1992) 339. M. Kibler and C. Campiogotto, Phys. Lett. [**A 181**]{} (1993) 1 and references contained therein. V.M. Villalba, Phys. Lett. [**A 193**]{} (1994) 218. Y. Aharonov, and D. Bohm, Phys. Rev. [**115**]{} (1959) 485. P.A.M. Dirac, Proc. R. Soc. London, [**A 133**]{} (1931) 60. L. I. Schiff, [*Quantum Mechanics*]{} (3rd Edition), McGraw-Hill, New York (1968); E. Merzbacher, [*Quantum Mechanics*]{} (2nd Edition) Wiley, New York (1970); L. Landau and E. M. Lifshitz, [*Quantum Mechanics, Non-relativistic Theory*]{}, (2nd Edition) Addison-Wesley, Reading, Mass. (1965). L. Chetonani, L. Guechi, and T.F. Harman, Jour. Math. Phys. [**30**]{} (1989) 655. Handbook of Mathematical Functions, M. Abramowitz and I.A. Stegun, Dover Publications, Inc., New York, 1970.\
Mathematical Methods for Physicists, George Arfken, Third Edition, Academic Press.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Given its low dollar and maintenance cost, RFID is poised to become the enabling technology for inventory control and supply chain management. However, as an outcome of its low cost, RFID based inventory control is susceptible to pernicious security and privacy threats. A deleterious attack on such a system is [*corporate espionage*]{}, where attackers through illicit inventorying infer sales and restocking trends for products. In this paper, we first present plausible aftermaths of corporate espionage using real data from online sources. Second, to mitigate corporate espionage in a retail store environment, we present a simple low-cost system called [[Mirage]{}]{}. [[Mirage]{}]{} uses additional programmable low cost passive RFID tags called honeytokens to inject noise in retail store inventorying. Using a simple history based algorithm that controls activation and de-activation of honeytokens, [[Mirage]{}]{} randomizes sales and restocking trends. We evaluate [[Mirage]{}]{} in a real warehouse environment using a commercial off-the-shelf Motorola MC9090 handheld RFID reader and over 450 Gen2 low cost RFID tags. We show that [[Mirage]{}]{} successfully flattens and randomizes sales and restocking trends while adding minimal cost to inventory control.\
[**Keywords:**]{} RFID, Illicit Inventorying, Honeytokens, Corporate Espionage
author:
- Jonathan White and Nilanjan Banerjee
bibliography:
- 'mirage.bib'
title: '[Mirage]{}: Mitigating Illicit Inventorying in a RFID Enabled Retail Environment'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A fundamental assumption in the theory of brane world is that all matter and radiation are confined on the four-dimensional brane and only gravitons can propagate in the five-dimensional bulk spacetime. The brane world theory did not provide an explanation for the existence of electromagnetic fields and the origin of the electromagnetic field equation. In this paper, we propose a model for explaining the existence of electromagnetic fields on a brane and deriving the electromagnetic field equation. Similar to the case in Kaluza-Klein theory, we find that electromagnetic fields and the electromagnetic field equation can be derived from the five-dimensional Einstein field equation. However, the derived electromagnetic field equation differs from the Maxwell equation by containing a term with the electromagnetic potential vector coupled to the spacetime curvature tensor. So it can be considered as generalization of the Maxwell equation in a curved spacetime. The gravitational field equation on the brane is also derived with the stress-energy tensor for electromagnetic fields explicitly included and the Weyl tensor term explicitly expressed with matter fields and their derivatives in the direction of the extra-dimension. The model proposed in the paper can be regarded as unification of electromagnetic and gravitational interactions in the framework of brane world theory.
Classical general relativity, higher-dimensional gravity, brane world theory
author:
- 'Li-Xin Li'
title: Electromagnetic Force on a Brane
---
Introduction
============
The theory of brane world was proposed to address the hierarchy problem in theoretical physics [@ran99a; @ran99b]. In the theory of brane world, the four-dimensional spacetime in which we live is assumed to be a hypersurface (or, a brane) embedded in a five-dimensional bulk spacetime. The gravitational field equation on the four-dimensional brane is derived from the Einstein field equation in the five-dimensional bulk space by the approach of projection [@shi00]. Standard model particles, including electromagnetic fields, strong and weak particles, are assumed to be confined on the four-dimensional brane. The assumption is motivated by D-branes in string theory, on which open strings representing the non-gravitational sector can end [@hor96; @luk99; @luk99b; @luk00]. For a recent and comprehensive review on the theory of brane world and its application in physics and cosmology, please refer to [@maa10].
Kaluza-Klein (KK) theory also attempts to interpret the physics in a four-dimensional spacetime as arising from the gravity in a five-dimensional bulk spacetime [@kal21; @kle26a; @kle26b; @bai87; @ove97]. In the KK theory, both the Maxwell equation and the four-dimensional Einstein field equation are derived from the five-dimensional Einstein field equation. In the KK theory the extra-dimension is assumed to be compactified to a very small scale so that the extra-dimension cannot be seen in normal physical experiments and hence gravity appears four-dimensional. However, in the brane theory the extra-dimension can be noncompact. It is the curvature of the bulk space that keeps gravity to be four-dimensional on scales larger than the curvature radius of the bulk space [@ran99b].
Although both attempt to interpret the four-dimensional physics as arising from the five-dimensional physics, KK and brane world theories are distinctly different in physics: they are defined on two different hypersurfaces in a five-dimensional spacetime and are not related by diffeomorphisms, as explained in detail in [@li15a]. As can be seen from the above description, electromagnetism has different origins in the two theories. In the KK theory, the Maxwell equation is derived from the five-dimensional Einstein field equation, hence electromagnetism and gravity have the same origin. In the brane world theory, electromagnetism is assumed to arise from open strings ending on D-branes hence has no relation to gravity arising from closed strings. In addition, as mentioned above, in the KK theory the extra spatial dimension must be compactified, but in the brane theory the extra spatial dimension can be noncompact.
It would be interesting to adapt the idea of deriving the Maxwell equation from the five-dimensional Einstein field equation in the KK theory to the brane world theory. However, this is not an easy task, since the 4+1 decomposition of the five-dimensional metric tensor adopted in the KK theory is different from that used for obtaining a metric tensor on a brane hypersurface, as explained in [@li15a]. In this paper, we propose a brane world model in which electromagnetic fields and the electromagnetic field equation on the brane are derived from the five-dimensional Einstein field equation. We will see that, this can be realized only if an appropriate boundary condition on the brane is adopted, and the derived electromagnetic field equation differs from the Maxwell equation by a curvature-coupled term. The boundary condition differs from the $Z_2$-symmetry boundary condition used in the standard brane world theory, as will be explained in the paper.
4+1 Decomposition of the Five-dimensional Einstein Field Equation
=================================================================
Assuming a five-dimensional spacetime in which gravity is described by the Einstein field equation $$\begin{aligned}
\t{R}_{ab}-\frac{1}{2}\t{R}\t{g}_{ab}=\t{\kappa}\t{T}_{ab} \;, \label{ein_eq5}\end{aligned}$$ where $\t{g}_{ab}$ is the metric tensor of the spacetime, $\t{R}_{ab}$ is the Ricci curvature tensor, $\t{R}\equiv \t{g}^{ab}\t{R}_{ab}$ is the Ricci scalar, $\t{T}_{ab}$ is the stress-energy tensor of matter, and $\t{\kappa}$ is gravitational coupling constant. The five-dimensional spacetime is sliced by a set of timelike hypersurfaces, and one of the hypersurfaces is just the four-dimensional spacetime in which we live. The hypersurface has a unit normal $n^a$, a metric tensor $g_{ab}\equiv \t{g}_{ab}-n_an_b$, and an extrinsic curvature tensor $K_{ab}\equiv\t{\pounds}_ng_{ab}$ where $\t{\pounds}_n$ denotes the Lie derivative with respect to $n^a$.
It is well known that the Einstein field equation (\[ein\_eq5\]) is equivalent to the following three equations expressed in terms of geometric quantities on the hypersurface [@mis73; @wal84; @shi00; @li15a]: $$\begin{aligned}
R +K_{ab}K^{ab}-K^2 = -2\t{\kappa}\t{T}_{ab}n^an^b \;, \label{scalar_eq}\end{aligned}$$ $$\begin{aligned}
\nabla_aK^{ab}-\nabla^bK = \t{\kappa}g^{ab}\t{T}_{ac}n^c \;, \label{vector_eq}\end{aligned}$$ and $$\begin{aligned}
G_{ab} = \t{\kappa}g_a^{\;\;c}g_b^{\;\;d}\t{T}_{cd} +g_{ab}{}{}^{cd}\t{\pounds}_nK_{cd} -\left(2K_a^{\;\;c}K_{cb}-KK_{ab}\right)+\frac{1}{2}\left(3K_{cd}K^{cd}-K^2\right)g_{ab} \;. \label{tensor_eq}\end{aligned}$$ Here, $g_{abcd}\equiv g_{ac}g_{bd}-g_{ab}g_{cd}$, $\nabla_a$ is the derivative operator associated with the metric $g_{ab}$, $R_{ab}$ is the Ricci curvature tensor of the brane hypersurface, $R\equiv R_a{}^a$ is the Ricci scalar, $G_{ab}\equiv R_{ab}-(1/2)Rg_{ab}$ is the Einstein tensor, and $K\equiv K_a{}^a$.
The scalar equation (\[scalar\_eq\]) is obtained by contraction of equation (\[ein\_eq5\]) with $n^an^b$. The vector equation (\[vector\_eq\]) is obtained by contraction of equation (\[ein\_eq5\]) with $n^b$, then projection onto the hypersurface by the projection operator $g_{ab}$. It is also called the Gauss-Codacci relation [@mis73; @wal84]. The tensor equation (\[tensor\_eq\]) is obtained by full projection of the equation (\[ein\_eq5\]) onto the hypersurface. We use normal letters to denote quantities on the brane hypersurface, and tilded letters to denote quantities defined in the bulk spacetime (except the normal vector $n^a$). The index of a tensor in the bulk spacetime is lowered (raised) by $\t{g}_{ab}$ ($\t{g}^{ab}$). The index of a tensor on the brane can be lowered (raised) by both $g_{ab}$ ($g^{ab}$) and $\t{g}_{ab}$ ($\t{g}^{ab}$), with the same result.
In the brane world theory with the $Z_2$-symmetry boundary condition, the tensor field equation (\[tensor\_eq\]) is interpreted as the Einstein field equation on the bane [@shi00]. Equation (\[tensor\_eq\]) agrees with the eq. 8 of [@shi00], if the traceless tensor $E_{ab}$ in the eq. 8 in [@shi00], which is defined by the Weyl tensor in the bulk space, is expressed in terms of the brane extrinsic curvature $K_{ab}$ and its derivative in the direction orthogonal to the brane. Note that, in derivation of equation (\[tensor\_eq\]), following [@shi00] we have assumed that $n^a$ is tangent to a geodesic so that the acceleration vector $a^b\equiv n^a\t{\nabla}_an^b=0$.
Electromagnetic Field Equations on a Brane
==========================================
Let us consider a discontinuous hypersurface (a brane) in a five-dimensional spacetime, which contains a surface stress-energy tensor as assumed in the theory of brane world. To have a well-defined four-dimensional geometry on the brane, the induced metric $g_{ab}$ must be continuous across it, hence the derivative operator $\nabla_a$ and the Riemann curvature defined by it. Then, by definition, the extrinsic curvature $K_{ab}$ does not contain a Dirac $\delta$-function, although it can be discontinuous across the brane. In fact, by the five-dimensional Einstein field equation (\[ein\_eq5\]), $K_{ab}$ must be discontinuous across the brane.
The bulk stress-energy tensor $\t{T}_{ab}$ must contain a $\delta$-function at the position of the brane. So, we can write $$\begin{aligned}
\t{T}_{ab}=-\frac{\t{\Lambda}}{\t{\kappa}}\t{g}_{ab}+\t{\cal T}_{ab}+\t{S}_{ab}\delta(n) \;, \label{tT_tSab}\end{aligned}$$ where $\t{\Lambda}$ is the cosmological constant in the bulk space, $\t{\cal T}_{ab}$ and $\t{S}_{ab}$ are regular tensors (i.e., they contain no $\delta$-function). Here we have written $n^a=(\partial/\partial n)^a$ and use $n=0$ to denote the position of the brane.
Let $[Q]$ denote the difference in the value of any quantity $Q$ on the two sides of the brane, i.e., $[Q]\equiv Q^+-Q^-$, $Q^+=Q(n=0^+)$, and $Q^-=Q(n=0^-)$. Integration of equation (\[tensor\_eq\]) across the brane hypersurface leads to the Israel junction condition [@isr66; @mis73] $$\begin{aligned}
[K_{ab}]=-\t{\kappa}\left(S_{ab}-\frac{1}{3}Sg_{ab}\right) \;, \label{join1}\end{aligned}$$ where $S_{ab}\equiv g_a^{\;\;c}g_b^{\;\;d}\t{S}_{cd}$ and $S=S_a{}^a$. Similarly, integration of equations (\[scalar\_eq\]) and (\[vector\_eq\]) across the brane leads to $$\begin{aligned}
n^cn^d\t{S}_{cd}=0 \hspace{0.5cm} \mbox{and} \hspace{1cm} g_b^{\;\;d}n^c\t{S}_{cd}=0 \;,\end{aligned}$$ which simply tells that the momentum flow represented by $\t{S}_{ab}$ is entirely in the brane [@mis73].
To introduce electromagnetic fields on the brane, in the neighborhood of the brane hypersurface we choose a general coordinate system $\{x^0,x^1,x^2,x^3,x^4\equiv w\}$ so that the brane is located at $w=0$. The coordinate vector $w^a=(\partial/\partial w)^a$ can be decomposed as $w^a=Nn^a+N^a$, where $N$ is the lapse function, and $N^a=g^a{}_bw^b$ is the shift vector [@wal84]. It can be verified that the acceleration vector $a^a=-\nabla^a\ln N$. Hence, the geodesic condition $a^a=0$ is identical to the condition $\nabla_aN=0$, i.e., $N$ can only be a function of $w$. Define $$\begin{aligned}
w^\prime=2\int Ndw \;, \hspace{0.5cm}
A^a=(2N)^{-1}N^a \;, \hspace{0.5cm}
H_{ab}=\nabla_aA_b+\nabla_bA_a=H_{ba} \;,\end{aligned}$$ the extrinsic curvature $K_{ab}$ can be expressed as $$\begin{aligned}
K_{ab} = \dot{g}_{ab}-H_{ab} \;, \label{Kab_H}\end{aligned}$$ where $\dot{g}_{ab}\equiv\partial g_{ab}/\partial w^\prime\equiv g_a^{\;\;c}g_b^{\;\;d}\t{\pounds}_{w^\prime}g_{cd}$.
Substituting equation (\[Kab\_H\]) into the vector equation (\[vector\_eq\]), we get $$\begin{aligned}
\nabla_aF^{ab}+2R^b{}_aA^a=-4\pi J^b \;, \label{meq1}\end{aligned}$$ where $F_{ab}\equiv\nabla_aA_b-\nabla_bA_a$, $$\begin{aligned}
J^a\equiv\frac{1}{4\pi}\left(\nabla_b\Phi^{ab}+\t{\kappa}g^{ab}\t{T}_{bc}n^c\right) \;, \label{Ja}\end{aligned}$$ and $$\begin{aligned}
\Phi_{ab}\equiv -g_{ab}{}{}^{cd}\dot{g}_{cd}=\Phi_{ba} \;. \label{Phi_ab}\end{aligned}$$ Equation (\[meq1\]) differs from the Maxwell equation only by a curvature-coupled term $2R^b{}_aA^a$, if $A^a$ is interpreted as the electromagnetic potential vector, and $J^a$ interpreted as the electric current density vector. When $R_{ab}=0$, it is exactly the Maxwell equation. Therefore, we can interpret equation (\[meq1\]) as generalization of the Maxwell equation in the brane world theory.
Einstein Field Equations on a Brane
===================================
Substituting equation (\[Kab\_H\]) in to the tensor equation (\[tensor\_eq\]), we get a four-dimensional Einstein field equation on the brane: $$\begin{aligned}
G_{ab}=\kappa T_{ab} \;, \label{ein_eq40}\end{aligned}$$ with the stress-energy tensor $$\begin{aligned}
T_{ab} = T_{\emm,ab}+T_{m,ab}+T_{\intt,ab}+\frac{\t{\kappa}}{\kappa}g_a{}^cg_b{}^d\t{T}_{cd} \;, \label{Tab}\end{aligned}$$ where $$\begin{aligned}
\kappa T_{\emm,ab} = 2\Psi_{ac}\Psi_b^{\;\;c}-\frac{2}{3}\Psi\Psi_{ab} -\frac{1}{2}\left(\Psi_{cd}\Psi^{cd}-\frac{1}{3}\Psi^2\right)g_{ab} -2\nabla^c\left(2A_{(a}\Psi_{b)c}-A_c\Psi_{ab}\right) \;, \label{Tab_em_M}\end{aligned}$$ $$\begin{aligned}
\kappa T_{m,ab} = -2\Phi_{ac}\Phi_b^{\;\;c} +\frac{1}{3}\Phi\Phi_{ab}-\frac{1}{2}\left(\Phi_{cd}\Phi^{cd}-\frac{1}{3}\Phi^2\right)g_{ab} -2\dot{\Phi}_{ab} \;, \label{Tab_m_M}\end{aligned}$$ and $$\begin{aligned}
\kappa T_{\intt,ab} = \frac{1}{3}\Phi\Psi_{ab}-\frac{2}{3}\Psi\Phi_{ab}-\left(\Phi_{cd}\Psi^{cd}-\frac{1}{3}\Phi\Psi\right)g_{ab} -2\nabla^c\left(2A_{(a}\Phi_{b)c}-A_c\Phi_{ab}\right)-2\dot{\Psi}_{ab} \;. \label{Tab_int_M}\end{aligned}$$ Here, $\kappa$ is the gravitational coupling constant in the four-dimensional spacetime, $$\begin{aligned}
\Psi_{ab}\equiv H_{ab}-Hg_{ab}=\Psi_{ba} \;, \label{Psi_ab}\end{aligned}$$ $H\equiv H_c{}^c$, $\Psi\equiv \Psi_c{}^c$, and $\Phi\equiv \Phi_c{}^c$. The braces in tensor indexes denote symmetrization of indexes.
According to [@li15a], $T_{\emm,ab}$ is interpreted as the stress-energy tensor of electromagnetic fields, $T_{m,ab}$ interpreted as the stress-energy tensor of the matter field associated with $\dot{g}_{ab}$, and $T_{\intt,ab}$ interpreted as the stress-energy tensor arising from the interaction between the electromagnetic field and the matter field. The $T_{\emm,ab}$ is related to the ordinary electromagnetic stress-energy tensor $$\begin{aligned}
\!\!~^{(0)}T_{\emm,ab}=\frac{2}{\kappa}\left(F_{ac}F_b{}^c-\frac{1}{4}g_{ab}F_{cd}F^{cd}\right) \label{T_em_0}\end{aligned}$$ by $$\begin{aligned}
T_{\emm,ab} =\, ^{(0)}\!T_{\emm,ab} + ^{(1)}\!T_{\emm,ab} \;,\end{aligned}$$ where $$\begin{aligned}
^{(1)}T_{\emm,ab} = -\frac{2}{\kappa}\left\{\nabla^c\nabla_c(A_aA_b)-2\nabla^c\nabla_{(a}(A_{b)}A_c)+4A^cR_{c(a}A_{b)}+g_{ab}\left[\nabla_c\nabla_d(A^cA^d)-R_{cd}A^cA^d\right]\right\}, \hspace{0.3cm} \label{T_em_1}\end{aligned}$$ which arises from the curvature-coupled term in the field equation (\[meq1\]).
Unlike $\!\!~^{(0)}T_{\emm,ab}$, the $\!\!~^{(1)}T_{\emm,ab}$ does not interact with electric charge and current, since $$\begin{aligned}
\nabla^a\!\!~^{(1)}T_{\emm,ab}=\frac{1}{2\pi}\left[F_b{}^cR_{cd}A^d-A_b\nabla^c\left(R_{cd}A^d\right)\right] . \hspace{0.2cm}\end{aligned}$$ In contrast, for the $\!\!~^{(0)}T_{\emm,ab}$, we have $$\begin{aligned}
\nabla^a\!\!~^{(0)}T_{\emm,ab}=-F_{ba}J^a-\frac{1}{2\pi}F_b{}^cR_{cd}A^d \;.\end{aligned}$$ When the electric charge is conserved, by equation (\[meq1\]) we have $\nabla^c\left(R_{cd}A^d\right)=-2\pi\nabla_aJ^a=0$. Then, for the total $T_{\emm,ab}$ we have just the Lorentz force law: $$\begin{aligned}
\nabla^aT_{\emm,ab}=-F_{ba}J^a \;.\end{aligned}$$ Note that, although the effect of $\!\!~^{(1)}T_{\emm,ab}$ cannot be measured by electromagnetic experiments, $\!\!~^{(1)}T_{\emm,ab}$ affects the spacetime geometry according to the Einstein field equation. Hence, $\!\!~^{(1)}T_{\emm,ab}$ represents a kind of dark electromagnetic energy and momentum [@li15b].
Boundary Conditions
===================
To have a well-defined electromagnetic field on the brane, $A^a$ and $\nabla_aA_b$ must be continuous across the brane. That is, we must have $[A^a]=0$, $[F_{ab}]=[H_{ab}]=[\Psi_{ab}]=0$. Then, by equation (\[Kab\_H\]), we have $[K_{ab}]=[\dot{g}_{ab}]$, and $[\Phi_{ab}]=-g_{ab}{}{}^{cd}[\dot{g}_{cd}]=-g_{ab}{}{}^{cd}[K_{cd}]$. By equation (\[join1\]), we then get $$\begin{aligned}
[\Phi_{ab}]=\t{\kappa}S_{ab} \;. \label{joinx}\end{aligned}$$ In the theory of brane world, it is usually assumed that $K_{ab}^+=-K_{ab}^-$, i.e., $K_{ab}$ is antisymmetric about the brane [@shi00; @maa10]. This $Z_2$-symmetry does not apply when the electromagnetic field is present, by equation (\[Kab\_H\]) and the condition $[H_{ab}]=0$. However, since $[K_{ab}]=[\dot{g}_{ab}]$, we can assume that $$\begin{aligned}
\dot{g}_{ab}^+=-\dot{g}_{ab}^- \;,\end{aligned}$$ i.e., $\dot{g}_{ab}$ is antisymmetric about the brane. This condition is equivalent to $\Phi_{ab}^+=-\Phi_{ab}^-$. With this assumption, from equation (\[joinx\]) we get the boundary condition $$\begin{aligned}
\Phi_{ab}^+=-\Phi_{ab}^-=\frac{1}{2}[\Phi_{ab}]=\frac{\t{\kappa}}{2}S_{ab} \;. \label{join2}\end{aligned}$$
By the symmetry properties of $\Psi_{ab}$ and $\Phi_{ab}$, we have $\dot{\Psi}_{ab}^+=-\dot{\Psi}_{ab}^-$, and $\dot{\Phi}_{ab}^+=\dot{\Phi}_{ab}^-$. Then, by equations (\[Tab\_em\_M\])–(\[Tab\_int\_M\]), we have $[T_{\emm,ab}]=[T_{m,ab}]=0$, and $$\begin{aligned}
\kappa T_{\intt,ab}^+ = -\kappa T_{\intt,ab}^-=-2\dot{\Psi}_{ab}^+-\t{\kappa}\nabla^c\left(2A_{(a}S_{b)c}-A_cS_{ab}\right) +\t{\kappa}\left[\frac{1}{6}S\Psi_{ab}-\frac{1}{3}\Psi S_{ab}-\frac{1}{2}\left(S_{cd}\Psi^{cd}-\frac{1}{3}S\Psi\right)g_{ab}\right] \;, \label{T_int_join}\end{aligned}$$ where we have omitted the indexes “$+$” and “$-$” for $\Psi_{ab}$ and $A_a$ since $\Psi_{ab}^+=\Psi_{ab}^-$ and $A_a^+=A_a^-$. To have a well-defined Einstein field equation on the brane, each term on the right-hand side of equation (\[Tab\]) must be symmetric about the brane, since $G_{ab}^+=G_{ab}^-$. So, we must have $\left[g_a{}^cg_b{}^d\t{T}_{cd}\right]=0$ and $\left[T_{\intt,ab}\right]=0$. Then, by equation (\[T\_int\_join\]), we must have $T_{\intt,ab}^+ = T_{\intt,ab}^-=0$, i.e., $$\begin{aligned}
\dot{\Psi}_{ab}^+=\t{\kappa}\left[-\frac{1}{2}\nabla^c\left(2A_{(a}S_{b)c}-A_cS_{ab}\right) +\frac{1}{12}S\Psi_{ab}-\frac{1}{6}\Psi S_{ab}-\frac{1}{4}\left(S_{cd}\Psi^{cd}-\frac{1}{3}S\Psi\right)g_{ab}\right]\;. \label{T_int_join2}\end{aligned}$$
Let us consider a simple case: $S_{ab}=-\lambda g_{ab}$, where $\lambda$ is constant on the brane but can be a function of $w^\prime$, and $\lambda(-w^\prime)=\lambda(w^\prime)$. That is, the brane has a positive tension (represented by $\lambda$) and is vacuum otherwise. By equation (\[T\_int\_join\]) and the above discussions, we get $\kappa T_{\intt,ab}^+ =-2\dot{\Psi}_{ab}^++(\t{\kappa}/3)\lambda\Psi_{ab}=0$, so we must have $\dot{\Psi}^+_{ab}=-\dot{\Psi}^-_{ab}=(\t{\kappa}/6)\lambda\Psi_{ab}$. Substituting $S_{ab}=-\lambda g_{ab}$ into equation (\[Tab\_m\_M\]), we get $\kappa T_{m,ab}=-2\dot\Phi_{ab}$. If we assume that the relation in equation (\[join2\]) holds in a small neighborhood of the brane, we get $2\dot{\Phi}_{ab}=-\t{\kappa}\left(\dot{\lambda}g_{ab}+\lambda\dot{g}_{ab}\right)=-\t{\kappa}\left(\dot{\lambda}-\t{\kappa}\lambda^2/6\right)g_{ab}$, and $$\begin{aligned}
\kappa T_{m,ab}=\t{\kappa}\left(\dot{\lambda}-\frac{1}{6}\t{\kappa}\lambda^2\right)g_{ab} \;,\end{aligned}$$ corresponding to a cosmological constant term. Then, we get the four-dimensional gravitational field equation on the brane $$\begin{aligned}
G_{ab}+\Lambda_\eff\, g_{ab}=\kappa T_{\emm,ab} \;, \label{ein_eq4a}\end{aligned}$$ where we have assumed $\t{\cal T}_{ab}=0$ (eq. \[tT\_tSab\]) and $$\begin{aligned}
\Lambda_\eff\equiv\t{\Lambda}-\t{\kappa}\left(\dot{\lambda}-\frac{\t{\kappa}}{6}\lambda^2\right) \;. \label{lam_eff}\end{aligned}$$ It is just the four-dimensional Einstein field equation with a cosmological constant and the stress-energy of electromagnetic fields as the source.
To cancel the five-dimensional cosmological constant $\t{\Lambda}$ and have $\Lambda_\eff=0$, we must have $\t{\kappa}\left(\dot{\lambda}-\t{\kappa}\lambda^2/6\right)=\t{\Lambda}$. This result agrees with that discussed in [@shi00] when $\dot{\lambda}=0$. If on the brane $\dot{\lambda}\neq 0$ but $\t{\kappa}^2\lambda^2=-6\t{\Lambda}$, we get a residual cosmological constant $\Lambda_\eff=-\t{\kappa}\dot{\lambda}$.
To find out the relation between the coupling constants $\kappa$ and $\t{\kappa}$, let us include a stress-energy tensor of normal matter in $S_{ab}$ and denote it by $\tau_{ab}$: $S_{ab}=-\lambda g_{ab}+\tau_{ab}$. Then we get $$\begin{aligned}
\kappa T_{m,ab}=\t{\kappa}\left(\dot{\lambda}-\frac{\t{\kappa}}{6}\lambda^2\right)g_{ab}+\frac{\t{\kappa}^2}{6}\lambda\tau_{ab}-\t{\kappa}\dot{\tau}_{ab}+\t{\kappa}^2\pi_{ab} ,\hspace{0cm} \nonumber\\ \label{T_m_ab}\end{aligned}$$ where $$\begin{aligned}
\pi_{ab}=-\frac{1}{2}\tau_{ac}\tau_b{}^c+\frac{1}{12}\tau\tau_{ab}-\frac{1}{8}\left(\tau_{cd}\tau^{cd}-\frac{1}{3}\tau^2\right)g_{ab} \hspace{0.4cm}\label{pi_ab}\end{aligned}$$ contains quadratic terms of $\tau_{ab}$. The requirement that the linear term of $\tau_{ab}$ in (\[T\_m\_ab\]) is identical to that contained in the standard four-dimensional Einstein field equation leads to $$\begin{aligned}
\kappa=\frac{1}{6}\t{\kappa}^2\lambda \;,\end{aligned}$$ in agreement with the result in [@shi00]. Hence, when $\Lambda_\eff=0$, we get the general four-dimensional gravitational field equation on the brane $$\begin{aligned}
G_{ab}=\kappa \left(T_{\emm,ab}+\tau_{ab}\right)+\t{\kappa}\left({\cal T}_{ab}-\dot{\tau}_{ab}\right)+\t{\kappa}^2\pi_{ab} \;, \hspace{0.2cm}\label{ein_eq4}\end{aligned}$$ where ${\cal T}_{ab}=g_a{}^cg_b{}^d\t{\cal T}_{cd}$ is the projection of the bulk stress-energy tensor.
The difference between the $\pi_{ab}$ in equation (\[pi\_ab\]) and the $\pi_{\mu\nu}$ in [@shi00] is caused by the fact that in our treatment the tensor $E_{ab}$ in [@shi00] has been expressed in terms of $K_{ab}$ and its derivative. In fact, the $\dot{\tau}_{ab}$ term in equation (\[ein\_eq4\]) arises from the expression for $E_{ab}$. (Details for the relation between $E_{ab}$ and $K_{ab}$ can be found in [@li15a].)
In equation (\[ein\_eq4\]), the term linear in $\tau_{ab}$ is the stress-energy tensor of normal matter (other than electromagnetic fields) on the brane, which has the same form as in the standard Einstein field equation. It guarantees that the Newtonian gravitational law can be obtained in linear perturbations. The $\pi_{ab}$, which contains quadratic terms of $\tau_{ab}$, is important only in high energy states [@shi00; @maa10]. The $T_{\emm,ab}$, defined by equation (\[Tab\_em\_M\]), represents the stress-energy tensor of electromagnetic fields, which differs from the standard electromagnetic stress-energy tensor by a term $\!\!~^{(1)} T_{\emm,ab}$ (eq. \[T\_em\_1\]). The $\dot{\tau}_{ab}$ is the gradient of $\tau_{ab}$ with respect to the extra dimension $w$, which should be small since its effect has never been detected in normal experiments of gravity. The ${\cal T}_{ab}$ arises from the stress-energy tensor of matter in the bulk space, whose effect on the brane may look like some kind of dark matter.
Relation to Other Work in the Literature
========================================
To understand the boundary condition better, we express $K_{ab}$ in terms of $\Phi_{ab}$ and $\Psi_{ab}$ $$\begin{aligned}
K_{ab}=-\hat{\Phi}_{ab}-\hat{\Psi}_{ab} \;, \label{Kab_Phi_Psi}\end{aligned}$$ where $$\begin{aligned}
\hat{\Phi}_{ab}\equiv\Phi_{ab}-\frac{1}{3}\Phi g_{ab} \;, \hspace{1cm}
\hat{\Psi}_{ab}\equiv\Psi_{ab}-\frac{1}{3}\Psi g_{ab} \;. \label{hat_Phi_Psi}\end{aligned}$$ In our model, we choose the boundary condition so that $\hat{\Psi}_{ab}$ is symmetric about the brane, but $\hat{\Phi}_{ab}$ is antisymmetric about the brane. The $\hat{\Psi}_{ab}$ is interpreted as representing electromagnetic fields on the brane. The $\hat{\Phi}_{ab}$ is related to the stress-energy tensor of matter on the brane through the Israel junction relation (eq. \[join2\]).
The boundary condition adopted in this paper is essentially a non-$Z_2$ symmetric boundary condition, which has been studied in the framework of an asymmetric brane world in the literature ([@bat01; @yam07], and references therein). In [@bat01] and [@yam07], the extrinsic curvature tensor $K_{ab}$ of the brane is separated into two parts, an antisymmetric part and a symmetric part about the brane. As usual, the antisymmetric part is related to the stress-energy tensor of matter confined in the brane by the Israel junction relation. The symmetric part is solved from a constraint equation derived from the requirement that $\left[R_{ab}\right]=0$, with a formal solution determined by the stress-energy tensor of matter confined in the brane and the antisymmetric part about the brane of the geometric quantity in the bulk space (eqs. 31 and 32 in [@bat01], eq. 3.17 in [@yam07]). We call it “a formal solution” because the $\left[{\cal F}_{ab}\right]$ in [@bat01; @yam07] contains the Weyl term $\left[{\cal E}_{ab}\right]$, which itself is a function of the $\langle K_{ab}\rangle$ and its derivative in the direction of the extra-dimension according to equation (B35) of [@li15a]. Electromagnetic fields are not discussed in [@bat01; @yam07].
Mathematically, the result in this paper agrees with that in [@bat01] and [@yam07]. The four-dimensional Einstein field equation (\[ein\_eq4\]) mathematically agrees with the four-dimensional Einstein field equation derived in [@bat01; @yam07]. The electromagnetic field equation (\[meq1\]) is mathematically equivalent to the Gauss-Codacci relation. This is not surprising, since in both models (the model in this paper and the model in [@bat01; @yam07]) the effective field equations on the brane hypersurface are derived from the five-dimensional Einstein field equation by projection, and the boundary conditions on the brane are mathematically equivalent. In particular, the electromagnetic field equation (\[meq1\]) is derived from the Gauss-Codacci relation.
However, the physics represented by our model is different from that represented by the model in [@bat01; @yam07]. In our model, aside from the part resulted from projection of the bulk stress-energy tensor, the stress-energy tensor in the Einstein field equation on the brane contains two parts. One part, represented by $\tau_{ab}$, $\dot{\tau}_{ab}$, and $\pi_{ab}$ in equation (\[ein\_eq4\]), is interpreted as the stress-energy tensor of normal matter confined in the brane and related to the antisymmetric part of $K_{ab}$ (i.e., $-\hat{\Phi}_{ab}$) by the Israel junction relation (eqs. \[Tab\_m\_M\] and \[T\_m\_ab\]). The other part, represented by $T_{\emm,ab}$ in equation (\[ein\_eq4\]), is interpreted as the stress-energy tensor of electromagnetic fields and related to the symmetric part of $K_{ab}$ (i.e., $-\Psi_{ab}$; see eq. \[Tab\_em\_M\]). The $T_{\emm,ab}$ differs from the standard electromagnetic stress-energy tensor $^{(0)}T_{\emm,ab}$ by a $^{(1)}T_{\emm,ab}$, see equations (\[T\_em\_0\])–(\[T\_em\_1\]). As explained in [@li15a; @li15b], the $^{(1)}T_{\emm,ab}$ arises from the curvature-coupled term in the electromagnetic field equation (\[meq1\]).
Our interpretation of $T_{\emm,ab}$ as the stress-energy tensor of electromagnetic fields is motivated by the fact that the vector field equation, i.e., the Gauss-Codacci relation (\[vector\_eq\]), can be expressed in a form very similar to the Maxwell equation (eq. \[meq1\]). By identities $$\begin{aligned}
\nabla_aK^{ab}-\nabla^bK=-\nabla_a\Psi^{ab}-\nabla_a\Phi^{ab}\end{aligned}$$ and $$\begin{aligned}
\nabla_a\Psi^{ab}=\nabla_aF^{ab}+2R^b{}_aA^a \;, \label{Psi_F}\end{aligned}$$ equation (\[meq1\]) is easily derived from equation (\[vector\_eq\]). The electromagnetic field $F_{ab}$ is related to the $\Psi_{ab}$ by equation (\[Psi\_F\]). Equation (\[meq1\]) differs from the Maxwell equation by a curvature-coupled term, $2R^b{}_aA^a$. In a flat spacetime with a vanishing spacetime curvature, equation (\[meq1\]) is exactly the Maxwell equation. Hence, it is natural to consider equation (\[meq1\]) as generalization of the Maxwell equation in a curved spacetime [@li15a; @li15b].
The Gauss-Codacci relation was mentioned and briefly discussed in [@bat01; @yam07], but its relation to the Maxwell equation was not noticed. The authors of [@bat01; @yam07] did not find the electromagnetic field contained in the field equations on the brane. Derivation of the electromagnetic field equation and identification of the electromagnetic part in the total effective stress-energy tensor in the four-dimensional Einstein field equation are the major new contribution of the present work.
Summary and Discussion
======================
We have shown that, similar to the case in the KK theory, electromagnetic fields on a four-dimensional brane can be derived from the gravity in the five-dimensional bulk spacetime. The electromagnetic field is contained in the extrinsic curvature tensor of the brane hypersurface and obeys the field equation (\[meq1\]), which differs from the Maxwell equation by a curvature-coupled term. Since by definition $N^a$ and $A^a$ are vectors tangent to the brane hypersurface, the electromagnetic field can only be seen on the brane and hence its effect is naturally confined on the brane. When $R_{ab}\neq 0$ the field equation is not gauge invariant. However, in a Ricci-flat spacetime with $R_{ab}=0$, the field equation (\[meq1\]) becomes the Maxwell equation and gauge symmetry is restored. Hence, the electromagnetic field equation (\[meq1\]) can be considered as generalization of the Maxwell equation to a curved spacetime, as an alternative to the Einstein-Maxwell equation. The curvature-coupled term $2R^b{}_aA^a$ can be regarded as a pseudo-charge current vector, whose effect is testable in an environment with high mass and energy density [@li15a; @li15b].
With appropriate boundary conditions (eq. \[join2\]), the four-dimensional Einstein field equation is derived, which is given by equation (\[ein\_eq4\]). The right-hand-side of the Einstein field equation explicitly contains the stress-energy tensor of electromagnetic fields defined by equation (\[Tab\_em\_M\]). By the relation $$\begin{aligned}
\t{R}=R-K_{ab}K^{ab}+K^2-2\t{\nabla}_av^a \;,\end{aligned}$$ where $v^a=Kn^a-a^a$, the five-dimensional Einstein-Hilbert action can be written as $$\begin{aligned}
S_\EH=\int \sqrt{-g}\left(L_G+L_\emm+L_m+L_\intt\right)d\t{V} \;, \end{aligned}$$ where $d\t{V}=2^{-1}dx^0dx^1dx^2dx^3dw^\prime$, $L_G=R$, $L_m=-g^{abcd}\dot{g}_{ab}\dot{g}_{cd}$, $L_\intt=-4g^{abcd}A_a\nabla_b\dot{g}_{cd}$, and $$\begin{aligned}
L_\emm = -4\sqrt{-g}\,\left(\frac{1}{4}F_{ab}F^{ab}-R_{ab}A^aA^b\right) \;.\end{aligned}$$ It can be verified that the variation of $S_\EH$ with respect to $A_a$ leads to the electromagnetic field equation (\[meq1\]). The variation of $S_\EH$ with respect to $g^{ab}$ leads to the four-dimensional Einstein field equation with the stress-energy tensor given by equations (\[Tab\_em\_M\])–(\[Tab\_int\_M\]) (see [@li15a] for detail).
The number of degrees of freedom (d.o.f) of gravity determined by the four-dimensional Einstein field equation is two. The number of d.o.f of the electromagnetic field determined by equation (\[meq1\]) is three, since the presence of the curvature-coupled term causes violation of gauge symmetry. This fact means that the curvature-coupled term in the electromagnetic field equation causes an effective mass to photons. Hence, the total number of d.o.f is five, which is equal to the number of d.o.f of the five-dimensional gravity.
So far we have not discussed the scalar constraint equation (\[scalar\_eq\]) yet. In fact, it can be replaced by another scalar equation obtained from the identity $E_a{}^a=0$, or equivalently, from substitution of equation (\[scalar\_eq\]) into the trace of equation (\[tensor\_eq\]). So, the scalar constraint equation gives essentially a constraint on the trace of $\t{\pounds}_nK_{ab}$ [@li15a]. It can be derived that equation (\[scalar\_eq\]) leads to $$\begin{aligned}
S_{ab}\Psi^{ab}-\frac{1}{3}S\Psi=-\left[\t{T}_{ab}n^an^b\right] \;,\end{aligned}$$ where the right-hand side is the difference in the bulk pressure acting on the two sides of the brane.
Similarly, from equation (\[vector\_eq\]) (i.e., eq. \[meq1\]) we can derive that $\nabla_aS^{ab}=0=\nabla_a\tau^{ab}$, if $\left[g^{ab}\t{T}_{ac}n^c\right]=0$. This is just the conservation equation for $S_{ab}$ and $\tau_{ab}$. Then $\nabla^a\Phi^+_{ab}=0$, and by equation (\[Ja\]) we have $J^a=(\t{\kappa}/4\pi)g^{ab}\t{T}_{bc}n^c$. When $J^a=0$, we have $\nabla^a{\cal T}_{ab}=0$ and $\nabla^aT_{\emm,ab}=0$, then by equation (\[ein\_eq4\]) we get $\nabla^a\dot{\tau}_{ab}=\t{\kappa}\nabla^a\pi_{ab}$.
This work was supported by the National Basic Research Program (973 Program) of China (Grant No. 2014CB845800) and the NSFC grant (No. 11373012).
[99]{}
L. Randall & R. Sundrum, Phys. Rev. Lett. [**83**]{}, 3370 (1999a) L. Randall & R. Sundrum, Phys. Rev. Lett. [**83**]{}, 4690 (1999b) T. Shiromizu, K. Maeda, & M. Sasaki, Phys. Rev. D [**62**]{}, 024012 (2000) P. Horava & E. Witten, Nucl. Phys. B [**460**]{}, 506 (1996) A. Lukas, B. A. Ovrut, K. S. Stelle, & D. Waldram, Phys. Rev. D [**59**]{}, 086001 (1999) A. Lukas, B. A. Ovrut, & D. Waldram, Phys. Rev. D [**60**]{}, 086001 (1999) A. Lukas, B. A. Ovrut, & D. Waldram, Phys. Rev. D [**61**]{}, 023506 (2000) R. Maartens & K. Koyama, Living Rev. Relativity [**13**]{}, 5 (2010) T. Kaluza, Sitzungsber. Preuss. Akad. Wiss, 966 (1921) O. Klein, Nature [**118**]{}, 516 (1926a) O. Klein, Zeit. Phys. [**37**]{}, 895 (1926b) D. Bailin & A. Love, Rep. Prog. Phys. [**50**]{}, 1087 (1987) J. M. Overduin & P. S. Wesson, Phys. Rep. [**283**]{}, 303 (1997) L.-X. Li, Front. Phys. [**11**]{}, 110402 (2016) (arXiv:1511.01260) C. W. Misner, K.S. Thorne, & J. A. Wheeler, Gravitation. W. H. Freeman, New York (1973) R. M. Wald, General Relativity. The University of Chicago Press, Chicago (1984) W. Israel, Il Nuovo Cimento B [**44**]{}, 1 (1966) L.-X. Li, Gen. Rel. Grav. [**48**]{}, 28 (2016) (arXiv:1508.06910) R. A. Battye, B. Carter, A. Mennim, & J.-P. Uzan, Phys.Rev. D [**64**]{}, 124007 (2001) D. Yamauchi & M. Sasaki, Prog.Theor.Phys. [**118**]{}, 245 (2007)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper applies G. Lyubeznik’s notion of $F$-finite modules to describe in a very down-to-earth manner certain annihilator submodules of some top local cohomology modules over Gorenstein rings. As a consequence we obtain an explicit description of the test ideal of Gorenstein rings in terms of ideals in a regular ring.'
address: |
Department of Pure Mathematics, University of Sheffield, Hicks Building, Sheffield S3 7RH, United Kingdom\
[*Fax number*]{}: +44-(0)114-222-3769
author:
- Mordechai Katzman
title: '$F$-stable submodules of top local cohomology modules of Gorenstein rings'
---
Introduction {#Section: Introduction}
============
*Throughout this paper $(R,\frak{m})$ will denote a regular local ring of characteristic $p$, and $A$ will be a surjective image of $R$. We also denote the injective hull of $R/m$ with $E$ and for any $R$-module $N$ we write $\operatorname{Hom}_R (N,E)$ as $N^\vee$.* We shall always denote with $f:R\rightarrow R$ the Frobenius map, for which $f(r)= r^p$ for all $r \in R$ and we shall denote the $e$th iterated Frobenius functor over $R$ with $F^e_R(-)$. As $R$ is regular, $F^e_R(-)$ is exact (cf. Theorem 2.1 in [@Kunz].)
For any commutative ring $S$ of characteristic $p$, the skew polynomial ring $S[T; f]$ associated to $S$ and the Frobenius map $f$ is a non-commutative ring which as a left $R$-module is freely generated by $(T^i)_{i \geq 0}$, and so consists of all polynomials $\sum_{i = 0}^n s_i T^i$, where $n\geq 0$ and $s_0,\ldots,s_n \in S$; however, its multiplication is subject to the rule $$Ts = f(s)T = s^pT \quad \mbox{~for all~} s \in S\/.$$
Any $A[T; f]$-module $M$ is a $R[T; f]$-module in a natural way and, as $R$-modules, $F^e_R(M)\cong RT^e \otimes_R M$.
It has been known for a long time that the local cohomology module $\H^{\dim A}_{\mathfrak{m}A} (A)$ has the structure of an $A[T; f]$-module and this fact has been employed by many authors to study problems related to tight closure and to Frobenius closure. Recently R. Y. Sharp has described in [@Sharp] the parameter test ideal of $F$-injective rings in terms of certain $A[T; f]$-submodules of $\H^{\dim A}_{\mathfrak{m}A} (A)$ and it is mainly this work which inspired us to look further into the structure of these $A[T; f]$-modules.
The main aim of this paper is to produce a description of the $A[T; f]$-submodules of $\H^{\dim A}_{\mathfrak{m}A} (A)$ in terms of ideals of $R$ with certain properties. We first do this when $A$ is a complete intersection. The $F$-injective case is described by Theorem \[Theorem: bijection\] and as a corollary we obtain a description of the parameter test ideal of $A$. Notice that for Gorenstein rings the test ideal the parameter test ideal coincide (cf. Proposition 8.23(d) in [@Hochster-Huneke-0] and Proposition 4.4(ii) in [@Smith2].) We then proceed to describe the parameter test ideal in the non-$F$-injective case (Theorem \[Theorem: non-$F$-injective\].) We generalise these results to Gorenstein rings in section \[Section: The Gorenstein case\].
Preliminaries: $F$-finite modules {#Section: Preliminaries: $F$-finite modules}
=================================
The main tool used in this paper is the notion of $F$-modules, and in particular $F$-finite modules. These were introduced in G. Lyubeznik’s seminal work [@Lyubeznik] and provide a very fruitful point of view of local cohomology modules in prime characteristic $p$.
One of the tools introduced in [@Lyubeznik] is a functor $\mathcal{H}_{R,A}$ from the category of $A[T; f]$-modules which are Artinian as $A$-modules to the category of $F$-finite modules. For any $A[T; f]$-module $M$ which is Artinian as an $A$-module the $F$-finite structure of $\mathcal{H}_{R,A}(M)$ is obtained as follows. Let $\gamma : R T \otimes_R M \rightarrow M$ be the $R$-linear map defined by $\gamma (r T \otimes m) = r T m$; apply the functor $^\vee$ to obtain $\gamma^\vee : M^\vee \rightarrow F_R(M)^\vee$. Using the isomorphism between $F_R(M)^\vee$ and $F_R(M^\vee)$ (Lemma 4.1 in [@Lyubeznik]) we obtain a map $\beta: M^\vee \rightarrow F_R(M^\vee)$ which we adopt as a generating morphism of $\mathcal{H}_{R,A}(M)$.
*We shall henceforth assume that the kernel of the surjection $R\rightarrow A$ is minimally generated by $\mathbf{u}=(u_1, \dots, u_n)$.* We shall also assume until section \[Section: The Gorenstein case\] that $A$ is a complete intersection. We shall write $u=u_1\cdot \ldots \cdot u_n$ and for all $t\geq 1$ we let $\mathbf{u}^t R$ be the ideal $u_1^t R + \dots + u_n^t R$.
To obtain the results in this paper we shall need to understand the $F$-finite module structure of $$\mathcal{H}_{R,A}\left( \H^{\dim A}_{\mathfrak{m}A} (A) \right) \cong \H^{\dim R - \dim A}_{\mathbf{u}R}(R) ;$$ this has generating root $$\frac{R}{\mathbf{u} R} \xrightarrow[]{u^{p-1}} \frac{R}{\mathbf{u}^p R}$$ (cf. Remark 2.4 in [@Lyubeznik].)
Define $\mathcal{I}(R,\mathbf{u})$ to be the set of all ideals $I\subseteq R$ containing $(u_1, \dots , u_n)R$ with the property that $$u^{p-1} \left(I + \mathbf{u}R\right) \subseteq I^{[p]}+\mathbf{u}^pR .$$
\[Lemma: generating morphisms\] Consider the $F_R$-finite $F$-module $M=\H^n_{\mathbf{u} R}(R)$ with generating root $$\frac{R}{\mathbf{u} R} \xrightarrow[]{u^{p-1}} \frac{R}{\mathbf{u}^pR} .$$
1. For any $I\in \mathcal{I}(R,\mathbf{u})$ the $F_R$-finite module with generating root $$\frac{I+ \mathbf{u}R}{\mathbf{u} R} \xrightarrow[]{u^{p-1}} \frac{I^{[p]}+\mathbf{u}^p R}{\mathbf{u}^p R} \cong F_R\left( \frac{I+\mathbf{u}R}{\mathbf{u}R} \right)$$ is an $F$-submodule of $M$ and every $F_R$-finite $F$-submodule of $M$ arises in this way.
2. For any $I\in \mathcal{I}(R,\mathbf{u})$ the $F_R$-finite module with generating morphism $$\frac{R}{I+\mathbf{u}R} \xrightarrow[]{u^{p-1}} \frac{R}{I^{[p]}+\mathbf{u}^p R} \cong F_R\left( \frac{R}{I+\mathbf{u} R}\right)$$ is an $F$-module quotient of $M$ and every $F_R$-finite $F$-module quotient of $M$ arises in this way.
\(a) For any $I\in \mathcal{I}(R,\mathbf{u})$, the map $$\frac{I+\mathbf{u}R}{\mathbf{u} R} \xrightarrow[]{u^{p-1}} \frac{I^{[p]}+\mathbf{u}^p R}{\mathbf{u}^p R}$$ is well defined and is injective; now the first statement follows from Proposition 2.5(a) in [@Lyubeznik]. If $N$ is any $F_R$-finite $F$-submodule of $M$, the root of $N$ is a submodule of the root of $M$, i.e., the root of $N$ has the form $(I+\mathbf{u}R)/\mathbf{u}R$ for some ideal $I\subseteq R$ (cf. [@Lyubeznik], Proposition 2.5(b)) and the structure morphism of $N$ is induced by that of $M$, i.e., by multiplication by $u^{p-1}$, so we must have $u^{p-1} I \subseteq I^{[p]}+\mathbf{u}^p R $, i.e., $I\in \mathcal{I}(R,\mathbf{u})$.
\(b) For any $I\in \mathcal{I}(R,\mathbf{u})$, the map $$\frac{R}{I+\mathbf{u}R} \xrightarrow[]{u^{p-1}} \frac{R}{I^{[p]}+\mathbf{u}^p R} \cong F_R\left( \frac{R}{I+\mathbf{u}R}\right)$$ is well defined and we have the following commutative diagram with exact rows $$\label{CD3}
\xymatrix{
0 \ar@{>}[r]^{} &
\displaystyle\frac{I+\mathbf{u}R}{\mathbf{u}R} \ar@{>}[r]^{} \ar@{>}[d]^{u^{p-1}} &
\displaystyle\frac{R}{\mathbf{u}R} \ar@{->}[r]^{} \ar@{>}[d]^{u^{p-1}} &
\displaystyle\frac{R}{I+\mathbf{u}R} \ar@{>}[r]^{} \ar@{>}[d]^{u^{p-1}} &
0 \\
0 \ar@{>}[r]^{} &
\displaystyle F_R\left(\frac{I+\mathbf{u}R}{\mathbf{u}R}\right) \ar@{>}[r]^{} \ar@{>}[d]^{u^{p(p-1)}} &
\displaystyle F_R\left(\frac{R}{\mathbf{u}R}\right) \ar@{>}[r]^{} \ar@{-}[d]^{u^{p(p-1)}} &
\displaystyle F_R\left(\frac{R}{I+\mathbf{u}R}\right) \ar@{>}[r]^{} \ar@{>}[d]^{u^{p(p-1)}} &
0 \\
& \vdots &\vdots&\vdots&\\
}$$ Taking direct limits of the vertical maps we obtain an exact sequence $0 \rightarrow M^{\prime} \rightarrow M \rightarrow M^{\prime\prime} \rightarrow 0$ which establishes the first statement of (b).
Conversely, if $M^{\prime\prime}$ is a $F$-module quotient of $M$, say, $M^{\prime\prime}\cong M/M^{\prime}$ for some $F$-submodule $M^{\prime}$ of $M$ use (a) to find a generating root of $M^{\prime}$ of the form $$\frac{I+\mathbf{u}R}{\mathbf{u} R} \xrightarrow[]{u^{p-1}} \frac{I^{[p]}+\mathbf{u}^pR}{\mathbf{u}^pR}$$ for some $I\in \mathcal{I}(R,\mathbf{u})$. Looking again at the direct limits of the vertical maps in (\[CD3\]) we establish the second statement of (b).
For all $I\in \mathcal{I}(R,\mathbf{u})$ we define $\mathcal{N}(I)$ to be the $F$-module quotient of $\H^n_{\mathbf{u}R}(R)$ with generating morphism $$\frac{R}{I+\mathbf{u}R} \xrightarrow[]{u^{p-1}} \frac{R}{I^{[p]}+\mathbf{u}^pR} \cong F_R\left( \frac{R}{I+\mathbf{u}R}\right) .$$
\[Lemma 2\] Assume that $R$ is complete. Let $H$ be an Artinian $A[T; f]$-module and write $M=\mathcal{H}_{R,A}(H)$. Let $N$ be a homomorphic image of $M$ with generating morphism $N_0$. Then $N_0^\vee$ is an $A[T; f]$-submodule of $H$ and $N\cong \mathcal{H}_{R,A}(N_0^\vee)$.
Notice that $M$ (and hence $N$) are $F$-finite modules (cf. [@Lyubeznik], Theorems 2.8 and 4.2). Let $N_0$ be root of $N$ and $M_0$ a root of $M$ so that we have a commutative diagram with exact rows $$\label{CD1}
\xymatrix{
M_0 \ar@{->>}[r]^{} \ar@{->}[d]^{\mu} & N_0 \ar@{->}[r]^{} \ar@{->}[d]^{\nu} & 0 \\
F_R(M_0) \ar@{->>}[r]^{} & F_R(N_0) \ar@{->}[r]^{} & 0
}$$ where the vertical arrows are generating morphisms. Apply the functor $\operatorname{Hom}(-,E)$ to the commutative diagram above to obtain the following commutative diagram with exact rows $$\label{CD2}
\xymatrix{
0 \ar@{->}[r]^{} & F_R(N_0)^\vee \ar@{->}[r]^{} \ar@{->}[d]^{\nu^\vee} & F_R(M_0)^\vee \ar@{->}[d]^{\mu^\vee}\\
0 \ar@{->}[r]^{} & N_0^\vee \ar@{->}[r]^{} & M_0^\vee
}$$ and recall that $M_0$ is isomorphic to $H^\vee$ (cf. [@Lyubeznik], Theorem 4.2). Since $R$ is complete, $\left(H^\vee\right)^\vee\cong H$ and we immediately see that $N_0^\vee$ is a $R$-submodule of $H$. We now show that $N_0^\vee$ is an $A[T;f]$ submodule of $H$ by showing that $T N_0^\vee \subseteq N_0^\vee$.
The construction of the functor $\mathcal{H}_{R,A}(-)$ is such that for any $h\in H \cong M_0^\vee$, $T h$ is the image of $T \otimes_R h$ under the map $$F_R(M_0)^\vee \xrightarrow[]{\mu^\vee} M_0^\vee$$ and so for $h\in N_0^\vee$, $T h$ is the image of $T \otimes_R h$ under the map $$F_R(N_0)^\vee \xrightarrow[]{\nu^\vee} N_0^\vee$$ and hence $T h\in N_0^\vee$.
Now the fact that $N\cong \mathcal{H}_{R,A} \left(N_0^\vee\right)$ follows the construction of the functor $\mathcal{H}_{R,A}(-)$.
Let $M$ be a left $A[T, f]$-module. We shall write $AT^\alpha M$ for the $A$-module generated by $T^\alpha M$. Note that $AT^\alpha M$ is a left $A[T, f]$-module. We shall also write $\displaystyle M^\bigstar=\bigcap_{\alpha\geq 0} AT^\alpha M$.
\[Lemma 4\] Assume that $R$ is complete. Let $H$ be an $A[T;f]$-module and assume that $H$ is $T$-torsion-free. Let $I,J \subseteq A$ be ideals. If, for some $\alpha\geq 0$, $$A T^\alpha \operatorname{ann}_H {I A[T; f]} = A T^\alpha \operatorname{ann}_H {J A[T; f]}$$ then $\operatorname{ann}_H { I A[T; f]} =\operatorname{ann}_H { J A[T; f]} $.
Both $A T^\alpha \operatorname{ann}_H {I A[T; f]} $ and $A T^\alpha \operatorname{ann}_H {J A[T; f]} $ are left $A[T; f]$-submodules. Now for every $T$-torsion-free $A[T;f]$-module $M$, and every ideal $K\subseteq A$, if $$\left(\bigoplus_{i\geq 0} K T^i\right) AT^\alpha M = \left(\bigoplus_{i\geq 0} K T^{i+\alpha}\right) M$$ vanishes then so does $$\left(\bigoplus_{i\geq 0} K^{[p^\alpha]} T^{i+\alpha}\right) M=
\left(\bigoplus_{i\geq 0} T^{\alpha} K T^{i}\right) M=
T^{\alpha} \left(\bigoplus_{i\geq 0} K T^{i}\right) M$$ and since $M$ is $T$-torsion-free, $$\left(\bigoplus_{i\geq 0} K T^{i}\right) M=0 .$$ We deduce that $\operatorname{gr-ann}A T^\alpha M=\operatorname{gr-ann}M$. Now $$\operatorname{gr-ann}A T^\alpha \big(\operatorname{ann}_H {I A[T; f]} \big) = \operatorname{gr-ann}\operatorname{ann}_H {I A[T; f]} ,$$ $$\operatorname{gr-ann}A T^\alpha \big(\operatorname{ann}_H {J A[T; f]} \big)= \operatorname{gr-ann}\operatorname{ann}_H {J A[T; f]}$$ and Lemma 1.7 in [@Sharp] shows that $\operatorname{ann}_H{ I A[T; f]} =\operatorname{ann}_H { J A[T; f]} $.
The $A[T;f]$ module structure of top local cohomology modules of $F$-injective Gorenstein rings
===============================================================================================
As in [@Smith2] we say that an ideal $I\subseteq A$ is an *$F$-ideal* if $\operatorname{ann}_{\H^{\dim(A)}_{\frak{m} A}(A)} I$ is a left $A[T; f]$-module, i.e., if $\operatorname{ann}_{\H^{\dim(A)}_{\frak{m} A}(A)} I=\operatorname{ann}_{\H^{\dim(A)}_{\frak{m} A}(A)} IA[T; f]$.
\[Theorem 1\] Assume that $R$ is complete. Consider the $F_R$-finite $F$-module $ M=\H^n_{\mathbf{u}R}(R)$ with generating root $$\frac{R}{\mathbf{u}R} \xrightarrow[]{u^{p-1}} \frac{R}{\mathbf{u}^pR}$$ and consider the Artinian $A[T; f]$ module $H=\H^{\dim(A)}_{\frak{m} A}(A)$. Let $N$ be a homomorphic image of $M$.
1. $M=\mathcal{H}_{R,A}(-)(H)$ and has generating root $H^\vee \cong R/\mathbf{u}R \xrightarrow[]{u^{p-1}} R/\mathbf{u}^p R \cong F_R(H^\vee)$.
2. If $N$ has generating morphism $$\frac{R}{I+\mathbf{u}R} \xrightarrow[]{u^{p-1}} \frac{R}{I^{[p]}+\mathbf{u}^pR}$$ then $I A$ is an $F$-ideal, $N \cong \mathcal{H}_{R,A}\big( \operatorname{ann}_H {IA[T;f]} \big)$. If, in addition, $H$ is $T$-torsion free then $\operatorname{gr-ann}\operatorname{ann}_H {IA[T;f]}=IA[T;f]$ and $I$ is radical.
3. Assume that $H$ is $T$-torsion free (i.e., $H_r=H$ in the terminology of [@Lyubeznik]). For any ideal $J\subset R$, the $F$-finite module $\mathcal{H}_{R,A}\big(\operatorname{ann}_H {JA[T;f]} \big)$ has generating morphism $$\frac{R}{I+\mathbf{u}R} \xrightarrow[]{u^{p-1}} \frac{R}{I^{[p]}+\mathbf{u}^pR}$$ for some ideal $I\in \mathcal{I}(R,\mathbf{u})$ with $\operatorname{ann}_H { I A[T; f]} =\operatorname{ann}_H { J A[T; f]} $.
The first statement is a restatement of the discussion at the beginning of section \[Section: Preliminaries: $F$-finite modules\].
Notice that Lemma \[Lemma: generating morphisms\] implies that $N$ must have a generating morphism of the form given in (b) for some $I\in \mathcal{I}(R,\mathbf{u})$.
Since $A$ is Gorenstein, $H$ is an injective hull of $A/\frak{m}A$ which we denote $\overline{E}$. Lemma \[Lemma 2\] implies that $ N \cong \mathcal{H}_{R,A}\left( L \right)$ where $\displaystyle L= \left( \frac{R}{I+\mathbf{u}R} \right)^\vee$ is a $A[T; f]$-submodule of $H=\overline{E}$. But $$\begin{aligned}
\left( \frac{R}{I+\mathbf{u}R} \right)^\vee&=& \operatorname{ann}_E \left(I+\mathbf{u}R\right) \\
&=& \operatorname{ann}_{\left( \operatorname{ann}_{\mathbf{u}R} E \right)} I \\
&=& \operatorname{ann}_{\overline{E}} I .\end{aligned}$$ But $L$ is a $A[T; f]$-submodule of $\overline{E}$ and so $I A$ is an $F$-ideal and $L= \operatorname{ann}_{\overline{E}} {I A[T; f]}$. Also, $$\begin{aligned}
\left( 0 :_R \operatorname{ann}_{\overline{E}} {I A[T; f]} \right) &= & \left( 0 :_R \operatorname{ann}_E I \right) \\
&=&\left( 0 :_R (R/I)^\vee \right) \\
&=&\left( 0 :_R (R/I) \right) \\
&=& I\end{aligned}$$ (where the third equality follows from 10.2.2 in [@Brodmann-Sharp]) If $H$ is $T$-torsion free, Proposition 1.11 in [@Sharp] implies that $I=\operatorname{gr-ann}\operatorname{ann}_{\overline{E}} {I A[T; f]}$ and Lemma 1.9 in [@Sharp] implies that $I$ is radical.
To prove part (c) we recall Lemma \[Lemma: generating morphisms\] which states that $\mathcal{H}_{R,A}\big(\operatorname{ann}_H {JA[T;f]} \big)$ has generating morphism $$\frac{R}{I+\mathbf{u}R} \xrightarrow[]{u^{p-1}} \frac{R}{I^{[p]}+\mathbf{u}^pR}$$ for some $I\in \mathcal{I}(R,\mathbf{u})$ and we need only show that $\operatorname{ann}_H { I A[T; f]} =\operatorname{ann}_H { J A[T; f]} $.
Part (b) implies that $\mathcal{H}_{R,A}\big(\operatorname{ann}_H {JA[T;f]} \big) = \mathcal{H}_{R,A} \big(\operatorname{ann}_H {IA[T;f]} \big)$ for some $I\in \mathcal{I}(R,\mathbf{u})$ and Theorem 4.2 (iv) in [@Lyubeznik] implies $$\bigcap_{i=0}^\infty A T^i \big( \operatorname{ann}_H {JA[T;f]} \big)= \bigcap_{i=0}^\infty A T^i \big( \operatorname{ann}_H {IA[T;f]} \big)$$ and since $H$ is Artinian there exists an $\alpha\geq 0$ for which $A T^\alpha \big( \operatorname{ann}_H {JA[T;f]} \big)= A T^\alpha \big( \operatorname{ann}_H {IA[T;f]} \big) $ and the result follows from Lemma \[Lemma 4\].
Theorem \[Theorem 1\] can provide an easy way to show that $H=\H^{\dim(A)}_{\frak{m} A}(A)$ is *not* $T$-torsion free. As an example consider $R=\mathbb{K}[\![ x,y, a,b ]\!]$, $u=x^2a-y^2b$ and $A=R/uR$. Its easy to verify that $(x,y,a^2)R\in \mathcal{I}(R,x^2a-y^2b)$ when $\mathbb{K}$ has characteristic 2, and we deduce that $\H^{3}_{(x,y,a,b) A}(A)$ is not $T$-torsion free.
\[Theorem2\] Assume that $R$ is complete and that $\H^{\dim(A)}_{\frak{m} A}(A)$ is $T$-torsion free.
1. For all $A[T;f]$-submodules $L$ of $\H^{\dim(A)}_{\frak{m} A}(A)$, $$L^\bigstar=\bigcap_{i=0}^\infty A T^i L$$ has the form $A T^{\alpha} M$ where $\alpha\geq 0$ and $M$ is a special annihilator submodule in the terminology of [@Sharp].
2. The set $\big\{ \mathcal{N}(I) \,|\, I\in \mathcal{I}(R,\mathbf{u}) \big\}$ is finite.
\(a) Let $L$ be a $A[T;f]$-submodule of $\H^{\dim(A)}_{\frak{m} A}(A)$. Pick a $I\in \mathcal{I}(R,\mathbf{u})$ such that $\mathcal{N}(I)=\mathcal{H}_{R,A}\left( L \right)$. Now use part (b) of Theorem \[Theorem 1\] and deduce that $\mathcal{N}(I)\cong \mathcal{H}_{R,A}\left( \operatorname{ann}_H {I A[T;f]} \right)$. Now the result follows from Theorem 4.2 (iv) in [@Lyubeznik].
\(b) Theorem \[Theorem 1\](b) implies that $$\big\{ \mathcal{N}(I) \,|\, I\in \mathcal{I}(R,\mathbf{u}) \big\}=
\left\{ \mathcal{H}_{R,A} \left(\operatorname{ann}_{\H^{\dim A}_{\frak{m} A} (A) } {I A[T; f]} \right)\,|\, I\in \mathcal{I}(R,\mathbf{u}) \right\} ;$$ now Corollary 3.11 and Proposition 1.11 in [@Sharp] imply that the set on the right is finite.
The following Theorem reduces the problem of classifying all $F$-ideals of $A$ (in the terminology of [@Smith2]) or all special $\H^{\dim(A)}_{\frak{m} A}(A)$-ideals (in the terminology of [@Sharp]) in the case where $A$ is an $F$-injective complete intersection, to problem of determining the set $\mathcal{I}(R,\mathbf{u})$.
\[Theorem: bijection\] Assume $H:=\H^{\dim(A)}_{\frak{m} A}(A)$ is $T$-torsion free and let $\mathcal{B}$ be the set of all $H$-special $A$-ideals (cf. §0 in [@Sharp])
1. The map $\Psi: \mathcal{I}(R,\mathbf{u}) \rightarrow \mathcal{B}$ given by $\Psi(I)=I A$ is a bijection.
2. There exists a unique minimal element $\tau$ in $\left\{I \,|\, I\in \mathcal{I}(R,\mathbf{u}) , \ \operatorname{ht}I A >0 \right \}$ and that $\tau$ is a parameter-test-ideal for $A$.
3. $A$ is $F$-rational if and only if $\mathcal{I}(R,\mathbf{u})=\{ 0, R \}$.
\(a) Assume first that $R$ is complete. Theorem \[Theorem 1\](b) implies that $\Psi$ is well defined, i.e., $\Psi(I)\in \mathbf{B}$ for all $I\in\mathcal{I}(R,\mathbf{u})$, and, clearly, $\Psi$ is injective. The surjectivity of $\Psi$ is a consequence of Theorem \[Theorem 1\](c).
Assume now that $R$ is not complete, denote completions with $\widehat{\phantom{h}}$ and write $\widehat{H}=\H^{\dim(\widehat{A})}_{\frak{m} \widehat{A}}(\widehat{A})$. If $I$ is a $\widehat{H}$-special $\widehat{A}$-ideal, i.e., if there exists an $\widehat{A}[T; f]$-submodule $N\subseteq \widehat{H}$ such that $\operatorname{gr-ann}N=I \widehat{A}[T; f]$ then $I=(0 :_{\widehat{A}} N)$ (cf. Definition 1.10 in [@Sharp]). But recall that $\widehat{H}=H$ and $N$ is a $A[T; f]$-submodule of $H$; now $I=(0 :_{\widehat{A}} N)=(0 :_A N) \widehat{A} $. If we let $\widehat{\mathcal{B}}$ be the set of $\H^{\dim(\widehat{A})}_{\frak{m} \widehat{A}}(\widehat{A})$-special $\widehat{A}$-ideals, we have a bijection $ \Upsilon : \mathcal{B} \rightarrow \widehat{\mathcal{B}}$ mapping $I$ to $I\widehat{A}$. This also shows that all ideals in $\mathcal{I}(\widehat{R},\mathbf{u})$ are expanded from $R$, and now since $\widehat{R}$ is faithfully flat over $R$, we deduce that all ideals in $\mathcal{I}(\widehat{R},\mathbf{u})$ have the form $I \widehat{R}$ for some $I \in \mathcal{I}(R,\mathbf{u})$. We now obtain a chain of bijections $$\mathcal{I}(R,\mathbf{u}) \longleftrightarrow
\mathcal{I}(\widehat{R},\mathbf{u}) \longleftrightarrow
\widehat{\mathcal{B}} \longleftrightarrow
\mathcal{B} .$$
\(b) This is immediate from (a) and Corollary 4.7 in [@Sharp].
\(c) If $A$ is $F$-rational, $\H^{\dim(A)}_{\frak{m} A}(A)$ is a simple $A[T;f]$-module (cf. Theorem 2.6 in [@Smith3]) and the only $H$-special $A$-ideals must be $0$ and $A$. The bijection established in (a) implies now $\mathcal{I}(R,\mathbf{u})=\{ 0, R \}$.
Conversely, if $\mathcal{I}(R,\mathbf{u})=\{ 0, R \}$, part (b) of the Theorem implies that $1\in A$ is a parameter-test-ideal, i.e., for all systems of parameters $\mathbf{x}=(x_1, \dots, x_d)$ of $A$, $\left( \mathbf{x} A \right)^*=\left( \mathbf{x} A \right)^F= \mathbf{x} A $ where the second equality follows from the fact that $\H^{\dim(A)}_{\frak{m} A}(A)$ is $T$-torsion free.
Examples
========
Throughout this section $\mathbb{K}$ will denote a field of prime characteristic.
\[Example 1\] Let $R$ be the localization of $\mathbb{K}[ x,y ]$ at $(x,y)$, $u=xy$ and $A=R/uR$. Then $\H^1_{xy R}(R)=\mathcal{H}_{R,A}\big(\H^1_{xA+yA}(A)\big)$ ought to have four proper $F$-finite $F$-submodules corresponding to the elements $0$, $xR$, $yR$ and $xR+yR$ of $\mathcal{I}(R,xy)$.
We verify this by giving an explicit description the $A[T;f]$-module structure of $$H:=\H^1_{xA+yA}(A)\cong
\lim_{\longrightarrow}
\left(
\frac{A}{(x-y) A } \xrightarrow[]{x-y}
\frac{A}{(x-y)^2 A} \xrightarrow[]{x-y}
\frac{A}{(x-y)^3 A} \xrightarrow[]{x-y} \dots
\right)$$ First notice that in $H$, for all $n\geq 1$ and $0<\alpha\leq n$, $x^\alpha + (x-y)^n A= x + (x-y)^{n-\alpha+1}$ and $y^\alpha + (x-y)^n A= y + (x-y)^{n-\alpha+1}$ so $H$ is the $\mathbb{K}$-span of $\left\{x + (x-y) A\right\} \cup X \cup Y \cup U$ where $$\begin{array}{l}
X=\big\{ x + (x-y)^n A \,|\, n\geq 2 \big\},\\
Y=\big\{ y + (x-y)^n A \,|\, n\geq 2 \big\},\\
U=\big\{ 1 + (x-y)^n A \,|\, n\geq 1 \big\}
\end{array}$$ and notice also that the action of the Frobenius map $f$ on $H$ is such that $ T\left( x^\alpha + (x-y)^n A\right) = x^{\alpha p} + (x-y)^{n p} A $ and $ T\left( y^\alpha + (x-y)^n A \right) = y^{\alpha p} + (x-y)^{n p} A $ for all $\alpha\geq 0$.
Next notice that any $A[T, f]$-submodule $M$ of $H$ which contains an element $1 + (x-y)^n A \in U$ must coincide with $H$: for $1\leq m<n$ we have $(x-y)^{n-m}\left( 1 + (x-y)^n A \right) = (x-y)^{n-m} + (x-y)^n A = 1 + (x-y)^m A$, whereas for $m>n$, pick an $e\geq 0$ such that $n p^e>m$, write $$T^e \left( 1 + (x-y)^n A \right)= 1 + (x-y)^{n p^e}A \in M$$ and use the previous case ($m<n$) to deduce that $1 + (x-y)^m A\in M$. Since now $U\subseteq M$, we see that $M=H$.
We now show that there are only three non-trivial $A[T, f]$-submodules of $H$, namely $\operatorname{Span}_\mathbb{K} X$ and $\operatorname{Span}_\mathbb{K} Y$, and $\operatorname{Span}_\mathbb{K} \left\{x + (x-y) A\right\} \cup X$. By symmetry, it is enough to show that, if $M$ is an $A[T, f]$-submodule of $H$ and $x + (x-y)^n A\in M$ for some $n\geq 2$, then $X\subset M$. If $1\leq m<n$, $$x^{n-m} \left( x + (x-y)^n A \right)=
x^{n-m+1} + (x-y)^n A=
x + (x-y)^{n-(n-m)}A=
x + (x-y)^m A$$ whereas, if $m>n\geq 2$, pick an $e\geq 0$ such that $n p^e-p^e+1>m$ and write $$T^e \left( x + (x-y)^n A \right)=
x^{p^e} + (x-y)^{n p^e} A =
x + (x-y)^{n p^e-p^e+1} A \in M$$ and using the previous case ($m<n$) we deduce that $x + (x-y)^m A \in M$.
Let $R$ be the localization of $\mathbb{K}[x,y,z]$ at ${\frak m}=(x,y,z)$, $u=x^2y+xyz+z^3$ and $A=R/uR$. Fedder’s criterion (cf. Propositon 2.1 in [@Fedder]) implies that $A$ is $F$-pure, and Lemma 3.3 in [@Fedder] implies that the $A[T;f]$ module $\H^1_{{\frak m}A}(A)$ is $T$-torsion-free.
Here $\mathcal{I}(R,u)$ contains the ideals $0$, $xR+zR$ and $xR+yR+zR$. We deduce that $A$ is not $F$-rational and that its parameter-test-ideal is $xR+zR$. Also, Theorem \[Theorem: bijection\](b) implies that the only proper ideals in $\mathcal{I}(R,u)$ are the ones listed above.
Let $R$ be the localization of $\mathbb{K}[ x,y,z ]$ at ${\frak m}=(x,y,z)$ and assume that $\mathbb{K}$ has characteristic $2$. Let $u=x^3+y^3+z^3+xyz$ and $A=R/uR$. Notice that we can factor $u=(x+y+z)(x^2+y^2+z^2+xy+xz+yz)$. Fedder’s criterion implies that $A$ is $F$-pure, and Lemma 3.3 in [@Fedder] implies that the $A[T;f]$ module $\H^1_{{\frak m}A}(A)$ is $T$-torsion-free.
Here $$\begin{aligned}
\mathcal{I}(R,u) & \supseteq & \big\{ 0, (x+y+z)R, (x^2+y^2+z^2+xy+xz+yz)R, \\
&& (x+z,y+z)R, (x+y+z,y^2+yz+z^2)R, \\
&& (x,y,z)R \big\} .\end{aligned}$$ The images in $A$ of the first three ideals have height zero while the images in $A$ of the fourth and fifth ideals have height $1$. Using \[Theorem: bijection\](b) we conclude that that the parameter test-ideal of $A$ is a sub-ideal of $$J=(x+z,y+z)A \cap (x+y+z,y^2+yz+z^2)A=(x^2+yx, y^2+xz, z^2+xy)A .$$ But this ideal defines the singular locus of $A$ and Theorem 6.2 in [@Hochster-Huneke-1] implies that the parameter test-element of $A$ contains $J$, so $J$ is the parameter test-ideal of $A$.
The non-$F$-injective case
==========================
In this section we extend the results of the previous section to the case where $A$ is not $F$-injective. First we produce a criterion for the $F$-injectivity of $A$.
Define $$\mathcal{I}_0(R,\mathbf{u})=\left\{ L\in \mathcal{I}(R,\mathbf{u}) \,|\,
u^{(p-1)(1+p+\dots+p^{e-1})} \in L^{[p^e]} + \mathbf{u}^{p^e} R \text{ for some } e\geq 1 \right\} .$$
1. For any $L\in \mathcal{I}(R,\mathbf{u})$, $\mathcal{N}(L)=0$ if and only if $L\in \mathcal{I}_0(R,\mathbf{u})$.
2. $\H^{\dim(A)}_{\frak{m} A}(A)$ is $T$-torsion free if and only if $\mathcal{I}_0(R,\mathbf{u})=\{R\}$.
\(a) Recall that the $F$-finite module $\mathcal{N}(L)$ has generating morphism $$\frac{R}{L+\mathbf{u}R} \xrightarrow[]{u^{p-1}}
\frac{R}{L^{[p]}+\mathbf{u}^pR} \cong F_R\left( \frac{R}{L+\mathbf{u}R}\right) .$$ Proposition 2.3 in [@Lyubeznik] implies that $\mathcal{N}(L)=0$ if and only if for some $e\geq 1$ the composition $$\frac{R}{L+\mathbf{u}R}
\xrightarrow[]{u^{p-1}}
\frac{R}{L^{[p]}+\mathbf{u}^pR}
\xrightarrow[]{u^{(p-1)p}}
\frac{R}{L^{[p]}+\mathbf{u}^{p^2}R}
\dots
\xrightarrow[]{u^{(p-1)p^{e-1}}}
\frac{R}{L^{[p]}+\mathbf{u}^{p^e}R}$$ vanishes, i.e., if and only if $u^{(p-1)(1+p+\dots+p^{e-1})} \in L^{[p^e]} + \mathbf{u}^{p^e} R $ for some $e\geq 1$.
\(b) Write $H=\H^{\dim(A)}_{\frak{m} A}(A)$. If $H$ is $T$-torsion free, the existence of the bijection described in Theorem \[Theorem: bijection\](a) implies that for any non-unit $L\in\mathcal{I}_0(R,\mathbf{u})$, $\operatorname{ann}_H {L A[T;f]}\neq \operatorname{ann}_H {A[T;f]}=0$. Theorem \[Theorem 1\](b) implies $\mathcal{N}(L)\cong \mathcal{H}_{R,A}\big( \operatorname{ann}_H {L A[T;f]} \big)$ so $\mathcal{H}_{R,A}\big( \operatorname{ann}_H {L A[T;f]} \big)=0$. But Theorem 4.2(ii) in [@Lyubeznik] now implies that $\operatorname{ann}_H {L A[T;f]}$ is nilpotent, a contradiction.
Assume now that $H$ is not $T$-torsion free, i.e., $H_n\neq 0$. The short exact sequence $$0 \rightarrow H_n \rightarrow H \rightarrow H/H_n \rightarrow 0$$ yields the short exact sequence $$0 \rightarrow \big(H/H_n\big)^\vee \rightarrow \frac{R}{\mathbf{u} R} \rightarrow H_n^\vee \rightarrow 0 .$$ Notice that as the functor $\operatorname{Hom}(-, E)$ is faithful, $H_n^\vee \neq 0$, and so $H_n^\vee\cong R/I$ for some ideal $\mathbf{u} R \subseteq I\varsubsetneq R$. Now $\mathcal{H}_{R,A}\big( H_n \big)$ is the $F$-finite quotient of $H$ with generating morphism $$\frac{R}{I} \xrightarrow[]{u^{p-1}} \frac{R}{I^{[p]}}$$ and this vanishes because of Theorem 4.2(ii) in [@Lyubeznik], i.e., $I\in \mathcal{I}_0(R,\mathbf{u})$.
We now describe the parameter test ideal of $A$. Henceforth we shall always denote $\H^{\dim(A)}_{\frak{m} A}(A)$ with $H$.
\[Theorem: non-$F$-injective\] Assume that $R$ is complete. The parameter test ideal of $A$ is given by $$\bigcap\big\{ I\in \mathcal{I}(R,\mathbf{u}) \,|\, \operatorname{ht}IA>0 \big\}.$$
Write $\overline{\tau}$ for the parameter test ideal of $A$ and let $\tau$ be its pre-image in $R$. Recall that $\overline{\tau}$ is an $F$-ideal (Proposition 4.5 in [@Smith2],) i.e., $\operatorname{ann}_H \overline{\tau}$ is an $A[T;f]$-submodule of $H$, and $\mathcal{H}_{R,A}\big( \operatorname{ann}_H \overline{\tau} \big)$ has generating morphism $$\left( \operatorname{ann}_H \overline{\tau} \right)^\vee \xrightarrow[]{u^{p-1}} F_R\left(\left( \operatorname{ann}_H \overline{\tau} \right)^\vee\right) .$$ But $$\left( \operatorname{ann}_H \overline{\tau} \right)^\vee \cong
\left( \left( A/\overline{\tau} \right)^\vee \right)^\vee \cong
R/(\tau+\mathbf{u}R)$$ so the generating morphism of $\mathcal{H}_{R,A}\big( \operatorname{ann}_H \overline{\tau} \big)$ is $$R/(\tau+\mathbf{u}R) \xrightarrow[]{u^{p-1}} R/(\tau^{[p]}+\mathbf{u}^p R)$$ and so we must have $\tau\in \mathcal{I}(R,\mathbf{u})$.
As $A$ is Cohen-Macaulay, $\overline{\tau}=\displaystyle (0 :_A 0^*_H) $ (cf. Proposition 4.4 in [@Smith2].)
By Theorem \[Theorem 1\](b), for each $I\in \mathcal{I}(R,\mathbf{u})$, the ideal $IA$ is an $F$-ideal and, if $\operatorname{ht}I>0$, $\displaystyle \operatorname{ann}_H IA= \operatorname{ann}_H I A[T;f] \subseteq 0^*_H $ and so $$\displaystyle \overline{\tau}=(0 :_A 0^*_H) \subseteq
\bigcap\big\{ (0 :_A \operatorname{ann}_H IA) \,|\, IA\in \mathcal{I}(R,\mathbf{u}), \operatorname{ht}IA>0 \big\} .$$ But $H$ is an injective hull of $A/\mathfrak{m}A$ so $$( 0 :_A \operatorname{ann}_H IA) = \left( 0 :_A \operatorname{Hom}(A/IA, H) \right)= (0 :_A A/IA)=IA$$ and $$\overline{\tau}\subseteq \bigcap\big\{IA \,|\, IA\in \mathcal{I}(R,\mathbf{u}), \operatorname{ht}IA>0 \big\} .$$ But as $\overline{\tau}$ is one of the ideals in this intersection, we obtain $\displaystyle \overline{\tau}=
\bigcap\big\{ IA\in \mathcal{I}(R,\mathbf{u}) \,|\, \operatorname{ht}IA>0 \big\}$.
The Gorenstein case {#Section: The Gorenstein case}
===================
In this section we generalise the results so far to the case where $A$ is Gorenstein.
Write $\delta=\dim R - \dim A$ and $\overline{E}=E_A(A/\mathfrak{m}A)$. Local duality implies $\displaystyle
\operatorname{Ext}^{\delta}_R(A, R)=
\H^{\dim A}_\mathfrak{m}(A)^\vee \cong
\operatorname{Hom}\left(\H^{\dim A}_{\mathfrak{m}A}(A),\overline{E} \right)$ and since $A$ is Gorenstein this is just $A=R/\mathbf{u}R$.
Now $\operatorname{Ext}^{\delta}_R \left( R/\mathbf{u}R, A \right)\cong R/\mathbf{u} R$, $\operatorname{Ext}^{\delta}_R \left( R/\mathbf{u}^pR, A \right)\cong R/\mathbf{u}^p R$ and $\displaystyle \mathcal{H}_{R,A}\big( \H^{\dim A}_{\mathfrak{m} A} \big)= \H^\delta_{\mathfrak{m}}(R)$ has generating morphism $R/\mathbf{u} \rightarrow R/\mathbf{u}^p R$ given by multiplication by some element of $R$ which we denote $\varepsilon(\mathbf{u})$ (this is unique up to multiplication by a unit.) Unlike the complete intersection case, the map $R/\mathbf{u} \xrightarrow[]{\varepsilon(\mathbf{u})} R/\mathbf{u}^p R$ may not be injective, i.e., this generating morphism of $\H^\delta_{\mathfrak{m}}(R)$ is not a *root*. However, if define $$\displaystyle K_\mathbf{u}:=
\bigcup_{e\geq 0} \big( \mathbf{u}^{p^{e+1}} R :_R \varepsilon(\mathbf{u})^{1+p+\dots+p^e} \big)$$ we obtain a root $R/K_\mathbf{u} \xrightarrow[]{\varepsilon(\mathbf{u})} R/K_\mathbf{u}^{[p]}$ (cf. Proposition 2.3 in [@Lyubeznik].)
We now extend naturally our definition of $\mathcal{I}(R, \mathbf{u})$ when $A$ is Gorenstein as follows.
If $A=R/\mathbf{u}R$ is Gorenstein we define $\mathcal{I}(R, \mathbf{u})$ to be the set of all ideals $I$ of $R$ containing $K_\mathbf{u}$ for which $\varepsilon(\mathbf{u}) I \subseteq I^{[p]}$.
Now a routine modification of the proofs of the previous sections gives the following two theorems.
\[Theorem: the Gorenstein case I\] Assume $A$ is Gorenstein and that $H^{\dim A}_{\mathfrak{m} A}(A)$ is $T$-torsion-free.
1. The map $I \mapsto IA$ is a bijection between $\mathcal{I}(R, \mathbf{u})$ and the $A$-special $H^{\dim A}_{\mathfrak{m} A}(A)$-ideals.
2. There exists a unique minimal element $\tau$ in $\left\{I \,|\, I\in \mathcal{I}(R,\mathbf{u}) , \ \operatorname{ht}I A >0 \right \}$ and that $\tau$ is a parameter-test-ideal for $A$.
3. $A$ is $F$-rational if and only if $\mathcal{I}(R,\mathbf{u})=\{ 0, R \}$.
\[Theorem: the Gorenstein case II\] Assume that $R$ is complete and that $A$ is Gorenstein. The parameter test ideal of $A$ is given by $$\bigcap\big\{ I\in \mathcal{I}(R,\mathbf{u}) \,|\, \operatorname{ht}IA>0 \big\}.$$
Acknowledgement {#acknowledgement .unnumbered}
===============
My interest in the study of local cohomology modules as modules over skew polynomial rings was aroused by many interesting conversations with Rodney Sharp, during one of which I learnt about Example \[Example 1\].
[HH2]{}
M. P. Brodmann and R. Y. Sharp. *Local cohomology: an algebraic introduction with geometric applications.* Cambridge Studies in Advanced Mathematics, [**60**]{}. Cambridge University Press, Cambridge, 1998.
R. Fedder. *$F$-purity and rational singularity.* Transactions of the AMS, [**278**]{} (1983), no. 2, pp. 461–480.
M. Hochster and C. Huneke. *Tight closure, invariant theory, and the Briançon-Skoda theorem.* Journal of the AMS [**3**]{} (1990), no. 1, 31–116.
M. Hochster and C. Huneke. *$F$-regularity, test elements, and smooth base change.* Transactions of the AMS, [**346**]{} (1994), no. 1, pp. 1–62.
E. Kunz. *Characterizations of regular local rings for characteristic $p$.* American Journal of Mathematics [**91**]{} (1969), pp. 772–784.
G. Lyubeznik. *$F$-modules: applications to local cohomology and $D$-modules in characteristic $p>0$.* J. Reine Angew. Math. [**491**]{} (1997), pp. 65–130.
R. Y. Sharp. *Graded annihilators of modules over the Frobenius skew polynomial ring, and tight closure.* Preprint.
K. E. Smith. *Test ideals in local rings.* Transactions of the AMS [**347**]{} (1995), no. 9, pp. 3453–3472.
K. E. Smith. *$F$-rational rings have rational singularities.* American Journal of Mathematics [**119**]{} (1997), no. 1, pp. 159–180.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We consider the problem of determining the energy distribution of quantum states that satisfy exponential decay of correlation and product states, with respect to a quantum local hamiltonian on a spin lattice. For a quantum state on a $D$-dimensional lattice that has correlation length $\sigma$ and has average energy $e$ with respect to a given local hamiltonian (with $n$ local terms, each of which has norm at most $1$), we show that the overlap of this state with eigenspace of energy $f$ is at most $exp(-((e-f)^2\sigma)^{\frac{1}{D+1}}/n^{\frac{1}{D+1}}D\sigma)$. This bound holds whenever $|e-f|>2^{D}\sqrt{n\sigma}$. Thus, on a one dimensional lattice, the tail of the energy distribution decays exponentially with the energy.
For product states, we improve above result to obtain a Gaussian decay in energy, even for quantum spin systems without an underlying lattice structure. Given a product state on a collection of spins which has average energy $e$ with respect to a local hamiltonian (with $n$ local terms and each local term overlapping with at most $m$ other local terms), we show that the overlap of this state with eigenspace of energy $f$ is at most $exp(-(e-f)^2/nm^2)$. This bound holds whenever $|e-f|>m\sqrt{n}$.
author:
-
bibliography:
- 'references.bib'
title: '**[Concentration bounds for quantum states with finite correlation length on quantum spin lattice systems]{}**\'
---
Introduction {#sec:intro}
============
A question of primary interest for local hamiltonian spin systems is to determine the energy distribution of natural class of states with respect to a given local hamiltonian. The knowledge of energy distribution reveals a lot of information about the nature of the state itself. As we shall discuss below, a *gaussian distribution* of energy can be associated to a product state. On the other hand, the well known entangled state $\frac{1}{\sqrt{2}}{\ensuremath{ \left| 0 \right\rangle }}^{\otimes n} + \frac{1}{\sqrt{2}}{\ensuremath{ \left| 1 \right\rangle }}^{\otimes n}$ (also termed as the ‘cat state’) has energy distribution peaking at opposite ends of the spectrum of the hamiltonian: $\sum_{i=1}^n {{\ensuremath{ \left| 1 \middle\rangle \middle\langle 1 \right| }}}_i$. Moreover, the knowledge of energy distribution plays an important role in the study of thermalization of quantum systems.
The aforementioned question has been well studied in classical setting, important examples of which are the Chernoff bound [@chernoff] and the Central limit theorem (which applies to asymptotic regime). Chernoff bound can be informally states as follows. Let $X_1,X_2\ldots X_n$ be independent and identically distributed random variables taking values in $[0,1]$ and each having average value $A$. Then $\text{Pr}(|X_1+X_2\ldots +X_n - nA| > {\varepsilon}) \leq e^{-c{\varepsilon}^2/n}$, where $c$ is a constant that depends on $A$.
One interpretation of this bound (which was the original motivation in [@chernoff]) is that it provides a recipe for distinguishing between two probability distributions $P{\ensuremath{ \stackrel{\mathrm{def}}{=} }}\sum_x p(x){{\ensuremath{ \left| x \middle\rangle \middle\langle x \right| }}}$ and $Q{\ensuremath{ \stackrel{\mathrm{def}}{=} }}\sum_x q(x){{\ensuremath{ \left| x \middle\rangle \middle\langle x \right| }}}$ with expectation values $A$ and $B$ respectively. Given $n$ independent samples $x_1,x_2\ldots x_n$ from either of these distributions, the sum $\sum_i x_i$ is highly likely to be concentrated around $nA$ if the underlying distribution is $P$ and around $nB$ if the underlying distribution is $Q$. A more precise formulation of this idea requires characterizing the trace distance between $P^{\otimes n}$ and $Q^{\otimes n}$ as $n$ becomes large, and it has been generalized to the quantum setting in [@audernaut].
Another interpretation of the Chernoff bound, which is the focus of present work, lies in the setting of ‘classical’ local Hamiltonian systems. Consider a product state $\rho^{\otimes n}$ on $n$ sites, where $\rho{\ensuremath{ \stackrel{\mathrm{def}}{=} }}\sum_x p(x){{\ensuremath{ \left| x \middle\rangle \middle\langle x \right| }}}$. Let $H$ be a $1$-local Hamiltonian $H{\ensuremath{ \stackrel{\mathrm{def}}{=} }}\sum_i h_i$, such that $h_i = \sum_x x{{\ensuremath{ \left| x \middle\rangle \middle\langle x \right| }}}$ acts non-trivially only on the site $i$ and is same for each site. If $A {\ensuremath{ \stackrel{\mathrm{def}}{=} }}{\ensuremath{ \mathrm{Tr} }}(\rho h_i)$ is the expectation value of $\rho$ with respect to $h_i$, then $nA$ is *average energy* of $\rho^{\otimes n}$ with respect to the hamiltonian $H$. Let $\Pi_{\geq nA+{\varepsilon}}$ be the projector onto eigenstates of $H$ with energy at least $nA+{\varepsilon}$. Then the Chernoff bound implies that ${\ensuremath{ \mathrm{Tr} }}(\rho^{\otimes n}\Pi_{\geq nA+{\varepsilon}}) \leq e^{-c{\varepsilon}^2/n}$. Thus, the energy distribution of $\rho^{\otimes n}$ is highly concentrated around the average energy $nA$.
The energy distribution of a product state for quantum lattice system with infinitely many sites was considered in [@goderis] (for translationally invariant systems) and in [@hartmann] (for non-translationally invariant systems). These results can be regarded as a generalization of the Central limit theorem to quantum systems. A non-asymptotic version of Central limit theorem is the Berry-Esseen theorem ([@berry],[@esseen]), which provides an upper bound on *trace distance* between energy distribution of product state and the *normal distribution* as a function of lattice size. This upper bound goes to zero as lattice size approaches infinity, thus recovering the Central limit theorem. For quantum states with finite correlation length (which includes product states) on finite sized lattice, a quantum version of Berry-Esseen theorem was recently shown to hold in [@brandao1],[@brandao2].
These results give a strong indication that states satisfying exponential decay of correlation behave similar to product states, even when their energy distributions are measured with respect to the eigenspectrum of a non-commuting (but local) hamiltonian. The work [@Keating2015] goes even further to show that non-commuting local hamiltonians themselves have energy spectrum that resemble that of a $1$-local hamiltonian (although, quite curiously, the same work shows that almost all eigenvectors of non-commuting local hamiltonians are highly entangled, in contrast with the eigenvectors of $1$-local hamiltonians).
Above mentioned results have added to the growing body of research on general properties of local hamiltonian systems, such as the Lieb-Robinson bound [@liebrobinson], exponential decay of correlation [@hastings04], the area laws [@Hastings2007; @ALV2012; @AKLV2013] and local reversibility [@aradkuwahara], to name a few. They have also found several applications in the problem of thermalization of many body systems. To start with, one of the first steps towards the problem of *locality of temperature* [^1] was taken in [@Hartmann04]. Crucially using the Central limit theorem obtained in [@hartmann], the authors characterized a set of conditions under which a given thermal state of a quantum local hamiltonian on a lattice would be close to a tensor product of thermal states on local subsystems on the lattice.
The work [@Cramer12] considered the problem of thermalization under random hamiltonians, where the hamiltonian was generated via a random unitary on a fixed local hamiltonian $H$. One of the main technical problems in this work was the study of the characteristic function ${\ensuremath{ \mathrm{Tr} }}(e^{iHt}\frac{I}{d})$, where $\frac{I}{d}$ is the maximally mixed state (which is also a product state on the lattice). The techniques were inspired from the proof of Central limit theorem in the works [@goderis],[@hartmann], where the characteristic function ${\ensuremath{ \mathrm{Tr} }}(e^{iHt}\rho)$ of a product state $\rho$ had been investigated in detail.
The quantum version of Berry-Esseen theorem [@brandao2] was used to show in [@brandao1] that Gibbs state of a local hamiltonian $H$ at sufficiently high temperature (high enough to ensure a clustering of correlation) is indistinguishable, over sufficiently large regions of lattice (that scale sub-linearly with lattice size), from the microcanonical ensemble of eigenstates of $H$ which have eigenvalues close to the average energy of the Gibbs state. This result bears close resemblance to the Eigenstate Thermalization Hypothesis [@Srednicki94; @Deutsch91], which is a stronger conjecture stating that every eigenstate of $H$ with eigenvalue close to the average energy of Gibbs state of $H$ is locally indistinguishable from this Gibbs state.
In present work, we provide further details on energy distribution of states satisfying exponential decay of correlation and product states. Our main results can be seen as an analogue of the Chernoff bound for quantum lattice systems.
Our first result concerns states that satisfy exponential decay of correlation on a $D$-dimensional lattice. Well known examples of such states include the ground states of gapped local hamiltonians [@hastings04] and Gibbs state above a finite temperature [@kliesch13]. In fact, it has been shown in [@Friesdorf15] that for local hamiltonians exhibiting many body localization and having non-degenerate energy spectrum, all eigenvectors satisfy exponential decay of correlation. Thus our result provides information about structure of eigenvectors of such hamiltonians and may have applications in the phenomena of many body localization.
Fix a $D$-dimensional lattice with spins of arbitrary local dimension sitting on each lattice site. Consider local hamiltonian $H$ on the lattice with $n$ local interaction terms, such that each local term has operator norm at most $1$ and its support is a hyper-cube of side length $2k$, hence containing $(2k)^D$ lattice sites (see Section \[sec:preliminaries\] and Figure \[fig:quantumlattice\] for a detailed description of $H$).
\[theo:expodecay\] Let $\rho$ be a quantum state with correlation length $\sigma$ and $\langle H\rangle_{\rho} {\ensuremath{ \stackrel{\mathrm{def}}{=} }}{\ensuremath{ \mathrm{Tr} }}(\rho H)$ be the average energy of $\rho$. Let $\Pi_{\geq f}$ ($\Pi_{\leq f}$) be the projection onto subspace which is union of eigenspaces of $H$ with eigenvalues $\geq f$ ($\leq f$).
For $a\geq \sqrt{\frac{2^{\mathcal{O}(D\log k)}}{n\sigma}}$ it holds that, $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + na}) \leq \mathcal{O}(\sigma) e^{\frac{2Dk}{\sigma}}\cdot e^{-\frac{(na^2\sigma)^{\frac{1}{D+1}}}{\mathcal{O}(1)D\sigma}} \text{ and } {\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\leq \langle H\rangle_{\rho} - na}) \leq \mathcal{O}(\sigma) e^{\frac{2Dk}{\sigma}}\cdot e^{-\frac{(na^2\sigma)^{\frac{1}{D+1}}}{\mathcal{O}(1)D\sigma}}.$$
Formal statement of the theorem is given in Theorem \[formaltheo:expodecay\]. Thus in one dimensional spin chain (with $D=1$), our upper bound decays exponentially with energy, rather than as a gaussian. The bound becomes weaker with higher dimensions and is depicted in Figure \[fig:distribution\].
Our second result concerns product states over a collection of spins and does not require any underlying lattice arrangement of these spins. It does impose, however, a locality constraint on the hamiltonian that acts on these spins. Consider a hamiltonian $H$ which is a sum of $n$ terms, each term being $k$-local (that is, it acts non-trivially on at most $k$ spins) and having operator norm at most $1$. Let $m$ be the maximum number of neighbours of any local term, where two local terms are neighbours if there is a spin on which both act non-trivially (See Section \[sec:momentbound\] and Figure \[fig:spincollection\] for detailed description of $H$). We show the following.
\[theo:qchernoff\] Consider a product state $\rho$ with average energy $\langle H\rangle_{\rho} {\ensuremath{ \stackrel{\mathrm{def}}{=} }}{\ensuremath{ \mathrm{Tr} }}(\rho H)$. Fix a real number $a\geq \sqrt{\frac{\mathcal{O}(m^2)}{n}}$. Let $\Pi_{\geq f}$ ($\Pi_{\leq f}$) be the projection onto subspace which is union of eigenspaces of $H$ with eigenvalues $\geq f$ ($\leq f$). It holds that $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle + na}) \leq e^{-\frac{na^2}{\mathcal{O}(m^2)}}$$ and $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\leq \langle H\rangle - na}) \leq e^{-\frac{na^2}{\mathcal{O}(m^2)}}.$$
Formal statement of the theorem is given in Theorem \[formaltheo:qchernoff\]. The energy distribution is depicted in Figure \[fig:distribution\]. The bound is not only independent of any underlying lattice structure, but is also independent of the locality $k$. This is not surprising, since the quantity $n$ that appears in the bound is the number of local terms in $H$, rather than number of spins on which $H$ acts. Following corollary is a restatement of above bound, in terms of number of spins (which we call $N$) and the maximum number of local terms that act on any given spin (which we call $g$). In the following, we also assume that each local term is exactly $k$-local.
\[cor:qchernoff\] Consider a product state $\rho$ with average energy $\langle H\rangle_{\rho} {\ensuremath{ \stackrel{\mathrm{def}}{=} }}{\ensuremath{ \mathrm{Tr} }}(\rho H)$. Fix a real number ${\varepsilon}\geq \sqrt{\mathcal{O}(g^3kN)}$. Let $\Pi_{\geq f}$ ($\Pi_{\leq f}$) be the projection onto subspace which is union of eigenspaces of $H$ with eigenvalues $\geq f$ ($\leq f$). It holds that $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle + {\varepsilon}}) \leq e^{-\frac{{\varepsilon}^2}{\mathcal{O}(g^3kN)}}$$ and $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\leq \langle H\rangle - na}) \leq e^{-\frac{{\varepsilon}^2}{\mathcal{O}(g^3kN)}}.$$
Formal statement of the corollary appears as Corollary \[formalcor:qchernoff\]. It shows a gaussian decay for tail of energy distribution of product states in the scenario where $g,k$ are constants [^2] independent of $N$ .
(0,4) – (0,0) – (5,0); (0.5,0.5) to \[out=5, in=210\] (2.4,3); (3,3) to \[out=330, in=175\] (4.9,0.5); (0.5,0.5) to \[out=5,in=240\] (1.4,1.2) – (1.4,0.5) – (0.5,0.5); (4,1.2) to \[out=300,in=175\] (4.9,0.5) – (4,0.5) – (4,1.2); (1.45,1) – (3.95,1); (2.4,3.1) – (3,3.1); (2.75,0.1) – (2.75,-0.1); at (2.8,-0.3) [$\langle H\rangle_{\rho}$]{}; at (2.8,1.3) [$\sqrt{n\log(n)}$]{}; at (2.7,3.4) [$\sqrt{n}$]{}; at (4.8,-0.3) [$e$]{}; at (-0.6,4) [${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_e)$]{};
(7,4) – (7,0) – (13,0); (7.2,0.8) to \[out=5, in=250\] (9.7,3); (10.3,3) to \[out=290, in=175\] (12.8,0.8); (7.2,0.79) to \[out=13,in=220\] (8.5,1.25) – (8.5,0.5) – (7.2,0.5) – (7.2,0.75); (11.5,1.25) to \[out=320,in=167\] (12.8,0.8) – (12.8,0.5) – (11.5,0.5) – (11.5,1.25); (8.55,1) – (11.45,1); (9.7,3.1) – (10.3,3.1); (10,0.1) – (10,-0.1); at (10,-0.3) [$\langle H\rangle_{\rho}$]{}; at (9.9,1.3) [$\sqrt{n(\sigma\log(n))^{D+1}}$]{}; at (10,3.4) [$\sqrt{\frac{n}{\sigma}}$]{}; at (13,-0.3) [$e$]{}; at (6.4,4) [${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_e)$]{};
Related recent works {#related-recent-works .unnumbered}
--------------------
A recent work [@kuwahara] has obtained a similar concentration result for product states (Lemma $4$ therein). The key idea is to split the hamiltonian $H$ as $H=H_1+H_2+\ldots$, where each $H_1,H_2\ldots $ is composed of local terms that are non-overlapping. Then from classical Chernoff bound, the product state exhibits a Gaussian decay in energy distribution for each of the hamiltonians $H_1,H_2\ldots$. Final step (which is also the main argument of the paper) is to combine these tails bounds to obtain a final bound for energy distribution with respect to the original hamiltonian $H$. Unfortunately, the techniques do not extend to states satisfying exponential decay of correlation. To establish a bound for energy distribution with respect to $H$, one needs the knowledge of bounds for energy distribution with respect to each of the ‘classical hamiltonians’ $H_1,H_2\ldots$. But even for these classical hamiltonians, no bound is known for states that satisfy exponential decay of correlation (apart from Theorem \[theo:expodecay\], to the best of author’s knowledge). We have provided further comparision of the bound in [@kuwahara] and Theorem \[theo:qchernoff\] in Subsection \[subsec:qchernoff\].
A concentration result has been noted in [@aradkuwahara] (Section $5$ in this reference) for ground states of gapped local hamiltonians on finite dimensional quantum lattice systems, which also exhibit exponential decay of correlation ([@hastings04]). In this work, the probability distribution has been shown to be concentrated about the *median* of the distribution with the weight of the distribution above energy ${\varepsilon}$ decaying as $e^{-|{\varepsilon}-f|/\mathcal{O}(1)\sqrt{n\sigma}}$ ($f$ being the median of the distribution, $n$ being the number of local terms in the local hamiltonian and $\sigma$ being the correlation length of the ground state). In comparison, we show a concentration about the *mean* of the distribution for all states satisfying exponential decay of correlation. While our bounds are weaker than those of [@aradkuwahara] in higher dimensions, it may be noted that we have considered a larger class of states that might possess weaker properties than the ground states of gapped local hamiltonians. This behaviour appears in the context of area laws as well: ground states of gapped local hamiltonians are known to have very good scaling of area laws with correlation length [@AKLV2013]; whereas a recent observation of Hastings [@Hastings2015] suggests that states satisfying exponential decay of correlation may have much weaker dependence of area law with correlation length [@brandaohorodecki].
Our technique and organisation {#our-technique-and-organisation .unnumbered}
------------------------------
The idea behind our approach is straightforward, to compute the moment generating function ${\ensuremath{ \mathrm{Tr} }}(H^r\rho)$ of the energy distribution and then use Markov’s inequality to upper bound the desired probability. Without loss of generality, we can assume that $H = \sum_w h_w$, where $w$ is a label for local terms and $\langle h_w\rangle_{\rho}{\ensuremath{ \stackrel{\mathrm{def}}{=} }}{\ensuremath{ \mathrm{Tr} }}(\rho h_w)=0$. Our key technical contribution is the combinatorial lemma (Lemma \[lem:combinatorial\]) which answers the following question: if we expand $H^r$ as a sum of product of local terms, that is $H^r = \sum_{w_1,w_2\ldots w_r}h_{w_1}h_{w_2}\ldots h_{w_r}$, how many terms $h_{w_1}h_{w_2}\ldots h_{w_r}$ make non-negligible (or non-zero) contribution to the moment generating function? We observe the terms that make non-negligible contribution possess a common property: there is no $h_{w_i}$ which is supported ‘far’ from all of $h_{w_1},h_{w_2},\ldots h_{w_{i-1}}, h_{w_{i+1}},\ldots h_{w_r}$. Making the notion of ‘far’ precise, we compute the number of such terms in Lemma \[lem:combinatorial\] and use it to bound the moment generating function.
The paper is organized as follows. We state basic facts and describe our physical set-up needed for Theorem \[theo:expodecay\] in section \[sec:preliminaries\]. We prove our combinatorial lemma in Section \[sec:combinatorial\]. In Section \[sec:expodecay\] we prove our bounds for states satisfying exponential decay of correlation. In Section \[sec:momentbound\], we introduce the physical set-up required for Theorem \[theo:qchernoff\] and provide the proof of the theorem. This proof also requires a variant of the combinatorial lemma (Lemma \[lem:combinatorial\]) which we prove in Appendix \[append:combinatorial\]. We conclude in Section \[sec:conclusion\] and address some questions left open by this work.
Physical set-up and basic facts {#sec:preliminaries}
===============================
In this section, we introduce the physical-set up required for Theorem \[theo:expodecay\]. For simplicity of the presentation, we shall assume that the spins are arranged on a square lattice, with a local interaction term acting between only those spins that are the vertices of a common ‘unit-hypercube’. We shall introduce the notion of a ‘dual lattice’ below, to formally and concisely represent these local interactions between the spins. It can be observed that more general local interactions on a square lattice can be put in this form by sufficient coarse-graining of lattice sites. The physical set-up for Theorem \[theo:qchernoff\] is relatively simple, and shall be introduced directly in Section \[sec:momentbound\].
Consider a $D$-dimensional real vector space $\mathbb{R}^D$. For a vector $v\in \mathbb{R}^D$, let $v_i$ represent its $i$-th component. For two vectors $v,v'\in \mathbb{R}^D$, define the ‘$1$-norm distance’ as $$\|v-v'\|{\ensuremath{ \stackrel{\mathrm{def}}{=} }}\sum_i |v_i-v'_i|.$$ It satisfies the triangle inequality: given $v,v',v" \in \mathbb{R}^D$, we have $$\|v-v"\|\leq \|v-v'\|+\|v'-v"\|.$$ For brevity, we shall refer to $1$-norm distance simply as *distance*.
For an integer $L>0$, define a lattice $\L_{D,L}$ as the set of all vectors $v\in \mathbb{R}^D$ that satisfy the following: for all $i\in \{1,2\ldots D\}$, it holds that $v_i$ is an integer and $0\leq v_i\leq L$. Two vectors $v,v'\in \L_{D,L}$ are *neighbours* if $\|v-v'\|=1$. Henceforth, the vectors belonging to $\L_{D,L}$ shall be referred to as *sites*.
For each site $v\in \L_{D,L}$, we associate a $d$-dimensional Hilbert space $\H_v^d$ and define the full Hilbert space as $\H = \otimes_{v\in \L_{D,L}}\H_v^d$. Local hamiltonian system is conveniently represented using the notion of *dual lattice*. Let $\bar{\L}_{D,L}$ be the set of vectors $w$ such that for all $i\in \{1,2\ldots D\}$, $0<w_i < L$ and $w_i$ is a half integer (that is, $w_i= k+\frac{1}{2}$, for $k$ an integer). For a fixed $w\in \bar{\L}_{D,L}$ and an integer $k$, let $\S(w,k)$ be the set of all sites $v\in \L_{D,L}$ such that: for all $i\in \{1,2\ldots D\}$, $|v_i - w_i| \leq k+\frac{1}{2}$
A local hamiltonian on $\L_{D,L}$ is defined as $H=\sum_{w\in \bar{\L}_{D,L}} h_w$, where $h_w$ is a ‘$(2k+2)^D$-local’ term that acts non-trivially only on sites in $\S(w,k)$ and acts as identity on rest of the sites. The number of sites in $\S(w,k)$ is at most $(2k+2)^D$, justifying $h_w$ as a ‘$(2k+2)^D$-local’ interaction. Following the physical motivation, we shall refer to vectors in $\bar{\L}_{D,L}$ as *interactions*. Figure \[fig:quantumlattice\] illustrates the notions introduced above for the case when $D=2$.
(0,0) grid (4,4); (0,0) circle \[radius=0.05\]; (0,1) circle \[radius=0.05\]; (0,2) circle \[radius=0.05\]; (0,3) circle \[radius=0.05\]; (0,4) circle \[radius=0.05\]; (1,0) circle \[radius=0.05\]; (1,1) circle \[radius=0.05\]; (1,2) circle \[radius=0.05\]; (1,3) circle \[radius=0.05\]; (1,4) circle \[radius=0.05\]; (2,0) circle \[radius=0.05\]; (2,1) circle \[radius=0.05\]; (2,2) circle \[radius=0.05\]; (2,3) circle \[radius=0.05\]; (2,4) circle \[radius=0.05\]; (3,0) circle \[radius=0.05\]; (3,1) circle \[radius=0.05\]; (3,2) circle \[radius=0.05\]; (3,3) circle \[radius=0.05\]; (3,4) circle \[radius=0.05\]; (4,0) circle \[radius=0.05\]; (4,1) circle \[radius=0.05\]; (4,2) circle \[radius=0.05\]; (4,3) circle \[radius=0.05\]; (4,4) circle \[radius=0.05\];
(0.5,0.5) circle \[radius=0.05\]; (0.5,1.5) circle \[radius=0.05\]; (0.5,2.5) circle \[radius=0.05\]; (0.5,3.5) circle \[radius=0.05\]; (1.5,0.5) circle \[radius=0.05\]; (1.5,1.5) circle \[radius=0.05\]; (1.5,2.5) circle \[radius=0.05\]; (1.5,3.5) circle \[radius=0.05\]; (2.5,0.5) circle \[radius=0.05\]; (2.5,1.5) circle \[radius=0.05\]; (2.5,2.5) circle \[radius=0.05\]; (2.5,3.5) circle \[radius=0.05\]; (3.5,0.5) circle \[radius=0.05\]; (3.5,1.5) circle \[radius=0.05\]; (3.5,2.5) circle \[radius=0.05\]; (3.5,3.5) circle \[radius=0.05\];
at (2.7,2.5) [$w$]{}; at (0.7,1.5) [$w'$]{}; at (0.2,2.1) [$v$]{}; at (2.2,2.1) [$v_1$]{}; at (2.8,2.1) [$v_2$]{}; at (2.2,2.85) [$v_3$]{}; at (2.8,2.85) [$v_4$]{};
Without loss of generality, we assume that the local terms $h_w$ are positive semi-definite matrices and $\|h_w\|_{\infty}\leq 1$, where $\|.\|_{\infty}$ is the operator norm. Given an operator $A$, *support* of $A$ (called $\text{supp}(A)$) is the set of sites in $\L_{D,L}$ on which $A$ acts non-trivially. We define the *distance* between two operators $A, B$ to be the minimum distance between their respective supports, that is, $\text{min}_{v\in \text{supp}(A), v'\in \text{supp}(B)}\|v-v'\|$.
For a quantum state $\rho \in \H$, the *reduced density matrix* of $\rho$ on a set $T$ of sites is represented as $\rho_T$. We define the *average energy* of $\rho$ to be ${\ensuremath{ \mathrm{Tr} }}(\rho H)$, and represent it as $\langle H\rangle_{\rho}$. For every local term $h_w$, let $\langle h_w\rangle_{\rho} {\ensuremath{ \stackrel{\mathrm{def}}{=} }}{\ensuremath{ \mathrm{Tr} }}(\rho_{\S(w)}h_w)$. Then we have $$\langle H\rangle_{\rho} = \sum_w \langle h_{w}\rangle_{\rho}.$$
A state $\rho \in \H$ satisfies $(C,l_0,\sigma)-$ decay of correlation if for any two operators $A,B$ such that distance between $A,B$ is $l\geq l_0$, it holds that $$|{\ensuremath{ \mathrm{Tr} }}(\rho A\otimes B)-{\ensuremath{ \mathrm{Tr} }}(\rho A){\ensuremath{ \mathrm{Tr} }}(\rho B)|\leq C\|A\|\|B\|e^{-\frac{l}{\sigma}}.$$
Define $\Pi_f$ to be the projector onto eigenspace of $H$ with eigenvalue (energy) equal to $f$. Let $\Pi_{\geq f}$ ($\Pi_{\leq f}$) be the projection onto the subspace which is union of eigenspaces of $H$ with eigenvalues greater (less) than $f$. The following fact follows from Markov’s inequality.
\[fact:markov\] For every $t,a>0$ and $r$ even, $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + a}) \leq \frac{{\ensuremath{ \mathrm{Tr} }}(\rho(H-\langle H\rangle_{\rho})^r)}{(a)^r}.$$
We have $(H-\langle H\rangle_{\rho})^r = \sum_f (f-\langle H\rangle_{\rho})^r\Pi_f$, which gives ${\ensuremath{ \mathrm{Tr} }}(\rho(H-\langle H\rangle_{\rho})^r) = \sum_f (f-\langle H\rangle_{\rho})^r{\ensuremath{ \mathrm{Tr} }}(\rho\Pi_f)$. This implies, $${\ensuremath{ \mathrm{Tr} }}(\rho(H-\langle H\rangle_{\rho})^r) \geq (a)^r\sum_{f>\langle H\rangle_{\rho} + a}{\ensuremath{ \mathrm{Tr} }}(\rho\Pi_f).$$
A combinatorial lemma {#sec:combinatorial}
=====================
In this section, we shall prove a combinatorial lemma, which we shall use in Section \[sec:expodecay\] to prove Theorem \[theo:expodecay\]. A slight variant of this lemma shall be proved in Appendix \[append:combinatorial\] and used in Section \[sec:momentbound\] to prove Theorem \[theo:qchernoff\]. We recall the definition of the set $\bar{\L}_{D,L}$ from Section \[sec:preliminaries\] and let the number of interactions in $\bar{\L}_{D,L}$ be $n$. It is easily seen that $n=(L-1)^D$.
Fix an integer $l$. An ordered set $\{w_1,w_2\ldots w_r\}$ of $r$ interactions in $\bar{\L}_{D,L}$ is said to satisfy a property $\P(l)$ if the following holds: for all $w_i$, there exists a $w_j$ such that $\|w_i-w_j\|\leq l$. Let the number of such ordered sets be $N_D(n,r,l)$.
Rest of the section is devoted to the proof of following lemma.
\[lem:combinatorial\] It holds that $N_D(n,r,l)\leq (4(4l)^Dnr)^{\frac{r}{2}}$.
We start with the following definition that we shall extensively use.
A *selection* is an ordered set $\{(b_1,x_1),(b_2,x_2)\ldots (b_r,x_r)\}$, where $b_i\in \{0,1\}$ and $x_i\in \bar{\L}_{D,L}$, that satisfies the following constraints:
1. If $b_i=0$, then $x_i$ can be any interaction in $\bar{\L}_{D,L}$ and if $b_i=1$, $x_i$ has to satisfy $\|x_i-x_j\|\leq 2\cdot l$ for some $j<i$.
2. Number of $i$ for which $b_i=0$ is at most $\frac{r}{2}$.
We show the following lemma from which the proof of Lemma \[lem:combinatorial\] shall follow immediately.
\[lem:selection\] Every ordered set $\{w_1,w_2\ldots w_r\}$ that satisfies property $\P(l)$ can be mapped to a *selection* in such a way that for any two distinct sets satisfying $\P(l)$, the corresponding *selections* are distinct.
We assign a *selection* to an ordered set $\{v_1,v_2\ldots v_r\}$ satisfying $\P(l)$ using the algorithm below.
We show that above algorithm terminates and assigns a *selection* to each ordered set satisfying property $\P(l)$.
1. Consider the running of algorithm during the step **Initialization**. Condition $1$ of a *selection* holds: for every $i$ for which there is a $j<i$ such that $\|x_i-x_j\|\leq l$, we have set $b_i=1$. But we haven’t constructed a *selection* yet, since condition $2$ may not be satisfied.
2. After the step **Pointer creation**, it may be possible that there exist indices $i_1,i_2\ldots i_s$ (for some $s< r$) such that $b_{i_1}=b_{i_2}=\ldots b_{i_s}=0$, $R(i_1)=R(i_2)=\ldots R(i_s) > 0$ and $i_s> i_{s-1}> \ldots i_1$. In this case, we find using triangle inequality that $\|w_{i_2}-w_{i_1}\|\leq \|w_{i_2}-w_{R(i_2)}\|+\|w_{R(i_2)}-w_{i_1}\|=\|w_{i_2}-w_{R(i_2)}\|+\|w_{R(i_1)}-w_{i_1}\|\leq 2l$. Similarly, $\|w_{i_3}-w_{i_1}\|\leq 2l, |w_{i_4}-w_{i_1}|\leq 2l, \ldots |w_{i_s}-w_{i_1}|\leq 2l$.
Thus, the step **Update** sets $b_{i_2}=b_{i_3}=\ldots b_{i_s}=1$, recognizing the fact that each of the points $w_{i_2},w_{i_3}\ldots w_{i_s}$ are at a lattice distance of at most $2l$ from $w_{i_1}$. This ensures that condition $1$ of *selection* is still satisfied.
3. After the step **Update** terminates, condition $2$ of *selection* is now satisfied as well. We now have that for every $i$ with $b_i=0$, there is no other $i'$ such that $R(i)=R(i')$ and $b_i=b_{i'}=0$. Furthermore, $b_{R(i)}=1$. Thus, number of $i$ with $b_i=0$ is at most as large as the number of $j$ with $b_j=1$.
Lemma follows as two distinct ordered sets satisfying $\P(l)$ are not assigned the same *selection*.
Now we prove Lemma \[lem:combinatorial\].
For $n\leq r(4l)^D$, we clearly have $N_D(n,r,l)\leq n^r \leq ((4l)^Dnr)^{\frac{r}{2}} < (4(4l)^Dnr)^{\frac{r}{2}}$.
So we assume $n> r(4l)^D$. We bound the number of *selections*, which gives the desired upper bound on $N_D(n,r,l)$ using Lemma \[lem:selection\].
Consider those *selections* for which number of $i$ such that $b_i=0$ is $u$. For each $i$ with $b_i=0$, number of possible choices of $x_i$ is $n$. For each $i$ with $b_i=1$, number of possible choices of $x_i$ is at most $(4l)^Dr$ (as there are at most $(4l)^D$ points $x_j\in \L_{D,L}$ that satisfy $\|x_i-x_j\|\leq 2l$ for a given $x_i$ [^3]). Hence total number of such *selections* is at most ${r \choose u}n^{u}((4l)^Dr)^{r-u}$. Since $u\leq \frac{r}{2}$, total number of *selections* is at most $$\sum_{u=0}^{\frac{r}{2}}{r \choose u}n^{u}((4l)^Dr)^{r-u} \leq \sum_{u=0}^{\frac{r}{2}}{r \choose u}n^{\frac{r}{2}}((4l)^Dr)^{\frac{r}{2}}<2^r((4l)^Dnr)^{\frac{r}{2}}.$$
This proves the lemma.
Energy distribution of states that satisfy an exponential decay of correlation {#sec:expodecay}
==============================================================================
Consider a state $\rho$ that satisfies $(C,l_0,\sigma)-$ decay of correlation and the hamiltonian $H=\sum_{w\in \bar{\L}_{D,L}} h_w$, where each term $h_w$ is $(2k+2)^D$-local, that is, it acts non-trivially only on sites in $\S(w,k)$. Let $g_w{\ensuremath{ \stackrel{\mathrm{def}}{=} }}h_w - {\ensuremath{ \mathrm{Tr} }}(\rho h_w)\text{I}$. We prove following bound on $r$-th moment.
\[expodecaymoment\] Given the state $\rho$ that satisfies $(C,l_0,\sigma)-$ decay of correlation, it holds that $${\ensuremath{ \mathrm{Tr} }}(\rho(H-\langle H\rangle)^r)\leq (4(4l_0+8Dk)^Dnr)^{\frac{r}{2}}+Ce^{\frac{2Dk}{\sigma}}\sigma\cdot(4(\frac{D\sigma}{2})^Dnr^{D+1})^{\frac{r}{2}}.$$
Consider,
$$\begin{aligned}
\label{eq:seriesexpansion}
{\ensuremath{ \mathrm{Tr} }}(\rho (\sum_{w\in \bar{\L}_{D,L}} g_w)^r) &=& \sum_{w_1,w_2\ldots w_r} {\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r})\end{aligned}$$
For every ordered set $\{w_1,w_2\ldots w_r\}$ define the quantity $D(w_1,w_2\ldots w_r){\ensuremath{ \stackrel{\mathrm{def}}{=} }}\text{max}_i (\text{min}_{j\neq i}|w_i-w_j|)$. This is the distance of farthest interaction from rest of the interactions in the ordered set.
For an integer $l>0$, define $\T(l)$ as the collection[^4] of all sets $\{w_1,w_2\ldots w_r\}$ that satisfy $D(w_1,w_2\ldots w_r)= l$. Now, fix a set $\{w_1,w_2\ldots w_r\}\in \T(l)$. Without loss of generality, suppose that $w_1$ is an interaction at the distance $l$ from rest of the interactions. The distance between operator $g_{w_1}$ and $g_{w_i}$, for any $i\neq 1$, is at least $l- 2Dk$, as the distance from $w_i$ to any site in $\S(w_i,k)$ is at most $Dk$. Then from $(C,\sigma,l_0)-$ decay of correlation and the relation ${\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1})=0$, it holds that $${\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r})\leq {\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1})\cdot{\ensuremath{ \mathrm{Tr} }}(\rho g_{w_2}\ldots g_{w_r})+ Ce^{-\frac{l-2Dk}{\sigma}} = Ce^{-\frac{l-2Dk}{\sigma}},$$ as long as $l-2Dk \geq l_0$.
Now, the number of sets in the collection $\T(l)$ is at most $N_D(n,r,l)$ which is upper bounded by $(4(4l)^Dnr)^{\frac{r}{2}}$ (Lemma \[lem:combinatorial\]). Thus we have
$$\begin{aligned}
\label{eq:expodecaycalc}
\sum_{w_1,w_2\ldots w_r} {\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r}) &=& \sum_l\sum_{(w_1,w_2\ldots w_r)\in \T(l)} {\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r}) \nonumber\\&=& \sum_{l\leq l_0+2Dk}\sum_{(w_1,w_2\ldots w_r)\in \T(l)} {\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r}) \nonumber\\&+& \sum_{l> l_0+2Dk}\sum_{(w_1,w_2\ldots w_r)\in \T(l)} {\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r}) \nonumber\\ &\leq& N_D(n,r,l_0+2Dk) + \sum_{l>l_0+2Dk} N_D(n,r,l)\cdot Ce^{-\frac{l-2Dk}{\sigma}}\nonumber\\ &\leq& (4(4l_0+8Dk)^Dnr)^{\frac{r}{2}}+Ce^{\frac{2Dk}{\sigma}}\sum_{ l\geq l_0+2Dk} (4(4l)^Dnr)^{\frac{r}{2}}e^{-\frac{l}{\sigma}}\nonumber\\&\leq& (4(4l_0+8Dk)^Dnr)^{\frac{r}{2}}+Ce^{\frac{2Dk}{\sigma}}(4nr)^{\frac{r}{2}}\sum_{l\geq 1} l^{\frac{rD}{2}}e^{-\frac{l}{\sigma}}\end{aligned}$$
Now, we evaluate $$\begin{aligned}
\sum_{l\geq 1} l^{\frac{rD}{2}}e^{-\frac{l}{\sigma}} &\leq& \int_{0}^{\infty} l^{\frac{rD}{2}}e^{-\frac{l}{\sigma}}dl = \sigma^{\frac{rD}{2}+1}\int_{0}^{\infty}s^{\frac{rD}{2}}e^{-s}ds \leq \sigma^{\frac{rD}{2}+1}(\frac{rD}{2})^{\frac{rD}{2}}. \end{aligned}$$
Using this in Equation (\[eq:expodecaycalc\]), we obtain $$\sum_{w_1,w_2\ldots w_r} {\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r})\leq (4(4l_0+8Dk)^Dnr)^{\frac{r}{2}}+Ce^{\frac{2Dk}{\sigma}}\sigma(4(\frac{D\sigma}{2})^Dnr^{D+1})^{\frac{r}{2}}.$$
Now we proceed to state Theorem \[theo:expodecay\] formally and provide its proof.
\[formaltheo:expodecay\] Consider a quantum state $\rho$ that satisfies $(C,l_0,\sigma)-$decay of correlation and has average energy $\langle H\rangle_{\rho}$.
For $\sqrt{\frac{8e(4l_0+8Dk)^{D+1}}{nD\sigma}}\geq a \geq \sqrt{\frac{8e(4l_0+8Dk)^D}{n}}$ (if the range exists) it holds that, $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + na}) \leq (1+C\sigma e^{\frac{2Dk}{\sigma}})e^{-\frac{na^2}{8e(4l_0+8Dk)^D}} \text{ and } {\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\leq \langle H\rangle_{\rho} - na}) \leq (1+C\sigma e^{\frac{2Dk}{\sigma}})e^{-\frac{na^2}{8e(4l_0+8Dk)^D}}.$$ For $a\geq \sqrt{\frac{(4l_0+8Dk)^{D+1}}{D\sigma n}}$ it holds that, $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + na}) \leq (1+C\sigma e^{\frac{2Dk}{\sigma}}) e^{-(\frac{na^2}{8e(D\sigma)^D})^{\frac{1}{D+1}}} \text{ and } {\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\leq \langle H\rangle_{\rho} - na}) \leq (1+C\sigma e^{\frac{2Dk}{\sigma}}) e^{-(\frac{na^2}{8e(D\sigma)^D})^{\frac{1}{D+1}}}.$$
Using Fact \[fact:markov\] and Lemma \[expodecaymoment\] we have, $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + na}) \leq (\frac{4(4l_0+8Dk)^Dr}{na^2})^{\frac{r}{2}}+Ce^{\frac{2Dk}{\sigma}}\sigma(\frac{4(\frac{D\sigma}{2})^D r^{D+1}}{na^2})^{\frac{r}{2}}$$
Consider the following two cases.
- $a\geq \sqrt{\frac{(4l_0+8Dk)^{D+1}}{D\sigma n}}$, or equivalently $\frac{(4l_0+8Dk)^{D+1}}{Dna^2\sigma}<1$
Then we set $r=2\lceil (\frac{na^2}{8e(D\sigma)^D})^{\frac{1}{D+1}} \rceil$, where $\lceil . \rceil$ denotes the ceiling operation (rounding to the nearest larger integer) to obtain
$$\begin{aligned}
{\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + na}) &\leq& (\frac{4(4l_0+8Dk)^Dr}{na^2})^{\frac{r}{2}}+Ce^{\frac{2Dk}{\sigma}}\sigma(\frac{4(\frac{D\sigma}{2})^D r^{D+1}}{na^2})^{\frac{r}{2}} \\ &\leq& (\frac{1}{e}(\frac{(4l_0+8Dk)^{D+1}}{Dna^2\sigma})^{\frac{D}{D+1}})^{\frac{r}{2}}+Ce^{\frac{2Dk}{\sigma}}\sigma(\frac{1}{e})^{\frac{r}{2}} \\ &\leq& (1+C\sigma\cdot e^{\frac{2Dk}{\sigma}}) e^{-(\frac{na^2}{8e(D\sigma)^D})^{\frac{1}{D+1}}}\end{aligned}$$
The last inequality follows from the assumption: $\frac{(4l_0+8Dk)^{D+1}}{Dna^2\sigma}<1$.
- $a\leq\sqrt{\frac{8e(4l_0+8Dk)^{D+1}}{nD\sigma}}$ and $a \geq \sqrt{\frac{8e(4l_0+8Dk)^D}{n}}$, or equivalently $\frac{na^2}{8e(4l_0+8Dk)^D}\geq1$ and $\frac{D\sigma na^2}{8e(4l_0+8Dk)^{(D+1)}}\leq1$.
We set $r=2\lceil \frac{na^2}{8e(4l_0+8Dk)^D}\rceil$ to obtain $$\begin{aligned}
{\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + na}) &\leq& (\frac{4(4l_0+8Dk)^Dr}{na^2})^{\frac{r}{2}}+Ce^{\frac{2Dk}{\sigma}}\sigma(\frac{4(\frac{D\sigma}{2})^D r^{D+1}}{na^2})^{\frac{r}{2}} \\ &\leq& (\frac{1}{e})^{\frac{r}{2}} + Ce^{\frac{2Dk}{\sigma}}\sigma(\frac{1}{e}(\frac{D\sigma na^2}{8e(4l_0+8Dk)^{(D+1)}})^D)^{\frac{r}{2}}\\ &\leq& (1+C\sigma e^{\frac{2Dk}{\sigma}})e^{-\frac{na^2}{8e(4l_0+8Dk)^D}}\end{aligned}$$
Last inequality follows from the assumption: $\frac{D\sigma na^2}{8e(4l_0+8Dk)^{(D+1)}}<1$.
For second part of the theorem, consider the hamiltonian $$H'=\sum_{w\in W_{k,m}}\text{I}-h_w.$$ Define $\langle H'\rangle_{\rho} {\ensuremath{ \stackrel{\mathrm{def}}{=} }}{\ensuremath{ \mathrm{Tr} }}(\rho H') = n-\langle H\rangle_{\rho}$. Let $\Pi'_{\geq f}$ be the projector onto subspace with eigenvalues of $H'$ larger than $f$. Same analysis as above for $H'$ in place of $H$, along with the relation $\Pi'_{\geq f} = \Pi_{\leq n-f}$ completes the proof.
Energy distribution of a product state {#sec:momentbound}
======================================
In this section, we introduce the physical set-up for Theorem \[theo:qchernoff\] and also provide its proof. We shall continue using the notations $H$ and $h$ for the hamiltonian and its local term, as this notation is restricted only to this section.
Consider a collection $\C$ of spins, such that a $d$-dimensional Hilbert space $\H_s^d$ is associated to each spin $s\in\C$. Let full Hilbert space $\H$ be defined as $\H=\otimes_s \H_s^d$. For an integer $k>0$, let $S_k$ be the set of all subsets of $\C$ of size at most $k$. For an integer $m>0$, let $W_{k,m}$ be a subset of $S_k$ defined as follows (note that $W_{k,m}$ is also a set of subsets of $\C$) : for each $w\in W_{k,m}$ the number of $w'\in W_{k,m}$ such that $|w'\cap w| > 0$ is at most $m$. For each $w\in W_{k,m}$, let $\N(w)$ be the set of all $w'\in W_{k,m}$ such that $|w\cap w'|>0$. Elements of $\N(w)$ shall be referred to as *neighbours* of $w$. The set-up has been depicted in Figure \[fig:spincollection\].
(0.5,1) circle \[radius=0.05\]; (0.5,2) circle \[radius=0.05\]; (1.3,3) circle \[radius=0.05\]; (2.6,3) circle \[radius=0.05\]; (3.5,2) circle \[radius=0.05\]; (3.5,1) circle \[radius=0.05\]; (2.6,0) circle \[radius=0.05\]; (1.3,0) circle \[radius=0.05\]; (1.4,1) circle \[radius=0.05\]; (2.5,1) circle \[radius=0.05\]; (1.4,2) circle \[radius=0.05\]; (2.5,2) circle \[radius=0.05\]; (0.5,1) – (1.3,0) – (1.4,1) – (0.5,1); (1.4,1) – (0.5,2) – (1.4,2) – (1.4,1); (1.4,2) – (1.3,3) – (2.5,2) – (1.4,2); (2.5,2) – (2.6,3) – (3.5,2) – (2.5,2); (3.5,2) – (3.5,1) – (2.5,1) – (3.5,2); (2.6,0) – (3.5,1) – (2.5,1) – (2.6,0); at (1.1,0.5) [$w_1$]{}; at (1.1,1.6) [$w_2$]{}; at (1.7,2.2) [$w_3$]{}; at (2.9,2.2) [$w_4$]{}; at (3.2,1.3) [$w_5$]{}; at (2.8,0.6) [$w_6$]{};
Let the hamiltonian $H$ be defined as: $$H=\sum_{w\in W_{k,m}} h_w,$$ where $h_w$ acts non-trivially only on spins in $w$ and acts trivially on rest of the spins. Further, we assume that $\|h_w\|_{\infty} \leq 1$.
The definition of $W_{k,m}$ thus translates to the assumption that:
1. Each ‘local term’ $h_w$ acts non-trivially on at most $k$ particle, and hence is $k$-local.
2. For each $h_w$, the number of $h_{w'}$ such that the supports of $h_w$ and $h_{w'}$ intersect, is at most $m$.
Let $\rho \in \H$ be a product state, that is, $\rho = \Pi_{s\in \C} \rho_{s}$ and support of each $\rho_s$ is exactly the spin $s$. Let the reduced density matrix of $\rho$ on a subset $T\subseteq \C$ of spins be denoted in the usual way as $\rho_T$.
We bound the moment function ${\ensuremath{ \mathrm{Tr} }}(\rho(H-\langle H\rangle_{\rho})^r)$ for an even $r$ to be chosen later and use it to prove Theorem \[theo:qchernoff\]. Define $g_w{\ensuremath{ \stackrel{\mathrm{def}}{=} }}h_{w}-\langle h_{w}\rangle_{\rho} \text{I}$.
We shall prove the following lemma.
\[lem:kmomentbound\] Let $n{\ensuremath{ \stackrel{\mathrm{def}}{=} }}|W_{k,m}|$ be the number of local terms. Given the product state $\rho = \Pi_{s\in \C} \rho_s$, it holds that $${\ensuremath{ \mathrm{Tr} }}(\rho(\sum_{w\in W_{k,m}}g_w)^r) \leq (4m^2nr)^{\frac{r}{2}}.$$
Consider,
$$\begin{aligned}
\label{eq:seriesexpansion}
{\ensuremath{ \mathrm{Tr} }}(\rho(\sum_{w\in W_{k,m}}g_w)^r) &=& \sum_{w_1,w_2\ldots w_r} {\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r}) \end{aligned}$$
Using ${\ensuremath{ \mathrm{Tr} }}(\rho g_w)= {\ensuremath{ \mathrm{Tr} }}(\rho_wg_w)=0$, we observe that the term ${\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_1}\ldots g_{w_r})$ is non-zero only if the ordered set $\{w_1,w_2\ldots w_r\}$ satisfies the following property $\Q$: for every $w_i$, there exists a $w_j$ such that $|w_i \cap w_j|> 0$. In other words, there is a $w_j\in W_{k,m}$ such that $w_i \in \N(w_j)$. Let number of ordered sets $\{w_1,w_2\ldots w_r\}$ that satisfy above property be $N_{k,m}(n,r)$. This gives us
$$\label{eq:boundpower}
{\ensuremath{ \mathrm{Tr} }}(\rho(\sum_{w\in \bar{\L}_{D,L}}g_w)\leq N_{k,m}(n,r) \text{max}_{w_1,w_2\ldots w_r}|{\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r})|$$
Setting the trivial bound $\text{max}_{w_1,w_2\ldots w_r}|{\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r})|\leq 1$, and using Lemma \[lem:combinatorial1\] below, the proof follows.
**Remark:** For the case of translationally invariant systems, where $h_w=h$ for all $w$, the bound $\text{max}_{w_1,w_2\ldots w_r}{\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r})\leq 1$ can be improved to $\text{max}_{w_1,w_2\ldots w_r}{\ensuremath{ \mathrm{Tr} }}(\rho g_{w_1}g_{w_2}\ldots g_{w_r})|\leq (1-\langle h\rangle_{\rho})^r$. This gives a minor improvement on the statement of Lemma \[lem:kmomentbound\].
Now we prove an upper bound on the quantity $N_{k,m}(n,r)$ in the following Lemma. Proof is very similar to Lemma \[lem:combinatorial\] and is deferred to Appendix \[append:combinatorial\] for completeness.
\[lem:combinatorial1\] It holds that $N_{k,m}(n,r)\leq (4m^2nr)^{\frac{r}{2}}$.
We restate Theorem \[theo:qchernoff\] formally and give its proof below.
\[formaltheo:qchernoff\] Given the product state $\rho = \Pi_{s\in \C}\rho_s$ with average energy $\langle H\rangle_{\rho}$, consider a real number $a\geq \sqrt{\frac{8em^2}{n}}$. It holds that $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + na}) \leq e^{-\frac{na^2}{4em^2}}$$ and $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\leq \langle H\rangle_{\rho} - na}) \leq e^{-\frac{na^2}{4em^2}}.$$
Lemma \[lem:kmomentbound\] gives the following upper bound on $r$-th moment:
$$\label{eq:momentbound}
{\ensuremath{ \mathrm{Tr} }}(\rho(H-\langle H\rangle_{\rho})^r) \leq (4m^2nr)^{\frac{r}{2}}$$
Using Fact \[fact:markov\], we have $$\begin{aligned}
{\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + na}) &\leq& \frac{(4m^2nr)^{\frac{r}{2}}}{(na)^r} = (\frac{4m^2r}{na^2})^{\frac{r}{2}}. \end{aligned}$$
Choosing $r=2\lceil\frac{na^2}{8em^2}\rceil$, we obtain for $a\geq \sqrt{\frac{8m^2e}{n}}$ $$\begin{aligned}
{\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + na}) &\leq& (\frac{8m^2\cdot na^2}{8em^2\cdot na^2})^{\frac{r}{2}} \leq e^{-\frac{na^2}{4em^2}}. \end{aligned}$$
For second part of the theorem, consider the hamiltonian $$H'=\sum_{w\in W_{k,m}}\text{I}-h_w.$$ Define $\langle H'\rangle_{\rho} {\ensuremath{ \stackrel{\mathrm{def}}{=} }}{\ensuremath{ \mathrm{Tr} }}(\rho H') = n-\langle H\rangle_{\rho}$. Let $\Pi'_{\geq f}$ be the projector onto subspace with eigenvalues of $H'$ larger than $f$. Same analysis as above for $H'$ in place of $H$ gives $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi'_{\geq \langle H'\rangle_{\rho} + na}) \leq e^{-\frac{na^2}{4em^2}}.$$ This completes the proof since $\Pi'_{\geq f} = \Pi_{\leq n-f}$.
Restatement of Theorem \[formaltheo:qchernoff\] in terms of number of spins {#subsec:qchernoff}
---------------------------------------------------------------------------
We introduce a new parameter that captures the number of local terms that act on any given spin. Define $$g_s {\ensuremath{ \stackrel{\mathrm{def}}{=} }}\sum_{w\in W_{k,m}: s\in w}1 , \quad \text{and } g{\ensuremath{ \stackrel{\mathrm{def}}{=} }}\text{max}_s g_s,$$ where $g_s$ is the maximum number of local terms that act non-trivially on spin $s$.
Now, we prove Corollary \[cor:qchernoff\]. Its formal statement is as follows, where we also assume that each local term is exactly $k$-local.
\[formalcor:qchernoff\] Let the hamiltonian $H$ be such that each term $h_w$ has locality equal to $k$. Let $N {\ensuremath{ \stackrel{\mathrm{def}}{=} }}|\C|$ be the number of spins. Given the product state $\rho = \Pi_{s\in \C}\rho_s$ with average energy $\langle H\rangle_{\rho}$, consider a real number ${\varepsilon}\geq \sqrt{8eg^3kN}$. It holds that $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho} + {\varepsilon}}) \leq e^{-\frac{{\varepsilon}^2}{4eg^3kN}}$$ and $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\leq \langle H\rangle_{\rho} - {\varepsilon}}) \leq e^{-\frac{{\varepsilon}^2}{4eg^3kN}}.$$
We set ${\varepsilon}{\ensuremath{ \stackrel{\mathrm{def}}{=} }}na$ as the energy with respect to $H$. Then the bound in Theorem \[formaltheo:qchernoff\] can be restated as: $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho}+{\varepsilon}}) \leq e^{-\frac{{\varepsilon}^2}{4enm^2}}.$$
Relation between $N$ and $n$ can be computed as follows. To each local term $h_w$, one can associate exactly $k$ spins on which $h_w$ acts non-trivially. On the other hand, to each spin $s$, one can associate at most $g$ local terms that contain $s$ in their support. From the first argument, the number of associations is exactly $k\cdot n$, whereas from the second argument, the number of associations is at most $g\cdot N$. Thus, $g\cdot N\geq k\cdot n$ which implies $n\leq \frac{gN}{k}$. Also, $m\leq k\cdot g$, since each local term is supported on $k$ spins, and each of these spins are in the support of at most $g$ other local terms. Collectively we obtain $nm^2 \leq Ng^3k$ and our bound takes the form:
$$\label{eq:qchernoffspinnumber}
{\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho}+{\varepsilon}}) \leq exp(-\frac{{\varepsilon}^2}{4eg^3kN}).$$
This completes the proof.
Above upper bound may be compared to Theorem $7$ in [@kuwahara]. In this reference, the notion of $g'$-extensitivity has been introduced (Definition $2$, [@kuwahara]), which is analogous to the locality parameter $g$ defined above. It is defined as follows: A local hamiltonian $H$ is $g'$-extensive if for every spin $s$, we have $\sum_{w\in W_{k,m}: s\in w} \|h_w\|\leq g'$. Using this, the following theorem has been shown in [@kuwahara]:
\[theo:kuwahara\] Given a $g'$-extensive local hamiltonian with locality $k$, it holds that $${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho}+{\varepsilon}}) \leq \mathcal{O}(1)exp(-\frac{{\varepsilon}^2}{cN\log(\frac{{\varepsilon}}{\sqrt{N}})}),$$ where $c$ is a $\mathcal{O}(1)$ constant that depends only on $k,g'$.
We observe that Equation (\[eq:qchernoffspinnumber\]) achieves a marginally better bound whenever the norm of each local term $h_w$ , that is $\|h_w\|$, is a constant independent of $w$. In such a case, $g'$ and $g$ are same up to the norm of local terms. In case the normalizations of each local term are different, it is not clear how $g,g'$ are related to each other. In such a case Equation (\[eq:qchernoffspinnumber\]) and Theorem \[theo:kuwahara\] may be viewed as complementary results.
Conclusion {#sec:conclusion}
==========
We have shown upper bounds on tail of energy distribution of states that satisfy exponential decay of correlation and product states, with respect to a local hamiltonian. Main technical tool we use is a combinatorial lemma that gives a non-trivial upper bound on the moments of the energy distribution. The results may have applications in the study of thermalization of many body quantum systems and also for many body localization, as noted in the Introduction. Main questions that we leave open are connected to tightness of our bounds, as we discuss below.
The bounds presented in Theorem \[theo:qchernoff\] can only be improved up to constants, since classical Chernoff bound also exhibits a Gaussian decay, which is known to be tight. More interesting situation occurs with the bounds presented in Theorem \[theo:expodecay\]. In one dimensional spin chain, our bound decays exponential with the energy. For gapped ground states, this is very similar to the behaviour noted in [@aradkuwahara] (Section $5$) using completely different techniques. This suggests that gapped ground states (such as the ground state of Transverse field Ising model, which is exactly solvable) are strong candidates for the study of tightness of above results. Our result for higher dimensions appears to be much weaker that those obtained in [@aradkuwahara] (Section $5$) for gapped ground states, and we expect further improvement using better combinatorial arguments.
An another interesting question is with respect to *Matrix product states* (with constant bond dimension) which are defined on one dimensional spin chain. It is well known that under reasonable assumptions (see Section $5.1.1$, [@Orus2014]) Matrix product states satisfy exponential decay of correlation. Furthermore, it has already been shown in [@Ogata2010] that given a Matrix product state $\rho$, if $n$ is large enough and energy ${\varepsilon}\approx \mathcal{O}(n)$, it holds that ${\ensuremath{ \mathrm{Tr} }}(\rho\Pi_{\geq \langle H\rangle_{\rho}+{\varepsilon}})\leq e^{-\mathcal{O}(n)}$. It is a strong indication that our bound (which applies for all energies ${\varepsilon}> \mathcal{O}(\sqrt{n}))$ may be considerably improved for this special, but important, class of states.
Acknowledgement {#acknowledgement .unnumbered}
===============
I thank Itai Arad for introducing me to the question considered here and helpful discussions. I thank Tomotaka Kuwahara for many insightful discussions and sharing an earlier version of his paper [@kuwahara] on the subject considered here. I also thank Fernando Brandao, Rahul Jain and Priyanka Mukhopadhyay for helpful discussions related to proof.
This work is supported by the Core Grants of the Center for Quantum Technologies (CQT), Singapore. Part of this work was done when author was visiting Institute for Quantum Computing (IQC), University of Waterloo, Waterloo.
Proof of Lemma \[lem:combinatorial1\] {#append:combinatorial}
=====================================
We repeat most of the proof of Lemma \[lem:combinatorial\], making changes wherever necessary.
We start with the following definition.
A *selection* is an ordered set $\{(b_1,x_1),(b_2,x_2)\ldots (b_r,x_r)\}$, where $b_i\in \{0,1\}$ and $x_i\in W_{k,m}$, that satisfies the following constraints:
1. If $b_i=0$, then $x_i$ can be any element of $W_{k,m}$ and if $b_i=1$, $x_i$ has to satisfy $|\N(x_i)\cap \N(x_j)|>0$ for some $j<i$.
2. Number of $i$ for which $b_i=0$ is at most $\frac{r}{2}$.
We show the following Lemma, from which the proof of Lemma \[lem:combinatorial\] follows easily.
\[lem:selection1\] Every ordered set $\{w_1,w_2\ldots w_r\}$ that satisfies property $\Q$ can be mapped to a *selection* in such a way that for any two distinct ordered sets satisfying $\Q$, the corresponding *selections* are distinct.
We assign a *selection* to an ordered set $\{w_1,w_2\ldots w_r\}$ satisfying $\Q$ using the algorithm below.
We show that above algorithm terminates and assigns a *selection* to each ordered set satisfying property $\Q$.
1. Consider the running of algorithm during the step **Initialization**. Condition $1$ of a *selection* holds: for every $i$ for which there is a $j<i$ such that $x_i\in \N(x_j)$, we have set $b_i=1$. But we haven’t constructed a *selection* yet, since condition $2$ may not be satisfied.
2. After the step **Pointer creation**, it may be possible that there exist indices $i_1,i_2\ldots i_s$ (for some $s\leq r$) such that $b_{i_1}=b_{i_1}=\ldots b_{i_s}=0$, $R(i_1)=R(i_2)=\ldots R(i_s) > 0$ and $i_s>i_{s-1}>\ldots i_1$. In this case, we find that $w_{i_2}\in \N(w_{R_{i_2}}), w_{i_1}\in \N(w_{R_{i_1}})$. But $\N(w_{R_{i_1}})=\N(w_{R_{i_2}})$, which implies that $|\N(i_1)\cap \N(i_2)| > 0$. Similarly, $|\N(i_1)\cap \N(i_3)| > 0$, …$|\N(i_1)\cap \N(i_s)| > 0$.
Thus, the step **Update** sets $b_{i_2}=b_{i_3}=\ldots b_{i_s}=1$, recognizing the fact that each of the points $w_{i_2},w_{i_3}\ldots w_{i_s}$ satisfy the property that the neighbourhood of each of them intersects with $\N(w_{i_1})$. This ensures that condition $1$ of *selection* is still satisfied.
3. After the step **Update** terminates, condition $2$ of *selection* is now satisfied as well. We now have that for every $i$ with $b_i=0$, there is no other $i'$ such that $R(i)=R(i')$ and $b_i=b_{i'}=0$. Furthermore, $b_{R(i)}=1$. Thus, number of $i$ with $b_i=0$ is at most as large as the number of $j$ with $b_j=1$.
Lemma follows as two distinct ordered tuples satisfying $\Q$ are not assigned the same *selection*.
Now we prove Lemma \[lem:combinatorial1\].
For $n\leq m^2r$, we clearly have $N_{k,m}(n,r)\leq n^r \leq (m^2nr)^{\frac{r}{2}} < (4m^2nr)^{\frac{r}{2}}$.
So we assume $n> m^2r$. We bound the number of *selections*, which gives the desired upper bound on $N_{k,m}(n,r)$ using Lemma \[lem:selection\].
Consider the collection of those *selections*, for which number of $i$ such that $b_i=0$ is $u$. For each $i$ with $b_i=0$, number of possible choices of $x_i$ is $n$. For each $i$ with $b_i=1$, number of possible choices of $x_i$ is at most $m^2r$ (as there are at most $m^2$ number of $x_j\in W_{k,m}$ that satisfy $|\N(x_i)\cap \N(x_j)|>0$ for a given $x_i$). Hence total number of such *selections* is at most ${r \choose u}n^{u}(m^2 r)^{r-u}$. Since $u\leq \frac{r}{2}$, total number of *selections* is at most $$\sum_{u=0}^{\frac{r}{2}}{r \choose u}n^{u}(m^2r)^{r-u} \leq \sum_{u=0}^{\frac{r}{2}}{r \choose u}n^{\frac{r}{2}}(m^2r)^{\frac{r}{2}}<2^r(m^2nr)^{\frac{r}{2}}.$$
This proves the lemma.
[^1]: which is roughly the problem of assigning a temperature to reduced density matrix of thermal state of a local hamiltonian, detailed discussion can be found in [@kliesch13]
[^2]: An interesting class of local hamiltonian system with constant $g,k$ is the family of hamiltonians defined on an expander graph, which has recently been a subject of interest with reference to the Quantum PCP conjecture [@Brandao2013; @Aharonov2013; @eldar2013; @Hastings2014; @eldar2015; @eldarharrow2015].
[^3]: This is a very crude upper bound and can be found as follows. The number of non-negative integers $\{a_1,a_2\ldots a_D\}$ such that $\sum_i a_i \leq 2l$ is at most $(2l)^D$. Thus, number of integers $\{a_1,a_2\ldots a_D\}$ such that $\sum_i |a_i| \leq 2l$ is at most $2^D(2l)^D$.
[^4]: to avoid confusion, we call $\T(l)$ a ‘collection’ instead of a ‘set’
| {
"pile_set_name": "ArXiv"
} |
In atom optics [@AtomOptics] it is usually desirable to separate atoms as far as possible from material objects in order to obtain pure and isolated quantum systems. With cooling and trapping techniques [@LasCool] being well established, there is now an interest in bringing the atoms close to material macroscopic objects. The proximity of the atoms to the object allows the design of tailored and easily controllable potentials which can be used to build novel atom optical elements.
In this letter we demonstrate two simple and versatile atom guides that are based on magnetic trapping potentials created by a thin current carrying wire: The ‘Kepler guide’ and the ‘side guide’. In our experiments we study the transport of cold lithium atoms from a magnetic-optical trap in these guiding potentials. We were able to measure scaling properties and extract characteristic atomic velocity distributions for each guide. The ‘side guide’ is especially interesting because it can easily be miniaturized and combined with other guides to form mesoscopic atom optical networks.
We start with discussing the interaction of a neutral atom and a current carrying wire and then describe our guiding experiments.
[**Kepler guide:** ]{}The magnetic field of a rectilinear current $I$ is given by: $$B = \frac{\mu_{0}}{2 \pi } \frac{I}{r} \hat{e}_{\varphi} ,
\label{MagField}$$ where $\hat{e}_{\varphi}$ is the circular unit vector in cylindrical coordinates. An atom with total spin $\vec{\bf S}$ and magnetic moment $\vec{\mu} = g_S \mu_B \vec{\bf S}$ experiences the interaction potential $V_{mag}= -\vec{\mu} \cdot \vec{\bf B} = - g_S
\mu_B \nF B$, where $\nF$ is the projection of $\vec{\bf S}$ on $\vec{\bf
B}$. In general the vector coupling $\vec{\mu} \cdot \vec{\bf B}$ results in a very complicated motion for the atom. However, in our experiments the Larmor precession ($\omega_{L}$) of the magnetic moment is much faster than the apparent change of direction of the magnetic field in the rest frame of the atom ($\omega_{B}$) and an adiabatic approximation can be applied. $\nF$ is then constant and the atom can be described as moving in a scalar 1/r potential.
For $\vec{\bf \mu}$ “parallel” to $\vec{\bf B}$, ($\vec{\mu} \cdot
\vec{B} > 0$), the atom is in its [*high field seeking state*]{}, and the interaction between the atom and the wire is attractive (see Fig. \[f:cw\_guides\]). The atoms in this state can be trapped and move in Kepler-like orbits around the wire [@Vlad61; @Schm92; @Schm96a; @CuWireTheory].
[**Side guide:** ]{}Combining the field of the current carrying wire with a homogeneous magnetic bias field $B_b$ perpendicular to the wire breaks the rotational symmetry resulting in unstable orbits for the strong field seeking atoms [@BiasWire]. In addition the bias field has the effect of exactly canceling the circular magnetic field of the wire along a line parallel to the wire at a distance $r_s = (\mu_0/2\pi)(I/B_b) $. Around this line the magnetic field increases in all directions and forms a tube with a magnetic field minimum in its center. Atoms in the [*low field seeking state*]{} ($\vec{\mu}\cdot\vec{B} < 0$) can be trapped in this tube and guided along the wire as shown in Fig. \[f:cw\_guides\]b.
Our experiments investigating the Kepler- and side guides are carried out in four steps:
[*(a)*]{} First we load about $2\times10^7$ lithium atoms into a magnetic optic trap (MOT) [@LiMOT], displaced typically 1 mm from a $50
\;
\mu$m thick and $10 \; cm$ long tungsten wire. The MOT is loaded at a distance to the wire in order to prevent trap losses due to atoms hitting the wire [@Den98].
[*(b)*]{} After loading the trap we shut off the slower beam and shift the atoms within $5$ ms to the position where they are loaded into the atom guide. This shifting is done by applying an additional magnetic offset field and moving the center of the magnetic quadrupole field, which defines the position of the MOT. Simultaneously the frequency and intensity of the trapping lasers are changed to control the size and temperature of the atom cloud (typically 1.6 mm diameter(FWHM) and T $\sim$ 200 $\mu$K which corresponds to a velocity of about 0.5 m/s).
[*(c)*]{} We then release the atoms from the MOT by switching off the laser light, the MOT magnetic fields, and the shifting fields within $
0.5$ ms. At this point the current through the wire (typically 1 A) and, if desired, a bias magnetic field is switched on within $
100\;\mu $s. From then on the atoms move in the tailored guiding potential. Starting from an initially well localized atom cloud the density distribution expands and changes shape according to the forces on the atoms.
[*(d)*]{} After a given guiding/trapping time, the spatial distribution of the atoms is measured by imaging the fluorescence from optical molasses [@LasCool] using a CCD camera. For this the guiding fields (current through the wire and bias field) are switched off and molasses laser beams are switched on for a short time (typically $< 1 \; $ms). Pictures are taken from above (looking in wire direction) and from the side (looking onto the wire from an orthogonal direction). This allows to study both, the radial confinement and the guiding of the atoms along the wire.
Typical pictures of atoms orbiting around the wire (Kepler guide) and being guided on the side of the current carrying wire (side guide) are shown in Fig. \[f:WireGuideExp\]. The left set of graphs visually demonstrates loading and guiding of atoms with the Kepler guide: The atoms are released from the MOT at $t=0$ in the center of the wire. Some fraction of these atoms will be bound by the guiding potential, the rest forms an expanding cloud that quickly fades away within about 15 ms. The bound atoms are guided along the wire corresponding to their initial velocity component in this direction. Consequently a cylindrical atomic cloud forms around the wire that expands along the wire. For long guiding times the bound atoms leave the field of view, and the fluorescence signal of the atoms decreases. The top view images show a round atom cloud that is centered on the wire suggesting that atoms circle around the wire.
The graphs on the right hand side of Fig. \[f:WireGuideExp\] show atoms that are bound to the side guide after 20 ms of guiding time. In the given examples two different currents (1 A and 0.5 A) were sent through the wire. The distance of the guide from the wire changes clearly with the current.
The CCD pictures can be used for further analysis as illustrated in figure \[f:MTech\] for the Kepler guide. For this the CCD images are integrated yielding a projection of the density distribution of the atoms in a direction perpendicular to the wire. Figure \[f:MTech\]a shows two such distributions: One corresponding to free expansion of the atomic cloud with no current through the wire. This yields a Gaussian distribution which is typical for atoms released from a MOT. The other one, corresponding to atoms interacting with a current carrying wire, exhibits a pronounced peak centered around the wire which can be attributed to trapped atoms orbiting the wire. The peak sits on top of the broad distribution of non-trapped atoms. In order to extract only the effects of the magnetic guiding potential on the atomic cloud, the two curves are subtracted from each other (Fig. \[f:MTech\]b). The peak of the trapped atoms now sits in a broader dip, which is caused by repulsion of atoms in low field seeking states from the wire and by the fact that atoms trapped around the wire are now missing from the expanding atomic cloud. Our experimental data agree well with numerical simulations of the atomic density distributions as shown in Fig. \[f:MTech\]c,d.
From fits to these experimental data we can extract quantitative information on the guiding of atoms. As an example Fig. \[f:fig4\] shows the number of detected atoms that are bound to the Kepler guide as a function of guiding time. The data is given as the fraction of the total number of atoms of the MOT. The observed number of guided atoms decreases with time mainly because the expanding atomic cloud leaves the detection region given by the 2 cm diameter laser beams. The solid line represents a corresponding calculation of how the atomic signal is expected to fall due to the free expansion of the atomic cloud along the wire and the falling under the influence of gravity. We attribute the additional loss observed in the experiment to magnetic stray fields that render the atomic orbits unstable, i.e. atoms will hit the wire and are lost. There will also be a small contribution to the losses because of a decrease in the wire current over time, caused by the increasing resistance due to ohmic heating of the wire.
From Fig. \[f:fig4\] we can also extract the absolute loading efficiency for atoms from the MOT into the guide. In agreement with Monte Carlo calculations we find loading efficiencies in the range of 10 % for the [*Kepler guide*]{}.
Using the same wire current, the loading of the [*side guide*]{} is less efficient (up to $4\%$ loading efficiency), mainly due to the smaller depth and ‘size’ of the side guide potential.
However, the side guide exhibits interesting scaling properties: Its trap depth is given by the [*magnitude*]{} of the bias field. With a fixed trap depth the trap size and its distance from the wire can be controlled by the current in the wire. The paradoxical situation arises that the trap gets smaller (size $\propto I/B_b$) and steeper (gradient $\propto B_b^2 / I$) for [*decreasing*]{} current in the wire. The smallest and steepest trap achievable with a fixed bias field is only limited by the requirement that it must be located outside the wire. A simple calculation shows for example that a trap with a gradient of over 1000 Gauss/cm can be achieved with a moderate current of 0.5 A and an offset field of 10 Gauss. The trap would be located 100 $\mu$m away from the wire center. Figure \[f:Compress\] illustrates the linear scaling of $r_s$, the distance of the side guide from the wire, as a function of the wire current (see also the right hand side of Fig. \[f:WireGuideExp\]).
Other interesting information about the transverse confinement can be extracted by measuring the momentum distribution of the trapped atoms. This can be accomplished by ballistic expansion after switching off the guiding potentials. After a few ms of expansion the spatial distribution well represents the velocity distribution of the atoms. Figure \[f:fig6\] shows the spatial atomic distribution 9 (7) ms after switching off the guides. A clear distinction can be seen between the two types of guides. Atoms in the Kepler guide, where atoms circle around the wire, expand in a ring (Fig. \[f:fig6\]a), showing clearly that there are no zero-velocity atoms. This is because in order to be trapped in stable orbits [*around*]{} the wire the atoms need sufficient angular momentum and therefore velocity. Atoms with too little angular momentum hit the wire and are lost. The low field seeker atoms of the side guide, however, do not have this constraint. Their velocity distribution is a standard Gaussian (Fig. \[f:fig6\]b).
In conclusion we presented two novel methods to build guides for cold atoms, by using potentials created by a current carrying wire: Guiding [*strong field seekers*]{} in Kepler-like orbits around a wire and guiding [*weak field seekers*]{} in a potential tube along the side of a wire. These atom guides represent promising techniques for future applications in atom optics, because of their simplicity and versatility. For example by combining different wires we can construct beam splitters [@test], interferometers and more complex matter wave networks [@Matter_Guides]. Achieving the ultra high vacuum (UHV) conditions required for coherent guiding is very simple, since the propagation of atoms is in the open and [*not*]{} in an enclosed space like for hollow optical fibers [@HollowFibers].
In addition side guides can be mounted on a surface, so that atoms are guided above the surface along the wires. This renders the wires more stable and at the same time allows for efficient cooling which enables also thin wires to support sizeable currents (see also [@SurfaceMounting]). Therefore we believe to have now a method for miniaturization and integration of many atom optical elements into one single quantum circuit in the near future, creating mesoscopic atom optics similar to mesoscopic quantum electronics.
We thank A. Zeilinger for his generous support throughout the work. This work was supported by the Austrian Science Foundation (FWF), project S065-05, the Jubiläums Fonds der Österreichischen Nationalbank, project 6400, and by the European Union, contract Nr. TMRX-CT96-0002.
[99]{}
For an overview see: C.S. Adams, M. Sigel, J. Mlynek, Phys. Rep. [**240**]{}, 143 (1994); [*Atom Interferometry*]{}, edited P. Berman (Academic Press, 1997) and references therein.
A good overview of laser cooling is given in: [*Laser Manipulation of Atoms and Ions*]{}, edited by E. Arimondo, W.D. Phillips, and F. Strumia (North Holland, 1992); and S. Chu, Rev. Mod. Phys. [**70**]{}, 685 (1998); C. Cohen-Tannoudji, Rev. Mod. Phys. [**70**]{}, 707 (1998); W.D. Phillips, Rev. Mod. Phys. [**70**]{}, 721 (1998).
V. V. Vladimirskii, Sov. Phys. JETP [**12**]{}, 740 (1961).
J. Schmiedmayer in [*XVIII International Conference on Quantum Electronics:*]{} Technical Digest, edited by G. Magerl (Technische Universität Wien, Vienna 1992), Series 1992, Vol. 9, 284 (1992); Appl. Phys. B [**60**]{}, 169 (1995); Phys. Rev. A [**52**]{}, R13 (1995).
Actually the adiabatic approximation results in an additional precession of the Kepler orbits around the wire as described in: J. Schmiedmayer and A. Scrinzi, Phys. Rev. A [**54**]{}, R2525 (1996); JEOS [**8**]{}, 693 (1996).
For quantum calculations see: G. P. Pron’kov and Yu. G. Stroganov, Zh. Eksp. Teor. Fiz. [ **72**]{}, 2048 (1977) \[Sov. Phys. JETP [**45**]{}, 1075 (1977)\]; R. Blümel and K. Dietrich, Phys. Lett. A [**139**]{}, 236 (1989); Phys. Rev. A [**43**]{}, 22 (1991); A. I. Voronin, Phys. Rev. A [**43**]{}, 29 (1991); L. Hau, J. Golovchenko, M. Burns, Phys. Rev. Lett. [**74**]{}, 3138 (1995).
With the rotational symmetry broken, angular momentum is no longer conserved. This results in unstable orbits for the strong field seeking atoms, and they will finally hit the wire.
The atoms are loaded into the MOT for 20 sec out of an effusive thermal beam at a red laser detuning of 25 MHz and a total beam power of about 150 mW. An electro-optic modulator produces sidebands of 812 MHz (30%) one of which is used as a repumper. To increase the loading rate by a factor of 5 we use an additional 20 mW slower beam with a red detuning of about 100 MHz directed through the MOT into the oven.
J. Denschlag, G. Umshaus, and J. Schmiedmayer, Phys. Rev. Lett. [**81**]{}, 737 (1998).
E.L. Raab [*et al.*]{}, Phys. Rev. Lett. [**59**]{}, 2631 (1987).
In our lab we are currently experimentally testing an atomic beamsplitter that is formed by two intersecting current carrying wires.
J. Schmiedmayer Habilitationsschrift Universität Innsbruck (1996); J. Denschlag and J. Schmiedmayer in [*Proceedings of the International Quantum Coherence Conference*]{}, (1997), Boston, World Scientific; J. Schmiedmayer, (1998), EPJ [**D**]{} in print.
Guiding in hollow fibers is a standard technology in neutron optics: M.A. Kumakhov, V.A. Sharov, Nature [**357**]{}, 390 (1992); H. Chen [*et al.*]{}, Nature [**357**]{}, 391 (1992). Guiding atoms in hollow optical fibers was proposed by: M.A. Ol’Shanii, Yu.B. Ovchinnikov, V.S. Letokhov, Opt. Comm. [**98**]{}, 77 (1993); S. Marksteiner, C.M. Savage, P. Zoller, S.L. Rolston, Phys. Rev. A [**50**]{}, 2680 (1994); and experimentally demonstrated by: M.J. Renn [*et al.*]{}, Phys. Rev. Lett. [**75**]{}, 3253 (1995). M. Drndic [*et al.*]{}, Appl. Phys. Lett. [**72**]{}, 2906 (1998).
| {
"pile_set_name": "ArXiv"
} |
Several years ago, D. Bessis conjectured on the basis of numerical studies that the spectrum of the Hamiltonian $H=p^2+x^2+ix^3$ is [*real and positive*]{} [@BESSIS]. To date there is no rigorous proof of this conjecture. We claim that the reality of the spectrum of $H$ is due to ${\cal PT}$ symmetry. Note that $H$ is invariant [*neither*]{} under parity ${\cal P}$, whose effect is to make spatial reflections, $p\to-p$ and $x\to-x$, [*nor*]{} under time reversal ${\cal T}$, which replaces $p\to-p$, $x\to x$, and $i\to-i$. However, ${\cal PT}$ symmetry is crucial. For example, the Hamiltonian $p^2+ix^3+ix$ has ${\cal PT}$ symmetry and our numerical studies indicate that its entire spectrum is positive definite; the Hamiltonian $p^2+ix^3+x$ is not ${\cal PT}$-symmetric, and the entire spectrum is complex.
The connection between ${\cal PT}$ symmetry and positivity of spectra is simply illustrated by the harmonic oscillator $H=p^2+x^2$, whose energy levels are $E_n
=2n+1$. Adding $ix$ to $H$ does not break ${\cal PT}$ symmetry, and the spectrum remains positive definite: $E_n=2n+{5\over4}$. Adding $-x$ also does not break ${\cal PT}$ symmetry if we define ${\cal P}$ as reflection about $x={1\over2}$, $x\to1-x$, and again the spectrum remains positive definite: $E_n=2n+{3\over4}$. By contrast, adding $ix-x$ [*does*]{} break ${\cal PT}$ symmetry, and the spectrum is now complex: $E_n=2n+1+{1\over2}i$.
The Hamiltonian studied by Bessis is just one example of a huge and remarkable class of non-Hermitian Hamiltonians whose energy levels are real and positive. The purpose of this Letter is to understand the fundamental properties of such a theory by examining the class of quantum-mechanical Hamiltonians $$\begin{aligned}
H=p^2+m^2 x^2-(ix)^N\quad(N~{\rm real}).
\label{e1}\end{aligned}$$ As a function of $N$ and mass $m^2$ we find various phases with transition points at which entirely real spectra begin to develop complex eigenvalues.
There are many applications of non-Hermitian ${\cal PT}$-invariant Hamiltonians in physics. Hamiltonians rendered non-Hermitian by an imaginary external field have been introduced recently to study delocalization transitions in condensed matter systems such as vortex flux-line depinning in type-II superconductors [@Hatano+Nelson], or even to study population biology [@Nelson+Shnerb]. Here, initially real eigenvalues bifurcate into the complex plane due to the increasing external field, indicating the unbinding of vortices or the growth of populations. We believe that one can also induce dynamic delocalization by tuning a physical parameter (here $N$) in a self-interacting theory.
Furthermore, it was found that quantum field theories analogous to the quantum-mechanical theory in Eq. (\[e1\]) have astonishing properties. The Lagrangian $L=(\nabla\phi)^2+m^2 \phi^2-g(i\phi)^N$ ($N$ real) possesses ${\cal PT}$ invariance, the fundamental symmetry of local self-interacting scalar quantum field theory [@PCT]. Although this theory has a non-Hermitian Hamiltonian, the spectrum of the theory appears to be positive definite. Also, $L$ is explicitly not parity invariant, so the expectation value of the field $\langle\phi\rangle$ is nonzero, even when $N=4$ [@BM]. Thus, one can calculate directly (using the Schwinger-Dyson equations, for example [@MILTON]) the (real positive) Higgs mass in a renormalizable theory such as $-g\phi^4$ or $ig\phi^3$ in which symmetry breaking occurs naturally (without introducing a symmetry-breaking parameter).
Replacing conventional $g\phi^4$ or $g\phi^3$ theories by $-g\phi^4$ or $ig
\phi^3$ theories has the effect of reversing signs in the beta function. Thus, theories that are not asymptotically free become asymptotically free and theories that lack stable critical points develop such points. For example, ${\cal PT}$-symmetric massless electrodynamics has a nontrivial stable critical value of the fine-structure constant $\alpha$ [@ALPHA].
Supersymmetric non-Hermitian, ${\cal PT}$-invariant Lagrangians have been examined [@BMsuper]. It is found that the breaking of parity symmetry does not induce a breaking of the apparently robust global supersymmetry. The strong-coupling limit of non-Hermitian ${\cal PT}$-symmetric quantum field theories has been investigated [@BBJM]; the correlated limit in which the bare coupling constants $g$ and $-m^2$ both tend to infinity with the renormalized mass $M$ held fixed and finite, is dominated by solitons. (In parity-symmetric theories the corresponding limit, called the Ising limit, is dominated by instantons.)
To elucidate the origin of such novel features we examine the elementary Hamiltonian (\[e1\]) using extensive numerical and asymptotic studies. As shown in Fig. \[fig1\], when $m=0$ the spectrum of $H$ exhibits three distinct behaviors as a function of $N$. When $N\geq2$, the spectrum is infinite, discrete, and entirely real and positive. (This region includes the case $N=4$ for which $H=p^2-x^4$; the spectrum of this Hamiltonian is positive and discrete and $\langle x\rangle\neq0$ in the ground state because $H$ breaks parity symmetry!) At the lower bound $N=2$ of this region lies the harmonic oscillator. A phase transition occurs at $N=2$; when $1<N<2$, there are only a [*finite*]{} number of real positive eigenvalues and an infinite number of complex conjugate pairs of eigenvalues. In this region ${\cal PT}$ symmetry is [*spontaneously broken*]{} [@BBM]. As $N$ decreases from $2$ to $1$, adjacent energy levels merge into complex conjugate pairs beginning at the high end of the spectrum; ultimately, the only remaining real eigenvalue is the ground-state energy, which diverges as $N\to1^+$ [@SH]. When $N\leq1$, there are no real eigenvalues. The massive case $m\neq0$ is even more elaborate; there is a phase transition at $N=1$ in addition to that at $N=2$.
=2.2truein 0.15truein
The Schrödinger eigenvalue differential equation corresponding to the Hamiltonian (\[e1\]) with $m=0$ is $$\begin{aligned}
-\psi''(x)-(ix)^N\psi(x)=E\psi(x).
\label{e2}\end{aligned}$$ Ordinarily, the boundary conditions that give quantized energy levels $E$ are that $\psi(x)\to0$ as $|x|\to\infty$ on the real axis; this condition suffices when $1<N<4$. However, for arbitrary real $N$ we must continue the eigenvalue problem for (\[e2\]) into the complex-$x$ plane. Thus, we replace the real-$x$ axis by a contour in the complex plane along which the differential equation holds and we impose the boundary conditions that lead to quantization at the endpoints of this contour. (Eigenvalue problems on complex contours are discussed in Ref. [@ROT].)
=2.2truein 0.15truein
The regions in the cut complex-$x$ plane in which $\psi(x)$ vanishes exponentially as $|x|\to\infty$ are [*wedges*]{} (see Fig. \[fig2\]); these wedges are bounded by the [*Stokes lines*]{} of the differential equation [@BO]. The center of the wedge, where $\psi(x)$ vanishes most rapidly, is called an [*anti-Stokes line.*]{}
There are many wedges in which $\psi(x)\to0$ as $|x|\to\infty$. Thus, there are many eigenvalue problems associated with a given differential equation [@ROT]. However, we choose to continue the eigenvalue equation (\[e2\]) away from the conventional harmonic oscillator problem at $N=2$. The wave function for $N=2$ vanishes in wedges of angular opening ${1\over2}\pi$ centered about the negative- and positive-real $x$ axes. For arbitrary $N$ the anti-Stokes lines at the centers of the left and right wedges lie at the angles $$\begin{aligned}
\theta_{\rm left}=-\pi+{N-2\over N+2}~{\pi\over2}\quad{\rm and}\quad
\theta_{\rm right}=-{N-2\over N+2}~{\pi\over2}.
\label{e3}\end{aligned}$$ The opening angle of these wedges is $\Delta=2\pi/(N+2)$. The differential equation (\[e2\]) may be integrated on any path in the complex-$x$ plane so long as the ends of the path approach complex infinity inside the left wedge and the right wedge [@QFT]. Note that these wedges contain the real-$x$ axis when $1<N<4$.
As $N$ increases from $2$, the left and right wedges rotate downward into the complex-$x$ plane and become thinner. At $N=\infty$, the differential equation contour runs up and down the negative imaginary axis and thus there is no eigenvalue problem at all. Indeed, Fig. \[fig1\] shows that the eigenvalues all diverge as $N\to\infty$. As $N$ decreases below $2$ the wedges become wider and rotate into the upper-half $x$ plane. At $N=1$ the angular opening of the wedges is ${2\over3}\pi$ and the wedges are centered at ${5\over6}\pi$ and ${1\over6}\pi$. Thus, the wedges become contiguous at the positive-imaginary $x$ axis, and the differential equation contour can be pushed off to infinity. Consequently, there is no eigenvalue problem when $N=1$ and, as we would expect, the ground-state energy diverges as $N\to1^+$ (see Fig. \[fig1\]).
To ensure the numerical accuracy of the eigenvalues in Fig. \[fig1\], we have solved the differential equation (\[e2\]) using two independent procedures. The most accurate and direct method is to convert the complex differential equation to a system of coupled, real, second-order equations which we solve using the Runge-Kutta method; the convergence is most rapid when we integrate along anti-Stokes lines. We then patch the two solutions together at the origin. We have verified those results by diagonalizing a truncated matrix representation of the Hamiltonian in Eq. (\[e1\]) in harmonic oscillator basis functions.
[*Semiclassical analysis:*]{} Several features of Fig. \[fig1\] can be verified analytically. When $N\geq2$, WKB gives an excellent approximation to the spectrum. The novelty of this WKB calculation is that it must be performed in the complex plane. The turning points $x_{\pm}$ are those roots of $E+(ix)^N=0$ that [*analytically continue*]{} off the real axis as $N$ moves away from $N=2$ (the harmonic oscillator): $$\begin{aligned}
x_-=E^{1/N}e^{i\pi(3/2-1/N)},\quad x_+=E^{1/N}e^{-i\pi(1/2-1/N)}.
\label{e5}\end{aligned}$$ These turning points lie in the lower-half (upper-half) $x$ plane in Fig. \[fig2\] when $N>2$ ($N<2$).
The leading-order WKB phase-integral quantization condition is $(n+1/2)\pi=\int_{x_-}^{x_+}dx\,\sqrt{E+(ix)^N}$. It is crucial that this integral follow a path along which the [*integral is real.*]{} When $N>2$, this path lies entirely in the lower-half $x$ plane and when $N=2$ the path lies on the real axis. But, when $N<2$ the path is in the upper-half $x$ plane; it crosses the cut on the positive-imaginary
[llddlldd]{} $N$ & $n$ & $E_{\rm exact}$ & $E_{\rm WKB}$ & $N$ & $n$ & $E_{\rm exact}$ & $E_
{\rm WKB}$\
3 & 0 & 1.1562& 1.0942& 4 & 0 & 1.4771 & 1.3765\
&1 & 4.1092& 4.0894& & 1 & 6.0033 & 5.9558\
&2 & 7.5621& 7.5489& & 2 & 11.8023 & 11.7689\
&3 & 11.3143& 11.3042& & 3 & 18.4590 & 18.4321\
&4 & 15.2916& 15.2832& & & &\
\[table1\]
axis and thus is [*not a continuous path joining the turning points.*]{} Hence, WKB fails when $N<2$.
When $N\geq2$, we deform the phase-integral contour so that it follows the rays from $x_-$ to $0$ and from $0$ to $x_+$: $(n+1/2)\pi=2\sin(\pi/N)E^{1/N+1/2}
\int_0^1 ds\,\sqrt{1-s^N}$. We then solve for $E_n$: $$\begin{aligned}
E_n\sim\left[{\Gamma(3/2+1/N)\sqrt{\pi}(n+1/2)\over\sin(\pi/N)\Gamma(1+1/N)}
\right]^{2N\over N+2}\quad(n\to\infty).
\label{e7}\end{aligned}$$ We perform a higher-order WKB calculation by replacing the phase integral by a [*closed contour*]{} that encircles the path in Fig. \[fig2\] (see Ref. [@BO; @BBM]). See Table I.
It is interesting that the spectrum of the $|x|^N$ potential is like that of the $-(ix)^N$ potential. The leading-order WKB quantization condition (accurate for $N>0$) is like Eq. (\[e7\]) except that $\sin(\pi/N)$ is absent. However, as $N\to\infty$, the spectrum of $|x|^N$ approaches that of the square-well potential \[$E_n=(n+1)^2\pi^2/4$\], while the energies of the $-(ix)^N$ potential diverge (see Fig. 1).
[*Asymptotic study of the ground-state energy near $N=1$*]{}: When $N=1$, the differential equation (\[e2\]) can be solved exactly in terms of Airy functions. The anti-Stokes lines at $N=1$ lie at $30^\circ$ and at $150^\circ$. We find the solution that vanishes exponentially along each of these rays and then rotate back to the real-$x$ axis to obtain $$\begin{aligned}
\psi_{\rm left,\,right}(x)=C_{1,\,2}\,{\rm Ai}(\mp xe^{\pm i\pi/6}
+Ee^{\pm 2i\pi/3}).
\label{e8}\end{aligned}$$ We must patch these solutions together at $x=0$ according to the patching condition $\left.{d\over dx}|\psi(x)|^2\right|_{x=0}=0$. But for real $E$, the Wronskian identity for the Airy function is $$\begin{aligned}
{d\over dx}|{\rm Ai}(xe^{-i\pi/6}+Ee^{-2i\pi/3})|^2\Bigm|_{x=0}=-{1\over2\pi}
\label{e9}\end{aligned}$$ instead of $0$. Hence, there is no real eigenvalue.
Next, we perform an asymptotic analysis for $N=1+\epsilon$, $-\psi''(x)-(ix)^{1+
\epsilon}\psi(x)=E\psi(x)$, and take $\psi(x)=y_0(x)+\epsilon y_1(x)+{\rm O}
(\epsilon^2)$ as $\epsilon\to0+$. We assume that $E\to\infty$ as $\epsilon\to0+$, let $C_2=1$ in Eq. (\[e8\]), and obtain $$\begin{aligned}
y_0(0)={\rm Ai}(Ee^{-2i\pi/3})\sim e^{i\pi/6}E^{-1/4}e^{{2\over3}E^{3/2}}/
(2\sqrt{\pi}).
\label{e10}\end{aligned}$$ We set $y_1(x)=Q(x)y_0(x)$ in the inhomogeneous equation $-y_1''(x)-ixy_1(x)-Ey_1(x)=ix\ln(ix)y_0(x)$ and get $$\begin{aligned}
Q'(0)={i\over y_0^2(0)}\int_0^{\infty}dx\,x\,\ln(ix)y_0^2(x).
\label{e11}\end{aligned}$$
Choosing $Q(0)=0$, we find that the patching condition at $x=0$ gives $1=2\pi\epsilon\left|y_0(0)\right|^2[Q'(0)+{Q^*}'(0)]$, where we have used the zeroth-order result in Eq. (\[e9\]). Using Eqs. (\[e10\]) and (\[e11\]) this equation becomes $$\begin{aligned}
1={\epsilon\over \sqrt{E}}e^{{4\over3}E^{3/2}}{\rm Re}\,\left[{i\over y_0^2(0)}
\int_0^{\infty}dx\,x\,\ln(ix)y_0^2(x)\right].
\label{e12}\end{aligned}$$ Since $y_0(x)$ decays rapidly as $x$ increases, the integral in Eq. (\[e12\]) is dominated by contributions near $0$. Asymptotic analysis of this integral gives an implicit equation for $E$ as a function of $\epsilon$ (see Table II): $$\begin{aligned}
1\sim\epsilon e^{{4\over3}E^{3/2}}E^{-3/2}[\sqrt{3}\ln(2\sqrt{E})
+\pi-(1-\gamma)\sqrt{3}]/8.
\label{e13}\end{aligned}$$
[*Behavior near $N=2$*]{}: The most interesting aspect of Fig. \[fig1\] is the transition that occurs at $N=2$. To describe quantitatively the merging of eigenvalues that begins when $N<2$ we let $N=2-\epsilon$ and study the asymptotic behavior as $\epsilon\to0+$. (A Hermitian perturbation causes adjacent energy levels to repel, but in this case the non-Hermitian perturbation of the harmonic oscillator $(ix)^{2-\epsilon}\sim x^2-\epsilon x^2[\ln(|x|+{1
\over2}i\pi\,{\rm sgn}(x)]$ causes the levels to merge.) A complete description of this asymptotic study is given elsewhere [@BBM].
The onset of eigenvalue merging is a phase transition that occurs even at the [*classical*]{} level. Consider the classical equations of motion for a particle of energy $E$ subject to the complex forces described by the Hamiltonian (\[e1\]). For $m=0$ the trajectory $x(t)$ of the particle obeys $\pm dx[E+
(ix)^N]^{-1/2}=2dt$. While $E$ and $dt$ are real, $x(t)$ is a path in the complex plane in Fig. \[fig2\]; this path terminates at the classical turning points $x_\pm$ in (\[e5\]).
When $N\geq2$, the trajectory is an arc joining $x_\pm$ in the lower complex plane. The motion is [*periodic*]{}; we have a complex pendulum whose (real) period $T$ is $$\begin{aligned}
T=2E^{2-N\over2N}\cos\left[{(N-2)\pi\over2N}\right]{\Gamma(1+1/N)\sqrt{\pi}\over
\Gamma(1/2+1/N)}.
\label{e14}\end{aligned}$$
At $N=2$ there is a global change. For $N<2$ a path starting at one turning point, say $x_+$, moves toward but [*misses*]{} the turning point $x_-$. This path spirals outward crossing from sheet to sheet on the Riemann surface, and eventually veers off to infinity asymptotic to the angle ${N\over2-N}\pi$. Hence, the period abruptly becomes infinite. The total angular rotation of the spiral is finite for all $N\neq2$ and as $N\to2^+$, but becomes infinite as $N\to2^-$. The path passes many turning points as it spirals anticlockwise from $x_+$. \[The $n$th turning point lies at the angle ${4n-N+2\over2N}\pi$ ($x_+$ corresponds to $n=0$).\] As $N$ approaches $2$ from below, when the classical trajectory passes a new
[ldd]{} $\epsilon=N-1$ & $E_{\rm exact}$ & Eq. (\[e13\])\
0.1 & 1.6837& 2.0955\
0.01 &2.6797&2.9624\
0.001 & 3.4947&3.6723\
0.0001 & 4.1753&4.3013\
0.00001 & 4.7798&4.8776\
0.000001 & 5.3383&5.4158\
0.0000001 & 5.8943&5.9244\
\[table2\]
turning point, there corresponds an additional merging of the quantum energy levels as shown in Fig. \[fig1\]). This correspondence becomes exact in the limit $N\to2^-$ and is a manifestation of Ehrenfest’s theorem.
[*Massive case*]{}: The $m\neq0$ analog of Fig. \[fig1\] exhibits a new transition at $N=1$ (see Fig. \[fig3\]). As $N$ approaches $1$ from above, the energy levels reemerge from the complex plane in pairs and at $N=1$ the spectrum is again entirely real and positive. Below $N=1$ the energies once again disappear in pairs, now including the ground state. As $N\to0$ the infinite real spectrum reappears again. The massive case is discussed further in Ref. [@BBM].
=2.2truein 0.15truein
We thank D. Bessis, H. Jones, P. Meisinger, A. Wightman, and Y. Zarmi for illuminating conversations. CMB thanks the Center for Nonlinear Studies, Los Alamos National Laboratory and STB thanks the Physics Department at Washington University for its hospitality. This work was supported by the U.S. Department of Energy.
D. Bessis, private discussion. This problem originated from discussions between Bessis and J. Zinn-Justin, who was studying Lee-Yang singularities using renormalization group methods. An $i\phi^3$ field theory arises if one translates the field in a $\phi^4$ theory by an imaginary term.
N. Hatano and D. R. Nelson, Phys. Rev. Lett. [**77**]{}, 570 (1996), and Phys. Rev. B [**56**]{}, 8651 (1997).
D. R. Nelson and N. M. Shnerb, condmat/9708071.
R. F. Streater and A. S. Wightman, [*PCT, Spin & Statistics, and all that*]{} (Benjamin, New York, 1964). There is no analog of the ${\cal C}$ operator in quantum systems having one degree of freedom.
C. M. Bender and K. A. Milton, Phys. Rev. D [**55**]{}, R3255 (1997). There is a factor of 2 error in Eqs. (15) and (16). C. M. Bender and K. A. Milton (submitted). C. M. Bender and K. A. Milton (submitted). C. M. Bender and K. A. Milton, Phys. Rev. D [**57**]{}, 3595 (1998).
C. M. Bender, S. Boettcher, H. F. Jones, and P. N. Meisinger, Phys. Rev. D (submitted).
C. M. Bender, S. Boettcher, and P. N. Meisinger, Phys. Rev. D (submitted).
It is known that the spectrum of $H=p^2-ix$ is null. See I. Herbst, Commun. Math. Phys. [**64**]{}, 279 (1979).
C. M. Bender and A. Turbiner, Phys. Lett. A [**173**]{}, 442 (1993). C. M. Bender and S. A. Orszag, [*Advanced Mathematical Methods for Scientists and Engineers*]{} (McGraw-Hill, New York, 1978).
In a Euclidean path integral representation for a quantum field theory, the (multiple) integration contour follows the same anti-Stokes lines. See Ref. [@BM].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A fast and efficient numerical-analytical approach is proposed for modeling complex behaviour in the BBGKY hierarchy of kinetic equations. We construct the multiscale representation for hierarchy of reduced distribution functions in the variational approach and multiresolution decomposition in polynomial tensor algebras of high-localized states. Numerical modeling shows the creation of various internal structures from localized modes, which are related to localized or chaotic type of behaviour and the corresponding patterns (waveletons) formation. The localized pattern is a model for energy confinement state (fusion) in plasma.'
address: |
IPME RAS, St. Petersburg, V.O. Bolshoj pr., 61, 199178, Russia\
E-mail: [email protected], [email protected]\
http://www.ipme.ru/zeitlin.html, http://www.ipme.nw.ru/zeitlin.html
author:
- 'Antonina N. Fedorova and Michael G. Zeitlin'
title: |
Classical and Quantum Ensembles via Multiresolution.\
I. BBGKY Hierarchy
---
[p[130mm]{}]{}
[**CLASSICAL AND QUANTUM ENSEMBLES**]{}\
[**VIA MULTIRESOLUTION.**]{}\
[**I. BBGKY HIERARCHY**]{}\
[**Antonina N. Fedorova, Michael G. Zeitlin**]{}\
[***IPME RAS, St. Petersburg, V.O. Bolshoj pr., 61, 199178, Russia***]{}\
[***e-mail: [email protected]***]{}\
[***e-mail: [email protected]***]{}\
[***http://www.ipme.ru/zeitlin.html***]{}\
[***http://www.ipme.nw.ru/zeitlin.html***]{}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
A fast and efficient numerical-analytical approach is proposed for modeling complex behaviour in the BBGKY hierarchy of kinetic equations. We construct the multiscale representation for hierarchy of reduced distribution functions in the variational approach and multiresolution decomposition in polynomial tensor algebras of high-localized states. Numerical modeling shows the creation of various internal structures from localized modes, which are related to localized or chaotic type of behaviour and the corresponding patterns (waveletons) formation. The localized pattern is a model for energy confinement state (fusion) in plasma.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
[Presented at IX International Workshop on Advanced]{}\
[Computing and Analysis Techniques in Physics Research]{}\
[ACAT03, December, 2003, KEK, Tsukuba, Japan]{}
[Nuclear Instruments and Methods in Physics Research A, in press]{}
INTRODUCTION
============
Kinetic theory is an important part of general statistical physics related to phenomena which cannot be understood on the thermodynamic or fluid models level \[1\]. In these two papers we consider the applications of a new numerical/analytical technique based on wavelet analysis approach for calculations related to the description of complex (non-equilibrium) behaviour of the corresponding classical and quantum ensembles. The classical ensembles in this part are considered in the framework of the general BBGKY hierarchy and the quantum ones in part 2 in the Wigner-Weyl approach \[1\]. We restrict ourselves to the rational/polynomial type of nonlinearities (with respect to the set of all dynamical variables) that allows to use our results from \[2\], which are based on the so called multiresolution framework \[3\] and the variational formulation of initial nonlinear (pseudodifferential) problems. Wavelet analysis is a set of mathematical methods which give a possibility to work with well-localized bases in functional spaces and provide the maximum sparse forms for the general type of operators (differential, integral, pseudodifferential) in such bases. It provides the best possible rates of convergence and minimal complexity of algorithms inside and, as a result, saves CPU time and HDD space \[3\]. Our main goals are an attempt of classification and construction of a possible zoo of nontrivial (meta) stable states: (a) high-localized (nonlinear) eigenmodes, (b) complex (chaotic-like or entangled) patterns, (c) localized (stable) patterns (waveletons). In case (c) an energy is distributed during some time (sufficiently large) between only a few localized modes (from point (a)). We believe, it is a good image for plasma in a fusion state (energy confinement). Our construction of cut-off of the infinite system of equations is based on some criterion of convergence of the full solution. This criterion is based on a natural norm in the proper functional space, which takes into account (non-perturbatively) the underlying multiscale structure of complex statistical dynamics. In Sec. 2 the kinetic BBGKY hierarchy is formulated. In Sec. 3 we present the explicit analytical construction of solutions of the hierarchy, which is based on tensor algebra extensions of bases generated by the hidden multiresolution structure and proper variational formulation leading to an algebraic parametrization of the solutions. So, our approach resembles Bogolyubov’s and related approaches but we don’t use any perturbation technique (like virial expansion) or linearization procedures. Numerical modeling as in general case as in particular cases of the Vlasov-like equations shows the creation of various internal structures from localized bases modes, which demonstrate the possiblity of existence of (metastable) pattern formation.
BBGKY HIERARCHY
===============
Let $M$ be the phase space of an ensemble of $N$ particles ($ {\rm dim}M=6N$) with coordinates $x_i=(q_i,p_i), \quad i=1,...,N, \quad
q_i=(q^1_i,q^2_i,q^3_i)\in R^3,\quad
p_i=(p^1_i,p^2_i,p^3_i)\in R^3,\quad
q=(q_1,\dots,q_N)\in R^{3N}$. Individual and collective measures are: $
\mu_i={\mathrm{d}}x_i={\mathrm{d}}q_i{\mathrm{d}}p_i,\quad \mu=\prod^N_{i=1}\mu_i
$. Our constructions can be applied to the following general Hamiltonians: $$H_N=
\sum^N_{i=1}\Big(\frac{p^2_i}{2m}+U_i(q)\Big)+
\sum_{1\leq i\leq j\leq N}U_{ij}(q_i,q_j)$$ where the potentials $U_i(q)=U_i(q_1,\dots,q_N)$ and $U_{ij}(q_i,q_j)$ are restricted to rational functions of the coordinates. Let $L_s$ and $L_{ij}$ be the Liouvillean operators (vector fields) $$\begin{aligned}
L_s=\sum^s_{j=1}\Big(\frac{p_j}{m}\frac{\partial}{\partial q_j}-
\frac{\partial U_j}{\partial q}\frac{\partial}{\partial p_j}\Big)-
\sum_{1\leq i\leq j\leq s}L_{ij}\end{aligned}$$ $$\begin{aligned}
L_{ij}=\frac{\partial U_{ij}}{\partial q_i}\frac{\partial}{\partial p_i}+
\frac{\partial U_{ij}}{\partial q_j}\frac{\partial}{\partial p_j}\nonumber\end{aligned}$$ and $
F_N(x_1,\dots,x_N;t)
$ be the hierarchy of $N$-particle distribution function. satisfying the standard BBGKY–hierarchy ($V$ is the volume) \[1\]: $$\begin{aligned}
\frac{\partial F_s}{\partial t}+L_sF_s=
\frac{1}{V^s}\int{\mathrm{d}}\mu_{s+1}
\sum^s_{i=1}L_{i,s+1}F_{s+1}\end{aligned}$$ In most cases, one is interested in a representation of the form $
F_k(x_1,\dots,x_k;t)=\prod^k_{i=1}F_1(x_i;t)+G_k(x_1,\dots,x_k;t)
$ where $G_k$ are correlators. Additional reductions often lead to simplifications, the simplest one, $G_k=0$, corresponding to the Vlasov approximation. Such physically motivated ansatzes for $F_k$ formally replace the linear (in $F_k$) and pseudodifferential (in general case) infinite system (2) by a finite-dimensional but nonlinear system with polynomial nonlinearities (more exactly, multilinearities \[3\]). Our key point in the following consideration is the proper nonperturbative generalization of the perturbative multiscale approach of Bogolyubov.
MULTISCALE ANALYSIS
===================
The infinite hierarchy of distribution functions satisfying system (2) in the thermodynamical limit is: $$F=\{F_0,F_1(x_1;t),\dots,
F_N(x_1,\dots,x_N;t),\dots\}$$ where $F_p(x_1,\dots, x_p;t)\in H^p$, $H^0=R,\quad H^p=L^2(R^{6p})$ (or any different proper functional space), $F\in$ $H^\infty=H^0\oplus H^1\oplus\dots\oplus H^p\oplus\dots$ with the natural Fock space like norm (guaranteeing the positivity of the full measure): $$\begin{aligned}
(F,F)=F^2_0+\sum_{i}\int F^2_i(x_1,\dots,x_i;t)\prod^i_{\ell=1}\mu_\ell.\end{aligned}$$ First of all we consider $F=F(t)$ as a function of time only, $F\in L^2(R)$, via multiresolution decomposition which naturally and efficiently introduces the infinite sequence of the underlying hidden scales \[3\]. Because the affine group of translations and dilations generates multiresolution approach, this method resembles the action of a microscope. We have the contribution to the final result from each scale of resolution from the whole infinite scale of spaces. We consider a multiresolution decomposition of $L^2(R)$ \[3\] (of course, we may consider any different and proper for some particular case functional space) which is a sequence of increasing closed subspaces $V_j\in L^2(R)$ (subspaces for modes with fixed dilation value): $$...V_{-2}\subset V_{-1}\subset V_0\subset V_{1}\subset V_{2}\subset ...$$ The closed subspace $V_j (j\in {\bf Z})$ corresponds to the level $j$ of resolution, or to the scale j and satisfies the following properties: let $W_j$ be the orthonormal complement of $V_j$ with respect to $V_{j+1}$: $
V_{j+1}=V_j\bigoplus W_j.
$ Then we have the following decomposition: $$\begin{aligned}
\{F(t)\}=\bigoplus_{-\infty<j<\infty} W_j =
\overline{V_c\displaystyle\bigoplus^\infty_{j=0} W_j}\end{aligned}$$ in case when $V_c$ is the coarsest scale of resolution. The subgroup of translations generates a basis for the fixed scale number: $
{\rm span}_{k\in Z}\{2^{j/2}\Psi(2^jt-k)\}=W_j.
$ The whole basis is generated by action of the full affine group: $$\begin{aligned}
&&{\rm span}_{k\in Z, j\in Z}\{2^{j/2}\Psi(2^jt-k)\}=\\
&&{\rm span}_{k,j\in Z}\{\Psi_{j,k}\}
=\{F(t)\}\nonumber\end{aligned}$$ Let the sequence $\{V_j^t\}, V_j^t\subset L^2(R)$ correspond to multiresolution analysis on the time axis, $\{V_j^{x_i}\}$ correspond to multiresolution analysis for coordinate $x_i$, then $
V_j^{n+1}=V^{x_1}_j\otimes\dots\otimes V^{x_n}_j\otimes V^t_j
$ corresponds to the multiresolution analysis for the $n$-particle distribution function $F_n(x_1,\dots,x_n;t)$. E.g., for $n=2$: $V^2_0=\{f:f(x_1,x_2)=$ $
\sum_{k_1,k_2}a_{k_1,k_2}\phi^2(x_1-k_1,x_2-k_2),\
a_{k_1,k_2}\in\ell^2(Z^2)\},
$ where $
\phi^2(x_1,x_2)=\phi^1(x_1)\phi^2(x_2)=\phi^1\otimes\phi^2(x_1,x_2),
$ and $\phi^i(x_i)\equiv\phi(x_i)$ form a multiresolution basis corresponding to $\{V_j^{x_i}\}$. If $\{\phi^1(x_1-\ell)\},\ \ell\in Z$ form an orthonormal set, then $\phi^2(x_1-k_1, x_2-k_2)$ form an orthonormal basis for $V^2_0$. So, the action of the affine group generates multiresolution representation of $L^2(R^2)$. After introducing the detail spaces $W^2_j$, we have, e.g. $
V^2_1=V^2_0\oplus W^2_0.
$ Then the 3-component basis for $W^2_0$ is generated by the translations of three functions: $
\Psi^2_1=\phi^1(x_1)\otimes\Psi^2(x_2), \Psi^2_2=\Psi^1(x_1)\otimes\phi^2(x_2),
\Psi^2_3=\Psi^1(x_1)\otimes\Psi^2(x_2).
$ Also, we may use the rectangle lattice of scales and one-dimensional wavelet decomposition: $$f(x_1,x_2)=\sum_{i,\ell;j,k}\langle f,\Psi_{i,\ell}\otimes\Psi_{j,k}\rangle
\Psi_{j,\ell}\otimes\Psi_{j,k}(x_1,x_2)$$ where the basis functions $\Psi_{i,\ell}\otimes\Psi_{j,k}$ depend on two scales $2^{-i}$ and $2^{-j}$. We obtain our multiscale/multiresolution representations (formulae (11) below) via the variational wavelet approach for the following formal representation of the BBGKY system (9) (or its finite-dimensional nonlinear approximation for the $n$-particle distribution functions) with the corresponding obvious constraints on the distribution functions. Let $L$ be an arbitrary (non)linear differential/integral operator with matrix dimension $d$ (finite or infinite), which acts on some set of functions from $L^2(\Omega^{\otimes^n})$: $\quad\Psi\equiv\Psi(t,x_1,x_2,\dots)=\Big(\Psi^1(t,x_1,x_2,\dots), \dots$, $\Psi^d(t,x_1,x_2,\dots)\Big)$, $\quad x_i\in\Omega\subset{\bf R}^6$, $n$ is the number of particles: [ $$\begin{aligned}
&&L\Psi\equiv L(Q,t,x_i)\Psi(t,x_i)=0,\\
&&Q\equiv Q_{d_0,d_1,d_2,\dots}(t,x_1,x_2,\dots,\nonumber\\
&&\partial /\partial t,\partial /\partial x_1,
\partial /\partial x_2,\dots,\int \mu_k)=\nonumber\\
&&\sum_{i_0,i_1,i_2,\dots=1}^{d_0,d_1,d_2,\dots}
q_{i_0i_1i_2\dots}(t,x_1,x_2,\dots)\nonumber\\
&&\Big(\frac{\partial}{\partial t}\Big)^{i_0}\Big(\frac{\partial}{\partial x_1}\Big)^{i_1}
\Big(\frac{\partial}{\partial x_2}\Big)^{i_2}\dots\int\mu_k\nonumber \end{aligned}$$ ]{} Let us consider now the $N$ mode approximation for the solution as the following ansatz: $$\begin{aligned}
\Psi^N(t,x_1,x_2,\dots)=
\sum^N_{i_0,i_1,i_2,\dots=1}a_{i_0i_1i_2\dots}\\
A_{i_0}\otimes
B_{i_1}\otimes C_{i_2}\dots(t,x_1,x_2,\dots)\nonumber\end{aligned}$$ We shall determine the expansion coefficients from the following conditions (different related variational approaches are considered in \[2\]): $$\begin{aligned}
&&\ell^N_{k_0,k_1,k_2,\dots}\equiv \end{aligned}$$ $$\int(L\Psi^N)A_{k_0}(t)B_{k_1}(x_1)C_{k_2}(x_2){\mathrm{d}}t{\mathrm{d}}x_1{\mathrm{d}}x_2\dots=0$$ Thus, we have exactly $dN^n$ algebraical equations for $dN^n$ unknowns $a_{i_0,i_1,\dots}$. So, the solution is parametrized by the solutions of two sets of reduced algebraical problems, one is linear or nonlinear (depending on the structure of the operator $L$) and the rest are linear problems related to the computation of the coefficients of the algebraic equations (10). which can be found by using the compactly supported wavelet basis functions for the expansions (9). As a result the solution of the equations (2) has the following multiscale decomposition via nonlinear high-localized eigenmodes $$F(t,x_1,x_2,\dots)=
\sum_{(i,j)\in Z^2}a_{ij}U^i\otimes V^j(t,x_1,\dots)$$ $$\begin{aligned}
V^j(t)=
V_N^{j,slow}(t)+\sum_{l\geq N}V^j_l(\omega_lt), \ \omega_l\sim 2^l \end{aligned}$$ $$U^i(x_s)=
U_M^{i,slow}(x_s)+\sum_{m\geq M}U^i_m(k^{s}_mx_s), \ k^{s}_m\sim 2^m$$ which corresponds to the full multiresolution expansion in all underlying time/space scales. The formulae (11) give the expansion into a slow and fast oscillating parts. So, we may move from the coarse scales of resolution to the finest ones for obtaining more detailed information about the dynamical process. In this way one obtains contributions to the full solution from each scale of resolution or each time/space scale or from each nonlinear eigenmode. It should be noted that such representations give the best possible localization properties in the corresponding (phase)space/time coordinates. Formulae (11) do not use perturbation techniques or linearization procedures.
{width="55mm"}
{width="55mm"}
{width="55mm"}
Numerical calculations are based on compactly supported wavelets and related wavelet families \[3\] and on evaluation of the accuracy on the level $N$ of the corresponding cut-off of the full system (2) regarding norm (4): $
\|F^{N+1}-F^{N}\|\leq\varepsilon.
$ We believe that the appearance of nontrivial localized patterns (a)-(c) demonstrated on Fig.1–Fig.3 constructed by these methods is a general effect which is also present in the full BBGKY hierarchy, due to its complicated intrinsic multiscale dynamics and it depends on neither the cut-off level nor the phenomenological-like hypothesis on correlators. So, representations like (11) and the prediction of the existence of the (asymptotically) stable localized patterns/states (energy confinement states) in BBGKY-like systems are the main results of this paper.
We are very grateful to Prof. Kaneko (KEK) and Prof. Perret-Gallix (CNRS) for kind help and attention during ACAT03 Workshop at KEK.
[3]{}
R. Balescu, [*Equilibrium and Nonequilibrium Statistical Mechanics*]{}, Wiley, New York, 1975.
A.N. Fedorova and M.G. Zeitlin, [*Progress in Nonequilibrium Green’s Functions II*]{}, Ed. M. Bonitz, 481, World Scientific, Singapore, 2003; arXiv: physics/0212066; quant-ph/0306197
Y. Meyer, [*Wavelets and Operators*]{}, Cambridge Univ. Press, 1990.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Let $F\langle X, \ast \rangle$ be the free associative algebra with involution $\ast$ over a field $F$ of characteristic zero. We study the asymptotic behavior of the sequence of $\ast$-codimensions of the T-$\ast$-ideal $\Gamma_{M+1,L+1}^\ast$ of $F\langle X, \ast \rangle$ generated by the $\ast$-Capelli polynomials $Cap^\ast_{M+1} [Y,X]$ and $Cap^\ast_{L+1} [Z,X]$ alternanting on $M+1$ symmetric variables and $L+1$ skew variables, respectively.
It is well known that, if $F$ is an algebraic closed field of characteristic zero, every finite dimensional $\ast$-simple algebra is isomorphic to one of the following algebras:
- $(M_{k}(F),t)$ the algebra of $k \times k$ matrices with the transpose involution;
- $(M_{2m}(F),s)$ the algebra of $2m \times 2m$ matrices with the symplectic involution;
- $(M_{h}(F)\oplus M_{h}(F)^{op}, exc)$ the direct sum of the algebra of $h \times h$ matrices and the opposite algebra with the exchange involution.
We prove that the $\ast$-codimensions of a finite dimensional $\ast$-simple algebra are asymptotically equal to the $\ast$-codimensions of $\Gamma_{M+1,L+1}^\ast$, for some fixed natural numbers $M$ and $L$. In particular: $$c^{\ast}_n(\Gamma^{\ast}_{\frac{k(k+1)}{2} +1,\frac{k(k-1)}{2}
+1})\simeq c^{\ast}_n((M_k(F),t));$$ $$c^{\ast}_n(\Gamma^{\ast}_{m(2m-1)+1,m(2m+1)+1})\simeq
c^{\ast}_n((M_{2m}(F),s));$$ and $$c^{\ast}_n(\Gamma^{\ast}_{h^2+1,h^2+1})\simeq
c^{\ast}_n((M_{h}(F)\oplus M_{h}(F)^{op},exc)).$$
address:
- |
Dipartimento di Matematica e Informatica\
Universitá di Palermo\
via Archirafi, 34\
90123 Palermo,Italy
- |
Dipartimento di Ingegneria\
Università di Palermo\
Viale delle Scienze\
90128 Palermo, Italy
author:
- 'F. S. Benanti'
- 'A. Valenti'
title: Asymptotics for Capelli Polynomials with Involution
---
[^1]
Introduction
============
Let $(A, \ast)$ be an algebra with involution $\ast$ over a field $F$ of characteristic zero and let $F\langle X, \ast
\rangle=F\langle x_1,x_1^\ast, x_2, x_2^\ast, \ldots\rangle$ denote the free associative algebra with involution $\ast$ generated by the countable set of variables $\{x_1, x_1^\ast, x_2, x_2^\ast,
\ldots\}$ over $F$. Recall that an element $f(x_1,x_1^\ast, \cdots,
x_n,x_n^\ast)$ of $ F\langle X,\ast \rangle$ is a $\ast$-polynomial identity (or $\ast$-identity) for $A$ if $f(a_1,a_1^\ast, \cdots,
a_n,a_n^\ast)=0$, for all $a_1, \ldots , a_n \in A$. We denote by $Id^\ast (A)$ the set of all $\ast$-polynomial identities satisfied by $A$ which is a T-$\ast$-ideal of $F\langle X, \ast \rangle,$ i.e., an ideal invariant under all endomorphisms of $F\langle X,
\ast \rangle$ commuting with the involution of the free algebra. For $\Gamma=Id^\ast (A)$ we denote by $\mathrm{var}^\ast(\Gamma) =
\mathrm{var}^\ast(A)$ the variety of $\ast$-algebras having the elements of $\Gamma$ as $\ast$-identities.
It is well known that in characteristic zero $Id^\ast (A)$ is completely determinated by the multilinear $\ast$-polynomials it contains. To the T-$\ast$-ideal $\Gamma=Id^\ast (A)$ one can associates a numerical sequence called the sequence of $\ast$-codimensions $c_n^\ast(\Gamma)=c_n^\ast(A)$ which is the main tool for the quantitative investigation of the $\ast$-polynomial identities of $A$. Recall that $c_n^\ast(A)$, $n=1,2,\ldots$, is the dimension of the space of multilinear polynomial in $n$-th variables in the corresponding relatively free algebra with involution of countable rank. Thus, if we denote by $P_n^\ast$ the space of all multilinear polynomials of degree $n$ in $x_1,x_1^\ast,
\cdots, x_n,x_n^\ast$ then $$c_n^\ast(A)= \mathrm{dim} P_n^\ast(A)=
\mathrm{dim }\frac{P_n^\ast}{P_n^\ast\cap Id^\ast (A)}.$$
A celebrated theorem of Amitsur [@Am] states that if an algebra with involution satisfies a $\ast$-polynomial identity then it satisfies an ordinary polynomial identity. At the light of this result in [@GR] it was proved that, as in the ordinary case, if $A$ satisfies a non trivial $\ast$-polynomial identity then $c_n^\ast(A)$ is exponentially bounded, i.e. there exist constants $a$ and $b$ such that $c_n^\ast(A)\leq ab^n$, for all $n \geq 1$. Later (see [@BGZ]) an explicit exponential bound for $c_n^\ast(A)$ was exhibited and in [@GZ0] a characterization of finite dimensional algebras with involution whose sequence of $\ast$-codimensions is polynomial bounded was given. This result was extended to non-finite dimensional algebras (see [@GM1]) and $\ast$-varieties with almost polynomial growth were classified in [@GM] and [@MV]. The asymptotic behavior of the $\ast$-codimensions was determined in [@BGR] in case of matrices with involution.
Recently (see [@GPV]), for any algebra with involution, it was studied the exponential behavior of $c_n^\ast(A)$, and it was showed that the $\ast$-exponent of $A$ $$\textrm{exp}^{\ast}(A)=\displaystyle{\lim_{n\rightarrow \infty}\root n
\of{c_n^{\ast}(A)}}$$ exists and is a non negative integer. It should be mentioned that the existence of the $\ast$-exponent was proved in [@GZ00] for finite dimensional algebra with involution.
Now, if $f\in F\langle X, \ast \rangle$ we denote by $\langle f\rangle^\ast$ the T-$\ast$-ideal generated by $f$. Also for a set of polynomials $V\subset F\langle X, \ast \rangle$ we write $\langle V\rangle^\ast$ to indicate the T-$\ast$-ideal generated by $V$.
An interesting problem in the theory of PI-algebras with involution $\ast$ is to describe the T-$\ast$-ideals of $\ast$-polynomial identities of $\ast$-simple finite dimensional algebras. Recall that, if $F$ is an algebraically closed field of characteristic zero, then, up to isomorphisms, all finite dimensional $\ast$-simple are the following ones (see [@Ro], [@GZ]):
- $(M_k(F),t)$ the algebra of $ k \times k$ matrices with the transpose involution;
- $(M_{2m}(F),s)$ the algebra of $ 2m \times 2m
$ matrices with the symplectic involution;
- $(M_h(F)\oplus M_h(F)^{op},exc)$ the direct sum of the algebra of $h \times h$ matrices and the opposite algebra with the exchange involution.
The aim of this paper is to find a relation among the asymptotics of the $\ast$-codimensions of the finite dimensional $\ast$-simple algebras and the T-$\ast$-ideals generated by the $\ast$-Capelli polynomials $Cap^\ast_{M+1} [Y,X]$ and $Cap^\ast_{L+1} [Z,X]$ alternanting on $M+1$ symmetric variables and $L+1$ skew variables, respectively.
More precisely, if $(A, \ast)$ is any algebra with involution $\ast$, let $A^+=\{a
\in A \, | \, a^\ast=a\}$ and $A^-=\{a \in A \, | \, a^\ast=-a\}$ denote the subspaces of symmetric and skew elements of $A$, respectively. Since char$F$=0, we can regard the free associative algebra with involution $F\langle X, \ast \rangle$ as generated by symmetric and skew variables. In particular, for $i=1, 2, \ldots $, we let $y_i=x_i+x_i^\ast$ and $z_i=x_i-x_i^\ast$, then we write $X=Y\cup Z$ as the disjoint union of the set $Y$ of symmetric variables and the set $Z$ of skew variables and $F\langle X, \ast \rangle=F\langle Y\cup Z\rangle$. Hence a polynomial $f=f(y_1, \ldots , y_m,z_1, \ldots , z_n)\in F\langle Y\cup Z\rangle$ is a $\ast$-polynomial identity of $A$ if and only if $f(a_1, \ldots , a_m,b_1, \ldots , b_n)=0$ for all $a_i \in A^+$, $b_i \in A^-$.
Let us recall that, for any positive integer $m$, the $m$-th Capelli polynomial is the element of the free algebra $F\langle X\rangle$ defined as $$Cap_m (t_1, \ldots , t_m ; x_1, \ldots , x_{m-1})=$$ $$= \sum_{\sigma \in S_m} \mathrm{(sgn
\sigma)}t_{\sigma(1)}x_1t_{\sigma(2)}\cdots t_{\sigma(m-1)}
x_{m-1}t_{\sigma(m)}$$ where $S_m$ is the symmetric group on $\{1, \ldots , m\}$. In particular we denote by $$Cap^\ast_m [Y,X]=Cap_m (y_1, \ldots ,
y_m ; x_1, \ldots , x_{m-1})$$ and $$Cap^\ast_m [Z,X]=Cap_m (z_1, \ldots , z_m ; x_1, \ldots ,
x_{m-1})$$ the $m$-th $\ast$-Capelli polynomial in the alternating symmetric variables $y_1, \ldots , y_m$ and skew variables $z_1, \ldots , z_m$, respectively ($x_1, \ldots ,x_{m-1}$ are arbitrary variables).
Let $Cap_m^+$ denote the set of $2^{m-1}$ polynomials obtained from $Cap_m [Y,X]$ by deleting any subset of variables $x_i$ (by evaluating the variables $x_i$ to $1$ in all possible way). Similarly, we define by $Cap_m^-$ the set of $2^{m-1}$ polynomials obtained from $Cap_m [Z,X]$ by deleting any subset of variables $x_i$.
If $L$ and $M$ are two natural numbers, we denote by $\Gamma_{M+1,L+1}^\ast =\langle Cap_{M+1}^+, Cap_{L+1}^- \rangle$ the T-$\ast$-ideal generated by the polynomials $Cap_{M+1}^+,
Cap_{L+1}^-$. We also write $\mathcal{U}^\ast_{M+1,L+1}$ $=\mathrm{var}^\ast(\Gamma_{M+1,L+1})$ for the $\ast$-variety generated by $\Gamma_{M+1,L+1}^\ast$.
In this paper we study the asymptotic behavior of the sequence of $
\ast$-codimensions of $\mathcal{U}^\ast_{M+1,L+1}.$
Recall that two sequences $a_n$, $b_n$, $n=1,2,\ldots$, are asymptotically equal, $a_n \simeq b_n$, if $ \lim_{n\rightarrow
+\infty}\frac{a_n}{b_n}=1. $ In the ordinary case (no involution) (see [@GZ1]) it was proved the asymptotic equality between the codimensions of the Capelli polynomials $Cap_{k^2+1}$ and the codimensions of the matrix algebra $M_k(F).$ In [@Be] these result was extended to finite dimensional simple superalgebras proving that the graded codimensions of the $T_2$-ideal generated by the graded Capelli polynomials $\Gamma_{M+1,L+1},$ for some fixed $M$, $L$, are asymptotically equal to the graded codimensions of a simple finite dimensional superalgebra. The link between the asymptotic of the codimensions of the Amitsur’s Capelli-type polynomials and the verbally prime algebras was studied in [@BS].
Here we characterize the T-$\ast$-ideal of $\ast$-identities of any $\ast$-simple finite dimensional algebra showing that
$$\Gamma^{\ast}_{\frac{k(k+1)}{2} +1,\frac{k(k-1)}{2}
+1}=Id^{\ast}((M_k(F),t)\oplus D');$$
$$\Gamma^{\ast}_{m(2m-1)+1,m(2m+1)+1}=Id^{\ast}((M_{2m}(F),s)\oplus
D'');$$
$$\Gamma^{\ast}_{h^2+1,h^2+1}=Id^{\ast}((M_{h}(F)\oplus M_{h}(F)^{op},exc)\oplus
D''')$$
where $D'$, $D''$ and $D'''$ are finite dimensional $\ast$-algebra with exp$^{\ast}(D')<k^2$, exp$^{\ast}(D'')$ $<(2m)^2$ and exp$^{\ast}(D''')<2h^2$. It follows that asymptotically
$$c^{\ast}_n(\Gamma^{\ast}_{\frac{k(k+1)}{2} +1,\frac{k(k-1)}{2}
+1})\simeq c^{\ast}_n((M_k(F),t));$$
$$c^{^\ast}_n(\Gamma^{\ast}_{m(2m-1)+1,m(2m+1)+1})\simeq
c^{\ast}_n((M_{2m}(F),s));$$
$$c^{^\ast}_n(\Gamma^{\ast}_{h^2+1,h^2+1})\simeq
c^{\ast}_n((M_{h}(F)\oplus M_{h}(F)^{op},exc)).$$
Preliminaries
=============
Let $F$ be a field of characteristic zero and let $G$ be the Grassmann algebra over $F$ generated by the elements $e_1,
e_2, \ldots$ subject to the following condition $e_ie_j = -e_je_i,$ for all $i, j \ge 1$. Recall that $G$ has a natural $Z_2$-grading $G= G_0 \oplus G_1$ where $G_0$ (resp. $G_1$) is the span of the monomials in the $e_i's$ of even length (resp. odd length). If $B=
B_0 \oplus B_1$ is a superalgebra, then the Grassmann envelope of $B$ is defined as $$G(B) = ( G_0 \otimes B_0)\oplus(G_1 \otimes
B_1).$$ The relevance of $G(A)$ relies in a result of Kemer ([@Ke Theorem 2.3]) stating that if $B$ is any PI-algebra, then its T-ideal of polynomial identities coincides with the T-ideal of identities of the Grassmann envelope of a suitable finite dimensional superalgebra. This result has been extended to algebras with involution in [@alj-gia-kar] and the following result holds
If $A$ is a $PI$-algebra with involution over a field $F$ of characteristic zero, then there exists a finite dimensional superalgebra with superinvolution $B$ such that $Id^*(A) = Id^*(G(B))$.
Recall that a superinvolution $*$ of $B$ is a linear map of $B$ of order two such that $(ab)^*= (-1)^{\mid a\mid\mid b\mid} b^*a^*,$ for any homogeneous elements $a,b \in B,$ where $\mid a\mid$ denotes the homogeneous degree of $a$. It is well known that in this case $B^*_0\subseteq B_0, B^*_1\subseteq B_1$ and we decompose $B=
B_0^+\bigoplus B_0^- \bigoplus B_1^+ \bigoplus B_1^-.$
We can define a superinvolution $*$ on $G$ by requiring that $e_i^*
= -e_i,$ for any $i \ge 1.$ Then it is easily checked that $G_0 =
G^+$ and $ G_1 = G^-.$ Now, if $B$ is a superalgebra one can perform its Grassmann envelope $G(B)$ and in [@alj-gia-kar] it was shown that if $B$ has a superinvolution $*$ we can regard $G(B)$ as an algebra with involution by setting $(g \otimes a)^* = g^*
\otimes a^*,$ for homogeneous elements $g \in G, a \in B.$
By making use of the previous theorem, in [@GPV] it was proved the existence of the $\ast$-exponent of a $PI$-algebra with involution $A$ and also an explicit way of computing exp$^\ast(A)$ was given. More precisely if $B$ is a finite dimensional algebra with superinvolution over an algebraic closed field of characteristic zero, then by [@GIL] we write $B= \bar{B}+J$ where $\bar{B}$ is a maximal semisimple superalgebra with induced superinvolution and $J =J(B)= J^*.$ Also we can write $$\bar{B} = B_1 \oplus \cdots \oplus B_k$$ where $B_1, \cdots, B_k$ are simple superalgebras with induced superinvolution. We say that a subalgebra $B_{i_1} \oplus \cdots \oplus B_{i_t},$ where $B_{i_1},\ldots, B_{i_t}$ are distinct simple components, is admissible if for some permutation $(l_1, \ldots, l_t)$ of $(i_1, \ldots, i_t)$ we have that $B_{l_1}JB_{l_2}J \cdots JB_{l_t} \ne 0.$ Moreover if $B_{i_1} \oplus \cdots \oplus B_{i_t}$ is an admissible subalgebra of $B$ then $B' = B_{i_1} \oplus \cdots \oplus B_{i_t}+J$ is called a reduced algebra. In [@GPV] it was proved that exp$^\ast(A)$ = exp$^\ast(G(B))= d$ where $d$ is the maximal dimension of an admissible subalgebra of $B.$
It follows immediately that
If $A$ is a $*$-simple algebra then $exp^\ast(A) = \textrm{dim}_F A$.
We next prove that the reduced algebras are basic elements of any $*$-variety. We start with the following
\[codimensioni\] Let $A$ and $B$ be algebras with involution satisfying a $\ast$-polynomial identity. Then
$$c_n^{\ast}(A), c_n^{\ast}(B) \leq c_n^{\ast}(A\oplus B) \leq
c_n^{\ast}(A)+c_n^{\ast}(B).$$
Hence the $\mathrm{exp}^{\ast}(A\oplus
B)=\mathrm{max}\{\mathrm{exp}^{\ast}(A), \mathrm{exp}^{\ast}(B)\}.$
[**Proof.**]{} It follows easily from the proof of the Lemma 1 in [@GZ1].
If $ \mathcal{V}= \mathrm{var}^{\ast}(A)$ is the variety of $\ast$-algebras generated by $A$ we write $Id^\ast (\mathcal{V})= Id^{\ast}(A)$, $c_n^{\ast}(\mathcal{V})=c_n^{\ast}(A)$ and $\textrm{exp}^{\ast}(\mathcal{V}) = \textrm{exp}^{\ast}(A)$.
We have the following
\[decomposizione\] Let $\mathcal{V}$ be a proper variety of $\ast$-algebras. Then there exists a finite number of reduced superalgebras with superinvolution $B_1, \ldots, B_t$ and a finite dimensional superalgebra with superinvolution $D$ such that
$$\mathcal{V}=\mathrm{var}(G(B_1)\oplus \cdots \oplus G(B_t) \oplus G(D))$$
with $\mathrm{exp}^{\ast}(\mathcal{V})=\mathrm{exp}^{\ast}(G(B_1))=\cdots=\mathrm{exp}^{\ast}G((B_t))$ and $\mathrm{exp}^{\ast}(G(D))<\mathrm{exp}^{\ast}(\mathcal{V}).$
[**Proof.**]{} The proof follows closely the proofs given in [@GZ1 Theorem 1] and in [@Be Theorem 3]. Let $A$ be a $\ast$-PI-algebra such that $\mathcal{V}=\mathrm{var}^\ast(A)$. By Theorem 1, there exists a finite dimensional superalgebra with superinvolution $B$ such that $Id^{\ast}(A) = Id^{\ast}(G(B))$. Also, by [@GIL Theorem 4.1], we may assume that
$$B=\bar{B}_1\oplus\cdots\oplus \bar{B}_s+J(B),$$ where $\bar{B}_i$ are simple $\ast$-superalgebras and $J^\ast=J$ is the Jacobson radical of $B$. Let $\textrm{exp}^{\ast}(A)=d$. Then, since $d$ is the maximal dimension of an admissible subalgebra of $A$ there exist distinct $\ast$-simple superalgebras $\bar{B}_{j_1}, \ldots \bar{B}_{j_k} $ such that
$$\bar{B}_{j_1}J \cdots J\bar{B}_{j_k}\neq 0 \,\,\,\,\,\,\,
\textrm{and} \,\,\,\,\,\,\, \textrm{dim}_F(\bar{B}_{j_1}\oplus
\cdots \oplus \bar{B}_{j_k})=d.$$
Let $\Gamma_1, \ldots , \Gamma_t$ be all possible subset of $\{1, \ldots , s\}$ such that, if $\Gamma_j=\{j_1, \ldots
,j_k\}$, then $dim_F(\bar{B}_{j_1}\oplus \cdots \oplus
\bar{B}_{j_k})=d$ and $\bar{B}_{\sigma(j_1)}J \cdots
J\bar{B}_{\sigma(j_k)}\neq 0$ for some permutation $\sigma \in S_k$. For any such $\Gamma_j$, $j=1, \ldots t,$ then we put $B_j=\bar{B}_{j_1}\oplus \cdots \oplus \bar{B}_{j_k} + J$. It follows, by the characterization of the $\ast$-exponent, that $$\mathrm{exp}^{\ast}(G(B_1))=\cdots
=\mathrm{exp}^{\ast}(G(B_t))=d=\mathrm{exp}^{\ast}(G(B)).$$ Let $D=D_1 \oplus \cdots \oplus D_p$, where $D_1, \ldots ,
D_p$ are all subsuperalgebras of $B$ with superinvolution of the type $\bar{B}_{i_1}\oplus \cdots \oplus \bar{B}_{i_r} + J$, with $1\leq i_1 <\cdots < i_r \leq s$ and $\textrm{dim}_F(\bar{B}_{i_1}\oplus \cdots \oplus \bar{B}_{i_r})<d
$. Then, by Lemma \[codimensioni\], we have $\mathrm{exp}^{\ast}(G(D))< \mathrm{exp}^{\ast}(G(B)).$
Now, we want to prove that $\mathrm{var}^\ast(G(B_1) \oplus \cdots \oplus G(B_t)\oplus
G(D))=\mathrm{var}^\ast(G(B))$. Since $G(D), G(B_i)\in
\mathrm{var}^\ast(A)$, $\forall i=1, \ldots , t,$ it follows that
$$\mathrm{var}^\ast(G(B_1) \oplus \cdots \oplus G(B_t)\oplus
G(D))\subseteq \mathrm{var}^\ast(G(B)).$$
Now, let $f=f(y_1^+, \ldots,y_n^+,y_1^-, \ldots,y_m^-,z_1^+, \ldots,
z_p^+,z_1^-, \ldots, z_q^-)$ be a multilinear polynomial such that $f\not\in Id^{\ast}(G(B))$. We shall prove that $f \not\in
Id^{\ast}(G(B_1) \oplus \cdots \oplus G(B_t)\oplus G(D))$. Since $f\not\in Id^{\ast}(G(B))$, there exist $$a_{1,0}^+\otimes g_{1,0},
\ldots , a_{n,0}^+\otimes g_{n,0} \in G(B)_0^+=B_0^+\otimes G_0,$$ $$a_{1,0}^-\otimes h_{1,0}, \ldots , a_{m,0}^-\otimes h_{m,0} \in
G(B)_0^-=B_0^-\otimes G_0,$$ $$b_{1,1}^-\otimes g_{1,1}, \ldots ,
b_{p,1}^-\otimes g_{p,1} \in G(B)_1^+=B_1^-\otimes G_1,$$ and $$b_{1,1}^+\otimes h_{1,1}, \ldots , b_{q,1}^+\otimes h_{q,1} \in
G(B)_1^-=B_1^+\otimes G_1$$ such that $$f(a_{1,0}^+\otimes g_{1,0}, \ldots , a_{n,0}^+\otimes g_{n,0},
a_{1,0}^-\otimes h_{1,0}, \ldots , a_{m,0}^-\otimes h_{m,0},$$ $$b_{1,1}^-\otimes g_{1,1}, \ldots , b_{p,1}^-\otimes
g_{p,1},b_{1,1}^+\otimes h_{1,1}, \ldots , b_{q,1}^+\otimes
h_{q,1})\neq 0.$$ It follows that $$0 \neq f(a_{1,0}^+\otimes g_{1,0}, \ldots , a_{n,0}^+\otimes g_{n,0},
a_{1,0}^-\otimes h_{1,0}, \ldots , a_{m,0}^-\otimes h_{m,0},$$ $$b_{1,1}^-\otimes g_{1,1}, \ldots , b_{p,1}^-\otimes
g_{p,1},b_{1,1}^+\otimes h_{1,1}, \ldots , b_{q,1}^+\otimes
h_{q,1})=$$ $$\tilde{f}(a_{1,0}^+,\ldots , a_{n,0}^+,a_{1,0}^-, \ldots ,
a_{m,0}^-,b_{1,1}^-, \ldots ,b_{p,1}^-,b_{1,1}^+, \ldots ,
b_{q,1}^+)\otimes$$ $$g_{1,0} \cdots g_{n,0}h_{1,0} \cdots
h_{m,0}g_{1,1} \cdots g_{p,1}h_{1,1} \cdots h_{q,1}$$
where $\tilde{f}$ is the multilinear polynomial introduced in [@GPV Lemma 1]. Clearly $\tilde{f}\neq 0$. From the linearity of $\tilde{f}$ we can assume that $a_{i,0}^+$, $b_{j,1}^-$, $a_{i,0}^-$ and $b_{j,1}^+$ $\in B_1 \cup \cdots \cup
B_s \cup J$. Since $B_iB_j=0$ for $i \neq j$, from the property of the $\ast$-exponent described above, we have that $$a_{1,0}^+, \ldots , a_{n,0}^+,a_{1,0}^-, \ldots ,
a_{m,0}^-,b_{1,1}^-, \ldots ,b_{p,1}^-,b_{1,1}^+, \ldots , b_{q,1}^+
\in B_{j_1}\oplus \cdots \oplus B_{j_k} + J$$ for some $B_{j_1}, \ldots , B_{j_k}$ such that dim$_F(B_{j_1}\oplus
\cdots \oplus B_{j_k})\leq d.$
Thus $f$ is not an identity for one of the algebras $G(B_1), \ldots , G(B_t), G(D)$. Hence $f \not\in
Id^{\ast}(G(B_1) \oplus \cdots \oplus G(B_t)\oplus G(D))$. In conclusion $$\mathrm{var}^\ast(G(B))\subseteq \mathrm{var}^\ast (G(B_1)\oplus
\cdots \oplus G(B_t) \oplus G(D))$$ and the proof is complete.
An application of Theorem \[decomposizione\] is given in terms of $\ast$-codimensions.
\[starcodimensioni\] Let $\mathcal{V}=\mathrm{var}^\ast(A)$ be a proper variety of $\ast$-algebras. Then there exists a finite number of reduced superalgebras with superinvolution $B_1, \ldots, B_t$ and a finite dimensional superalgebra with superinvolution $D$ such that $$c_n^{\ast}(A)\simeq c_n^{\ast} (G(B_1)\oplus \cdots \oplus G(B_t)).$$
[**Proof.**]{} By Theorem \[decomposizione\], there is a finite number of reduced superalgebras with superinvolution $B_1,
\ldots, B_t$ such that $$\mathcal{V}=\mathrm{var}^\ast(A)= \mathrm{var}^\ast(G(B_1)\oplus
\cdots \oplus G(B_t)\oplus G(D))$$
with $\textrm{exp}^{\ast}(A)=\textrm{exp}^{\ast}(G(B_1))=\cdots=\textrm{exp}^{\ast}(G(B_t))$ and $\textrm{exp}^{\ast}(G(D))< \textrm{exp}^{\ast}(A)$. Then, by Lemma \[codimensioni\] $$c_n^{\ast}(G(B_1)\oplus \cdots \oplus G(B_t)) \leq
c_n^{\ast}(G(B_1)\oplus \cdots \oplus G(B_t)\oplus G(D))
\leq$$ $$c_n^{\ast}(G(B_1)\oplus \cdots \oplus G(B_t))+c_n^{\ast}(G(D)).$$
Recalling that $\textrm{exp}^{\ast}(G(D))<\textrm{exp}^{\ast}(G(B_1))=\mathrm{exp}^{\ast}(G(B_1)\oplus
\cdots \oplus G(B_t))$ we have that
$$c_n^{\ast}(A)\simeq c_n^{\ast} (G(B_1)\oplus \cdots \oplus G(B_t))$$
and the proof of the corollary is complete.
If $A$ is a finite dimensional $\ast$-algebra we obtain a simplified form of the previous theorem and corollary. Let us recall that an algebra with involution can be regarded as a superalgebra with superinvolution with trivial grading. We have the following
\[decomposizionefinita\] Let $A$ be a finite dimensional $\ast$-algebra. Then there exists a finite number of reduced $\ast$-algebras $B_1, \ldots, B_t$ and a finite dimensional $\ast$-algebra $D$ such that $$\mathrm{var}^\ast(A)=\mathrm{var}^\ast(B_1\oplus \cdots \oplus
B_t\oplus D)$$ $$c_n^{\ast}(A)\simeq c_n^{\ast} (B_1\oplus \cdots \oplus B_t)$$ and $$\mathrm{exp}^\ast(A)= \mathrm{exp}^\ast(B_1)=\cdots =
\mathrm{exp}^\ast(B_t),\,\, \mathrm{exp}^\ast(D)<\mathrm{exp}^\ast(A).$$
The following results give us a characterization of the $\ast$-varieties satisfying a Capelli identity. Let’s start with the
\[capelliidentity\] Let $M$ and $L$ be two natural numbers. If $A$ is an algebra with involution satisfying the $\ast$-Capelli polynomials $Cap^\ast_M
[Y,X]$ and $Cap^\ast_L [Z,X]$, then $A$ satisfies the Capelli identity $Cap_{M+L} (x_1, \ldots , x_{M+L} ; \bar{x}_1, \ldots ,
\bar{x}_{M+L-1})$.
[**Proof.**]{} To obtain the thesis it is sufficient to observe that $$Cap_{M+L} (x_1, \ldots , x_{M+L} ; \bar{x}_1, \ldots ,
\bar{x}_{M+L-1})=$$ $$Cap_{M+L} (\frac{x_{1}+x_{1}^\ast}{2}+\frac{x_{1}-x_{1}^\ast}{2},
\ldots ,
\frac{x_{M+L}+x_{M+L}^\ast}{2}+\frac{x_{M+L}-x_{M+L}^\ast}{2} ;
\bar{x}_1, \ldots , \bar{x}_{M+L-1})$$
is a linear combinations of $\ast$-Capelli polynomials alternating either in $m\geq M$ symmetric variables or in $l\geq L$ skew variables.
The proof of the next result follows closely the proof given in [@GZ Theorem 11.4.3]
\[finitelygenerated\] Let $\mathcal{V}$ be a variety of $\ast$-algebras. If $\mathcal{V}$ satisfies the Capelli identity of some rank then $\mathcal{V}=\mathrm{var}^\ast(A)$, for some finitely generated $\ast$-algebra $A$.
Let $M$, $L$ be two natural numbers. Let $A=A^{+}\oplus A^{-}$ be a generating $\ast$-algebra of $\mathcal{U}^{\ast}_{M+1,L+1}$. By remark \[capelliidentity\], $A$ satisfies a Capelli identity. Hence by the previous theorem, we may assume that $A$ is a finitely generated $\ast$-algebra. Moreover by [@Sv Theorem 1] we may consider $A$ as a finite-dimensional $\ast$-algebra. Since any polynomial alternating on $M+1$ symmetric variables vanishes in $A$ (see [@GZ Proposition 1.5.5]), we get that $\mathrm{dim}\,
A^{+}\leq M$. Similarly we get that $\mathrm{dim}\, A^{-}\leq L$ and $\mathrm{exp}^\ast(A)\leq \mathrm{dim}\, A \leq M+L$. Thus we have the following
\[esponente\] $\mathrm{exp}^\ast(\mathcal{U}^{\ast}_{M+1,L+1})\leq M+L$.
The algebra $UT^\ast(A_1, \ldots, A_n)$
========================================
In this section we recall the construction of the $\ast$-algebra $UT^\ast(A_1, \ldots, A_n)$ given in Section 2 of [@DvLs]. Let $A_1, \ldots, A_n$ be a $n$-tuple of finite dimensional $\ast$-simple algebras, then $A_i=(M_{d_i},\mu_i)$, where $\mu_i$ is the transpose or the symplectic involution, or $A_i=(M_{d_i}\oplus M_{d_i}^{op},exc)$, where $exc$ is the exchange involution.
Let $\gamma_d$ be the orthogonal involution defined on the matrix algebra $M_d(F)$ by putting, for all $a \in M_d(F)$, $$a^{\gamma_d}=g^{-1}a^tg=ga^tg,$$ where $$g=\left(
\begin{array}{ccccc}
0 & & \ldots & & 1 \\
& & &\cdot & \\
& &\cdot & & \\
&\cdot & & & \\
1 & &\ldots & & 0 \\
\end{array}
\right)$$ and $a^t$ is the transposed of the matrix $a$. $\gamma_d$ acts on matrix units $e_{pq}$ of $M_d$ by sending it to $e^{\gamma_d}_{pq} =
e_{d-q+1,d-p+1}$ (it is the reflection along the secondary diagonal).
If $d=\sum_{i=1}^n \textrm{dim}_FA_i$, then we have an embedding of $\ast$-algebras
$$\Delta : \bigoplus_{i=1}^n A_i \rightarrow
(M_{2d}(F), \gamma_{2d})$$ defined by
$$(a_1, \ldots, a_n)\rightarrow \left(
\begin{array}{cccccc}
\bar{a}_1 & & & & & \\
& \ddots & & & & \\
& & \bar{a}_n & & & \\
& & & \bar{b}_n & & \\
& & & & \ddots & \\
& & & & & \bar{b}_1 \\
\end{array}
\right)$$
where, if $a_i \in A_i=(M_{d_i},\mu_i)$, then $\bar{a}_i=a_i$ and $\bar{b}_i=a_i^{\mu_i\gamma_{d_i}}$, and if $a_i=(\tilde{a}_i, \tilde{b}_i)\in A_i=(M_{d_i}\oplus
M_{d_i}^{op},exc)$, then $\bar{a}_i=\tilde{a}_i$ and $\bar{b}_i=\tilde{b}_i$.
Let denote by $D=D(A_1, \ldots, A_n)\subseteq M_{2d}(F)$ the $\ast$-algebra image of $\bigoplus_{i=1}^n A_i$ by $\Delta$ and let $U$ be the subspace of $M_{2d}(F)$ so defined:
$$\left(
\begin{array}{cccccccc}
0 & U_{12} & \cdots & U_{1t} & & & & \\
& \ddots & \ddots & \vdots & & & & \\
& & 0 & U_{t-1t} & & & & \\
& & & 0 & & & & \\
& & & & 0 & U_{tt-1} & \cdots & U_{t1} \\
& & & & & \ddots & \ddots & \vdots \\
& & & & & & 0 & U_{21} \\
& & & & & & & 0 \\
\end{array}
\right)$$
where, for $1 \leq i,j \leq n$, $i\neq j$, $U_{ij}$ denote the vector space of the rectangular matrices of dimensions $d_i
\times d_j.$ Let define
$$UT^\ast (A_1, \ldots, A_n)=D\oplus U \subseteq M_{2d}(F)$$
(see section 2 of [@DvLs]).
It is easy to show that $UT^\ast (A_1, \ldots, A_n)$ is a subalgebra with involution of $(M_{2d}(F),\gamma_{2d})$ in which the algebras $A_i$ are embedded as $\ast$-algebras and whose $\ast$-exponent is given by $$\mathrm{exp}^\ast(UT^\ast (A_1, \ldots, A_n))=\sum_{i=1}^n \mathrm{dim}_FA_i.$$
In [@DvS1] and [@DvS2] the link between the degrees of $\ast$-Capelli polynomials and the $\ast$-polynomial identities of $UT^\ast (A_1, \ldots, A_n)$ was investigated.
If we set $d^+:=\sum_{i=1}^n \mathrm{dim}_FA_i^+$ and $d^-:=\sum_{i=1}^n
\mathrm{dim}_FA_i^-, $ then the following result applies (see [@DvS])
\[capelliidentita\] Let $R=UT^\ast (A_1, \ldots, A_n).$ Then $Cap^\ast_M [Y,X]$ and $Cap^\ast_L [Z,X]$ are in $Id^\ast(R)$ if and only if $M\geq d^++n$ and $L\geq d^-+n$.
Asymptotics for $\mathcal{U}^\ast_{\frac{k(k+1)}{2} +1,\frac{k(k-1)}{2} +1}$ and $(M_k(F),t)$
=============================================================================================
Let $A= \bar{A} \oplus J$ where $\bar{A}$ is a $\ast$-simple finite dimensional algebra and $J=J(A)$ is its Jacobson radical. It is well known that the Jacobson radical $J$ is a $\ast$-ideal of $A$.
We start with the following key lemmas that hold for any $\ast$-simple finite dimensional algebra.
\[radicaltrasposta\] Let $A= \bar{A} \oplus J$ where $\bar{A}$ is a $\ast$-simple finite dimensional algebra and $J=J(A)$ is its Jacobson radical. Then $J$ can be decomposed into the direct sum of four $\bar{A}$-bimodules $$J=J_{00}\oplus J_{01} \oplus J_{10} \oplus J_{11}$$ where, for $p,q \in \{0,1\}$, $J_{pq}$ is a left faithful module or a $0$-left module according to $p=1$, or $p=0$, respectively. Similarly, $J_{pq}$ is a right faithful module or a $0$-right module according to $q=1$ or $q=0$, respectively. Moreover, for $p,q,i,l \in
\{0,1\}$, $J_{pq}J_{ql}\subseteq J_{pl}$, $J_{pq}J_{il}=0$ for $q\neq i$ and there exists a finite dimensional nilpotent $\ast$-algebra $N$ such that $J_{11}\cong \bar{A}\otimes_F N$ (isomorphism of $\bar{A}$-bimodules and of $\ast$-algebras).
[**Proof.**]{} It follows from the proof of Lemma 2 in [@GZ1].
Notice that $J_{00}$ and $J_{11}$ are stable under the involution whereas $J_{01}^\ast=J_{10}$.
\[identitytrasposta\] Let $\bar{A}$ be a $\ast$-simple finite dimensional algebra. Let $M
=\textrm{dim}_F \bar{A}^+$ and $L=\textrm{dim}_F \bar{A}^-$. Then $\bar{A}$ does not satisfy $Cap^\ast_{M}[Y,X]$ and $Cap^\ast_{L}[Z,X]$.
[**Proof.**]{} The result follows immediately from [@DvS Lemma 3.1].
From now on we assume that $A=M_{k}(F)+ J$, where $J=J(A)$ is the Jacobson radical of the finite dimensional $\ast$-algebra $A$ and $(M_k(F),t)$ is the $\ast$-algebra of matrices with transpose involution.
\[j10trasposta\] Let $M=k(k+1)/2$ and $L=k(k-1)/2$ with $k \in \mathbb{N}$, $k>0$. If $\Gamma^\ast_{M+1,L+1}\subseteq Id^{\ast}(A)$, then $J_{10}=J_{01}=(0)$.
[**Proof.**]{} By Lemma \[identitytrasposta\], $M_{k}(F)$ does not satisfy the $\ast$-Capelli polynomial $Cap^\ast_{M}[Y,X]$. Then, there exist elements $a_1^+, \ldots ,
a_M^+ \in M_{k}(F)^{+}$ and $b_1, \ldots , b_{M-1} \in M_{k}(F)$ such that $$Cap^\ast_{M}(a_1^{+}, \ldots , a_M^{+};b_1, \ldots ,
b_{M-1})=e_{1,k},$$ where the $e_{i,j}$’s are the usual matrix units. Let $d \in
J_{01},$ then $d^\ast \in J_{10}$ and $d+d^\ast \in (J_{01} \oplus
J_{10})^{+}$. Since $\Gamma^\ast_{M+1, L+1}\subseteq Id^{\ast}(A)$ we have $$0=Cap^\ast_{M+1}(a_1^{+}, \ldots , a_M^{+},d+d^\ast;b_1, \ldots ,
b_{M-1},e_{k,k})=de_{1,k}\pm e_{1,k}d^\ast.$$ Hence $de_{1,k}\pm e_{1,k}d^\ast=0$ and, so, $de_{1,k} =\mp
e_{1,k}d^\ast \in J_{01}\cap J_{10}=(0)$. Then $d=0$, for all $d \in
J_{01}$. Thus $J_{01}=(0)$ and $J_{10}=(0)$.
\[Ntrasposta\] Let $M=k(k+1)/2$ and $L=k(k-1)/2$ with $k \in \mathbb{N}$, $k>0$. Let $J_{11} \cong M_{k}(F) \otimes_F N$, as in Lemma \[radicaltrasposta\]. If $\Gamma_{M+1,L+1}\subseteq Id^{\ast}(A)$, then $N$ is commutative.
[**Proof.**]{} Let $N$ be the finite dimensional nilpotent $\ast$-algebra of the Lemma \[radicaltrasposta\]. Write $N=N^+\oplus N^-$, where $N^+$ and $N^-$ denote the subspaces of symmetric and skew elements of $N$ respectively. Let $e_1^+,\ldots
, e_M^+$ be an ordered basis of $M_{k}(F)^{+}$ consisting of symmetric matrices $e_h^+ \in \{e_{i,i}\, | \, i=1,\ldots k\}\cup \{
e_{i,j}+ e_{j,i}\, | \, i < j , \, i,j= 1,\ldots k\} $ such that $e_1^+= e_{1,1}$ and let $a_0,a_1, \ldots , a_M \in M_{k}(F)$ be such that
$$a_0e_1^+a_1\cdots a_{M-1} e_M^+ a_M=e_{1,1}$$ and $$a_0e_{\sigma (1)}^+a_1\cdots a_{M-1} e_{\sigma (M)}^+ a_M=0$$ for any $\sigma \in S_M$, $\sigma \neq id$.
Consider $d^{+}_1, d^{+}_2 \in N^{+}$ and set $c^{+}_1=
e_{1,1} d^{+}_1$ and $c^{+}_2= e_{1,1} d^{+}_2$. Notice that, since $N$ commutes with $M_{k}(F)$, $c^{+}_1, c^{+}_2\in A^{+}$ and $$Cap^\ast_{M+2}(c^{+}_1,e_1^+,\ldots ,e_M^+, c^{+}_2; a_0, \ldots , a_M )=$$ $$c^{+}_1e_{1,1}c^{+}_2-c^{+}_2e_{1,1}c^{+}_1-e_{1,1}c^{+}_1c^{+}_2+$$ $$c^{+}_2c^{+}_1e_{1,1} +e_{1,1}c^{+}_2c^{+}_1-c^{+}_1c^{+}_2e_{1,1}=$$ $$[c^{+}_2,c^{+}_1]e_{1,1}=[d^{+}_2,
d^{+}_1]e_{1,1}.$$ Since $Cap^\ast_{M+1}[Y;X]\subseteq Id^{\ast}(A)$ we have $[d^{+}_1,
d^{+}_2]=0$. Thus $d^{+}_1d^{+}_2=d^{+}_2d^{+}_1$, for all $d^{+}_1,
d^{+}_2 \in N^{+}$.
Now, let $e_1^-,\ldots , e_L^-$ be an ordered basis of $M_{k}(F)^{-}$ consisting of skew matrices $e_h^- \in \{ e_{i,j}-
e_{j,i}\, | \, i < j , \, i,j= 1,\ldots k\} $ such that $e_1^-=
e_{1,2}-e_{2,1}$. We consider $b_0, \ldots , b_L \in M_{k}(F)$ such that $b_0=e_{1,1}$, $b_L=e_{k,k},$ $$b_0e_1^-b_1 \cdots b_{L-1}e_L^-b_L=e_{1,k}$$ and $$b_0e_{\tau (1)}^-b_1\cdots b_{L-1} e_{\tau (L)}^- b_L=0$$ for all $\tau \in S_L$, $\tau \neq id$.
Let $d^{-}_1, d^{-}_2 \in
N^{-}$ and put $c^{-}_1= (e_{1,2}+e_{2,1})d_1^-$ and $c^{-}_2=
(e_{1,2}+e_{2,1})d_2^-$. Since $N$ commutes with $M_{k}(F)$ then $c^{-}_1, c^{-}_2\in A^{-}$. As above we compute $$Cap^\ast_{L+2}(c^{-}_1,e_1^-,\ldots ,e_L^-, c^{-}_2;b_0, \ldots , b_L )=$$ $$[c^{-}_1, c^{-}_2]e_{1,k}=(e_{1,1}+e_{2,2})e_{1,k}[d^{-}_1,
d^{-}_2]=e_{1,k}[d^{-}_1, d^{-}_2].$$ Since $Cap^\ast_{L+1}[Z;X]\subseteq Id^{\ast}(A)$ we get that $[d^{-}_1,d^{-}_2]=0$, then $d^{-}_1d^{-}_2=d^{-}_2d^{-}_1$, for all $d^{-}_1, d^{-}_2 \in N^{-}$.
Next we show that $N^{+}$ commutes with $N^{-}$. Take $e_1^+,\ldots
, e_M^+$ an ordered basis of $M_{k}(F)^{+}$ such that $e_1^+= e_{1,1}$ and let $a_1, \ldots , a_{M} \in
M_{k}(F)$ be such that
$$a_1e_1^+a_2\cdots a_{M} e_{M}^+=e_{1,k}$$ and $$a_1e_{\rho (1)}^+a_2\cdots a_{M} e_{\rho(M)}^+ =0$$ for any $\rho \in S_M$, $\rho \neq id$. Notice that $a_1=
a_2=e_{1,1}$. Let $d^{+}_1 \in N^{+}$ and $d^{-}_2 \in N^{-}$. We set $c^{+}_1=(e_{1,2}+ e_{2,1}) d^{+}_1$ and $\bar{a}_2= e_{1,1}
d_2^{-}$. Notice that $c^{+}_1 \in A^{+}$. Then, since $Cap^\ast_{M+1}[Y;X]\subseteq Id^{\ast}(A)$, we obtained $$0=Cap^\ast_{M+1}(c^{+}_1,e_1^+,\ldots ,e_M^+; a_1, \bar{a}_2, a_3, \ldots
, a_{M})=[d^{+}_1, d^{-}_2]e_{2,k}.$$ Thus $d^{+}_1d^{-}_2=d^{-}_2d^{+}_1$, for all $d^{+}_1 \in N^{+}$, $d^{-}_2 \in N^{-}$ and we are done.
Now we are able to prove the main result about $Id^{\ast}((M_{k}(F),t))$ and the T-$\ast$-ideal generated by the $\ast$-Capelli polynomials $Cap^+_{\frac{k(k+1)}{2}+1}$, $Cap^-_{\frac{k(k-1)}{2}+1}.$
First we prove the following
\[esponentetrasposta\] Let $M=k(k+1)/2$ and $L=k(k-1)/2$ with $k \in \mathbb{N}$, $k>0$. Then $$\mathrm{exp}^\ast(\mathcal{U}^{\ast}_{M+1,L+1})=M+L=k^2=\mathrm{exp}^\ast((M_{k}(F),t)).$$
[**Proof.**]{} The exponent of $\mathcal{U}^{\ast}_{M+1,L+1}$ is equal to the exponent of some minimal variety lying in $\mathcal{U}^{\ast}_{M+1,L+1}$ ( for the definition of minimal variety see [@GZ]). By [@DvS1 Theorem 1.2] and Lemma \[capelliidentita\] we have that $$\mathrm{exp}^\ast(\mathcal{U}^{\ast}_{M+1,L+1})=$$ $$\mathrm{max}\{\mathrm{exp}^{\ast}(UT^\ast(A_1, \dots ,A_n)) \, | \,
UT^\ast(A_1, \dots ,A_n) \, \mathrm{satisfies} \, Cap^+_{M+1} \,
\mathrm{and}\, Cap^-_{L+1}\}.$$
Let $d^+:=\sum_{i=1}^n \mathrm{dim}_FA_i^+$ and $d^-:=\sum_{i=1}^n
\mathrm{dim}_FA_i^-, $ then $$\mathrm{exp}^\ast(\mathcal{U}^{\ast}_{M+1,L+1})=
\mathrm{max}\{\mathrm{exp}^{\ast}(UT^\ast(A_1, \dots ,A_n)) \, | \,
d^++n\leq M+1 \, \mathrm{and}\, d^-+n\leq L+1\}\geq$$ $$\mathrm{exp}^\ast(UT^\ast (M_k(F)))=k^2=M+L.$$
Since by Lemma \[esponente\], $\mathrm{exp}^\ast(\mathcal{U}^{\ast}_{M+1,L+1})\leq M+L$ then the proof is completed.
\[teorematrasposta\] Let $M=k(k+1)/2$ and $L=k(k-1)/2$ with $k \in \mathbb{N}$, $k>0.$ Then
$$\mathcal{U}^{\ast}_{M+1,L+1}=\mathrm{var}^\ast(\Gamma^{\ast}_{M+1,L+1})=\mathrm{var}^\ast(M_{k}(F)\oplus
D'),$$
where $D'$ is a finite dimensional $\ast$-algebra such that $\mathrm{exp}^{\ast}(D')<M+L$. In particular $$c^{\ast}_n(\Gamma^{\ast}_{M+1,L+1})\simeq c^{\ast}_n(M_{k}(F)).$$
[**Proof.**]{} By Lemma \[esponentetrasposta\] we have that $\mathrm{exp}^{\ast}(\mathcal{U}^{\ast}_{M+1,L+1})=M+L$.
Let $A=A^+\oplus A^-$ be a generating $\ast$-algebra of $\mathcal{U}^{\ast}_{M+1,L+1}$. As remarked before we can assume that $A$ is finite dimensional. Thus, by Corollary \[decomposizionefinita\], there exists a finite number of reduced $\ast$-algebras $B_1,\ldots , B_s$ and a finite dimensional $\ast$-algebra $D'$ such that
$$\mathcal{U}^{\ast}_{M+1,L+1}=\mathrm{var}^\ast(A)=\mathrm{var}^\ast(B_1\oplus
\cdots \oplus B_s \oplus D'). \eqno (1)$$
Moreover $$\mathrm{exp}^{\ast}(B_1)=\cdots
=\textrm{exp}^{\ast}(B_s)=
\textrm{exp}^{\ast}(\mathcal{U}^{\ast}_{M+1,L+1})=M+L$$ and
$$\textrm{exp}^{\ast}(D')< \textrm{exp}^{\ast}(\mathcal{U}^{\ast}_{M+1,L+1})= M+L.$$
Next, we analyze the structure of a finite dimensional reduced $\ast$-algebra $R$ such that $\textrm{exp}^{\ast}(R)= M+L=
\textrm{exp}^{\ast}(\mathcal{U}^{\ast}_{M+1,L+1})$ and $\Gamma^\ast_{M+1,L+1}\subseteq Id^{\ast}(R) $. We can write
$$R=R_1
\oplus \cdots \oplus R_q + J ,$$ where $R_i$ are simple $\ast$-subalgebras of $R$, $J=J(R)$ is the Jacobson radical of $R$ and $R_1J\cdots J R_q\neq 0$.
Recall that every $\ast$-algebra $R_i$ is isomorphic to one of the following algebras :$(M_{k_i}(F),t)$ or $(M_{2m_i}(F),s)$ or $(M_{h_i}(F)\oplus M_{h_i}(F)^{op},exc).$
Let $t_1$ be the number of $\ast$-algebras $R_i$ of the first type, $t_2$ the number of $\ast$-algebras $R_i$ of the second type and $t_3$ the number of $R_i$ of the third type, with $t_1+t_2+t_3=q$.
By [@DvLs Theorem 4.5] and [@DvLs Proposition 4.7] there exists a $\ast$-algebra $\overline{R}$ isomorphic to the $\ast$-algebra $UT^\ast(R_1, \ldots, R_q)$ such that
$$\mathrm{exp}^{\ast}(R)=\mathrm{exp}^{\ast}(\overline{R})=\mathrm{exp}^{\ast}(UT^\ast(R_1,
\ldots, R_q)).$$
Let observe that $$k^2=M+L=\mathrm{exp}^{\ast}(R)=\mathrm{exp}^{\ast}(\overline{R})=\mathrm{exp}^{\ast}(UT^\ast(R_1,
\ldots, R_q)) =$$ $$\mathrm{dim}_FR_1+ \cdots + \mathrm{dim}_F R_q=
k_1^2 + \cdots + k_{t_1}^2+ (2 m_1)^2+ \cdots + (2 m_{t_2})^2 +
2h_1^2 + \cdots + 2h_{t_3}^2.$$
Let $d^{\pm}=\mathrm{dim}_F(R_1\oplus \cdots \oplus
R_q)^{\pm}$ then $$d^++d^-=d=\mathrm{dim}_F(R_1\oplus \cdots \oplus
R_q)=\mathrm{exp}^{\ast}(\overline{R})=M+L.$$ By [@DvS1 Lemma 3.2] $\overline{R}$ does not satisfy the $\ast$-Capelli polynomials $Cap^\ast_{d^++q-1}[Y;X]$ and $Cap^\ast_{d^-+q-1}[Z;X]$, but $\overline{R}$ satisfies $Cap^\ast_{M+1}[Y;X]$ and $Cap^\ast_{L+1}[Z;X]$. Thus $d^++q-1\leq M$ and $d^-+q-1\leq L$. Hence $d^++ d^- + 2q-2\leq M+L$. Since $d^++d^-= M+L$ we obtain that $2q-2= 0$ and so $1=q= t_1+t_2 +t_3$. Since $t_1$, $t_2$ and $t_3$ are nonnegative integers, we have the following three possibilities
1. $t_1=1$ and $t_2=t_3=0$;
2. $t_2=1$ and $t_1=t_3=0$
3. $t_3=1$ and $t_1=t_2=0$.
If $t_2=1$, then $R=(M_{2m}(F), s)+J$ and $\mathrm{exp}^\ast(R)=4m^2$. Thus $$k^2=M+L=\mathrm{exp}^{\ast}(R)=4m^2$$ and so $k=2m$. By hypothesis $R$ satisfies $Cap^\ast_{L+1}[Z;X]$ but, since $Id^\ast(R)\subseteq
Id^\ast(\overline{R})$, $R$ does not satisfy $Cap^\ast_{d^-}[Z;X],$ where $d^- = m(2m+1).$ It follows that $$L+1=k(k-1)/2+1=m(2m-1)+1<m(2m+1)=d^-,$$ for $m\geq 1,$ and this is a contradiction.
Let assume $t_3=1$. Then $R=(M_{h}(F)\oplus
M_{h}(F)^{op})+J$ and $\mathrm{exp}^{\ast}(R)=2h^2$. Thus $k^2=M+L=\mathrm{exp}^\ast(R)=2h^2$ and this is impossible.
Then $t_1 =1$ and in this case $$R\cong M_k(F)+J.$$ From Lemmas \[radicaltrasposta\], \[j10trasposta\], \[Ntrasposta\] we obtain
$$R\cong (M_k(F)+J_{11})\oplus J_{00}\cong (M_k(F)\otimes
N^\sharp)\oplus J_{00}$$
where $N^\sharp$ is the algebra obtained from $N$ by adjoining a unit element. Since $N^\sharp$ is commutative, it follows that $M_k(F) + J_{11}$ and $M_k(F)$ satisfy the same $\ast$-identities. Thus var$^\ast$(R)=var$^\ast(M_k(F)\oplus J_{00})$ with $J_{00}$ a finite dimensional nilpotent $\ast$-algebra. Hence, recalling the decomposition given in (1), we get
$$\mathcal{U}^{\ast}_{M+1,L+1}=\mathrm{var}^{\ast}(\Gamma_{M+1,L+1})=\mathrm{var}^{\ast}(M_k(F)\oplus
D'),$$
where $D'$ is a finite dimensional $\ast$-algebra with exp$^{\ast}(D')<M+L$. Then, from Corollary \[codimensioni\] we have
$$c^{\ast}_n(\Gamma_{M+1,L+1})\simeq c^{\ast}_n(M_k(F))$$
and the theorem is proved.
Asymptotics for $\mathcal{U}^\ast_{m(2m-1)+1,m(2m+1)+1}$ and $(M_{2m}(F),s)$
============================================================================
Throughout this section we assume that $A=M_{2m}(F)+ J$, where $J=J(A)$ is the Jacobson radical of the finite dimensional $\ast$-algebra $A$ and $(M_{2m}(F),s)$ is the algebra of matrices with symplectic involution.
\[j10simplettica\] Let $M=m(2m-1)$ and $L=m(2m+1)$ with $m \in \mathbb{N}$, $m>0$. If $\Gamma^\ast_{M+1,L+1}\subseteq Id^{\ast}(A)$, then $J_{10}=J_{01}=(0)$.
[**Proof.**]{} By Lemma \[identitytrasposta\], $(M_{2m}(F),s)$ does not satisfy the $\ast$-Capelli polynomial $Cap^\ast_{M}[Y,X]$. Also there exist elements $a_1^+, \ldots ,
a_M^+ \in M_{2m}(F)^{+}$ and $b_1, \ldots , b_{M-1} \in M_{2m}(F)$ such that $$Cap^\ast_{M}(a_1^{+}, \ldots , a_M^{+};b_1, \ldots ,
b_{M-1})=e_{1,2m},$$ where the $e_{i,j}$’s are the usual matrix units. Let $d \in
J_{01},$ then $d+d^\ast \in (J_{01} \oplus J_{10})^{+}$. Since $\Gamma^\ast_{M+1, L+1}\subseteq Id^{\ast}(A)$ we have $$0=Cap^\ast_{M+1}(a_1^{+}, \ldots , a_M^{+},d+d^\ast;b_1, \ldots ,
b_{M-1},e_{2m,2m})=de_{1,2m}\pm e_{1,2m}d^\ast.$$ Hence $de_{1,2m} =\mp e_{1,2m}d^\ast \in J_{01}\cap J_{10}=(0)$ and so $d=0$, for all $d \in J_{01}$. Thus $J_{01}=(0)=J_{10}$ and we are done.
\[Nsimplettica\] Let $M=m(2m-1)$ and $L=m(2m+1)$ with $m \in \mathbb{N}$, $m>0$. Let $J_{11} \cong M_{2m}(F) \otimes_F N$, as in Lemma \[radicaltrasposta\]. If $\Gamma_{M+1,L+1}\subseteq
Id^{\ast}(A)$, then $N$ is commutative.
[**Proof.**]{} Let $N$ be the finite dimensional nilpotent $\ast$-algebra of the Lemma \[radicaltrasposta\]. As in Lemma \[Ntrasposta\] we can consider an ordered basis of $M_{2m}(F)^{+}$ consisting of symmetric matrices $e_h^+ \in \{e_{i,j}+e_{m+j,m+i}\,
| \, i,j=1,\ldots m\}\cup \{ e_{i,m+j}- e_{j,m+i}\, | \, 1 \leq i <
j \leq m\} \cup \{ e_{m+i,j}- e_{m+j,i}\, | \, 1 \leq i < j \leq
m\}$ with $e_1^+= e_{1,1}+e_{m+1,m+1}$ and $a_0,a_1, \ldots , a_M
\in M_{2m}(F)$ such that $$a_0e_1^+a_1\cdots a_{M-1} e_M^+ a_M=e_{1,1}$$ and $$a_0e_{\sigma (1)}^+a_1\cdots a_{M-1} e_{\sigma (M)}^+ a_M=0$$
for any $\sigma \in S_M$, $\sigma \neq id$. Consider $d^{+}_1, d^{+}_2 \in N^{+}$ and set $c^{+}_1= (e_{1,1}+e_{m+1,m+1})
d^{+}_1$ and $c^{+}_2= (e_{1,1}+e_{m+1,m+1}) d^{+}_2$. Since $N$ commutes with $M_{2m}(F)$, $c^{+}_1, c^{+}_2\in A^{+}$ and we obtain $$Cap^\ast_{M+2}(c^{+}_1,e_1^+,\ldots ,e_M^+, c^{+}_2; a_0, \ldots , a_M )=
[c^{+}_2,c^{+}_1]e_{1,1}=[d^{+}_2, d^{+}_1]e_{1,1}.$$
Since $Cap^\ast_{M+1}[Y;X]\subseteq Id^{\ast}(A)$ we have $d^{+}_1d^{+}_2=d^{+}_2d^{+}_1$, for all $d^{+}_1, d^{+}_2 \in
N^{+}$.
Now, let $e_1^-,\ldots , e_L^-$ be an ordered basis of $M_{2m}(F)^{-}$ consisting of skew matrices, $e_h^- \in \{e_{i,j}-e_{m+j,m+i}\, | \,
i,j=1,\ldots m\}\cup \{ e_{i,m+j}+ e_{j,m+i}\, | \, 1 \leq i < j
\leq m\}\cup \{ e_{i,m+i}\, | \, i=1,\ldots , m\} \cup \{ e_{m+i,j}+
e_{m+j,i}\, | \, 1 \leq i < j \leq m\}\cup \{ e_{m+i,i}\, | \,
i=1,\ldots , m\}$ such that $e_1^-= e_{1,1}-e_{m+1,m+1}$ .
We consider $b_0, \ldots , b_L \in M_{2m}(F)$ such that $b_0=e_{1,1}$, $b_L=e_{2m,2m}$ $$b_0e_1^-b_1 \cdots b_{L-1}e_L^-b_L=e_{1,2m}$$ and $$b_0e_{\tau (1)}^-b_1\cdots b_{L-1} e_{\tau (L)}^- b_L=0$$ for all $\tau \in S_L$, $\tau \neq id$. Let $d^{-}_1, d^{-}_2 \in
N^{-}$. Let $c^{-}_1= (e_{1,1}+e_{m+1,m+1}) d^{-}_1$ and $c^{-}_2=
(e_{1,1}+e_{m+1,m+1}) d^{-}_2$. Since $N$ commutes with $M_{2m}(F)$ then $c^{-}_1, c^{-}_2\in A^{-}$. As above we obtain
$$Cap^\ast_{L+2}(c^{-}_1,e_1^-,\ldots ,e_L^-, c^{-}_2;b_0, \ldots , b_L )=$$ $$[c^{-}_1, c^{-}_2]e_{1,2m}=(e_{1,1}+e_{m+1,m+1})e_{1,2m}[d^{-}_1,
d^{-}_2]=e_{1,2m}[d^{-}_1, d^{-}_2].$$
Since $Cap^\ast_{L+1}[Z;X]\subseteq Id^{\ast}(A)$ we get that $[d^{-}_1,d^{-}_2]=0$ and so $d^{-}_1d^{-}_2=d^{-}_2d^{-}_1$, for all $d^{-}_1, d^{-}_2 \in N^{-}$.
Next we show that $N^{+}$ commutes with $N^{-}$.
Take $e_1^+,\ldots
, e_M^+$ an ordered basis of $M_{2m}(F)^{+}$ such that $e_1^+= e_{1,1}+e_{m+1,m+1}$. Let $a_1, \ldots ,
a_{M} \in M_{2m}(F)$ be such that $$a_1e_1^+a_2\cdots a_{M} e_{M}^+=e_{1,2m}$$ and $$a_1e_{\rho (1)}^+a_2\cdots a_{M} e_{\rho(M)}^+ =0$$ for any $\rho \in S_M$, $\rho \neq id$. Notice that $a_1=
a_2=e_{1,1}$. Let $d^{+}_1 \in N^{+}$ and $d^{-}_2 \in N^{-}$. We set $c^{+}_1=(e_{1,2}+ e_{m+2,m+1}) d^{+}_1$ and $\bar{a}_2= e_{1,1}
d_2^{-}$, then $c^{+}_1 \in A^{+}$. Since $Cap^\ast_{M+1}[Y;X]\subseteq Id^{\ast}(A)$, we obtain
$$0=Cap^\ast_{M+1}(c^{+}_1,e_1^+,\ldots ,e_M^+; a_1, \bar{a}_2, a_3, \ldots
, a_{M})=[d^{+}_1, d^{-}_2]e_{2,2m}.$$
Thus $d^{+}_1d^{-}_2=d^{-}_2d^{+}_1$, for all $d^{+}_1
\in N^{+}$, $d^{-}_2 \in N^{-}$ and we are done.
\[esponentesimplettica\] Let $M=m(2m-1)$ and $L=m(2m+1)$ with $m \in \mathbb{N}$, $m>0$. Then
$$\mathrm{exp}^\ast(\mathcal{U}^{\ast}_{M+1,L+1})=M+L=4m^2=\mathrm{exp}^\ast((M_{2m}(F),s)).$$
[**Proof.**]{} The proof is the same of that of Lemma \[esponentetrasposta\].
Now we are able to prove the following
\[teoremasimplettica\] Let $M=m(2m-1)$ and $L=m(2m+1)$ with $m \in \mathbb{N}$, $m>0$. Then
$$\mathcal{U}^{\ast}_{M+1,L+1}=\mathrm{var}^\ast(\Gamma_{M+1,L+1})=\mathrm{var}^\ast(M_{2m}(F)\oplus
D''),$$
where $D''$ is a finite dimensional $\ast$-algebra such that $\mathrm{exp}^{\ast}(D'')<M+L$. In particular
$$c^{\ast}_n(\Gamma_{M+1,L+1})\simeq c^{\ast}_n(M_{2m}(F)).$$
[**Proof.**]{} The first part of the proof follows step by step that of Theorem \[teorematrasposta\] and we obtain
$$\mathcal{U}^{\ast}_{M+1,L+1}=\mathrm{var}^\ast(B_1\oplus
\cdots \oplus B_s \oplus D''),$$
where $B_1,\ldots , B_s$ are reduced $\ast$-algebras and $D''$ is a finite dimensional $\ast$-algebra such that $\mathrm{exp}^\ast(D'')<\mathrm{exp}^\ast(B_i)=M+L$, for all $i=1,\ldots s$.
Let $R$ be a finite dimensional reduced $\ast$-algebra with $\mathrm{exp}^\ast(R)=M+L=\mathrm{exp}^\ast(\mathcal{U}_{M+1,L+1})$ and $\Gamma^\ast_{M+1,L+1}\subseteq Id^{\ast}(R)$. We can write $R=R_1
\oplus \cdots \oplus R_q + J$, where $R_i$ are simple $\ast$-subalgebras of $R$, $J=J(R)$ is the Jacobson radical of $R$. Let $t_1$ be the number of $\ast$-algebras $R_i$ isomorphic to $(M_{k_i}(F),t)$, $t_2$ the number of $\ast$-algebras $R_i$ isomorphic to $(M_{2m_i}(F),s)$ and let $t_3$ be the number of $R_i$ isomorphic to $(M_{h_i}(F)\oplus M_{h_i}(F)^{op},exc),$ where $t_1+t_2+t_3=q$. Hence, as in Theorem \[teorematrasposta\], there exists a $\ast$-algebra $\overline{R}$ isomorphic to the $\ast$-algebra $UT^\ast(R_1, \ldots, R_q)$ such that $\mathrm{exp}^{\ast}(R)=\mathrm{exp}^{\ast}(\overline{R})=\mathrm{exp}^{\ast}(UT^\ast(R_1,
\ldots, R_q))$ and $$4m^2=M+L=
\mathrm{exp}^{\ast}(R)=\mathrm{exp}^{\ast}(\overline{R})=\mathrm{exp}^{\ast}(UT^\ast(R_1,
\ldots, R_q))=$$ $$k_1^2 + \cdots + k_{t_1}^2+ (2 m_1)^2+ \cdots + (2 m_{t_2})^2 +
2h_1^2 + \cdots + 2h_{t_3}^2.$$
As in the proof of the Theorem \[teorematrasposta\] we have $q=1$ and so we obtain only three possibilities: $t_1=1$ and $t_2=t_3=0$ or $t_2=1$ and $t_1=t_3=0$ or $t_3=1$ and $t_1=t_2=0$.
If $t_1=1$, then $R=(M_{k}(F), t)+J$ and $\mathrm{exp}^\ast(R)=k^2$. Thus $4m^2=M+L=\mathrm{exp}^{\ast}(R)=k^2$ and so $k=2m$. Hence $R$ satisfies $Cap^\ast_{M+1}[Y;X]$ and, since $Id^\ast(R)\subseteq Id^\ast(\overline R)$, $R$ does not satisfy $Cap^\ast_{d^+}[Y;X],$ where $d^+ = k(k+1)/2$. It follows that $M+1=m(2m-1)+1<m(2m+1)=k(k+1)/2=d^+$ for $m\geq 1,$ and this is a contradiction.
Let assume $t_3=1$. Then $R=(M_{h}(F)\oplus M_{h}(F)^{op})+J$ and $\mathrm{exp}^{\ast}(R)=2h^2$. Thus $4m^2=M+L=\mathrm{exp}^\ast(R)=2h^2$ and this is impossible.
Finally, let $t_2 =1$ and $t_1=t_3=0$. Then $R\cong M_{2m}(F)+J$. By Lemmas \[radicaltrasposta\], \[j10simplettica\], \[Nsimplettica\] we obtain
$$R\cong (M_{2m}(F)+J_{11})\oplus J_{00}\cong (M_{2m}(F)\otimes
N^\sharp)\oplus J_{00}$$
where $N^\sharp$ is the algebra obtained from $N$ by adjoining a unit element. Since $N^\sharp$ is commutative, it follows that $M_{2m}(F) + J_{11}$ and $M_{2m}(F)$ satisfy the same $\ast$-identities. Thus var$^\ast$(R)=var$^\ast(M_{2m}(F)\oplus
J_{00})$ with $J_{00}$ a finite dimensional nilpotent $\ast$-algebra. Hence by (1) we get
$$\mathcal{U}^{\ast}_{M+1,L+1}=\mathrm{var}^{\ast}(\Gamma_{M+1,L+1})=\mathrm{var}^{\ast}(M_{2m}(F)\oplus
D''),$$
where $D''$ is a finite dimensional $\ast$-algebra with exp$^{\ast}(D'')<M+L$. Then, from Corollary \[codimensioni\], we have $$c^{\ast}_n(\Gamma_{M+1,L+1})\simeq c^{\ast}_n(M_{2m}(F))$$
and the theorem is proved.
Asymptotics for $\mathcal{U}^\ast_{h^2+1,h^2+1}$ and $(M_{h}(F)\oplus M_{h}(F)^{op}, exc)$
==========================================================================================
Throughout this section we assume that $A=(M_{h}(F)\oplus M_{h}(F)^{op})+ J$, where $J=J(A)$ is the Jacobson radical of the finite dimensional $\ast$-algebra $A$ and $(M_{h}(F)\oplus M_{h}(F)^{op}, exc)$ is the direct sum of the algebra of $h \times h$ matrices and the opposite algebra with the exchange involution.
We start with the following lemmas
\[j10scambio\] Let $M=L=h^2$ with $h \in \mathbb{N}$, $h>0$. If $\Gamma^\ast_{M+1,L+1}\subseteq Id^{\ast}(A)$, then $J_{10}=J_{01}=(0)$.
[**Proof.**]{} By Lemma \[identitytrasposta\], $(M_{h}(F)\oplus M_{h}(F)^{op}, exc)$ does not satisfy the $\ast$-Capelli polynomial $Cap^\ast_{M}[Y,X]$. Also, there exist elements $a_1^+, \ldots ,
a_M^+ \in (M_{h}(F)\oplus M_{h}(F)^{op})^{+}$ and $b_1, \ldots , b_{M-1} \in M_{h}(F)\oplus M_{h}(F)^{op}$ such that $$Cap^\ast_{M}(a_1^{+}, \ldots , a_M^{+};b_1, \ldots ,
b_{M-1})=\tilde{e}_{1,h},$$ where the $\tilde{e}_{i,j}=(e_{i,j}, e_{j,i})$ and $e_{i,j}$’s are the usual matrix units. Let $d \in J_{01},$ then $d+d^\ast \in
(J_{01} \oplus J_{10})^{+}$. Since $\Gamma^\ast_{M+1, L+1}\subseteq
Id^{\ast}(A)$ it follows $$0=Cap^\ast_{M+1}(a_1^{+}, \ldots , a_M^{+},d+d^\ast;b_1, \ldots ,
b_{M-1},\tilde{e}_{h,h})\tilde{e}_{h,h}=d\tilde{e}_{1,h}.$$ Hence $d=0$, for all $d \in J_{01}$. Thus $J_{01}=(0)$ and $J_{10}=(0)$.
\[Nscambio\] Let $M=L=h^2$ with $h \in \mathbb{N}$, $h>0$. Let $J_{11} \cong (M_{h}(F)\oplus M_{h}(F)^{op}) \otimes_F N$, as in Lemma \[radicaltrasposta\]. If $\Gamma_{M+1,L+1}\subseteq
Id^{\ast}(A)$, then $N$ is commutative.
[**Proof.**]{} Let $N$ be the finite dimensional nilpotent $\ast$-algebra of the Lemma \[radicaltrasposta\]. Let now $v_1^+, \ldots , v_M^+$ an ordered basis of $(M_{h}(F)\oplus M_{h}(F)^{op})^+$ consisting of all $e_{i,j}^+=(e_{i,j},e_{i,j})$ such that $v_1^+=e_{1,1}^+$ and let $a_0, \dots , a_M \in M_{h}(F)\oplus M_{h}(F)^{op}$ be such that $$a_0 v_1^+ a_1 \cdots a_{M-1}v_M^+a_M=e_{1,1}^+$$ and $$a_0 v_{\sigma(1)}^+ a_1 \cdots a_{M-1}v_{\sigma(M)}^+a_M=(0,0)$$ for any $\sigma \in S_M$, $\sigma \neq id$.
Now let $d_1^+$, $d_2^+\in N^+$ and set $c_1^+=e_{1,1}^+d_1^+$ and $c_2^+=e_{1,1}^+d_2^+$. Recalling that $N$ commutes with $M_{h}(F)\oplus M_{h}(F)^{op}$ we have that $c_1^+$, $c_2^+\in A^+$ and, as in Lemma \[Ntrasposta\], $$Cap^\ast_{M+2}(c^{+}_1,v_1^+,\ldots ,v_M^+, c^{+}_2; a_0, \ldots , a_M )=
[c^{+}_2,c^{+}_1]e_{1,1}^+=[d^{+}_2,d^{+}_1]e_{1,1}^+.$$
Since $Cap^\ast_{M+1}[Y;X]\subseteq Id^\ast(A)$ it follows $d_1^+d_2^+=d_2^+d_1^+$, for all $d_1^+$, $d_2^+\in N^+$.
Now, let $v_1^-, \ldots , v_L^-$ be an ordered basis of $(M_{h}(F)\oplus M_{h}(F)^{op})^-$ consisting of all $e_{i,j}^-=(e_{i,j},-e_{i,j})$ such that $v_1^-=e_{1,1}^-$ and let $b_0, \dots , b_L \in M_{h}(F)\oplus M_{h}(F)^{op}$ be such that $$b_0 v_1^+ b_1 \cdots b_{L-1}v_L^+b_L=\left\{
\begin{array}{ll}
e_{1,1}^+, & \hbox{if L is even;} \\
e_{1,1}^-, & \hbox{if L is odd.}
\end{array}
\right.$$ and $$b_0 v_{\sigma(1)}^+ b_1 \cdots b_{L-1}v_{\sigma(L)}^+b_L=(0,0)$$ for any $\sigma \in S_M$, $\sigma \neq id$. Now, we take $d_1^-$, $d_2^-\in N^-$ and set $c_1^-=e_{1,1}^+d_1^-$ and $c_2^-=e_{1,1}^+d_2^-$. Since $N$ commutes with $M_{h}(F)\oplus M_{h}(F)^{op}$, we have that $c_1^-$, $c_2^-\in A^-$. Thus, since $Cap^\ast_{L+1}[Z,X]\subseteq Id^\ast(A)$, we obtain $$0\! = \! Cap^\ast_{L+2}(c^{-}_1,v_1^-,\ldots ,v_L^-, c^{-}_2;b_0, \ldots , b_L)\! = \! \left\{
\begin{array}{ll}
[c^{-}_2,c^{-}_1]e_{1,1}^+= [d^{-}_2,d^{-}_1]e_{1,1}^+, & \!\!\hbox{if L is even;} \\
\hbox{$[c^{-}_2,c^{-}_1]$}e_{1,1}^-=[d^{-}_2,d^{-}_1]e_{1,1}^-, &\!\! \hbox{if L is odd.}
\end{array}
\right.$$ In conclusion $[d^{-}_2,d^{-}_1]=0$, for all $d_1^-$, $d_2^-\in N^-$.
Finally, we consider $d_1^+ \in N^+$, $d_2^-\in N^-$ and set $c^+_1=e_{1,1}^+d_1^+$ and $c^+_2=e_{1,1}^-d_2^-$. As above, we have that $c^+_1$, $c^+_2 \in A^+$. Then $$0=Cap^\ast_{M+2}(c^{+}_1,v_1^+,\ldots ,v_M^+, c^{+}_2;a_0, \ldots , a_M)=[c^{+}_2,c^{+}_1]e_{1,1}^+= [d^{-}_2,d^{+}_1]e_{1,1}^-$$ and $[d^{-}_2,d^{+}_1]=0$ for all $d_1^+ \in N^+$, $d_2^-\in N^-$ and the lemma is proved.
\[esponentescambio\] Let $M=L=h^2$ with $h \in \mathbb{N}$, $h>0$. Then $$\mathrm{exp}^\ast(\mathcal{U}^{\ast}_{M+1,L+1})=M+L=2h^2=\mathrm{exp}^\ast((M_{h}(F)\oplus M_{h}(F)^{op},exc)).$$
[**Proof.**]{} The proof is the same of that of Lemma \[esponentetrasposta\].
\[teoremascambio\] Let $M=L=h^2$ with $h \in \mathbb{N}$, $h>0$. Then
$$\mathcal{U}^{\ast}_{M+1,L+1}=\mathrm{var}^\ast(\Gamma_{M+1,L+1})=\mathrm{var}^\ast((M_{h}(F)\oplus M_{h}(F)^{op})\oplus
D'''),$$
where $D'''$ is a finite dimensional $\ast$-algebra such that $\mathrm{exp}^{\ast}(D''')<M+L$. In particular $$c^{\ast}_n(\Gamma_{M+1,L+1})\simeq c^{\ast}_n(M_{h}(F)\oplus M_{h}(F)^{op}).$$
[**Proof.**]{} The proof proceeds as in that of Theorem \[teorematrasposta\]. Let $R$ be a finite dimensional reduced $\ast$-algebra such that $\mathrm{exp}^\ast(R)=M+L=\mathrm{exp}^\ast(\mathcal{U}^\ast_{M+1,L+1})$ and $\Gamma^\ast_{M+1,L+1}\subseteq Id^{\ast}(R)$. We can write $R=R_1 \oplus \cdots \oplus R_q + J$, where $R_i$ are simple $\ast$-subalgebras of $R$, $J=J(R)$ is the Jacobson radical of $R$. If $t_1$ denotes the number of algebras $R_i$ isomorphic to $(M_{k_i}(F),t)$, $t_2$ the number of algebras $R_i$ isomorphic to $(M_{2m_i}(F),s)$ and $t_3$ the number of $R_i$ isomorphic to $(M_{h_i}(F)\oplus
M_{h_i}(F)^{op},exc)$, then we have $t_1+t_2+t_3=q$. As in Theorem \[teorematrasposta\], there exists a $\ast$-algebra $\overline{R}$ isomorphic to the $\ast$-algebra $UT^\ast(R_1, \ldots, R_q)$ such that $\mathrm{exp}^{\ast}(R)=\mathrm{exp}^{\ast}(\overline{R})=\mathrm{exp}^{\ast}(UT^\ast(R_1,
\ldots, R_q))$. Then
$$2h^2=M+L=
\mathrm{exp}^{\ast}(R)=\mathrm{exp}^{\ast}(\overline{R})=\mathrm{exp}^{\ast}(UT^\ast(R_1,
\ldots, R_q))=$$ $$k_1^2 + \cdots + k_{t_1}^2+ (2 m_1)^2+ \cdots + (2 m_{t_2})^2 +
2h_1^2 + \cdots + 2h_{t_3}^2$$
and we have only three possibilities: $t_1=1$ and $t_2=t_3=0$ or $t_2=1$ and $t_1=t_3=0$ or $t_3=1$ and $t_1=t_2=0$.
If $t_1=1$, then $R=(M_{k}(F), t)+J$ and $\mathrm{exp}^\ast(R)=k^2$. Thus $2h^2=M+L=\mathrm{exp}^{\ast}(R)=k^2$ and this is a contradiction.
If $t_2=1$. Then $R=(M_{2m}(F),s)+J$ and $\mathrm{exp}^{\ast}(R)=4m^2$. Thus $2h^2=M+L=\mathrm{exp}^\ast(R)=4m^2$ and so $h^2=2m^2$, contradiction.
Finally, let $t_3 =1$ and $t_1=t_2=0$. Then $R\cong (M_{h}(F)\oplus M_{h}(F)^{op})+J$. As in Theorem \[teorematrasposta\], by Lemmas \[radicaltrasposta\], \[j10scambio\], \[Nscambio\] we obtain
$$R\cong ((M_{h}(F)\oplus M_{h}(F)^{op})+J_{11})\oplus J_{00}\cong
((M_{h}(F)\oplus M_{h}(F)^{op})\otimes N^\sharp)\oplus J_{00}$$
where $N^\sharp$ is the algebra obtained from $N$ by adjoining a unit element. Since $N^\sharp$ is commutative, we have that $(M_{h}(F)\oplus M_{h}(F)^{op}) + J_{11}$ and $M_{h}(F)\oplus
M_{h}(F)^{op}$ satisfy the same $\ast$-identities. Thus var$^\ast$(R)=var$^\ast((M_{h}(F)\oplus M_{h}(F)^{op})\oplus
J_{00})$ with $J_{00}$ a finite dimensional nilpotent $\ast$-algebra. We get
$$\mathcal{U}^{\ast}_{M+1,L+1}=\mathrm{var}^{\ast}(\Gamma_{M+1,L+1})=\mathrm{var}^{\ast}((M_{h}(F)\oplus M_{h}(F)^{op})\oplus
D'''),$$
where $D'''$ is a finite dimensional $\ast$-algebra with exp$^{\ast}(D''')<M+L$. Then, from Corollary \[codimensioni\], we have $$c^{\ast}_n(\Gamma_{M+1,L+1})\simeq c^{\ast}_n((M_{h}(F)\oplus M_{h}(F)^{op}))$$
and the proof is completed.
[10]{}
E. Aljadeff, A. Giambruno and Y. Karasik, *Polynomial identities with involution, superinvolutions and the Grassmann envelope*, Proc. Amer. Math. Soc. **145** (2017), 1843-1857.
S. A. Amitsur, *Identities in rings with involution*, Israel J. Algebra **7** (1969), 63–68.
Y. Bahturin, A. Giambruno and M. Zaicev, *G-identities on associative algebras*, Proc. Amer. Math. Soc. **127** (1999), 63–69.
F. Benanti, *Asymptotics for Graded Capelli Polynomials*, Algebra Repres. Theory **18** (2015), 221–233.
F. Benanti and I. Sviridova *Asymptotics for Amitsur’s Capelli-type polynomials and verbally prime PI-algebras*, Israel J. Math. **156** (2006), 73–91.
A. Berele, A. Giambruno and A. Regev, *Involution codimensions and trace codimensions of matrices are asymptotically equal*, Israel J. Math. **96** (1996), 49–62.
O. M. Di Vincenzo and R. La Scala *Minimal algebras with respect to their $*$-exponent*, J.Algebra **317** (2007), 642-657.
O. M. Di Vincenzo and E. Spinelli, *Some results of $\ast$-minimal algebras with involution*, in *Group, Rings and Group Rings* 2008 (A. Giambruno et al.eds.), International Conference, Ubatuba, Brazil, July 27 August 2, 2008, Contemporary Mathematics **499**, American Mathematical Society, Providence, RI, (2009), 75–88.
O. M. Di Vincenzo and E. Spinelli, *On $\ast$-minimality of algebras with involution*, J. Algebra **323** (2010), 121–131.
O. M. Di Vincenzo and E. Spinelli, *A characterization of $\ast$-minimal algebras with involution*, Israel J. Math. **186** (2011), 381–400.
A. Giambruno and S. Mishchenko, *On star-varieties with almost polynomial growth*, Algebra Colloq. **8** (2001), 33–42.
A. Giambruno and S. Mishchenko, *Polynomial growth of the $\ast$-codimensions and Young diagrams*, Comm. Algebra **29** (2001), 277–284.
A. Giambruno, A. Ioppolo and D. La Mattina, *Varieties of Algebras with Superinvolution of Almost Polynomial Growth*, Algebr. Represent. Theory, [**19**]{} (2016), 599–611.
A. Giambruno, C. Polcino Milies and A. Valenti, *Star-polynomial identities: Computing the exponential growth of the codimensions*, J. Algebra **469** (2017), 302–322.
A. Giambruno and A. Regev, *Wreath products and P.I.algebras*, J. Pure Appl. Algebra **35** (1985), 133–149.
A. Giambruno and M. Zaicev, Polynomial Identities and Asymptotics Methods, Surveys, vol. 122, American Mathematical Society, Providence, RI, 2005.
A. Giambruno and M. Zaicev, *Involution codimensions of finite dimensional algebras and exponential growth*, J. Algebra **222** (1999), 471–484.
A. Giambruno and M. Zaicev, *A characterization of algebras with polynomial growth of the codimensions*, Proc. Amer. Math. Soc. **129** (2001), 59–67.
A. Giambruno and M. Zaicev, *Asymptotics for the Standard and the Capelli Identities*, Israel J. Math. **135** (2003), 125–145.
A. R. Kemer, Ideals of Identities of Associative Algebra, Amer. Math. Soc. Translations of Math. Monographs 87, Providence, RI, 1991.
S. Mishchenko and A. Valenti *A star-variety with almost polynomial growth*, J. Algebra **223** (2000), 66–84. L. H. Rowen, Polynomial Identities in Ring Theory, Academic Press, New York, 1980.
I. Sviridova, *Finitely generated algebras with involution and their identities*, J. Algebra **383** (2013), 144–167.
[^1]: The authors were partially supported by INDAM-GNSAGA of Italy.
| {
"pile_set_name": "ArXiv"
} |
[**[Some new lacunary $f$-statistical $A$-convergent sequence spaces of order $\alpha$ ]{}**]{}
0.5 cm
Ekrem Sava$\c{s}$
Istanbul Commerce University, 34840 Istanbul, Turkey
E-mail : [email protected]
0.5 cm
Stuti Borgohain $^*$[[^1]]{}
Department of Mathematics
Indian Institute of Technology, Bombay
Powai:400076, Mumbai, Maharashtra; INDIA. 0.5 cm
E-mail : stutiborgohain$@$yahoo.com 1 cm
[ We study the concept of density for sets of natural numbers in some lacunary $A$-convergent sequence spaces. Also we are trying to investigate some relation between the ordinary convergence and module statistical convergence for evey unbounded modulus function. Morever we also study some results on the newly defined lacunary $f$-statistically $A$-convergent sequence spaces with respect to some Musielak-Orlicz function.]{}\
Introduction
============
In order to extend the notion of convergence of sequences, Fast [@Fast] and Schoenberg [@Schoenberg] independently introduced the concept of statistical convergence. Later on it was studied from sequence space point of view and also linked with summability theory by many mathematicians including $\check{S}$al$\grave{a}$t [@Salat], Fridy [@Fridy], Tripathy and Borgohain [@Tripathy] and many more. Kolk [@Kolk] began to study the applications of statistical convergence to Banach spaces. Connor et. al [@Connor] proved some important results that relate the statistical convergence to classical properties of Banach spacs.\
The notion depends on the idea of asymtotic density of subsets of the set $\mathbb{N}$ of natural numbers. A subset $A$ of $\mathbb{N}$ is said to have natural density ${\delta(A)}$ if
$${\delta(A)}={\lim_{n \rightarrow \infty}}{\frac{1}{n}}{\sum\limits_{k=1}^n}{\chi_A}(k).$$
where $\chi_A$ is the characteristic function of $A$.\
A sequence $(x_n)_n$ is said to be statistically convergent to $L$, if for any $\varepsilon > 0$,we have $\delta\{(k \in \mathbb{N}:\vert x_k - L \vert \geq \varepsilon )\} = 0.$ Analogously, $(x_n)_n$ is said to be statistically Cauchy if for each $\varepsilon>0$ and $n \in \mathbb{N}$ there exists an integer $m \geq n$ such that $d(\{i \in \mathbb{N}:\Vert x_i - x_m \Vert < \varepsilon \})=1.$\
Fridy [@Fridy] proved that in a Banach space, a sequence is statistically convergent if and only if it is statistically Cauchy. Fast [@Fast] proved that st$\displaystyle\lim_n x_n =x$ if and only if there exists $A \subset \mathbb{N}$ with $d(A)=1$ and $\displaystyle\lim_{n \in A} x_n =x$.\
The notion of a modulus function was introduced by Nakano [@Nakano]. Ruckle [@Ruckle] and Maddox [@Maddox] have introduced and discussed some properties of sequence spaces defined by using a modulus function. By the definition of modulus function, we mean a function $f: R^+ \rightarrow R^+ $ which satisfies:
1. $f(x)=0$ if and only if $x=0$.
2. $f(x+y) \leq f(x)+f(y)$ for every $x,y \in R^+$.
3. $f$ is increasing.
4. $f$ is continuous from the right at 0.
A lacunary sequence is defined as an increasing integer sequence $\theta = (k_r)$ such that $k_0=0$ and $h_r=k_r-k_{r-1} \rightarrow \infty$ as $r \rightarrow \infty.$ Throughout this paper, the intervals determined by $\theta$ will be denoted by $J_r=(k_{r-1}, k_r]$ and the ratio $\frac{k_r}{k_{r-1}}$ will be defined by $\phi_r$.\
In this paper, we study on a new lacunary $f$-statistical $A$-convergent sequence space of order $\alpha$ with respect to the Musielak-Orlicz fuction. We also investigate some results on a new concept of nonmatrix convergence which is intermediate between the ordinary convergence and the statistical convergence.
Definitions and basic results
=============================
By the definition of $f$-density of a set $A \subseteq \mathbb{N}$ , we mean $d_f(A)=\displaystyle\lim_n \frac{f(\vert A(n) \vert)}{f(n)}$, (in case this limit exists )where $f$ is an unbounded modulus function.\
Let $X$ be a normed space and let $(x_n)_n$ be a sequence in $X$. If if for each $\varepsilon > 0$, $d_f(\{i \in \mathbb{N}: \Vert x_i-x \Vert > \varepsilon \})=0$, then it is said that the $f$-statistical limit of $(x_n)_n$ is $x \in X$, and we write it as $f$-stlim$x_n=x$.\
It is clear that $d(A)=1-d(\mathbb{N} \backslash A)$.\
Let us assume that $A \subseteq \mathbb{N}$ and $d_f(A)=0$. For every $n \in N$ we have $f(n) \leq f(\vert A(n)\vert)+f(\vert (\mathbb{N} \backslash A)(n)\vert )$ and so
$$1 \leq \frac{f(\vert A(n) \vert)}{f(n)} +\frac{f(\vert ( \mathbb{N} \backslash A)(n)\vert )}{f(n)} \leq \frac{f(\vert A(n) \vert)}{f(n)}+1$$
By taking limits we deduce that $d_f(\mathbb{N} \backslash A)=1$.\
[**[Corollary 2.1. ]{}**]{}Let $f, g$ be unbounded moduli, $X$ a normed space, $(x_n)_n$ a sequence in $X$ and $x, y \in X$. We have
1. The $f$-statistical convergence implies the statistical convergence (to the same limit).
2. The $f$-statistical limit is unique whenever it exists.
3. Moreover, two different methods of statistical convergence are always compatible, which means that if $f$-st$\displaystyle\lim x_n =x$ and $g$-st$\displaystyle\lim x_n =y$ then $x=y$.
By an Orlicz function , we mean a function $M: [0,\infty )\rightarrow [0,\infty )$, which is continuous, non-decreasing and convex with $M(0) = 0, M(x)>0$, for $x>0$ and $M(x)\rightarrow \infty$, as $x\rightarrow \infty$.\
The idea of Orlicz function is used to construct the sequence space, (see Lindenstrauss and Tzafriri [@Lindenstrauss]),
$$\ell_M=\left\{ (x_k) \in w: \displaystyle\sum_{k=1}^\infty M \left(\frac{\vert x \vert}{\rho}\right) < \infty, \mbox{~for some~} \rho>0 \right\}.$$
This space $\ell_M$ with the norm,
$$\Vert x \Vert = \mbox{inf}\left\{ \rho>0: \displaystyle\sum_{k=1}^\infty M \left(\frac{\vert x_k \vert}{\rho}\right) \leq 1 \right\}$$
becomes a Banach space which is called an Orlicz sequnce space.\
Musielak [@Musielak] defined the concept of Musielak-Orlicz function as $\mathscr{M}=(M_k)$.\
A sequence $\mathscr{N}=(N_k)$ defined by $$N_k(v)=\sup \{ \vert v \vert u -M_k(u): u \geq 0 \}, k=1,2,..$$
is called the complementary function of a Musielak-Orlicz function $\mathscr{M}$. The Musielak-Orlicz sequence space $t_\mathscr{M}$ and its subspace $h_\mathscr{M}$ are defined as follows:
$$t_\mathscr{M}=\{ x \in w: I_\mathscr{M}(cx) < \infty \mbox{~for some ~} c>0\},$$ $$h_\mathscr{M}=\{x \in w:I_\mathscr{M}(cx) < \infty, \forall c >0\},$$
where $I_\mathscr{M}$ is a convex modular defined by,
$$I_\mathscr{M}(x) = \displaystyle\sum_{k=1}^\infty M_k(x_k), x=(x_k) \in t_\mathscr{M}.$$
It is considered $t_\mathscr{M}$ equipped with the Luxemberg norm
$$\Vert x \Vert = \inf \left\{ k>0: I_\mathscr{M}\left(\frac{x}{k}\right) \leq 1 \right\}$$
or equiped with the Orlicz norm
$$\Vert x \Vert^0 = \inf \left\{ \frac{1}{k} (1+I_\mathscr{M}(kx)): k>0 \right\}.$$
A Musielak-Orlicz function $(M_k)$ is said to satisfy $\Delta_2$-condition if there exist constants $a, K>0$ and a sequence $c=(c_k)_{k=1}^\infty \in \ell_+^1$ ( the positive cone of $\ell^1$) such that the inequality
$$M_k(2u) \leq KM_k(u)+c_k$$
holds for all $k \in N$ and $u \in R_+$, whenever $M_k(u) \leq a$.
For any lacunary sequence $\theta= (k_r)$, the space $N_\theta$ defined as, (Freedman et al.\[1\]) $$N_\theta=\left\{(x_k): \displaystyle\lim_{r \rightarrow \infty} h_r^{-1} \displaystyle\sum_{k \in J_r} \vert x_k -L \vert =0, \mbox{~for some~} L \right\}.$$
The space $N_\theta$ is a $BK$ space with the norm, $$\Vert (x_k) \Vert_\theta= \displaystyle\sup_r h_r^{-1} \displaystyle\sum_{k \in J_r} \vert x_k \vert.$$
A sequence $(x_i)$ is said to be lacunary $f$-statistically $A$-convergent to $L$ if, $$d_f\left(\left\{ k \in N: \displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left(\frac{\vert A_i(x) -L \vert}{\rho^{(k)}} \right) > \varepsilon, \mbox{~for some~} L \mbox{~and~} \rho^{(k)}>0 \right\} \right) =0,$$
where $I_r = (i_{i-1}, i_r]$ , $A=(A_i(x))$ such that $A_i x =\displaystyle\sum_{k=1}^\infty a_{ik} x_k$ converges for each $i$ , $h_r^\alpha$ denotes the $\alpha$-th power $(h_r^\alpha)$ of $h_r$, that is, $h^\alpha=(h_r^\alpha)=(h_1^\alpha, h_2^\alpha,...h_r^\alpha,...)$ and $\mathscr{M}=(M_k)$ is a Musielak-Orlicz function and it is written as $f_\theta^\alpha(A, \mathscr{M})$-statistically convergentto $L$ with respect to the Musielak-Orlicz function $\mathscr{M}$.\
Some particular cases :\
If we take $\alpha=1$, then the lacunary $f$-statistically $A$-convergence of order $\alpha$ reduces to the lacunary $f$-statistically $A$-convergent to $L$ ($f_\theta(A, \mathscr{M})$-statistically convergent to $L$) , i.e. $$d_f\left(\left\{ k \in N: \displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r} \displaystyle\sum_{i \in I_r} M_i \left(\frac{\vert A_i(x) -L \vert}{\rho^{(k)}} \right) > \varepsilon, \mbox{~for some~} L \mbox{~and~} \rho^{(k)}>0 \right\} \right) =0.$$\
If $\theta=(2^r)$ and $\alpha=1$ , then the sequence $(x_i)$ is said to be $f(A,\mathscr{M})$-statistically convergent to $L$ if , $$d_f\left(\left\{ k \in N: \displaystyle\sum_{i \in I_r} M_i \left(\frac{\vert A_i(x) -L \vert}{\rho^{(k)}} \right) > \varepsilon, \mbox{~for some~} L \mbox{~and~} \rho^{(k)}>0 \right\} \right) =0.$$
If $M_i(x)=x$, $\theta=(2^r)$ and $\alpha=1$, then we can say the sequence $(x_i)$ is $f$-statistically $A$-convergent to $L$ , if $$d_f\left(\left\{ k \in N: \displaystyle\sum_{i \in I_r} \vert A_i(x) -L \vert > \varepsilon, \mbox{~for some~} L \right\}\right)=0.$$
Let us define a space of lacunary $A$-convergent sequences of order $\alpha$ defined by Musielak-Orlicz function as, $$w_\theta^\alpha(A, \mathscr{M})=\left\{ (x_i):\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left(\frac{\vert A_i(x) - L \vert}{\rho^{(i)}} \right)=0, \mbox{~for some~} L \mbox{~and~} \rho^{(i)}>0\right\}.$$
In particular, if we have $M_i(x)=x$, $\theta=(2^r)$ and $\alpha=1$, then $w_\theta^\alpha(A, \mathscr{M})$ reduces to, $$w(A)=\left\{ (x_i): \displaystyle\sum_{i \in I_r} \vert A_i(x) -L \vert =0, \mbox{~for some~} L \right\}.$$
Main Results
============
[**[Theorem 3.1.]{}**]{} Let $(x_n)_n$ be a sequence in a normed space $w_\theta^\alpha(A, \mathscr{M})$ and $f$ an unbounded modulus. Then $f$-st$\displaystyle\lim x_i=L$ if and only if there exists $X \subseteq \mathbb{N}$ such that $d_f(X)=0$ and $\displaystyle\lim_{i \in \mathbb{N} \backslash X} x_i =L $.\
[*[Proof]{}*]{}: Let $B_j = \left\{ i \in \mathbb{N}:\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left(\frac{ \vert A_i(x)- L \vert}{\rho^{(i)}} \right) > \frac{1}{j} \right\}$, for every $j \in \mathbb{N}$.\
Since $B_j \subset B_{j+1}$ and $d_f(B_j)=0$, it is required to prove the case where some of the $B_j$’s are non-empty, in particular we can assume that $B_1 \neq \phi$.\
Choose any $r_1 \in B_1$. Now, by taking $r_2 \in B_2$ with $r_2 > r_1$ and $\frac{f(\vert B_2(i) \vert)}{f(i)} \leq \frac{1}{2}$, if $i \geq r_2$.\
Inductively we obtain $r_1 < r_2 < r_3 ...$ such that $r_j \in B_j$ and $\frac{f(\vert B_j(i) \vert)}{f(i)} \leq \frac{1}{j}$ whenever $i \geq r_j$.\
Now consider $X= \displaystyle\cup_{j \in \mathbb{N}} ([r_j, r_{j+1}) \cap B_j)$. Then for every $i \geq r_1$ there exists $j \in \mathbb{N}$ such that $r_j \leq i \leq r_{j+1}$ and if $n \in X(i)$ then $n < r_{j+1}$, which implies $n \in B_j$. Therefore $X(i) \subseteq B_j(i)$ and thus,
$$\frac{f(\vert X(i) \vert )}{f(i)} \leq \frac{f(\vert B_j(i) \vert )}{f(i)} \leq \frac{1}{j}$$
which follows $d_f(X)=0$.\
For $\varepsilon > 0$ and $j \in \mathbb{N}$ such that $\frac{1}{j} < \varepsilon$, we have $i \in \mathbb{N} \backslash X$ and $i \geq r_j$ for which there exists $k \geq j$ with $r_k \leq i \leq r_{k+1}$ and this implies $i \notin B_k$, so,
$$\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left( \frac{\vert A_i(x) - L \vert}{\rho^{(i)}} \right) \leq \frac{1}{k} \leq \frac{1}{j} < \varepsilon.$$
We conclude $\displaystyle\lim_{i \in \mathbb{N} \backslash X} x_i =L$.\
Conversely, let us assume that $X \in \mathbb{N}$ satisfies $\displaystyle\lim_{i \in \mathbb{N} \backslash X} x_i=L$ and $d_f(X)=0$. For $\varepsilon > 0$, there exists $i_0 \in \mathbb{N}$ such that if $i > i_0$ and $i \in \mathbb{N} \backslash X$ then
$$\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left( \frac{\vert A_i(x) - L \vert}{\rho^{(i)}} \right) \leq \varepsilon.$$
This implies,
$$\left\{i \in \mathbb{N}: \displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left( \frac{\vert A_i(x) - L \vert}{\rho^{(i)}} \right) > \varepsilon \right\} \subseteq X \cup \{1,..i\}$$
and then $d_f\left(\left\{i \in \mathbb{N}: \displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left( \frac{\vert A_i(x) - L \vert}{\rho^{(i)}} \right) > \varepsilon \right\}\right)=0$.\
0.3 cm
[**[Definition 3.2.]{}**]{} The sequence $(x_n)_n$ is $f$-statistically Cauchy if for every $\varepsilon > 0$ there exists $N \in \mathbb{N}$ such that $d_f(\{i \in \mathbb{N} : \Vert x_i - x_N \Vert > \varepsilon \})= 0$.\
It is well known that if a sequence is $f$-statistically convergent then it is $f$-statistically Cauchy. The converge is true if the space is complete and this result is a particular case of filter convergence.\
[**[Thoerem 3.3.]{}**]{} For the Banach space $w_\theta^\alpha(A,\mathscr{M})$ where $f$ be an unbounded modulus and $(x_n)_n$ is an $f$-statistically Cauchy sequence. Then $(x_n)_n$ is $f$-statistically convergent.\
[*[Proof]{}*]{}: For every $k \in N$, let $x_{N_k}$ be such that
$$d_f\left(\left\{i \in \mathbb{N}: \displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left( \frac{\vert A_i(x) - A_{N_k}(x) \vert}{\rho^{(i)}} \right)> \frac{1}{k} \right\}\right)=0.$$
Consider the sets $I_k = \cap_{j \leq k} B(x_{N_j}, \frac{1}{j})$ and $J_k=\{ i \in \mathbb{N}: x_i \notin I_k\}.$\
Then for each $k \in \mathbb{N}$ we have $\mbox{diam}(I_k) \leq \frac{2}{k}$ and
$$J_k=\cup_{j \leq k} \left\{i \in \mathbb{N}: \displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left( \frac{\vert A_i(x) - A_{N_j}(x) \vert}{\rho^{(i)}} \right) > \frac{1}{j} \right\},$$
which implies $d_f(J_k)=0$.\
Consequently each $I_k \neq \phi$.\
Sincet $I_1 \supseteq I_2 \supseteq ...$ and $J_1 \subseteq J_2 \subseteq...$ so as in the proof of the previous theorem, we can find a sequence of natural numbers $r_1<r_2<....$ such that $r_k \in J_k$ and if $i \geq r_k$ then $\frac{f(\vert J_k(i)\vert )}{f(i)} \leq \frac{1}{k}.$\
Considering $X= \cup_{k \in \mathbb{N}} ([r_k, r_{k+1}) \cap J_k)$, we get $d_f(X)=0$.\
Since $w_\theta^\alpha(A,\mathscr{M})$ is complete, then $\cap_{k \in \mathbb{N}} I_k$ has exactly one element, say $x$.\
We have to prove that $\displaystyle\lim_{i \in \mathbb{N} \backslash X} x_i =L$.\
For $\varepsilon > 0$, choose $j \in \mathbb{N}$ with $\frac{2}{j} < \varepsilon$. If $i \geq r_j$ and $i \in \mathbb{N} \backslash X$ then there exists $k \geq j$ such that $r_k \leq i < r_{k+1}$ and then $i \notin J_k$, which implies $x_i \in I_k$ and thus,
$$\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left( \frac{\vert A_i(x) - A_{N_k}(x) \vert}{\rho^{(i)}} \right) \leq \frac{2}{k} \leq \frac{2}{j} < \varepsilon.$$
This completes the proof of the theorem.\
[**[Theorem 3.4.]{}**]{} Let $(x_n)_n$ be a sequence in $w_\theta^\alpha(A,\mathscr{M})$. If for every unbounded modulus $f$ there exists $f$-st$\displaystyle\lim x_i$ then all these limits are the same $x \in w_\theta^\alpha (A, \mathscr{M})$ and $(x_n)_n$ also converges to $L$ in the norm topology.\
[*[Proof]{}*]{} : It is proved that for $f,g$ two unbounded moduli, $X$ a normed space, $(x_n)_n$ a sequence in $X$ and $L, M \in X$, the $f$-statistical limit is unique whenever it exists.\
If it is false that $\displaystyle\lim x_i =L$, there exists $\varepsilon > 0$ such that
$$X=\left\{ i \in \mathbb{N}: \displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left( \frac{\vert A_i(x) - L \vert}{\rho^{(i)}} \right)> \varepsilon \right\}$$
is infinite.\
Now, by choosing an unbounded modulus $f$ which will satisfy $d_f(X)=1$, then this clearly contradicts the assumption that $f$-stlim $x_i=L$, which completes the proof of the Theorem.\
**Theorem 3.5.** Let $\mathscr{M}$ be a Musielak-Orlicz function, $x=(x_i)$ be a bounded sequence and $\theta=(i_r)$ be a lacunary sequence. If $\displaystyle\lim_{r \rightarrow \infty} \frac{h_r}{h_r^\alpha}=1$,then $x \in f_\theta^\alpha(A, \mathscr{M}) \Rightarrow x \in w_\theta^\alpha (A, \mathscr{M})$.\
[*[Proof]{}*]{}: Suppose that $x=(x_i)$ be a bounded sequence and $ f_\theta^\alpha(A,\mathscr{M})-\mbox{lim} x_i =L$.\
Then we have, $$d_f \left( \left\{ i \in \mathbb{N}: \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left(\frac{\vert A_i(x) - L \vert}{\rho^{(i)}}\right) > \varepsilon \right\}\right) =0.$$
For $\varepsilon >0$ given, let us denote $\sum_1$ as the sum over $i \in I_r$, $\vert A_i(x)-L \vert \geq \varepsilon$ and $\sum_2$ denotes the sum over $i \in I_r$, $\vert A_i(x)-L \vert < \varepsilon$ respectively.\
As $x \in \ell_\infty$, we have a constant $T>0$ such that $\vert x_i \vert \leq T$. Now, for $\varepsilon > 0$, we have,\
$\frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left(\frac{\vert A_i(x) - L \vert}{\rho^{(i)}}\right)$\
$\frac{1}{h_r^\alpha} \sum_1 M_i\left(\frac{\vert A_i(x)-L \vert}{\rho^{(i)}}\right) +
\frac{1}{h_r^\alpha} \sum_2 M_i\left(\frac{\vert A_i(x)-L \vert}{\rho^{(i)}}\right)$\
$\leq \frac{1}{h_r^\alpha} \sum_1 \mbox{max} \left\{M_i\left(\frac{T}{\rho^{(i)}}\right), M_i\left(\frac{T}{\rho^{(i)}}\right) \right\} + \frac{1}{h_r^\alpha} \sum_2 M_i\left(\frac{\varepsilon}{\rho^{(i)}} \right)$\
$\leq \mbox{max}\{M_i(K),M_i(K) \} \frac{1}{h_r^\alpha} \vert \{ i \in I_r: \vert A_i(x)-L \vert \geq \varepsilon \} \vert + \frac{h_r}{h_r^\alpha} M_i(\varepsilon_1) , ~~\frac{T}{\rho^{(i)}}=K, \frac{\varepsilon}{\rho^{(i)}}=\varepsilon_1.$\
Hence, $x \in w_\theta^\alpha(A, \mathscr{M})$.\
[**[Theorem 3.6. ]{}**]{} Let $\mathscr{M}=(M_i)$ be a Musielak-Orlicz function where $(M_i)$ is pointwise convergent. Then, $w_\theta^\alpha(A, \mathscr{M}) \subset f_\theta^\alpha(A, \mathscr{M})$ if and only if $\displaystyle\lim_i M_i \left(\frac{ \nu}{\rho^{(i)}}\right) > 0$ for some $\nu>0, \rho^{(i)}>0$.\
[*[Proof]{}*]{} : Let $\varepsilon >0$ and $x \in w_\theta^\alpha(A, \mathscr{M})$.\
If $\displaystyle\lim_i M_i\left( \frac{\nu}{\rho^{(i)}}\right) >0$, then we have a number $c > 0$ such that
$$M_i \left(\frac{\nu}{\rho^{(i)}}\right) \geq c, \mbox{~for~} \nu > \varepsilon.$$
Let us define, $I_r^1=\left\{ i \in I_r: M_i \left(\frac{\vert A_i(x)-L \vert}{\rho^{(i)}}\right) \geq \varepsilon \right\}$.\
Then, $$\begin{aligned}
\frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left(\frac{\vert A_i(x)-L \vert}{\rho^{(i)}} \right)
&\geq &
\frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r^1} M_i \left(\frac{\vert A_i(x)-L \vert}{\rho^{(i)}} \right)\\
&\geq &
c \frac{1}{h_r^\alpha} \vert A_0(\varepsilon) \vert\end{aligned}$$
Hence, it follows that $ x \in f_\theta^\alpha(A, \mathscr{M})$.\
Conversely, let the condition do not exist. Now, for a number $\nu>0$ , let $\displaystyle\lim_i M_i\left(\frac{\nu}{\rho^{(i)}}\right)=0$ for some $\rho>0$.
Select a lacunary sequence $\theta=(n_r)$ such that $M_i \left(\frac{\nu}{\rho^{(i)}}\right)< 2^{-r}$ for any $i > n_r$.\
Also, if $A=I$ , then we can define a sequence $x$ by putting,
$$A_i(x) = \left\{ \begin{array}{ll}
\nu & \mbox{if $n_{r-1} < i \leq \frac{n_r+n_{r-1}}{2}$};\\
\theta & \mbox{if $\frac{n_r+n_{r-1}}{2} < i \leq n_r$}.\end{array} \right.$$
Therefore,
$$\begin{aligned}
\frac{1}{h_r^\alpha}\displaystyle\sum_{i \in I_r} M_i \left( \frac{\vert A_i(x) \vert}{\rho^{(i)}} \right)
&=&
\frac{1}{h_r^\alpha} \displaystyle\sum_{n_{r-1} < i \leq \frac{(n_r+n_{r-1})}{2}} M_i \left( \frac{\nu}{\rho^{(i)}} \right)\\
&<&
\frac{1}{h_r^\alpha} \frac{1}{2^{r-1}} \left[ \frac{n_r+n_{r-1}}{2} - n_{r-1} \right]\\
&=&
\frac{1}{2^r} \rightarrow 0 \mbox{~as~} r \rightarrow \infty.\end{aligned}$$
Thus we have $x \in w_\theta^{\alpha 0} (A, \mathscr{M})$.\
But, $$\begin{aligned}
\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \left\vert \left\{ i \in I_r: \displaystyle\sum_{i \in I_r} M_i \left(\frac{\vert A_i(x) \vert}{\rho^{(i)}} \right) \geq \varepsilon \right\} \right\vert
&=&
\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \left \vert \left \{ i \in \left( n_{r-1}, \frac{n_r + n_{r-1}}{2} \right) : \displaystyle\sum_{i \in I_r} M_i \left( \frac{\nu}{\rho^{(i)}} \right) \geq \varepsilon \right\} \right\vert\\
&=&
\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \frac{n_r-n_{r-1}}{2} \\
&=&
\frac{1}{2}.\end{aligned}$$
So, $x \notin f_\theta^\alpha(A, \mathscr{M})$.\
[**[Theorem 3.7. ]{}**]{} Let $\mathscr{M}=(M_i)$ be a Musielak-Orlicz function. Then $f_\theta^\alpha(A, \mathscr{M}) \subset w_\theta^\alpha(A, \mathscr{M})$ if and only if $\displaystyle\sup_\nu \displaystyle\sup_i M_i \left(\frac{\nu}{\rho^{(i)}}\right) < \infty$.\
Proof: Let $x \in f_\theta^\alpha(A, \mathscr{M})$. Suppose $h(\nu)=\displaystyle\sup_i M_i \left(\frac{\nu}{\rho^{(i)}} \right)$ and $h=\displaystyle\sup_\nu h(\nu)$. Let
$$I_r^2=\left\{i \in I_r: M_i \left(\frac{\vert A_i(x)-L \vert}{\rho^{(i)}} \right) < \varepsilon \right\}.$$
Now, $M_i(\nu) \leq h$ for all $i, \nu >0$. So, $$\begin{aligned}
\frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i \left(\frac{\vert A_i(x)-L \vert}{\rho^{(i)}} \right)
&= &
\frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r^1} M_i \left(\frac{\vert A_i(x)-L \vert}{\rho^{(i)}} \right)\\
&+&
\frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r^2} M_i \left(\frac{\vert A_i(x)-L \vert}{\rho^{(i)}} \right)\\
&\leq &
h \frac{1}{h_r^\alpha} \vert A_0(\varepsilon) \vert + h(\varepsilon).\end{aligned}$$
Hence, as $\varepsilon \rightarrow 0$, it follows that $x \in w_\theta^\alpha(A, \mathscr{M})$.\
Conversely, suppose that $$\displaystyle\sup_\nu \displaystyle\sup_i M_i \left( \frac{\nu}{\rho^{(i)}} \right) =\infty.$$
Then, we have $$0< \nu_1< \nu_2 <...< \nu_{r-1}< \nu_r<...$$
so that $M_{n_r} \left(\frac{\nu_r}{\rho^{(i)}} \right) \geq h_r^\alpha$ for $r \geq 1$. For $A=I$, we set a sequence $x=(x_i)$ by, $$A_i(x) = \left\{ \begin{array}{ll}
\nu_r & \mbox{if $i=n_r$ for some $r=1,2,..$};\\
\theta & \mbox{otherwise}.\end{array} \right.$$
Then,
$$\begin{aligned}
\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \left\vert \left\{ i \in I_r : \displaystyle\sum_{i \in I_r} M_i \left(\frac{\vert A_i(x) \vert}{\rho^{(i)}} \right) \geq \varepsilon \right\} \right \vert
&=&
\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \\
&=&
0\end{aligned}$$
Hence, $x \in f_\theta^\alpha(A, \mathscr{M})$.\
But, $$\begin{aligned}
\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} \displaystyle\sum_{i \in I_r} M_i\left( \frac{\vert A_i(x)-L \vert}{\rho^{(i)}}\right)
&=&
\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} M_{n_r}\left(\frac{\vert \nu_r- L\vert}{\rho^{(i)}}\right)\\
&\geq &
\displaystyle\lim_{r \rightarrow \infty} \frac{1}{h_r^\alpha} h_r^\alpha\\
&=&1\end{aligned}$$
So, $x \in w_\theta^\alpha(A, \mathscr{M})$.
[100]{}
A. Aizpuru, M.C. List$\acute{a}$n-Garc$\acute{i}$ and F. Rambla-Barreno, Density by Moduli and Statistical Convergence, [*[Quaestiones Mathematicae]{}*]{} 2014: 1-6. http://dx.doi.org/10.2989/16073606.2014.894682
A. Esi and E. Savas, On Lacunary Statistically Convergent Triple Sequences in Probabilistic Normed Space, [*[Applied Mathematics and Information Sciences]{}*]{}, 9(5)(2015), 2529-2534.
B. C. Tripathy and S. Borgohain , Statistically Convergent Difference Sequence spaces of Fuzzy Real numbers defined by Orlicz Function, [*[Thai Journal of Mathematics]{}*]{} , 11(2) (2013), 357-370.
E. Kolk, The statistical convergence in Banach spaces, [*[Acta Et Commentationes Univ. Tartuensis]{}*]{}, 928(1991), 41-52.
E. Savas and S. Borgohain , On strongly almost lacunary statistical $A$-convergence defined by Musielak-Orlicz function. (coomunicated)
E. Savas, On $I$-lacunary Statistical Convergence of Order $\alpha$ for Sequences of Sets, [*[Filomat]{}*]{}, 29(6)(2015), 1223–1229.
E. Savas, Double almost lacunary statistical convergence of order $\alpha$., [*[ Advances in Difference Equations]{}*]{} (254)(2013), 10 pp.
E. Savas, On lacunary statistically convergent double sequences of fuzzy numbers, [*[Applied Mathematics Letters]{}*]{}, 21(2-21)(2008), 134-141.
E. Savas, On $I$-asymptotically lacunary statistical equivalent sequences, [*[Advances in Difference Equations]{}*]{}, 11 (2013), 0-0
E. Savas and R. Savas, On some sequence spaces and lacunary $\sigma$-statistical convergence, [*[Mathematical and Computational Applications]{}*]{}, 8(2)(2003).
E. Savas and R. F. Patterson, Lacunary statistical convergence of multiple sequences, [*[Appl. Math. Lett.]{}*]{} 19(6)(2006), 527–534.
E. Savas and H. Gumus, A generalization on $I$-asymptotically lacunary statistical equivalent sequences, [*[Journal of Inequalities and Applications]{}*]{}, 270(2013), 9 pp.
E. Savas and R. F. Patterson, On $I$-asymptotically lacunary statistical equivalent sequences, [*[Central European Journal of Mathematics]{}*]{} 4(4)(2006), 648-655.
E. Savas and V. Karakaya, Some new sequence spaces defined by lacunary sequences, [*[Mathematica Slovaca]{}*]{}, 57(4)(2007), 393-399.
S. Eren and E. SAVAS, Double lacunary statistical convergence of order $\alpha$,[*[ Indian Journail of Mathematics]{}*]{}, 57(1)(2015), 0-0
H. Fast, Sur la convergence statistique, [*[ Colloq. Math.]{}*]{} 2(1951), 241-244.
H. Nakano, Concave modulars, [*[J. Math. Soc. Japan.]{}*]{}, 5 (1953), 29–49.
I. J. Maddox , Sequence Spaces Defined by a modulus,[*[ Mathematical Proceedings of the Cambridge Philosophical Society]{}*]{}, 100 (1986), 161-166.
I.J. Schoenberg, The integrability of certain functions and related summability methods, [*[Amer. Math. Monthly]{}*]{}, 66(1959), 361-375.
J.A. Fridy, On statistical convergence, Analysis, 5(1985), 301-313.
Lindenstrauss, J. and Tzafriri,L., On Orlicz sequence Spaces, [*[Israel Journal of Mathematics]{}*]{}, 10(1971), 379-390.
Fridy, J.A. and Orhan, C., Lacunary statistically convergence, [*[ Pacific Journal of Mathematics]{}*]{}, 160(1)(1993), 43-51.
J. Connor, M.Ganichev and V. Kadets, A characterization of Banach spaces with separable duals via weak statistical convergence, [*[J. Math. Anal. Appl.]{}*]{}, 244 (2000), 251-261.
J. Connor and E. SAVAS, Lacunary statistical and sliding window convergence for measurable functions, [*[Acta Mmathematica Hungarica]{}*]{}, 145(2)(2015), 416-432.
J. Musielak, Orlicz spaces and modular spaces, [*[Lecture Notes in Mathematics]{}*]{}, Springer, Berlin, Germany, 1034(1983).
M. Mursaleen, A. Alotaibi and S. K. Sharma, Some new lacunary strong convergent vector-valued sequence spaces, [*[Abstract and Applied Analysis]{}*]{}, volume 2014, Article ID 858504, 8 pages.
M. Mursaleen, A. Alotaibi and S. K. Sharma, Some new lacunary strong convergent vector-valued sequence spaces, [*[ Abstract and Applied Analysis]{}*]{}, Volume 2014, Article ID 858504, 8 pages.
Patterson, F. and Savas, E., On asymptotically lacunary statistical equivalent sequences. [*[Thai J. Math.]{}*]{}, 4 (2) (2006), 267–272.
T. $\check{S}$al$\grave{a}$t, On statistically convergent sequences of fuzzy real numbers, [*[Math. Slovaca]{}*]{}, 30(1980), 139-150.
U. Ulusu and E. Savas, An extension of asymptotically lacunary statistical equivalence set sequences, [*[Journal of Inequalities and Applications]{}*]{}, 1(2014), 0-0
V. Karakaya and E. Savas, On almost $p$-bounded variation of lacunary sequences, [*[Computers and Mathematics with Applications]{}*]{} 61(6)(2011), 1502-1506.
W. H. Ruckle, FK spaces in which the sequence of coordinate vectors is bounded, [*[Canad. J. Math.]{}*]{}, 25 (1973), 973–978.
[^1]: The work of the authors was carried under the Post Doctoral Fellow under National Board of Higher Mathematics, DAE, project No. NBHM/PDF.50/2011/64
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper discusses Parisian ruin problem with capital injection for Lévy insurance risk process. Capital injection takes place at the draw-down time of the surplus process when it drops below a pre-specified function of its last record maximum. The capital is continuously paid to keep the surplus above the draw-down level until either the surplus process goes above the record high or a Parisian type ruin occurs, which is announced at the first instance the surplus process has undergone an excursion below the record for an independent exponential period of time consecutively since the time the capital was first injected. Some distributional identities concerning the excursion are presented. Firstly, we give the Parisian ruin probability and the joint Laplace transform (possibly killed at the first passage time above a fixed level of the surplus process) of the ruin time, surplus position at ruin, and the total capital injection at ruin. Secondly, we obtain the $q$-potential measure of the surplus process killed at Parisian ruin. Finally, we give expected present value of the total discounted capital payments up to the Parisian ruin time. The results are derived using recent developments in fluctuation and excursion theory of spectrally negative Lévy process and are presented semi explicitly in terms of the scale function of the Lévy process. Some numerical examples are given to facilitate the analysis of the impact of initial surplus and frequency of observation period to the ruin probability and to the expected total capital injection.'
address: 'Budhi Surya: School of Mathematics and Statistics, Victoria University of Wellington, New Zealand; Email address: [email protected]. Wenyuan Wang: School of Mathematical Sciences, Xiamen University, Fujian 361005, People’s Republic of China; Email address: [email protected]. Xianghua Zhao: School of Mathematical Sciences, Qufu Normal University, Qufu 273165, China; Email address: [email protected]. Xiaowen Zhou: Department of Mathematics and Statistics, Concordia University, Montreal, Quebec, Canada H3G 1M8; Email address: [email protected].'
author:
- 'Budhi Surya, Wenyuan Wang, Xianghua Zhaoand Xiaowen Zhou'
title: 'Parisian excursion with capital injection for draw-down reflected Lévy insurance risk process'
---
Introduction
=============
Parisian ruin has been actively studied since its introduction by Chesney et al. (1997). In contrary to the classical ruin model of Cramér-Lundberg in which case ruin/default occurs at the first instance the underlying (surplus) process crossing below a threshold, Parisian ruin is announced at the first time the process has undertaken an excursion below the default level for a fixed consecutive period of time. It has been applied in finance, among others, by Francois and Morellec (2004), Broadie et al. (2007), and recently by Antill and Grenadier (2019), in particular for studying firm’s optimal capital structure in the presence of Chapters 7 and 11 (default and reorganization proceeding). The default level may be determined endogenously in the sense of Leland and Toft (1996) by maximizing the firm’s equity value. Under Chapter 11 the firm is granted a dilution period during which the firm is given the opportunity to operate and reorganize itself until its asset value goes above the default level, or otherwise liquidated. In their recent work, Palmowski et al. (2020) revisit the Leland-Toft model under spectrally negative Lévy process, discussed earlier in Hilberink and Rogers (2002) and Kyprianou and Surya (2007), by considering information of the firm’s asset only available periodically at Poisson time. They showed using the results of Albrecher et al. (2016) that under Poisson observation of the firm’s assets, nonzero credit spreads holds for firm with lower initial endowment than default level, and in particular, the default time corresponds to the Parisian ruin time with an independent exponential excursion period.
It was introduced to insurance/actuarial science literature started by the work of Dassios and Wu (2008) for compound Poisson process, then extended to spectrally negative Lévy process subsequently by Czarna and Palmowski (2011), Loeffen et al. (2013), among others. Further distributional identities concerning Parisian ruin with independent exponential excursion period were discussed in Baurdoux et al. (2016). The results are generalized to fixed excursion period for a class of penalty functions by Loeffen et al. (2018) which identify known results on Parisian ruin.
It is worth mentioning that the Parisian excursion discussed in the above literature gets started at the first passage below a threshold (default level) of the underlying process. In the past decades some discussions have been developed towards risk protection mechanism against certain financial assets’ outperformance over their last record maximum, also referred to as high-water mark or draw-down. We refer interested readers to the works by Goetzmann et al. (2003) and Agarwal et al. (2009). Default is triggered when the underlying process has gone below a specified level from its last record maximum. Distributional identities regarding first-passage above a threshold for draw-down process were presented in Avram et al. (2004, 2007) and used for pricing Russian options under randomized maturity, and solving optimal dividend problem where the cumulative paid dividend is given by the running supremum of the surplus process. First-passage identities for draw-down process were later extended to a more general form of threshold boundary for e.g. by Zhou (2007), Wang and Zhou (2018), and Li et al. (2019). Parisian excursion below a fixed level from the last record maximum of the surplus process was considered in Surya (2019).
The introduction of capital injection to the firm may prevent the firm from going default at the time its asset value decreases below a certain threshold and the firm needs to meet its commitment to pay dividends to the shareholders. We refer among others to Kulenko and Schmidli (2008), Tao et al. (2011), Avanzi et al. (2011), Bayraktar et al. (2013), and Wang et al. (2019).
In this paper we study Parisian ruin from the last record maximum of surplus process with capital injection. We assume there is no dividend payment in the model. As the source of uncertainty in the surplus process is the downward jumps of the Lévy insurance risk process. Capital is provided to the firm by the stakeholder as soon as the surplus process goes below a given function of the last record maximum of the surplus process. It is continuously provided to the firm to keep the surplus above the draw-down level until the surplus goes back to above the last record or the ruin occurs. If the surplus process has stayed below the record since the first capital injection longer than an independent exponential period of time, the firm faces the credit event and is liquidated. Distributional identities concerning the Parisian ruin are presented.
The rest of the paper is arranged as follows. In Section \[2\] we introduce the model and formulate the problems we are interested. The main results and proofs are presented in Section \[3\]. Section \[4\] presents some numerical examples of the main results. In particular, they are presented to analyze the impact of observation frequency and initial value of surplus to various shapes of ruin probability and expected nett present value of the total capital injection. Section \[5\] concludes this paper.
Spectrally negative Lévy process and its reflected processes
============================================================
\[2\]
Write $X\equiv\{X(t);t\geq0\}$, defined on a probability space $(\Omega, \mathcal{F},(\mathcal{F}_t)_{t\geq 0},\mathbb{P})$ with probability laws $\{\mathbb{P}_{x};x\in\mathbb{R}\}$ denoting the family of probability laws of $X$ such that $X_0=x$, with $\mathbb{P}=\mathbb{P}_0$, and natural filtration $\{\mathcal{F}_{t};t\geq0\}$ of $X$ satisfying the usual conditions of right-continuity and completeness. In particular, we exclude a spectrally negative Lévy process that is not a purely increasing linear drift or the negative of a subordinator. The Lévy-Itô sample paths decomposition of the Lévy process is given by $$\label{eq:LevyIto}
\begin{split}
X_t=\mu t + \sigma B_t &+\int_0^t\int_{\{x<-1\}} x\mathcal{N}(\mathrm{d}x,\mathrm{d}s) + \int_0^t\int_{\{-1\leq x <0\}} x\big(\mathcal{N}(\mathrm{d}x,\mathrm{d}s)-\nu(\mathrm{d}x)\mathrm{d}s\big),
\end{split}$$ where $\mu\in\mathbb{R}$, $\sigma\geq0$ and $(B_t)_{t\geq0}$ is standard Brownian motion, whilst $\mathcal{N}(\mathrm{d}x,\mathrm{d}t)$ denotes the Poisson random measure associated with the jumps process $\Delta X_t:=X_t-X_{t-}$ of $X$. This Poisson random measure has compensator given by $\nu(\mathrm{d}x)\mathrm{d}t$, where $\nu$ is the Lévy measure satisfying the condition: $$\label{eq:intcond}
\int_{-\infty}^0 (1\wedge x^2)\nu(\mathrm{d}x)<\infty.$$ In the expression (\[eq:LevyIto\]), $X$ may define the surplus process of an insurance firm in which $\mu$ represents the premium rate charged on the insurance holder and $\sigma B_t$ is of volatile trading uncertainties which results from the firm investing in financial market. The other two jump components correspond to the compensated small claims and uncompensated big claims from the insurance holder whose distribution is given by the Lévy measure $\nu$ satisfying the integral constraint (\[eq:intcond\]). In the classical model of Cramer-Lundberg risk process with positive drift $c>0$, $c= \mu -\int_{-\infty}^0 x \mathbf{1}_{\{x>-1\}} \nu(\mathrm{d}x)$ with $\nu(\mathrm{d}x)=\beta F(\mathrm{d}x)$ for Poisson claim arrival intensity $\beta>0$ and distribution $F$ of the claim size.
Denote the running supremum process $\overline{X}\equiv\{\sup\limits_{0\leq s\leq t}X(s),\,t\geq0\}$ with $\overline{X}(
0)=x$ under $\mathbb{P}_{x}$. Given a value $a\in\mathbb{R}$, the process $X$ reflected from below at the level $a$ is defined as $$X(t)-(\underline{X}(t)-a)\wedge 0, \,\, t\geq 0$$ where $\underline{X}(t):=\inf_{0\leq s\leq t}X(s)$ with $\underline{X}(
0)=x$ under $\mathbb{P}_{x}$, denotes the running infimum process. Let $\{Y(t), t\geq0\}$ be the process $X$ reflected from below at the level $0$ (cf., Pistorius (2004)).
The draw-down time associated to a draw-down function $\xi$ on $(-\infty,\infty)$ satisfying $\xi(x)<x, \,\, x\in(-\infty,\infty)$, the $\xi$-draw-down time in short, is defined as $$\tau_{\xi}\equiv \tau_{\xi}(X):=\inf\{t\geq0: X(t)<\xi(\overline{X}(t))\}$$ with the convention $\inf\emptyset:=\infty$. We define the process $X$ reflected at the $\xi$-draw-down time $\tau_{\xi}$ as $$X(t)-\mathbf{1}_{[\tau_\xi,\infty)}(t)\left(\inf_{\tau_\xi\leq s\leq t}{X}(s)- \xi(\overline{X}(\tau_\xi))\right)\wedge 0, \,\, t\geq 0,$$ where we call $\xi(\overline{X}(\tau_\xi))$ the draw-down level at the draw-down time $\tau_\xi$.
We now define the draw-down reflected process $U$ for $X$. Intuitively, the process $U$ initially agrees with $X$ until the first draw-down time of $U$. Then it starts to evolve according to $X$ reflected at the draw-down level until the next draw-down time of $U$ when it is reflected at the draw-down level again, and so on. Then given that $U(s)=\overline{U}(s):=\sup_{0\leq t\leq s}U(t)$, the process $\{U(t);t\geq s\}$ evolves without reflection until the next draw-down time $\tau_\xi$; and given that $\overline{U}(s)> U(s)$, the process $\{U(t); t\geq s\}$ is reflected from below at the current draw-down level $\xi(\overline{U}(s))$ until it comes back to the level $\overline{U}(s)$. Note that the process $U$ is not a Markov process in general, but the process $(U,\overline{U})$ is Markovian. Write $\mathbb{P}_{x,y}$ and $\mathbb{E}_{x,y}$ for the law of $(U,\overline{U})$ such that $U(0)=x$ and $\overline{U}(0)=y$. For simplicity, denote $\mathbb{P}_{x}=\mathbb{P}_{x,x}$ and $\mathbb{E}_{x}=\mathbb{E}_{x,x}$.
To be more precise, define $T_0:=0$ and $U(T_0):=X(0) $. Suppose first that for $n\geq 1$, $U(t)$ has been defined on $[0, T_n]$ for $T_n<\infty$, $n\geq 1$. Let $X_{n+1}$ be an independent copy of $X$ starting at $U(T_n)$ and $U_{n+1}$ be the process $X_{n+1}$ reflected at its $\xi$-draw-down time $\tau_{\xi}({X}_{n+1})$. If $\tau_{\xi}(X_{n+1})=\infty$, let $T_{n+1}:=\infty$, and if $\tau_{\xi}(X_{n+1})<\infty$, let $$T_{n+1}:=T_n+\inf\{t\geq 0: U_{n+1}(t)>\overline{X}_{n+1}(\tau_{\xi}(X_{n+1})) \},$$ where $\overline{X}_{n+1}(t):=\sup\limits_{0\leq s\leq t}X_{n+1}(s)$. Observe that $T_{n+1}<\infty$ if $\tau_{\xi}(X_{n+1})<\infty $. Then define $$U(T_n+t):=U_{n+1}(t) \quad\text{for}\quad t\in [0, T_{n+1}-T_n) \quad\text{and}\quad U(T_{n+1}):=U_{n+1}(T_{n+1}-T_n)\quad\text{if}\quad T_{n+1}<\infty.$$ Suppose now that $U(t)$ has been defined on $[0, T_n=\infty)$ for $ n\geq 0$. For convenience, let $T_{n+1}:=\infty$. For the well-definedness of the process $U$, we are referred to Wang and Zhou (2019).
For the process $X$, define its first up-crossing time of level $b\in(-\infty,\infty)$ and first down-crossing time of level $c\in(-\infty,\infty)$, respectively, by $$\begin{aligned}
\tau^{+}_{b}:=\inf\{t\geq0: X(t)>b\}\,\,\, \text{and}\,\,\, \tau_{c}^{-}:=\inf\{t\geq0: X(t)<c\}.\nonumber\end{aligned}$$
For the processes $Y$ and $U$, their first up-crossing times of $b\in(-\infty,\infty)$ are defined respectively by $$\begin{aligned}
\sigma^{+}_{b}:=\inf\{t\geq0: Y(t)>b\} \,\,\, \text{and} \,\,\,\kappa_{b}^{+}:=\inf\{t\geq0: U(t)>b\}.\nonumber\end{aligned}$$ The Parisian ruin time $\theta_{\lambda}$ of the process $U$ is defined as $$\begin{aligned}
N_{\lambda}:=\inf\{n\geq 1: \,T_{n}-T_{n-1}-\tau_{\xi}(X_{n})>e_{\lambda}^{(n)}\},\quad \theta_{\lambda}:=T_{N_{\lambda}-1}+\tau_{\xi}(X_{N_{\lambda}})+e_{\lambda}^{(N_{\lambda})},\nonumber\end{aligned}$$ where $\{e_{\lambda}^{(n)};n\geq 1\}$ is a sequence of i.i.d. exponential randoms with parameter $\lambda>0$, and is independent of $X$.
Due to the absence of positive jumps, it is therefore sensible to define $$\begin{aligned}
\psi(\theta):=\ln \mathbb{E}_{x}\left(\mathrm{e}^{\theta (X_{1}-x)}\right)=\mu\theta+\frac{1}{2}\sigma^{2}\theta^{2}+\int_{(-\infty,0)}\left(\mathrm{e}^{\theta x}-1-\theta x\mathbf{1}_{(-1,0)}(x)\right)\nu(\mathrm{d}x),\nonumber\end{aligned}$$ It is known that $\psi(\theta)$ is finite for $\theta\in[0,\infty)$ in which case it is strictly convex and infinitely differentiable. As in Bertoin (1996), the $q$-scale functions $\{W_{q};q\geq0\}$ of $X$ are defined as follows. For each $q\geq0$, $W_{q}:\,[0,\infty)\rightarrow[0,\infty)$ is the unique strictly increasing and continuous function with Laplace transform $$\begin{aligned}
\int_{0}^{\infty}\mathrm{e}^{-\theta x}W_{q}(x)\mathrm{d}x=\frac{1}{\psi(\theta)-q},\quad \mbox{for }\theta>\Phi(q),\label{eq:scale}\end{aligned}$$ where $\Phi(q)$ is the largest solution of the equation $\psi(\theta)=q$. Further define $W_{q}(x)=0 $ for $x<0$, and write $W$ for the $0$-scale function $W_{0}$.
It is known that $W_{q}(0)=0 $ if and only if process $X$ has sample paths of unbounded variation. If $X$ has sample paths of unbounded variation, or if $X$ has sample paths of bounded variation and the Lévy measure has no atoms, then the scale function $W_{q}$ is continuously differentiable over $(0, \infty)$. By Loeffen (2008), if $X$ has a Lévy measure which has a completely monotone density, then $W_{q}$ is twice continuously differentiable over $(0, \infty)$ when $X$ is of unbounded variation. Moreover, if process $X$ has a nontrivial Gaussian component, then $W_{q}$ is twice continuously differentiable over $(0, \infty)$. Below is an example of the scale function $W_q$ for downward jump-diffusion process with exponentially distributed jumps, which will be used for numerical examples discussed in Section 4.
[cc]{}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Scale function $W_q(x)$ and $W^{\Phi(q)}(x)$ for downward jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$. []{data-label="fig:SF"}](Wqx.pdf "fig:"){width="40.00000%"}
\#2
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
&
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Scale function $W_q(x)$ and $W^{\Phi(q)}(x)$ for downward jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$. []{data-label="fig:SF"}](Wx.pdf "fig:"){width="40.00000%"}
\#2
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Scale function $W_q(x)$ and $W^{\Phi(q)}(x)$ for downward jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$. []{data-label="fig:SF"}](WqxBV.pdf "fig:"){width="40.00000%"}
\#2
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
&
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Scale function $W_q(x)$ and $W^{\Phi(q)}(x)$ for downward jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$. []{data-label="fig:SF"}](WxBV.pdf "fig:"){width="40.00000%"}
\#2
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
\[ex:Ex1\]Consider one-sided jump-diffusion process $X$ with $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for all $\theta\in\mathbb{R}$ s.t. $\theta\neq -c$. It is known that the inverse of the Laplace transform (\[eq:scale\]) for $q>0$ is $$\begin{aligned}
\label{eq:ScaleF}
W_q(x)=\frac{\mathrm{e}^{-\beta_2 x}}{\psi^{\prime}(-\beta_2)} + \frac{\mathrm{e}^{-\beta_1 x}}{\psi^{\prime}(-\beta_1)} + \frac{\mathrm{e}^{\Phi(q)x}}{\psi^{\prime}(\Phi(q))}, \quad \forall x\geq 0,\end{aligned}$$ where $-\beta_1$, $-\beta_2$, and $\Phi(q)$ denotes three roots of $\psi(\theta)=q$ s.t. $-\beta_2<-c<-\beta_1<0<\Phi(q)$. It is straightforward to check by taking Laplace transform that $W_q(x)=\mathrm{e}^{\Phi(q)x}W^{\Phi(q)}(x)$ where $$\begin{aligned}
\label{eq:ScaleF2}
W^{\Phi(q)}(x)=\frac{\mathrm{e}^{-(\beta_2+\Phi(q)) x}}{\psi^{\prime}(-\beta_2)} + \frac{\mathrm{e}^{-(\beta_1+\Phi(q)) x}}{\psi^{\prime}(-\beta_1)} + \frac{1}{\psi^{\prime}(\Phi(q))}, \quad \forall x\geq 0,\end{aligned}$$ with $W^{\Phi(q)}(x)=0$ for $x<0$. It is known that $\psi^{\prime}(-\beta_2)<0$, $\psi^{\prime}(-\beta_1)<0$ and $\psi^{\prime}(\Phi(u))>0$. In fact $W^{\Phi(q)}(x)$ plays the role of $W(x)$ under the Esscher transform of measure $\mathbb{P}^{\Phi(q)}$ defined by $\frac{\mathrm{d}\mathbb{P}^{\Phi(q)}}{\mathrm{d}\mathbb{P}}\Big\vert_{\mathcal{F}_t}=\mathrm{e}^{\Phi(q) X_t -qt}$. It is straight forward to check that $W^{\Phi(q)}(x)$ is increasing for $x\geq 0$, and so is $W_q(x)$, concave and is bounded from above by $1/\psi^{\prime}(\Phi(q))$.
The interested readers are referred to Chan et al. (2011) and Kuznetsov et al. (2012) for more detailed discussions on the smoothness of scale functions. For results on numerical computation of the scale function, the readers are referred to Surya (2008), Hubalek and Kyprianou (2011) and the references therein.
Further define $$\begin{aligned}
Z_{q}(x):=1+q\int_{0}^{x}W_{q}(z)\mathrm{d}z,\quad x\geq0,\nonumber\end{aligned}$$ and $$\begin{aligned}
Z_{q}(x,\theta):=\mathrm{e}^{\theta x}\left(1-\left(\psi(\theta)-q\right)\int_{0}^{x}\mathrm{e}^{-\theta z}W_{q}(z)\mathrm{d}z\right),\quad \theta\geq0, \,x\geq0,\nonumber\end{aligned}$$ with $Z(x,\theta):=Z_{0}(x,\theta)$, and $$\overline{W}_{q}(x):=\int_{0}^{x}W_{q}(z)\mathrm{d}z,\quad q\geq0, x\geq0,$$ and $$\overline{Z}_{q}(x):=\int_{0}^{x}Z_{q}(z)\mathrm{d}z=x+q\int_{0}^{x}\int_{0}^{z}W_{q}(w)\mathrm{d}w\mathrm{d}z,\quad q\geq0, x\geq0.$$
In the sequel, without loss of generality we assume $X_{1}\equiv X$. By Li et al. (2017), we have $$\begin{aligned}
\label{part1}
\mathbb{E}_{x}(\mathrm{e}^{-q\kappa_{b}^{+}}\mathbf{1}_{\{\kappa_{b}^{+}<\tau_{\xi}\}})=
\mathbb{E}_{x}\left(\mathrm{e}^{-q\tau_{b}^{+}}\mathbf{1}_{\{\tau_{b}^{+}<\tau_{\xi}\}}\right)
=\exp\left(-\int_{x}^{b}\frac{W_{q}^{\prime}(\overline{\xi}\left(z\right))}
{W_{q}(\overline{\xi}\left(z\right))}\mathrm{d}z\right),\end{aligned}$$ where $\overline{\xi}(z)=z-\xi(z)$. For $x\in[0,b]$ and $q\geq0$, from Proposition 2 in Pistorius (2004) we have $$\begin{aligned}
\label{two.sid.exit.Y}
\mathbb{E}_{x}(\mathrm{e}^{-q\sigma^{+}_{b}})=\frac{Z_{q}(x)}{Z_{q}(b)}.\end{aligned}$$
By Kyprianou (2006), the resolvent measure corresponding to $X$ is absolutely continuous with respect to the Lebesgue measure with density given by $$\begin{aligned}
\label{h2}
\hspace{-0.3cm}&&\hspace{-0.3cm}\int_{0}^{\infty}\mathrm{e}^{-qt}\mathbb{P}_{x}(X(t)\in \mathrm{d}y;t<\tau_{c}^{-}\wedge \tau_{b}^{+})\mathrm{d}t
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\left(\frac{W_{q}(x-c)}{W_{q}(b-c)}W_{q}(b-y)-W_{q}(x-y)\right)\mathbf{1}_{(c,b)}(y)\mathrm{d}y,\end{aligned}$$ for $x\in(c, b)$. By Pistorius (2004), the resolvent measure corresponding to $Y$ is also absolutely continuous with respect to the Lebesgue measure and has a version of density given by $$\begin{aligned}
\label{reso.meas.Y}
\hspace{-0.3cm}&&\hspace{-0.3cm}\int_{0}^{\infty}\mathrm{e}^{-qt}\mathbb{P}_{x}(Y(t)\in \mathrm{d}y,t<\sigma^{+}_{b})\mathrm{d}t
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}\left(\frac{Z_{q}(x)}{Z_{q}(b)}W_{q}(b-y)-W_{q}(x-y)\right)
\mathbf{1}_{[0,b)}(y)\mathrm{d}y,\end{aligned}$$ where $x\in[0, b)$.
Define the total amount of capital injections made until time $t$ for the draw-down reflected process as $$\begin{aligned}
R(t)\hspace{-0.3cm}&:=&\hspace{-0.3cm}
-\sum_{k=1}^{N-1}\mathbf{1}_{[T_{k-1}+\tau_{\xi}(X_k),\infty)}(t)\left( \inf_{\tau_\xi(X_k)\leq s\leq T_k\wedge t-T_{k-1}}{X}_{k}(s)-\xi(\overline{X}_{k}(\tau_{\xi}(X_k)))\right)\wedge 0.
\nonumber
$$ where $N:=\inf\{n: T_n=\infty\}=\inf\{n: \tau_\xi(X_{n})=\infty\}$.
In this paper, we are interested in evaluating:
- Expectation of the net present value of principal payment of one unit at time $\kappa_b^+\wedge\theta_{\lambda}$: $$\begin{aligned}
U_{\xi}(x;b)=\mathbb{E}_x\Big(\mathrm{e}^{-q(\kappa_b^+\wedge\theta_{\lambda})}\Big). \nonumber\end{aligned}$$
- The joint Laplace transform of $\kappa_{b}^{+}\wedge \theta_{\lambda}$, $U(\kappa_{b}^{+}\wedge \theta_{\lambda})$ and $R(\kappa_{b}^{+}\wedge \theta_{\lambda})$, i.e. $$\begin{aligned}
\quad \quad G_\xi(x;b)
\hspace{-0.3cm}&
=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{-q \left(\kappa_{b}^{+}\wedge \theta_{\lambda}\right)+u U\left(\kappa_{b}^{+}\wedge \theta_{\lambda}\right)-v R\left(\kappa_{b}^{+}\wedge \theta_{\lambda}\right)}\right),\quad q,\lambda, u,v\in \mathbb{R}_{+},\, b\in\mathbb{R},\, x\in(-\infty,b]
.\nonumber\end{aligned}$$
- The potential measure of $U$ involving the Parisian ruin time, i.e. $$\begin{aligned}
\label{}
\int_{0}^{\infty}\mathrm{e}^{-qt}\mathbb{P}_{x}\left(U(t)\in \mathrm{d}u,t<\kappa_{b}^{+}\wedge \theta_{\lambda}\right)\mathrm{d}t,\quad q,\lambda\in \mathbb{R}_{+},\, b\in\mathbb{R},\,x,u\in(-\infty,b].
\nonumber\end{aligned}$$
- The expectation of the total discounted capital injections until $\kappa_{b}^{+}\wedge \theta_{\lambda}$, i.e. $$\begin{aligned}
V_{\xi}(x;b)\hspace{-0.3cm}&=&\hspace{-0.3cm}\mathbb{E}_{x}\left(\int_{0}^{\kappa_{b}^{+}\wedge \theta_{\lambda}}\mathrm{e}^{-q t}\mathrm{d}R(t)\right),\quad q,\lambda\in \mathbb{R}_{+},\, b\in\mathbb{R},\,x\in(-\infty,b].
\nonumber\end{aligned}$$
We also briefly recall concepts in excursion theory for the reflected process $\{\overline{X}(t)-X(t);t\geq0\}$, and we refer to Bertoin (1996) for more details. For $x\in(-\infty,\infty)$, the process $\{L(t):= \overline{X}(t)-x, t\geq0\}$ serves as a local time at $0$ for the Markov process $\{\overline{X}(t)-X(t);t\geq0\}$ under $\mathbb{P}_{x}$. Let the corresponding inverse local time be defined as $$L^{-1}(t):=\inf\{s\geq0: L(s)>t\}=\sup\{s\geq0: L(s)\leq t\}.$$ Further let $L^{-1}(t-):=\lim\limits_{s\uparrow t}L^{-1}(s)$. Define a Poisson point process $\{(t, e_{t}); t\geq0\}$ as $$e_{t}(s):=X(L^{-1}(t))-X(L^{-1}(t-)+s), \,\,s\in(0,L^{-1}(t)-L^{-1}(t-)],$$ whenever the lifetime of $e_{t}$ is positive, i.e. $L^{-1}(t)-L^{-1}(t-)>0$. Whenever $L^{-1}(t)-L^{-1}(t-)=0 $, define $e_{t}:=\Upsilon$ with $\Upsilon$ being an additional isolated point. A result of Itô states that $e$ is a Poisson point process with characteristic measure $n$ if $\{\overline{X}(t)-X(t);t\geq0\}$ is recurrent; otherwise $\{e_{t}; t\leq L(\infty)\}$ is a Poisson point process stopped at the first excursion of infinite lifetime. Here, $n$ is a measure on the space $\mathcal{E}$ of excursions, i.e. the space $\mathcal{E}$ of càdlàg functions $f$ satisfying $$\begin{aligned}
&&f:\,(0,\zeta)\rightarrow (0,\infty)\,\quad \mbox{for some } \zeta=\zeta(f)\in(0,\infty],
\nonumber\\
&&f:\,\{\zeta\}\rightarrow (0,\infty)\,\,\,\,\,\,\quad \mbox{if } \zeta<\infty,
\nonumber\end{aligned}$$ where $\zeta=\zeta(f)$ is the excursion length or lifetime; see Definition 6.13 of Kyprianou (2006) for the definition of $\mathcal{E}$. Denote by $\varepsilon(\cdot)$, or $\varepsilon$ for short, a generic excursion belonging to the space $\mathcal{E}$ of canonical excursions. The excursion height of a canonical excursion $\varepsilon$ is denoted by $\overline{\varepsilon}=\sup\limits_{t\in[0,\zeta]}\varepsilon(t)$. The first passage time of a canonical excursion $\varepsilon$ is defined by $$\rho_{b}^{+}\equiv\rho_{b}^{+}(\varepsilon) :=\inf\{t\in[0,\zeta]: \varepsilon(t)>b\},$$ with the convention $\inf\emptyset:=\zeta$.
Denote by $\varepsilon_{g}$ the excursion (away from $0$) with left-end point $g$ for the reflected process $\{\overline{X}(t)-X(t);t\geq0\}$, and by $\zeta_{g}$ and $\overline{\varepsilon}_{g}$ the excursion’s lifetime and the excursion’s height, respectively; see Section IV.4 of Bertoin (1996).
Main results
============
\[3\]
In this section we present the main results concerning the general draw-down reflected process $U$ with Parisian stopping. For preparation, we recall the following result of Wang and Zhou (2019).
\[lemma2\] Given $\theta,\,q\in(0,\infty)$ and measurable function $\phi:\,(-\infty,\infty)\rightarrow(-\infty,\infty)$, we have $$\begin{aligned}
\label{10}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{\xi}}\,\mathrm{e}^{\theta X(\tau_{\xi})}\,\phi\left(\overline{X}(\tau_{\xi})\right); \tau_{\xi}<\tau_{b}^{+}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\int_{x}^{b}\phi\left(s\right)\mathrm{e}^{\theta \xi(s)}
\exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}\left(z\right))}
{W_{q}(\overline{\xi}\left(z\right))}\mathrm{d}z\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
\times\left(\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}Z_{q}(\overline{\xi}(s),\theta )-\theta Z_{q}(\overline{\xi}(s),\theta )-(q-\psi(\theta ))W_{q}(\overline{\xi}(s))\right)
\mathrm{d}s,\quad x\in(-\infty, b].\end{aligned}$$ In particular, we have $$\begin{aligned}
\label{12}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{\xi}}\,\phi\left(\overline{X}(\tau_{\xi})\right); \tau_{\xi}<\tau_{b}^{+}\right)
=
\int_{x}^{b}\phi\left(s\right)
\exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}\left(z\right))}
{W_{q}(\overline{\xi}\left(z\right))}\mathrm{d}z\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{3.8cm}
\times\left(\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}Z_{q}(\overline{\xi}(s))-qW_{q}(\overline{\xi}(s))\right)
\mathrm{d}s,\quad x\in(-\infty, b]
,\end{aligned}$$ and $$\begin{aligned}
\label{13}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{\theta X(\tau_{\xi})}\,\phi\left(\overline{X}(\tau_{\xi})\right); \tau_{\xi}<\tau_{b}^{+}\right)
=
\int_{x}^{b}\phi\left(s\right)\mathrm{e}^{\theta \xi(s)}
\exp\left(-\int_{x}^{s}\frac{W^{\prime}(\overline{\xi}\left(z\right))}
{W^{}(\overline{\xi}\left(z\right))}\mathrm{d}z\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{3cm}
\times\left(\frac{W^{\prime}(\overline{\xi}(s))}{W^{}(\overline{\xi}(s))}Z^{}(\overline{\xi}(s),\theta )-\theta Z^{}(\overline{\xi}(s),\theta )+\psi(\theta )W^{}(\overline{\xi}(s))\right)
\mathrm{d}s,\quad x\in(-\infty, b]
,\end{aligned}$$ and $$\begin{aligned}
\label{14}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{\xi}}\left(\xi\left(\overline{X}(\tau_{\xi})\right)-X(\tau_{\xi})\right); \tau_{\xi}<\tau_{b}^{+}\right)
=\int_{x}^{b}
\exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}\left(z\right))}
{W_{q}(\overline{\xi}\left(z\right))}\mathrm{d}z\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{0.3cm}
\times\left(Z_{q}(\overline{\xi}(s))-\psi^{\prime}(0+)W_{q}(\overline{\xi}(s))-\frac{\overline{Z}_{q}(\overline{\xi}(s))-\psi^{\prime}(0+)
\overline{W}_{q}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}W_{q}^{\prime}(\overline{\xi}(s))\right)\mathrm{d}s,\quad x\in(-\infty, b]
.\end{aligned}$$
We start with the Laplace transform of the upper exiting time for the process $U$.
\[3.1\] For $q,\lambda\in(0,\infty)$ we have $$\begin{aligned}
\label{upper.lower.boun.resu.}
\mathbb{E}_{x}\left(\mathrm{e}^{-q\kappa_{b}^{+}}\mathbf{1}_{\{\kappa_{b}^{+}<\theta_{\lambda}\}}\right)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\exp\left(-\int_{x}^{b}\ell_{1}(w)\mathrm{d}w\right),\quad x\in(-\infty,b],\end{aligned}$$ where $$\begin{aligned}
\ell_{1}(w)=\frac{W_{q}^{\prime}(\overline{\xi}(w))}
{W_{q}(\overline{\xi}(w))}
\left(1-\frac{Z_{q}(\overline{\xi}(w))}{Z_{q+\lambda}(\overline{\xi}(w))}\right)
+\frac{qW_{q}(\overline{\xi}(w))}{Z_{q+\lambda}(\overline{\xi}(w))}.\nonumber\end{aligned}$$
Denote by $f(x)$ the left hand side of (\[upper.lower.boun.resu.\]). We have $$\begin{aligned}
\label{part.}
f(x)=\mathbb{E}_{x}\left(\mathrm{e}^{-q\kappa_{b}^{+}}\mathbf{1}_{\{\kappa_{b}^{+}<\tau_{\xi}\}}\right)
+\mathbb{E}_{x}\left(\mathrm{e}^{-q\kappa_{b}^{+}}\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+}<\theta_{\lambda}\}}\right),\quad x\in(-\infty,b].\end{aligned}$$ Note that by definition, $\tau_{\xi}<\kappa_{b}^{+}$ implies $\overline{X}(\tau_{\xi})<b$ which further implies $T_{1}<\kappa_{b}^{+}$. Hence, taking use of and , we get for $x\in(-\infty,b]$ $$\begin{aligned}
\label{part2}
\hspace{-0.3cm}&&\hspace{-0.3cm}\mathbb{E}_{x}\left(\mathrm{e}^{-q\kappa_{b}^{+}}\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+}<\theta_{\lambda}\}}\right)
=\mathbb{E}_{x}\left(\mathrm{e}^{-q\kappa_{b}^{+}}\mathbf{1}_{\{\tau_{\xi}<T_{1}<\kappa_{b}^{+}<\theta_{\lambda}\}}
\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}} \left[\left.\mathbb{E}_{}\left(\mathrm{e}^{-q \sigma^{+}_{z}}\mathbf{1}_{\{\sigma^{+}_{z}<e_{\lambda}\}}\right)\right|_{z=\overline{\xi}(\overline{X}(\tau_{\xi}))}\right]f(\overline{X}(\tau_{\xi}))\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}} \frac{f(\overline{X}(\tau_{\xi}))}{Z_{q+\lambda}(\overline{\xi}(\overline{X}(\tau_{\xi})))}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\int_{x}^{b}\frac{f(s)}{Z_{q+\lambda}(\overline{\xi}(s))}
\exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}(z))}{ W_{q}(\overline{\xi}(z))}\mathrm{d}z\right)
\left(\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}Z_{q}(\overline{\xi}(s))-qW_{q}(\overline{\xi}(s))\right)\mathrm{d}s.\end{aligned}$$ Combining (\[part1\]), (\[part.\]) and (\[part2\]), we obtain for $x\in(-\infty, b]$ $$\begin{aligned}
\label{f(x)}
f(x)\hspace{-0.3cm}&=&\hspace{-0.3cm}\exp\left(-\int_{x}^{b}\frac{W_{q}^{\prime}(\overline{\xi}\left(z\right))}
{W_{q}(\overline{\xi}\left(z\right))}\mathrm{d}z\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.5cm}
+\int_{x}^{b}\frac{f(s)}{Z_{q+\lambda}(\overline{\xi}(s))}
\exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}(z))}{ W_{q}(\overline{\xi}(z))}\mathrm{d}z\right)
\left(\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}Z_{q}(\overline{\xi}(s))-qW_{q}(\overline{\xi}(s))\right)\mathrm{d}s.\end{aligned}$$ Taking derivative on both sides of (\[f(x)\]) with respect to $x$, we have $$\begin{aligned}
\label{f'(x)}
f^{\prime}(x)\hspace{-0.3cm}&=&\hspace{-0.3cm}
\frac{W_{q}^{\prime}(\overline{\xi}\left(x\right))}
{W_{q}(\overline{\xi}\left(x\right))}
f(x)
-\frac{f(x)}{Z_{q+\lambda}(\overline{\xi}(x))}\left(\frac{W_{q}^{\prime}(\overline{\xi}(x))}{W_{q}(\overline{\xi}(x))}Z_{q}(\overline{\xi}(x))-qW_{q}(\overline{\xi}(x))\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\ell_{1}(x)f(x),\quad x\in(-\infty,b].\end{aligned}$$ Solving (\[f’(x)\]) we obtain for $x\in(-\infty,b]$ $$\begin{aligned}
\label{fwithC}
f(x)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
C\exp\left(-\int_{x}^{b}\ell_{1}(w)\mathrm{d}w\right),\end{aligned}$$ for some constant $C$. The boundary condition $f(b)=1$ together with (\[fwithC\]) yields (\[upper.lower.boun.resu.\]).
For $q,\lambda\in (0,\infty)$, we have for $x\in(-\infty,b]$, $$\begin{aligned}
\label{eq:main1}
\mathbb{E}_x\Big(\mathrm{e}^{-q\theta_{\lambda}}\mathbf{1}_{\{\theta_{\lambda}<\kappa_b^+\}}\Big)=\int_x^b \exp\Big(-\int_x^y \ell_{1}(w)dw\Big)\overline{\ell}_{1}(y)\mathrm{d}y,\end{aligned}$$ where $\ell_{1}(x)$ is given in (\[upper.lower.boun.resu.\]), while the function $\overline{\ell}_{1}(x)$ is defined by $$\begin{aligned}
\overline{\ell}_{1}(x)
=
\frac{\lambda}{q+\lambda}\left(1 - \frac{1}{Z_{q+\lambda}(\overline{\xi}(x))} \right)\left(\frac{W_q^{\prime}(\overline{\xi}(x))}{W_q(\overline{\xi}(x))}Z_q(\overline{\xi}(x)) - qW_q(\overline{\xi}(x))\right).\end{aligned}$$
Denote by $\overline{f}_{\xi}(x;b)$ the left hand side of . We have $$\begin{aligned}
\overline{f}_{\xi}(x;b) \hspace{-0.3cm}&=&\hspace{-0.3cm}\mathbb{E}_x\Big( \mathrm{e}^{-q \theta_{\lambda}}\mathbf{1}_{\{\theta_{\lambda}<\kappa_b^+, \tau_{\xi}<\tau_b^+\}} \Big) \nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm} \mathbb{E}_x\Big( \mathrm{e}^{-q \theta_{\lambda}}\mathbf{1}_{\{\theta_{\lambda}<\kappa_b^+, \tau_{\xi}<\tau_b^+, T_1-\tau_{\xi} > e_{\lambda}\}} \Big) + \mathbb{E}_x\Big( \mathrm{e}^{-q \theta_{\lambda}}\mathbf{1}_{\{\theta_{\lambda}<\kappa_b^+, \tau_{\xi}<\tau_b^+,T_1-\tau_{\xi} \leq e_{\lambda}\}} \Big). \label{eq:eq1}\end{aligned}$$ By the strong Markov property of the bi-variate process $(U,\overline{U})$ one can obtain $$\begin{aligned}
\mathbb{E}\Big( \mathrm{e}^{-q\theta_{\lambda}}\mathbf{1}_{\{\theta_{\lambda}<\kappa_b^+, \tau_{\xi}<\tau_b^+, T_1-\tau_{\xi} > e_{\lambda}\}}\Big\vert \mathcal{F}_{\tau_{\xi}}\Big)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_b^+\}}
\left[\left.\mathbb{E}_{}\Big(\mathrm{e}^{-q e_{\lambda}}\mathbf{1}_{\{ e_{\lambda}<\sigma_{z}^+\}}\Big)\right|_{z=\overline{\xi}(\overline{X}_{\tau_{\xi}})}\right] \nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_b^+\}}\frac{\lambda}{q+\lambda}\Big(1-\Big[\mathbb{E}\Big(\mathrm{e}^{-(q+\lambda)\sigma_z^+}\Big)\Big\vert_{z=\overline{\xi}(\overline{X}_{\tau_{\xi}})}\Big] \Big) \nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_b^+\}}\frac{\lambda}{q+\lambda}\Big(1-\frac{1}{Z_{q+\lambda}(\overline{\xi}(\overline{X}_{\tau_{\xi}}))} \Big), \label{eq:eq1a}\end{aligned}$$ and $$\begin{aligned}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\mathbb{E}\Big(\mathbb{E}\Big( \mathrm{e}^{-q\theta_{\lambda}}\mathbf{1}_{\{\theta_{\lambda}<\kappa_b^+, \tau_{\xi}<\tau_b^+, T_1-\tau_{\xi} \leq e_{\lambda}\}}\Big\vert \mathcal{F}_{T_{1}}\Big)\Big\vert \mathcal{F}_{\tau_{\xi}}\Big)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_b^+\}}
\left[\left.\mathbb{E}_{}\Big(\mathrm{e}^{-q \sigma_{z}^+}\mathbf{1}_{\{ \sigma_{z}^+\leq e_{\lambda}\}}\Big)\right|_{z=\overline{\xi}(\overline{X}_{\tau_{\xi}})}\right]\mathbb{E}_{\overline{X}_{\tau_{\xi}}}\Big(\mathrm{e}^{-q\theta_{\lambda}}\mathbf{1}_{\{\theta_{\lambda}<\kappa_b^+\}}\Big)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_b^+\}}\frac{\overline{f}_{\xi}(\overline{X}_{\tau_{\xi}};b)}{Z_{q+\lambda}(\overline{\xi}(\overline{X}_{\tau_{\xi}}))}\label{eq:eq1b}.\end{aligned}$$ Following the two identities (\[eq:eq1a\]) and (\[eq:eq1b\]), we obtain from (\[eq:eq1\]) and (\[12\]) the equation $$\begin{aligned}
&&\overline{f}_{\xi}(x;b)=\frac{\lambda}{q+\lambda}\mathbb{E}_x\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_b^+\}}\Big(1-\frac{1}{Z_{q+\lambda}(\overline{\xi}(\overline{X}_{\tau_{\xi}}))} \Big) \right) \nonumber\\
&&\hspace{2cm}+ \mathbb{E}\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_b^+\}}\frac{\overline{f}_{\xi}(\overline{X}_{\tau_{\xi}};b)}{Z_{q+\lambda}(\overline{\xi}(\overline{X}_{\tau_{\xi}}))}\right) \nonumber\\
&&= \frac{\lambda}{q+\lambda} \int_x^b \Big[ 1- \frac{1}{Z_{q+\lambda}(\overline{\xi}(s))} \Big]\exp\left(-\int_x^s \frac{W_q^{\prime}(\overline{\xi}(z))}{W_q(\overline{\xi}(z))}dz\right)\left(\frac{W_q^{\prime}(\overline{\xi}(s))}{W_q(\overline{\xi}(s))} Z_q(\overline{\xi}(s)) - q W_q(\overline{\xi}(s)) \right)\mathrm{d}s\\
&& \hspace{1cm}+ \int_x^b\frac{\overline{f}_{\xi}(s;b)}{Z_{q+\lambda}(\overline{\xi}(s))}\exp\left(-\int_x^s \frac{W_q^{\prime}(\overline{\xi}(z))}{W_q(\overline{\xi}(z))}dz\right)\left(\frac{W_q^{\prime}(\overline{\xi}(s))}{W_q(\overline{\xi}(s))} Z_q(\overline{\xi}(s)) - q W_q(\overline{\xi}(s)) \right)\mathrm{d}s.\end{aligned}$$ After taking partial derivative w.r.t $x$ on both sides, we have by the Leibniz integral rule, $$\begin{aligned}
\overline{f}_{\xi}^{\prime}(x;b)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
-\frac{\lambda}{q+\lambda}\Big[1-\frac{1}{Z_{q+\lambda}(\overline{\xi}(x))}\Big]
\left(\frac{W_q^{\prime}(\overline{\xi}(x))}{W_q(\overline{\xi}(x))} Z_q(\overline{\xi}(x)) - q W_q(\overline{\xi}(x)) \right) \nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-\frac{\overline{f}_{\xi}(x;b)}{Z_{q+\lambda}(\overline{\xi}(x))}\left(\frac{W_q^{\prime}(\overline{\xi}(x))}{W_q(\overline{\xi}(x))} Z_q(\overline{\xi}(x)) - q W_q(\overline{\xi}(x)) \right) + \frac{W_q^{\prime}(\overline{\xi}(x))}{W_q(\overline{\xi}(x))}\overline{f}_{\xi}(x;b) \nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\ell_1(x)\overline{f}_{\xi}(x;b) - \overline{\ell}_{1}(x)\label{eq:eq3}.\end{aligned}$$ Identity (\[eq:main1\]) follows by solving the differential equation (\[eq:eq3\]) subject to the boundary condition $\overline{f}_{\xi}(b;b)=0$.
For $\lambda\in (0,\infty)$, the Parisian ruin probability is given by $$\begin{aligned}
\label{eq:ruinprobab}
\mathbb{P}_x\big(\theta_{\lambda}<\infty)=1-\exp\Big(-\int_x^{\infty} \ell_1(w;\lambda)dw\Big), \quad x\in\mathbb{R} ,\end{aligned}$$ with $$\begin{aligned}
\ell_1(x;\lambda)=\frac{W^{\prime}(\overline{\xi}(x))}{W(\overline{\xi}(x))}\left(1-\frac{1}{Z_{\lambda}(\overline{\xi}(x))}\right).\end{aligned}$$
For $q,\lambda\in(0,\infty)$, we have for $x\in(-\infty,b]$ that $$\begin{aligned}
U_{\xi}(x;b)=\int_x^b \exp\Big(-\int_x^y \ell_{1}(w)dw\Big)\overline{\ell}_{1}(y)\mathrm{d}y + \exp\Big(-\int_x^b \ell_{1}(w)dw\Big).\end{aligned}$$
The proof follows from combining the two results (\[upper.lower.boun.resu.\]) and (\[eq:main1\]).
The next result gives an expression of the joint Laplace transform concerning $\kappa_{b}^{+}\wedge \theta_{\lambda}$.
\[3.4\] For any $q, u,v,\lambda\in(0,\infty)$ with $u\in(0,\Phi_{q+\lambda})$, we have $$\begin{aligned}
\label{lap.of.exp.tot.dis.cap.inj.}
\quad\quad G_\xi(x;b)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\left(\mathrm{e}^{ub}-\int_{x}^{b}\overline{\ell}_{2}(z)\,\exp\left(\int_{z}^{b}\ell_{2}(w)\mathrm{d}w\right)\mathrm{d}z\right)\exp\left(-\int_{x}^{b}\ell_{2}(w)\mathrm{d}w\right),\quad x\in(-\infty, b],\end{aligned}$$ where $$\begin{aligned}
\ell_{2}(w)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\frac{W_{q}^{\prime}(\overline{\xi}(w))}
{W_{q}(\overline{\xi}(w))}
\left(1-\frac{Z_{q}(\overline{\xi}(w),v)}
{Z_{q+\lambda}(\overline{\xi}(w),v)}\right)+
\frac{v Z_{q}(\overline{\xi}(w),v)+(q-\psi(v ))W_{q}(\overline{\xi}(w))}
{Z_{q+\lambda}(\overline{\xi}(w),v)},
\nonumber\\
\overline{\ell}_{2}(w)\hspace{-0.3cm}&=&\hspace{-0.3cm}
-\mathrm{e}^{u\xi(w)}
\hbar(\overline{\xi}(w))
\left(\frac{W_{q}^{\prime}(\overline{\xi}(w))}{W_{q}(\overline{\xi}(w))}Z_{q}(\overline{\xi}(w),v)-v Z_{q}(\overline{\xi}(w),v)-(q-\psi(v ))W_{q}(\overline{\xi}(w))\right)
,\nonumber
\\
\label{hbar}
\hbar(w)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\frac{\lambda}{\Phi_{q+\lambda}-u}
\left(
W_{q+\lambda}(0+)
+\int_{y<0}\mathrm{e}^{u y+(v-u) y}\left(W_{q+\lambda}^{\prime}(-y)-\Phi_{q+\lambda}W_{q+\lambda}(-y)\right)
\mathrm{d}y\right.
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-\frac{1}{Z_{q+\lambda}(w,v)}
\bigg(
W_{q+\lambda}(0+)\,\mathrm{e}^{uw}
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
\left.\left.+\int_{y<w}\mathrm{e}^{u y+(v-u) (y\wedge 0)}\left(W_{q+\lambda}^{\prime}(w-y)-\Phi_{q+\lambda}W_{q+\lambda}(w-y)\right)
\mathrm{d}y\right)\right).\nonumber\end{aligned}$$
Given $q, u,v\in(0,\infty)$ and $x\in(-\infty, b]$, applying the Markov property of the process $(U,\overline{U})$ we have $$\begin{aligned}
\label{add.1}
\hspace{-0.3cm}&&\hspace{-0.3cm}
G_\xi(x;b):=
\mathbb{E}_{x}\left(\mathrm{e}^{-q \left(\kappa_{b}^{+}\wedge \theta_{\lambda}\right)+u U(\kappa_{b}^{+}\wedge \theta_{\lambda})-v R\left(\kappa_{b}^{+}\wedge \theta_{\lambda}\right)}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbb{E}_{x}\left(\left.\mathrm{e}^{-q \,(\tau_{\xi}+e_{\lambda}^{(1)})+u U(\tau_{\xi}+e_{\lambda}^{(1)})-v R(\tau_{\xi}+e_{\lambda}^{(1)})}\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+},\,T_{1}-\tau_{\xi}>e_{\lambda}^{(1)}\}}\right|\mathcal{F}_{\tau_{\xi}}
\right)\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+
\mathbb{E}_{x}\left(\mathbb{E}_{x}\left(\left.\mathrm{e}^{-q \left(\kappa_{b}^{+}\wedge \theta_{\lambda}\right)+u U\left(\kappa_{b}^{+}\wedge \theta_{\lambda}\right)-v R\left(\kappa_{b}^{+}\wedge \theta_{\lambda}\right)}\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+},\,T_{1}-\tau_{\xi}\leq e_{\lambda}^{(1)}\}}\right|\mathcal{F}_{T_{1}}
\right)\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+
\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{b}^{+}+u X(\tau_{b}^{+})-v R(\tau_{b}^{+})}\mathbf{1}_{\{\tau_{b}^{+}<\tau_{\xi}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(
\mathrm{e}^{-q \tau_{\xi}+(u-v)\xi(\overline{X}(\tau_{\xi}))+vX(\tau_{\xi})}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}
\left(\left.\mathbb{E}_{}\left(\mathrm{e}^{-q e_{\lambda}+u Y(e_{\lambda})+v(\underline{X}(e_{\lambda}))}\mathbf{1}_{\{\sigma_{z}^{+}>e_{\lambda}\}}
\right)\right|_{z=\overline{\xi}(\overline{X}(\tau_{\xi}))}\right)\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+
\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+}\}}\mathrm{e}^{-v(\xi(\overline{X}(\tau_{\xi}))-X(\tau_{\xi}))}
\,G_{\xi}(\overline{X}(\tau_{\xi});b)\,\mathbb{E}_{x}\left(\left.
\mathrm{e}^{-v (R(T_{1})-R(\tau_{\xi}))-(q+\lambda) (T_{1}-\tau_{\xi})}\right|\mathcal{F}_{\tau_{\xi}}\right)\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+
\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{b}^{+}+u X(\tau_{b}^{+})-v R(\tau_{b}^{+})}\mathbf{1}_{\{\tau_{b}^{+}<\tau_{\xi}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(
\mathrm{e}^{-q \tau_{\xi}+(u-v)\xi(\overline{X}(\tau_{\xi}))+vX(\tau_{\xi})}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}
\left(\left.\mathbb{E}_{}\left(\mathrm{e}^{-q e_{\lambda}+u Y(e_{\lambda})+v(\underline{X}(e_{\lambda}))}\mathbf{1}_{\{\sigma_{z}^{+}>e_{\lambda}\}}
\right)\right|_{z=\overline{\xi}(\overline{X}(\tau_{\xi}))}\right)\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+
\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+}\}}\mathrm{e}^{vX(\tau_{\xi})}
\mathrm{e}^{-v\xi(\overline{X}(\tau_{\xi}))}
\,G_{\xi}(\overline{X}(\tau_{\xi});b)\left(\left.\mathbb{E}_{}\left(
\mathrm{e}^{v(\underline{X}(\sigma_{z}^{+}))-(q+\lambda) \sigma_{z}^{+}}\right)\right|_{z=\overline{\xi}(\overline{X}(\tau_{\xi}))}\right)\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+
\mathrm{e}^{u b}\,\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{b}^{+}}\mathbf{1}_{\{\tau_{b}^{+}<\tau_{\xi}\}}\right).\end{aligned}$$ In addition, $$\begin{aligned}
\label{w36}
\hbar(z)\hspace{-0.3cm}&:=&\hspace{-0.3cm}\mathbb{E}_{}\left(\mathrm{e}^{-q e_{\lambda}+u Y(e_{\lambda})+v(\underline{X}(e_{\lambda})\wedge 0)}\mathbf{1}_{\{\sigma_{z}^{+}>e_{\lambda}\}}
\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}\mathbb{E}_{}\left(\mathrm{e}^{-q e_{\lambda}+u Y(e_{\lambda})+v(\underline{X}(e_{\lambda})\wedge 0)}\right)
-\mathbb{E}_{}\left(\mathrm{e}^{-q e_{\lambda}+u Y(e_{\lambda})+v(\underline{X}(e_{\lambda})\wedge 0)}\mathbf{1}_{\{\sigma_{z}^{+}<e_{\lambda}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\lambda\,\mathbb{E}_{}\left(
\int_{0}^{\infty}\mathrm{e}^{-(q+\lambda) t}\,\mathrm{e}^{u Y(t)+v(\underline{X}(t)\wedge 0)}
\mathrm{d}t
-\int_{\sigma_{z}^{+}}^{\infty}\mathrm{e}^{-(q+\lambda) t}\,\mathrm{e}^{u Y(t)+v(\underline{X}(t)\wedge 0)}
\mathrm{d}t\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\frac{\lambda}{q+\lambda}\mathbb{E}_{}\left(
\mathrm{e}^{u Y(e_{q+\lambda})+v(\underline{X}(e_{q+\lambda})\wedge 0)}\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-
\lambda\,\mathbb{E}_{}\left(\int_{0}^{\infty}
\mathbb{E}_{}\left(\left.\mathrm{e}^{-(q+\lambda) (\sigma_{z}^{+}+t)}\,\mathrm{e}^{u Y(\sigma_{z}^{+}+t)+v(\underline{X}(\sigma_{z}^{+}+t)\wedge 0)}
\right|\mathcal{F}_{\sigma_{z}^{+}}\right)\mathrm{d}t\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\frac{\lambda}{q+\lambda}\mathbb{E}_{}\left(
\mathrm{e}^{u Y(e_{q+\lambda})+v(\underline{X}(e_{q+\lambda})\wedge 0)}\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-\frac{\lambda}{q+\lambda}\mathbb{E}_{}\left(
\mathrm{e}^{-(q+\lambda) \sigma_{z}^{+}+v(\underline{X}(\sigma_{z}^{+}))}
\,\mathbb{E}_{z}\left(\mathrm{e}^{u Y(e_{q+\lambda})+v(\underline{X}(e_{q+\lambda})\wedge 0)}
\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\frac{\lambda}{q+\lambda}\mathbb{E}_{}\left(
\mathrm{e}^{u X(e_{q+\lambda})+(v-u)(\underline{X}(e_{q+\lambda})\wedge 0)}\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-
\frac{\lambda}{q+\lambda}\,\mathbb{E}_{}\left(
\mathrm{e}^{-(q+\lambda) \sigma_{z}^{+}+v(\underline{X}(\sigma_{z}^{+}))}
\right)
\,\mathbb{E}_{z}\left(\mathrm{e}^{u X(e_{q+\lambda})+(v-u) (\underline{X}(e_{q+\lambda})\wedge 0)}
\right),\quad z\in(0,\infty),\end{aligned}$$ where for the fifth equality we have taken use of $$\begin{aligned}
\mathbb{E}_{}\left(\left.\mathrm{e}^{u Y(\sigma_{z}^{+}+t)+v(\underline{X}(\sigma_{z}^{+}+t)\wedge 0)}
\right|\mathcal{F}_{\sigma_{z}^{+}}\right)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{}\left(\left.\mathrm{e}^{(u-v) Y(\sigma_{z}^{+}+t)+v(X(\sigma_{z}^{+}+t)-X(\sigma_{z}^{+})+z+X(\sigma_{z}^{+})-z)}
\right|\mathcal{F}_{\sigma_{z}^{+}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathrm{e}^{v(X(\sigma_{z}^{+})-z)}
\mathbb{E}_{}\left(\left.\mathrm{e}^{(u-v) Y(\sigma_{z}^{+}+t)+v(X(\sigma_{z}^{+}+t)-X(\sigma_{z}^{+})+z)}
\right|\mathcal{F}_{\sigma_{z}^{+}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathrm{e}^{v(Y(\sigma_{z}^{+})+(\underline{X}(\sigma_{z}^{+})\wedge 0)-z)}
\mathbb{E}_{z}\left(\mathrm{e}^{(u-v) Y(t)+vX(t)}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathrm{e}^{v(\underline{X}(\sigma_{z}^{+})\wedge 0)}
\mathbb{E}_{z}\left(\mathrm{e}^{u Y(t)+v(\underline{X}(t)\wedge 0)}\right),
\nonumber\end{aligned}$$ which holds true since $$\begin{aligned}
Y(\sigma_{z}^{+}+t)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
X(\sigma_{z}^{+}+t)-X(\sigma_{z}^{+})-\inf_{s\leq t}(X(\sigma_{z}^{+}+s)-X(\sigma_{z}^{+}))
\wedge (-Y(\sigma_{z}^{+}))
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
X(\sigma_{z}^{+}+t)-X(\sigma_{z}^{+})+z-\inf_{s\leq t}(X(\sigma_{z}^{+}+s)-X(\sigma_{z}^{+})+z)
\wedge 0
\nonumber\\
\hspace{-0.3cm}&:=&\hspace{-0.3cm}
\widetilde{X}(t)-\underline{\widetilde{X}}(t)\wedge 0,
\nonumber\end{aligned}$$ where $\underline{\widetilde{X}}$ denotes the running infimum process of the process $\widetilde{X}=\{X(\sigma_{z}^{+}+t)-X(\sigma_{z}^{+})+z; t\geq 0\}$ which starts from $z\in(0,\infty)$, is independent of $\mathcal{F}_{\sigma_{z}^{+}}$ and is identical in law to $(X, \mathbb{P}_{z})$.
By adapting (24) in Albrecher et al. (2016) one has $$\begin{aligned}
\label{w37}
\mathbb{E}_{}\left(
\mathrm{e}^{-(q+\lambda) \sigma_{z}^{+}+v(\underline{X}(\sigma_{z}^{+}))}
\right)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{}\left(
\mathrm{e}^{v(\underline{X}(\sigma_{z}^{+}))}
\mathbf{1}_{\{\sigma_{z}^{+}<e_{q+\lambda}\}}
\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\frac{1}{Z_{q+\lambda}(z,v)},\quad z\in(0,\infty).\end{aligned}$$ In addition, by Lemma 1 of Bertoin (1997) with minor adaptation one has $$\begin{aligned}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\mathbb{P}_{z}\left(X(e_{q+\lambda})\in \mathrm{d}x,
\underline{X}(e_{q+\lambda})\geq y\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
(q+\lambda)\left(\mathrm{e}^{-\Phi_{q+\lambda}(x-y)}W_{q+\lambda}(z-y)-\mathbf{1}_{\{z\geq x\}}W_{q+\lambda}(z-x)\right)\mathrm{d}x,\quad y\in(-\infty, x\wedge z],\, z\in(0,\infty),\nonumber\end{aligned}$$ which implies that $$\begin{aligned}
\label{w38}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\mathbb{P}_{z}\left(X(e_{q+\lambda})\in \mathrm{d}x,
\underline{X}(e_{q+\lambda})\in \mathrm{d}y\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
(q+\lambda)\,\mathrm{e}^{-\Phi_{q+\lambda}(x-y)}\left(W_{q+\lambda}^{\prime}(z-y)-\Phi_{q+\lambda}W_{q+\lambda}(z-y)\right)
\mathbf{1}_{\{y<z\}}\mathbf{1}_{\{y\leq x\}}\mathrm{d}x\mathrm{d}y
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+(q+\lambda)\,\mathrm{e}^{-\Phi_{q+\lambda}(x-z)}
W_{q+\lambda}(0+)\mathbf{1}_{\{z\leq x\}}\delta_{z}(\mathrm{d}y) \mathrm{d}x,\quad y\in(-\infty, z],\,x\in [y,\infty),\, z\in(0,\infty).\end{aligned}$$ Combining , and , we have for $z\in(0,\infty)$ $$\begin{aligned}
\hbar(z)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\lambda
\left(W_{q+\lambda}(0+)\int_{0}^{\infty}
\mathrm{e}^{u x}
\mathrm{e}^{-\Phi_{q+\lambda}x} \mathrm{d}x
\right.
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
\left.+\int_{y<0}\int_{x\geq y}\mathrm{e}^{u x+(v-u) (y\wedge 0)}\mathrm{e}^{-\Phi_{q+\lambda}(x-y)}\left(W_{q+\lambda}^{\prime}(-y)-\Phi_{q+\lambda}W_{q+\lambda}(-y)\right)
\mathrm{d}x\mathrm{d}y\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-\frac{\lambda}{Z_{q+\lambda}(z,v)}
\left(W_{q+\lambda}(0+)\int_{z}^{\infty}
\mathrm{e}^{u x}
\mathrm{e}^{-\Phi_{q+\lambda}(x-z)} \mathrm{d}x
\right.
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
\left.+\int_{y<z}\int_{x\geq y}\mathrm{e}^{u x+(v-u) (y\wedge 0)}\mathrm{e}^{-\Phi_{q+\lambda}(x-y)}\left(W_{q+\lambda}^{\prime}(z-y)-\Phi_{q+\lambda}W_{q+\lambda}(z-y)\right)
\mathrm{d}x\mathrm{d}y\right)
.\nonumber\end{aligned}$$ By , and one can rewrite as $$\begin{aligned}
\label{G.int.}
G_\xi(x;b)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(
\mathrm{e}^{-q \tau_{\xi}+vX(\tau_{\xi})}
\mathrm{e}^{(u-v)\xi(\overline{X}(\tau_{\xi}))}
\hbar(\overline{\xi}(\overline{X}(\tau_{\xi})))
\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+
\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{\xi}+vX(\tau_{\xi})}
\frac{\mathrm{e}^{-v\xi(\overline{X}(\tau_{\xi}))}G_{\xi}(\overline{X}(\tau_{\xi});b)}{Z_{q+\lambda}(\overline{\xi}(\overline{X}(\tau_{\xi})),v)}
\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+}\}}
\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+
\mathrm{e}^{u b}\,\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{b}^{+}}\mathbf{1}_{\{\tau_{b}^{+}<\tau_{\xi}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\int_{x}^{b}\left(\mathrm{e}^{u\xi(s)}
\hbar(\overline{\xi}(s))+
\frac{G_{\xi}(s;b)}{Z_{q+\lambda}(\overline{\xi}(s),v)}
\right)
\exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}\left(z\right))}
{W_{q}(\overline{\xi}\left(z\right))}\mathrm{d}z\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
\times\left(\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}Z_{q}(\overline{\xi}(s),v)-v Z_{q}(\overline{\xi}(s),v)-(q-\psi(v ))W_{q}(\overline{\xi}(s))\right)
\mathrm{d}s
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+\mathrm{e}^{ub}\exp\left(-\int_{x}^{b}\left(\frac{W_{q}^{\prime}(\overline{\xi}\left(z\right))}
{W_{q}(\overline{\xi}\left(z\right))}
\right)\mathrm{d}z\right).\end{aligned}$$ Differentiating with respect to $x$ gives $$\begin{aligned}
\label{G.dif.}
G_{\xi}^{\prime}(x,b)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\ell_{2}(x)G_{\xi}(x,b)+\overline{\ell}_{2}(x).\end{aligned}$$ Solving (\[G.dif.\]) with boundary condition $G_{\xi}(b;b)=\mathrm{e}^{ub}$, we obtain (\[lap.of.exp.tot.dis.cap.inj.\]).
We then obtain an expression of the resolvent density for the process $U$.
\[3.2\] \[reso.meas.U.\] For $q\in (0,\infty)$, the resolvent measure of $U$ is absolutely continuous with respect to the Lebesgue measure with density given by $$\begin{aligned}
\label{resovent.meas.}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\int_{0}^{\infty}\mathrm{e}^{-qt}\mathbb{P}_{x}\left(U(t)\in \mathrm{d}u,t<\kappa_{b}^{+}\wedge \theta_{\lambda}\right)\mathrm{d}t
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
W_{q}(0)
\exp\left(-\int_{x}^{y}
\ell_{1}(w)
\mathrm{d}w\right)
\mathbf{1}_{(x,b)}(u)\mathrm{d}u
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+\int_{x}^{b}
\exp\left(-\int_{x}^{y}
\ell_{1}(w)
\mathrm{d}w\right)
\ell_{3}(y,u)
\mathbf{1}_{(\xi(y),y)}(u)\mathrm{d}y \mathrm{d}u,
\quad x,u\in(-\infty, b],\end{aligned}$$ where $\ell_{1}$ is defined as in Theorem \[3.1\], and $$\begin{aligned}
\ell_{3}(y,u)\hspace{-0.3cm}&=&\hspace{-0.3cm}\left(\frac{W_{q}^{\prime}(\overline{\xi}(y))}{W_{q}(\overline{\xi}(y))}Z_{q}(\overline{\xi}(y))-qW_{q}(\overline{\xi}(y))\right)
\frac{Z_{\lambda}(u-\xi(y))
W_{q}(y-u)}
{Z_{\lambda}(\overline{\xi}(y))Z_{q+\lambda}(\overline{\xi}(y))}
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+
W_{q}^{\prime}(y-u)-\frac{W_{q}^{\prime}(\overline{\xi}(y))}{W_{q}(\overline{\xi}(y))}W_{q}(y-u).\nonumber\end{aligned}$$
Recall $\overline{U}(t)=\sup_{s\in[0,t]}U(s)$ and let $e_{q}$ be an exponential random variable independent of $X$. For $q>0$, $x\leq b$ and any continuous, non-negative and bounded function $h$, let $$\begin{aligned}
\label{gene.reso.meas.}
\hspace{-0.3cm}&&\hspace{-0.3cm}qg(x):=\int_{0}^{\infty}q \mathrm{e}^{-qt}\mathbb{E}_{x}\left(h(U(t));t<\kappa_{b}^{+}\wedge \theta_{\lambda}\right)\mathrm{d}t
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(h(X(e_{q}))\mathbf{1}_{\{X(e_{q})<\overline{X}(e_{q}),\,e_{q}<\tau_{b}^{+}\wedge \tau_{\xi}\}}\right)
+\mathbb{E}_{x}\left(h(U(e_{q}))\mathbf{1}_{\{U(e_{q})<\overline{U}(e_{q}),\tau_{\xi}<e_{q}<\kappa_{b}^{+}\wedge \theta_{\lambda}\}}\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+\mathbb{E}_{x}\left(\int_{0}^{\infty}q \mathrm{e}^{-qt}h(X(t))\mathbf{1}_{\{X(t)=\overline{X}(t),\,t<\tau_{b}^{+}\wedge \tau_{\xi}\}}\mathrm{d}t\right)
+\mathbb{E}_{x}\left(h(U(e_{q}))\mathbf{1}_{\{U(e_{q})=\overline{U}(e_{q}),\tau_{\xi}<e_{q}<\kappa_{b}^{+}\wedge \theta_{\lambda}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=:&\hspace{-0.3cm}\,
qg_{1}(x)+qg_{2}(x)+qg_{3}(x)+qg_{4}(x).\nonumber\end{aligned}$$
Note that $\int_{0}^{t}\mathbf{1}_{\{X(s)=\overline{X}(s)\}}\mathrm{d}s=W_{q}(0) \,\overline{X}(t)$ under $\mathbb{P}_{0}$; see Chapters IV and VII of Bertoin (1996), the proof of Part (ii) of Theorem 1 in Pistorius (2004) or the first three paragraphs in Section 5 of Li et al. (2019). By we have $$\begin{aligned}
\label{qg3}
qg_{3}(x)\hspace{-0.3cm}&=&\hspace{-0.3cm}\mathbb{E}_{x}\left(\int_{0}^{\infty}q \mathrm{e}^{-qL^{-1}(L(t))}h(X(L^{-1}(L(t))))\mathbf{1}_{\{X(t)=\overline{X}(t),\,L^{-1}(L(t))<\tau_{b}^{+}\wedge \tau_{\xi}\}}\mathrm{d}t\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
W(0)\mathbb{E}_{x}\left(\int_{0}^{\infty}q \mathrm{e}^{-qL^{-1}(L(t))}h(X(L^{-1}(L(t))))\mathbf{1}_{\{L^{-1}(L(t))< \tau_{b}^{+}\wedge \tau_{\xi}\}}\mathrm{d}L_{t}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q W(0)\int_{0}^{b-x} \mathbb{E}_{x}\left(\mathrm{e}^{-qL^{-1}(t)}\mathbf{1}_{\{L^{-1}(t)< \tau_{\xi}\}}\right)h(x+t)\mathrm{d}t
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q W(0)\int_{x}^{b} \exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}\left(z\right))}
{W_{q}(\overline{\xi}\left(z\right))}\mathrm{d}z\right)h(s)\mathrm{d}s,\end{aligned}$$ where we have used the fact that $L^{-1}(t)$ has the same law as the first exit time $\tau_{x+t}^{+}$ under $\mathbb{P}_{x}$.
By the strong Markov property of $(U,\overline{U})$, the definition of $T_{1}=\inf\{t\geq 0: U(t)>\overline{X}(\tau_{\xi}) \}$ (i.e., $\tau_{\xi}<T_{1}$ holds implicitly) given in the definition of $U$ in Section \[2\], the memoryless property of the exponentially distributed random variable, as well as and , one has $$\begin{aligned}
\label{25}
\hspace{-0.2cm}
qg_{4}(x)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\left.\mathbb{E}_{x}\left(h(U(e_{q}))\mathbf{1}_{\{U(e_{q})=\overline{U}(e_{q}),\tau_{\xi}<T_{1}\leq e_{q}<\kappa_{b}^{+}\wedge \theta_{\lambda}\}}\right|\mathcal{F}_{T_{1}}\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbf{1}_{\{T_{1}<e_{q}\wedge \kappa_{b}^{+}\wedge \theta_{\lambda}\}}
\mathbb{E}_{\overline{X}(\tau_{\xi})}\left(h(U(e_{q}))\mathbf{1}_{\{U(e_{q})=\overline{U}(e_{q}),e_{q}<\kappa_{b}^{+}\wedge \theta_{\lambda}\}}\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbb{E}_{x}\left(\left.\mathrm{e}^{-q T_{1}}\mathrm{e}^{-\lambda (T_{1}-\tau_{\xi})}\mathbf{1}_{\{T_{1}< \kappa_{b}^{+}\}}\left(qg_{3}(\overline{X}(\tau_{\xi}))+qg_{4}(\overline{X}(\tau_{\xi}))\right)\right|\mathcal{F}_{\tau_{\xi}}\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q\mathbb{E}_{x}\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+}\}}\frac{1}{Z_{q+\lambda}(\overline{\xi}(\overline{X}(\tau_{\xi})))}\left(g_{3}(\overline{X}(\tau_{\xi}))+g_{4}(\overline{X}(\tau_{\xi}))\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q\int_{x}^{b}\frac{g_{3}(s)+g_{4}(s)}{Z_{q+\lambda}(\overline{\xi}(s))}
\exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}(z))}{ W_{q}(\overline{\xi}(z))}\mathrm{d}z\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
\times\left(\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}Z_{q}(\overline{\xi}(s))-qW_{q}(\overline{\xi}(s))\right)\mathrm{d}s,\end{aligned}$$ where we also used the fact that $\tau_{\xi}<\kappa_{b}^{+}$ implies $T_{1}<\kappa_{b}^{+}$ (see also ), and the fact that $\tau_{\xi}<e_{q}$ combined with $U(e_{q})=\overline{U}(e_{q})$ implies $T_{1}\leq e_{q}$. By the compensation formula, the memoryless property for exponential random variable and , $qg_{1}(x)$ can be expressed as $$\begin{aligned}
\label{qg1.raw.}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\int_{0}^{\infty}\sum_{g}\mathrm{e}^{-qg}\prod\limits_{r<g}\mathbf{1}_{\{\overline{\varepsilon}_{r}\leq \overline{\xi}(x+L(r)),\,L(g)\leq b-x\}}\,h\left(x+L(g)-\varepsilon_{g}(t-g)\right)
\right.
\nonumber\\
\hspace{-0.3cm}&&\hspace{0.5cm}
\left.
\times q\mathrm{e}^{-q (t-g)}\mathbf{1}_{\{g<t<g+\zeta_{g}\wedge \rho_{\overline{\xi}(x+L(g))}^{+}(g)\}}\mathrm{d}t\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\sum_{g}\mathrm{e}^{-qg}
\prod\limits_{r<g}\mathbf{1}_{\{\overline{\varepsilon}_{r}\leq \overline{\xi}(x+L(r)),\,L(g)\leq b-x\}}\right.
\nonumber\\
\hspace{-0.3cm}&&\hspace{0.5cm}
\left.
\times\int_{0}^{\infty}q \mathrm{e}^{-q s}h\left(x+L(g)-\varepsilon_{g}(s)\right)\mathbf{1}_{\{s<\zeta_{g}\wedge\rho_{\overline{\xi}(x+L(g))}^{+}(g)\}}
\mathrm{d}s\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\int_{0}^{\infty}\mathrm{e}^{-qt}\prod\limits_{r<t}\mathbf{1}_{\{\overline{\varepsilon}_{r}\leq \overline{\xi}(x+L(r)),\,L(t)\leq b-x\}}\right.
\nonumber\\
\hspace{-0.3cm}&&\hspace{0.5cm}
\times
\left.\left(\int_{\mathcal{E}}
\int_{0}^{\infty}q \mathrm{e}^{-q s}h\left(x+L(t)-\varepsilon(s)\right)\mathbf{1}_{\{s<\zeta\wedge\rho_{\overline{\xi}(x+L(t))}^{+}\}}
\mathrm{d}s \,n\left(\mathrm{d}\varepsilon\right)\right)\mathrm{d}L(t)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q\int_{0}^{b-x}\mathbb{E}_{x}\left(\mathrm{e}^{-qL_{t-}^{-1}}
\mathbf{1}_{\{L_{t-}^{-1}< \tau_{\xi}\}}\right)
\int_{0}^{\infty}n\left(\mathrm{e}^{-qs}h(x+t-\varepsilon(s))\mathbf{1}_{\{s<\zeta\wedge\rho^{+}_{\overline{\xi}(x+t)}\}}\right)
\mathrm{d}s\mathrm{d}t
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q\int_{x}^{b}\exp\left(-\int_{x}^{t}\frac{W_{q}^{\prime}(\overline{\xi}(z))}{ W_{q}(\overline{\xi}(z))}\mathrm{d}z\right)\int_{0}^{\infty}
n\left(\mathrm{e}^{-qs}h(t-\varepsilon(s))\mathbf{1}_{\{s<\zeta\wedge\rho^{+}_{\overline{\xi}(t)}\}}\right)\mathrm{d}s\mathrm{d}t,\end{aligned}$$ where $g$ is the left-end point of the excursion $\varepsilon_{g}$, as introduced at the end of Section 2. Applying the same arguments as in (\[qg3\]) and (\[qg1.raw.\]) we have $$\begin{aligned}
\label{h1}
\hspace{-0.3cm}&&\hspace{-0.3cm}\mathbb{E}_{x}\left(h(X(e_{q}))\mathbf{1}_{\{e_{q}<\tau_{b}^{+}\wedge \tau_{c}^{-}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(h(X(e_{q}))\mathbf{1}_{\{X(e_{q})=\overline{X}(e_{q}),\,e_{q}<\tau_{b}^{+}\wedge \tau_{c}^{-}\}}\right)+\mathbb{E}_{x}\left(h(X(e_{q}))\mathbf{1}_{\{X(e_{q})<\overline{X}(e_{q}),\,e_{q}<\tau_{b}^{+}\wedge \tau_{c}^{-}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q\int_{x}^{b}\frac{W_{q}(x-c)}{ W_{q}(t-c)}\left(W(0)h(t)+\int_{0}^{\infty}n\left(\mathrm{e}^{-qs}h(t-\varepsilon(s))\mathbf{1}_{\{s<\zeta\wedge\rho^{+}_{t-c}\}}\right)\mathrm{d}s\right)\mathrm{d}t,\end{aligned}$$ where the identity $$\mathbb{E}_{x-c}\left(\mathrm{e}^{-q \tau_{t-c}^{+}};\tau_{t-c}^{+}<\tau_{0}^{-}\right)=\frac{W_{q}(x-c)}{ W_{q}(t-c)},\quad -\infty<c\leq x\leq t<\infty,$$ is used. Equating the right hand sides of (\[h1\]) and (\[h2\]) and then differentiating the resulting equation with respect to $b$ gives $$\begin{aligned}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\frac{W_{q}(x-c)}{ W_{q}(b-c)}\left(W(0)h(b)+\int_{0}^{\infty}n\left(\mathrm{e}^{-qs}h(b-\varepsilon(s))\mathbf{1}_{\{s<\zeta\wedge\rho^{+}_{b-c}\}}\right)\mathrm{d}s\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\frac{W_{q}(x-c)}{ W_{q}(b-c)}\left(h(b)W(0)+\int_{c}^{b}h(y)\left(W_{q}^{\prime}(b-y)-\frac{W_{q}^{\prime}(b-c)}{W_{q}(b-c)}W_{q}(b-y)\right)\mathrm{d}y\right),\nonumber\end{aligned}$$ or equivalently, $$\begin{aligned}
\label{impo.iden.for.n.01}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\int_{0}^{\infty}n\left(\mathrm{e}^{-qs}h(b-\varepsilon(s))\mathbf{1}_{\{s<\zeta\wedge\rho^{+}_{b-c}\}}\right)\mathrm{d}s
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\int_{c}^{b}h(y)\left(W_{q}^{\prime}(b-y)-\frac{W_{q}^{\prime}(b-c)}{W_{q}(b-c)}W_{q}(b-y)\right)\mathrm{d}y.\end{aligned}$$ Combining (\[impo.iden.for.n.01\]) and (\[qg1.raw.\]), we get $$\begin{aligned}
\label{qg1}
g_{1}(x)\hspace{-0.3cm}&=&\hspace{-0.3cm}\int_{x}^{b}\mathrm{e}^{-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}(z))}{ W_{q}(\overline{\xi}(z))}\mathrm{d}z}
\int_{\xi(s)}^{s}h(y)\left(W_{q}^{\prime}(s-y)-\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}W_{q}(s-y)\right)\mathrm{d}y
\mathrm{d}s.\end{aligned}$$
Using the memoryless property of exponential random variable, $qg_{2}(x)$ can be rewritten as $$\begin{aligned}
\label{qg2}
qg_{2}(x)\hspace{-0.3cm}&=&\hspace{-0.3cm}\mathbb{E}_{x}\left(h(U(e_{q}))\mathbf{1}_{\{U(e_{q})<\overline{U}(e_{q}),\,\tau_{\xi}<e_{q}<T_{1}<\kappa_{b}^{+}\wedge \theta_{\lambda}\}}\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+\mathbb{E}_{x}\left(h(U(e_{q}))\mathbf{1}_{\{U(e_{q})<\overline{U}(e_{q}),\,\tau_{\xi}<T_{1}<e_{q}<\kappa_{b}^{+}\wedge \theta_{\lambda}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}\mathbb{E}_{x}\left(h(U(e_{q}))\mathbf{1}_{\{\tau_{\xi}<e_{q}<T_{1}<\kappa_{b}^{+}\wedge \theta_{\lambda}\}}\right)
+\mathbb{E}_{x}\left(h(U(e_{q}))\mathbf{1}_{\{U(e_{q})<\overline{U}(e_{q}),\,\tau_{\xi}<T_{1}<e_{q}<\kappa_{b}^{+}\wedge \theta_{\lambda}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=:&\hspace{-0.3cm}\,qg_{21}(x)+qg_{22}(x),\end{aligned}$$ where we also took use of the fact that $\tau_{\xi}<\kappa_{b}^{+}$ implies $T_{1}<\kappa_{b}^{+}$ as in .
Using once again the facts that $\tau_{\xi}<T_{1}$ holds implicitly and $\tau_{\xi}<\kappa_{b}^{+}$ implies $T_{1}<\kappa_{b}^{+}$, we have by (\[reso.meas.Y\]) and $$\begin{aligned}
\label{qg21}
qg_{21}(x)
\hspace{-0.3cm}&=&\hspace{-0.3cm}\mathbb{E}_{x}\left(\mathbb{E}_{x}\left(\left.h(U(e_{q}))\mathbf{1}_{\{\tau_{\xi}<e_{q}<T_{1}<\kappa_{b}^{+}\wedge \theta_{\lambda}\}}\right|\mathcal{F}_{\tau_{\xi}}\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbf{1}_{\{\tau_{\xi}<e_{q}\wedge\kappa_{b}^{+}\}}\mathbb{E}_{x}\left(\left.h(U(e_{q}))\mathbf{1}_{\{e_{q}<T_{1}< \theta_{\lambda}\}}\right|\mathcal{F}_{\tau_{\xi}}\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbf{1}_{\{\tau_{\xi}<e_{q}\wedge\kappa_{b}^{+}\}}\left.\mathbb{E}_{}\left(h(\xi(z)+Y(e_{q}))\mathbf{1}_{\{e_{q}<\sigma^{+}_{\overline{\xi}(z)}<e_{\lambda}\}}\right)\right|_{z=\overline{X}(\tau_{\xi})}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q\mathbb{E}_{x}\left(\mathrm{e}^{-q \tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}\int_{0}^{\overline{\xi}(\overline{X}(\tau_{\xi}))}\frac{Z_{\lambda}(y)\,h(\xi(\overline{X}(\tau_{\xi}))+y)}{Z_{\lambda}(\overline{\xi}(\overline{X}(\tau_{\xi})))}
\frac{W_{q+\lambda}(\overline{\xi}(\overline{X}(\tau_{\xi}))-y)}{Z_{q+\lambda}(\overline{\xi}(\overline{X}(\tau_{\xi})))}\mathrm{d}y\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q\int_{x}^{b}
\exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}(z))}{ W_{q}(\overline{\xi}(z))}\mathrm{d}z\right)
\left(\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}Z_{q}(\overline{\xi}(s))-qW_{q}(\overline{\xi}(s))\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
\,\,\,\,\,\,\,\,\,\times\int_{0}^{\overline{\xi}(s)}\frac{Z_{\lambda}(y)h(\xi(s)+y)
W_{q+\lambda}(\overline{\xi}(s)-y)}
{Z_{\lambda}(\overline{\xi}(s))Z_{q+\lambda}(\overline{\xi}(s))}\mathrm{d}y\mathrm{d}s,\end{aligned}$$ where for the fourth equation the following equation is applied $$\begin{aligned}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\mathbb{E}_{}\left(h(Y(e_{q}))\mathbf{1}_{\{e_{q}<\sigma^{+}_{a}<e_{\lambda}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{}\left(\mathrm{e}^{-\lambda \sigma^{+}_{a}}h(Y(e_{q}))\mathbf{1}_{\{e_{q}<\sigma^{+}_{a}\}}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q\int_{0}^{\infty}\mathrm{e}^{-(q+\lambda)t}\,\mathbb{E}_{}\left(h(Y(t))\mathbf{1}_{\{t<\sigma^{+}_{a}\}}\mathbb{E}_{}\left(\left.\mathrm{e}^{-\lambda \left(\sigma^{+}_{a}-t\right)}\right|\mathcal{F}_{t}\right)\right)
\mathrm{d}t
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q\int_{0}^{\infty}\mathrm{e}^{-(q+\lambda)t}\,\mathbb{E}_{}\left(\frac{Z_{\lambda}(Y(t))}{Z_{\lambda}(a)}\,h(Y(t))\mathbf{1}_{\{t<\sigma^{+}_{a}\}}\right)
\mathrm{d}t, \quad a\in(0,\infty).\nonumber\end{aligned}$$
In addition, observing that $\tau_\xi<T_1$ for $\tau_\xi<\infty$, by (\[two.sid.exit.Y\]) and one can rewrite $qg_{22}(x)$ as $$\begin{aligned}
\label{qg22}
qg_{22}(x)\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbb{E}_{x}\left(\left.h(U(e_{q}))\mathbf{1}_{\{U(e_{q})<\overline{U}(e_{q}),\,\tau_{\xi}<T_{1}<e_{q}<\kappa_{b}^{+}\wedge \theta_{\lambda}\}}\right|\mathcal{F}_{T_{1}}\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbf{1}_{\{T_{1}<e_{q}\wedge\kappa_{b}^{+},\,T_{1}-\tau_{\xi}<e_{\lambda}\}}
\left(qg_{1}(\overline{X}(\tau_{\xi}))+qg_{2}(\overline{X}(\tau_{\xi}))\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbb{E}_{x}\left(\left.\mathrm{e}^{-qT_{1}}
\mathrm{e}^{-\lambda (T_{1}-\tau_{\xi})}\mathbf{1}_{\{T_{1}<\kappa_{b}^{+}\}}
\left(qg_{1}(\overline{X}(\tau_{\xi}))+qg_{2}(\overline{X}(\tau_{\xi}))\right)\right|\mathcal{F}_{\tau_{\xi}}\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}\frac{1}{Z_{q+\lambda}(\overline{\xi}(\overline{X}(\tau_{\xi})))}\left(qg_{1}(\overline{X}(\tau_{\xi}))+qg_{2}(\overline{X}(\tau_{\xi}))\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
q\int_{x}^{b}
\frac{g_{1}(s)+g_{2}(s)}{Z_{q+\lambda}(\overline{\xi}(s))}\mathrm{e}^{-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}(z))}{ W_{q}(\overline{\xi}(z))}\mathrm{d}z}
\left(\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}Z_{q}(\overline{\xi}(s))-qW_{q}(\overline{\xi}(s))\right)\mathrm{d}s.\end{aligned}$$
Combining (\[qg3\]), , (\[qg1\]), (\[qg2\]), (\[qg21\]) and (\[qg22\]), we obtain the following differential equation on $g(x)$ with boundary condition $g(b)=0$. $$\begin{aligned}
\label{g'}
g^{\prime}(x)\hspace{-0.3cm}&=&\hspace{-0.3cm}\frac{W_{q}^{\prime}(\overline{\xi}(x))}{ W_{q}(\overline{\xi}(x))}g(x)- W(0)h(x)
-\frac{g_{3}(x)+g_{4}(x)}{Z_{q+\lambda}(\overline{\xi}(x))}
\left(\frac{W_{q}^{\prime}(\overline{\xi}(x))}{W_{q}(\overline{\xi}(x))}
Z_{q}(\overline{\xi}(x))-qW_{q}(\overline{\xi}(x))\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-\int_{\xi(x)}^{x}h(y)\left(W_{q}^{\prime}(x-y)-\frac{W_{q}^{\prime}(\overline{\xi}(x))}{W_{q}(\overline{\xi}(x))}W_{q}(x-y)\right)\mathrm{d}y
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-\frac{g_{1}(x)+g_{2}(x)}{Z_{q+\lambda}(\overline{\xi}(x))}
\left(\frac{W_{q}^{\prime}(\overline{\xi}(x))}{W_{q}(\overline{\xi}(x))}Z_{q}(\overline{\xi}(x))-qW_{q}(\overline{\xi}(x))\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-\left(\frac{W_{q}^{\prime}(\overline{\xi}(x))}{W_{q}(\overline{\xi}(x))}Z_{q}(\overline{\xi}(x))-qW_{q}(\overline{\xi}(x))\right)
\int_{0}^{\overline{\xi}(x)}\frac{Z_{\lambda}(y)h(\xi(x)+y)
W_{q+\lambda}(\overline{\xi}(x)-y)}
{Z_{\lambda}(\overline{\xi}(x))Z_{q+\lambda}(\overline{\xi}(x))}\mathrm{d}y
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\ell_{1}(x)g(x)- W(0)h(x)
-\int_{\xi(x)}^{x}h(y)\left(W_{q}^{\prime}(x-y)-\frac{W_{q}^{\prime}(\overline{\xi}(x))}{W_{q}(\overline{\xi}(x))}W_{q}(x-y)\right)\mathrm{d}y
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-\left(\frac{W_{q}^{\prime}(\overline{\xi}(x))}{W_{q}(\overline{\xi}(x))}Z_{q}(\overline{\xi}(x))-qW_{q}(\overline{\xi}(x))\right)
\int_{\xi(x)}^{x}\frac{Z_{\lambda}(y-\xi(x))h(y)
W_{q}(x-y)}
{Z_{\lambda}(\overline{\xi}(x))Z_{q+\lambda}(\overline{\xi}(x))}
\mathrm{d}y.\end{aligned}$$ Solving equation (\[g’\]) yields $$\begin{aligned}
\label{eq.gene.reso.meas.}
g(x)\hspace{-0.3cm}&=&\hspace{-0.3cm}\int_{0}^{\infty}\mathrm{e}^{-qt}\mathbb{E}_{x}\left(h(U(t));t<\kappa_{b}^{+}\wedge \theta_{\lambda}\right)\mathrm{d}t
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
W_{q}(0)\int_{x}^{b}h(y)
\exp\left(-\int_{x}^{y}
\ell_{1}(w)
\mathrm{d}w\right)\mathrm{d}y
+\int_{x}^{b}\exp\left(-\int_{x}^{y}
\ell_{1}(w)
\mathrm{d}w\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
\times\left(\int_{\xi(y)}^{y}h(z)\left(W_{q}^{\prime}(y-z)-\frac{W_{q}^{\prime}(\overline{\xi}(y))}{W_{q}(\overline{\xi}(y))}W_{q}(y-z)\right)\mathrm{d}z\right.
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
\left.+
\left(\frac{W_{q}^{\prime}(\overline{\xi}(y))}{W_{q}(\overline{\xi}(y))}Z_{q}(\overline{\xi}(y))-qW_{q}(\overline{\xi}(y))\right)
\int_{\xi(y)}^{y}\frac{Z_{\lambda}(z-\xi(y))h(z)
W_{q}(y-z)}
{Z_{\lambda}(\overline{\xi}(y))Z_{q+\lambda}(\overline{\xi}(y))}
\mathrm{d}z\right) \mathrm{d}y.\end{aligned}$$ The expression (\[resovent.meas.\]) follows immediately from (\[eq.gene.reso.meas.\]).
The following result gives an expression of the expectation of the total discounted capital injections until time $\kappa_{b}^{+}\wedge \theta_{\lambda}$.
\[3.3\] For $q\in (0,\infty)$, we have $$\begin{aligned}
\label{expr.of.exp.tot.dis.cap.inj.}
V_\xi(x;b)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\int_{x}^{b}
\exp\left(-\int_{x}^{y}
\ell_{1}(w)
\mathrm{d}w\right)
\ell_{4}(y)\mathrm{d}y,\quad x\in(-\infty,b],\end{aligned}$$ where $\ell_{1}$ is defined as in Theorem \[3.1\], and $$\begin{aligned}
\ell_{4}(y)\hspace{-0.3cm}&=&\hspace{-0.3cm}Z_{q}(\overline{\xi}(y))-\psi^{\prime}(0+)W_{q}(\overline{\xi}(y))-\frac{\overline{Z}_{q}(\overline{\xi}(y))-\psi^{\prime}(0+)\overline{W}_{q}(\overline{\xi}(y))}{W_{q}(\overline{\xi}(y))}W_{q}^{\prime}(\overline{\xi}(y))
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+
\frac{\overline{Z}_{q+\lambda}(\overline{\xi}(y))-\psi^{\prime}(0+)\overline{W}_{q+\lambda}(\overline{\xi}(y))}{Z_{q+\lambda}(\overline{\xi}(y))}
\left(\frac{W_{q}^{\prime}(\overline{\xi}(y))}{W_{q}(\overline{\xi}(y))}Z_{q}(\overline{\xi}(y))-qW_{q}(\overline{\xi}(y))\right).\nonumber\end{aligned}$$
For $q\in(0,\infty)$ and $x\in(-\infty,b]$, we have $$\begin{aligned}
\label{disc.total.cap.}
V_\xi(x;b)\hspace{-0.3cm}&=&\hspace{-0.3cm}\mathbb{E}_{x}\left(\int_{0}^{\kappa_{b}^{+}\wedge \theta_{\lambda}}\mathrm{e}^{-q t}\mathrm{d}R(t)\right)
=\mathbb{E}_{x}\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}\left(\xi(\overline{X}(\tau_{\xi}))-X(\tau_{\xi})\right)\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+\mathbb{E}_{x}\left(\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}\int_{\tau_{\xi}+}^{T_{1}\wedge \theta_{\lambda}}\mathrm{e}^{-q t}\mathrm{d}R(t)\right)
+\mathbb{E}_{x}\left(\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+}\}}\int_{T_{1}\wedge \theta_{\lambda}}^{\kappa_{b}^{+}\wedge \theta_{\lambda}}\mathrm{e}^{-q t}\mathrm{d}R(t)\right)
\nonumber\\
\hspace{-0.3cm}&=:&\hspace{-0.3cm}V_{1}(x;b)+V_{2}(x;b)+V_{3}(x;b).\end{aligned}$$
By , $V_{1}(x;b)$ can be expressed as $$\begin{aligned}
\label{l1}
V_{1}(x;b)
\hspace{-0.3cm}&=&\hspace{-0.3cm}\int_{x}^{b}
\exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}\left(z\right))}
{W_{q}(\overline{\xi}\left(z\right))}\mathrm{d}z\right)
\Bigg(Z_{q}(\overline{\xi}(s))-\psi^{\prime}(0+)W_{q}(\overline{\xi}(s))
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}\hspace{4cm}
-\frac{\overline{Z}_{q}(\overline{\xi}(s))-\psi^{\prime}(0+)\overline{W}_{q}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}W_{q}^{\prime}(\overline{\xi}(s))\Bigg)\mathrm{d}s.\end{aligned}$$ By the Markov property for the reflected process $(U,\overline{U})$, $$\begin{aligned}
\label{v0(0}
V_{2}(x;b)\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+},\,e_{\lambda}^{(1)}\geq T_{1}-\tau_{\xi}\}}\int_{\tau_{\xi}+}^{T_{1}}\mathrm{e}^{-q t}\mathrm{d}R(t)\right)
+\mathbb{E}_{x}\left(\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+},\,e_{\lambda}^{(1)}<T_{1}-\tau_{\xi}\}}\int_{\tau_{\xi}+}^{\tau_{\xi}+e_{\lambda}^{(1)}}\mathrm{e}^{-q t}\mathrm{d}R(t)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}\left(\mathbb{E}\left(\left.\mathrm{e}^{-\lambda \sigma_{z}^{+}}\int_{0}^{\sigma_{z}^{+}}
\mathrm{e}^{-qt}\mathrm{d}\left(-\underline{X}(t)\right)\right)\right|_{z=\overline{\xi}(\overline{X}(\tau_{\xi}))}
\right)\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
+\mathbb{E}_{x}\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}\left(\mathbb{E}\left(\left.\mathbf{1}_{\{e_{\lambda}<\sigma_{z}^{+}\}}\int_{0}^{e_{\lambda}}
\mathrm{e}^{-qt}\mathrm{d}\left(-\underline{X}(t)\right)\right)\right|_{z=\overline{\xi}(\overline{X}(\tau_{\xi}))}
\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}\left(-\frac{\psi^{\prime}(0+)}{q+\lambda}+\frac{\overline{Z}_{q+\lambda}(\overline{\xi}(\overline{X}(\tau_{\xi})))+\frac{\psi^{\prime}(0+)}{q+\lambda}}{Z_{q+\lambda}(\overline{\xi}(\overline{X}(\tau_{\xi})))}
\right)\right),\end{aligned}$$ where the following equations are applied $$\begin{aligned}
\hspace{-0.3cm}&&\hspace{-0.3cm}
\mathbb{E}\left(\mathbf{1}_{\{e_{\lambda}<\sigma_{z}^{+}\}}\int_{0}^{e_{\lambda}}
\mathrm{e}^{-qt}\mathrm{d}\left(-\underline{X}(t)\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}\left(\int_{0}^{\sigma_{z}^{+}} \int_{0}^{t}\mathrm{e}^{-q s}\mathrm{d}\left(-\underline{X}(s)\right)\mathrm{d}\left(-\mathrm{e}^{-\lambda t}\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
- \mathbb{E}\left(\mathrm{e}^{-\lambda \sigma_{z}^{+}}
\int_{0}^{\sigma_{z}^{+}}\mathrm{e}^{-q t}\mathrm{d}\left(-\underline{X}(t)\right)\right)
+
\mathbb{E}\left(\int_{0}^{\sigma_{z}^{+}}
\mathrm{e}^{-(q+\lambda) t}\mathrm{d}\left(-\underline{X}(t)\right)\right),\quad z\in(0,\infty),\nonumber\end{aligned}$$ and $$\mathbb{E}\left(\int_{0}^{\sigma_{z}^{+}}
\mathrm{e}^{-(q+\lambda)t}\mathrm{d}\left(-\underline{X}(t)\right)\right)=-\frac{\psi^{\prime}(0+)}{q+\lambda}+\frac{\overline{Z}_{q+\lambda}(z)+\frac{\psi^{\prime}(0+)}{q+\lambda}}{Z_{q+\lambda}(z)},\quad z\in(0,\infty),$$ which can be found in the proof of Theorem 1 of Avram et al. (2007). Combining and yields $$\begin{aligned}
\label{l2.}
V_{2}(x;b)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\int_{x}^{b}\left(-\frac{\psi^{\prime}(0+)}{q+\lambda}+\frac{\overline{Z}_{q+\lambda}(\overline{\xi}(s))+\frac{\psi^{\prime}(0+)}{q+\lambda}}{Z_{q+\lambda}(\overline{\xi}(s))}\right)
\exp\left(-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}(z))}{ W_{q}(\overline{\xi}(z))}\mathrm{d}z\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}\,\,\,\,\,\,\,\times
\left(\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}Z_{q}(\overline{\xi}(s))-qW_{q}(\overline{\xi}(s))\right)\mathrm{d}s.\end{aligned}$$
Making use of again, one can get $$\begin{aligned}
\label{l3}
V_{3}(x;b)\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbb{E}_{x}\left(\left.\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+},\,T_{1}-\tau_{\xi}<e_{\lambda}^{(1)}\}}\int_{T_{1}}^{\kappa_{b}^{+}\wedge \theta_{\lambda}}\mathrm{e}^{-q t}\mathrm{d}R(t)\right|\mathcal{F}_{T_{1}}\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathbb{E}_{x}\left(\left.\mathrm{e}^{-q\tau_{\xi}}\,\mathrm{e}^{-(q+\lambda)(T_{1}-\tau_{\xi})}\mathbf{1}_{\{\tau_{\xi}<\kappa_{b}^{+}\}}V_\xi(\overline{X}(\tau_{\xi});b)\right|\mathcal{F}_{\tau_{\xi}}\right)\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\mathbb{E}_{x}\left(\mathrm{e}^{-q\tau_{\xi}}\mathbf{1}_{\{\tau_{\xi}<\tau_{b}^{+}\}}\frac{V_\xi(\overline{X}(\tau_{\xi});b)}{Z_{q+\lambda}(\overline{\xi}(\overline{X}(\tau_{\xi})))}\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\int_{x}^{b}\frac{V_\xi(s;b)}{Z_{q+\lambda}(\overline{\xi}(s))}
\mathrm{e}^{-\int_{x}^{s}\frac{W_{q}^{\prime}(\overline{\xi}(z))}{ W_{q}(\overline{\xi}(z))}\mathrm{d}z}
\left(\frac{W_{q}^{\prime}(\overline{\xi}(s))}{W_{q}(\overline{\xi}(s))}Z_{q}(\overline{\xi}(s))-qW_{q}(\overline{\xi}(s))\right)\mathrm{d}s.\end{aligned}$$
Denote by $V_{\xi}^{\prime}(x;b)$ the derivative of $V_{\xi}(x;b)$ with respect to its first argument. Combining (\[disc.total.cap.\]), (\[l1\]), (\[l2.\]) and (\[l3\]) we have $$\begin{aligned}
\label{l'}
V_{\xi}^{\prime}(x;b)
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\frac{W_{q}^{\prime}(\overline{\xi}(x))}{ W_{q}(\overline{\xi}(x))}V_\xi(x;b)
-\frac{V_\xi(x;b)}{Z_{q+\lambda}(\overline{\xi}(x))}
\left(\frac{W_{q}^{\prime}(\overline{\xi}(x))}{W_{q}(\overline{\xi}(x))}Z_{q}(\overline{\xi}(x))-qW_{q}(\overline{\xi}(x))\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-\left(Z_{q}(\overline{\xi}(x))-\psi^{\prime}(0+)W_{q}(\overline{\xi}(x))-\frac{\overline{Z}_{q}(\overline{\xi}(x))-\psi^{\prime}(0+)\overline{W}_{q}(\overline{\xi}(x))}{W_{q}(\overline{\xi}(x))}W_{q}^{\prime}(\overline{\xi}(x))\right)
\nonumber\\
\hspace{-0.3cm}&&\hspace{-0.3cm}
-\left(-\frac{\psi^{\prime}(0+)}{q+\lambda}+\frac{\overline{Z}_{q+\lambda}(\overline{\xi}(x))+\frac{\psi^{\prime}(0+)}{q+\lambda}}{Z_{q+\lambda}(\overline{\xi}(x))}\right)\left(\frac{W_{q}^{\prime}(\overline{\xi}(x))}{W_{q}(\overline{\xi}(x))}Z_{q}(\overline{\xi}(x))-qW_{q}(\overline{\xi}(x))\right)
\nonumber\\
\hspace{-0.3cm}&=&\hspace{-0.3cm}
\ell_{1}(x)V_\xi(x;b)-\ell_{4}(x).\end{aligned}$$ Solving (\[l’\]) with boundary condition $V_\xi(b, b)=0 $, we obtain (\[expr.of.exp.tot.dis.cap.inj.\]).
Numerical examples {#4}
==================
To exemplify the main results, we discuss some numerical examples for one-sided jump-diffusion process $X$ with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for all $\theta\in\mathbb{R}$ s.t. $\theta\neq -c$. See Example \[ex:Ex1\] for the corresponding scale function. We set $\mu=0.075$, $a=0.5$ and $c=9$ (on average once every two years the firm suffers an instantaneous loss of $10\%$ of its value), and $q=5\%$.
Figures \[fig:exit\] and \[fig:Vxi\] display various shapes of ruin probability $\mathbb{P}_x(\theta_{\lambda}<\infty)$ (\[eq:ruinprobab\]) and the expected nett present value $V_{\xi}(x):=V_{\xi}(x;b=\infty)$ of the total amount of required capital injection to the firm as a function of initial surplus $x$ and the default monitoring frequency $\lambda$. The computation was performed for different forms of drawdown function: $\xi(x)=Kx$, with $0<K<1$, and $\xi(x)=\min\{1,Kx\}$. The former dictates that default is announced and followed by capital injection as soon as the surplus process has crossed below $K\%$ of its last record high, whereas the latter deals with the case of injecting capital to the firm as soon as the surplus is less than one dollar or below $K\%$ of the last record high, whichever is smaller. The results are presented for two cases: $\sigma=0.2$ (the process has paths of unbounded variation) and $\sigma=0$ (paths with bounded variation).
Over all, we observe that the ruin probability $\mathbb{P}_x(\theta_{\lambda}<\infty)$ decreases as the initial surplus increases. This implies that firms with higher initial endowment/surplus has lower probability of ruin than those with lower value of surplus at the beginning. Furthermore, when $\sigma\neq 0$ in which case the firm has additional (immediate) exposure to risk, say from investing in financial market, the ruin probability is higher than those firms which do not have any risk exposure ($\sigma=0$) other than the claim from the insurance holder, which arrives at exponential random time. From the sample paths point of view, the presence of Brownian motion allows the paths to reach a new maximum level and then makes an excursion below that level before a jump arrives. As a result, such movement triggers the observation clock start to run and put the firm into a risky position of getting default. Moreover, we also notice from the figure that the higher the observation frequency $\lambda$, the higher the ruin probability. This is to say that the larger $\lambda$ gives the firm lesser time to come out of dilution period during which the firm is in financial distress, resulting in a higher chance of going ruin. The choice of drawdown level $\xi(x)$ also determines the shape of the ruin probability. Under the drawdown level $Kx$, the ruin probability is higher for larger value of $K$. This is due to the fact that the higher value of $K$ sets the default level higher causing the firm to go default/ruin sooner than lower value of $K$, in particular when $\sigma\neq 0$. As $\min\{1,Kx\}\leq Kx$, default is expected to occur earlier under drawdown level $\xi(x)=Kx$ than $\xi(x)=\min\{1,Kx\}$ for the same reason explained. For the latter, the ruin probability decreases linearly when the surplus is about less than one unit, then decreases at exponential rate when the surplus is larger than that value. Similar observation is observed for $\sigma=0$ with only exception that the curve has lower degree of smoothness than the case $\sigma\neq 0$ in which case the scale function is twice continuously differentiable over $(0,\infty)$.
Similar feature is exhibited by the expected total amount of discounted cash for capital injection function $V_{\xi}(x)$. Healthier firm with larger initial surplus $x$ requires less capital injection in total than that of unhealthier firm with lower surplus. When the firm has more uncertainties as a result of investing in financial market, the case $\sigma\neq 0$, the firm requires more capital injection than those firm which do not have riskier exposure ($\sigma=0$) other than the claim from the insurance holder arriving in exponential random time. As explained above, the presence of Brownian motion puts the firm in riskier situation with higher probability of ruin when the surplus process makes an excursion from its last record high. This induces the firm to ask for more financial coverage (injection) to deal with during the dilution period. Moreover, the longer the observation window (the lower the value of $\lambda$), the firm receives more capital injection from stakeholder than otherwise. The choice of drawdown level $\xi(x)$ also shifts the curve $\xi(x)$. For $\xi(x)=Kx$, the curve is higher for larger value of $K$. The upward shift of the curve is attributed by the fact that the larger value of $\lambda$ increases the default threshold to higher level giving the firm higher chance of default resulting the firm in asking for more capital to prevent ruin to occur at earlier stage, in particular when $\sigma\neq 0$. As $\min\{1,Kx\}\leq Kx$, default is expected to occur later for $\xi(x)=\min\{1,Kx\}$ than that of under the drawdown level $\xi(x)=Kx$ which in turn requiring lesser amount of capital injection. Under $\xi(x)=\min\{1,Kx\}$, the function $V_{\xi}(x)$ decreases linearly with quite high slope (rate) when the surplus is about less than one unit, then decreases exponentially at lower rate when the surplus is larger than that value. Overall, unlike the ruin probability $\mathbb{P}_x(\theta_{\lambda}<\infty)$, the function $V_{\xi}(x)$ retains its convexity in all cases having zero value at infinity implying the function to be positive for all $x$.
The above analyses conclude our numerical study on the ruin probability and nett present value of the capital injection which summarizes their various shape w.r.t changing the value of observation frequency $\lambda$ and initial surplus $x$.
[cc]{}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Ruin probability $\mathbb{P}_x(\theta_{\lambda}<\infty)$ for downward jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:exit"}](pdefubv1.pdf "fig:"){width="46.50000%"}
\#2
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
&
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Ruin probability $\mathbb{P}_x(\theta_{\lambda}<\infty)$ for downward jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:exit"}](pdefbv1.pdf "fig:"){width="46.50000%"}
\#2
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Ruin probability $\mathbb{P}_x(\theta_{\lambda}<\infty)$ for downward jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:exit"}](pdefubv2.pdf "fig:"){width="46.50000%"}
\#2
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
&
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Ruin probability $\mathbb{P}_x(\theta_{\lambda}<\infty)$ for downward jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:exit"}](pdefbv2.pdf "fig:"){width="46.50000%"}
\#2
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Ruin probability $\mathbb{P}_x(\theta_{\lambda}<\infty)$ for downward jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:exit"}](pdefubv2b.pdf "fig:"){width="46.50000%"}
\#2
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
&
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Ruin probability $\mathbb{P}_x(\theta_{\lambda}<\infty)$ for downward jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:exit"}](pdefbv2b.pdf "fig:"){width="46.50000%"}
\#2
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
[cc]{}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Expected capital injection $V(x):=V_{\xi}(x;\infty)$ for jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:Vxi"}](capubv1.pdf "fig:"){width="46.50000%"}
\#2
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
&
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Expected capital injection $V(x):=V_{\xi}(x;\infty)$ for jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:Vxi"}](capbv1.pdf "fig:"){width="46.50000%"}
\#2
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Expected capital injection $V(x):=V_{\xi}(x;\infty)$ for jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:Vxi"}](capubv2.pdf "fig:"){width="46.50000%"}
\#2
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
&
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Expected capital injection $V(x):=V_{\xi}(x;\infty)$ for jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:Vxi"}](capbv2.pdf "fig:"){width="46.50000%"}
\#2
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Expected capital injection $V(x):=V_{\xi}(x;\infty)$ for jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:Vxi"}](capubv2b.pdf "fig:"){width="46.50000%"}
\#2
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
&
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Expected capital injection $V(x):=V_{\xi}(x;\infty)$ for jump-diffusion process with Laplace exponent $\psi(\theta)=\mu\theta+\frac{\sigma^2}{2}\theta^2-\frac{a\theta}{\theta+c}$ for $\mu=0.075, a=0.5, c=9, q=0.05$.[]{data-label="fig:Vxi"}](capbv2b.pdf "fig:"){width="46.50000%"}
\#2
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
Conclusion {#5}
==========
This paper presents some distributional identities concerning excursion below the last record high of surplus process, driven by downward jumps Lévy process, with capital injection. Capital injection is provided to the firm as soon as the process goes below a drawdown level and is continuously paid until the process goes above the record or ruin occurs, which is announced at the first time the process has undertaken an excursion below the record high longer than an independent exponential period of time. Identities are given explicitly in terms of the scale function of the Lévy process. The latter makes possible to have a fast numerical computation of the identities and do analysis on the impact of observation frequency and initial surplus to the ruin probability and the expected nett present value of the required total capital injection. The choice of some drawdown functions was made to study various shapes of the ruin probability and the nett present value of the capital injection. Numerical study shows that the results implied by the model is found to be consistent with an observation one would have in financial market. We leave this for further research.
Acknowledgements {#acknowledgements .unnumbered}
================
Wenyuan Wang and Xianghua Zhao thank Concordia University where the first draft of this paper was finished during their visits. Wenyuan Wang acknowledges the support of the National Natural Science Foundation of China (No. 11601197). Wenyuan Wang and Xiaowen Zhou are supported by a National Sciences and Engineering Research Council of Canada grant (No. RGPIN-2016-06704). Xiaowen Zhou is supported by National Natural Science Foundation of China (No. 11771018). Part of this work was carried out while Budhi Surya was visiting Department of Satistics of New York University Stern School of Business in September 2019. He acknowledges Faculty Strategic Research Grant No. 20859 of Victoria University and hospitality provided by the NYU Stern.
[99]{}
Agarwal, V., Daniel, N. and Naik, N. 2009. Role of managerial incentives and discretion in hedge fund performance. *Journal of Finance*, 64(5), 2221-2256.
Albrecher, H., Ivanovs, J. and Zhou, X., 2016. Exit identities for Lévy processes observed at Poisson arrival times. *Bernoulli*, [**22(3)**]{}, 1364-1382.
Antill, S. and Grenadier, S., 2019. Optimal capital structure and bankruptcy choice: Dynamic bargaining versus liquidation. *Journal of Financial Economics*, [**133(1)**]{}, 198-224.
Asmussen, S., 1989. Applied Probability and Queues, *Wiley series in probability*.
Avanzi, B., Shen, J. and Wong, B., 2011. Optimal dividends and capital injections in the dual model with diffusion. *ASTIN Bulletin: The Journal of the IAA* [**41**]{}(2), 611-644.
Avram, F., Palmowski, Z. and Pistorius, M., 2007. On the optimal dividend problem for a spectrally negative Lévy process. [*Annals of Applied Probability*]{} [**17,**]{} 156–180.
Avram, F., Kyprianou, A. and Pistorius, M., 2004. Exit problems for spectrally negative Lévy processes and applications to (Canadized) Russian options. *Annals of Applied Probability* [**14,**]{} 215-238.
Avram, F. Vu, N. and Zhou, X., 2017. On taxed spectrally negative Lévy processes with draw-down stopping. *Insurance, Mathematics and Economics* [**76**]{}, 69-74.
Baurdoux, E., Pardo, J., Pérez, J. and Renaud, J., 2016. Gerber-Shiu distribution at Parisian ruin for Lévy insurance risk processes. *Journal of Applied Probability* [**53**]{}, 572-584.
Baurdoux, E., 2007. Fluctuation theory and stochastic games for spectrally negative Lévy processes. Doctoral Thesis, de Universiteit Utrecht.
Baurdoux, E. and Kyprianou, A. 2008. The McKean stochastic game driven by a spectrally negative Lévy process. *Electronic Journal of Probability*, [**13 (8),**]{} 173-197.
Bayraktar, E., Kyprianou, A. and Yamazaki, K., 2013. On optimal dividends in the dual model. *ASTIN Bulletin: The Journal of IAA* [**43**]{}(3), 359-372.
Bertoin, J., 1996. Lévy Processes. *Cambridge University Press*.
Bertoin, J., 1997. Exponential decay and ergodicity of completely asymmetric Lévy processes in a ?nite interval. *Annals of Applied Probability*, [**7**]{}, 156-169.
Borodin, A. and Salminen, P., 2012. Handbook of Brownian motion-facts and formulae. *Birkh?user*.
Borovkov, A., 1976. Stochastic Processes in Queueing Theory, *Springer-Verlag*.
Broadie M., Chernov M. and Sundaresan S., 2007. Optimal debt and equity values in the presence of Chapter 7 and Chapter 11. *Journal of Finance*, LXII (3), 1341-1377.
Chan, T., Kyprianou, A. and Savov, M., 2011. Smoothness of scale functions for spectrally negative Lévy processes. *Probability Theory and Related Fields*, 150 (3-4), 691-708.
Chesney, M., Jeanblanc-Picqué M. and Yor, M., 1997. Brownian excursions and Parisians options. *Advances in Applied Probability*, 29, 165-184.
Choi, B. and Roh, J., 2013. On the trivariate joint distribution of Brownian motion and its maximum and minimum. *Statistics and Probability Letters*, 83(4), 1046-1053.
Czarna, I. and Palmowski, Z. 2011. Ruin probability with Parisian delay for a spectrally negative Lévy process. *Journal of Applied Probability* **48**, 984-1002.
Dassios, A. and Wu, S., 2009. Parisian ruin with exponential claims. *Working paper, LSE London*. Available at http://stats.lse.ac.uk/angelos.
De Finetti, 1957. Su un’impostazione alternativa della teoria colletiva del rischio. *Trans. XV Intern. Congress Act.* [**2,**]{} 433-443.
Dickson, D. and Waters, H., 2004. Some optimal dividends problems. *ASTIN Bulletin*, [**34 (1)**]{}, 49-74.
Francois P. and Morellec E., 2004. Capital structure and asset prices: Some effects of bankruptcy procedures. *Journal of Business*, 77, 387-411.
Gerber, H., 1990. When does the surplus reach a given target? *Insurance Mathematics and Economics* [**9,**]{} 115-119.
Goetzmann, W., Ingersoll Jr, J., and Ross, S., 2003. High-water marks and hedge fund management contracts. *Journal of Finance* [**58**]{}, 1685-1717.
Hilberink, B. and Rogers, L., 2002. Optimal capital structure and endogenous bankruptcy. *Finance and Stochastics* [**6**]{}, 237-263.
Hubalek, F. and Kyprianou, E., 2011. Old and new examples of scale functions for spectrally negative Lévy processes. *Seminar on stochastic analysis, random fields and applications VI. Springer, Basel*, 119-145.
Kuznetsov, A., Kyprianou, A. and Rivero, V. 2012. The theory of scale functions for spectrally negative Lévy processes. In Lévy Matters II (Lecture Notes Math. 2061), *Springer*, Heidelberg, 97-186.
Kulenko, N. and Schmidli, H. 2008. Optimal dividend strategies in a Cramér-Lundberg model with capital injections. *Insurance: Mathematics and Economics* [**43**]{}, 270-278.
Kyprianou, A., 2006. Introductory Lectures on Fluctuations of Lévy Processes with Applications. *Springer*, Berlin.
Kyprianou, A. and Pistorius, M., 2003. Perpetual options and Canadization through fluctuation theory. *Annals of Applied Probability* [**13,**]{} 1077-1098.
Kyprianou, A. and Surya, B., 2007. Principles of smooth and continuous fit in the determination of endogeneous bankruptcy levels. *Finance and Stochastics* [**11**]{}, 131-152.
Kyprianou, A. and Zhou, X., 2009. General tax structures and the Lévy insurance risk model. *Journal of Applied Probability*, [**46**]{}, 1146-1156.
Li, B., Vu, N. and Zhou, X. 2019. Exit problems for general draw-down times of spectrally negative L[é]{}vy processes. *Journal of Applied Probability*, 56, 441-457.
Lehoczky, J., 1977. Formulas for stopped diffusion processes with stopping times based on the maximum. *Annals of Applied Probability* [**5,**]{} 601-607.
Leland, H. and Toft, K., 1996. Optimal capital structure, endogenous bankruptcy, and the term structure of credit spreads. *Journal of Finance* [**51**]{}, 987-1019.
Loeffen, R., 2008. On optimality of the barrier strategy in de Finetti’s dividend problem for spectrally negative Lévy processes. *The Annals of Applied Probability*, [**18**]{}, 1669-1680.
Loeffen, R., Czarna, I. and Palmowski, Z., 2013. Parisian ruin probability for spectrally negative Lévy processes. *Bernoulli* **19**(2), 599-609.
Loeffen, R., Palmowski, Z. and Surya, B., 2018. Discounted penalty function at Parisian ruin for Lévy insurance risk process. *Insurance: Mathematics and Economics*, 83, 190-197.
Palmowski, Z. and Surya, B., 2020. Optimal valuation of American callable credit default swaps under drawdown of Lévy insurance risk process. Forthcoming in *Insurance: Mathematics and Economics*.
Palmowski, Z., Pérez, J., Surya, B. and Yamazaki, K., 2020. The Leland-Toft optimal capital structure model under Poisson observations. Forthcoming in *Finance and Stochastics*.
Pistorius, M., 2004. On exit and ergodicity of the spectrally one-sided Lévy process reflected at its infimum. *Journal of Theoretical Probability*, [**17(1)**]{}, 183-220.
Pistorius, M., 2007. An excursion-theoretical approach to some boundary crossing problems and the skorokhod embedding for refrected Lévy processes. In *Séminaire de Probabilités* XL, 287-307. *Springer*.
Prabhu, N., 1997. Insurance, Queues, Dams. *Springer-Verlag*.
Shepp, L. and Shiryaev, A., 1994. A new look at pricing the “Russian option”. *Theory of Probability and Applications* 39, 103-119.
Surya, B., 2019. Parisian excursion below a fixed level from the last record maximum of Lévy insurance risk process. *In: Wood D., de Gier J., Praeger C., Tao T. (eds) 2017 MATRIX Annals. MATRIX Book Series*, Vol 2, Springer Nature, 311-326.
Surya, B., 2008. Evaluating scale functions of spectrally negative Lévy processes. *Journal of Applied Probability*, **45**, 135-149.
Wang, W., Wang, Y., and Wu, X., 2019. Dividend and capital injection optimization with transaction cost for spectrally negative Lévy risk processes. *arXiv:1807.11171*
Wang, W. and Zhou, X., 2018. General draw-down based de Finetti optimization for spectrally negative Lévy risk processes. *Journal of Applied Probability*, [**55(2)**]{}, 513-542.
Wang, W. and Zhou, X., 2019. A draw-down reflected spectrally negative Lévy process. [*Journal of Theoretical Probability*]{}, 1-24.
Yao, D., Yang, H., and Wang, R. 2011. Optimal dividend and capital injection problem in the dual model with proportional and fixed transaction costs. *European Journal of Operational Research* [**211**]{}(3), 568-576.
Zhou, X., 2007. Exit problems for spectrally negative Lévy processes reflected at either the supremum or the infimum. *Journal of Applied Probability*, [**44**]{}, 1012-1030.
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.